Numerical Approaches To Combustion Modelling

Numerical
Approaches
to
Combustion
Modeling
Edited by
Elaine S. Oran and Jay P. Boris
Naval Research Laboratory
Washington, DC
Volume 135
PROGRESS IN
ASTRONAUTICS AND AERONAUTICS
A. Richard Seebass, Editor-in-Chief

University of Colorado at Boulder
Boulder, Colorado
Published by the American Institute of Aeronautics and Astronautics, Inc.

370 L'Enfant Promenade, SW, Washington, DC, 20024-2518
Data and information appearing in this Book are for informational purposes only.
AIAA is not responsible for any injury or damage resulting from use or reliance, nor
does AIAA warrant that use or reliance will be free from privately owned rights.
American Institute of Aeronautics and Astronautics, Inc.

Washington, DC
Library of Congress Cataloging-in-Publication Data
Numerical approaches to combustion modeling / edited by Elaine S. Oran
and Jay P. Boris,
p. cm. — (Progress in astronautics and aeronautics ; v. 135)
Includes bibliographical references.
ISBN 1-56347-004-7
1. Combustion—Computer simulation. I. Oran, Elaine S.
II. Boris, Jay P. III. Series.
QD516.N84 1991 621.402'3—dc20 91-17893
Copyright © 1991 by the American Institute of Aeronautics and Astronautics, Inc.
All rights reserved. Reproduction or translation of any part of this work beyond
that permitted by Sections 107 and 108 of the U.S. Copyright Law without the
permission of the copyright owner is unlawful. The code following this statement
indicates the copyright owner's consent that copies of articles in this volume may
be made for personal or internal use, on condition that the copier pay the per-
copy fee ($2.00) plus the per-page fee ($0.50) through the Copyright Clearance
Center, Inc., 21 Congress Street, Salem, Mass. 01970. This consent does not extend
to other kinds of copying, for which permission requests should be addressed to
the publisher. Users should employ the following code when reporting copying
from this volume to the Copyright Clearance Center:
1-56347-004-7/91 $2.00+.50
Progress in Astronautics and Aeronautics
Editor-in-Chief
A. Richard Seebass
University of Colorado at Boulder
Editorial Board
Richard G. Bradley John L. Junkins
General Dynamics Texas A&M University
John R. Casani John E. Keigler

California Institute of Technology General Electric Company
Jet Propulsion Laboratory Astro-Space Division
Alien E. Fuhs Daniel P. Raymer

Carmel, California Lockheed Aeronautical Systems
Company
George J. Gleghorn
TR W Space Joseph F. Shea
and Technology Group Massachusetts Institute
of Technology
Dale B. Henderson
Los Alamos National Laboratory Martin Summerfield
Princeion Combustion Research
Carolyn L. Huntoon Laboratories, Inc.
NASA Johnson Space Center
Charles E. Treanor
Reid R. June Arvin/Calspan
Boeing Military Airplane Company Advanced Technology Center
Norma J. Brennan
Director, Editorial Department
AIAA
Jeanne Godette
Series Managing Editor
AIAA
This page intentionally left blank
In memory of our colleagues, Pavel Chushkin, Jace Nunziato,
and Fritz Krause
Table of Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xix
Part 1. Chemistry in Combustion Modeling
Background
Chapter 1. Ab Initio Quantum Chemistry for Combustion . . . . . . . . . . . 3

Michael Page, Naval Research Laboratory, Washington, DC, and Byron H.
Lengsfield III, Lawrence Livermore Laboratory, Livermore, California
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
Electronic Structure Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .5
Bond Breaking and the MCSCF Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . .6
Configuration Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9
Basis Sets. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Perturbation Methods and Bond-Additive-Correction Schemes . . . . . . . . . . . 11
Potential Energy Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Born-Oppenheimer Potential Energy Derivatives. . . . . . . . . . . . . . . . . . . . . . 15
Locating Critical Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Reaction Path Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Reaction Path Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Following Reaction Pathways . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .22
Computational Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 3
Architectural Considerations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 3
Theoretical Developments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 7
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 0
Chapter 2. Rate Coefficient Calculations for Combustion

Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 7
Nancy J. Brown, Lawrence Berkeley Laboratory, University of California, Berkeley,
California
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 7
Potential Energy Surfaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Transition State Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39
Theoretical Considerations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .39
Application of TST . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Symmetry Corrections for Rate Coefficient Calculations . . . . . . . . . . . . . . . .44
VII
Tunneling C o r r e c t i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Complex Reactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Molecular Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .47
Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Chapter 3. Numerical Modeling of Combustion of Complex

Hydrocarbon.................................................57
Charles K. Westbrook and William J. Pitz, Lawrence Livermore Laboratory,
Livermore, California
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
Autoignition and Engine Knock . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Temperature H i s t o r y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
Fuel Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Comprehensive Reaction Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
High- and Intermediate-Temperature Ignition . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Chain Branching Mechanisms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Detailed Reaction Paths . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Extensions to Low T e m p e r a t u r e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Examples of Model Tests in Different Environments. . . . . . . . . . . . . . . . . . . . .72
Shock Tubes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
Flow Reactors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Low-Temperature Reactors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 4
Knocking E n g i n e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Chapter 4. Combustion Kinetics and Sensitivity Analysis

Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 3
K. Radhakrishnan, NASA-Lewis Research Center, Cleveland, Ohio
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Combustion Kinetics Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
Computational Difficulty—Stiffness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Sensitivity Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Scope of Chapter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
Methods for Stiff ODE's. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 1
Linear Multistep Methods: Backward Differentiation Formulas. . . . . . . . . . .91
Runge-Kutta M e t h o d s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Extrapolation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .94
Hybrid Methods: Selected Asymptotic Integration Method . . . . . . . . . . . . . .96
Exponential Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Published Numerical Comparisons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Local Sensitivity Analysis Methods. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Brute Force Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Direct M e t h o d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Green's Function Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
viii
Decoupled Direct Method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
Other Local Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ill
Comparison of Local Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1 1
Test Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
Concluding Remarks. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Chapter 5. Reduction of Chemical Reaction Models . . . . . . . . . . . . . . . 1 2 9

Michael Frenklach, Pennsylvania State University, State College, Pennsylvania
Introduction: Why Reduction? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Modeling S t r a t e g y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Global M o d e l i n g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Empirical Fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Reduction by Approximations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Lumping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
Response Modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
Detailed R e d u c t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Chemical Lumping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
Statistical Lumping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Summary: Use of Reduction in Combustion Modeling . . . . . . . . . . . . . . . . . . 148
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
Part 2. Flames and Flame Structure
Background
Chapter 6. Length Scales in Laminar and Turbulent Flames . . . . . . . 1 5 5

Norbert Peters, RWTH Aachen, Aachen, Germany
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
Length Scales in Laminar F l a m e s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Premixed F l a m e s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
Diffusion Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160
Length Scales in Turbulent Flames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Turbulent Premixed Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
Turbulent Diffusion Flames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
Chapter 7. Numerical Modeling of Laminar Diffusion Flames . . . . . . 1 8 3

Mitchell D. Smooke, Yale University, New Haven, Connecticut
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
Counterflow Diffusion Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Flame-Sheet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
Method of Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
Double-Jet Counterflow Diffusion Flame . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
ix
Axisymmetric Coflow Diffusion Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
Flame-Sheet Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
Method of Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
Confined Axisymmetric Laminar Diffusion Flame . . . . . . . . . . . . . . . . . . . . 214
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
Chapter 8. Laminar Flames in Premixed G a s e s . . . . . . . . . . . . . . . . . . . 2 2 5

K. Kailasanath, Naval Research Laboratory, Washington, DC
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 5
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 6
Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 2 6
Modeling of Laminar Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
Numerical Approaches to Modeling of Flames. . . . . . . . . . . . . . . . . . . . . . . . . 229
Physical and Chemical Processes to be Modeled . . . . . . . . . . . . . . . . . . . . . 229
Basic Numerical Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
Different Numerical Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
Description of Two Detailed Flame M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
FLAME1D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
FLIC2D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Input Parameters to Detailed Flame Models . . . . . . . . . . . . . . . . . . . . . . . . 237
Numerical Studies of Laminar Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
Steady-State Flames and the Determination of Burning Velocities . . . . . . . 238
Ignition and Quenching of Premixed Gases . . . . . . . . . . . . . . . . . . . . . . . . . 240
Burning Velocity of Curved and Unsteady F l a m e s . . . . . . . . . . . . . . . . . . . . 242
Effects of Curvature and Dilution of Flame Propagation . . . . . . . . . . . . . . . 244
Flammability Limits. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
Flame Instabilities and the Cellular Structure of Flames . . . . . . . . . . . . . . . 248
Effects of Gravity on Multidimensional Flame Structure . . . . . . . . . . . . . . . 250
Current Status and Future Needs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 5 2
Chapter 9. Spectral Simulations of Turbulent Reacting F l o w s . . . . . . . 2 5 7

Patrick A. McMurtry, University of Utah, Salt Lake City, Utah, and Peyman
Givi, State University of New York at Buffalo, Amherst, New York
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
Spectral Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
Spectral Galerkin Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
Spectral Collocation (Pseudospectral) Method . . . . . . . . . . . . . . . . . . . . . . . 263
Spectral Elements M e t h o d s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
Chemical Reactions in Homogeneous Turbulence . . . . . . . . . . . . . . . . . . . . 273
Chemically Reacting Mixing Layers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
Concluding R e m a r k s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
Chapter 10. Vortex Simulation of Reacting Shear Flow . . . . . . . . . . . . 3 0 5
Ahmed F. Ghoniem, Massachusetts Institute of Technology, Cambridge,
Massachusetts
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Numerical Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
Vortex Element Method in Two D i m e n s i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Numerical S c h e m e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
Primary Shear Flow Instabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
Vortex Element Method in Three Dimensions. . . . . . . . . . . . . . . . . . . . . . . . . 313
Propagation and Instability of Vortex R i n g s . . . . . . . . . . . . . . . . . . . . . . . . . 315
Secondary Instabilities of Shear Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
Transport Element Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 322
Scalar Mixing in Shear Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
Entrainment Enhancement due to Three-Dimensional Instabilities . . . . . . . 325
Vortex Methods for Variable-Density Flows . . . . . . . . . . . . . . . . . . . . . . . . . . 326
Vortex Methods for Reacting F l o w s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
Reacting Shear Layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 332
Three-Dimensional Reacting Shear L a y e r s . . . . . . . . . . . . . . . . . . . . . . . . . . 333
Reacting J e t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
Premixed Shear L a y e r . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 4 4
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
Chapter 11. Combustion Modeling Using PDF M e t h o d s . . . . . . . . . . . . 3 4 9

S. B. Pope, Cornell University, Ithaca, New York
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 349
PDF Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Definitions and Properties. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
Composition Joint PDF Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 352
Velocity-Composition Joint PDF E q u a t i o n . . . . . . . . . . . . . . . . . . . . . . . . . . 353
Lagrangian V i e w p o i n t . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Stochastic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
Numerical Solution Algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
Monte Carlo Method for the Velocity-Composition Joint PDF . . . . . . . . . . 356
Monte Carlo Algorithms for the Composition Joint PDF. . . . . . . . . . . . . . . 357
Turbulent Flame C a l c u l a t i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Turbulent Diffusion Flames. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
Turbulent Premixed Flames . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
Discussion and Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 360
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 362
XI
Part 3. High-Speed Reacting Flows
Background
Chapter 12. Supersonic Reacting Internal Flow Fields . . . . . . . . . . . . . 3 6 5

J. Phillip Drummond, NASA-Langley Research Center, Hampton, Virginia
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Governing Equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
Thermodynamics Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 372
Chemistry Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 373
Molecular Diffusion Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374
Turbulent Diffusion Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
Discretization of the Governing E q u a t i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Numerical A l g o r i t h m s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
Conventional Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
Alternate Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 415
Chapter 13. Studies of Detonation Initiation, Propagation, and

Quenching..................................................421
E. S. Oran, J. P. Boris, and K. Kailasanath, Naval Research Laboratory,
Washington, DC
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421
Numerical C o n s i d e r a t i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 423
Coupling Convection and C h e m i s t r y . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
Convection Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424
Solving the Chemistry Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 427
Solving the Temperature Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429
Phenomenologies for Energy Release . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
Variable and Adaptive G r i d d i n g . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 430
Weak and Strong Ignition Behind Reflected Shocks . . . . . . . . . . . . . . . . . . . . 432
Multidimensional Detonation Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
Detonation Cells and Unreacted Pockets . . . . . . . . . . . . . . . . . . . . . . . . . . . 439
Irregularity in the Structure of Detonation Waves . . . . . . . . . . . . . . . . . . . . 440
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 442
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
Chapter 14. Numerical Modeling of Heterogeneous Detonations.... .447

Martin Sichel, University of Michigan, Ann Arbor, Michigan
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 447
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 448
xii
Structure of Heterogeneous Detonations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451
Detonations in Sprays, ZND Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 453
Dust Detonations, ZND Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 457
Blast Wave Models of Detonation Structure. . . . . . . . . . . . . . . . . . . . . . . . . 462
Direct Initiation of Heterogeneous Detonations. . . . . . . . . . . . . . . . . . . . . . . . 466
Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 6 6
Shock-Fitting Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 468
Shock-Capturing M e t h o d s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 473
Discussion and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 476
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 7 7
Chapter 15. Deflagration-to-Detonation Transition in Reactive

Granular Materials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 8 1
Melvin R. Baer and Jace W. Nunziato, Sandia National Laboratories,
Albuquerque, New Mexico, and Pedro F. Embid, University of New Mexico,
Albuquerque, New Mexico
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 8 1
Theory for Reactive Two-Phase Flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 483
Equations of Motion and Thermodynamics . . . . . . . . . . . . . . . . . . . . . . . . . 483
Constitutive Models and Closure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 485
One-Dimensional Initial Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 489
Field Equations and Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . 489
Mathematical Structure of Reactive Multiphase Flow Equations . . . . . . . . . 490
Numerical Methods in One D i m e n s i o n . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 492
DDT in Granular Explosive HMX . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 497
Two-Dimensional Computations of Flame Propagation . . . . . . . . . . . . . . . . . . 503
Unresolved Technical Issues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 506
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 508
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 509
Chapter 16. Toward a Microscopic Theory of Detonations

in Energetic Crystals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1 3
Michel Peyrard, Laboratoire ORC, Dijon, France and Simone Odiot,
Universite de Pierre and Marie Curie, Paris France
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 513
First Molecular Dynamics Investigations of Energetic Materials . . . . . . . . . . . 514
Molecular Dynamics Investigations of Shock Waves in Crystals. . . . . . . . . . 514
First Investigations of Energetic Materials . . . . . . . . . . . . . . . . . . . . . . . . . . 516
Role of the Crystal Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
One-Dimensional M o d e l . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 517
Two-Dimensional M o d e l . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 523
Effects of the Chemistry . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 531
New Approach: Microscopic Modeling. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 536
Discussion and Conclusion: Toward a Microscopic Theory . . . . . . . . . . . . . . . 538
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 540
xiii
Part 4. (Even More) Complex Combustion Systems
Background
Chapter 17. An Overview of Spray M o d e l i n g . . . . . . . . . . . . . . . . . . . . . 5 4 3

Clayton T. Crowe, Washington State University, Pullman, Washington
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 545
Chapter 18. Liquid Drop Behavior in Dense and Dilute Clusters . . . .547
Josette Bellan, Jet Propulsion Laboratory, Pasadena, California
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 547
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 548
Survey of Drop Interaction Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 551
Models of Drop Evaporation, Ignition, and Combustion Based on
Multiple-Drop Interactions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 555
Constant-Volume (Variable-Pressure) Cluster Models . . . . . . . . . . . . . . . 557
Constant-Pressure (Variable-Volume) Cluster Models . . . . . . . . . . . . . . . 565
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 577
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 580
Chapter 19. Spray Combustion in Idealized Configurations:

Parallel Drop S t r e a m s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 8 5
R. H. Rangel and W. A. Sirignano, University of California, Irvine,
Irvine, California
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 585
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 586
Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Comprehensive Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 588
Nonvaporizing Case: Analytical Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 591
Vaporizing Spray. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 596
Steady Reactive Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 598
Unsteady Reactive Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 606
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 610
Future Challenges and Concluding Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . 611
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 612
Chapter 20. Comparisons of Deterministic and Stochastic

Computations of Drop Collisions in Dense Sprays . . . . . . . . . . . . . . . . 6 1 5
J. M. Maclnnes and F. V. Bracco, Princeton University, Princeton, New Jersey
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 615
Calculation Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
Stochastic M e t h o d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 616
Deterministic Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 621
Stochastic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 624
Selected Spray and Earlier Comparisons with Experiments . . . . . . . . . . . . . 624
xiv
Stochastic Results with Calculated Gas Field . . . . . . . . . . . . . . . . . . . . . . . . 627
Stochastic Results with Fixed Gas Field . . . . . . . . . . . . . . . . . . . . . . . . . . . . 632
Deterministic Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634
Comparison of Deterministic and Stochastic Results . . . . . . . . . . . . . . . . . . 635
Higher Resolution Computations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 637
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 640
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 641
Chapter 21. Ignition and Flame Spread Across Solid Fuels . . . . . . . . . 6 4 3

Colomba Di Blasi, Universita di Napoli, Napoli, Italy
Physical Description of P r o b l e m . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 643
Classification of Ignition and Flame-Spread Problems . . . . . . . . . . . . . . . . . . . 644
Modeling Solid-Fuel Combustion Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . 646
Gas-Phase M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 647
Solid-Phase M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 650
Thermally Thick Fuels. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Thermally Thin Fuels . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 651
Initial and Boundary Conditions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 652
Finite-Difference Approximations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 654
Numerical Solution Procedures. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 655
Steady Gas-Phase Equations in Thermodiffusive Models . . . . . . . . . . . . . . . 656
Solution of Momentum Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 656
Numerical Simulations of Ignition and Flame Spread . . . . . . . . . . . . . . . . . . . 657
Ignition and Opposed-Flow Flame Spread . . . . . . . . . . . . . . . . . . . . . . . . . . 657
Flow-Assisted Flame Spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 664
Conclusions and Further Developments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 666
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 669
Chapter 22. Pulse Combustor Dynamics: A Numerical Study . . . . . . . 6 7 3

Pamela K. Barr, Sandia National Laboratories, Livermore, California and
Harry A. Dwyer, University of California, Davis, California
Nomenclature . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 7 3
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 7 4
Problem Formulation, Basic Equations, and Numerical M e t h o d . . . . . . . . . . . 677
Governing Equations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 677
Numerical Method of Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 678
Friction, Heat-Transfer, and Energy-Release Models. . . . . . . . . . . . . . . . . . 680
Boundary Conditions—Basic Inflow and Outflow . . . . . . . . . . . . . . . . . . . . 681
Valve Dynamics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Models for Industrial Furnace Decouplers . . . . . . . . . . . . . . . . . . . . . . . . . . 683
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
Simulation of a Pulse Combustor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 686
Rayleigh's Criterion and Unsteady Energy Release . . . . . . . . . . . . . . . . . . . 692
Use of Rayleigh's Criterion to Determine the Optimal Design
Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 694
Combustion Chamber Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 696
Tail-Pipe Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 698
xv
Influence of an Exhaust Decoupler. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 702
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 709
Chapter 23. Mathematical Modeling of Enclosure Fires . . . . . . . . . . . . 7 1 1

Henri E. Mitler, National Institute of Standards and Technology,
Gaithersburg, Maryland
Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 711
Reasons for Fire Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 712
Brief History of Fire Science and Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 712
General Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
Physics of F i r e . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 713
Field Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 715
Turbulence Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 716
Solving the Model Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 719
General Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 720
Solving the Difference Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 721
Radiation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
Calculations in Three D i m e n s i o n s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725
General-Purpose Programs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Zone M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Description of M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 726
Formulation of the Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 728
Structure of the Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 732
Numerical Methods for Solving the Model Equations . . . . . . . . . . . . . . . . . 733
Harvard Mark 5. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 735
FIRST Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 743
Alternative M e t h o d . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
Harvard Mark 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 744
Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 746
Chapter 24. Nuclear S y s t e m s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 5 5

J. R. Travis, W. S. Gregory, and F. R. Krause, Los Alamos National Laboratory,
Los Alamos, New Mexico
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 755
Overview of Numerical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Traditional Fire Models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 757
Desired Capabilities for Nuclear Facility Models . . . . . . . . . . . . . . . . . . . . . 760
Available Advanced Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 762
Nonreactor Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 763
Numerical Method and M o d e l s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 764
Applications of Heat-Transfer and Material-Transport Experiments . . . . . . 772
xvi
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785
Nuclear Reactor Containment Systems. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785
Mathematical and Physical Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 787
Computational Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 795
Application to Full-Scale Fire Experiments . . . . . . . . . . . . . . . . . . . . . . . . . 798
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
Research Directions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
Nonreactor Combustion Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 808
Reactor Combustion Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 809
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1 1
Author Index for Volume 135 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815

List of Series V o l u m e s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8 1 7
XVII
Preface
Perhaps the most appealing intellectual reason for studying combustion is
the wide range of phenomena and interactions that can honestly be called
"crucial to combustion." Combustion research covers everything from nu-
clear processes, to radiation transport, to macroscopic fluid dynamics. It
encompasses complex multiphase processes that occur in diesel engines,
room fires, and even in supernova explosions. The more microscopic stud-
ies investigate the actual quantum chemical dynamics of reacting molecular
species leading to the macroscopic models which represent chemical energy
conversion—the change in the macroscopic energy balance through the
nuclear, atomic, or molecular reactions that absorb or release energy. The
second level of complication involves the macroscopic interaction of the
energy released with the background fluids and solid structures in the
vicinity. When chemical energy release is slow enough, the result is usually
a subsonic reaction front such as a flame or deflagration. When the energy
release is fast or the flow is fast, shock and detonation waves become
important parts of the system.
Effective research in any field entails the most modern tools. In com-
bustion, these tools include powerful new theoretical and computational
capabilities as well as the development of new chemical and physical di-
agnostic techniques. In particular, over the last two decades computation
has become an independent way of pursuing knowledge about physical
systems. The computational approach has certain aspects that are common
to both theory and experiment, but it differs appreciably from each in
critical ways. This book spans the breadth of computational research in
combustion by describing how the latest approaches have been applied to
various levels of combustion and reactive-flow problems. What can be
simulated about combustion systems has expanded exponentially, reflect-
ing the growth of computational capability and the general availability of
large-scale scientific computers.
The technical advances that have so greatly expanded what we can cal-
culate also cover a wide range. New numerical algorithms are faster and
more accurate than in the past, and new software allows us to use computers
more easily. User-friendly graphics packages permit easy display and anal-
ysis of the huge files of data that constitute today's computed results. In
addition, the advent of supercomputers has made large-scale computing
something that most scientists can do. The latest advances in comprehen-
sive graphical displays and the emphasis on very fast vector and highly
parallel computer systems promise orders-of-magnitude increases in what
we can calculate in the next few years. Today, three-dimensional com-
putations can be pursued with about the same spatial resolution as state-
of-the-art two-dimensional computations ten years ago. This resolution of
a couple of hundred grid points in each dimension, in turn, is about the
same as used in the ground-breaking reactive-flow models two decades
ago.
xix
This book presents a series of topics ranging from microscopic combus-
tion physics to several aspects of macroscopic reactive-flow modeling. As
the reader progresses into the book, the successive chapters generally in-
clude a wider range of physical and chemical processes in the mathematical
model. Including more processes, however, usually means that they will
be represented phenomenologically at a cruder level. In practice, the de-
tailed microscopic models and simulations are often used to develop and
calibrate the phenomenologies used in the macroscopic models. The book
first describes computations of the most microscopic chemical processes,
then considers laminar flames and detonation modeling, and ends with
computations of complex, multiphase combustion systems. Many impor-
tant new computations cannot be included here. We hope, however, that
both the introductions to each section and the various chapters themselves
will help the reader to place the missing topics into perspective.
Finally, the editors and authors would like to acknowledge Sandra Harris
of the Naval Research Laboratory for her major contributions to the or-
ganization, copy editing, and uniform high quality of this book.
Elaine S. Oran
Jay P. Boris
May 1991
xx
Background
This section contains a series of five chapters describing applications of
numerical methods to problems in combustion chemistry. It begins with
the most microscopic considerations of fundamental chemical processes
and the transition to research on statistical combinations of the fundamental
processes to construct multireaction chemical mechanisms. Next it treats
the various algorithms to perform numerical integration of these complex
mechanisms, and then finishes with a discussion of techniques aimed at
reducing these complex mechanisms to simpler mathematical models that
lend themselves to faster computation and to use in reactive-flow models.
These papers show a progression often repeated in research: a detailed
and quite fundamental treatment of a class of problems naturally leads to
a complex but generally predictive model of the phenomena involved. To
use the detailed models, specialized numerical techniques and algorithms
must be developed to solve complex equations, but these are expensive
and relatively impractical to use. The process of developing and using these
detailed models, however, increases our understanding to the point that
we see how to absorb and ''average over" the complexities. This results
in more compact and computationally less expensive phenomenologies that
nevertheless contain the essence of the underlying detailed models. In the
context of combustion chemistry, the search for these more efficient, if
less accurate, reduced chemical mechanisms is driven both by the necess-
ities of treating more and more reactions and chemical species and of
describing more spatial dimensions of variation.
The first two chapters (Page and Lengsfield, Brown) describe how the-
oretical quantum chemistry can be used to determine quantitative infor-
mation crucial to predicting fundamental chemical reactions and to
interpreting combustion chemistry experiments. Page and Lengsfield de-
scribe one of the most fundamental problems of quantum chemistry: com-
puting the potential energy surface based on the Born-Oppenheimer
approximation and using the most recent and advanced numerical algo-
rithms. For each possible arrangement of nuclei in a molecule or in a small
set of reacting molecules, a potential energy can be determined for the
system by solving for the quantum mechanical wave function of the elec-
trons, assuming that the nuclei move so much slower than the electrons
that they can be assumed to be fixed in space. Determining this electronic
wave function and the corresponding lowest energy state proves a very
difficult and expensive computation.
The "potential-energy surface" (PES) is the function of the atomic nu-
clear coordinates that describes how the composite electronic and nuclear
potential energy changes as the nuclei move relative to one another. A
combination of nuclear positions giving a minimum on the energy surface,
for example, represents a local equilibrium of the molecular configuration.
A path between two local minima can represent a reaction. Such paths of
lowest total energy pass through a saddle point connecting the two minima.
More is needed than the potential-energy surface, however; the dynamics
and temperature dependence of a reaction must be determined by other
considerations.
Applications of quantum chemistry to combustion problems typically
involve electronic-structure calculations to map important regions of the
potential-energy surface, and then statistical or dynamical models to obtain
quantities such as the molecular heat of formation or the reaction-rate
coefficients. In addition to these kinetic and thermodynamic parameters,
quantum chemistry can provide spectral information that can be used as
diagnostics for experiments. Page and Lengsfield describe both the theo-
retical and computational aspects of ab initio quantum chemistry.
Brown describes how numerically determined potential-energy surfaces
are used as input to transition-state and molecular-dynamic calculations,
plus elements of different approaches to the computation of reaction-rate
coefficients. Some of these are statistical or transition-state approaches in
which the dynamical aspects of the collisions and interactions can be by-
passed. These approaches require only limited information about the po-
tential-energy surfaces. Another class of approach, based on molecular
dynamics, uses time-dependent scattering calculations to simulate the prog-
ress of molecular collisions or rearrangements. The PES is actually used
as an interaction potential for these calculations, which are repeated over
and over with different impact parameters and molecular orientations until
useful statistics for the average reaction rates can be obtained. This ap-
proach is what chemists mean by the term "molecular dynamics." (Mo-
lecular dynamics has a different meaning to physicists, as shown in the
chapter by Peyrard and Odiot.)
Elementary rates of reaction between molecules and molecular frag-
ments (computed from methods such as those described by the first two
papers or measured from experiments) can be combined to model the
chemical reaction sequences in a multispecies combustion system. For ex-
ample, even the simplest reaction, H2 + iO2 —» H2O, actually occurs as
a sequence of elementary reactions among the reactants (H2 and O2) to
form intermediates (H, O, OH, HO2, and H2O2) and subsequent rear-
rangements and recombinations eventually to form the product (H2O).
Each of the elementary reactions may be endothermic or exothermic, but
result in an overall exothermicity for the combustion system. In the past
ten years, Westbrook and coworkers and Warnatz and coworkers (see Ref.
1, for example) have compiled a large body of chemical reaction rates, and
analyses of them, describing complex hydrocarbon chemistry. In the chap-
ter in this book, Westbrook and Pitz describe work they have done to
construct and use complex chemical reactions mechanisms to describe the
combustion of hydrocarbon fuels. With the combined increase in our
knowledge of chemical rates and in computational power available, sig-
nificant changes are occurring in the complexity of fuels that can be mod-
eled and the range of pressure and temperature over which these mechanisms
are valid. Westbrook and Pitz describe reaction mechanisms that involve
reaction of carbon-containing molecules with four to eight carbon atoms.
In particular, they focus on the low-temperature regime and the application
of the complex kinetics to the study of engine knock.
To solve some of the dynamics problems described by Brown and the
time-dependent chemical kinetics described by Westbrook and Pitz, it is
necessary to solve sets of coupled, nonlinear ordinary differential equa-
tions. Numerous methods for doing this are enumerated and explained by
Radhakrishnan, who also describes sensitivity-analysis techniques for de-
termining which of the reactions or species are most important, and for
quantifying exactly how important they are. Radhakrishnan primarily de-
scribes the solution of sets of ordinary differential equations describing
multireaction kinetics problems made difficult by the presence of reactions
occurring on a wide range of timescales (stiff equations). Generally speak-
ing, fluid-dynamic considerations are restricted to the local choices of con-
stant pressure, constant temperature, or constant density. Spatial variation,
convection, and diffusion generally appear in these ordinary differential
equations only through the choice of constant-density, constant-tempera-
ture, or constant-pressure systems. Subsequent chapters of this book, such
as those by Kailasanath and Smooke, address techniques for solving or-
dinary differential equations suitable for use in problems where diffusion
or convection must also be included in the model.
In addition to discussing methods for solving ordinary differential equa-
tions, Radhakrishnan briefly describes recent work applying sensitivity
analysis to these complex chemical rate schemes. This approach is designed
to help determine which rates are the most important, and therefore which
are the controlling ones in the multireaction mechanism. Among the other
approaches to sensitivity analysis described, Rabitz's has been broadly
applied to combustion problems. 2
Even with the best methods and the fastest computers, it is not yet
practical to solve detailed chemical reaction mechanisms embracing hundreds
of reaction rates among dozens to hundreds of chemical species with mul-
tidimensional fluid dynamics and diffusion. Substantial effort by various
groups has therefore gone into reducing the complete chemical reaction
mechanisms to subset approximations that capture a usefully large fraction
of the complete behavior. Frenklach classifies these reduction methods and
describes some new approaches. He distinguishes between global modeling
that describes the overall conversion of reactants to products, and chemical
lumping in which subsets of the reactions within the overall mechanism
are combined (as dictated by the particular reactions and the prevailing
conditions) to yield a significantly reduced mechanism. A more specialized
form of chemical lumping called "statistical lumping" is considered for the
linear kinetics of polymer growth.
Through the chapters in this section, then, the reader sweeps over the
most fundamental quantum chemical computations of the potential-energy
surface, to the computation of chemical reaction rates and other chemical
parameters, to the construction of complex chemical reaction-rate mech-
anisms and the analysis of their structure, and finally to various attempts
to reduce these complex reaction mechanisms for specific applications.
This section thus provides the fundamental chemistry background for the
subsequent description of what happens when other physical processes
enter the overall combustion model.
Frenklach, M., "Modeling," Combustion Chemistry, edited by W.C. Gardner, Springer-
Verlag, New York, 1984, pp. 423-453.
2
Rabitz, H., Kramer, M., and Dacol, D., "Sensitivity Analysis in Chemical Kinetics,"
Annual Review of Physical Chemistry, Vol. 34, 1983, pp. 419-461; also, Kramer, M.A.,
Rabitz, H., Calo, J.M., and Kee, R.J., "Sensitivity Analysis in Chemical Kinetics: Recent
Developments and Computational Comparisons," International Journal of Chemistry and
Kinetics, Vol. 16, 1984, pp. 559-578.
Chapter 1
Ab Initio Quantum Chemistry for Combustion
Michael Page and Byron H. Lengsfield III
I. Introduction
A DVANCES in theoretical and computational methods, coupled with

the rapid development of powerful and inexpensive computers, fuel
the current rapid development in computational quantum chemistry (QC).
Nowhere is this more evident than in the areas of QC most relevant to
combustion: the description of bond breaking and rate phenomena. Al-
though the development of faster computers with larger memories has had
a major impact on the scope of problems that can be addressed with QC,
the development of new theoretical techniques and capabilities is respon-
sible for adding new dimensions in QC and has paved the way for the
unification of QC electronic structure calculations with statistical and dy-
namical models of chemical reactions. These advances will be stressed in
this chapter. This is not a review. We describe past accomplishments se-
lectively to set the stage for discussion of ideas or techniques that we believe
will have significant impact on combustion research. Thus, the focus of the
chapter is as much on the future as it is on the past.
Central to the theoretical description of chemical phenomena is the Born-
Oppenheimer approximation, 1 - 2 which asserts that the motion of the elec-
trons in a molecule is uncoupled from the motion of the nuclei. The Born-
Oppenheimer approximation leads to the concept of a potential energy
surface (PES). For each possible arrangement of the nuclei in a molecule,
a potential energy can be determined by solving the quantum mechanical
equations of motion for the electrons. The PES is the function that describes
how this potential energy changes as the nuclei move relative to one an-
other. It is the multidimensional generalization of the familiar diatomic
potential energy curve. Application of QC to combustion problems typi-
cally involves performing electronic structure calculations that serve to map
out important regions of the PES, and then using statistical or dynamical
models to obtain such quantities as molecular heats of formation and re-
action rate coefficients.
This paper is declared a work of the U.S. Government and is not subject to copyright
protection in the United States.
4 AB INITIO QUANTUM CHEMISTRY FOR COMBUSTION
QC impacts combustion modeling in three nearly distinct ways. To ap-

preciate this, it helps to consider the two different types of chemical input
to a combustion model: thermodynamic parameters for each of the chem-
ical species involved and kinetic parameters for reactions among these
species. This distinction is illustrated in Fig. 1. QC contributes directly, as
shown, to the determination of each of the two types of chemistry input:
thermodynamic and kinetic parameters. In addition, QC can contribute
indirectly to the experimental determination of kinetic data through cal-
culation of spectra and properties of transient species pertinent to exper-
imental diagnostics. We focus primarily on the direct theoretical determination
of crucial combustion model input.
The most important distinction, considering the electronic structure cal-
culations, between the theoretical determination of thermodynamic and
kinetic parameters is that notably different regions of the PES are required.
Thermodynamic parameters, i.e., molecular heats of formation, depend
on energies and other properties of molecules at or near their equilibrium
configurations on the PES. Kinetic parameters, on the other hand, are
most sensitive to the PES in the transition state region, which is charac-
terized by partially formed and partially broken chemical bonds. These
are entirely different requirements, and different electronic structure tech-
niques are best suited to each of these tasks.
This chapter first describes the electronic structure methods most ame-
nable to the equilibrium and the transition state regions of the PES. Next,
in Sec. Ill, potential energy surfaces will be discussed more generally,
stressing the computational advances that have been developed to inves-
tigate important regions for molecules with several atoms. Reaction path
Quantum chemistry ( H ¥ = E ¥
ABC
A + BC, AB + C
reaction coordinate
Variational transition
state theory
Heats of formation
of reactant and
product species
Computer modeling
input
Fig. 1 Schematic representation of link between first-principles quantum chemistry

and combustion modeling input.
M. PAGE AND B. H. LENGSFIELD Mi 5
approaches will be described in Sec. IV. These approaches incorporate the

advances discussed in the previous sections and yield a tractable and ef-
ficient means for generating PES information for use in statistical and
dynamical reaction rate theories. Section V discusses computational issues
with an emphasis on the limitations of current methods and how recent
advances in computer architectures and theoretical methods can be used
to address these deficiencies.
II. Electronic Structure Methods

Within the Born-Oppenheimer approximation and neglecting spin-orbit
and other small terms in the Hamiltonian, the electronic wave function is
determined by solving the time-independent Schroedinger equation for the
electrons in the coulomb field of the fixed nuclei. Most of the methods
discussed in this chapter are based on the Raleigh-Ritz variational principle,
which states that the energy calculated from an approximate trial wave
function is an upper bound to the true electronic energy. Thus, determining
a wave function is typically a linear or nonlinear optimization problem.
Given an initial-guess trial wave function, it is generally straightforward to
calculate an energy associated with that guess. Electronic wave functions
are obtained by minimizing the computed energy with respect to some type
of functional flexibility introduced in the trial wave function through the
variational parameters. Methods differ by the constrained functional form
of the trial wave function and the nature and extent of the flexibility. The
Hartree-Fock wave function, which reflects an independent particle ap-
proximation, is both the conceptual and computational starting point for
many types of electronic wave functions. The ^-electron Hartree-Fock wave
function, ^HF? is written as an antisymmetrized product of one-electron
functions called molecular orbitals (MO's),
^HF = 1^(1)^(2)^(3)^(4) ... <$>,,(2n - 1H,,(2*)| (1)
The determinantal form of the Hartree-Fock wave function ensures that
the wave function is antisymmetric with respect to interchange of any two
electrons, as it must be for a system of indistinguishable fermions. The
overbar indicates that the spatial MO is multiplied by a p spin function
(spin down). No bar indicates an a spin function (spin up). Because each
MO appears twice in the Hartree-Fock wave function multiplied by a
different spin function, the MO's are occupied by two electrons (doubly
occupied) by construction. Other MO's, obtained in the calculation but
not appearing in the HF wave function, are unoccupied by electrons and
are often called virtual orbitals. The Hartree-Fock method is old, dating
back to around 1930 and the work of Hartree, 3 Fock,4 and Slater.5 The
birth of the modern computational era, however, came in 1951 with the
contribution of Roothaan. 6 Roothaan introduced variational flexibility into
the Hartree-Fock wave function through a finite basis set expansion of the
MO's,
m
4>i = E c^ (2)
Given a set of fixed basis functions x, the /th MO is thus specified by a

column vector of MO expansion coefficients ch which are taken as varia-
tional parameters. What was previously an awkward variational problem—
minimize the energy with respect to an orbital defined on a grid of points
in space—now becomes a problem that can be solved by using the methods
of linear algebra. Applying the variational condition to this wave function
leads to the Roothaan equations for the optimum MO expansion coeffi-
cients. In matrix form, these can be written collectively as a pseudoeigen-
value equation,
FC = SCe (3)
where S is the basis function overlap matrix. The Fock matrix F depends
on the MO coefficients C, so this is a nonlinear set of equations and must
be solved iteratively. The result at convergence is an approximation to the
Hartree-Fock wave function since the chosen finite basis set is only an
approximation to a complete set of basis functions.
The Hartree-Fock approximation is an independent particle approxi-
mation; each electron moves under the influence of an average field gen-
erated by the other electrons. Thus, the motion of one electron is not
correlated with the motion of any other electron, with the exception of the
correlation implicitly introduced by the antisymmetrization of the wave
function. The difference between the true electronic energy and the Har-
tree-Fock energy is called the correlation energy, and electronic structure
methods that go beyond the Hartree-Fock approximation are referred to
as correlated methods. Correlated methods are needed to obtain reliable
heats of reaction and reaction rates. Electronic structure methods that
incorporate electron correlation are discussed in the following subsections.
A. Bond Breaking and the MCSCF Method

The Hartree-Fock method enjoys its greatest success in calculations on
molecules at or near their equilibrium configurations on the PES. On the
other hand, the method fails miserably for the description of the simple
homolytic cleavage of a two-electron chemical bond such as
H20 ——> H + OH
or
C2H6 ——> 2CH3
In these cases, the constraint of doubly occupied MO's breaks down, as
the known products of these reactions include single, unpaired electrons
on each of the species, each best described by a separate, singly occupied
MO.
A simple, effective remedy for this problem, one that allows the occu-
pation numbers of key MO's to change continuously as a reaction proceeds,
is to choose the wave function as a linear combination of two determinants
of the form of Eq. (1). In the second, additional determinant, the MO
M. PAGE AND B. H. LENGSFIELD
corresponding to the two-electron bond that is being broken is replaced

by the previously unoccupied antibonding MO for this bond,
V = A! |<|>1(l)*i(2)c|)2(3H2(4) ... 4>n(2n - l)$n(2n)\
+ A2 |c|>1(l)$1(2)<|>2(3)$2(4) ... 4>n + l(2n - l)$ll + 1(2n)| (4)
If the coefficients Al and A2 are determined variationally, the result is

a configuration interaction (CI) wave function. The resulting CI wave
function is not, however, the best wave function of the form of Eq. (4).
The MO's appearing in Eq. (4), which were variationally optimized for
the single-determinant case, are no longer the best choice when the second
configuration is included. If the MO's and the CI coefficients are simul-
taneously determined variationally, the result is a multiconfiguration self-
consistent field (MCSCF) wave function. The variational problem for the
MCSCF wave function is more difficult than it is for HF or CI wave
functions. There are many parameters that must be determined variation-
ally, including CI coefficients and orbital expansion coefficients. There are
also many constraints among these parameters due to orthogonality and
normalization requirements. Early attempts to formulate the orbital op-
timization portion of the problem using Lagrange multiplier techniques,
as in the Hartree-Fock-Roothaan method, were plagued by convergence
difficulties.7 (See Ref. 7 for a good discussion of early attempts to solve
the MCSCF problem.) Significant success in determining MCSCF wave
functions did not come until the late 1970s and early 1980s, with the de-
velopment of techniques using a subset of parameters that can undergo
unconstrained variation and still satisfy the orthonormality constraints by
construction.8-9 Further success came with the realization of the importance
of the coupling between the MO's and the CI coefficients. 10 Most MCSCF
wave functions are now determined using techniques based on the Newton-
Raphson method, and either simultaneous optimization of all variational
degrees of freedom11 or serial optimization steps of MO parameters and
CI parameters using a restricted form of the MCSCF wave function for
which these techniques are rapidly convergent. 12
In addition to providing the best wave function of the form of Eq. (4),
variational determination of the MO's leads to distinct advantages in the
calculation of energy derivatives with respect to positions of the atoms.
These advantages are important and will be discussed in a later section.
The MCSCF wave function is compact and easily interpreted. It forms the
basis of many of the recent theoretical treatments of important combustion
reactions and is likely to become well entrenched as a method of choice
in this area. In most applications, the choice of which configurations to
include in the MCSCF wave function is made fairly automatic by using the
contributions of Ruedenberg 13a4 and Roos et al.12 Their idea is to identify
an active set of molecular orbitals for which electron correlation effects are
expected to be prominent. For a combustion reaction, this typically includes
the electrons and the MO's that are part of the chemical bonds being broken
and formed in the reaction. A full CI is then constructed in the active
space: all possible configurations of a given space and spin symmetry that
can be constructed by distributing the active electrons among the active

MO's are included in the wave function. If the MO's and the CI coefficients
are optimized simultaneously, the result is a restricted class of MCSCF
wave function called complete active space (CASSCF). The number of
configurations in the CAS expansion grows rapidly with the size of the
active space, and a practical limit is reached at about 10-12 electrons
distributed among 10-12 orbitals. Fortunately, for most organic or com-
bustion reactions, relatively modest active spaces, ranging from about two
to eight electrons distributed among roughly an equivalent number of or-
bitals, can be identified.
With the advent of new MCSCF techniques, it was hoped that MCSCF
methods would replace CI techniques and become the method of choice
for accurate molecular calculations. Although MCSCF wave functions now
play a key role in accurate electronic structure calculations, this early
philosophy largely has been abandoned. MCSCF calculations are now gen-
erally used when the Hartree-Fock wave function is expected to be inad-
equate and is thought of more as being an extension of the Hartree-Fock
method than as a replacement for CI methods. The generally accepted
view15 is that MCSCF should be used only to obtain a qualitatively correct
reference description and not to describe extensive dynamical correlation
effects. These ideas will be illustrated by the following examples. Consider
the calculation of the potential energy curve for internal twisting about the
carbon-carbon double bond in ethylene. The proper CASSCF reference
description includes two electrons distributed among two active orbitals.
This description correctly describes the TT bond for which the two methylene
groups are parallel to one another. It also correctly describes the biradical
region for which the methylenes are perpendicular as well as regions in
between, which are characterized by partial TT bonding and partial biradical
character. If, on the other hand, the problem of interest is the internal
rotation about the single bond in ethane, then the proper reference de-
scription is a Hartree-Fock wave function. This is not a discontinuity in
approach; rather, the Hartree-Fock wave function is the limit of the CASSCF
concept for zero active space. In both cases, the calculation is qualitatively
correct but provides only a zeroth-order description of the chemistry. An
alternative reference wave function, one which has been used by the the-
oretical chemistry group at Argonne for many important contributions to
combustion chemistry, is the generalized valence bond wave function pi-
oneered by Bobrowitz and Goddard.16 Although derived from a different
perspective, this wave function is equivalent to a restricted MCSCF.
Taken by itself, the absolute value of the calculated energy of a molecule
with respect to separated nuclei and electrons is of minimal interest. To
extract chemical information from QC calculations, we must compare the
energies of two or more points on the PES. The calculated PES is always
higher in energy than the actual surface; however, if the surfaces are every-
where parallel, then all of the chemical information extracted from the
calculated surface will be correct. If the stabilization energy due to electron
correlation was constant across the PES, then the Hartree-Fock PES would
be correct, in the sense of being parallel to the true surface, even though
electron correlation is neglected. The Hartree-Fock PES is not correct,
M. PAGE AND B. H. LENGSFIELD III 9
however, and the MCSCF method has been developed to account for the
major changes in electron correlation across the surface.
Quantitative accuracy requires that differential electron correlation ef-
fects involving all of the electrons be described in more detail. The best
way to capture these differential correlation effects in combustion reactions
is to perform large-scale configuration interaction calculations, with the
important requirement that the CI must be based on a qualitatively correct
reference description.
B. Configuration Interaction
For a given one-electron basis set, only a finite number of MO's can be
constructed, equal in number to the number of linearly independent basis
functions. These MO's are different, depending on whether they are de-
termined by a Hartree-Fock calculation, by an MCSCF calculation, or by
some other means. If, however, one were to perform a full CI, including
all possible configurations distributing all electrons among all MO's, then
the resulting wave function is the same, regardless of which set of MO's
is used. In fact, the full CI wave function is the exact solution of Schroe-
dinger's equation in a given basis set.
Full CI calculations are prohibitive however, and in practical calculations
with basis sets of useful size the configuration list is truncated in some way.
This is where the importance of the reference wave function enters the
calculations. Because the truncated CI is no longer invariant to the choice
of reference MO's, but depends critically on the choice, expending some
effort in calculating a good reference wave function is a necessity. The
truncated CI can be made more compact and provides a consistent treat-
ment of the relevant regions of the PES.
A widely used truncated CI is obtained by including all configurations
that can be constructed by either a single- or a double-electron substitution
from a Hartree-Fock reference configuration. 17 This wave function, singles
and doubles configuration interaction (SDCI), provides a good way to
include dynamical correlation effects in cases where the Hartree-Fock wave
function represents an adequate reference description. If the Hartree-Fock
reference is inadequate, as is the case for most combustion reaction path-
ways, then one can use the multireference counterpart of SDCI. A trun-
cated CI that includes all configurations that are either a single or a double
substitution from a selected set of reference configurations, usually formed
with MCSCF MO's, is called a multireference configuration interaction
(MRCI).
In an important series of benchmark calculations, Bauschlicher and co-
workers at NASA Ames Research Center have performed extensive com-
parisons of various correlated methods with full CI calculations. 18~21 The
full CI calculations all employed atomic basis sets of double-zeta plus
polarization quality, which is flexible enough to allow a meaningful com-
parison of a variety of theoretical methods. The comparisons typically
involved stretching single bonds both individually and simultaneously,
stretching multiple bonds, evaluating molecular properties, and computing
barriers to reaction. Their results underscore the importance of choosing
an adequate reference description. Over a wide range of problems, the
CASSCF/MRCI wave function was found to be in excellent agreement

with the full CI calculations and, thus, provides a balanced treatment of
the electron correlation problem in chemical reactions.
The size of the MRCI expansion grows very rapidly as the number of
atoms in the problem increases. Therefore, it is desirable to find approx-
imations to the full MRCI wave function that reduce the number of var-
iational parameters. The polarization CI (POL-CI) method of Hay and
Dunning22-23 was designed for this purpose and has been used for a number
of combustion studies involving a limited number of heavy atoms. In a
POL-CI calculation, only single excitations from the reference space into
the virtual space are included in the CI expansion.
New CI techniques that include contributions from double excitations
in an approximate way by restricting groups of configurations to appear
with fixed relative importance have also been developed. The contribution
that this group of configurations makes to the wave function is then de-
termined variationally. This is the idea behind the externally24-25 and
internally26'27 contracted CI (CCI) schemes, which will be discussed in more
detail in Sec. V.
C. Basis Sets
All of the one-electron molecular orbitals that form the foundation of
the electronic structure methods described in this chapter are represented
as linear sums of fixed basis functions [Eq. (2)]. This is often referred to
as the linear combination of atomic orbitals approximation. Although atomic
orbitals may be a judicious choice, their use is no more than computa-
tionally convenient. Any set of functions satisfying the correct boundary
conditions and forming a reasonable approximation to a complete set of
functions can be used. Molecules are, however, essentially perturbed at-
oms, so that functions that describe the atoms well are useful in molecular
calculations.
Whatever basis functions are chosen, and whatever electronic structure
method is used, the two-electron integrals involving the electron-electron
repulsion term in the Hamiltonian (l/r12) will have to be computed,
l) - X,(2)X8(2) dTl dT2 (5)

rl2
where xa is tne ath basis function and the integral is over all space for the
coordinates of electrons 1 and 2. The evaluation of these integrals is com-
plicated by the fact that the basis functions are, in general, centered at
different points in space; they are typically centered on the atoms in the
molecule.
Virtually all basis functions used in polyatomic calculations are comprised
of Gaussian functions for which these multicenter integrals can be evaluated
efficiently.2Si29 Since each integral has four basis function indices, there
are, ignoring permutation symmetry, N4 of these integrals, where N is the
number of basis functions. The number of operations required to process
these integrals in a Hartree-Fock calculation also scales as TV4. Thus, there
is a severe dependence of the cost of the calculation on the length of the
basis set expansion, providing a strong motivation for the development of

compact basis sets. The problem is worse for MCSCF and second-order
CI calculations, which have N5 and TV6 steps, respectively.
The crudest basis set includes one function for each orbital in an occupied
or partially occupied subshell of the free atom. This would include five
basis functions for carbon or oxygen and one basis function for hydrogen,
for example. The next level of flexibility is to include two functions of
different size for each valence orbital, e.g., two functions for each of the
2s and 2p atomic orbitals of carbon or oxygen. An example of such a split
valence set is the widely used 6-31G basis set.30 The spatial extent of a
Gaussian basis function is determined by the coefficient in the exponent,
which, by convention, is referred to as zeta. Two or three functions for
each atomic orbital are thus referred to as double-zeta or triple-zeta basis
sets, respectively.31 Higher angular momentum functions, e.g., d functions
on carbon, are necessary to describe properly the polarization of the elec-
tron distribution when atoms are brought together to form molecules. A
double-zeta basis set with polarization functions (DZP) is the minimum
level for semiquantitative calculations.
Most basis sets used in the past decade have been optimized for Hartree-
Fock calculations on the free atoms. This is a sound approach when the
basis sets are used for molecular calculations at the Hartree-Fock, or even
MCSCF, level. The inadequacies of these basis sets for the description of
dynamical correlation effects have recently received attention. Almlof and
Taylor32 have presented large contracted Gaussian basis sets based on
atomic natural orbitals (ANO's). Similarly, Dunning 33 has recently pre-
sented new basis sets for use in correlated calculations, placing emphasis
on reducing the number of Gaussian primitives without sacrificing the
accuracy of the ANO basis sets.
D. Perturbation Methods and Bond-Additive-Correction Schemes

The primary focus of discussion up to now has been electronic structure
methods capable of describing chemical reactions, i.e., chemical bonds
being broken and formed. Equilibrium thermochemistry presents a whole
class of problems important to combustion that do not explicitly involve
the creation or destruction of chemical bonds. Instead, these problems
involve only energy differences between molecules in their equilibrium
configurations and separated noninteracting fragments. For the most part,
molecules at or near their equilibrium configurations on the PES are well
described by a Hartree-Fock wave function. Thus, correlated methods
based on a Hartree-Fock reference are expected to be adequate for cal-
culating bond strengths, energy differences between isomers, and molec-
ular heats of formation.
Perhaps the most widely used correlated method is Moller-Plesset per-
turbation theory. This is Raleigh-Schroedinger perturbation theory with
the zeroth-order Hamiltonian chosen as a sum of one-electron Fock op-
erators. In this implementation, the first-order correction recovers the
Hartree-Fock energy. Higher-order terms in the perturbation expansion
systematically include the results of electron correlation. Calculations are
practical through fourth order (MP4) and are widely used at second order
(MP2). Perturbation theory methods are not variational: the computed

energy is not an upper bound to the true energy. On the other hand,
perturbation theory methods have the desirable feature of being size ex-
tensive, i.e., the computed energy of n separated noninteracting fragments
is n times the computed energy of one fragment. In general, this is not the
case for variational methods such as configuration interaction.
Molecular heats of formation are largely molecular energies relative to
the energies of the molecule's constituent atoms in their predefined stand-
ard states. Computationally, molecular energies are determined relative
to gas-phase atomic species and then these atomic species are related to
their standard state by experimental information. Therefore, accurate cal-
culation of molecular heats of formation requires accurate calculation of
atomization energies, or successive bond dissociation energies.
The first principles calculation of bond dissociation energies is a noto-
riously difficult problem in QC because there is substantially more cor-
relation energy in the molecule than there is in the separated fragments.
Quantum chemical calculations that include electron correlation at a high
level and employ a DZP quality basis set typically underestimate bond
energies by several kcal/mole for single bonds, and even more for multiple
bonds. This is an unacceptable error, particularly if several chemical bonds
are broken, such as in a calculation of a molecular atomization energy.
Calculations at this level of theory contain a greater predictive capability
than would be implied by these statements, however. Pople and cowork-
ers,34 for example, have focused attention on isodesmic chemical reactions,
for which the number of chemical bonds is conserved. Similarly, and more
recently, attention has been focused on isogyric energy comparisons, for
which the number of unpaired electrons (or, equivalently, the number of
paired electron spins) is conserved. These successful ideas are both at-
tempts to exploit a cancellation of the errors associated with bond disso-
ciation energies, which tend to be of the same sign and, to a first
approximation, of similar magnitude.
Accurate results for the atomization energies for AH,7 hydrides, where
A is a second row atom, have been obtained using isogyric energy com-
parisons.35-36 In these calculations, a large basis set was used on the central
atom and a DZP quality basis set was used on the hydrogen atoms. Atom-
ization reactions were made isogyric by including an appropriate number
of H2 molecules on the side of the atoms and twice that number of H atoms
on the side of the molecule,
nH + AH,, -» A + nH2
This is equivalent to a direct atomization energy calculation incorporating

a bond-additive-correction (BAC) scheme, if the correction factor is taken
to be the dissociation energy defect found for H2. This defect, or bond
correction factor, can be determined by comparing the calculated DZP H2
bond energy to the experimentally known value.
If, in contrast to using a large basis set on the central atom, as was done
in the preceding study, a smaller DZP quality basis set is used on the
nonhydrogen atoms as well, then one would expect the bond correction
factor to depend on the identity of the central atom. This is the situation
in the BAC-MP4 method of Melius and Binkley. 37 By comparison with
experiment, a different bond correction factor is found for each type of
chemical bond. Atomization energies calculated with a reasonable, but
tractable, level of theory are then combined with the BAC's to obtain
semiempirical estimates of molecular heats of formation. Refinements to
this method have included an exponential bond length dependence of the
bond correction factor, which implicitly reflects the dependence of the
bond correction factor on bond order. Further refinements reflect a de-
viation from simple bond additivity by including small corrections for sec-
ond bonding neighbors. Useful results have been obtained for hundreds
of molecules using the BAC-MP4 method.
The bond correction schemes just discussed all depend on a comparison
of calculated results to accurately known experimental quantities. For classes
of compounds for which this accurate experimental data are unavailable,
bond correction factors can still be determined by performing very high
level calculations on small prototype molecules and comparing these results
with the results of lower level calculations. Theoretically determined bond
correction factors have been determined and used to calculate heats of
formation for small boranes.38
III. Potential Energy Surfaces

Perhaps the most useful mental construct in all of theoretical chemistry
is the representation of chemical dynamics, the complex simultaneous mo-
tion of several atoms in a molecule, by the rolling or sliding of a single
particle on a multidimensional surface.39-40 This concept is brought to life
through our everyday experiences with gravitational potential energy. Words
and phrases such as hills, valleys, mountain passes, and plateaus are per-
vasive in discussions involving PES's. Many important dynamical concepts
can be illustrated, or at least can be imagined, using the two dimensions
of the gravitational potential energy construct. In the remainder of this
section and into the next section, we discuss the PES and, in particular,
how to explore dynamically useful PES regions. During this discussion, it
is important to keep in mind that the PES is vast and that the determination
of a single point on the PES implies an extensive electronic structure
calculation.
Describing the orientation of an N atom molecule, the position in three-
dimensional space of each of the atoms, requires 3N numbers. Discarding
six degrees of freedom for overall rotation and translation of the molecule
leaves, for a nonlinear molecule, 3N—6 internal degrees of freedom. Imag-
ine how one might map out a PES. The water molecule is an easy case,
with only three atoms and three internal degrees of freedom. Fixing the
HOH bond angle, one might construct a grid of potential energy as a
function of the two OH bond lengths. Performing an electronic structure
calculation for each of, say, 10 values of each of the bond lengths leads to
100 calculations (ignoring molecular symmetry). By repeating this proce-
dure for 10 values of the bond angle, a rough grid of the PES is obtained,
but it took 1000 electronic structure calculations! A similar grid for a four-
atom molecule requires 106 calculations; a five-atom system, 109 calcula-

tions.
Clearly, mapping out the full potential energy surface is tractable only
for the smallest molecules. Even once such a grid is available, the problem
of fitting the discrete points to an analytic functional form that can be used
in subsequent dynamical studies is not straightforward. Despite a great
deal of work in this area, there is no systematic, widely applicable approach
to this problem.41 Many global PES's, which are most prevalent for three-
atom systems, have been fit to some combination of ab initio energy points,
experimental information about the PES, and semiempirical calculations.
They may also have adjustable parameters that are chosen to reproduce
some desired dynamical result. The functional forms to which the PES
data are fit vary considerably but are generally physically motivated.
The value of much of the research involving PES global fits is not pre-
dictive capability on specific molecular systems. Rather, the value lies in
the contribution of this research to the generic understanding of the dy-
namics of energy transfer, and in the testing of simplified models based
only on selected PES regions.
For molecules with many atoms, it is not possible to map out the full
PES, but a great deal about a molecule and its behavior can be learned
from knowledge of a few selected PES regions. Since stable isomers of a
molecule are represented as local minima on the PES, it is reasonable to
expect that properties of these species are determined by the PES shape
in the vicinity of these minima. The local minima are fundamental points
on the PES. They determine the molecular structure and the moments of
inertia by which the rotational spectra can be estimated. The curvature of
the multidimensional PES about the local minima determines the vibra-
tional properties of the molecule, including the normal modes and har-
monic vibrational frequencies.
Another fundamental point on the PES is the lowest energy saddle point
connecting two local valleys or minima. The saddle point is the highest
point on a lowest total energy pathway connecting these minima. It has
the property that it is a maximum with respect to one degree of freedom
and is a minimum with respect to all other degrees of freedom. The saddle
point is an approximate dynamical bottleneck, a point of no return, for a
transformation from the vicinity of one local minimum or valley to another.
The probability (or rate) of reaction is influenced very strongly by the
height in energy of this saddle point. If the energy barrier is high and
peaked, then the saddle point generally provides an excellent dynamical
bottleneck or transition state: molecular trajectories that cross the saddle
point in the reactant to product direction are very unlikely to return to
reactants. If the barrier is broad, then one has to pay more attention to
the way in which the other degrees of freedom are changing in the vicinity
of the saddle point to determine a dynamical bottleneck, but this search,
in practice, still begins with the location of the saddle point.
Both the minima and the saddle points are stationary points on the PES:
the derivative of the energy with respect to each of the geometrical co-
ordinates, and thus the force on each of the atoms, is zero. One can imagine
locating a minimum on the PES by systematically minimizing the energy
with respect to first one bond length or bond angle and then another,
calculating a series of energy points each time. This is cumbersome, how-
ever, since the degrees of freedom are coupled to one another. Depending
on the initial parameter guesses, this process will converge slowly, if at all.
There are fairly advanced algorithms for locating minima of functions of
several variables based on function evaluations only, but these are still
prohibitive for molecules with many atoms. The location of saddle points
is even a more difficult problem than the location of minima.
A. Born-Oppenheimer Potential Energy Derivatives

In a recent review, Schaefer and Yamaguchi 42 state that "Analytic de-
rivative methods have, over the last fifteen years, revolutionized the way
in which quantum chemistry is done." To illustrate the impact that analytic
derivative techniques have on the theoretical study of chemical reactions,
we consider methanol (CH3OH), with 6 atoms and 18 Cartesian degrees
of freedom. For some arbitrary orientation of the molecule, we can perform
an electronic structure calculation and obtain a point on the PES; but we
learn nothing about the dynamics of the molecule unless we consider how
the PES changes as the atoms move. The slope of the PES along any
direction is minus the component of the force in that direction, and the
gradient of the PES—a vector containing the slope in all of the coordinate
directions—gives the forces on all of the atoms. The gradient can be
obtained by a finite-difference procedure: compute the energy for very
small displacements in each of the coordinate directions and see how much
it changes. Using a two-point difference method for acceptable accuracy,
the Cartesian gradient takes, in this case, 36 energy evaluations (reduced
to 24 by exploiting the invariance of the energy to overall translations and
rotations43). The second derivatives of the energy with respect to nuclear
coordinates, which, at equilibrium, are the force constants determining
small vibrations, can be obtained by finite differencing the gradients. These
quantities—the energy, gradient vector, and force constant matrix—are
the fundamental tools of PES exploration. Yet, for methanol, if an energy
evaluation takes one unit of effort, then a finite-difference gradient eval-
uation takes 24, and a force constant calculation about 600 units. In con-
trast, determining these quantities by the direct evaluation of the algebraically
differentiated electronic energy expressions takes roughly one, three, and
ten units of effort for evaluating the same quantities! Furthermore, the
ratio of these efforts—force constants to gradient or gradient to energy—
is only weakly dependent on the number of atoms for analytic evaluation,
in contrast to being directly proportional to the number of atoms for the
finite-difference procedure.
Clearly, for a particular application for which the energy evaluation
takes, for example, 15 min of computer time, PES exploration involving
many repeated calculations of gradients and force constants is out of the
question unless analytic derivatives are available. Applications that are
opened up by using analytical derivatives are not just bigger and better
examples of the same types of problems that were done before, but include
whole new classes of problems.
In analytic derivative evaluation, the energy expression for a particular

electronic structure method, the expectation value of the Hamiltonian, is
first differentiated and then the resulting expression is evaluated. Although
straightforward in principle, the algebra tends to be cumbersome and there
can be data handling problems. Fortunately, there are some major sim-
plifications for wave functions that have been determined by the variational
principle. If we separate the parameters determining the wave function
into those that are determined by the variational principle, {q}, and those
that are not variational, {p}, then the derivative of the Born-Oppenheimer
potential energy with respect to a nuclear coordinate, x, can be written as
d , / / . / ^ idE\ dEdp dE dq
- - + - + - ( 6 )
The difficult parts of the calculation are evaluating the derivatives of the
parameters {p} and {q} with respect to x. Fortunately, since the parameters
{q} are determined through the variational principle by the requirement
that they minimize the energy, dE/dq = 0, the last term in the preceding
expression vanishes. For example, the derivatives of the molecular orbitals
with respect to nuclear coordinates do not have to be determined to cal-
culate gradients for Hartree-Fock or MCSCF wave functions, since the
orbitals are determined variationally. Similarly, the derivatives of the CI
coefficients for MCSCF or CI wave functions do not have to be determined.
For CI wave functions, however, for which the MO's have not been de-
termined variationally, the simplification involving the MO's is no longer
available.
A major part of the derivative evaluation for all of the wave-function
formulations considered here is the evaluation of the derivatives of the
two-electron integrals with respect to nuclear displacements. These are
included in Eq. (6) as part of {dp/dx}. These integrals are the basic building
blocks of the electronic energy expressions. The Hartree-Fock energy
expression for a calculation with order of 102 basis functions, for example,
has 6(108) terms, each involving orbital expansion coefficients multiplied
by an integral over up to four basis functions. Since the basis functions are
centered on the atoms, these integrals depend on the positions of the atoms
and must be differentiated. The derivative of a two-electron integral is
itself a small linear combination of new two-electron integrals involving
slightly more complicated functions. Second derivatives of the energy in-
volve second derivatives of the integrals, which leads to more integrals that
need to be evaluated. Recent advances in the efficient evaluation of these
derivative integrals44"46 are a major catalyst for the rapid development of
derivative methods.
A lucid, detailed review of analytic derivative techniques has recently
been presented by Pulay,47 who was the original pioneer of derivative
methods in the late 1960s48 and has remained very active in this area.
B. Locating Critical Points

The greatest application of analytical potential energy gradients over the
past 15 years undoubtedly arises from their role in the theoretical deter-
M. PAGE AND B. H. LENGSFIELD Ml 17
mination of molecular structure. Equilibrium molecular structures are those

geometries for which the potential energy is a local minimum, i.e., a local
valley on the PES. Given the gradient, finding a minimum energy structure
is straightforward in principle. Since the negative of the gradient points in
the steepest downhill direction, one can step some small distance along
this direction, calculate a new gradient, and begin again. This steepest
descent procedure is not the most efficient way to find the minimum,
however. Fortunately, through chemical intuition and specific knowledge
about related compounds, a reasonably good initial guess for the structure
of a molecule usually can be obtained. Given a reasonable guess, it is useful
to expand the energy in a Taylor series about the guessed geometry. The
coefficients of the expansion to second order are the gradient vector and
the force constant matrix, respectively. If {x} is a set of coordinates, either
Cartesian coordinates of the atoms or internal bond lengths and bond
angles, then the energy can be expanded as
E(x) = E(x0) + gt(x - *0) + H* - x0yF0(x ~ *o) + - (7)
Truncating the expansion at second order and solving for the point at
which the gradient vanishes gives the Newton-Raphson expression:
x - x0 = -(Fo)-1^ (8)
If the initial-guess geometry is close to the minimum, so that the surface

is truly quadratic, then the Newton-Raphson step gives the minimum in
one step. Otherwise, finding the minimum is an iterative process. Note
that, for Eq. (7) to approach a minimum, indeed to guarantee that the
step direction is initially downhill at all, the matrix F0 must be positive
definite, i.e., the surface must curve upward in all of the normal coordinate
directions at the expansion point. Unless analytical second derivatives are
available, the force constant matrix is expensive to calculate and, therefore,
the matrix F0 is often—in fact, usually—replaced by some approximation
to it. If F0 is replaced with any positive-definite matrix, e.g., a unit matrix,
then the iterative search technique is termed a quasi-Newton method. Dur-
ing the iterative cycles, the inverse of the matrix can be updated by using
the newly calculated information to approximate more closely the inverse
of the force constant matrix. These are called variable metric methods.
There are some advantages to using internal coordinates for performing
quasi-Newton optimizations, particularly if one starts with a diagonal guess
to the force constant matrix. Bond lengths and bond angles are typically
less coupled to one another than are Cartesian coordinates; therefore, the
force constant matrix is more dominant diagonally. In addition, internal
coordinates are physically motivated; thus, sound initial guesses for the
diagonal second derivatives can be made.
There is a rich literature on the subject of unconstrained minimization
of nonlinear functions of many variables, 49 but the concerns involved in
optimizations on PES's are often different than those assumed in the lit-
erature. The cost of simply evaluating the function, i.e., calculating the
energy, is much higher than usually assumed. Therefore, the number of
function evaluations must be kept to a minimum. In addition, the cost of
the matrix manipulations, e.g., diagonalizing or inverting the force constant

matrix, is trivial compared with the cost of one function evaluation. A
thorough review of geometry optimization methods in quantum chemistry
has recently been presented by Schlegel.50
The optimization of minimum energy structures through the methods
mentioned herein is now fairly routine. Many thousands of optimum struc-
tures have been reported in the literature. The accuracy of the structural
determinations depends on the electronic structure method and the quality
of the basis set, but this dependence is well documented. Many molecules
have secondary local minima on the PES. These isomeric forms are often
unstable and experimentally inaccessible. Nonetheless, they may play a
key role as intermediate species in chemical reaction mechanisms. Quan-
tum chemical calculations play a key role in the identification of these
intermediate and unstable species and in the determination of their prop-
erties.
As much as the optimization of minima is routine, the location of saddle
points, or, less precisely, transition states, is a vexatious problem. The
objective is to minimize the energy with respect to all degrees of freedom
except one, and to maximize with respect to that degree of freedom. The
problem is that the coordinate direction for which the energy is to be
maximized is known only after the problem is solved. In a broader sense,
the problem is compounded by the more stringent requirements on the
electronic structure methods for investigating transition regions of chemical
reactions.
At a saddle point, the surface has zero gradient and one, and only one,
negative eigenvalue of the force constant matrix. Given a reasonable guess
to the saddle-point structure, the Newton-Raphson method [Eq. (7)] can
be used to locate the stationary point. The radius of convergence of the
Newton-Raphson iterative procedure is not very large, however. Guessing
a structure that is close enough to the saddle point is difficult since chemical
intuition is less well developed for transition state regions than it is for the
vicinity of minima. Guessing diagonal elements of the initial force constant
matrix, as is often done to initiate searches for minima, is less useful because
the internal coordinates tend to be strongly coupled to one another when
bonds are being broken and formed. Methods have been proposed that
seek to minimize the norm of the gradient, 51 climb uphill from the vicinity
of local minima, 52 53 and maximize along a simple prescribed path while
minimizing perpendicular to the path.54-55 None of these methods is fool-
proof, and finding difficult saddle-point structures is somewhat of an art.
However, locating a saddle point is well worth the effort, since that one
point on the PES brings quantum chemistry from the arena of thermo-
dynamics—relative stability and equilibrium populations—to that of ki-
netics. One of the most enduring and useful theories in all of chemical
physics is the transition state theory (TST) of chemical reaction rates.
Imagine two molecules approaching each other. They can either scatter
off one another and go off in different directions (a nonreactive collision)
or they can react in some way, e.g., exchange atoms. These processes can
be viewed as the motion of a single particle on one PES. In the nonreactive
case, the particle approaches a mountain pass from the vicinity of a valley
and either does not have enough energy to reach the top of the mountain
pass, or has enough energy but missed the passageway and so came back
down in the vicinity of the original (reactant) valley. In the reactive case,
the particle goes over the mountain pass and ends up in the new (product)
valley. Consider a large number of such trajectories at the same temper-
ature (a canonical ensemble). The probability that any one trajectory is
reactive is related to the probability that it has at least the energy of the
saddle point (£ f ), which is just the Boltzmann term exp[ — E*/(kT)]. It is
also related to the probability that it finds it way through the mountain
pass, given that it has enough energy. Whether the pass is narrow or broad
in the vicinity of the saddle point depends on the eigenvalues of the force
constant matrix for the directions perpendicular to the reactive direction.
These are the basic ingredients for a simple application of TST: energies,
structures, and force constants for reactants and the saddle point. From
these quantities, the rate coefficient can be calculated over a wide tem-
perature range. Since the rate constant is sensitive to the precise value of
the activation energy, the activation energy is sometimes adjusted so that
the calculated rate constant agrees with experiment at, for example, room
temperature, and then the TST calculations are used to extrapolate the
rate to experimentally inaccessible temperatures.
IV. Reaction Path Approaches

Reaction rate theories are described in more detail in the next chapter
by Brown. We briefly note here two inadequacies in conventional TST,
describe what PES information is needed to resolve them, and describe
how to get this information efficiently from QC calculations. Implicit in
conventional TST is the assumption that the saddle point, or, more pre-
cisely, a plane passing through the saddle point separating reactants from
products, provides a perfect dynamical bottleneck for the reaction: no
trajectory that passes through this plane from reactants to products ever
bounces back. The saddle point is a logical place to choose such a transition
state. Certainly, in one dimension the top of the hill is the dynamical
bottleneck. In many dimensions, however, the width of the pass has to be
taken into consideration. If the pass significantly narrows as one proceeds
downhill from the saddle point, and the energy has not dropped signifi-
cantly, then the dynamical bottleneck may be shifted away from the saddle
point. This is more likely to be the case when the energy profile in the
saddle-point vicinity is broad and drops off slowly as one moves away from
the saddle point.
The best transition state, at least in classical mechanics, is the one for
which the calculated reaction rate is a minimum. The idea behind varia-
tional transition state theory (VTST) is to move the transition state dividing
plane (or, more generally, surface) away from the saddle point toward
reactants and products until the best bottleneck is found. Application of
VTST requires more PES information than does conventional TST. Truhlar
et al.56 have done much over the last decade toward developing VTST into
a practical tool for applications, such as combustion reactions using ab
initio PES information.
Another limitation of conventional TST occurs when motion across the

saddle point involves light atoms, such as H, particularly if the barrier is
peaked. In this case, the assumption that motion across the saddle point
is well represented by classical mechanics breaks down.
For both of these problems—the variational determination of the tran-
sition state and the incorporation of quantum effects for reaction coordinate
motion—the PES information can be obtained by using the idea of a
reaction path.39 In this idea, one degree of freedom is separated out as a
reaction coordinate. One widely used path, referred to as the intrinsic
reaction coordinate57 or minimum energy path (MEP),56 is defined as the
path of steepest descent in a mass-weighted Cartesian coordinate system
that passes through the saddle point. In mass-weighted coordinates, the
individual atomic Cartesian displacements are scaled by the square root of
the atomic mass. It is this mass scaling that allows us to view the dynamics
of N particles in three dimensions as the dynamics of one particle in 3N
dimensions. The MEP satisfies the following differential equation:
v(s) = (s) = - = (s) (9)
where the normalization factor c is defined as

c(s) = VE'VE (10)
If one imagines a trajectory starting at the saddle point and proceeding

downhill along the MEP, then partway down the hill, as the MEP curves,
the trajectory would overshoot the MEP due to inertia effects. The MEP
is unique, but does not correspond to any physical trajectory. It can be
identified with a hypothetical trajectory for which the kinetic energy is
removed constantly. It is physically analogous to the path that water—or,
better yet, molasses—takes flowing downhill.
It is customary to make a quadratic approximation to the PES about the
saddle point. This is done in conventional TST, for example. Given the
force constants at the saddle point, a normal coordinate analysis can be
performed. The harmonic vibrational frequencies are obtained as the square
root of the eigenvalues in each of the normal mode directions. Since the
saddle point has one direction of negative curvature, one of the harmonic
frequencies is imaginary. In fact, the direction of the eigenvector associated
with the imaginary frequency is the initial direction of the MEP toward
reactants and products.
A. Reaction Path Hamiltonian

In 1980, Miller et al.58 further entrenched the reaction path idea with
the derivation of the classical Hamiltonian for a simple potential based on
the MEP. A number of dynamical models based on this Hamiltonian have
since been proposed. Conceptually, the idea is to consider the potential
as a trough or stream bed plus 37V-7 harmonic walls that are free to
contract or expand as one proceeds along the trough. The PES is approx-
M. PAGE AND B. H. LENGSFIELD ill 21
imated as the potential energy of the MEP [V0(s)] plus a quadratic ap-
proximation to the energy in directions perpendicular to the MEP,
37V-7
V(s, Q1 ... Q3N.7) = V0(s) + ^ ±<4Qi (11)
k= 1
Here Q is the generalized normal coordinate and co the associated harmonic

frequency. They are obtained at each point on the path by diagonalizing
the force constant matrix for which the reaction path direction, as well as
directions corresponding to rotations and translations, have been projected
out.
To obtain the reaction path potential, one must begin at the saddle point
and numerically integrate Eq. (9) to obtain the MEP. The force constant
matrix is then needed at several points along the path to perform the normal
mode analyses to obtain the generalized normal modes.
To obtain the Hamiltonian function for this reaction path potential, it
is necessary to express the kinetic energy in terms of the momenta conjugate
to the reaction path coordinates. The necessary coordinate transformation
is presented in Ref. 58, as well as a second transformation to the harmonic
action-angle variables. This second transformation is particularly instruc-
tive since the Hamiltonian reduces to a simple Hamiltonian for just the
reaction path degree of freedom with an effective one-dimensional poten-
tial if certain coupling terms are set to zero. That one-dimensional potential
is the vibrationally adiabatic potential. If the motion along the reaction
coordinate is slow compared with motion in the transverse directions, the
transverse vibrations are abiabatic. The energy in these vibrations is not
conserved, but rather the vibrational action, which is the ratio of the energy
to the frequency, is conserved. As the transverse frequencies change along
the MEP, energy leaks into or out of these degrees of freedom giving rise
to an effective one-dimensional potential. Quantum mechanically, the vi-
brational action corresponds to the vibrational quantum number, and vi-
brational adiabaticity means that the transverse vibrations remain in the
same quantum state as the system proceeds along the reaction coordinate.
The coupling terms that appear in the reaction path Hamiltonian can be
given a geometrical interpretation. The most important coupling terms are
those that couple the transverse vibrational modes directly to the reaction
path. These are called the curvature coupling elements because they meas-
ure the degree to which the reaction path curves into a particular transverse
mode as the reaction coordinate is traversed. The curvature coupling ele-
ments are also an important ingredient in the description of quantum me-
chanical tunneling effects for the reaction path degree of freedom in VTST
calculations. The curvature of the reaction path can be obtained by dif-
ferentiating Eq. (9) with respect to s. The result is59
^1 = [FV - (v'Fv)v]/Vc (12)
The scalar curvature is just the length of this vector, and the curvature
coupling elements in the reaction path Hamiltonian are just elements of
the curvature vector expressed in the basis of generalized normal modes.
Note that the curvature can be calculated from only first and second de-
rivatives at a point on the path. The classical notion that a trajectory will
overshoot the path and climb the wall if the path curves on the way down
the hill is a reflection of this curvature coupling. Climbing the wall in a
transverse direction is tantamount to exchanging energy between the re-
action path and the transverse vibration. The curvature coupling elements
are also very important for the treatment of quantum effects in reaction
coordinate motion. If the path is curved in the vicinity of the saddle point,
then a quantum particle might be able to cut the corner and tunnel through
a narrower section of the potential. However, it should be cautioned that
Eq. (12) becomes indeterminate at the saddle point and the curvature must
be obtained from a limiting equation based on 1'Hospital's rule.59
Determining the curvature at the saddle point, as well as the remaining
coupling elements in the reaction path Hamiltonian, require knowing lim-
ited information about the third derivatives on the PES. Specifically, they
require the derivative of the force constant matrix with respect to the
reaction coordinate direction, which can be obtained by finite differences.
These remaining coupling elements can be identified with the direct cou-
pling of transverse vibrational modes (the mode-mode coupling elements)
and with the derivatives of the transverse frequencies with respect to s.
B, Following Reaction Pathways

The most computationally intensive step in VTST calculations or in
dynamical studies based on the reaction path Hamiltonian is the deter-
mination of the reaction path by numerical integration of Eq. (9) and the
evaluation of potential energy derivatives along the path. Most reaction
paths are obtained by using methods based on simple Euler integration of
Eq. (9). Beginning at the saddle point, one first steps along the path
tangent, which, at the saddle point, is an eigenvector of the force constant
matrix. Subsequent small, straight-line steps are taken along the local steepest
descent direction, which requires only the gradient. The Euler method
tends to oscillate about the true MEP. A number of devices based on
constrained energy minimization have been proposed to return to the path
after too large an Euler step has been taken. 50
If the reason for following a reaction path is to check for intervening
minima or barriers, or simply to visualize how a molecule might change
during a reaction, then the Euler-type methods may be adequate. For the
most part, however, reaction paths are followed to obtain properties such
as transverse vibrational frequencies, which depend on second energy de-
rivatives. Recently, methods have been proposed59-60 that efficiently use
the second derivative information when it is available to follow the path.
These methods are based on a Taylor series representation of the path in
the arc length about a point on the path,
X(S) = X(S0) + vV\S - S0) + W»(S - Jo)2
+ ••• (lln\)v(n-l\s - s0)n (13)

The coefficients v(n\ which are all vectors, depend only on energy de-
rivatives evaluated at the point of expansion. Truncating the expansion at
M. PAGE AND B. H. LENGSFIELD ill 23
first order yields an Euler step. Truncating the path at second order yields
a simple quadratic step, and determining the coefficient of the second-
order term requires second energy derivatives. If one makes a local quad-
ratic approximation (LQA) to the energy, however, then the first two terms
in the Taylor series are given correctly and there is a nonzero contribution
at every other order. Given partial third energy derivative information,
further contributions to the coefficients in Eq. (13) can be determined. 59
This approximate third derivative information is available if successive
force constant matrices have been determined along the path. Thus, by
using a judicious combination of the LQA method and approximate eval-
uation of higher order terms in the Taylor series expansion of the path, 60
the information that is already available for performing VTST calculations
can be used to follow the path more accurately.
V. Computational Issues
In this section, we discuss the limitations of the current computational
methods and how recent advances in computer architectures and theoretical
methods can be used to address these deficiencies. First, we discuss the
computational bottlenecks present in the current theoretical methods. The
notion of a direct algorithm is introduced, and we will discuss how this
technique can be used in conjunction with vector and parallel computer
architectures to redress the problems encountered when one seeks to use
the current theoretical methods to study larger molecules. In the second
portion of this section, we discuss the new theoretical methods that are
evolving to treat larger molecular systems and to study reactions that occur
in the liquid or solid phase. CCI and reaction field methods are described
briefly, and the type of impact these new developments are likely to have
on current research efforts is discussed.
A. Architectural Considerations
The first step in an electronic structure calculation is to evaluate matrix
elements of the nonrelativistic Born-Oppenheimer Hamiltonian operator,
//, in a basis of atomic orbitals, x- The Hamiltonian is given by
electrons _ > 2 electrons nuclei _ ^
2m, ' i j Ru
nuclei 7- >y electrons i
+ g ~f + g ^ (14)
where the four terms are the one-electron kinetic energy operator, the
electron-nuclear attraction operator, the nuclear-nuclear repulsion oper-
ator, and the electron-electron repulsion operator, respectively. The dis-
tance between two particles, e.g., electron / and nucleus /, is represented
by RJJ. It is convenient to express this Hamiltonian as the sum of nuclear
repulsion, one- and two-electron operators,
electrons electrons
H = ^ hf + ^ g, + NR (15)
where
/J..^v?
n
' 2m, V' ++ t
S^Ru 8 ^~
" r,
and
nuclei -7- ^
NR=
Within the Born-Oppenheimer approximation, for which the motion of

the electrons is assumed to be decoupled from the motion of the nuclei,
we solve for the electronic energy as a function of nuclear position. Thus
the electronic wave function is determined with the nuclear coordinates
held fixed, and the only remaining two-body operator is the electron-
electron repulsion term. By evaluating matrix elements of the one- and
two-electron operators comprising this Hamiltonian, we obtain a two-di-
mensional array of one-electron integrals and a four-dimensional array of
two-electron integrals,
1) di-! (16a)
j
and
gmnop -
l2
From the definition of these arrays, the following permutation symmetry

is apparent:
and
5 opmn •> etc. (17)

Thus, for N basis functions, there are N x (N + 1)12 unique one-electron
integrals and [N x (N + 1)12] x {[TV x (N + 1)12] + l}/2 or approximately
7V4/8 unique two-electron integrals. In an iterative SCF or MCSCF calcu-
lation, these integrals must be processed repeatedly. The computation and
storage of the two-electron integrals clearly dominate this stage of the
calculation.
In a typical electronic structure calculation, one employs a DZP61 quality
atomic basis set to achieve semiquantitative results. In such a basis, 16
functions are centered on each first row atom and 5 functions are centered
on each hydrogen atom. An electronic structure calculation on the nitro-
methane molecule (CH3NO2), for example, would contain 79 basis func-
tions and give rise to about 5 million unique two-electron integrals. A

similar calculation on the propellant RDX (C3N6O6H6) would contain 270
basis functions and 20 billion unique integrals. Of these unique integrals,
perhaps 30% are large enough to merit retention. Thus, in a traditional
electronic structure calculation, the number of two-electron integrals ex-
ceeds the high-speed computer memory available, and these integrals must
be stored on a peripheral device, usually a disk drive.
The large number of two-electron integrals that must be processed gives
rise to an input/output (I/O) bottleneck in subsequent stages of the cal-
culation since there is a disparity between the CPU time required to process
the integrals and the time required to move these integrals from a disk
drive to the computer's memory. At most computer centers, the available
disk space for integral storage is exceeded when the basis set approaches
150 functions. We should also note that the CPU time required to generate
the two-electron integrals associated with a basis set of 150 basis functions
is substantial. Current algorithms62"64 generate about 50,000 two-electron
integrals per second. The gist of this discussion is that the computation of
two-electron integrals is inherently a parallelizable task and the I/O bot-
tleneck can be circumvented by using a direct algorithm where the two-
electron integrals, or, in general, any matrix elements, are recomputed as
they are needed and never stored. Direct algorithms become more attrac-
tive as our ability to compute rapidly the matrix element of interest in-
creases, so a parallelizable method of constructing the matrix elements is
ideally suited for a direct algorithm.
A direct SCF algorithm was first implemented by Almlof.65 In each
iteration of a closed shell SCF calculation, the two-electron integrals are
contracted with a density matrix to generate a Fock matrix, F,
tnn '^mn / ^ opL omnop ^\ \&monp &mpnoj\ V /

op ^
where an element of the density matrix, dop, is constructed from the coef-
ficients of the occupied MO's [see Eq. (2)J as
occupied
C
dop = S oaCpa (19)
a
The SCF wave-function optimization is a nonlinear procedure, so this step

is repeated until the change in the MO's falls below a prescribed tolerance.
In Almlof s original work, 65 published in 1982 before parallel computers
were readily available, the two-electron integrals were evaluated during
each iteration, but integral construction was not parallelized.
We stated earlier that the generation of two-electron integrals is inher-
ently parallelizable since the construction of these integrals naturally con-
sists of a series of independent calculations. One simply assigns a different
group of integrals to be computed in each CPU. This does not require the
reorganization of existing computational methods or the development of
any new methods to take advantage of the additional CPU's available on
a parallel computer. The ability to easily parallelize the generation of two-
electron integrals was recently exploited in a traditional SCF algorithm by

Watt and Dupuis.66 In this work, a large number of independent CPU's
and disk drives were employed in SCF calculations involving several hundred
basis functions. A block of integrals was computed in each CPU. Each
group of integrals was then stored on a disk drive and later used to build
a portion of the Fock matrix. This method also parallelizes the I/O in the
SCF calculations. These direct methods have also been extended to include
MP2 calculations.67-68 However, as stated earlier, a parallel direct SCF
method avoids the necessity of obtaining large quantities of disk space,
and such an algorithm is being developed by several groups.69
Almlof s direct SCF algorithm has been generalized by Taylor70 to treat
MCSCF wave functions. With a parallel version of this algorithm, SCF and
MCSCF calculations on molecules as large as RDX will be practical. It is
important to bear in mind that for these methods to be widely applicable,
we must be able to differentiate these energy expressions with respect to
nuclear displacements. These derivatives allow equilibrium and transition
state geometries to be rapidly located on the molecular PES and ultimately
allow us to obtain the kinetic and thermodynamic quantities of interest.
To obtain the first derivative of the SCF, MCSCF, or CI energy, an expres-
sion similar to Eq. (18), but where the integrals have been replaced by
derivative integrals, must be evaluated. 4871 Since the computation of these
derivative integrals involves the same methods used to compute the inte-
grals themselves,70-71 this step is also inherently parallelizable. Thus, with
a parallelized integral program, the first derivative portion of an electronic
structure calculation is also tractable.
The most commonly employed direct algorithm in quantum chemistry
is the direct CI method pioneered by Roos and Siegbahn.72 This method
has evolved a great deal with the advent of unitary group,73 74 symbolic
matrix CI,75 and full CI76"79 methods, but the essential feature remains
unchanged. In a CI calculation, matrix elements of the Hamiltonian are
computed in a basis of configuration state functions (CSF's). The first few
eigenvalues and eigenvectors of this matrix are then needed to define
potential energy surfaces, to compute excitation energies, and to evaluate
molecular properties. In a direct CI method, the eigenvector of the CI
Hamiltonian is obtained by a scheme due to Davidson,80 where one iter-
atively multiplies this matrix times a trial vector. The construction and
storage of the CI Hamiltonian can be avoided because only the vector
resulting from the matrix multiplication is needed. CI expansions consisting
of several million CSF's can be treated with these methods.81
A CI calculation with 1 million CSF's might require an hour of CPU
time on a Cray-XMP to complete. There has been a great deal of effort
devoted to vectorizing these methods,82 and these schemes can be paral-
lelized; however, there are a number of hurdles that must be overcome
before very large systems can be studied. The major problem is that a
viable method of directly processing the integrals in a CI scheme has not
been demonstrated, although at least one such method has been pro-
posed.69 The most efficient CI methods currently available require at least
a partial transformation of the two-electron integrals from the atomic or-
bital (AO) basis to an MO basis. Currently, the best prospect for treating
a very large system, such as RDX, is to employ a localization scheme83

where only a small number of MO's must be considered in the correlation
problem. For example, to study the cleavage of one of the three equivalent
nitro groups from the RDX chemical ring, one might include in the cor-
relation calculation MO's localized on the nitro group and the neighboring
chemical bonds in the ring along with the associated virtual orbitals. The
problems associated with transforming the very large number of two-elec-
tron integrals from the AO basis to the basis defined by the smaller set of
localized MO's is still substantial, but this problem is manageable and
transformation algorithms designed for parallel computers exist.68
The application of parallel CI algorithms on current and future gener-
ations of supercomputers will have a dramatic effect on our ability to study
accurately small and midsized problems. Larger AO basis sets and CI
expansions will be tractable in these studies. Jobs that required 10-20 h
of CPU time to complete will take less than an hour. This dramatically
impacts our ability to respond rapidly to questions posed by experimen-
talists. So we should see much closer ties between experimental and the-
oretical efforts in the future.
We also note that there are "size-consistency"84 and "size-extensivity"83-86
problems associated with using a CI method (short of a full CI method,
which is impractical for these applications) to correlate a large number of
electrons. These problems, which are related to the correct scaling of the
method with size, are well understood and can be addressed — although
at some computational expense — with multireference coupled cluster,87
modified coupled pair functional 88 89 or averaged coupled pair functional 90
methods.
B. Theoretical Developments
Contracted CI Methods
In our earlier discussion of correlation methods, we noted that the size
of multireference CI calculations grows very rapidly as the size of the
molecule increases. Recall that, in a multireference CI calculation, one
employs a series of reference configurations that qualitatively describes the
interaction of interest and generates a multireference CI configuration list
by performing single and double replacements of the occupied orbitals in
the reference configurations with virtual orbitals. These virtual orbitals are
usually obtained as the unoccupied orbitals from an earlier SCF or MCSCF
calculation. To illustrate these points, we consider the dissociation of the
diatomic molecule, HF. Here we would employ three reference configu-
rations:
nd ) (20a
R2 = îÎ^sCHF^^^CHF™,^^) (20b
(20c)
where, for example, <$:(Fls), the first MO, is occupied by two electrons
and consists predominantly of the Is atomic orbital on the F atom.
We would then generate a group of configurations by a series of single

and double excitations of the occupied orbitals in the reference configu-
rations to the virtual orbitals,
R,(3 -* 8) = 4>?cf>i<f>M<!>3<!>8} (21)

We can construct a compact representation of this list of configurations
by denoting three things: the reference configuration, the occupied orbital
we are exciting from, and the virtual orbital we are exciting to;
R,(2 -» 7) - /?,(2; 7)
K,(3 -» 8) - tf :(3; 8) (22)
Note that excitations from the doubly occupied or inactive orbitals (or-
bitals 1-4 in our reference configurations) generate a unique list of con-
figurations, whereas excitations from the partially occupied or active orbitals
(orbitals 5 and 6 in our reference configurations) need not produce a unique
configuration. Similarly, we can use this notation to label the doubly excited
configurations,
R,(2 -> 7, 3 -> 8) = tfbfolfafo^*}
= fli(2, 3; 7, 8) (23)
Multireference CI calculations, including all single and double substi-
tutions, on molecules the size of nitromethane with a DZP basis set and
perhaps 10 reference configurations require substantial quantities of com-
puter time on the current generation of supercomputers. Furthermore, the
configuration list consisting of all single and double substitutions from all
of the reference configurations grows very quickly when either the number
of orbitals in the inactive, active, or virtual spaces is increased, or when
the number of reference configurations is increased and rapidly becomes
prohibitive. Recently, this problem has been addressed with contracted CI
methods.91"94 There are two such methods: the externally contracted CI
method and the internally contracted method. In the externally contracted
CI method, developed by Siegbahn,91 one contracts all of the configurations
with the same occupied orbital indices into a single configuration,
<&/(*, b) = 2 Ct(ab, cd)Rl(ab, cd) (24)
cd
The contraction coefficients Cf(ab, cd) are obtained with little computa-
tional expense from a perturbation theory expression.91 The dimension of
the CI problems is reduced by a factor of approximately 3 x Nv x (NL,
+ l)/2, where Nv is the number of virtual orbitals. Therefore, for a mol-
ecule with 55 virtual orbitals, such as a DZP calculation of nitromethane,
this factor is 4620 and a CI problem with 8 million CSF's is reduced to a
problem with less than 2000 CSF's.
About 60% of the computational expense of generating this contracted
Hamiltonian matrix is associated with a single iteration of the Davidson
algorithm (multiplying the Hamiltonian matrix times a trial vector) used

to obtain one eigenvector of the uncontracted Hamiltonian matrix. From
7 to 12 cycles are usually required to obtain convergence with Davidson's
method, thus an externally contracted CI is from 10 to 15 times faster than
a conventional CI calculation. However, since there are fewer variational
degrees of freedom, this technique is also less accurate than a conventional
CI calculation. The success of this technique depends on traditional per-
turbation theory arguments; the reference configurations define the zeroth-
order space in perturbation theory, and if this space represents a good
description of the interactions being studied, then this contracted CI tech-
nique produces very good results. It has been possible to determine such
a reference space in the vast majority of calculations performed to date.
In an internally contracted CI calculation, 26 92~93 one generates all single
and double excitations from a single-reference wave function, but this wave
function is composed of a linear combination of Slater determinants with
fixed expansion coefficients. This technique is the generalization of the
single-reference CI wave function where one employs an MCSCF wave
function in place of the SCF wave function as a reference configuration.
Thus, in an internally contracted CI calculation, one contracts over the
reference configurations to define the new configuration list,
r(ab, cd) = ^ Câb, cd) (25)
/
The contraction coefficients C, are obtained either from an MCSCF cal-
culation or determined variationally. This technique has been discussed by
several authors, but practical calculations have been possible only recently
with the work of Werner and Knowles.27 The computational effort saved
in an internally contracted CI scheme is clearly dependent on the number
of references one employs. The power of this method lies in the ability to
efficiently address problems where a large number of reference configu-
rations are employed. Since there are still a large number of variational
parameters present in this calculation, the CI results are quite accurate.
In a recent paper by Werner and Knowles,27 they report results from a
calculation on Cr2 where 78 million CSF's were present in the uncontracted
configuration list and 2.4 million CSF's were employed in the contracted
calculation. This calculation was completed in 9 h of CPU time on a Cray-
XMP.
For these methods to be fully utilized in combustion studies, one must
be able to efficiently locate the stable points on the potential energy surfaces
generated by these methods. The required derivative methods are well
understood,94'95 but efficient algorithms have not yet been demonstrated.
Derivative versions of these methods will likely be completed in the near
future.
Reaction Field Methods

It is a common practice in electronic structure calculations to exclude
the inner shell electrons from the CI portion of the calculation. These inner
shell electrons are constrained to be doubly occupied in the CI calculation,
and the interaction of these electrons with the valence or chemically active
electrons can be accounted for by adding a Fock operator [as in Eq. (18)]
to the one-electron integrals. This approximation is justified as these inner
shell electrons have a very different spatial extent, being much closer to
the nucleus, than the valence electrons. The effect of including these elec-
trons in the CI calculation is to shift uniformly the whole PES and not to
change the interaction energies. Thus, none of the relevant thermodynamic
or kinetic quantities of interest would be changed, but the resulting cal-
culation would be computationally very expensive and, in most cases, in-
tractable.
To further reduce the computational expense associated with treating
these inner shell electrons, pseudopotentials96 have been used to modify
the one-electron integrals in place of the Fock operator mentioned earlier.
The atomic basis functions used to account for the motion of these inner
shell electrons can be eliminated from the calculation, so one needs to
construct and manipulate much smaller two-electron integral files. Pseu-
dopotential methods have been widely used in the past few years to reduce
the computational expense of treating systems with a large number of atoms
and to account for relativistic effects97 in molecular calculations involving
heavy atoms. In a reaction field method,98"102 these pseudopotential ideas
are extended to represent the interaction of molecules or fragments of
molecules with a localized, chemically active group of electrons. Thus, one
could study a reaction occurring in a solution or in the solid state by using
ab initio methods to treat the molecules undergoing the reaction and mo-
lecular pseudopotentials to account for the solvent or nearest neighbor
interactions. To date, polarizable point charges have been employed in
reaction field calculations to account for these nearest neighbor interac-
tions; however, calculations describing true molecular pseudopotentials
will be reported in the near future. 103 We believe a great deal of effort
will be devoted to improving these techniques in the next few years, and
these efforts will have a positive impact on the type of combustion problems
that can be treated successfully with theoretical methods.
Acknowledgments
This work was supported by the Office of Naval Research through the
Naval Research Laboratory and by the Department of Energy through
Lawrence Livermore National Laboratory under Contract W-7405-Eng-
48.
References
ôrn, M. and Oppenheimer, J. R., "Quantum Theory of Molecules," Annales
de Physique (Paris), Vol. 84, 1927, p. 457.
2
Eyring, H., Walter, J., and Kimball, G. E., Quantum Chemistry, Wiley, New
York, 1944, pp. 190-192.
3
Hartree, D. R., "The Wave Mechanics of an Atom with Non-Coulomb Central
Field. I. Theory and Methods," Proceedings of the Cambridge Philosophical So-
ciety, Vol. 24, 1928, p. 89.
4
Fock, V., "Naherungsmethode zur Losung des quantenmechanischen Mehr-
korperproblems," Zeitschrift fuer Physik, Vol. 61, 1930, p. 126.
5
Slater, J. C, "Note on Hartree's Method," Physical Review, Vol. 35, 1930, p.
210.
6
Roothaan, C. C. J., "New Developments in Molecular Orbital Theory," Reviews
of Modern Physics, Vol. 23, 1951, p. 69.
7
Werner, H. and Meyer, W., ''A Quadratically Convergent Multiconfiguration-
Self-Consistent-Field Method with Simultaneous Optimization of Orbitals and CI
Coefficients" Journal of Chemical Physics, Vol. 73, 1980, p. 2342.
8
Levy, B. and Berthier, G., "Generalized Brillouin Theorem for Multicon-
figurational SCF Theories," International Journal of Quantum Chemistry, Vol. 2,
1968, p. 307.
9
Dalgaard, E. and Jorgensen, P., "Optimization of Orbitals for Multiconfigur-
ational Reference States," Journal of Chemical Physics, Vol. 68, 1978, p. 3833.
10
Lengsfield, B. H., Ill, "General Second-Order MCSCF Theory: A Density
Matrix Directed Algorithm," Journal of Chemical Physics, Vol. 73, 1980, p. 382.
H
Shepard, R., "The Multiconfiguration Self-Consistent Field Method," Ab Initio
Methods in Quantum Chemistry, Pt. II, edited by K. P. Lawley, Wiley, Chichester,
England, 1987, p. 63.
12
Roos, B. O., Taylor, P. R., and Siegbahn, P. E. M., "A Complete Active
Space SCF Method (CASSCF) using a Density Matrix Formulated Super-CI Ap-
proach," Chemical Physics, Vol. 48, 1980, p. 152.
13
Ruedenberg, K., Cheung, L. M., and Elbert, S. T., "MCSCF Optimization
Through Combined Use of Natural Orbitals and the Brillouin-Levy-Berthier Theo-
rem," International Journal of Quantum Chemistry, Vol. 16, 1979, p. 1069.
14
Cheung, L. M., Sundberg, K. R., and Ruedenberg, K., "Electronic Rear-
rangements During Chemical Reaction. II. Planar Dissociation of Ethylene," In-
ternational Journal of Quantum Chemistry, Vol. 16, 1979, p. 1103.
15
Roos, B. O., "The Complete Active Space Self-Consistent Field Method and
Its Applications in Electronic Structure Calculations," Ab Initio Methods in Quan-
tum Chemistry, Pt. II, edited by K. P. Lawley, Wiley, Chichester, England, 1987,
p. 399.
16
Bobrowitz, F. W. and Goddard, W. A., "The Self-Consistent Field Equations
for Generalized Valence Bond and Open-Shell Hartree—Fock Wave Functions,"
Methods of Electronic Structure Theory, edited by H. F. Schaefer, Plenum, New
York, 1977, p. 79.
17
Shavitt, L, "The Method of Configuration Interaction," Methods of Electronic
Structure Theory, edited by H. F. Schaefer, Plenum, New York, 1977, p. 189.
18
Bauschlicher, C. W. and Taylor, P. R., "Full CI Studies of the Collinear
Transition State for the Reaction F + H2 —> HF + H," Journal of Chemical
Physics, Vol. 86, 1987, pp. 858 and 5600.
19
Bauschlicher, C. W., Langhoff, S. R., Taylor, P. R., Handy, N. C., and
Knowles, P. W., "Full CI Benchmark Calculations on CH,.," Chemical Physics
Letters, Vol. 85, 1986, p. 1469.
20
Bauschlicher, C. W., Langhoff, S. R., Partridge, H., and Taylor, P. W.,
"Benchmark Full Configuration Interaction Calculations on HF and NH 2 ," Journal
of Chemical Physics, Vol. 85, 1986, p. 3407.
21
Bauschlicher, C. W. and Langhoff, S. R., "On the Electron Affinity of the
Oxygen Atom," Journal of Chemical Physics, Vol. 86, 1986, p. 5595.
22
Hay, P. J. and Dunning, T. H., "Polarization CI Wavefunction: The Valence
States of the NH Radical," Journal of Chemical Physics, Vol. 64, 1976, p. 5077.
23
Dunning, T. H., "The Low-Lying States of Hydrogen Fluoride: Potential En-
ergy Curves for the X'2 + , 3£ + , 3H, and 'II states," Journal of Chemical Physics,
Vol. 65, 1976, p. 3854.
24
Siegbahn, P. E. M., "The Direct Configuration Interaction Method with a
Contracted Configuration Expansion," Chemical Physics, Vol. 25, 1977, p. 197.
25
Siegbahn, P. E. M., "The Externally Contracted CI Method Applied to N 2 ,"
International Journal of Quantum Chemistry, Vol. 23, 1983, p. 1869.
26
Siegbahn, P. E. M., "Direct Configuration Interaction with a Reference State
Composed of Many Reference Configurations," International Journal of Quantum

Chemistry, Vol. 18, 1980, p. 1229.
27
Werner, H. and Knowles, P. W., "An Efficient Internally Contracted Multi-
configuration-Reference Configuration Interaction Method," Journal of Chemical
Physics, Vol. 89, 1988, p. 5803.
28
Dupuis, M., Rys, J., and King, H. F., "Evaluation of Molecular Integrals Over
Gaussian Basis Functions," International Journal of Chemical Physics, Vol. 65,
1976, p. 111.
29
King, H. F. and Dupuis, M., "Numerical Integration Using Rys Polynomial,"
Journal of Computational Physics, Vol. 21, 1976, p. 124.
30
Hehre, W. J., Ditchfield, R., and Pople, J. A., "Self-Consistent Molecular
Orbital Methods XII. Further Extensions of Gaussian-Type Basis Sets for Use in
Molecular Orbital Studies of Organic Molecules," Journal of Chemical Physics,
Vol. 56, 1972, p. 2257.
31
Dunning, T. H. and Hay, P. J., "Gaussian Basis Sets for Molecular Calcula-
tions," Methods of Electronic Structure Theory, edited by H. F. Schaefer, Plenum,
New York, 1977, p. 1.
32
Almlof, J. and Taylor, P. R., "General Contraction of Gaussian Basis Sets. I.
Atomic Natural Orbitals for First- and Second-Row Atoms," Journal of Chemical
Physics, Vol. 86, 1987, p. 4070.
33
Dunning, T. H., "Gaussian Basis Sets for Use in Correlated Molecular Cal-
culations. I. The Atoms Boron Through Neon and Hydrogen," Journal of Chemical
Physics, Vol. 90, 1989, p. 1007.
34
Hehre, W. J., Radom, L., Schleyer, P. V. R., and Pople, J. A., Ab Initio
Molecular Orbital Theory, Wiley, New York, 1986, pp. 270-308.
35
Pople, J. A., Frisch, M. J., Luke, B. T., and Binkley, J. S., "A Moller-Plesset
Study of the Energies of AH n Molecules (A = Li to F)," International Journal of
Quantum Chemistry, Vol. 17, 1983, p. 307.
36
Pople, J. A., Luke, B. T., Frisch, M. J., and Binkley, J. S., "Theoretical
Thermochemistry. I. Heats of Formation of Neutral AH n Molecules (A = Li to
Cl)," Journal of Physical Chemistry, Vol. 89, 1985, p. 2198.
37
Melius, C. F. and Binkley, J. S., "Energetics of the Reaction Pathways for
NH2 + NO —» Products and NH + NO —» Products," Twentieth Symposium
(International) on Combustion, The Combustion Institute, Pittsburgh, PA, 1984,
p. 575.
38
Page, M. and Adams, G. F. (to be published).
39
Schaefer, H. F., Chemistry in Brit., Vol. 11, 1975, p. 227.
40
Murrell, J. N., Carter, S., Farantos, S. C., Huxley, P., and Varandas, A. J. C.,
Molecular Potential Energy Surfaces, Wiley, New York, 1984.
41
Truhlar, D. G., Steckler, R., and Gordon, M. S., "Potential Energy Surfaces
for Polyatomic Reaction Dynamics," Chemical Reviews, Vol. 87, 1987, p. 217.
42
Schaefer, H. F. and Yamaguchi, Y., "A New Dimension to Quantum Chem-
istry. Methods for the Analytical Evaluation of First, Second, and Third Derivatives
of the Molecular Electronic Energy with Respect to Nuclear Coordinates," Journal
of Molecular Structure (Theochem), Vol. 135, 1986, p. 391.
43
Page, M., Saxe, P., Adams, G. F., and Lengsfield, B. H., "Exploiting Rota-
tional and Translational Invariance of the Energy in Derivative Evaluations in
Quantum Chemistry," Chemical Physics Letters, Vol. 104, 1984, p. 587.
44
Dupuis, M. and King, H. F., "Molecular Symmetry and Closed-Shell SCF
Calculations. I," International Journal of Quantum Chemistry, Vol. 11, 1977, p.
613.
45
Saxe, P., Yamaguchi, Y., and Schaefer, H. F., "Analytical Second Derivatives
in Restricted Hartree-Fock Theory. A Method for High-Spin Open-Shell Molecular
Wave Functions," Journal of Chemical Physics, Vol. 77, 1982, p. 5647.
46
Schlegel, H. B., "An Efficient Algorithm for Calculating Ab Initio Energy
Gradients Using s,p Cartesian Gaussians," Journal of Chemical Physics, Vol. 77,
1982, p. 3676.
47
Pulay, P., "Analytical Derivative Methods in Quantum Chemistry," Ab Initio
Methods in Quantum Chemistry, Pt. II, edited by K. P. Lawley, Wiley, Chichester
England, 1987, p. 241.
48
Pulay, P., "Ab Initio Calculation of Force Constants and Equilibrium Geo-
metries in Polyatomic Molecules. I. Theory," Molecular Physics, Vol. 17, 1969, p.
197.
49
Fletcher, R., Practical Methods of Optimization, Wiley, Chichester, England,
1981.
50
Schlegel, H. B., "Optimization of Equilibrium Geometries and Transition
Structures," Ab Initio Methods in Quantum Chemistry, Pt. I, edited by K. P.
Lawley, Wiley, Chichester, England, 1987, p. 249.
51
McIver, J. M. and Komornicki, A., "Structure of Transition States in Organic
Reactions. General Theory and an Application to the Cyclobutene-Butadiene Iso-
merization Using a Semiempirical Molecular Orbital Method," Journal of the Amer-
ican Chemical Society, Vol. 94, 1972, p. 2625.
52
Cerjan, C. and Miller, W. H., "On Finding Transition States," Journal of
Chemical Physics, Vol. 75, 1981, p. 2800.
53
Banerjee, A., Adams, N., Simons, J., and Shepard, R., "Search for Stationary
Points on Surfaces," Journal of Physical Chemistry, Vol. 89, 1985, p. 52.
54
Halgren, T. A. and Lipscomb, W. N., "The Synchronous-Transit Method for
Determining Reaction Pathways and Locating Molecular Transition States," Chem-
ical Physics Letters, Vol. 49, 1977, p. 225.
55
Bell, S. and Crighton, J. S., "Locating Transition States," Journal of Chemical
Physics, Vol. 80, 1980, p. 2464.
56
Truhlar, D. G., Isaacson, A. D., and Garrett, B. C., The Theory of Chemical
Reaction Dynamics, Vol. 4, edited by M. Baer, CRC Press, Boca Raton, FL, 1985,
p. 65.
57
Fukui, K., "The Path of Chemical Reactions—The IRC Approach," Accounts
of Chemical Research, Vol. 14, 1981, p. 363.
58
Miller, W. N., Handy, N. C., and Adams, J. A., "Reaction Path Hamiltonian
for Polyatomic Molecules," Journal of Chemical Physics, Vol. 72, 1980, p. 99.
59
Page, M. and Mclver, J. W., "On Evaluating the Reaction Path Hamiltonian,"
Journal of Chemical Physics, Vol. 88, 1988, p. 922.
60
Page, M., Doubleday, C., and Mclves, J.W., "Following Steepest Descent
Reaction Paths. The Use of Higher Energy Derivations with Ab Initio Electronic
Structure Methods," Journal of Chemical Physics, Vol. 93, 1990, p. 5634.
61
Dunning, T. H., "Gaussian Basis Functions for Use in Molecular Calculations
I. Contraction of (955p) Atomic Basis Sets for the First-Row Atoms," Journal of
Chemical Physics, Vol. 53, 1970, p. 2823; "Gaussian Basis Functions for Use in
Molecular Calculations III. Contractions of (1056p) Atomic Basis Sets for the First
Row Atoms," Vol. 55, 1971, p. 3958.
62
McMurchie, L. E. and Davidson, E. R., "One- and Two-Electron Integrals
Over Cartesian Gaussian Function," Journal of Chemical Physics, Vol. 26, 1978,
p. 218.
63
Obari, S. and Saika, A., "Efficient Recursive Computation of Molecular In-
tegrals Over Cartesian Gaussian Functions," Journal of Chemical Physics, Vol. 84,
1986, p. 3963.
64
Head-Gordon, M. and Pople, J. A., "A Method for Two-Electron Gaussian
Integral and Integral Derivative Evaluation Using Recurrence Relations," Journal
of Chemical Physics, Vol. 89, 1988, p. 5777.
65
Almlof, J., Faegri, K., and Korsel, K. "Principles for Direct SCF Approach
to LCAO-MO Ab Initio Calculations," Journal of Computational Chemistry, Vol.

3, 1982, p. 385.
66
Watt, J. and Dupuis, M., "Parallel Computation of Molecular Energy Gra-
dients on the Loosely Coupled Array of Processors (LCAP)," Theoretica Chimica
Ada, Vol. 71, 1987, p. 91.
67
Watt, J. and Dupuis, M., "Parallel Computation of the Moller-Plesset Second-
Order Contribution to the Electronic Correlation Energy," Journal of Computa-
tional Chemistry, Vol. 9, 1988, p. 158.
68
Head-Gordon, M., Pople, J., and Frisch, M., Chemical Physics Letters, Vol.
153, 1989, p. 503.
69
Whiteside, R., Binkley, J. S., Colvin, M., and Schaefer, H. F., "Parallel Al-
gorithms for Quantum Chemistry. I. Integral Transformations on a Hypercube
Multiprocessor," Journal of Chemical Physics, Vol. 86, 1987, p. 2185.
70
Taylor, P. R., "Integral Processing in Beyond-Hartree-Fock Calculations,"
71
Pople, J. A., Krishnan, R., Schlegel, H. B., and Binkley, J. S., "Derivative
Studies in Hartree-Fock and M011er-Plesset Theories," International Journal of
Quantum Chemistry Symposium, Vol. 13, 1979, p. 225.
72
Roos, B. O. and Siegbahn, P. E. M., "The Direct Configuration Interaction
Method from Molecular Integrals," Methods of Electronic Structure Theory, edited
by F. Schaefer, Plenum, New York, 1977.
73
Shavitt, I., Advanced Theories, Computational Approaches to the Electronic
Structure of Molecules, edited by C. E. Dykstra, Reidel, Dordrecht, The Neth-
erlands, 1983.
74
Paldus, J. Theoretical Chemistry: Advances and Perspectives, edited by H.
Eyring and D. Henderson, Academic, New York, 1976.
75
Liu, B. and Yoshimine, M., "The Alchemy Configuration Interaction Method.
I. The Symbolic Matrix Method for Determining Elements of Matrix Operators,"
76
Siegbahn, P. E. M., "A New Direct CI Method for Large CI Expansions in a
Small Orbital Space," Chemical Physics Letters, Vol. 109, 1984, p. 417.
77
Knowles, P. J. and Handy, N. C., "A New Determinant-Based Full Config-
uration Interaction Method," Chemical Physics Letters, Vol. Ill, 1984, p. 315.
78
Olsen, J., Roos, B., Jorgensen, P., and Jensen, H., "Determinant Based Con-
figuration Interaction Algorithms for Complete and Restricted Configuration In-
teraction Spaces," Journal of Chemical Physics, Vol. 89, 1988, p. 2185.
79
Knowles, P. J. and Handy, N. C., "Theoretical Studies of the First- and Second-
Row Transition-Metal Methyls and Their Positive Ions," Journal of Chemical Phys-
ics, Vol. 91, 1989, p. 2396.
80
Davidson, E. R., "The Iterative Calculation of a Few of the Lowest Eigenvalues
and Corresponding Eigenvectors of Large Real-Symmetric Matrices," Journal of
Computational Physics, Vol. 17, 1975, p. 87.
81
Bauschlicher, C. W., Langhoff, S. R., andTaylor, P. R., Advances in Chemical
Physics, Vol. 77.
82
Saunders, V. R. and Van Lenthe, J. H., "The Direct CI Method. A Detailed
Analysis," Molecular Physics, Vol. 48, 1983, p. 923.
83
Saebo, S. and Pulay, P., "The Local Correlation Treatment. II. Implementation
and Tests," Journal of Chemical Physics, Vol. 88, 1988, p. 1884.
84
Pople, J. A., Binkley, J. S., and Seeger, R., "Theoretical Models Incorporating
Electron Correlation," International Journal of Quantum Chemistry Symposium,
Vol. 10, 1976, p. 1.
85
Paldus, J. and Cizek, J., "Time-Independent Diagramatic Approach to Per-
turbation Theory of Fermion Systems," Advances in Quantum Chemistry, Vol. 9,
1975, p. 105.
86
Bartlett, R. J., Annual Review of Physical Chemistry, Vol. 32, 1981, p. 359.
87
Rittby, M., Pal, S., and Bartlett, R. J., "Multireference Coupled Cluster Method,
lonization Potentials and Excitation Energies for Ketene and Diazomethane,"
88
Chong, D. P. and Langhoff, S. R., "A Modified Coupled Pair Functional
Approach," Journal of Chemical Physics, Vol. 84, 1986, p. 5606.
89
Ahlrichs, R., Scharf, P., and Ehrhardt, C., "The Coupled Pair Functional
(CPF). A Size Consistent Modification of the CI(SD) Based on an Energy Func-
tional," Journal of Chemical Physics, Vol. 82, 1985, p. 890.
90
Glanditz, R. J. and Ahlrichs, R., "The Averaged Coupled-Pair Functional
(ACPF): A Size-Extensive Modification of MR CI(SD)," Chemical Physics Letters,
Vol. 143, 1988, p. 413.
91
Siegbahn, P. E. M., "The Externally Contracted CI Method Applied to N 2 ,"
92
Werner, H. J. and Reinsert, E.-A., "The Self-Consistent Electron Pairs Method
for Multiconfiguration Reference State Functions," Journal of Chemical Physics,
Vol. 76, 1982, p. 3144.
93
Werner, H. J., "Matrix-Formulated Direct Multiconfiguration Self-Consistent
Field and Multiconfiguration Reference Configuration Interaction Methods," Ad-
vances in Chemical Physics, Vol. 59, 1987, p. 1.
94
Rice, J. E. and Amos, R. D., "On the Efficient Evaluation of Analytical Energy
Gradients," Chemical Physics Letters, Vol. 122, 1985, p. 585.
95
Page, M., Saxe, P., Adams, G. F., and Lengsfield, B. H., "Multireference CI
Gradients and MCSCF Second Derivatives," Journal of Chemical Physics, Vol.
81, 1984, p. 434.
96
Krauss, M. and Stevens, W. J., "Relativistic Quantum Chemistry," Annual
Review of Physical Chemistry, Vol. 35, 1984, p. 5357.
97
Balasubramanian, K. and Pitzer, K. S., Ab Initio Methods in Quantum Chem-
istry, Pt. I, edited by K. P. Lawley, Wiley, Chichester, England, 1987.
98
Thole, V. D., "On the Quantum Mechanical Treatment of Solvent Effects,"
Theoretica Chimica Ada, Vol. 55, 1980, p. 307.
"Tapia, O., Molecular Interactions, edited by H. Ratajczak and W. J. Orville-
Thomas, Wiley, New York, 1980.
100
Thole, V. D., "The Direct Reaction Field Hamiltonian: Analysis of the Dis-
persion Term and Application to the Water Dimer," Chemical Physics, Vol. 71,
1982, p. 211.
101
Thole, V. D., "Molecular Polarizabilities Calculated with a Modified Dipole
Interaction," Chemical Physics, Vol. 59, 1981, p. 341.
102
Medina-Llanos, C., Agren, H., Mikkelsen, K. V., and Jensen, H., "Self
Consistent Reaction Field Calculations of Photoelectron Binding Energies for Sol-
uated Molecules," Journal of Chemical Physics, Vol. 90, 1989, p. 6422.
103
Stevens, W. J., private communication, 1989.
Chapter 2
Rate Coefficient Calculations for

Combustion Modeling
Nancy J. Brown
I. Introduction
T HE intent of this chapter is to describe elements of different approaches

used to calculate rate coefficients. Such calculations range from rough
back-of-the-envelope estimates to highly sophisticated quantum mechan-
ical calculations, and they include a variety of approaches that are classified
as either statistical or dynamical. Statistical theories are based on the laws
of statistical mechanics and require only limited information about the
potential energy surface (PES). Often they are constructed to describe the
probabilities of possible results of single well-defined collisions rather than
equilibrium ensemble-averaged values. Dynamical theories are scattering
calculations that involve simulating the progress of molecular collisions or
rearrangements. The type of dynamical theory depends on the type of
mechanics used to describe molecular motion. Partition functions, numbers
of states, and densities of states are the ingredients of statistical theory.
This chapter is not a review of this vast subject because extensive reviews
covering the various methods already exist. The purpose here is to provide
a list of the essential ingredients, a description of the approximations,
limitations, and recent progress for rate coefficient calculations. Fre-
quently, combustion modelers need values of rate coefficients for elemen-
tary reactions for which experimentally measured values are nonexistent
or poor. Theory can be used to make very good rate coefficient estimates
and to reduce error.
II. Potential Energy Surfaces

The importance of potential energy surfaces, techniques for their cal-
culation, and their use in combustion modeling have been discussed in the
chapter by Page and Lengsfield. Here, we will briefly mention some details
of potential energy systems that are important in rate coefficient calcula-
tions. References 1 and 2 describe the important PES features.
Copyright © 1991 by Nancy J. Brown. Published by the American Institute of Aeronautics

and Astronautics with permission.
37
38 RATE COEFFICIENT CALCULATIONS
In general, a PES is a function of 3N - 5 or 37V - 6 internal coordinates,

where TV is the number of atoms in the reacting system. Most surfaces are
calculated by invoking the Born-Oppenheimer (BO) approximation, and
this implies that the wave function is a product of an electronic wave
function and a nuclear wave function (the vibration/rotation wave func-
tion). Within the BO approximation, the electronic energy, which depends
on the location of the nuclei, provides the PES on which the nuclei move.
Within the clamped nuclei (BO) approximation, where the electronic en-
ergy is computed parametrically as a function of fixed nuclear positions,
wave functions that include spin/orbit interactions frequently are referred
to as the adiabatic basis. However, the term adiabatic is also used to refer
to the wave functions that are eigenfunctions of the electronic Hamiltonian
for the clamped nuclei that exclude the spin/orbit terms. This is the less
correct but more commonly used definition of adiabatic and the one that
is used in most PES calculations. Reactions on such surfaces are called
adiabatic reactions because the electronic state of the system throughout
the course of reaction is associated with a single PES. Nonadiabatic surfaces
are usually called diabatic surfaces.
The information that one needs about a PES depends on the type of
rate coefficient calculation being pursued. Dynamics calculations require
the most detailed information about the PES, reaction path calculations
require less, and transition state calculations, in their simplest form, require
the least amount of detail.
In full dynamics calculations, whether they are quantum, semiclassical,
or quasiclassical, the PES must be represented by a global function of the
37V — 5 or 37V — 6 coordinates. These global functions must span the
configuration space of nuclear motion. Frequently, ab initio calculations
are performed to calculate a limited number of points on the surface, and
an interpolation procedure is used to span configuration space. An excel-
lent discussion of various criteria that a successful interpolating function
must satisfy is given in Ref. 3. In Ref. 4, a review of reactive molecular
collisions discusses these criteria and previous research concerned with
fitting the ab initio points to a global function to generate the PES.
Another approach to obtaining a global PES is to use analytical functions
to approximate the overlap, coulomb, and exchange integrals required for
the surface calculation. This was pursued in Ref. 5 for the H3 system and
in Ref. 6 for H4. In both approaches, the integral approximations were
based on exact values of the integrals obtained directly in ab initio cal-
culations. Figure 1 shows the PES for collinear HA + HBHc-^> HAHB +
Hc as calculated in Ref. 5. The contour lines join points of equal potential
energy. The dashed line indicates the reaction path of minimum energy;
the cross indicates the saddle point, the highest point on this path.
In reaction path treatments, the potential energy must be known in
analytical form along the reaction path. The reaction path has been defined
in Ref. 7, and is, in mass-weighted Cartesian coordinates, the path of
steepest descent from the saddle point to reactants and products. The mass-
weighted reaction coordinate is the reaction path and the distance along
it. There are accurate and efficient techniques for the ab initio quantum
mechanical calculation of the PES gradient with respect to the nuclear
N. J. BROWN 39
0.20 -
0.15 -
0.10
0.05 -
0.05
/nm
Fig. 1 Potential energy surface for collinear HA + HBHC —» HAHB + Hc as cal-

culated by Porter and Karplus.5 The contour lines join points of equal potential
energy. The dashed line indicates the reaction path of minimum energy, and the
cross indicates the transition state, the highest point on this path.
coordinates. These are used most frequently to determine the saddle point,
and the same techniques can be used to calculate the reaction path. Miller
et al.8 describe the construction of a classical Hamiltonian for a general
molecular system based on the reaction path and a harmonic approximation
to the PES about the reaction path.
Finally, in conventional transition state theory (TST), where the saddle
point is used as the transition state dividing surface, the barrier height,
geometry of the saddle point, and the vibrational frequencies of the tran-
sition state are required. In generalized TST, there must be enough PES
data to define and characterize the transition state that is not necessarily
chosen as the saddle point as it is in conventional TST. Figure 2 shows the
relationship between the PES and the various types of rate coefficient
calculations and their relationship to sensitivity analysis calculations.
Although not concerned specifically with potential energy surfaces, Ref.
9 is a veritable treasure house of information relating molecular structure
characteristics to chemical reactivity, providing an impressive discussion
of semiempirical methods for estimating Arrhenius parameters for a num-
ber of different types of reactions. It also gives additivity rules for estimating
thermochemical properties. In this author's opinion, Ref. 9 is an absolute
requirement for every combustion modeler's bookshelf.
III. Transition State Theory

A. Theoretical Considerations
Transition state theory has been used extensively for the calculation of
rate coefficients for over 50 years. Combustion modelers most frequently
Fig. 2 Schematic diagram showing the relationship between the PES and the various
types of rate coefficient calculations and their relationship to sensitivity analysis
calculations.
use TST to calculate rate coefficients. Transition state theory has been the
object of renewed theoretical interest since the early 1970s. There are a
number of texts that describe TST.10"19 Pechukas20 21 has written two ex-
cellent reviews that describe characteristics of transition state theory and
research in the area. References 22 and 23 are more specialized reviews
concerned with TST. Reference 22 is especially useful in providing detailed
information about methodology for actual rate coefficient determinations.
Reference 24 presents a historical review of the theoretical developments
of TST.
Microcanonical refers to an isolated system with a given value of energy,
volume, and number of particles. Canonical refers to an isolated system
with a given value of temperature (average energy), volume, and number
of particles.
TST is paradoxical in that it is simple in construction and subtle in its
consequences. To paraphrase some of the excellent material of Pechu-
kas,20'21 TST is a statistical theory, and the statistics in this theory involve
counting the number of ways a system can pass through a transition state.
Although statistical concepts are employed to avoid the computational
intensity of dynamical treatments, the essence of TST is founded in dy-
namical concepts. The idea of a transition state invokes dynamical consid-
erations, specifically the dynamical instability that occurs in passing through
the transition state where there are reactants on one side and products on
the other.
There are two major assumptions that characterize TST. First, a local
equilibrium is assumed to exist between reactants and species that originate
as reactants and become transition state species. Second, it is assumed that
any species passing through the transition state does so only once. The
N. J. BROWN 41
first of these is called the local equilibrium approximation and the second
is the no-recrossing assumption. Within the framework of classical me-
chanics, the net rate of reaction in the forward direction is given by the
flux of trajectories from the reactant phase space passing through the
transition state dividing surface to the product phase space. The transition
state is truly a point of no return, and implicit in the theory is the idea
that once the system passes through the transition state from the reactant
side of the PES, it is, by definition, in product space forever more. The
exactness of TST at a fixed energy E, microcanonical TST, can be verified
by following all trajectories that leave the transition state, and if there is
no recrossing, the theory is exact. Canonical TST is exact if, and only if,
no trajectory of any energy crosses the transition state more than once, a
significantly more stringent requirement. The rate coefficients from clas-
sical TST are verified with a dynamical criterion, and they must be com-
pared with rate coefficients calculated entirely classically, without
quasiclassical quantization of initial states using standard trajectory cal-
culations. When one speaks about TST being exact, it implies exact agree-
ment between the rate coefficient calculated from TST and a purely classical
trajectory calculation of the rate coefficient.
For the prototypical reaction A -h B —> C + D, the TST expression for
the forward rate coefficient kf is
kBT D'
k
f= -T
where kB is the Boltzmann constant, h is Planck's constant, and T is tem-
perature (in degrees Kelvin). The partition function for molecule A per
unit volume is given by
QA = h-^ J dl^ exp[-HA/kBT\ (2)

where NA is equal to the number of atoms in A, HA is the Hamiltonian of
the isolated molecule A, and the volume element dP^ is such that A remains
in a single electronic state. The volume element is restricted such that the
center of mass of A is confined to a unit volume. The partition function
QB is defined in a similar manner. The partition function Q; of the transition
state per unit volume is slightly more complicated because it is not asso-
ciated with a stable entity but rather that portion of phase space that the
reactant system N = NA + NB must pass through to become products.
For conventional TST, the dividing surface is the saddle-point region. With
the reaction coordinate and its conjugate momentum separated out, Q*
can be written as
r
dP exp[ — H'/kBT] (3)
where the integral is over the remaining N - 1 coordinates and N — 1
conjugate momenta. The transition state has a fixed value of the reaction
coordinate and, thus, one degree of freedom less than the reactants. Note
that factors such as h~3NA are included in the classical partition function
to give an approximate counting of the quantum states. It is also assumed

that all atoms are distinguishable or individually countable.
If TST calculations of the rate coefficient differ from values determined
from experimentally measured values, the differences are due to errors in
the PES, in the theoretical formulation, or in the experiment. If TST is in
error relative to an exact classical dynamics calculation, it is because re-
crossing trajectories, which increase with energy in excess of the classical
threshold, exist. This implies that classical TST provides an upper bound
to the true classical rate coefficient, and this provides a variational criterion
for TST, implying that one should choose the dividing transition state
surface to minimize the reactant flux through it. Note that the variational
criterion exists only with respect to the "true" classical rate coefficient
calculated solely with classical mechanics with no quasiclassical consider-
ations. In conventional TST, for a PES with a saddle point, the phase-
space dividing surface is chosen to be a function of coordinates only (mo-
menta are excluded) and is located so that it passes through the saddle
point. The transition state is a plane determined by normal coordinate
analysis of small vibrations around the saddle point with displacement along
the reaction normal coordinate set equal to zero. Any other dividing surface
is by definition a generalized transition state. Variational TST is the name
we apply to theory that uses the minimum flux (microcanonical) or max-
imum free energy of activation (canonical) criterion for locating the divid-
ing surface. The minimum free energy criterion was misinterpreted in Ref.
15 to imply a minimum density of states criterion, and invoking this does
not yield the correct variational rate coefficient.
The best dividing surface for variational microcanonical TST for collinear
atom-diatom reactions was determined by Pechukas,20 Pollak and Pechu-
kas,25 and Sverdlik and Koeppl.26 They found that the dividing surface is
a curve joining the two equipotentials that bound the classically allowed
region of the coordinate plane. Possible candidates for the best dividing
surfaces are curves traced out by classical trajectories vibrating between
the two equipotentials. If there is only one such vibrational energy, then
that path is automatically the best dividing surface. If there are more than
one, the one with the least reactant flux associated with it is best.
When the potential expands in either direction away from the saddle
point, the best choice of the transition state dividing surface is in the vicinity
of the saddle point. What is the optimum dividing surface for more com-
plicated reactions in three dimensions? Considerable research on varia-
tional TST for atom/molecule reactions in three dimensions and for some
molecule/molecule reactions has been summarized and discussed in Refs.
22 and 23.
B. Application of TST
Our discussion has been confined to classical TST to illustrate to the
reader the theoretical foundations of the theory, and the limits of appli-
cability, and to avoid possible misconceptions. The next question to address
is how to calculate rate coefficients using TST. The most popular approach
is to perform a type of hybrid quantum calculation. In the quantization of
N. J. BROWN 43
conventional TST, there are three assumptions. First, we assume that mo-
tion associated with the reaction coordinate is separable from the other
degrees of freedom. As stated most clearly in Ref. 14, this approximation
is reasonable as long as the PES near the saddle point is well approximated
by a quadratic function over an area whose linear dimensions are much
larger than the de Broglie wavelength. Second, we assume that the reaction
coordinate motion can be treated classically and that the energy levels
associated with the remaining bound degrees of freedom of the transition
state and those of the reactant are quantized. Third, we frequently assume
that the rate coefficient can be multiplied by a transmission coefficient to
correct for nonclassical motion along the reaction coordinate. There is also
the possibility of quantum mechanical tunneling through the saddle-point
barrier at energies below the classical threshold and the possibility of re-
flection from the barrier at energies above threshold.
When the hybrid quantum approach is used, TST no longer provides an
upper bound to the rate coefficient for a given PES. There is no variational
principle associated with a rate coefficient calculated in this manner. This
is also not a true quantum mechanical approach because the fundamental
assumption of TST violates the uncertainty principle by requiring that the
position and momentum along the reaction coordinate must be specified
simultaneously.
There are a number of requirements for both hybrid conventional and
generalized transition state theories. Rule 1 is to know what the zero of
energy is for the reactant partition functions and for the transition state
partition functions. Be internally consistent. If there are more than one
reactant, a relative translational partition function per unit volume,
QA B, must be included in the reactant partition function. This, for three-
dimensional motion, is
GJ.fl(T) = (2^kBTlh^ (4)
where |JL is the reduced mass. Since the relative translational partition
function is not a function of the reaction coordinate, it is the same for
conventional and generalized transition states. The same is true for elec-
tronic partition functions, but remember that the electronic energy is meas-
ured relative to the zero of energy, which for the transition state is the
transition state energy. The expression for the electronic partition function
for species A is
QeA(T) = S dA^ exp[-e^/^71 (5)
a
where the degeneracy (multiplicity) is dA a and the energy of the state e^ a

is measured relative to the ground electronic state, which is the zero of
energy. It is also important to include the electronic partition function for
atoms if they are reactants.
The geometry and frequencies of the saddle point must be determined
to evaluate rotational and vibrational partition functions in conventional
TST. The imaginary frequency associated with the reaction coordinate
motion is required to correct reaction coordinate motion for nonclassical
effects, which frequently are not large for temperatures of interest in com-
bustion studies. Truhlar et al.22 have given excellent descriptions of tech-
niques for locating generalized transition states. Once transition states are
defined, the moments of inertia and vibrational frequencies must be eval-
uated. Then, the rotational partition function for a nonlinear molecule A
(assuming temperature to be in the range appropriate for combustion) is
Q« = (vl«r\2kBTlh^'\IAIBlcy* (6)
where IA,1B-, and Ic are the equilibrium moments of inertia associated with
the principal axis of rotation. This formula holds for spherical top mole-
cules, IA = IB = 7C, or symmetric top molecules, IA < !B = Ic or IA =
IB < Ic. The rotational partition function for a linear molecule is
QKA = ± (2kBT/h2)I (7)
The symmetry number cr must be included with the rotational partition

functions, and this corrects for repeated counting of indistinguishable con-
figurations in the classical phase integral. The vibrational partition function
for a molecule with m normal modes is
Q\ = [1 exp[~hcvJ2kBT] (1 - exp[-hci>m/kBT]) (8)
m
The zero of energy is assumed to be located at the bottom of the potential

well where the internuclear separation is the equilibrium value, vm is the
vibrational frequency (in cm" 1 ) for the rath oscillator, and c is the velocity
of light. Anharmonic effects become important, and neglecting them results
in rate coefficients that are too large. In particular, for stretching modes,
anharmonic effects are especially important for tight transition states at
low temperatures. For bending modes, anharmonicity becomes important
at higher temperatures. Anharmonic effects can be accounted for using
the Pitzer-Gwinn approximation (PGA).27 The PGA is based on the idea
that the ratio of the anharmonic vibrational partition function to the har-
monic value is given correctly by the classical approximation at both the
low and high temperature limits. The PGA is to assume that the ratio is
correct at all temperatures.
C. Symmetry Corrections for Rate Coefficient Calculations

Symmetry must be properly accounted for in rate coefficient calculations.
Although this section is concerned with TST calculations, the symmetry
corrections described here are relevant to dynamics calculations as well.
The literature contains many contradictions associated with this subject,
and many of the standard texts in chemical kinetics are in error. Papers
by Pollak and Pechukas25 and Coulson28 describe symmetry corrections
correctly. These papers are important because they clarify the misconcep-
tions and capture the essence of the problem. Briefly, nuclear spin degen-
eracy can be ignored because nuclei are the same in the transition state as
in the reactants and product and, thus, the degeneracies cancel. Symmetry
numbers should be used rather than reaction path degeneracies or statistical
N. J. BROWN 45
factors. The symmetry number is the number of different ways in which

the molecule can achieve, by rotation, the same orientation in space by
counting similar atoms as indistinguishable. Rotational partition functions
should be divided by molecular symmetry numbers to account for the effect
of quantum statistics. Symmetry numbers should appear in both classical
and quantum partition functions. Symmetry numbers also should be used
for the rotational partition functions of the transition state. Pechukas20 21
and Metiu et al.29 provide excellent discussions of reaction path symmetry
that should be used as input when transition state properties are deduced
semiempirically. For conventional TST, the symmetry of the transition
state is limited to the joint symmetries of the reactant and product unless
they are physically indistinguishable or symmetry-related, i.e., optical iso-
mers, and for these the transition state may have additional symmetries.
D. Tunneling Corrections
Frequently, transition state calculations include a tunneling correction,
which is a quantum mechanical correction to reaction coordinate motion.
The tunnel effect is the name given to describe the entire quantum me-
chanical barrier crossing problem, which includes barrier penetration or
transmission for energies less than the barrier height and barrier reflection
corrections for energies that are greater. Tunneling is usually important
for light atoms and molecules at low temperatures, and for combustion
applications, it may be important for reactions involving H atoms.
Johnston14 very carefully gives some general criteria for assessing the
nature of quantum effects on reaction coordinate motion. He suggests
performing a normal coordinate analysis along the reaction coordinate to
obtain the mass m* and then using m* to compute the Boltzmann-averaged
de Broglie wavelength as
A* - (h2l2vm*kBT)m (9)
If the potential is flat over distances that are large compared with A*, then
the reaction coordinate motion is considered classical. If it is flat over a
distance that is small compared with A*, then the reaction coordinate
motion is nonseparable and very difficult to treat. If the PES near the
saddle point is well approximated as a quadratic function over an area
whose linear dimensions are comparable to A*, then the motion is separable
but not classical and requires a tunneling correction.
There are many approaches for treating tunneling. An especially fine
description is given in Ref. 22. Herein we will mention the more commonly
used approaches. The imaginary frequency associated with reaction co-
ordinate motion must be determined for all of these corrections.
The simplest and most common method of approximating tunneling is
the semiclassical Wigner30 approximation:
K W - 1 + (1/24) \fiw'/kT\2 (10)
This tunneling correction is the transmission coefficient obtained by Boltz-
mann averaging of the semiclassical barrier penetration probabilities. The
symbol vvf is the imaginary frequency associated with reaction coordinate
motion. For this approximation to be valid, tunneling contributions must

come entirely from the saddle-point region, implying that transverse vi-
brations do not vary appreciably and the potential along the reaction path
is well approximated by an inverted parabola. Wigner corrections are most
likely to be valid when the correction is small and less than 2.
Other commonly used tunneling corrections are based on assumptions
of barrier shapes in the saddle-point region. Two very common barriers
worth noting are parabolic and Eckart barriers. The assumption of a fixed
shape barrier sometimes leads to large errors in the tunneling correction.
There is an excellent discussion of Eckart and parabolic barriers in Chap.
2 of Ref. 14 and in Ref. 31. [Note: For those interested in using an Eckart-
type correction, Eq. (2.22) in Ref. 14 is in error. The 2ir2 should be replaced
by 47T2.]
E. Complex Reactions
Statistical theory for reactions that proceed through a long-lived collision
complex is less rigorous, in its dynamical foundations, than the TST of
direct reactions. Many important combustion reactions have multiple tran-
sition states, and the dynamics in the region between the outermost tran-
sition states involve passing through a deep well in the PES, and these
reactions are very often complex rather than direct. In complex reactions,
the variational principle associated with classical TST may fail. Assuming
that potential wells are deep enough for at least one bound state, reactions
along pathways with multiple transition states result in complex formation
at some fixed energy. Formation of the complex is a bimolecular reaction,
and its dissociation is a unimolecular one. There is much confusion in the
literature about complex reactions, and much of it concerns the bound-
edness of the calculated rate coefficient. References 32-34 investigate the
bounds associated with rate coefficients of complex reactions in detail and
are careful in discussing the constraints on the reactive system that must
occur for the rate coefficient to be bounded as it is for a direct reaction.
The importance of complex reactions to combustion is illustrated by a
set of rate coefficient calculations by Miller and colleagues (see, for ex-
ample, Refs. 35-37). Melius and Binkley3839 used the BAC-MP4 method
to calculate transition state geometries, vibrational frequencies, and en-
thalpies for each of the molecular species. Rate coefficients were calculated
by using a hierarchy of statistical approximations: canonical theory, can-
onical theory with a Wigner tunneling correction, microcanonical theory,
microcanonical theory with angular momentum conservation, and micro-
canonical theory with angular momentum conservation with one-dimen-
sional tunneling. The tunneling correction for the microcanonical rate
coefficients was determined using the method recommended in Ref. 40,
in which a quantum mechanical flux through the dividing surface was ap-
proximated. Three assumptions were invoked to calculate the rate coef-
ficients: the RRKM or strong coupling assumption, the TST assumption
of no crossing of the final transition state leading to products, and steady-
state approximations for all complexes. Rate coefficients and branching
N. J. BROWN 47
ratios were calculated for the product channels of the reactions O + HCN,
OH + HCN, and OH 4- C2H2.
IV. Molecular Dynamics
Dynamical calculations are classified as being quantum, semiclassical,
or classical. Quantum dynamics are based on exact or approximate solu-
tions to Schroedinger's equation. Classical mechanics and, most frequently,
the solution of Hamilton's equations, describe the dynamics in a classical
calculation. Semiclassical treatments of dynamics use real- and complex-
valued solutions of Hamilton's equations to construct an asymptotic so-
lution of Schroedinger's equation. The proper limit of a semiclassical cal-
culation is classical mechanics. Semiclassical calculations, which will not
be discussed further herein, have the particular advantage of extending
classical theory to describe classically forbidden events.
Schroedinger's equation is the only rigorous and valid approach for treat-
ing dynamics and reactivity on the atomic scale. Three-dimensional exact
quantum mechanical calculations of reactivity have been performed to
study reactivity and energy exchange in H + H2 and its isotopic analogs.
These calculations are described in Refs. 4 and 41. Currently, there is
substantial research activity (see, for example, Refs. 42-44) devoted to
improving three-dimensional quantum mechanical calculations on the H
+ H2 system, and exact calculations are currently under way on the F +
H2 system. The developments in the field of molecular quantum dynamics
are occurring at an exciting pace, and they will provide some intriguing
results that should stimulate further progress in PES construction and in
state-to-state experiments measuring reaction attributes.
Quantum mechanical calculations are not practical for determining ther-
mal rate coefficients for combustion applications due to the higher energies
associated with combustion and the concomitantly large number of coupled
opened and closed channels that must be included in the wave-function
expansion. Moreover, quantum effects such as tunneling, interference phe-
nomena, and resonances are not likely to be important for reactions at
temperatures above 1000 K. For predicting ignition characteristics at tem-
peratures lower than 1000 K, quantum effects, although usually small, may
be important. The effects of combustion (i.e., higher energies and higher
masses) tend to decrease the de Broglie wavelength of the reactive system,
X - hip (11)
wherep is the momentum. As X decreases to a size smaller than the distance
over which the potential changes significantly, the system can be treated
classically. Furthermore, the amount of averaging necessary to compute a
thermal rate coefficient from state-to-state rate coefficients is a smoothing
process that removes much of the quantum behavior while retaining the
features produced by a classical calculation. Quantum mechanics is nec-
essary for the calculation of transition probabilities for reactions where the
energy is close to but less than the barrier height (for so-called classically
forbidden processes) and for describing interference phenomena.
Classical mechanics is a very popular and often quite accurate approach

for treating reactions between an atom and a molecule and between two
molecules. Hamilton's equations
-*
must be solved. In Eq. (12), H(p,q) is the system Hamiltonian; q( and p(
are the iih coordinate and momentum, and the dot indicates the time
derivative. For a given PES and selected values of pt and q( at some initial
time, the classical laws of motion completely define the subsequent dy-
namical behavior of a system and the classical path or trajectory of the
collision can be calculated by numerically integrating the 6N - 6 differ-
ential equations (Hamilton's equations), where N is the number of atoms
in the system. The calculation is called quasiclassical if the molecule is
initially given energy that corresponds to a defined vibrational/rotational
state. Trajectory calculations are useful because they yield reaction cross
sections, angular distributions, product energy and angular momentum
distributions, final energy and angular momentum distributions in non-
reactive collisions, and they also provide insight into the nature of the
reactions.
Reference 45 is a seminal paper in the field of quasiclassical dynamics
that describes calculations on the H + H2 reaction. There are two excellent
reviews that describe how to perform a quasiclassical calculation of a rate
coefficient. The first is by Truhlar and Muckerman 46 and is concerned with
all aspects of quasiclassical calculations for the system A + BC. The second
is by Porter and Raff, 47 and it describes the calculation of reaction attributes
by the quasiclassical method for A + B, A + BC, AB + CD, and A +
BCD collisions; it is particularly good in describing the selection of initial
conditions. Reviews have also been written by Connor4 and by Walker
and Light41; these are useful because they survey work that uses dynamical
approaches to studies of reactivity and energy transfer. References 19, 48,
and 49 are highly recommended as sources describing the molecular dy-
namics (quasiclassical) approach to rate coefficient calculations.
There are three parts to a molecular dynamics calculation: 1) specifi-
cation of the initial conditions, 2) integration of Hamilton's equations, and
3) determination of the molecular properties at the end of the collision
when the collision partners are no longer interacting. Monte Carlo tech-
niques are then used to average over an ensemble of trajectories, N tra-
jectories, each belonging to the same "state." The initial values of the
nonconsequential (i.e., those not crucial to defining the state) coordinates
and momenta are selected randomly according to their distribution in phase
space. The number of trajectories required for convergence of the ensemble
average depends on the precision required for calculation of the variable
of interest. Important work on improving sampling to render more rapid
convergence has been reported in Refs. 50 and 51.
N. J. BROWN 49
Specification of initial conditions is relatively straightforward for A 4

B, A 4 BC, and AB 4 CD, and is more complicated for A 4 £CZ).
Reference 47 provides an excellent discussion of the specification of initial
conditions.
A quasiclassical trajectory is defined as one starting from a defined
vibrational/rotational state. The initial values of the coordinates and mo-
menta, which define the trajectory, fix the energy in the system and de-
termine the system's relative orientation. An ensemble of trajectories are
TV trajectories that have the same value of translational energy, rotational
energy, and vibrational energy. The ensemble average is performed by
using Monte Carlo techniques whereby coordinates and momenta are se-
lected randomly according to their distribution in phase space with energies
fixed at values defining the ensemble. Figure 3 is an Arrhenius plot of
calculated values of the H2 4 OH —» H2O 4- H rate coefficient. The
Zellner values are the experimental values due to Zellner and Steinert, 52
the MCQCD values are Monte Carlo quasiclassical dynamics values, and
the MJCT values are the results of TST calculations that are thermally
averaged microcanonical values calculated with full angular momentum
conservation with approximate tunneling calculations. The calculated val-
ues are due to Brown et al.,53 the designation "ab initio" indicated the
PES transition state properties of Walch and Dunning, 54 and "fitted" is
due to fitted potential parameters of Schatz and Elgersma.55 Note the very
good agreement among the different theoretical approaches and experi-
ments at combustion temperatures.
There are a number of ordinary-differential-equation solvers available,
and an investigation of mathematical libraries is recommended. In partic-
ular, there has been a great deal of research on efficient algorithms for
stiff ordinary differential equations at the Lawrence Livermore National
Laboratory. It is important to check the solution of Hamilton's equations,
and an obvious way to do this is to verify that the three components of
angular momentum and total energy are conserved. This is a necessary but
insufficient check of the integration. The relationship between difference
equations and differential equations is not unique, therefore, additional
verification is required. An excellent check is to integrate the trajectory
backward to replicate the initial conditions. Backward integration and con-
servation of energy and angular momentum are sufficient checks. Ob-
viously, backward integration is not possible for all trajectories, but it
should be attempted for a representative sample that includes long-lived
trajectories, if they exist.
Each trajectory is integrated in time until the collision partners are no
longer interacting, which implies that the ratio of translational energy to
potential energy of interaction is quite large. A decision regarding whether
or not the system has reacted is usually made prior to the termination of
the trajectory. Reaction is often defined in terms of the extension of some
bond coordinate. In the case of A 4 BC reactions, the pair of atoms with
the smallest internuclear separation is usually considered a stable molecule,
and this is confirmed by calculating the internal energy of the molecule.
All of the so-called integration parameters—e.g., step size, separation of
T (K) / 100
24 10 6 4 3
-10 T
HH + OH ~> H + HOH
A MJCT-fittQd
-11 + MJCT-ab initio
a MCQCD
-12
o
01
o
-14-
OT
O
-15-
-16-
-171
1000 K / T
Fig. 3 Arrhenius plot of calculated values of the H2 + OH H2O + H rate
coefficient.
reactants or products at the termination of the trajectory, and maximum

impact parameter—must be determined for each system.
After the trajectory is terminated, the so-called collision attributes have
to be determined for the reactants if there is no reaction and for the
products if there is. A reaction cross section is calculated for an ensemble
of trajectories and is given by the following formula:
S = b2mPr (13)
where bm is the maximum value of the impact parameter for which reaction
occurs, and Pr is the reaction probability for the ensemble. The reaction
probability is given by the ratio of the numbers of reactive to total collisions.
Other frequently calculated attributes are the scattering angle, the final
N. J. BROWN 51
molecular energy and its decomposition into vibrational and rotational

energy, the final orbital and molecular angular momentum, and the final
translational energy. These attributes can be used to calculate cross sections
and other derived quantities for the ensemble or a state-to-state rate coef-
ficient. To calculate a thermal rate coefficient, it is necessary to average
the state-to-state rate coefficients according to their equilibrium distribu-
tion at a given temperature.
In quasiclassical calculations, it is often important to separate the ro-
tational and vibrational contributions to molecular energy. For collisions
involving one or more diatomic molecules, a rigid rotator Morse oscillator
model is used frequently for the internal motion of the diatom. The internal
energy calculated using the model is equated to a spectroscopically derived
value of the energy for a given value o f v , the vibrational quantum number,
and /, the rotational quantum number. The turning points /?+ and R —
are solved for, and these and the angular molecular momentum Jh are
used to describe the initial state of the molecule. To separate the total
molecular energy into rotational and vibrational energies, the following
relationship must be used:
erot = min{VD(r) + [(Jr • Jr)/2^r2]} - VD(re) (14)
where e rot is the average rotational energy, VD(re) the classical zero of
internal energy (usually for a Morse potential), r the internuclear sepa-
ration, and subscript e denotes equilibrium separation. The minimum in
the effective potential given by the expression in brackets is usually solved
for using the Newton-Raphson iteration technique. The vibrational energy
is obtained from the total molecular energy minus the calculated average
rotational energy. When the molecule is a triatomic or greater, the situation
is more complex because the Coriolis interaction couples the vibrational
and rotational degrees of freedom. For studies with triatomics, the most
efficient coordinate system for investigating collisions is the space-fixed
frame. However, when a more rigorous separation between rotational and
vibrational motion is required, the individual contributions to the energy
cannot be separated using the space-fixed frame because of the Coriolis
interaction. Suzukawa et al.56 provide a good description of vibrational
and rotational motion separation in triatomic molecules. Sometimes it is
not necessary to know the individual contributions. In their studies of the
unimolecular dissociation of HO2, Miller and Brown57 and Brown and
Miller58 required only values of total angular momentum and total molec-
ular energy.
V. Sensitivity Analysis
An important question to answer in rate coefficient calculations is how
sensitive the calculated rate coefficient is to the PES and where in config-
uration space do we have to know the PES best. The answer, of course,
depends on the characteristics of the PES, the types of molecules reacting,
and the initial states of the react ants. Frequently, answers to these questions
have been sought by attempting systematically to vary features of the PES
and determining the concomitant effect on the rate coefficient. There are
endless variations (many of which are not unique) that can be attempted,
and it is quite difficult to organize the results of such an analysis in a very
systematic manner. Isotopic substitution has also been used to gain insight
into how different general parts of the PES affect observables. This is
useful because the PES is the same for all of the isotopes, but for each of
them the dynamics probes different regions of the potential.
A more systematic approach to the problem of determining how structure
in the PES maps itself into the observables associated with a reacting system
is provided by functional sensitivity analysis. The combustion community
is familiar with parametric sensitivity analysis through its use in the detailed
modeling of species and temperature profiles in flames and from modeling
of temporal species profiles in well-stirred and plug flow reactors. This
approach, particularly that associated with the use of Green's function,
has been richly developed by Rabitz59 and his colleagues. In these appli-
cations, parametric sensitivity analysis frequently has been employed since
the rate and transport coefficients are usually thought of as discrete pa-
rameters. In rate coefficient calculations, the PES is actually a function of
37V - 5 or 37V - 6 coordinates despite the fact that it may be parameterized
in numerical calculations. The PES is treated as a function, and functional
sensitivity analysis can be used to determine how the observables associated
with the reacting system depend on features of the PES. In using functional
sensitivity analysis, one calculates the gradient of an observable with re-
spect to a perturbation to the potential at a particular point in configuration
space. The sensitivity function directly indicates how a perturbation in the
potential will be translated into a change in the observable. The application
of functional sensitivity analysis to classical dynamics of molecular systems
was initiated by Judson and Rabitz. 60
Functional sensitivity analysis has been used to treat inelastic collisions
in the H2 + H2 system and its isotopic analogs by Judson et al.61 They
also used functional sensitivity analysis to investigate reactivity in the col-
linear F + H2 system.62 In both of these calculations, classical mechanics
has been used to treat the dynamics. Both of these calculations have shown
that the physics governing collisions in these problems are complex and
that there is a highly structured relationship between details in the potential
energy system and the observables. The highly structured nature of the
sensitivities implies that there are localized regions of the potential that
are especially important to the dynamics. These localized regions are those
where a high density of ab initio points should be calculated and where
minimal error should be introduced by fitting procedures.
Although studies of the dynamics using sensitivity analysis are still in
the nascent stage, sensitivity analysis provides a set of powerful tools to
analyze models for their physical content and mathematical behavior. It
provides a means in all of its applications to probe the interrelationship
between input (PES) and output (observables associated with the reacting
system) functions. The development and application of functional sensi-
tivity analysis to studies of reactive scattering will provide exciting new
kinds of information regarding the mechanisms of reactivity on the micro-
scopic level.
N. J. BROWN 53
Acknowledgments
This work was supported by the Director, Office of Energy Research,
Office of Basic Energy Sciences, Chemical Sciences Division of the U.S.
Department of Energy under Contract DE-AC3-76SF0098.
References
Simons, J., Energetic Principles of Chemical Reactions, Jones and Bartlett Pub-
lishers, Boston, MA, 1983.
2
Murrell, J. N., Carter, S., Farantos, S. C., Huxley, P., and Varandas, A. J.
C., Molecular Potential Energy Functions, Wiley, New York, 1984.
3
Wright, J. S. and Gray, S. K., Journal of Chemical Physics, Vol. 69, 1978,
p. 67.
4
Connor, J. N. L., Computer Phys. Comm., Vol. 17, 1979, p. 117.
5
Porter, R. N. and Karplus, M., Journal of Chemical Physics, Vol. 40, 1964,
p. 1105.
6
Silver, D. M. and Brown, N. J., Journal of Chemical Physics, Vol. 72, 1980,
p. 3859.
7
Fukui, K., Kato, S., and Fujimoto, H., Journal of the American Chemical
Society, Vol. 97, 1975, p. 1.
8
Miller, W. H., Handy, N. C., and Adams, J. E., Journal of Chemical Physics,
Vol. 72, 1980, p. 99.
9
Benson, S. W., Thermo chemical Kinetics, 2nd ed., Wiley, New York, 1976.
10
Glasstone, S., Laidler, K. J., and Eyring, H., The Theory of Rate Processes,
McGraw-Hill, New York, 1941.
H
Laidler, K. J., Theories of Chemical Reaction Rates, McGraw-Hill, New York,
1969.
12
Eyring, H., Lin, S. H., and Lin, S. M., Basic Chemical Kinetics, Wiley, New
York, 1980.
13
Eyring, H., Walter, J., and Kimball, G. E., Quantum Chemistry, Wiley, New
York, 1944.
14
Johnston, H. S., Gas Phase Reaction Rate Theory, Ronald Press, New York,
1966.
15
Bunker, D. L. and Pattengill, M., Journal of Chemical Physics, Vol. 48, 1968,
p. 772.
16
Weston, R. E., Jr. and Schwartz, H. A., Chemical Kinetics, Prentice-Hall,
Englewood Cliffs, NJ, 1972.
17
Nikitin, E. E., Theory of Elementary Atomic and Molecular Processes in Gases,
Oxford University Press, London, England, 1974.
18
Smith, I. W. M., Kinetics and Dynamics of Elementary Gas Reactions, But-
terworths, London, England, 1980.
19
Levine, R. and Bernstein, R. B., Molecular Reaction Dynamics and Chemical
Reactivity, Oxford University Press, New York, 1987.
20
Pechukas, P., "Statistical Approximation in Collision Theory," Dynamics of
Molecular Collisions, Part B, edited by W. H. Miller, Plenum, New York, 1976,
p. 269.
21
Pechukas, P., Annual Review of Physical Chemistry, Vol. 32, 1981, p. 159.
22
Truhlar, D. G., Isaacson, A. D., and Garrett, B. C., "Generalized Transition
State Theory," The Theory of Chemical Reaction Dynamics, Vol. 4, 1985, p. 65.
23
Truhlar, D. G. and Garrett, B. C., Annual Review of Physical Chemistry, Vol.
35, 1984, p. 159.
24
Laidler, K. J. and King, M. C, Journal of Physical Chemistry, Vol. 87, 1983,
p. 2657.
25
Pollak, E. and Pechukas, P., Journal of the American Chemical Society, Vol.
100, 1978, p. 2984.
26
Sverdlik, D. I. and Koeppl, G. W., Chemical Physics Letters, Vol. 59, 1978,
p. 449.
27
Pitzer, K. S. and Gwinn, W. D., Journal of Chemical Physics, Vol. 10, 1942,
p. 428.
28
Coulson, D. R., Journal of the American Chemical Society, Vol. 100, 1978,
p. 2992.
29
Metiu, H. J., Ross, R. S., and George, T. F., Journal of Chemical Physics,
Vol. 61, 1974, p. 3200.
30
Wigner, E. P., Zeitschrift fuer Physikalische Chemie Abt. B, Vol. 19, 1932,
p. 203.
31
Truhlar, D. G. and Kuppermann, A., Journal of the American Chemical
Society, Vol. 93, 1971, p. 1840.
32
Pollak, E., Child, M. S., and Pechukas, P., Journal of Chemical Physics, Vol.
72, 1980, p. 1669.
33
Pollak, E. and Child, M. S., Journal of'Chemical Physics, Vol. 73,1980, p. 4373.
34
Pechukas, P. and Pollak, E., Journal of Chemical Physics, Vol. 67, 1977,
p. 5976.
35
Miller, J. A., Parrish, C., and Brown, N. J., Journal of Physical Chemistry,
Vol. 90, 1986, p. 3339.
36
Miller, J. A. and Melius, C. F., Twenty-First Symposium (International) on
Combustion, The Combustion Institute, Pittsburgh, PA, 1988, p. 919.
37
Miller, J. A. and Melius, C.F., Twenty-Second Symposium (International) on
Combustion, The Combustion Institute, Pittsburgh, PA, 1989.
38
Melius, C. F. and Binkley, J. S., ACS Combustion Symposium, 1984, p. 103.
39
Melius, C. F. and Binkley, J. S., Twentieth Symposium (International) on
40
Miller, W. H., Journal of the American Chemical Society, Vol. 101, 1979,
p. 6810.
41
Walker, R. B. and Light, J. C., Annual Review of Physical Chemistry, Vol.
31, 1980, p. 401.
42
Pack, R. T. and Parker, G. A., Journal of Chemical Physics, Vol. 90, 1989,
p. 3511.
43
Pack, R. T. and Parker, G. A., Journal of Chemical Physics, Vol. 87, 1987,
p. 443888.
Zhang, J. Z. H. and Miller, W. H., Chemical Physics Letters, Vol. 140, 1987,
p. 329, also, Zhang, J. Z. H, Kouri, D. J., Haug, K., Schwenke, D. W., Shima,
Y., and Truhlar, D. G., Journal of Chemical Physics, 1988, p. 2492.
45
Karplus, M., Porter, R. N., and Sharma, R. D., Journal of Chemical Physics,
Vol. 43, 1965, p. 3259.
46
Truhlar, D. G. and Muckerman, J. T., "Reactive Scattering Cross Sections
III: Quasiclassical and Semiclassical Methods," Atom-Molecule Collision Theory,
A Guide for the Experimentalist, edited by R. B. Bernstein, Plenum, New York,
1979,
47
pp. 505-566.
Porter, R. N. and Raff, L. M., "Classical Trajectory Methods in Molecular
Collisions," Dynamics of Molecular Collisions, Part B, edited by W. H. Miller,
Plenum, New York, 1976, p. 1.
48
Bernstein, R. B., Atom-Molecule Collision Theory, A Guide for the Experi-
mentalist, Plenum, New York, 1979.
N. J. BROWN 55
49
Miller, W. H., Dynamics of Molecular Collisions, Modern Theoretical Chem-
istry, Volumes 1 and 2, Plenum, New York, 1976.
50
Faist, M. B., Muckerman, J. T., and Schubert, F. E., Journal of Chemical
Physics, Vol. 69, 1978, p. 4087.
51
Muckerman, J. T. and Faist, M. B., Journal of Physical Chemistry, Vol. 83,
1979, p. 79.
52
Zellner, R. and Steinert, W., Chemical Physics Letters, Vol. 81, 1981, p. 568.
53
Brown, N. J. (to be published) 1991.
54
Walch, S. P. and Dunning, T. H., Journal of Chemical Physics, Vol. 72, 1980,
p. 1303.
55
Schatz, G. and Elgersma, H., Chemical Physics Letters, Vol. 73, 1980, p. 21.
56
Suzukawa, H. H., Jr., Wolfsberg, M., and Thompson, D. L., Journal of Chem-
ical Physics, Vol. 68, 1978, p. 455.
57
Miller, J. A. and Brown, N. J., Journal of Physical Chemistry, Vol. 86, 1982,
p. 772.
58
Brown, N. J. and Miller, J. A., Journal of Chemical Physics, Vol. 80, 1984,
p. 5568.
59
Rabitz, H., Frontiers in Applied Mathematics, edited by J. Buckmaster, SIAM,
Philadelphia, PA, 1985.
60
Judson, R. S., Rabitz, H., and Brown, N. J., Journal of Chemical Physics,
Vol. 86, 1987, p. 3886.
61
Judson, R. S. and Rabitz, H., "A Classical Functional Sensitivity Analysis of
Coplanar Inelastic Scattering for H2 + H2 and Its Isotopic Analogs," Journal of
Physical Chemistry, Vol. 93, 1989, p. 2400.
62
Judson, R. S. and Rabitz, H., Journal of Chemical Physics, Vol. 90, 1989,
p. 2283.
Bibliography
Bernstein, R. B., Chemical Dynamics via Molecular Beam and Laser Techniques,
Oxford University Press, New York, 1982.
Bowman, J. M. and Kuppermann, A., Chemical Physics Letters, Vol. 34, 1975,
p. 523.
Brown, N. J. and Rashed, O., Journal of Chemical Physics, Vol. 85, 1986,
p. 4348.
Bunker, D. L., Theory of Elementary Gas Reaction Rates, Pergamon, Oxford,
England, 1966.
Dunning, T. H., Jr., Harding, L. B., Bair, R. A., Eades, R. A., and Shepard,
R. L., Journal of Physical Chemistry, Vol. 90, 1986, p. 344.
Dunning, T. H., Jr., Harding, L. B., Wagner, A. F., Schatz, G. C., and Bowman,
J. M., Science, Vol. 240, 1988, p. 453.
Forst, W., Theory of Unimolecular Reactions, Academic, New York, 1973.
Geiger, L. C. and Schatz, G. C., Chemical Physics Letters, Vol. 114, 1985, p. 520.
Golden, D. M. and Larson, C. W., Twentieth Symposium (International) on
Combustion, 1984, p. 595.
Ho, P., Coltrin, M. E., Binkley, J. S., and Melius, C. P., Journal of the American
Chemical Society, Vol. 89, 1985, p. 4647.
Levine, R. D., Quantum Mechanics of Molecular Rate Processes, Clarendon
Press, Oxford, England, 1969.
Levine, R. D., Quantum Mechanics of Molecular Rate Processes, Oxford Uni-
versity Press, London, England, 1969.
Miller, J. A., Journal of Chemical Physics, Vol. 75, 1981, p. 5349.
Miller, W. H., Dynamics of Molecular Collisions, Pts. A and B, Plenum, New

York, 1976.
Miller, W. H., Accounts of Chemical Research, Vol. 9, 1976, p. 306.
Miller, W. H., Journal of Chemical Physics, Vol. 61, 1974, p. 1823.
Parker, G. A., Pack, R. T., Archer, B. J., and Walker, R. B., Chemical Physics
Letters, Vol. 137, 1987, p. 564.
Pechukas, P. and Pollak, E., Journal of Chemical Physics, Vol. 71, 1979, p. 2062.
Rabitz, H., Chemical Reviews, Vol. 87, 1987, p. 101.
Webster, F. and Light, J.C., Journal of Chemical Physics, Vol. 90, 1989, p. 265.
Webster, F. and Light, J. C, Journal of Chemical Physics, Vol. 90, 1989,
p. 300.
Whetten, R. L., Ezra, G. S., and Grant, E. R., Annual Review of Physical
Chemistry, Vol. 36, 1985, p. 277.
Chapter 3
Numerical Modeling of Combustion of Complex

Hydrocarbon Fuels
Charles K. Westbrook and William J. Pitz
I. Introduction
N UMERICAL modeling of chemical kinetics of hydrocarbon fuel oxi-

dation is currently undergoing a major revolution. Three intercon-
nected new themes can be identified. Significant changes are occurring in
the types of fuels that can be treated in numerical models; reaction mech-
anisms for many hydrocarbon fuels are being extended to regimes of tem-
perature, pressure, and other parameters that have not been studied
numerically in the past; and kinetic models are being used to examine
practical problems of much greater complexity than those addressed pre-
viously. All three of these trends have been accompanied, and in some
sense made possible, by a rapid growth in the size and speed of available
computers on which to integrate the kinetic rate equations and the devel-
opment of improved algorithms for integrating the equations. There is
every reason to believe that this rate of increase in computer power and
algorithm capability will continue.
There have been several recent reviews of detailed chemical kinetic
reaction rates and mechanisms for hydrocarbon fuels containing as many
as three carbon atoms (C3), such as propane. 1 2 However, extensions to
much larger fuels, and the numerical models for oxidation of these larger
fuels, create new computational problems that must be addressed. This
discussion will deal with extensions of reaction mechanisms to C4 and other
fuels as large as octane (C8H18) and to their treatment in numerical models.
In addition, for hydrocarbon fuels from methane to octane, the majority
of modeling studies in the past have considered problems such as flame
propagation, turbulent flow reactor oxidation, and shock-tube ignition,
which are applications that require only intermediate- and high-tempera-
ture reaction mechanisms. However, an entire new class of applications
Copyright © 1991 by the American Institute of Aeronautics and Astronautics. No copyright

is asserted in the United States under Title 17, U.S. Code. The U.S. Government has a
royalty-free license to exercise all rights under the copyright claimed herein for Governmental
purposes. All other rights are reserved by the copyright owner.
57
58 NUMERICAL MODELING OF COMPLEX HYDROCARBON FUELS
for hydrocarbon oxidation models requires an understanding of reaction

processes at much lower temperatures than those in older models. Ac-
cordingly, this chapter will describe the application of numerical modeling
techniques to extensions of these reaction mechanisms to lower tempera-
tures. Finally, the combination of reaction mechanisms for much larger
hydrocarbon fuel molecules and the capability to study hydrocarbon oxi-
dation over very wide ranges of temperature and pressure have made it
possible to address some new practical combustion problems for the first
time.
One such application area—the problem of engine knock—will be used
to illustrate the new capabilities made possible by recent developments in
the field of kinetic modeling. Computer models of chemical kinetics of
autoignition have been used in recent years to help understand the chemical
processes that take place in the combustion chamber of an internal com-
bustion engine. In particular, autoignition plays a large role in the problem
of engine knock. The degree of kinetic sophistication can vary widely from
one model to another, depending on the way in which that model is in-
tended to be integrated into the overall description of the engine com-
bustion problem.
In the past, the most fruitful type of kinetics modeling for autoignition
and engine knock was developed at the Shell Thornton Research Centre3^5
using a generalized reaction mechanism. The major kinetic processes in
the Shell model are all based on actual reactions observed in the lower
temperature oxidation of hydrocarbon fuels. Most important, the principal
step in the Shell model involves the addition of molecular oxygen to the
alkyl radical. This is followed by degenerate chain branching, leading even-
tually to a thermal explosion. Other important steps include internal H
atom abstraction (alkylperoxy isomerization), and formation of dihydro-
peroxy radicals. The Shell model and others based on it (see, for example,
Refs. 6 and 7) have been very successful in many cases at helping to describe
the dependence of engine knock on engine operating parameters.
A parallel effort over the past several years8"12 has emphasized a more
detailed chemical kinetic modeling approach than the Shell model. Al-
though more costly in terms of development time and requiring much more
powerful computers, these detailed models promise an improved ability to
treat such effects as fuel molecular structure, mixtures of fuels, and the
use of fuel additives. In the Shell model, the global or representative
reactions have rates that must be adjusted for each fuel to be analyzed, a
process that requires extensive experimental testing of each fuel. Further-
more, such models are generally suspect in any application outside their
specific regimes of calibration. In contrast, a detailed kinetic model uses
reaction rate data that are more fundamentally based than those in a global
model and, thus, this model is often more suited to new predictions. How-
ever, these models are generally orders of magnitude larger and more
complex than global models. Current research being carried out toward a
detailed modeling goal is described in this paper.
We will also show how this model evolution has been accompanied or,
in fact, driven by ever-increasing computational resources, including hard-
ware (both CPU performance and storage capacity) and development of
C. K. WESTBROOK AND W. J. PITZ 59
numerical solution techniques. The growth in the sophistication and com-

plexity of kinetic reaction mechanisms has followed very closely the growth
in available computer power and other computational resources.
II. Autoignition and Engine Knock

A typical engine cycle begins when unburned fuel and air are introduced
into the combustion chamber. This charge is compressed by the piston and
then spark ignited at a time usually close to top dead center (TDC), the
point of maximum compression when the piston reaches the highest lo-
cation in the combustion chamber. A flame then begins to propagate across
the chamber, steadily converting fuel and air to combustion products. The
gases that have not yet been consumed (the "end gases") are not chemically
inert during this time interval; they are reacting steadily but at rather low
rates. Under normal operation, the end gases are eventually consumed by
the flame propagating through the engine chamber. However, under ex-
treme conditions of pressure and temperature, it is possible for these end
gases to autoignite spontaneously prior to the arrival of the flame. When
autoignition produces a sufficiently rapid pressure rise and involves a sig-
nificant fraction of the end gas, engine knock results. Numerical simulations
of this autoignition process place great demands on the chemical kinetic
description of the hydrocarbon fuel oxidation mechanism.
Temperature History
-During the combustion of the fuel-air mixture in a typical spark ignition
engine, the unburned portion of the charge is subjected to a considerable
increase in pressure and temperature as a result of compression, both by
the engine piston and by the expansion of the burned product gases. Three
temperature regimes can be defined for the temperature history of the end
gas. The first kinetic regime encountered is at low temperature. When the
intake valve opens, the unburned mixture is at approximately atmospheric
temperature and pressure. As the end gas is compressed by the piston, it
spends a large fraction of its time in the temperature range of 450-800 K;
at low engine speeds, the end gases may be in this range for as long as 70
ms. 9 ~ n Although the overall rate of combustion is slow in this temperature
range, the residence time is long enough for some fuel consumption and
heat release to occur.
As the temperature of the end gas increases above about 800-850 K,
the dominant kinetic paths change markedly. At the lower temperatures,
the overall reaction takes place primarily through the formation of various
alkylperoxy radicals via addition of molecular oxygen to alkyl radicals,
followed by production of alkyl hydroperoxides and subsequent reactions.
At higher temperatures, however, the alkylperoxy radicals become un-
stable, creating a region of negative temperature coefficient (NTC). The
NTC region is characterized by a curious and unexpected feature, whereby
an increase in the gas temperature produces a reduced rate of fuel con-
sumption. This phenomenon is responsible for a large variety of nonlinear
responses in combustion systems, including the existence of so-called "cool
flames." Eventually, the NTC region disappears as the temperature is
increased further, due to the emergence of new and rapid reaction se-
quences that begin to appear. At this point, the combustion becomes con-
trolled by the reactions of the hydroperoxyl radical and hydrogen peroxide.
As the temperature increases further to values above about 1000-1100
K, the dominant reaction path changes again, now being controlled by the
production of H atoms, which react with molecular oxygen to produce both
oxygen atoms and OH radicals. The problem of modeling the chemical
kinetics of end-gas autoignition is considerably more difficult than modeling
the kinetics of most other types of experiments because the end gases
spend significant periods of time in all of these temperature regimes. The
kinetic model must be able to describe all of the important reaction paths
and rates in each regime, and the numerical integration must be efficient
enough to treat the relatively long times involved.
The history of the end-gas temperature as a function of time is illustrated
by the solid curve in Fig. 1; we will return later to a discussion of the other
curves. In this particular example, the experimental data are taken from
an engine operating at 600 rpm with a compression ratio of 6.63.13 A time
of zero corresponds to bottom dead center (BDC), the point at which the
piston is at its lowest point and the combustion-chamber volume has its
greatest value. The initial temperature is close to room temperature, and
spark ignition occurs at TDC, 50 ms later. From this figure, it is clear that
1200
1100-
0> 1000-
900-
0)
800-
700-
600
0.035 0.040 0.045 0.050 0.055 0.060
Time (sec)
Fig. 1 End-gas temperature in an engine operating at 600 rpm. Compression ratio
set for 90-octane fuel conditions; solid line represents experimental values, and other
curves show computed results for various mixtures of iso-octane and n-heptane.
the end-gas temperature lies between 300 and 600 K for nearly 40 ms, and
it exceeds 750 K only after a total elapsed time of 50 ms. The decreasing
value of temperature after 53 ms assumes that the end gas is not consumed
by the flame but is cooled by the expansion stroke of the piston. Although
reaction rates are small at temperatures below 750 K, there is a considerable
amount of time available for reactions to affect the end-gas composition
and temperature. In contrast, the end-gas residence time above 750 K is
relatively short, although the reaction rates for the higher temperature are
much faster.
Fuel Structure
Another significant difficulty in modeling the autoignition of typical hy-
drocarbon fuels under knocking conditions results from the composition
of the fuel itself. Conventional automotive fuel is usually a mixture of
different constituents, including saturated and unsaturated straight- and
branched-chain species as well as aromatic and other types of molecular
structures. Very few of these species have been described by detailed
chemical kinetic reaction mechanisms in any of the temperature and pres-
sure regimes that are important for knock.
Figure 2 shows two isomeric forms of octane, the straight-chain mole-
cule n-octane and the branched-chain form, iso-octane, also referred to as
2,2,4-trimethylpentane. Both fuels have the same numbers of C and H
atoms, both have very nearly the same heat of reaction, and the burning
velocities of flames through mixtures of both fuels with air are very similar.
However, the autoignition properties of the two fuels are dramatically
different.
n-octane
H H H H H H H H
I I I I I I I
H—C™C—C—C—C—C—C—C—H
! I I ! I I I
H H H H H H H H
iso-octane
H H
HCH HCH
H | H | H
C — C~~C^C — C
H | H H H
HCH
H
Fig. 2 Schematic diagrams for n-octane and iso-octane.

The octane number of a given fuel is a measure of its autoignition rate

and tendency to knock under engine conditions, with a large number in-
dicating good resistance to knock and a small number indicating a tendency
to knock. Octane ratings in this study are taken from several sources.1445
Iso-octane is very knock resistant, with an octane number equal to 100.
Iso-octane is used to define this point of 100 on the octane scale. In contrast,
n-octane has a very high knock tendency, with an octane number of less
than zero. The similar fuel, n-heptane, has, by definition, an octane number
of exactly zero, and n-octane is slightly more prone to knock than is n-
heptane. Therefore, the vast difference in knocking tendency between the
two isomers, n-octane and iso-octane, can be attributed only to the dif-
ferences in the structure of the two molecules.
Other examples are given in Figs. 3 and 4, in which five different isomeric
forms of hexane and three forms of pentane are shown schematically.
Again, the flame temperatures and burning velocities are all very similar.
The octane numbers for the compounds are shown in Table 1, together
with the octane numbers of other related fuel molecules. It is clear that
structural factors can significantly influence the autoignition properties for
these fuels.
A factor that can have a comparable influence on autoignition and knock
is the absolute size of the fuel molecule. For straight-chain hydrocarbons
from n-butane to n-octane, the octane numbers are 92, 62, 25, 0, and less
than zero, respectively, as illustrated in Fig. 5.
Comprehensive Reaction Mechanisms

A chemical kinetic reaction mechanism is a part of a numerical modeling
attempt to describe the actual reaction history of the reacting gas sample.
When these reacting gases encounter a wide range of operating conditions,
a great demand is made on the reaction mechanism. Some time ago, West-
Table 1 Octane numbers for sample fuels

Fuel Octane no.
QHi4 n-hexane 25
QH14 2-methyl pentane 73
QFL14 3-methyl pentane 74
QHl4 2-2 dimethyl butane 92
QHl4 2-3 dimethyl butane 99
QH12 cyclohexane 80
C5H12 n-pentane 62
C5H12 2-2 dimethyl propane 90
QF[12 2-methyl butane 92
C5H10 cyclo-pentane 93
C4H10 n-butane 92
C4H10 iso-butane 100
C8H18 n-octane <0
QH18 iso-octane 100
QHl6 n-heptane 0
1) n-hexane Octane no. = 25
H H H H H H
H—C—C—C—C—C—C—H
H H H H H H
2) 2-methyl pentane Octane no. = 73
H
H CH
H | H H H
H— C— C—C—C—C—H
H H H H H
3) 3-methyl pentane Octane no. = 74
H
H CH
H H | H H
H—C—C—C—C—C—H
H H H H H
4) 2-2-dimethyl butane Octane no. = 92
H
H CH
H | H H
H—C—C—C—C—H
H | H H
H CH
H
5) 2-3-dimethyl butane Octane no. = 99
H
H CH
H | H H
H^C—C—C~C—H
H H | H
H CH
H
Fig. 3 Schematic diagrams for hexane isomers.
6) ii-peiitaiie Octane no. = 62
H H H H H
H—C—C—C—C—C—H
H H H H H
7) 2-methyl butane Octane no. = 92
H
H CH
H H H
H—C—C—C—C—H
H H H H
8) 2-2-dimethyl propane Octane no. — 90
H
HCH
H H
H—C—C—C—H
H | H
HCH
H
Fig. 4 Schematic diagrams for pentane isomers.
brook and Dryer outlined the properties of a "comprehensive" reaction

mechanism,1-16 designed to be valid over such ranges of operating condi-
tions. The reaction mechanism should be tested over the entire range of
conditions of interest, using experimental data from a variety of separate
experiments under disparate conditions. Therefore, mechanistic simplifi-
cations that are valid only within a limited range of conditions cannot be
made. Comprehensive mechanisms have been developed for a considerable
number of hydrocarbon fuels to date.17"21
Modeling combustion in an internal combustion engine—in particular,
the process of autoignition that leads to engine knock—requires a com-
prehensive reaction mechanism that is even more general than any devel-
oped to date. All such mechanisms previously considered combustion
problems at temperatures above about 900 K, using experimental data
from such environments as shock tubes, laminar flames, turbulent flow
reactors, and high-temperature stirred reactors. Reactions and products
important at lower temperatures but with no importance at 900 K and
above were not included in these previous models. However, oxidation
processes at temperatures as low as 500 K can play an important role in
the autoignition and engine knock problem and, thus, cannot be neglected.
In fact, the dominant reaction steps between 500 and 800 K are distinctly
n-butane Octane no. = 92
H H H H
I I I I
H—C—C^C—C—H
I I I I
H H H H
n-peiitane Octane no. —62
H H H H H
I I I I i
H—C—C—C^C—C~H
I I I i I
H H H H H
n-hexane Octane no. = 25
H H H H H H
i I I I I I
H-^C^C — C —C — C —C — H
I I I I I I
H H H H H H
n-heptane Octane no. = 0
H H H H H H H
I i I I I I I
H— C — C-^C^C — C — C^C — H
I ! I I I -I I
H H H H H H H
n-octane Octane no. < 0
H H H H H H H H
I I I I I I i I
H — C —C — C—C~C^-C — C — C— H
\ \ I I I I I I
H H H H H H H H
Fig. 5 Schematic diagrams for C4-C8 n-alkanes.
different from those that are important at higher temperatures, requiring

entirely different elementary reactions and chemical species. Therefore,
reaction mechanisms developed for these problems must be valid for tem-
peratures from 500 to 1500 K and more.
Finally, the fuel molecules to be considered in a study of engine knock
are considerably larger and structurally more complex than those examined
in earlier comprehensive reaction mechanisms. In particular, reaction
mechanisms for the ignition of fuels such as n-heptane and iso-octane, the
primary reference fuels for knock rating of hydrocarbon fuels in automotive
engines, are extremely desirable. The development of mechanisms for these
larger fuels is described subsequently.
The principal difficulty in modeling the combustion of large fuel mole-
cules in computer simulations is that the models require large amounts of
computer time, memory, and extremely sophisticated numerical algo-
rithms. Furthermore, the comprehensive models required for autoignition
are even larger and more expensive because no restricting assumptions can
be made. Many more species and reactions need to be included in the
model, driving the computational costs even higher. In many ways, our
ability to compute the evolution of complex hydrocarbon fuel systems has
paralleled the explosive growth in computer resources over the past 10
years. This parallel is expected to continue in the future as we continue to
address progressively more realistic combustion problems.
III. High- and Intermediate-Temperature Ignition
Chain Branching Mechanisms

At temperatures in excess of about 1000-1100 K, the ignition of typical
hydrocarbon fuels is conceptually quite simple. Following a brief initiation
period during which a free radical pool is established, the fuel molecule is
systematically broken down into smaller fragments. The principal types of
reactions responsible for this fuel consumption process are H atom ab-
stractions from the fuel and other intermediate species, and thermal de-
composition reactions of the larger hydrocarbon radicals that are produced.
Once the larger fuel and intermediate hydrocarbon species have been con-
verted to CO and H2, these intermediates are oxidized rapidly to CO2 and
H2O. This releases a large amount of thermal energy, raises the temper-
ature and pressure of the reacting gas mixture, and effectively completes
the overall reaction.
In this high-temperature regime, the chain branching features of the
combustion reaction mechanism are dominated by the single reaction
H + O2 - O 4- OH
which consumes one H atom and produces two radical species, which then
accelerate the overall fuel consumption process. This reaction has a fairly
high activation energy of about 17 kcal/mole,22 so it requires a high tem-
perature to proceed rapidly. A second reaction,
H + O2 + M = HO2 4- M
can, under some conditions, compete effectively with the chain branching
reaction, thereby reducing the overall rate of reaction. This recombination
reaction is pressure dependent, and its rate has very little temperature
dependence. Therefore, as the reacting gas temperature decreases to about
1000 K, or as the pressure is increased significantly, this second reaction
competes for the available H atoms and tends to reduce the overall rate
of oxidation.
At temperatures in excess of 1000-1100 K, where the H + O2 = O + OH
chain branching reaction is the dominant reaction, the overall rate of re-
action and ignition is very fast compared with the typical residence times
in the engine chamber. That is, once the end-gas temperatures have reached
this level, ignition is essentially instantaneous. Earlier modeling studies8 23
showed that the autoignition process occurs on a time scale shorter than
a millisecond when the gas temperature is greater than 1100 K.
At end-gas temperatures between about 800 and 1100 K, the reaction
H + O2 + M = HO2 + M
is the dominant reaction for H atoms. Its relative importance is increased

by the high pressures (greater than 20 atm) encountered in the combustion
chamber of a typical automotive engine. At these intermediate tempera-
tures, another cycle of reactions becomes responsible for the majority of
chain branching observed in the engine chamber. This sequence of chain
branching reactions is dominated by
fuel + HO2 - alkyl + H2O2
alkyl + O2 = olefin + HO2
H2O2 + M - OH + OH + M
This sequence converts one HO2 radical to two OH radicals.8 The HO2
radical consumed in the first step is also regenerated, making this a very
efficient chain branching sequence. These reactions require a sufficiently
high temperature to dissociate the H2O2 molecule and are, therefore, neg-
ligibly slow below 750-800 K. However, as the temperature increases
above about 1000 K, two factors combine to alter the major reaction
sequence. First, at higher temperatures, the competing reaction
H + O2 = O + OH
becomes important. In addition, instead of reacting with O 2 , alkyl and
other hydrocarbon radicals are consumed primarily by thermal decom-
position, yielding H atoms, as in the following examples:
C2H5 + M = C2H4 + H + M
iC3H7 - C3H6 + H
HCO 4 - A f - H + CO + M
tC4H9 - H + iC4H8
Therefore, there is a fairly narrow range of temperatures over which the

temperature is high enough to dissociate H2O2 and still low enough that
the H + O2 = O + OH reaction and the alkyl radical decomposition
reactions are unimportant.
Under most practical engine conditions, the temperature at which the
end gases ignite lies between 850 and 1100 K, almost exactly the range
over which the HO2-H2O2 reactions dominate. Previous work8 has shown
that antiknock additives such as tetraethyl lead (TEL) act within this same
temperature range by removing HO2 and H2O2 from the chain branching
reaction sequence through catalytic heterogeneous reactions on the sur-
faces of solid lead oxide particles.
Detailed Reaction Paths

At temperatures above about 1000 K, combustion occurs through a
sequence of reactions that begin with the thermal decomposition of the
fuel molecule. This decomposition is different for each type of hydrocarbon
fuel; some of the dominant decomposition reactions and their reaction
products are summarized in Table 2.
After a brief initiation period, a radical pool is established and the process
of fuel consumption takes place primarily through abstraction of H atoms
from the fuel molecule. The alkyl radicals that are produced then react by
using a variety of processes, including thermal decomposition and reactions
with oxygen molecules. The complexity of these processes can be illustrated
by using the diagrams for two isomers of octane, as shown in Fig. 2. The
first isomer is the straight-chain fuel n-octane, and the second form is the
multiply branched iso-octane, also denoted as 2,2,4-trimethylpentane. For
n-octane, four different, logically distinct H atoms can be abstracted from
the fuel molecule, depending on how far from the end of the chain the H
atom is located. There are also four logically distinct H atom sites in the
iso-octane molecule. The alkyl radicals produced by H atom abstraction
from those sites subsequently decompose into different sets of products.24
Each distinct set of products has a different effect on the subsequent re-
action history for that fuel, so it is essential to keep these paths separate
in a computational model.
Table 2 Thermal decomposition reactions

of selected alkane fuels
nC4H10 = C,H5 + C,H,

iC4H10 = nC3H7 f CH3
nC8H18 - pC4H9 f pC4H9
iC8H18 - tC4H9 -f- iC4H9
nC6H14 = nC3H7 f nC3H7
2-methyl pentane = nC3H7 f iC3H7
3-methyl pentane = C2H5 + sC4H9
2-2 dimethyl butane = tC4H9 4- C2HS
2-3 dimethyl butane = iC3H7 4 iC3H7
This variability can be shown for the four logically distinct alkyl radicals
produced from n-octane. If we denote the alkyl radical produced by the
abstraction of one of the six H atoms bonded to the C atoms at either end
of the chain as 1-C8H17, the alkyl radical produced by abstraction of one
of the four H atoms bonded to the C atoms once removed from the end
of the chain as 2-C8H17, and similarly for the other alkyl radicals, we can
summarize the major decomposition products for each of the alkyl radicals
as follows:
1-C8H17 - C2H4 + C2H4 + C2H4 + C2H4 + H
2-C8H17 - C3H6 + C2H4 + C2H4 + CH3
3-C^8Jrii7 — jC^4H8 ~h O2H4 ~h C^2ri4 H- H
4-C8H17 = iC5H10 + C2H4 + CH3

As we have already seen, production of H atoms has a much different
influence on the rate of reaction than the production of methyl and other
radical species. In addition, the various olefin species produced are dif-
ferent for each path. The same type of behavior is seen for the alkyl radicals
in iso-octane:
a-C8H17 - iC4H8 + C3H6 + CH3
b-C8H17 = C7H14 + CH3
c-C8H17 - iC4H8 + iC4H8 + H
d-C8H17 = iC4H8 + C3H6 + CH3
where the four logically distinct alkyl radicals in iso-octane have been
indicated by the prefixes a, b, c, and d.
The rates of abstraction of H atoms are determined primarily by the
nature of the C-H bonds that must be broken. Three types of C-H bonds
are encountered in saturated hydrocarbons, such as those shown in Figs.
2-5. Primary bonds are the strongest, secondary bonds are next in strength,
and tertiary bonds are weakest. These factors affect the ease of abstraction
of the H atom from the parent fuel. Furthermore, each radical species
abstracts H atoms at different rates. Distinctions among the varying rates
of H atom abstraction by radical attack, depending on the type of radical
and the type of C-H bond, are easily incorporated into numerical models,
as are the separate accounting of the logically distinct alkyl radicals and
their different paths for decomposition. However, this degree of complexity
significantly increases the magnitude of the computational problem, adding
to the computer resources required to set up and then integrate the kinetic
rate equations.
This type of modeling analysis has been applied to ignition of large fuels
behind shock waves25"27 and for high-temperature well-stirred reactors,28
where the high-temperature reaction mechanism is of primary importance.
All of the initiation and fuel decomposition reactions previously described
occur at high rates and determine the overall reaction rate. The present
discussion has been greatly simplified, emphasizing only the most dominant
features of the high- and intermediate-temperature kinetics mechanisms.
Many additional features can play significant roles, and these additional
details are discussed in depth in a number of previous publications.16<29-30
This type of simulation is very costly to perform in terms of computer
time requirements and other resource needs. If the time evolution of each
variety of radical is to be followed independently, and the subsequent
reactions of each of these species are then traced, the number of species
in the reaction mechanism can become very large. Reaction mechanisms
for the high-temperature oxidation of simple fuels such as methane or
methanol involve 20-30 chemical species, but reaction mechanisms for
octanes require more than 100 different species, even for the qualitatively
simple high-temperature mechanisms. For large fuels such as hexane or
octane, reaction mechanisms that cover both the low- and high-temperature
regimes require more than 200 species. In a computer model of this type,
the costs are related to the number of chemical species to be evolved. Each
species is described by a time-dependent differential equation, transformed
into a difference equation for numerical solution, and the total amount of
CPU and storage costs typically scale roughly with A/ 2 , where N is the
number of equations (species) to be solved. Only with current and future
computers and algorithms can the very large systems of equations be solved
efficiently to deal with these practical fuels.
In many cases, the rate of high-temperature ignition correlates well with
the octane number. For example, the octane number of iso-octane, which
we have seen ignites relatively slowly at high temperatures, has the very
high octane number of 100, while the octane number for n-octane is very
low, actually less than zero. There are other examples of this type of overall
correlation between the rate of high temperature and the octane number.
However, there are also a large number of cases in which this agreement
breaks down. As an example of this type of problem, we can consider the
cases of neopentane and 2,2-dimethylbutane, whose structures are sum-
marized in Figs. 3 and 4. If we follow the product distributions of the alkyl
radical decomposition reactions, it becomes clear that all 12 H atom ab-
straction possibilities in neopentane lead to methyl radicals and isobutane.
Therefore, the high-temperature ignition of this fuel should be very slow.
Furthermore, the knock resistance and octane number of neopentane are
very high, with an octane rating of 90. The second fuel shown in Fig. 3 is
2,2-dimethylbutane, which is really equivalent to neopentane with a methyl
group replaced by an ethyl radical. Of its 14 H atoms, abstraction of 12
of them leads to H atom production and high-temperature chain branching.
As a result, the high-temperature ignition of this fuel would be expected to
be very rapid. However, the octane number of this fuel is very high, with
a value of 92, even higher than the value of 90 observed for neopentane.
There are many other examples of this lack of correlation between the
rate of H atom production and high-temperature ignition rate with the
octane number. The immediate conclusion is that the octane number and
knock susceptibility are not determined by the high-temperature reaction
mechanisms alone. Other chemical reactions must control the rate of com-
bustion at lower temperatures.
IV. Extensions to Low Temperature

Below temperatures of about 800 K, alkyl radicals produced by H atom
abstraction reactions from the fuel molecule thermally decompose at a
relatively slow rate. Instead, these radicals react with O2 to produce al-
kylperoxy radicals (RO2). These RO2 radicals are involved in reaction paths
that lead to significant chain branching at low temperatures. These chain
branching paths must be taken into account in the chemical kinetic mech-
anism, making the mechanism considerably more complex. Alkylperoxy
radicals react in a variety of ways:
R + O2 = RO| - olefin + HO2
- RO2
- R + O2
The simplest result of such reactions, other than redissociation back to
R + O2, is the production of an olefin species and an HO2 radical.31*32
Alternatively, the alkylperoxy radical, RO 2 , can then react with the fuel
or undergo an internal H atom abstraction (isomerization), both of which
lead to the formation of hydroperoxides, ROOH and QOOH,
RO2 + R'H = R' + ROOH
or
RO2 = QOOH
where R'H is the fuel, and the species Q is the radical species R minus
one H atom. Both hydroperoxides can decompose to form OH radicals,
ROOH - RO + OH
QOOH - epoxide + OH
with ROOH producing two radicals and QOOH producing only one rad-
ical. The hydroperoxide radical QOOH can undergo further reaction with
molecular oxygen and eventually yield two OH radicals and the radical
Q'O,
QOOH + 02 - 02QOOH
02QOOH = aldehyde + Q'O + OH + OH
Therefore, if the alkyl radical R eventually produces the alkyl hydrope-

roxide ROOH, three radicals—R', RO, and OH—are produced and one
R radical is consumed. The amount of radical multiplication for this re-
action sequence is a factor of 3. If the R radical follows the path of the
hydroperoxide radical QOOH, either one OH radical is produced or three
radicals (Q'O, OH, and OH) are produced. Both of these reaction paths
yield significant radical multiplication at low temperatures that must be
included in the chemical kinetic mechanism.
For practical hydrocarbon fuel molecules of the size range being dis-
cussed here, such as hexanes, heptanes, and octanes, there are a bewil-
dering array of possible isomerization processes. The rates of these RO2
reactions depend on the type of C-H bond broken in the internal H atom
abstraction step, on the distance within the molecule between the last O
atom and the H atom being abstracted, and on the number of different H
atoms available for abstraction. Again, all of these factors can be param-
eterized and included in a detailed kinetic model, but the very large number
of species and reactions needed to deal with this degree of detail make it
computationally demanding. For example, the low-temperature reaction
mechanism for n-heptane adds 61 new species, including 10 epoxide spe-
cies, whereas the same additions for iso-octane include 48 species with 8
epoxides. Thus, the extensions to low temperature approximately double
the size of the reaction mechanism from about 100 species (differential
equations) to more than 200 species (equations), multiplying the computer
time requirements by approximately a factor of 4 and greatly increasing
the storage and software requirements. The justification for increasing the
costs by a factor this large is that the predictive powers of the mechanism
make it possible to interpret large quantities of experimental data that were
not possible to understand without such a mechanism.
V. Examples of Model Tests in Different Environments

Before a kinetic model can be applied to a practical problem such as
engine knock, first it must be validated through a series of numerical tests
against experimental data obtained under somewhat more idealized con-
ditions. In most cases, any of these experiments will cover a rather limited
range of operating conditions. However, that experiment generally will
provide a rather demanding test of the reaction mechanism under those
limited conditions. In order to "pass" the test, the reaction mechanism
must accurately reproduce the observed experimental results under each
of the sets of available conditions. Only when the reaction mechanism has
passed such tests, under an entire range of conditions that spans those of
concern in the practical problem, can the model be expected to provide
realistic and reliable predictions for the practical, applied problem, such
as the end-gas autoignition in an engine chamber.
Shock Tubes
Shock-tube experiments are an excellent test of the high-temperature
reaction mechanism for hydrocarbon oxidation. 1 16 The nearly instanta-
neous attainment of high temperature (i.e., >1200 K) and fairly high pres-
sure (i.e., > 3 atm) means that the results are independent of the low-
and intermediate-temperature-regime reaction subsets. The most common
observable quantity in such experiments is the ignition delay time, generally
defined as the time at which noticeable heat release or pressure rise is
observed in the reacting gases.
In a series of shock-tube experiments, Burcat et al.33 used mixtures of
n-alkanes, oxygen, and argon in proportions that were intended to simulate
fuel-air conditions. The ratio of argon to oxygen was 5:1 rather than the
high value that is typical of many very dilute shock-tube experiments, so

the total heat release and temperature increase resulting from the ignition
will be substantial. In the experimental paper, the n-alkanes from Cl to
C5 were considered; methane was found to be anomalously slow to ignite,
whereas all of the other n-alkanes were very similar in their rate of ignition,
with ethane being slightly faster to ignite than the C3-C5 fuels. Computer
modeling results were compared with the experimental results,27 with the
overall performance of the numerical model being quite satisfactory. These
results are summarized in Fig. 6, with the results for ethane not shown,
but with computed results for isobutane included. Following the chain
branching arguments presented earlier, ethane reacts most rapidly because
virtually all H abstraction reactions lead to ethyl radicals and H atoms,
which accelerate the rate of ignition. Propane, n-butane, and n-pentane
produce a mixture of methyl radicals and H atoms, so their ignition delay
periods are very similar. A later study examined the shock-tube ignition
of mixtures of n-octane, n-heptane, and iso-octane with oxygen and argon,
under conditions equivalent to those studied previously for the smaller
fuels.26'27*34
In the high-temperature regime, at pressures less than 10 atm (1 MPa),
kinetic analysis in terms of H atoms, methyl radicals, intermediate olefins,
and their relative rates of production provides a rather simple, straight-
forward, uncomplicated picture of hydrocarbon fuel ignition. The straight-
chain fuels produce relatively large amounts of H atoms through alkyl
Temperature (K)
1900 1700 1500 1300 1150
nC5H12
<D .
E 10 CH'4
\A
A/
"•
A/
, n C4H1 0
<D
.2 iff 8
O)
C~H
3 8
10'
0.5 0.6 0.7 0.8 0.9
3
10 /T
10 /T (K)
(K)
Fig. 6 Experimental shock-tube ignition delay time measurements (indicated by
symbols) and predicted model results for methane, propane, n-butane, and n-pen-
tane. Also shown are computed predictions for isobutane.
fragmentation into ethyl radicals. These H atoms provide large rates of

chain branching and rapid rates of fuel ignition. In strong contrast, the
branched-chain iso-octane results in few H atoms and a low rate of chain
branching, with a correspondingly slower rate of ignition.
Flow Reactors
The turbulent flow reactor provides an experimental environment that
is complementary to that of the shock tube for carrying out development
of detailed chemical kinetic models. The temperatures are considerably
lower than those of the shock tube and fall in exactly the same range (900-
1200 K) in which the reaction mechanism undergoes the transition from
HO2-dominated chain branching to that dominated by the H + O2 =
O + OH reaction. In contrast to the shock tube, initiation reactions play
virtually no part in the flow reactor, but the period of fuel consumption
and production of stable intermediates is easily accessible.
The most prominent experimental program in recent years devoted to
exploring hydrocarbon kinetics in the flow reactor has been that at Prince-
ton University. The experimental facility is described in Ref. 35, and de-
tailed results from this facility have been used in a large number of studies
in recent years17"20-2536 to develop the understanding of intermediate- and
high-temperature kinetic processes that control hydrocarbon combustion.
The turbulent flow reactor can study temperatures as low as approxi-
mately 900 K. As the temperature is reduced further, the overall rate of
fuel conversion becomes quite small, and the extent of reaction within the
test section of the reactor becomes too low to be able to detect the changes
in composition. It is then necessary to switch to an experimental config-
uration that is better suited to following the much longer time scales of
the hydrocarbon oxidation process.
Low-Temperature Reactors
Three types of experimental facilities are generally most useful to study
oxidation at temperatures between about 500 and 900 K. From a numerical
modeling point of view, these three types of environments are interesting
because they involve three different sets of boundary conditions. These
are the static reactor, the continuously stirred tank reactor (CSTR), and
the rapid compression machine. Historically, the static reactor has been
most prominent for many years, and perhaps thousands of experimental
studies in static reactors have been carried out to provide a wealth of data
and insight into oxidation kinetics. Computationally, the appropriate con-
ditions are constant volume, together with a submodel for heat transfer
between the reacting gases and the problem boundaries. From the point
of view of kinetic mechanism development, these experiments have been
essential in terms of qualitative information; from a quantitative aspect,
modeling these experiments is often difficult because of the influence of
the walls, both due to their kinetic activity and to the fact that heat transfer
through the walls to the outside is difficult to quantify. Recent contributions
to the kinetic mechanism development described herein have been pro-
vided by static reactor experiments carried out by Kaiser et al.37 and Wilk
et al.38'39
The CSTR is a more recent development, described in detail in Refs.

40 and 41. This system is a flowing one in which a reaction chamber has
fresh reactants flowing into it, reaction is taking place in the reactor, and
reaction products are flowing out of the chamber. The contents of the
reactor are well stirred, so that spatial homogeneity is maintained in the
reactor. As a result, the system approaches the ideal conditions where
there is a single reaction zone with steady flow in and out of that zone.
Such a system is quite simple to simulate numerically, involving one bound-
ary through which a prescribed mass flux passes at a given temperature,
together with an outflow boundary beyond which the pressure is assumed
fixed and at which temperature and mole fraction gradients are assumed
to be zero. An important advantage of both the model and the experiment
is that more complex classes of behavior can be studied than those acces-
sible in the static reactor (see, for example, Ref. 42), including periodic
cool flames, periodic ignition, and a wealth of other periodic behavior.
The rapid compression machine has been valuable for reaction kinetics
development in several ways. A vast number of these experiments were
used to develop and test the Shell model,3"7 discussed earlier. These ex-
periments have been used extensively by Griffiths et al. (see, for example,
Refs. 43 and 44) to study hydrocarbon fuel oxidation, especially in the
negative temperature coefficient region, kinetic information which is es-
sential in developing kinetic models for the lower temperature regime.
Another advantage of the rapid compression machine data is that, in ad-
dition to the temperature range of the experiments, the pressures attained
are significantly higher than those commonly found in static reactors and
the CSTR and, therefore, are much closer to those encountered in knocking
engines. Numerically, the problems are characterized by a variable but
prescribed volume, together with heat transfer to the combustion-chamber
walls.
Knocking Engine
In a typical internal combustion engine cycle, the reactive end gases are
compressed and heated both by piston motion and by propagation of a
flame after spark ignition. This sequence of compressions is nearly adi-
abatic, except for heat transfer to the engine chamber walls. Experimental
and modeling investigations into actual engines allow one to address the
entire temperature history of the end gas from 400 K to burned gas tem-
peratures. Green and co-workers9"11 experimentally investigated engine
knock in the Sandia optical research engine, measuring autoignition times
of hydrocarbon-air mixtures, primarily with n-butane and isobutane as
fuels. In addition, gas sampling techniques were used to determine species
concentration histories in the end gas. Numerical analyses of these engine
data have provided considerable insight into end-gas chemistry leading to
engine knock.
In a second study, two different sets of measured pressure histories from
actual engines13'45 were used to develop and test detailed kinetic models.
The modeling analyses followed mixtures of n-heptane and iso-octane through
prescribed pressure histories, and the times at which autoignition was pre-
dieted were compared with experimental results. Heat losses to the engine
chamber walls were estimated by using a Newtonian heat loss term, with
the heat loss rate adjusted so that a reference mixture of 90% iso-octane
and 10% n-heptane ignited at a time of approximately 22-deg after TDC,
consistent with experimental observation. 13
A representative series of computational results is illustrated in Fig. 1,
showing the end-gas temperature histories for five different stoichiometric
fuel-air mixtures. Each end-gas sample was forced to follow the pressure
history in an engine operating at 600 rpm, with a compression ratio of 6.63,
corresponding to an octane number of 89.7. Spark ignition occurs at TDC,
at a time of 0.050 s, with time measured from BDC. The solid curve
discussed earlier shows the end-gas temperature, which results from a
computation in which only the influence of end-gas compression and heat
losses are considered, neglecting heat release from reactions. This curve
was established by requiring the model to reproduce the experimentally
measured pressure history in this engine. The other four curves show com-
puted results for mixtures with octane number of 0, 50, 90, and 100, as
indicated by the Fig. 1 legend. Recall that, for mixtures of iso-octane and
n-heptane, the percentage of iso-octane in the mixture is equal to the octane
number. Thus, a fuel mixture of 60% iso-octane and 40% n-heptane is,
by definition, a 60 octane number mixture. All of these reactive end-gas
cases produce heat during the period of 0.048-0.052 s; only one curve is
shown in Fig. 1, but this one dot-dash curve applies equally to all four
reactive fuel mixtures. These computed results show clearly that the time
of autoignition, indicated by the rapid growth in end-gas temperature, is
a monotonic function of octane number; reducing the octane number re-
sults in an earlier autoignition of the end gases while increasing the octane
number leads to later autoignition.
Careful analysis of the modeling results provides some insight into the
specific kinetic processes taking place. Most significant, the production
rate of OH radicals decreases with increasing octane number. At a time
of approximately 0.054 s, the OH concentration in the RON50 case is twice
that in the RON100 model. The fastest reactions consuming the fuel are
its reactions with OH, so this difference in OH level produces a corre-
spondingly large difference in the rate of fuel consumption. Autoignition
does not occur until the fuel has been consumed, so enhanced OH pro-
duction leads directly to advanced autoignition. Also important, the same
reactions with OH produce H2O with a significant release of chemical
energy, raising the temperature of the end gases and accelerating the overall
rate of autoignition. The OH radicals are themselves the result of decom-
position of H2O2, produced from reactions of HO2 with the fuel and from
HO2 + HO2 - H2O2 + O2
The largest sensitivity coefficients for ignition of n-heptane and iso-octane

at 800 K are for conversion of HO2 radicals to H2O2 and the decomposition
of H2O2 to produce OH + OH.
The kinetic factors that are responsible for the differences in radical
production rates, particularly for OH, between straight-chain fuels such as
n-heptane or n-octane and branched-chain fuels such as iso-octane, are

quite complex. However, the general principles are quite clear. The initial
process of H atom abstraction from the fuel does not depend strongly on
fuel structure; iso-octane has one tertiary H atom that is easily abstracted,
but it also has 15 primary H atoms that are relatively difficult to abstract,
whereas n-heptane has a large number (10) of secondary H atoms, with
abstraction rates somewhat intermediate between the primary and tertiary
sites. However, in branched-chain fuels such as iso-octane, virtually all of
the isomerization pathways that produce OH radicals at temperatures be-
low 1000 K are inhibited because some or all of the steps require the
abstraction of primary H atoms with their relatively large activation energy
barriers. In contrast, many such paths in the straight-chain fuels involve
only secondary H atom sites with significantly lower energy barriers. These
paths, dominated at low and intermediate temperatures by internal H atom
abstractions involving R, RO2, and OOROOH radical species, proceed
much more rapidly when only secondary H atom sites are involved, re-
sulting in an enhanced yield of OH radicals and a more rapid rate of fuel
consumption.
VI. Summary
The discussion presented in this chapter has shown how detailed chemical
kinetic models can be constructed to address problems associated with
hydrocarbon autoignition and engine knock. Engine knock represents a
practical combustion problem of great technical and economic importance;
thus, any insight or progress in dealing with this problem is extremely
valuable. Kinetic models of the present type are making significant progress
in addressing these problems. There is every expectation that, combined
with careful experimentation, they will continue to be essential elements
in practical combustion research.
Numerical modeling of this scale requires substantial computer resources
in CPU speed, large-scale storage, and robust and fast numerical solution
algorithms. The rate of growth in computer resources has paralleled the
expansion in the scope and size of the reaction mechanisms that have been
used in combustion simulations. This growth is illustrated in Fig. 7, in
which the mechanism size, measured by the number of chemical species
included in mechanisms taken from the current literature, is plotted as a
function of calendar year. The rate of growth is approximately exponential
during the period from about 1978 to the present. This parallel growth can
be expected to continue. This will permit larger, more complex problems
to be addressed with current reaction mechanisms, and it will permit current
problems to be addressed using even more complex fuels. One example
of this latter type of problem is the extension of reaction mechanisms to
describe aromatic hydrocarbon fuels, none of which has been described
kinetically in the detailed way described earlier for paraffinic hydrocarbon
fuels. Aromatic fuels make up a large fraction of most practical fuels such
as gasoline, and aromatic components are commonly used to increase the
octane rating of a given fuel mixture, so they must be described in kinetic
models eventually. Current growth in the speed and power of computer
Number of Species
1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 19831984 1985 1986 1987
Fig. 7 Schematic graph of the sizes of typical hydrocarbon oxidation reaction

mechanisms over the past 20 years, showing the approximately exponential growth
in the number of species (equations) to be solved.
resources makes it certain that even very large aromatic fuel mechanisms
will be soluble when they are developed.
Extensions of current kinetic modeling techniques to multidimensional
geometries are not yet feasible for the largest current reaction mechanisms.
Even one-dimensional problems are very costly with reaction mechanisms
for fuels such as octanes, and increases of factors of 30-50 in the number
of equations to be solved. Increases for two-dimensional applications are
much too expensive on current computers. Development of improved nu-
merical algorithms will provide some of this increase, but computer hard-
ware advances are needed to impact this large class of very difficult, but
potentially very significant, computer modeling problems.
Acknowledgments
The authors appreciate the extensive discussions, collaborative inter-
actions, and constructive criticisms from a large number of colleagues. In
particular, thanks go to E. I. Axelsson, K. Brezinsky, N. P. Cernansky,
M. Croudace, F. L. Dryer, R. M. Green, P. Jessup, W. R. Leppard, J.
Warnatz, and R. D. Wilk. This work was supported by the U.S. Depart-
ment of Energy, Division of Energy Conversion and Utilization Technol-
ogies, and the Office of Basic Energy Sciences, Division of Chemical Sciences.
This work was carried out under the auspices of the U.S. Department of
Energy by the Lawrence Livermore National Laboratory under Contract
W-7405-Eng-48.
References
!
Westbrook, C. K., and Dryer, F. L., "Chemical Kinetics of Hydrocarbon Com-
bustion," Progress in Energy and Combustion Science, Vol. 10, 1984, p. 1.
2
Warnatz, J., Combustion Chemistry, edited by W. C. Gardiner, Jr., Springer-
Verlag, New York, 1984.
3
Halstead, M. P., Kirsh, L. J., and Quinn, C. P., "The Autoignition of Hydro-
carbon Fuels at High Temperatures and Pressures—Fitting of a Mathematical
Model," Combustion and Flame, Vol. 30, 1977, p. 45.
4
Cox, R. A., and Cole, J. A., "Chemical Aspects of the Autoignition of Hydro-
carbon Air Mixtures," Combustion and Flame, Vol. 60, 1985, p. 109.
5
Kirsch, L. J., and Quinn, C. P., "Progress Towards a Comprehensive Model
of Hydrocarbon Autoignition," Journal de Chimie Physique, Vol. 82, 1985, p. 459.
6
Natarajan, B., and Bracco, F. V., "On Modeling Auto-Ignition in Spark-
Ignition Engines," Combustion and Flame, Vol. 57, 1984, p. 179.
7
Hu, H., and Keck, J. C., "Autoignition of Adiabatically Compressed Com-
bustible Gas Mixtures," Society of Automotive Engineers, Warrendale, PA, Paper
SAE-872110, 1987.
8
Pitz, W. J., and Westbrook, C. K., "Chemical Kinetics of the High Pressure
Oxidation of n-Butane and Its Relation to Engine Knock," Combustion and Flame,
Vol. 63, 1986, p. 113.
9
Cernansky, N. P., Green, R. M., Pitz, W. J., and Westbrook, C. K., "Chemistry
of Fuel Oxidation Preceding End-Gas Autoignition," Combustion Science and
Technology, Vol. 50, 1986, p. 3.
10
Green, R. M., Parker, C. D., Pitz, W. J., and Westbrook, C. K., "The Auto-
ignition of Isobutane in a Knocking Engine," Society of Automotive Engineers,
Warrendale, PA, Paper SAE-870169, 1987.
H
Smith, J. R., Green, R. M., Westbrook, C. K., and Pitz, W. J., "An Exper-
imental and Modeling Study of Engine Knock," Twentieth Symposium (Interna-
tional) on Combustion, The Combustion Institute, Pittsburgh, PA, 1984, p. 91.
12
Westbrook, C. K., Warnatz, J., and Pitz, W. J., "A Detailed Chemical Kinetic
Reaction Mechanism for the Oxidation of Iso-Octane and n-Heptane Over an
Extended Temperature Range and Its Application to Analysis of Engine Knock,"
Twenty-Second Symposium (International) on Combustion, The Combustion In-
stitute, Pittsburgh, PA, 1989, pp. 893-901.
13
Leppard, W. R., personal communication, 1987.
14
Obert, E. F., Internal Combustion Engines and Air Pollution, Harper & Row,
New York, 1973, pp. 234-235.
15
"Knocking Characteristics of Pure Hydrocarbons," American Society for Test-
ing Materials, Philadelphia, PA, ASTM Special Pub. 225, 1958.
16
Westbrook, C. K., and Dryer, F. L., "Chemical Kinetics and Modeling of
Combustion Processes," Eighteenth Symposium (International) on Combustion,
The Combustion Institute, Pittsburgh, PA, 1981 p. 749.
17
Westbrook, C. K., and Dryer, F. L., "A Comprehensive Mechanism for Meth-
anol Oxidation," Combustion Science and Technology, Vol. 20, 1979, p. 125.
18
Westbrook, C. K., Dryer, F. L., and Schug, K. P., "A Comprehensive Mech-
anism for the Pyrolysis and Oxidation of Ethylene," Nineteenth Symposium (Inter-
national) on Combustion, The Combustion Institute, Pittsburgh, PA, 1983, p. 153.
19
Westbrook, C. K., and Pitz, W. J., "A Comprehensive Chemical Kinetic Reac-
tion Mechanism for Oxidation and Pyrolysis of Propane and Propene," Combustion
Science and Technology, Vol. 37, 1984, p. 117.
20
Pitz, W. J., Westbrook, C. K., Proscia, W. M., and Dryer, F. L., "A Com-
prehensive Chemical Kinetic Reaction Mechanism for the Oxidation of n-Butane,"
Twentieth Symposium (International) on Combustion, The Combustion Institute,
Pittsburgh, PA, 1985, p. 831.
21
Miller, J. A., Mitchell, R. E., Smooke, M. D., and Kee, R. J., "Toward a
Comprehensive Chemical Kinetic Mechanism for the Oxidation of Acetylene: Com-
parison of Model Predictions with Results from Flame and Shock Tube Experi-
ments," Nineteenth Symposium (International) on Combustion, The Combustion
Institute, Pittsburgh, PA, 1983, p. 181.
22
Baulch, D. L., Drysdale, D. D., Home, D. G., and Lloyd, A. C., Evaluated
Kinetic Data for High Temperature Reactions, Vols. 1 and 2, Butterworths, London,
England, 1973.
23
Pitz, W. J., and Westbrook, C. K., "Interactions Between a Laminar Flame
and End Gas Autoignition," Dynamics of Reactive Systems Part II: Modeling and
Heterogeneous Combustion, Vol. 105, Progress in Astronautics and Aeronautics,
AIAA, New York, 1986, p. 293.
24
Dryer, F. L., and Glassman, L, "Combustion Chemistry of Chain Hydrocar-
bons," Alternative Hydrocarbon Fuels; Combustion and Chemical Kinetics, Vol.
62, edited by C. T. Bowman and J. Birkeland, Progress in Astronautics and Aer-
onautics, AIAA, New York, 1978, pp. 255-306.
25
Axelsson, E. L, Brezinsky, K., Dryer, F. L., Pitz, W. J., and Westbrook, C.
K., "Chemical Kinetic Modeling of the Oxidation of Large Alkane Fuels: N-Octane
and Iso-Octane," Twenty-First Symposium (International) on Combustion, The
Combustion Institute, Pittsburgh, PA, 1988, p. 783.
26
Westbrook, C. K., and Pitz, W. J., "Kinetic Modeling of Autoignition of Higher
Hydrocarbons—n-Heptane, n-Octane, and iso-Octane," Complex Chemical Re-
action Systems Mathematical Modeling and Simulation, edited by J. Warnatz and
W. Jager, Springer-Verlag, Heidelberg, West Germany, 1985, pp. 39-54.
27
Westbrook, C. K., and Pitz, W. J., "Chemical Kinetics of the Influence of
Molecular Structure on Shock Tube Ignition Delay," Shock Waves and Shock
Tubes, edited by D. Bershader and R. Hanson, Stanford University Press, Stanford,
CA, 1986, p. 287.
28
Westbrook, C. K., Pitz, W. J., Thornton, M. M., and Malte, P. C., "A Kinetic
Modeling Study of N-Pentane Oxidation in a Well-Stirred Reactor," Combustion
and Flame, Vol. 72, 1988, p. 45.
29
Westbrook, C. K., "Chemical Kinetic Modeling of Higher Hydrocarbon Fuels,"
AIAA Journal, Vol. 24, 1986, p. 2002.
30
Westbrook, C. K., "Numerical Simulation of Chemical Kinetics of Combus-
tion," Chemical Kinetics of Small Organic Radicals, edited by Z. Alfassi, Chemical
Rubber Co. Press, Boca Raton, FL, 1988, pp. 59-95.
31
Slagle, I. R., Park, J. Y., and Gutman, D., "Experimental Investigation of the
Kinetics and Mechanism of the Reaction of n-Propyl Radicals with Molecular
Oxygen from 297 K to 653 K," Twentieth Symposium (International) on Combus-
tion, The Combustion Institute, Pittsburgh, PA, 1984, p. 91.
32
Baldwin, R. R., Dean, C. E., and Walker, R. W., "Relative Rate Study of
the Addition of HO2 Radicals to C2H4 and C3H6," Journal of the Chemical Society,
Faraday Transactions, Ser. 2, Vol."82, 1986, p. 1445.
33
Burcat, A., Scheller, K., and Lifshitz, A., "Shock-Tube Investigation of Com-
parative Ignition Delay Times for C,-C5 Alkanes," Combustion and Flame, Vol.
16, 1971, p. 29.
34
Westbrook, C. K., "An Analytical Study of the Shock Tube Ignition of Mixtures
of Methane and Ethane," Combustion Science and Technology, Vol. 20, 1979,
p. 5.
35
Dryer, F. L., and Glassman, I., "High-Temperature Oxidation of CO and
CH4," Fourteenth Symposium (International) on Combustion, The Combustion
36
Westbrook, C. K., Creighton, J., Lund, C., and Dryer, F., "A Numerical
Model of Chemical Kinetics of Combustion in a Turbulent Flow Reactor," Journal
of Physical Chemistry, Vol. 81, 1977, p. 2542.
37
Kaiser, E. W., Westbrook, C. K., and Pitz, W. J., "Acetaldehyde Oxidation
in the Negative Temperature Coefficient Regime: Experimental and Modeling
Results," International Journal of Chemical Kinetics, Vol. 18, 1986, p. 655.
38
Wilk, R. D., Cernansky, N. P., Pitz, W. J., and Westbrook, C. K., wt Propene
Oxidation at Low and Intermediate Temperatures: A Detailed Chemical Kinetic
Study," The Combustion Institute Spring Meeting, Paper 87-42, April 1987.
39
Wilk, R. D., Cernansky, N. P., and Cohen, R. S., "An Experimental Study
of Propene Oxidation at Low and Intermediate Temperatures," Combustion Sci-
ence and Technology, Vol. 52, 1987, p. 39.
40
Griffiths, J. F., "The Fundamentals of Spontaneous Ignition of Gaseous Hydro-
carbons and Related Organic Compounds," Advances in Chemical Physics, Vol.
64, 1986, p. 203.
41
Lignola, P. G., and Reverchon, E., "Cool Flames," Progress in Energy and
Combustion Science, Vol. 13, 1987, p. 75.
42
Lignola, P. G., Reverchon, E., Autuori, R., Insola, A., and Silvestre, A. M.,
"Propene Combustion Process in a CSTR," Combustion Science and Technology,
Vol. 44, 1985, p. 1.
43
Frank, J., Griffiths, J. F., and Nimmo, W., "The Control of Spontaneous
Ignition under Rapid Compression," Twenty-First Symposium (International) on
44
Griffiths, J. F., and Hasko, S. M., "Two-Stage Ignitions During Rapid
Compression: Spontaneous Combustion in Lean Fuel-Air Mixtures," Proceedings
of the Royal Society of London, Series A: Mathematical and Physical Sciences, Vol.
393, 1984, p. 371.
45
Pinchon, M., Institute Francaise du Petrole, manuscript in preparation, 1989.
Chapter 4
Combustion Kinetics and Sensitivity

Analysis Computations
K. Radhakrishnan
I. Introduction
C OMBUSTION reactions, which are accompanied by rapid and large

increases in temperature, occur in many practical devices, such as gas
turbine combustors and internal combustion engines. Detailed modeling
of such devices requires knowledge of the reaction mechanism, i.e., the
set of elementary chemical reactions that together produce combustion and
determine its rate. 1 - 2 Mathematical descriptions of reaction mechanisms by
means of the principle of mass action result in systems of coupled, nonlin-
ear, first-order ordinary differential equations (ODE's). 13
In this chapter, we focus attention on two topics: 1) solving homoge-
neous, static (i.e., no transport of energy or matter), combustion kinetics
problems, and 2) methods for sensitivity analysis that provide the effects
of uncertainties or changes in the input parameters of the problem on the
solutions. Both subjects are relevant to the development and validation of
reaction mechanisms and to the development of simplified reaction mech-
anisms for specific applications.
Combustion Kinetics Equations

The ODE's describing homogeneous, static, gas-phase elementary chem-
ical reactions are given by 4 " 6
^ = /(K},7», i,k=i,...,NS (i)

where cr, is the mole number of species / (moles of species / per gram of
mixture), t the time, T the temperature, p the mixture mass density, and
NS the total number of chemical species.
83
84 COMBUSTION KINETICS AND SENSITIVITY ANALYSIS
The net rate of formation of species /, fh is given by6

NR
IJ IJ J
y=l ~J
In Eq. (2), NR is the total number of distinct chemical reactions; v'tj is the
stoichiometric coefficient (i.e., number of moles) of reactant species / in
reaction/; v'-j is the stoichiometric coefficient of product species / in reaction
y; and the molar forward (/?y) and reverse (/?_/) rates per unit volume for
reaction y are given by the law of mass action7
j = 1, ..., NR (3)
(=1
NS
R-j = k-i II (P<re)£ i = 1, »., NR (4)
Here kj and k_}- are, respectively, the forward and reverse rate constants
for reaction y. Each kj is a function of temperature and is usually given by
the empirical expression2
fy = AjT"> exp(-EjlRT) (5)
where the pre-exponential factor A^ the temperature exponent n^ and the
activation energy Ej are constants, and R is the gas constant.
For a reversible elementary chemical reaction, /c_ y and kf obey the prin-
ciple of microscopic reversibility or detailed balancing 1 7
k.j = kj/KCJ (6)
where Kcj, the concentration equilibrium constant for reaction y, is a func-

tion of only temperature for ideal gas mixtures. 1 6
To solve for the temporal evolution of the temperature during the com-
bustion process, appropriate forms of the energy conservation equation
are used. For example, the energy equation for adiabatic constant-pressure
problems is
NS
^ &{&{ — const (7)
where ^ is the molar-specific enthalpy of species €. Time differentiation

of Eq. (7) provides the following ODE for the temperature:
j j, NS I NS
where cpfi is the molar-specific heat at constant pressure of species €. For

ideal gases, the thermodynamic properties fi( and cp^ are functions of
temperature alone. Their temperature dependencies are usually described
by polynomials in engineering calculations.8 9
K. RADHAKRISHNAN 85
For adiabatic constant-pressure problems, either Eq. (7) or (8) can be

used to compute the temperature. If Eq. (7) is used, T is calculated from
the {cr,} and the initial mixture enthalpy. An iterative technique is em-
ployed, and Tis adjusted until Eq. (7) is satisfied. We explore the use of
both calculation procedures for T. For clarity in presentation, the two
temperature calculation methods are labeled A and B, with method A
using the algebraic equation [Eq. (7)], and method B using the ODE [Eq.
(8)].
For nonadiabatic reactions, models for heat transfer are required, and
the temperature ODE, which is somewhat more involved than Eq. (8),
must be solved.6 Variable-density problems require an ODE or an algebraic
equation for the density. For such problems, the pressure is specified as a
function of time, and the density, or its ODE, is obtained from the ideal
gas law.6
The governing ODE's can be written in vector form as
(9a)
y (* = t0) = y() = given (9b)

where y is an TV-dimensional vector with
yc = a,-, i = 1, ..., NS (10a)
y»s+i = T (lob)
y*s+2 = P (10c)
TV - NS + 2 (lOd)
The computational problem is to solve for y at the end (/cnd) of a prescribed

time interval, given the initial conditions, the reaction mechanism, and the
rate constant parameters.
The number of ODE's can be large because of the numerous chemical
species involved in the combustion process. Validating a reaction mecha-
nism requires repetitive solutions of the governing ODE's for a variety of
reaction conditions. Consequently, there is a need for fast and reliable
solution techniques for combustion kinetics problems. For reacting flow
problems, the need for fast kinetics solvers is even more critical because
the rate equations may require integration at several thousand grid points.
In fact, solution of the chemical kinetics equations is the most expensive
part of detailed modeling of reactive flowfields. 10
Computational Difficulty — Stiffness

Figure 1 illustrates the solution to a typical adiabatic combustion kinetics
problem. This constant-pressure problem, taken from Ref. 11, describes
the ignition and subsequent combustion of a carbon monoxide-hydrogen-
air mixture at an initial temperature of 1000 K and a pressure of 10 atm.
The reaction mechanism consists of 12 reactions among 11 species and is

given by Radhakrishnan. 4
Three distinct regimes, commonly denoted as induction, heat release,
and equilibration, are apparent in Fig. 1. The induction regime is the period
of time immediately following some form of reaction initiation, such as the
passage of a shock wave through the fuel-air mixture. During the induction
period, concentrations of chemically active species (such as O, H, OH)
increase from very small initial values to values sufficiently large to initiate
oxidation reactions. The induction period ends and the heat-release period
begins when sharply defined changes in temperature and concentrations
occur. During heat release, the temperature and the concentrations of
combustion products and active intermediates increase very rapidly. The
heat-release period ends and the equilibration period begins after the active
intermediate species have reached their maximum values. The equilibration
regime is characterized by the monotonic, asymptotic approach of all var-
iables toward their chemical equilibrium values.
Solution of the relatively small problem illustrated in Fig. 1 was at-
tempted with a classical numerical method, the explicit fourth-order Runge-
Kutta method. [Detailed discussion of classical numerical methods for ODE's
is available in any of the numerous books 12 ~ 2: " or review articles 26 " 28 pub-
lished on this subject.] In general, numerical methods replace the ODE's
with difference equations, which are solved step by step. Starting with the
initial conditions at time /0, approximations Yn [= V/(O; 'l = 1> ..., W] to
the exact solutiony(t n ) [= }>/(O; / = 1, ..., N] of the ODE's are generated
at discrete points in time, tn (n = 1 , 2 , . . . ) . The spacing between any two
3000
- "2
2500
U v /
co^
n-2
2000
n-4
1500 —
INDUCTION f*HEAT RELEASE*)—————EQUILIBRATION
10"r5 ~
1000 '— ,-6 ^ I i I i 1 1 hi I / lAylvKI/ I i I iI

10,-7 n-6 10" n-3
TIME, s
Fig. 1 Species mole fractions and temperature profiles for constant-pressure (10-
atm) combustion of CO-H2-air mixture. Initial temperature is 1000 K.
K. RADHAKRISHNAN 87
points is called the "step size," or "step length," and is denoted by h,,,
where
The step length can vary from one step to the next, or can be constant,
and is usually computed by the numerical integration method.
Figure 2 presents a plot of the step length successfully used by the explicit
Runge-Kutta method in solving the problem illustrated in Fig. I. 29 During
induction and heat release, the step lengths selected by the method are
small because the solution changes very rapidly in these regimes. During
late heat release and equilibration, however, when the variables change
slowly, larger step lengths should be possible. Surprisingly, for reasons
discussed subsequently, the step lengths continue to remain small and are
of the same magnitude as the step lengths used in the earlier regimes.
Consequently, the number of steps and computer time (i.e., CPU time)
required for the problem are large—approximately 10,700 steps and 1 min,
respectively, on an IBM 370/3033 computer. 29 Such large computational
work required by classical ODE solution methods has impeded the de-
velopment of complex reaction mechanisms.
The first numerical integration method suitable for chemical kinetics
problems was reported by Curtiss and Hirschfelder 30 in 1952. They used
the term "stiff" to describe the ODE's arising in chemical kinetics. For
reasons discussed by Shampine, 31 stiffness does not have a simple definition
involving only the mathematical problem [Eq. (9)]. However, stiff equa-
tions are characterized by time constants of widely varying magnitude. 32 - 33
In addition, Shampine and Gear34 discuss some fundamental issues related
to stiffness and how it can be recognized.
10
EXPLICIT RUNGE-KUTTA
~ — — —— LSODE
10"
10"
10"
^
UJ
5^
10"7
10"9 HEAT RELEASE -v vEQUILIBRATION

INDUCTION
10,-10
' ) i hlilJ | ihlilil | ilihlJ I ihhlJ I i hiilJ I ihlihl I ihlilil
n-10
10" IU
103 10 ° n-7
10 ' 10 10~3 10 10 -3
TIME, s
Fig. 2 Step length variation with time for kinetics

test problem illustrated in Fig. 1.
From a practical viewpoint, numerical stability requirements limit the

size of the step lengths used by classical ODE solution methods to integrate
stiff problems. Because of the approximate nature of numerical solutions,
errors are inevitably introduced at every time step. Numerical stability
requires that the errors remain bounded as the calculation proceeds in
time. Stiff problems force classical ODE solution methods to use unac-
ceptably small step lengths to maintain numerical stability. 35
An alternative practical description of a stiff problem is one whose so-
lution by classical ODE methods is significantly more expensive than its
solution by stiff methods (i.e., methods designed specifically for stiff prob-
lems).36-37 For example, Fig. 2 shows the step lengths selected by a stiff
ODE solver, LSODE,38 39 through the course of solving the problem il-
lustrated in Fig. 1. Significantly larger step lengths are selected by the stiff
ODE solver, especially during late heat release and equilibration. Con-
sequently, the computational work required by the stiff ODE solver—
approximately 90 steps and 0.35 s 29 —was significantly less than that re-
quired by the explicit Runge-Kutta method.
One requirement for a problem to be stiff is that the ODE's are sta-
^^31,34-37,40-42 m tê sense fa^ different solution curves to the ODE's
approach one another for increasing ^17.21,23,28.37 jn t h e example problem
(Fig. 1) during heat release, especially in the early part, many of the species
and the temperature increase exponentially and are described by unstable
ODE's; therefore, this regime is not stiff. During late heat release and
equilibration, however, the governing ODE's are stable, and these regimes
are stiff.
In addition to being stable, the numerical method needs to be accurate,
especially to track the unstable modes. Accuracy of a numerical method
refers to the size of the error introduced in a single step, or, more precisely,
the local truncation or discretization error. The local truncation error at
time tn is the difference between the exact solution and the computed
approximation obtained using exact past values. The accuracy of a nu-
merical method is usually estimated by its order. A method is said to be
of order q if the local truncation error varies as hf/t +l.
Since 1952, when the first paper30 recognizing stiffness in chemical ki-
netics was published, many methods have been developed for the solution
of stiff ODE's. The literature in the field is vast, with information available
in many books,10 < 17 - 19 - 23 ^ 43 - 53 review articles, 28 - 34 - 37 - 54 ~ 56 and journal ar-
ticles (see Refs. 10, 28, 37, 41, and 53-57 for extensive lists). We partic-
ularly recommend Ref. 53 for an exhaustive review of the state-of-the-art
and practice of stiff computations. The review article by Byrne and
Hindmarsh37 is more recent, and the authors discuss most modern methods
for stiff ODE's. Details of solution techniques developed for chemical
kinetics calculations are provided in Refs. 10, 44, 45, 51, and 54; for
reviews, analyses, and comparisons, see Refs. 4, 29, 41, and 58-68.
Sensitivity Analysis
The solutions of the ODE's describing combustion chemistry depend on
a number of input parameters, including, for example, the initial conditions
and rate constant parameters. It is often necessary to know what effect
K. RADHAKRISHNAN 89
variations of input parameters have on the solutions. Such a need arises

in the development of reaction mechanisms.64 The rate constants are often
not known with great precision, and, in general, the experimental data are
not detailed sufficiently to estimate accurately the rate constant parameters.
The development of a reaction mechanism is facilitated by a systematic
sensitivity analysis, which provides the basic methods to study parameter
sensitivities — i.e., the changes in model behavior due to parameter vari-
ations.69 Sensitivity analysis in chemical kinetics provides relationships be-
tween the predictions of a kinetic model and the input parameters of the
problem. It helps to determine the effects of uncertainties in rate constant
parameters or errors in initial conditions on the model predictions. Thus,
sensitivity analysis helps to identify parameters that have to be determined
accurately because they have a large effect on the solution. Rate parameters
that show little or no effect need not be known with great precision. In
some cases, these latter reactions may even be eliminated, thus simplifying
the mechanism. Perhaps the most significant utility of sensitivity analysis
is the increased understanding it provides of a complex reaction mecha-
nism. Other advantages and applications of sensitivity analysis are discussed
in Refs. 3 and 70.
Sensitivity analysis methods fall into two categories: 1) local methods
and 2) global methods. Local methods typically produce the effect of a
small change in one parameter about its nominal value by computing the
first-order sensitivity coefficient dyjdi\j evaluated at the nominal values
{f)^}. Here y{ is a variable of interest (e.g., a concentration or temperature)
and r\j a rate constant parameter or an initial condition value. The first-
order sensitivity coefficient vector 5y can be defined by
, j=l,...,M (12)
I/ all other j}k held constant
In Eq. (12), y is the TV-dimensional vector containing the dependent var-

iables, concentrations, temperature, and density; r\j is either a rate constant
parameter [Eq. (5)] or an initial condition value; M is the total number of
independent parameters; and Sy is a column vector of length TV, with the
ith component, 5/y, given by
Most local methods of sensitivity analysis obtain Sf- by solving ODE's

for the first-order sensitivity coefficients. To derive these ODE's, we re-
write Eq. (9) as
=
'jE) (14a)
y(t = t0) = >>„ (i4b)

to emphasize the functional dependence of y(t) on the rate constant pa-
rameters A, n, and E, each of which has NR components. For isothermal
reactions, the rate constants {kj} are time invariant, and there is then no
need to consider the individual rate constant parameters. For such prob-
lems, Eq. (14a) simplifies to
f = f(y,k) (15)
where k is the rate constant vector containing the NR rate constants.

The ODE for Sp obtained by differentiating Eq. (14a) with respect to
r\j and interchanging the order of differentiation with respect to t and iq7,
is
= js, + - (16a)
^** \Jy
where / is the Jacobian matrix of size N x N, with element Jmn given by

N
'- - & •-- - ' ••••
In Eq. (16a), the first term on the right-hand side accounts for the implicit
dependence of/on r\j through j, and the second term takes into account
any explicit dependence o f / o n r\. If T]y is an initial condition value,/is
not a function of r|y, and Eq. (16a) reduces to
The M vector equations [Eqs. (16)] are independent of each other. In

addition, although the model [Eq. (14a)] can be highly nonlinear, Eqs.
(16) are linear ODE's with time-dependent coefficients. The initial value
of Sj is equal to the null vector if r)y is a rate constant parameter and is
equal to trie yth column of the identity matrix if r\j is the yth element of y().
Each parameter T); results in N ODE's given by Eqs. (16). Hence, for
M independent parameters, NM ODE's have to be solved to obtain all
first-order sensitivity coefficients.
For meaningful comparisons with respect to different T)y values, sensi-
tivity coefficients are usually presented in normalized form. A simple way
of accomplishing the normalization is to use the relation
T,^ = ^
J
yfd^j d/,,T]j
where <S/y) is the normalized sensitivity coefficient 5,y. This procedure works
when the r]y values are nonzero. For example, for isothermal problems,
the rate constants {&,} [Eq. (5)] are time invariant, and setting T]; = kj
presents no difficulty with the use of Eq. (18). The (5iy) can then be in-
terpreted as the percent change in y, due to a 1% change, or uncertainty,
in kj. However, for nonconstant-temperature problems, the rate constants
K. RADHAKRISHNAN 91
{kj} vary with time, so that sensitivities with respect to the individual rate
constant parameters {A}}, {«,}, and {£,} [see Eq. (5)] must be computed.
Since the temperature exponents and activation energies can be zero, Eq.
(18) cannot be used, and alternative normalization procedures have been
developed.6'71
Some local methods have also been used to examine the simultaneous
variation of two parameters by calculating the second-order sensitivity
coefficient d^/d^l/dr^, where r\j and j]k are different parameters. The first-
and second-order sensitivity coefficients are the first two terms of a Taylor
series expansion of the solution about the parameters' nominal values.
Such an expansion is strictly valid only for small changes in the parameters.
The restriction to small changes does not apply to global methods, which
can account for simultaneous parameter variations of arbitrary magnitudes.
These techniques provide an average effect of all uncertainties examined.
The sensitivities obtained are, therefore, different from the sensitivity coef-
ficients defined earlier. For small variations, however, there is a simple
relationship between the coefficients calculated by global methods and the
linear sensitivity coefficients [Eq. (12)].72 Since global methods require
repeated solution of the chemical kinetics equations, their computational
cost can become prohibitive as the number of independent parameters
increases.73"75 Global methods have been reviewed by Cukier et al.,76
Tilden et al., 77 Rabitz et al.,70 and briefly by Oran and Boris.10
Scope of Chapter
The remainder of the chapter is organized as follows. In Sec. II, the stiff
ODE solution methods more commonly used for chemical kinetics prob-
lems are examined. In addition, the efficiency and accuracy of several
solution methods are compared by application to two combustion kinetics
problems. In Sec. Ill, the local sensitivity analysis methods developed to
date are outlined, and their advantages and difficulties are discussed briefly.
The section concludes with some examples and comparisons among the
different techniques. Section IV presents a summary and a discussion of
areas that need further development.
II. Methods for Stiff ODE's

The solution methods examined here can be classified conveniently into
two types—single-step and multistep. A single-step method provides a rule
for computing the approximation Yn at the current time tn by using the step
length hn and information only from one previous time / „ _ , . Multistep
methods, on the other hand, use hn and information from several previous
time steps to generate Yn.
Linear Multistep Methods: Backward Differentiation Formulas

The general linear multistep method may be written in the form 37
where F; [=/(f/,y/)] is an approximation to the exact derivative y(tj)

[=f[tj,y(tj)]}. The method is called linear because the Yj and Fy- occur lin-
early. The coefficients an€ and $ne and the constants A^ and K2 are asso-
ciated with a particular method. For example, the choice Kl = 1, K2 =
q — 1 results in the popular implicit Adams, or Adams-Moulton, method
of order q. The method is called implicit because it uses the as yet unknown
Yn to compute Yn.
The choice K1 = q, K2 = 0 results in the backward differentiation
formula (BDF) of order q. BDF's have much better stability characteristics
than the Adams-Moulton method and are, therefore, preferred for stiff
problems. The BDF of order q is given by
Yn = 2 a*<r*-< + h,$n[?n (20)

t =1
The coefficients ani and p/l0 are predetermined constants corresponding to
the order q, if the step size hn is constant. Otherwise, they can vary from
one step to the next and have to be computed at each step. If p //0 = 0,
the method is explicit because it involves only known quantities, Yn_l,
F n _ 2 , etc., and Eq. (20) is easy to solve. Otherwise, the method is implicit
and, in general, solution of Eq. (20) is expensive; because /is nonlinear,
some type of iterative procedure is usually used to solve Eq. (20). Never-
theless, implicit methods are preferred because they can use much larger
step lengths than explicit methods and are also more accurate. 17 - 19 Explicit
methods are used as predictors, which generate an initial guess for Yn. The
implicit method corrects the initial guess iteratively and provides a rea-
sonable approximation to the solution of Eq. (20).
Many researchers have made significant contributions to the develop-
ment of BDF's, and numerous computer codes exist.37 However, the meth-
ods developed by Gear17-32 as implemented by Hindmarsh in the codes
GEAR,78 EPISODE,79 and LSODE38-39 are among the most popular in
use today for combustion kinetics calculations. In GEAR and LSODE,
the step size is fixed for a prescribed number of steps; in EPISODE,
however, the step size is allowed to change at each step. All three codes
use a standard predictor formula and include several corrector iteration
techniques — Newton-Raphson,49 functional, 49 and Jacobi-Newton.80 New-
ton-Raphson iteration is given by the recursive relation
q
Q Fy(m + l) _ My(m)l _ V y , + h Q nf\t F ( m ) l ~ MF("° (21)
w
L n n J Z~i ^*-ntLn — t ' 'lnrJn()J lln->£ n J n V^1/
e= i
Here (rri) and (m + 1) denote the iteration numbers, and the iteration
matrix G, of size TV x TV, is given by
G = I - h,$MJ (22)
where / is the identity matrix and / the Jacobian matrix [Eq. (17)].
This iteration technique requires much computational work to form the
Jacobian matrix and to perform the linear algebra necessary to solve Eq.
(21). To reduce the computational work associated with matrix algebra,
K. RADHAKRISHNAN 93
the iteration matrix G is updated only when the solution given by Eq. (21)
does not converge.
Functional and Jacobi-Newton iteration techniques do not involve any
matrix algebra and, therefore, require much less work per step than New-
ton-Raphson iteration.81 However, for successful convergence of functional
and Jacobi-Newton iterations, the step length may be restricted to very
small values. Consequently, these two iteration techniques require pro-
hibitive amounts of computer time for combustion kinetics problems, and
we do not recommend their use (see Ref. 6 for details).
A useful feature of the GEAR, EPISODE, and LSODE codes is that
they will estimate the elements of the Jacobian matrix by finite-difference
approximations if the user chooses not to provide analytical expressions
for them. However, we recommend using analytical Jacobians because
they result in significantly smaller CPU times than numerical Jacobians
(see Ref. 6 for details).
Irrespective of the iteration method selected, the test for convergence
of the iterates Y(nm} is based on the magnitudes of some successive differ-
ences such as [y,("7) — y^""1*]. After successful convergence, an estimate
of the local truncation error is made; it is accepted if it is less than a user-
specified bound. Periodically, the codes adjust both the step size and the
order to minimize computational work while maintaining prescribed ac-
curacy.
Runge-Kutta Methods
Runge-Kutta (R-K) methods are single-step methods, but they involve
intermediate points, or stages, in the step. They are designed to agree with
Taylor series expansions without requiring evaluation of derivatives higher
than the first. As discussed in the previous section, explicit R-K methods
are unsuitable for chemical kinetics applications because of their poor
stability characteristics. Implicit R-K methods, on the other hand, are
attractive for stiff problems because they are highly stable and can attain
higher orders of accuracy than explicit R-K methods. The general form of
an r-stage implicit R-K method is10-18-19-28-37-49
Yn = r,,-i + hn 2 b<k( (23)

i =i
where € labels the stage and
r \
-i + c<hn, ¥„_, + E a,,.*,-, € = 1, ..., r (24)
The quantities {a^, {b^, and {q} are constants, obtained by requiring Eq.
(23) to agree with a specified number of terms in a Taylor series. The
{Q} satisfy the constraints
Q = 2 atj, € = 1, . . . , r (25)
For explicit R-K methods, the upper limit in the summation for k{ [Eq.
(24)] is € - 1. Each k( is then obtained explicitly in terms of the k values
at the previous stages. For implicit R-K methods, however, each k is defined
implicitly; as a result, the r vector equations [Eq. (24)] are more difficult
to solve.
Because of the enormous difficulty and expense of having to solve the
rN nonlinear equations [Eq. (24)], several special cases have been stud-
|ecj 28,37,54,56 Methods that have been applied successfully to chemical ki-
netics use semi-implicit R-K formulas. Here, the upper limit in the summation
for kt [Eq. (24)] is set equal to €. However, the approach still requires the
solution of r systems of TV nonlinear equations. To avoid iterative solution
for kf, Rosenbrock82 linearized Eq. (24) to give, in autonomous form, 56 - 83
(/ - atihfj)k( - f Yn.1 + hn a(jkj, € = 1, ..., r (26)

V y=i /
Although Eq. (26) is a linear system of equations, r systems of N equations

must be solved. The method requires the evaluation o f / at each step and
one Newton-Raphson iteration of Eq. (26). Many similar methods have
been proposed or developed (for details see Refs. 37, 49, 56, and 83). For
example, Rosenbrock-Wanner (ROW) methods (see, for example, Refs.
84 and 85) were developed to extend the stability of the Rosenbrock meth-
ods. In the ROW methods, terms involving the Jacobian matrix are added
to Eq. (26) and are given by
Yn-i + hn ^ <*tjkj + hnJ 2 dij

7=1 /=1
€ = 1, ..., r (27)
The coefficients {d^} are chosen to optimize order and stability properties.
All of the afc values are set equal to one another to reduce the required
matrix algebra.
The ROW method was used for chemical kinetics problems by Gottwald
and Wanner. 85 Their code, ROW4A, is based on the code GRK4A of
Kaps and Rentrop. 84 ROW4A uses a four-stage method and is fourth-
order accurate. The local truncation error is estimated using an embedded
third-order method. The error estimate provides the step size to be at-
tempted on the next step.
Extrapolation Methods
The basic idea behind extrapolation methods is to regard the solution
Yn as an analytic function of the step size h and to evaluate the function
at h = 0. Such methods involve solving the ODE's repeatedly over the
same time interval [£„_!,*„]. However, each integration is done with a
different step size. The solutions obtained with different step sizes h can
be fit to an analytic form, usually a polynomial or a rational function. The
numerical approximation corresponding to h = 0 is then obtained by ex-
K. RADHAKRISHNAN 95
trapolation. The difference in two successive numerical approximations

corresponding to h = 0 gives an estimate of the local truncation error.
High accuracy can be obtained by extrapolation, although the basic
integration method used to obtain successive approximations is of low
order. The reason is that the combination of numerical results obtained
with different step lengths results in the cancelation of leading error terms
in the asymptotic error expansion of the numerical method. 12 1924 - 28 Thus,
extrapolation gives a higher-order approximation to the true solution.
The accuracy of the extrapolation procedure depends on the basic nu-
merical method used. Because explicit numerical methods are unsuitable
for stiff ODE's, both implicit and semi-implicit extrapolation methods have
been proposed. Recently, Deuflhard 86 has reviewed extrapolation meth-
ods.
For chemical kinetics applications, Deuflhard et al. 87 ~ 89 have developed
the LARKIN package. The code includes two integration methods: 1) a
semi-implicit midpoint rule due to Bader and Deuflhard, 90 together with
an h2 extrapolation, and 2) a semi-implicit Euler method with h extrapo-
lation.86 The choice of integration method is left to the user.
In the semi-implicit midpoint rule, the basic step size hn is divided into
an even number, L, of subintervals, each of size /z, called the "internal
step size." A two-step method is used to advance the solution by intervals
of 2h. A semi-implicit Euler method is used to start the process, and the
integration is completed with a final smoothing step. We now illustrate
how the method advances the solution from tn_l to tn. For clarity in pres-
entation, we denote the approximations obtained at the subinterval tn_l
+ fh(( = 0, 1, ..., L) by Z(. The sequence for computing Yn from Yn_l
is as follows:
1) Starting step:
*o = Yn-i (28a
Azo - zl - zo = (I ~ hJ)-1 A/(z0) (28b
2) Semi-implicit midpoint steps (€ = 1, ..., L — 1):
Zt = zt-t + Az<_! (29a
1
Az, = Az,_! + 2(1 - hJ)- [hf(z<) - Az,_,] (29b)
3) Final smoothing step:
ZL = zL-j + Az L _! (30a)
AzL = Az L _! + (/ - hJ)~l W(zL) - Az^] (30b)
ZL = Z L _! + AzL (30c)
The Jacobian matrix is computed at the start of the step. The method is
called semi-implicit because it uses only one Newton-Raphson iteration for
each subinterval.
The integration over the basic step length hn is repeated for various
values of L, and Yn is obtained by polynomial extrapolation to h = 0.
Bader and Deuflhard90 have shown that the asymptotic error expansion of
the midpoint implicit method contains only even powers of h. Hence, if
the integration over hn is carried out q times, the local error in Yn is
0(/zJ? + 1). The local truncation error estimate is used to obtain the step size
and the order to be attempted on the next step.87 91
Hybrid Methods: Selected Asymptotic Integration Method
As discussed in the previous section, not all of the ODE's arising in
combustion chemistry are stiff. For nonstiff ODE's, classical solution meth-
ods, which require less work per step than implicit methods, may work
satisfactorily. However, stiff ODE's must be solved by a stable method.
Hybrid methods, which use stable methods for stiff ODE's and classical
methods for nonstiff ODE's, have been used by Keneshea, 92 Keneshea
and Swider,93 Young and Boris,94-95 and Young. 96
In these methods, the ODE's are expressed as differences between two
positive-definite terms as follows:
-II = f. = Q. - &i = Q. - L.y. = Q. - y.fr.
i = 1, ..., N (31)
where Q, and Dt are, respectively, the production and destruction rates of
species /, and Lt—called the 'loss coefficient"—is obtained simply by
dividing Di by yt. The temperature ODE can also be cast in a similar
form.4'68 For constant Qt and Lh Eq. (31) can be solved over the time step
Dn-i, tn] to give
Equation (32) shows that T, (= 1/L,-) controls how quickly the variable yt
reaches its equilibrium value.
If a classical method were applied to Eq. (31), stability requirements
would restrict hn to satisfy (/J,/T,) < 1 for all /. For each step n, with a
given step size hn, we can classify ODE's as stiff or normal (i.e., nonstiff)
according to the criteria
hnhf < 1 normal (33a)
and
hnht > I stiff (33b
For equations classified as normal, classical ODE methods can be used
for the step. For stiff equations, a stable method is required. For example,
a widely used hybrid method, the selected asymptotic integration method
(SAIM),94"96 solves individual stiff equations in the implicit form10-94-95
yitn - yt,n-i _ Q,n + 6/.,,-i y,> + y/.»-i
hn 2 T,-,, 4- T / t / I _ !
by using a predictor-corrector technique. The code CHEMEQ94 95 uses the

SAIM for stiff equations. For nonstiff equations, the improved Euler method49
K. RADHAKRISHNAN 97
is used. Young96 subsequently revised the algorithm for stiff ODE's by

writing Eq. (34) in terms of the loss coefficient L,:
*i,n ~ */.M-l Qi,n + Qi,n-l _ *i,n ~^~ * / ' . » - ! ^/'./; ~^~ ^ / . n - 1
The two sets of equations, stiff and nonstiff, are solved simultaneously.
The corrector equations are iterated until all variables have converged,
i.e., the normalized difference in successive iterates is less than a user-
supplied tolerance. After successful convergence, the step size to be at-
tempted on the next step is computed from the converged integration
cycle.94"96
Exponential Methods
The components of the solution of a stiff system of ODE's behave as
decaying exponential functions. Because exponentials are represented poorly
by polynomials when the step size is larger than the decay time, exponential
functions have been proposed for stiff ODE's. For example, Keneshea 92
suggested using exponential functions in the equilibration regime of a chem-
ical kinetics problem. Here the variables exhibit asymptotic behavior (Fig.
1), which is well represented by an exponential function with a negative
time constant. The rapid increase of reactive species during induction sug-
gests the use of exponentials in this regime also.41
Exponential integration methods can be obtained by curve fitting. For
example, we assume that in the interval |/,,__i, tn] (i.e., locally) each com-
ponent yf varies exponentially. In particular, we consider
y.(t) = A, + BteZlt, i = 1, ..., N (36)
where {At}, {£,}, and {Z,} have to be determined. By requiring each yf(t)
to satisfy the constraints ^(^-i) = YLn_lt .y/fe-i) = A/i-i andy,(O =
y /iW , we can solve for Ai and Bt in terms of Z,. The result is
Yitn = n«-i + &«//,„-! [(eZihn ~ 1)/ZM i = 1, ...,N (37)
which can be called an exponential difference equation. 97 The parameter
{Z,} can be obtained in a number of ways, 40 as, for example,
Explicitly:
Zt =/L-i//;>-i, i = 1, ....N (38)
Implicitly:
Zf = ^(fJf^-iVhn, /=!,...,# (39)
where /' denotes d//cU.
Although Eq. (37) is stable for all values of hn (for Z, < 0), it is not
particularly useful in this form. 40 A more useful approach to solving the
equations implicitly is given by exponential fitting, 97 " 10° which is illustrated
for a generalized, tunable, single-step implicit procedure.
(40)
where Uf is a degree-of-implicitness factor. The parameter U,- is obtained

by exponentially fitting Eq. (40), i.e., equating Eqs. (37) and (40), which
gives
Uf = HZfin + [1/(1 - eZift»)], i = !,...,# (41)
Equations (40) and (41) constitute an exponentially fitted trapezoidal
method. The method is used in the code CREK1D, 40 - 421()1 - 102 which has
been designed for adiabatic, constant-pressure combustion reactions. The
implicit equation (40) is solved iteratively, but with different iteration tech-
niques in the various regimes of the combustion problem. During induction
and early heat release when small step lengths are required for solution
accuracy, Jacobi-Newton iteration 80 is used to minimize computational
work. During late heat release and equilibration, when larger step lengths
can be used, Newton-Raphson iteration is preferred, since it has better
convergence properties. The selection of the iteration technique is auto-
mated and is based on species production and destruction rates. However
iterated, the test for convergence is based on the difference in successive
iterates. After successful convergence, the step size to be attempted on
the next step is obtained from an estimate of the local truncation error.
Comparison of Methods
Among the methods examined, BDFs are multistep formulas, whereas
the rest are single-step methods. Since information is available only at one
point, t = f () , at the outset of the integration, multistep methods are not
self-starting. These methods have to be "primed" by values generated with
a single-step algorithm. The extrapolation method examined here expe-
riences starting difficulty at each basic step because of the two-step pro-
cedure used on the internal steps. Alternatively, as in the BDF methods,
the difficulty at the initial point is resolved by starting with a single-step,
first-order method. As the integration proceeds, the solutions that the code
generates provide necessary values for the multistep method. In either
case, it can be expensive to start multistep methods. Since they usually
begin the integration with very small step lengths, frequent restarting—
e.g., over discontinuities, or in coupled reactive flow calculations—can
make BDF's very expensive. In addition, for methods based on fixed step-
size formulas such as GEAR, 78 changing the step size too frequently can
lead to numerical instability. Because multistep methods have to store
information from several previous steps, they have high storage require-
ments. For the same reason step length changing requires much compu-
tation. Finally, the stability properties of multistep methods deteriorate as
the order is increased. The implicit R-K methods, on the other hand, do
not suffer from this difficulty and are able to achieve high order and good
stability properties simultaneously. 37 For example, BDF's of high order
have poor stability characteristics when the eigenvalues of the Jacobian
matrix are close to the imaginary axis; i.e., the solution is oscillatory. The
implicit R-K methods, on the other hand, are stable for such problems.
Favoring multistep methods is their ability to vary the method order
easily and dynamically, which can be very important in solving stiff ODE's.26 37
Implicit R-K methods, on the other hand, are usually of fixed order. 28 The
K. RADHAKRISHNAN 99
local error can be estimated and controlled inexpensively. Compared with

the implicit R-K methods, much smaller systems of nonlinear algebraic
equations have to be solved at each step.
Single-step methods have the great advantage of being self-starting. Be-
cause they have no memory of previous values (other than the immediate
one) the step size can be changed frequently. For the same reason, storage
and computational overhead are less with these methods. Also, changes
in step length and, in the event of a failed step, retrieval of original in-
formation can both be made with much less complication than for multistep
methods. Implicit R-K methods require more storage than the other single-
step methods because of the large systems of nonlinear equations [Eq.
(24)] that must be solved at each step. However, not storing the previous
values can be wasteful because they can be used in developing the solution
further, as done in multistep methods. The result of this extrapolation
provides an accurate predicted value. Consequently, the number of cor-
rector iterations and, hence, the number of derivative evaluations [see Eq.
(21)], is usually less than two.34'37 A significant feature of multistep methods
is that the number of derivative evaluations per step is essentially inde-
pendent of the order used.28 In contrast, for the R-K and extrapolation
methods, the number of derivative evaluations per step increases with the
order of the approximation. Consequently, when the derivative functions
are expensive to calculate, multistep methods are superior. Another im-
portant advantage of generating the predicted value is that the local error
can be reliably and inexpensively estimated. For single-step methods, the
local error is estimated either by performing the computation over each
interval with two different step lengths or, more commonly, by embedding
a lower-order method (however, see Ref. 56).
An important difference between the selected asymptotic integration
method and the other methods concerns the Jacobian matrix, /. SAIM is
the only method that does not require the evaluation of / and inversion
of the matrix G [Eq. (22)]. For a system of N ODE's, / is of size N x N
so that N2 terms have to be evaluated. In practice, matrix inversion is
never attempted because of prohibitive costs,23 and modern methods such
as Lower-Upper (LU) decomposition103 are used. In this method, G is
factored into the product of a lower triangular matrix L and an upper
triangular matrix U. The solution of the linear system of equations, such
as Eq. (21) for the BDF, then reduces to the fairly simple solution of two
triangular linear systems in succession. The LU method requires on the
order of N3/3 + N2 operations. Thus, SAIM requires less work per step
than the other methods. However, when larger steps are possible with the
other methods, they can be more efficient for the complete problem.
Even among methods that require /, there is a large difference in the
work related to matrix algebra. BDF methods reduce the work associated
with computing / and LU-decomposing G by using an old G for as long
as possible. Because G involves the step length, hn, even modest changes
in hn may require the re-evaluation of G and its LU decomposition, even
though J is changed only slightly. Hence, codes such as EPISODE,79 which
is based on variable-step formulas, may require more frequent evaluations
of G and LU decompositions than codes such as GEAR78 and LSODE,38-39
which are based on fixed-step formulas. In contrast to the BDF methods,

the ROW methods [Eq. (27)] require evaluation of G and its LU decom-
position at every step. The semi-implicit Rosenbrock methods [Eq. (26)]
may require several matrix evaluations and LU decompositions per step.
For implicit R-K methods, the size of the nonlinear system [Eq. (24)] to
be solved is rN. Therefore, the size of the matrix to be computed and LU-
decomposed at each step can be very large, with attendant increases in
storage requirements and computational cost. The implicit R-K methods
must be able to use much larger step lengths than the BDF methods to be
more efficient. However, if the solution changes character from a smooth
region to rapidly varying transients, such as during the onset of heat release
(Fig. 1), the ability to use larger step lengths is wasteful. 34
Similar remarks apply to extrapolation methods, whose efficiency derives
from their ability to use large basic step lengths. The matrices used in such
methods involve the internal step length h [Eqs. (28-30)]. Hence, even if
the same Jacobian can be used for the whole basic step, several matrix
evaluations and LU decompositions are required because the method uses
successively smaller internal step lengths. Unless large basic step lengths
are possible, the method can become very expensive. However, very high
accuracy can be obtained with these methods.
The code CREKID40-42-101-102 attempts to minimize the cost associated
with matrix algebra by using Newton-Raphson iteration only when large
step lengths are possible, such as during equilibration. During induction
and early heat release, when the rapidly varying solution requires small
step lengths, Jacobi-Newton iteration is used to minimize computational
work.
Another method to reduce the cost associated with matrix algebra ex-
ploits any special structure of the problem. For example, for constant-
density isothermal kinetics problems, the Jacobian matrix has many zero
elements because the species derivatives [Eq. (2)] involve only a few other
species. Such matrices are called sparse because they have few nonzero
elements. Clearly, they require less storage if only the nonzero elements
are stored. Methods have been devised for storing only the nonzero ele-
ments of sparse matrices and performing the LU decomposition in such a
way that few nonzero elements are introduced. Such special-purpose de-
composition requires extra effort and storage, but significant savings can
be realized.34 The code LSODES39 has been designed specifically for prob-
lems with sparse Jacobian matrices. The chemical kinetics code
LARKIN110"112 uses sparse matrix routines.
A final point to be considered is the specification of output stations, i.e.,
the time values at which the solution is required. 28 34 The methods examined
here automatically adjust the step length hn for optimal efficiency and
generate solutions at discrete points in time, tn. If the output stations
coincide with these times, the efficiency of the algorithm is not affected.
However, requesting the solution at values other than the naturally oc-
curring tn can significantly affect the computational work. For BDF meth-
ods, the cost is minimal because they obtain results at intermediate points
by interpolation. In contrast, the other methods must obtain the results by
integration. Hence, if output is required at many intermediate points, the
K. RADHAKRISHNAN 101
computational cost can be high, especially for the implicit R-K and ex-
trapolation methods, which require considerable computational work per
step.
Published Numerical Comparisons

Systematic testing and comparison of numerical methods assist in se-
lecting the integration method for a particular problem. However, it is
important that such studies be performed, if not on that particular problem,
at least on a similar class of problems. Evaluation and comparison of
methods are difficult because they can involve many factors, not all of
which have equal importance for all users, and not all of which can be
quantified. 22 However, the main factors that can be assessed and usually
reported are reliability, accuracy, and efficiency. The efficiency comparison
generally is made by measuring the CPU time required by each technique
to solve the problem of interest. Occasionally, memory requirements must
be included in the comparisons.
We now present some published numerical comparisons of integration
methods and computer codes for solving the ODE's arising in combustion
chemistry. Although we focus attention on methods and codes used for
nonisothermal combustion kinetics calculations, we briefly summarize pub-
lished results for isothermal problems.
Numerical methods for constant-temperature chemical kinetics problems
have been compared by various investigators.54-59 6()-63-83-85-87-88-"-1()4-1()5 Careful
examination of these studies leads to the following three observations for
isothermal problems:
1) Implicit Runge-Kutta methods presently are not suitable for chem-
istry applications.
2) The backward differentiation formulas as implemented by Hindmarsh
are better than, or competitive with, the other methods and continue to
be the benchmark against which new methods and codes should be com-
pared.
3) The use of sparse matrix techniques can result in significant savings
of computer time.
More recently, the accuracy and efficiency of currently available tech-
niques for nonisothermal combustion kinetics were examined by applica-
tion to two problems.4-29'66"68 The codes studied in these comparisons were
EPISODE,79 LSODE,38 39 CHEMEQ94 95 CREK1D,101 and GCK84,5 which
uses a modified version of GEAR.78 The test problems describe adiabatic
combustion reactions at constant pressure, and are summarized in Table
1. Both cases included all three combustion regimes: induction, heat re-
lease, and equilibration.
Figures 3 and 4 present the variation of the CPU time required by the
codes with the local error bound, EPS, for test problems 1 and 2, respec-
tively. All results were generated on the NASA Lewis Research Center's
IBM 370/3033 computer using single-precision accuracy, with the exception
of GCKP84, which uses double-precision arithmetic. The suffix A or B
attached to some methods indicates the temperature method used (see Sec.
I). No suffix is attached to CREK1D and GCKP84 because the two codes
have built-in procedures for computing the temperature.
Table 1 Kinetics test problems

Problem Description
1 Ignition and subsequent combustion of a carbon monoxide-hydrogen-
air mixture at an initial temperature of 1000 K and a pressure of
10 atm. The reaction mechanism included 12 reactions among 11
species.
2 Ignition and subsequent combustion of a stoichiometric hydrogen-air
mixture at an initial temperature of 1500 K and a pressure of 2
atm. The reaction mechanism involved 30 reactions and 15 species.
a
Both problems described adiabatic, gas-phase combustion reactions at constant pressure.
For reaction mechanisms and rate constants, see Ref. 4.
The parameter EPS in Figs. 3 and 4 is related to the local error. However,
because the local error control is performed differently in each code, EPS
does not have the same meaning for all codes. In a typical combustion
kinetics problem, the solutions for the species mole numbers and temper-
ature vary widely (Fig. 1). Hence, pure relative error control is appro-
priate—EPS is then a measure of the number of accurate significant figures
in the numerical solution. However, many species have zero initial con-
centrations. For such variables, relative error is undefined, and the error
control used in GCKP84 and EPISODE is pure absolute—EPS is then a
measure of the largest number that may be neglected. Hence, the error
control performed by EPISODE and GCKP84 is semirelative—absolute
for variables with zero initial values, and relative otherwise.68 The error
control performed by LSODE was mixed relative/absolute. The code re-
quired values for the local relative error tolerance (EPS) and the local
absolute error tolerance (ATOL) for each variable. The codes CHEMEQ
and CREKID do not control the local truncation error. In the two codes,
the solution is accepted if it converges, and EPS is a local relative con-
vergence criterion.68
Very small values for EPS had to be used with EPISODE-A and EPI-
SODE-B to solve both test problems. Values larger than those shown in
Figs. 3 and 4 resulted in significant inaccuracies, 66 ~ 68 and for test problem
2, some runs resulted in increased CPU times.106 The run using GCKP84
with an EPS value of 10 ~ 2 for test problem 1 exhibited serious instability
and was terminated.
Two important conclusions can be made from Figs. 3 and 4. First, LSODE
is the fastest code for solving combustion kinetics rate equations. Second,
the difference in CPU times required by temperature methods A and B is
small for test problem 1, and for test problem 2 with large values of EPS.
But for small values of EPS, the differences in CPU times are marked,
with method A being significantly superior to method B.
Because EPS has different meanings for different codes, no conclusions
can be made about the relative accuracy of results obtained by two different
codes using the same value for EPS. In addition, because the codes deal
only with local errors, not much can be said about the actual or global
error (i.e., deviation of numerical solution from exact solution), which
accumulates in a complicated way from the local errors. Since the user
102
10'
10" n-6
10~^ 10"M 10" 10,-2
LOCAL ERROR BOUND, EPS
Fig. 3 Variation of CPU time with local error bound for kinetics test problem 1.
All runs on IBM 370/3033.
METHOD
——V—— GCKP84
——O—— CREK1D
— -O- - LSODE-A
——•—— LSODE-B
—-O-- EPISODE-A
——•—— EPISODE-B
——A-- CHEMEQ-A
CHEMEQ B
10 10 10
Fig. 4 Variation of CPU time with local error bound for kinetics test problem 2.
All runs on IBM 370/3033.
usually wants to control the global error, it must be measured. In the

absence of exact solutions, one way to estimate the global error associated
with a particular value of EPS is to compare the solution generated with
that value for EPS with the solution obtained with a reduced EPS. For
each code and test problem, the most accurate solution generated by that
code (i.e., with the smallest EPS used) was used as a basis for comparison
and is referred to as its standard solution. For solutions generated with
larger values of EPS, the global error et(t) in variable / at time t was
estimated by using the formula
e,(0 = [Y,(0/nsT«] - 1, I' = 1, ..., MS + 1 (42)
For each code, YiST(t) is its standard solution value at time t for variable
/, and Yf(i) is the solution for variable / at time t obtained by that code
using the value of EPS for which the global error is to be estimated.
Detailed plots of et(f) are presented in Ref. 68. The plots show that
CHEMEQ is the most accurate code during induction and early heat re-
lease; in the later regimes, however, the other codes are more accurate.
To provide a more comprehensive measure of the accuracy, a mean
integrated rms error, £rms, was calculated as follows:
*rms(0 df (43)
where erms(0 is the rms estimated global error at time t, and is given by
[ NS+l
2
I
ef(t)/(NS + 1)1
1 1/2
(44)
Equation (43) provides a single quantity that is a measure of the global

error incurred in solving the complete problem.
Figures 5 and 6 present the variation of £rms with the user-supplied local
error bound, EPS. The significant discrepancies between the values spec-
ified for EPS and the errors actually obtained illustrate the inadequacy of
local error measures and control used in current codes; clearly, some con-
trol of the global error is desirable.
Two conclusions are apparent from Figs. 5 and 6. First, LSODE is the
most accurate code examined here. Second, temperature method A is as
accurate as method B and, in many cases, is significantly more accurate.
For test problem 2 and EPS = 10~ 4 , LSODE-B is more accurate than
LSODE-A because it used a much smaller ATOL. When both codes use
the same value of ATOL, LSODE-A is more accurate. 6768
Figures 7 and 8 show the variation of the CPU time with the mean
integrated global error, £rms. These figures illustrate the large differences
in the computational work required by the different techniques to attain
comparable accuracy. For both test problems, LSODE is the most efficient
code in the sense that it requires the least CPU time to attain a specified
accuracy. Another important conclusion is that, for comparable accuracy,
the CPU times required by method A are less than, or compare favorably
with, those required by method B.
GCKP84
CREK1D
——Q.—— LSODE-A
——•——— LSODE-B
—-•A--- CHEMEQ-A
——A—— CHEMEQ-B
I i I iI I I I
10° 10~z 10"
Fig. 5 Variation of mean integrated global error with local error bound for kinetics
test problem 1.
III. Local Sensitivity Analysis Methods
Brute Force Method

The simplest method for estimating the effect of an uncertainty in any
parameter r|y is to solve the governing ODE's with two different values for
T]y—say, fjy and f|y + Air]y—where fjy is the generally accepted nominal
value for r\j. The change, Ay^/), in the solution at any time t is then simply
given by
+
Ay/(0 ~ yf^\j Ain/,0 — y/CÔ (45
where yXrjy + Air]y,f) and .y/(TJy,0 are the solutions at time t obtained with
the two different values of r\j. It is clear that Air]y can be as large as the
user desires, so that there is no restriction on the size of the change AT^.
However, the sensitivity coefficient 5/y is well approximated by Ay/Airiy for
only small Aiqy.
The technique is very simple to use because no additional programming
is required—the only changes made are to the input data set. The method
also has the advantage that several parameters can be varied simultaneously
by arbitrary magnitudes. However, to obtain the sensitivity coefficients 5/y-,
the parameters must be varied one at a time. The brute force method has
10° p-
GCKP84
CREK1D
- LSODE-A
——•—— LSODE-B
——0—— EPISODE-A
——•—— EPISODE-B
—A— CHEMEQ-A
——A—— CHEMEQ-B
10 10'3 10 10
Fig. 6 Variation of mean integrated global error with local error bound for kinetics
test problem 2.
10'
10U
10" 10'-3 10-2 10" 10U

MEAN INTEGRATED ERROR, Er,
Fig. 7 Variation of CPU time with mean integrated global error for kinetics test
problem 1. All runs on IBM 370/3033. For symbol key, see Fig. 5.
^•--^
10'
10"
i i I i 11
10" 10,-3 10i 2 10,-1 10°
MEAN INTEGRATED ERROR, E pMS
Fig. 8 Variation of CPU time with mean integrated global error for kinetics test
problem 2. All runs on IBM 370/3033. For symbol key, see Fig. 6.
been used successfully , 72J()7 ~ 109 but it can become tedious and expensive
when the number of parameters is large. For M parameters, the simulation
program must be run at least M + 1 times to get all first-order sensitivity
coefficients, and the choice of ATI, must be made carefully to ensure ac-
curacy of the estimated coefficients.70-110
Direct Method
The direct method (DM),111 also called the direct variational method, 112
computes the first-order sensitivity coefficients by simultaneously solving
the model equations [Eqs. (14)] and sensitivity equations [Eqs. (16)]. The
method requires the solution of 2N ODE's for one sensitivity parameter—
N ODE's for the combustion chemistry and TV ODE's for the sensitivity
coefficients. Because the original ODE system given by Eqs. (14) is stiff
and generally nonlinear, the combined problem is also stiff and nonlinear.
Hence, implicit methods are preferred for solution efficiency. However,
they require the evaluation and inversion of the iteration matrix for the
system, analogous to Eq. (22). However, the iteration matrix for the com-
bined system of Eqs. (14) and (16) results in a particularly simple form,
and its inversion requires only the inversion of G [Eq. (22)], the iteration
matrix for the original ODE system.111 Note that the Jacobian for the
sensitivity ODE's [Eqs. (16)] given by 3(dSj/dt')/dSj is simply /. Therefore,
the DM is more efficient than the brute force method, which will require
the formation and inversion of G approximately twice as many times be-
cause the model equations must be solved twice.
For more than one parameter, the combined system of ODE's can be
solved in two ways. For each parameter, the combined system of 2N ODE's
can be solved to generate sensitivity coefficients for that individual param-
1 08 COMBUSTION KINETICS AND SENSITIVITY ANALYSIS
eter, which will therefore require solving 2N ODE's M times to generate

sensitivity coefficients for all M parameters. Alternatively, sensitivity coef-
ficients for all M parameters can be solved in a single calculation, which
requires solving N + MN = (M + l)N ODE's. Although this is a much
larger system, the inversion of its iteration matrix still requires the inversion
of only G.111 The DM can also be used to compute higher-order sensitivity
coefficients.
An alternative to the direct approach is, first, to solve Eqs. (14) and
store the solution, Y(f). The stored information is used to approximate J
and df/dr\j by interpolation, and Eqs. (16) are solved for the sensitivity
coefficients.113 The decoupled method requires the solution of TV ODE's
once (for Y) followed by N ODE's M times (for the sensitivity coefficients).
Because the ODE's for the sensitivity coefficients [Eqs. (16)] are linear,
they are less expensive to solve than the nonlinear ODE's for the com-
bustion chemistry [Eqs. (14)]. Hence, the decoupled version of the DM is
also more efficient than the brute force method. The procedure can be
employed to compute the second-order derivatives as well, by storing val-
ues for the required first-order sensitivity coefficients.
The decoupling procedure can result in interpolation errors with at-
tendant accuracy and stability problems. However, it is significantly more
expensive to solve for 2N nonlinear ODE's M times than to solve the N
ODE's for the chemistry once, followed by the N linear ODE's for the
sensitivity coefficients M times.
Green's Function Methods

A procedure that has proven more efficient than the direct method for
M > N uses the associated Green's function to solve for the sensitivity
coefficients.73-114"116 Since Eq. (16a) is a linear differential equation, it can
be transformed into the linear integral equation
S/f)
J = K(t,t0)S;(t0) + dT K(t,T) - (T) (46)
Jti) dr\j
where K is the Green's function matrix of size N x N. The advantage of

the Green's function method (GFM) is that K is independent of the non-
homogeneous term df/di}^ therefore, there is only one Green's function
for the problem. The solution to all possible problems with different func-
tions df/dr\j is simply given by Eq. (46).
The Green's function K(t,t) is obtained by solving the following ODE's
for the adjoint Green's function
- Kt(t,t) + Kt(Tj)J(t) = 0, 0 < T< t (47 a)
Kt(-r,t) = /, T= t (47b)
Equations (47) are solved backward in T, and /Ct(i,/) is obtained at
various values of T for a fixed value of t. K(t^) is given simply by
r) (48)
The GFM can be summarized as follows. Let r 1? t2, ... be the times at
which the sensitivity coefficients are desired. The TV ODE's for the com-
bustion chemistry are solved up to the final time with a standard solver
for stiff ODE's, such as the GEAR package discussed in the previous
section, and the solution Y(t) is saved for later use to approximate /. The
generally stiff system of N2 ODE's given by Eqs. (47) is solved backward
from t = tl to t = tQ by using the same solver, and the solution /Ct(j,r) is
saved at some, or all, of the integration points. The saved values are then
used to evaluate numerically the integral in Eq. (46), usually through simple
quadrature schemes.115-116 If r\j is an initial value, the integration is un-
necessary because df/dr\j = 0.
The GFM requires the solution of N ODE's for the combustion chemistry
followed by the TV2 ODE's given by Eqs. (47). Hence, the total number
of ODE's required by the GFM is equal to N2 + N = N(N + 1). This
number is considerably less than the N(M + 1) ODE's required by the
DM, for TV < M, which is generally true in combustion kinetics. Of course,
the integrals in Eq. (46) must be performed as well to evaluate the sen-
sitivity coefficients with respect to the rate constant parameters.
To obtain sensitivity coefficients at times greater than ^, the following
simple recursion formula is used:
S,{tk) = K&.t^JSjfr^) + _ ( dT K(tk,i) (T) (49)
where tk is the current time at which S; is required, tk_1 is the previous

time at which Sy- was computed, and K(tk,rr) in the interval [^_i,/*] is
obtained by solving Eqs. (47) backward in time from tk to tk_l.
To reduce the cost associated with the solution of the TV2 ODE's [Eqs.
(47)], Kramer et al.118 replaced the implicit method algorithm with the
piecewise Magnus method119 algorithm suggested by Chan et al.120 In the
new procedure, the total integration interval is divided into smaller steps
of size Ar, and in each interval the Green's function kernel K(t + Ar,f) is
approximated by118
p + Ar
K(t + Ar,r) = exp I J(s)ds (50)
where J(s) is the Jacobian matrix df/dy at time s. The step size Af is selected
to ensure that the local error in K, normalized by an appropriate measure,
is less than a user-specified tolerance.
A second improvement incorporated into the technique, called the an-
alytically integrated Magnus (AIM) modification of the Green's function
method, or GFM/AIM, concerns the evaluation of the sensitivity integrals
[Eq. (49)]. The simple numerical quadrature schemes used in the GFM
were replaced with analytical expressions, obtained with an assumed ap-
proximation for df/dr\j in the interval [t,t + Ar]. However, the procedure
requires the specification of another local tolerance to test the adequacy
of the approximation for df/d^.
The GFM/AIM was implemented by Kramer et al.121 into the code AIM,
which was combined with a modified LSODE algorithm. The combined
package, called CHEMSEN,122 can solve the chemical kinetics ODE's and
first-order sensitivities with respect to both initial conditions and rate con-
stants. It is, however, restricted to constant-density isothermal problems.
Modified GFM methods have also been developed by Hwang123-124 and
Hwang and Chang.125 Their methods use a different calculation procedure
for K(I,T) than the use of Eqs. (47) and (48).
Decoupled Direct Method

In the decoupled direct method (DDM), the ODE's for the sensitivity
coefficients are solved separately from, but sequentially with, those de-
scribing the combustion chemistry. The same algorithm that solves the
ODE's for the chemistry also solves for the sensitivity coefficients. The
rationale for using the same solver is that the two systems of ODE's have
the same Jacobian. The decoupling procedure has been suggested by a
number of researchers (see, for example, Refs. 126-129) and is illustrated
below by using the BDF's as implemented in LSODE.38 39
If BDF's are used to solve for the sensitivity coefficients, the resulting
formula, corresponding to Eq. (20), is
In Eq. (51), S,(O is the sensitivity coefficient at the current time tn\ Sy (*„_<-)
are the Sy at previous times tn_^ Sj(tn) = dSy/df(f,,); and the other terms
have been defined in the previous section. Writing out the expression for
Sj(tn) and collecting terms gives
[/ - hn^J(tn}} Sj(tn) = 2 c^S/f,,.,) + A«fc,o -f (tn) (52)

1= 1 or\j
Equations (21) and (52) show the similarity between the model and the
sensitivity equations. The DDM exploits this similarity by alternating the
solution of Eq. (52) with that of Eq. (20). The solution Y for the model
problem is advanced from time tn_1 to the current time, tn. The solution
Yn at the new time is used in Eq. (52) to advance Sj from time tn_ { to time
tn. The solution procedure for 5; does not require either a predictor or an
iterative process because the ODE is linear. The process of advancing Y
and then 5; by the same time step is repeated until the end of the integration
interval.
During the course of solving the model problem, modern integration
packages automatically generate values for qn, qn +l, ... and for hn, hn +l,
... (and, hence, for /„, tn +l, ...). In solving Eq. (52) for the sensitivity
coefficients, the DDM uses exactly the same time sequence tn, tn +l, ... and
the same order sequence qn, qn +1 , . . . . Thus, the error control in the solution
of Eq. (52) is determined by the error control in the solution of Eq. (20).11()
Hence, the DDM needs no additional tolerances beyond those required
for the model system, provided J(t) is updated before the sensitivities are
computed.
Although the Gear method has been used to illustrate a decoupled direct
method, the DDM can be used with any other ODE integration technique.
However, because of its efficiency for combustion kinetics calculations,
the Gear method, or variants of it, have been the only methods employed
for sensitivity analysis in chemical kinetics. 6 - 7K11(U30 ~ 136
The DDM, together with a modified version of LSODE, has been im-
plemented in an efficient computer program, CHEMDDM, by Dunker. 130
However, it is restricted to constant-density, constant-temperature prob-
lems. CHEMDDM was modified extensively by Radhakrishnan and Bittker133
(see also Ref. 71), who developed the code GCKP86 for sensitivity analysis
of nonisothermal combustion kinetic rate equations. GCKP86 has since
undergone several revisions to expand its options, and the current version,
named LSENS,6 can be used for a wide variety of general kinetics and
sensitivity calculations. The DDM was subsequently used by Leis and
Kramer134-135 in their general-purpose code ODESSA and by Lutz et al.136
in their chemical kinetics code SENKIN, which also works for noniso-
thermal problems.
Other Local Methods

A method for obtaining sensitivities of objective functions is presented
by Cacuci137-138 and Seigneur et al.139 The functional can involve any num-
ber of dependent variables. The objective function can be a single varia-
ble—such as a species concentration or temperature—or it may involve
several variables—such as a class of species such as unburnt hydrocarbons
or pollutant species. The rationale for the variational method is to reduce
the amount of sensitivity information needed and generated. The calcu-
lation procedure requires the solution of adjoint equations and is similar
to the GFM. Koda et al.140 give a detailed discussion of the variational
approach.
Another method, based on polynomial approximations, has been pro-
posed by Hwang.141 The essence of the polynomial approximation method
is to subdivide the time domain of interest into subintervals and represent
the system temporal behavior in each subinterval by low-degree poly-
nomials. The numerical procedure is facilitated by the use of Lagrange
polynomials.
Comparison of Local Methods

The simplest local method to use is the brute force method. Each pa-
rameter is varied in turn by an arbitrary amount. If sensitivity calculations
with respect to M parameters are needed, the method would require at
least M + 1 solutions of the system ODE's and can, therefore, become
very expensive. The brute force method does have the benefit of requiring
little or no additional programming.
The direct method also requires very little additional programming. De-
pending on how the method is applied, however, it requires the solution
of either 2N ODE's M times or MTV ODE's once. Dickinson and Gelinas111
have used the DM to study atmospheric ozone kinetics but do not present
a timing comparison with the brute force method. The DM, however, has
been found to be quite inefficient and unstable, or has failed completely,
on several stiff problems.115-118 The decoupled approach of implementing

the DM is, in theory, more efficient than the coupled DM. However, the
decoupled solution procedure has also proven unstable on some stiff prob-
lems.118 Furthermore, surprisingly, it has shown no efficiency gains over
the coupled procedure on some problems.115-118
The GFM is more efficient than the DM for N < M because it requires
the solution of fewer ODE's. Most of the computer work required by the
GFM is associated with calculating the Green's function matrix. Once the
Green's function matrix is computed, calculating the rate constant sensi-
tivities requires little additional work. Each sensitivity coefficient vector
requires two matrix-vector multiplications and N integrations [Eq. (49)].
The GFM has been shown to be more efficient than the DM115; however,
it still can be very expensive to use.118-123 The computational work require-
ment depends on the number of output stations. In addition, the GFM
can produce significant inaccuracies when the solution changes rapidly. 118
To track the solution in these regions more accurately, the GFM requires
so many time steps that it loses its computational advantage over the DM.
The GFM/AIM has been shown by Kramer et al.118 to be significantly
faster than both the GFM and the DM. The GFM/AIM does, however,
share with the GFM the disadvantage of the computational work being a
function of the number of times the sensitivity information is required.
Dunker110 showed that a poor selection of the output times produces sub-
stantial increases in computational effort and also results in significant
inaccuracies. The GFM/AIM also requires the specification of error tol-
erances in addition to those required by the ODE solver. Kramer et al.75
showed that small values for the error tolerances can result in much larger
computational costs than those of the DM.
The DDM alternates the solution for the model problem and sensitivity
coefficients. Much of the work demands of the DDM is associated with
calculating the first sensitivity coefficient because it requires forming and
decomposing the matrix [/ - hn$nQJ]. Since all sensitivity coefficients need
the same matrix, calculating the second and subsequent sensitivity coef-
ficients takes much less work.
Because the DDM solves the model and sensitivity equations in tandem,
the calculated solutions to the model equations do not have to be stored
for the entire integration interval, which consists of many steps. The DDM
also avoids the added expense and inherent errors associated with the later
approximation of the Jacobian matrix from the stored solution, as done in
the Green's function methods. The GFM/AIM typically requires fewer
steps than the DDM.110 Nevertheless, for applications to date, the DDM
has been shown to be more efficient than the Green's function meth-
ods.6-71'110
Test Problems
The accuracy and efficiency of the techniques and codes previously dis-
cussed can be evaluated by application to test problems from combustion
kinetics. In particular, we have applied the codes LSENS and CHEMDDM
to an isothermal problem that has been studied extensively. Results ob-
tained with other techniques and codes and the required execution times
are therefore available, permitting a comparison among techniques and

codes. A second test problem is used to demonstrate the capability of the
decoupled direct method to perform sensitivity analysis of nonisothermal
combustion kinetics. All calculations with LSENS and CHEMDDM were
performed in double precision on the NASA Lewis Research Center's IBM
370/3033 computer.
Pyrolysis of Ethane
Test problem 1, taken from Ref. 110, describes the constant-density,
constant-temperature pyrolysis of ethane at a temperature of 923 K. The
reaction mechanism, rate constants, and initial conditions are given in
Table 2. Because this is a constant-temperature problem, the rate constants
kj [Eq. (5)] are time invariant. Hence, sensitivity coefficients need to be
computed only with respect to kj.
The reaction mechanism is quite small (Table 2), but the problem is very
stiff, and the DM with Eqs. (14) and (16) decoupled produced very in-
accurate results, as did the GFM.118 The coupled DM, however, produced
accurate solutions; but not all sensitivity coefficients could be obtained.
The codes LSENS and CHEMDDM were applied to test problem 1,
and sensitivity coefficients with respect to all rate constants and all initial
conditions were computed successfully. Normalized sensitivity coefficients
with respect to rate constant kl (d/^Yj/d/^k^ calculated at two different
times, t — 1 and 20 s, are given in Tables 3 and 4, respectively. Also
presented in these tables are solutions obtained by Dunker, 110 using the
DDM, the GFM/ AIM, and, to assess their accuracy, a brute force method,
by means of the finite-difference approximation
dk, 28*!
with 8 - 0.05.
The results obtained with the coupled DM, GFM, and GFM/ AIM by
Kramer et al.118 were normalized by the kinetic solutions generated by
LSENS and are given in Table 3 for t = 1 s.
Table 2 Reaction mechanism for sensitivity test

problem 1
Rate
Reaction8 constant,17 /c;
1) C2H6 -* CH3 + CH3 1.14 x 10-2
2) CH3 -1- C2H6 -» CH4 + C2H<; 1.19 X 106
3) C2H5 -> C2H4 + H 1.57 x 103

4) H + C2H6 -> H2 + C2H, 9.72 x 108
5) H + H -> H, 6.99 x 1013
a
See Ref. 110. The initial concentration of C2H6 is 5.951 x
10~ 6 mole-cm3; all other initial concentrations are zero.
b
Units are mole, centimeters and seconds, and the tem-
perature is 923 K.
o
o
^
03
C
Table 3 Normalized sensitivity coefficients with respect to rate constant kl for sensitivity test problem 1 at time = 1 s O
d/.Yi/d/,* ! at 1 s
m
Brute force
Species LSENS CHEMDDM DDM a
GFM/AIM a
methoda DM h
GFM b
GFM/AIM b
o
O)
C2H6 -0.044 -0.044 -0.044 -0.044 -0.044 -0.044 -0.044 -0.045
CH3 1.000 1.000 1.000 0.999 1.000 1.000 1.000 0.988
CH4 0.976 0.976 0.976 0.977 0.978 0.976 0.976 0.969 0)
C2H5 0.662 0.662 0.662 0.688 0.663 0.662 0.659 0.665 m
C2H4 0.680 0.680 0.681 0.682 0.681 __ c —c —c
H2 0.602 0.602 0.602 0.604 0.602 —c __ c —c
H 0.478 <
0.478 0.478 0.493 0.480 0.478 0.620 0.478 H
a b c
From Ref. 110. From Ref. 118. Not given.
V)
O)
Table 4 Normalized sensitivity coefficients with respect to rate constant kl for

sensitivity test problem 1 at time = 20 s
at 20s
Brute force
Species LSENS CHEMDDM DDMa GFM/AIMa method a
C2H6 -0.819 -0.819 -0.820 -0.824 -0.823
CH3 1.000 1.000 1.000 0.692 0.999
CH4 0.644 0.644 0.643 0.652 0.646
C2H5 -0.210 -0.210 -0.210 -0.266 -0.211
C2H4 0.324 0.324 0.323 0.325 0.323
H2 0.221 0.221 0.221 0.220 0.220
H 0.091 0.091 0.090 0.080 0.091
a
From Ref. 110.
Tables 3 and 4 show the excellent agreement between the results gen-
erated with LSENS and CHEMDDM. In addition, both codes agree well
with the DDM and brute force method results of Dunker. 110 In contrast,
the sensitivity coefficient for the hydrogen atom (H) generated with the
GEM is significantly inaccurate (Table 3). The DM and GFM/AIM are
both in agreement with the brute force method results at t = I s . However,
at t = 20 s, the GFM/AIM results are inaccurate, especially for CH3. Even
by decreasing the error tolerances required by GFM/AIM, Dunker 110 was
unable to obtain better agreement for sensitivities with respect to k^ In
fact, the error tolerance reduction significantly worsened the agreement
for the other rate constant sensitivities. However, when the number of
output stations was increased, GFM/AIM was more accurate, but the DDM
or brute force method results were not affected. The dependence of the
GFM/AIM results on the number of output stations at which sensitivities
are required and the lack of accuracy improvement with error tolerance
reduction have been attributed to difficulties in choosing step sizes and in
evaluating some integrals.110
A comparison of the computational work required by the codes ESENS
and CHEMDDM to solve this problem showed LSENS to be significantly
more efficient. For two output stations at 1 and 20 s, LSENS required
approximately 0.6 s to solve for the composition and sensitivity coefficients
with respect to all seven initial conditions and all five rate constants. In
contrast, CHEMDDM required approximately 1.4 s. For four output sta-
tions between 0 and 20 s, Dunker110 obtained execution times of 1.25 and
1.55 s, respectively, for the DDM and GFM/AIM on an IBM 370 computer.
Kramer et al.118 report execution times of 1.8, 12.5, and 79 s, respectively,
for the GFM/AIM, DM, and GFM.
To further explore the efficiency differences among the different tech-
niques, after Dunker,110 execution times were measured for computing 1)
the concentrations alone, Tconc; 2) the concentrations and all initial con-
dition sensitivities, TIC; 3) the concentrations and all rate constant sensi-
tivities, TRC; and 4) the concentrations, all initial condition sensitivities,
and all rate constant sensitivities, T ALL .
These execution times are given in Table 5 for LSENS and CHEMDDM,
together with the execution times obtained by Dunker110 for DDM and
GFM/AIM. All execution times are much smaller for the decoupled direct
method and, in particular, for the code LSENS. One additional important
difference between LSENS and CHEMDDM deserves mention. LSENS
can be used to generate any number of sensitivity coefficients, from just
one initial condition or one rate constant parameter to the full set of all
initial conditions and all rate constant parameters. CHEMDDM does not
have this feature, and all initial conditions and/or all rate constant sensi-
tivities are computed. This difference can result in significantly smaller
execution times for LSENS, if not all sensitivity coefficients are required.
For example, for test problem 1, LSENS required only about 0.26 s to
generate one initial condition sensitivity, instead of the 0.44 s required for
all initial condition sensitivities. For one rate constant sensitivity, the CPU
time was about 0.27 s. Comparing these execution times for one sensitivity
with the times given in Table 5 for all sensitivities shows that LSENS
requires less than 0.04 s for each additional sensitivity. The brute force
method requires at least one additional calculation of the concentrations
for each sensitivity coefficient. Table 5 shows the cost to be about 0.13 s,
which is significantly more expensive. The brute force method would re-
quire approximately 1 s for all initial condition sensitivities, and approxi-
mately 0.8 s for all rate constant sensitivities.
As discussed previously, much of the work required by the DDM is
associated with calculating the first sensitivity coefficient; subsequent sen-
sitivity coefficients require much less work. Similarly, calculating the Green's
function matrix [i.e., the initial condition sensitivities; see Eq. (46)] takes
up much of the computational work required by the GFM and the GFM/
AIM. Calculating the rate constant sensitivities requires much less work.
A useful measure for comparing efficiency is, therefore, the difference
T
ALL ~ Tio tne additional time needed to calculate the rate constant
sensitivities.110 Examination of Table 5 shows that, by this measure also,
the DDM is the more efficient method and LSENS is the most efficient
Table 5 Execution times needed for sensitivity test problem 1

Execution time, s
Initial
condition
and
Initial Rate rate
Method/ Concentrations condition constant constant
code alone sensitivitiesa sensitivities3 sensitivities11
LSENS5 0.13 0.44 0.41 0.62
CHEMDDMb 0.26 1.01 1.02 1.38
DDMC 0.21 0.90 __ d
1.25
GFM/AIMC 0.21 1.10 __ d
1.55
a
lncludes execution time for concentrations. b Execution'times on an IBM 370/3033
computer. cFrom Ref. 110; execution times on an IBM 370 computer. d Not given.
code. In addition, if only rate constant sensitivities are required, the DDM
is even more efficient than the GFM and GFM/AIM, because the DDM
does not then have to calculate the initial condition sensitivities, and its
execution time is only T RC . The GFM and GFM/AIM, however, still require
approximately TALL because the Green's function matrix must be computed
before the rate constant sensitivities can be evaluated.
Nonisothermal Reaction
A novel feature of the LSENS code is its capability to perform sensitivity
analysis of nonisothermal combustion kinetics. To illustrate this capability,
we consider a simple problem for which an analytical solution is known,
permitting evaluation of the decoupled direct method for nonisothermal
problems. Test problem 2 involves a first-order irreversible reaction
k
A-> B
k = AT" exp[-E/RT], A = 1, n = 1, E = 0
CTA(V = Q) = 1, or B (r = 0) = 0, T(t = 0) = 1000 K
To solve the problem analytically, the following simplifying assumptions
were made: 1) constant pressure, adiabatic reaction, and 2) constant and
equal specific heats cp for species A and B. The solution is
aA(0 = C<rA(0)e-AC'/[C - Xa A (0)(l - e^°)]
<rB(0 = MO) + <rB(0) - <JA(0

7X0 = 7X0) + ^ K(0) - aA(0]
where
C = 7X0) + KaA(0)
X = Gc/{c>A(0) + <rB(0)]}
where Qc is the heat of combustion, which dictates the temperature rise
due to the reaction.
The analytical and computed normalized sensitivity coefficients with re-
spect to the rate constant parameter A are given in Table 6. The solution
was obtained with values of Qc = 5000 cal/mole and cp = 5 cal/mole/K,
which together result in a 1000-K temperature rise. Also presented in Table
6 is the mixture temperature, which indicates the extent of the reaction.
The agreement between the analytical and computed results is excellent
at all levels of reactedness, illustrating the accuracy of the decoupled direct
method for sensitivity analysis of nonisothermal combustion kinetics rate
equations. The execution time required to solve for the dependent variables
(concentrations, temperature, and density) was approximately 0.09 s. The
execution time for the dependent variables and sensitivities with respect
to all four initial conditions and all three rate constant parameters was
approximately 0.3 s. The brute force method would have required at least
o
o
00
C
CO
o
Table 6 Sensitivity coefficients with respect to rate constant parameter A for sensitivity test problem 2 7s
t-pi
O
EXACT LSENS EXACT LSENS EXACT LSENS
io-56 1001 -1.001 x io-32 -1.001 x IO-32 1.000 1.000 9.990 x io-34 9.990 x io-34
io- 1010 -1.010 x io- -1.010 x 10~ 1.000 1.000 9.900 x io- 9.899 x io-2
io-34 1100 -0.110 -0.110 0.993 0.992 9.003 x io-2 8.978 x io- m
io-2 1761 -1.762 -1.755 0.551 0.550 0.238 0.238
io- 2000 -20.00 -20.09 8.244 x IO-8 8.150 x 10~8 4.122 x io-8 4.075 x io-8
0.7 s, which is significantly more. Although this is a simple problem, se-

lected because it permits comparison of the computed results with the exact
solution, LSENS has been demonstrated to be efficient on more realistic
problems.6 For example, a test problem describing the oxidation of ben-
zene-oxygen-argon mixtures involves 40 species among 115 reactions. This
constant-density problem required approximately 8.3 s to solve for the
composition and temperature. The execution time for the dependent var-
iables and sensitivities with respect to all 41 initial conditions and all 345
rate constant parameters was approximately 265 s, which may be compared
with the more than 3200 s that the brute force method would require. More
significantly, unless automated, the brute force method would need 387
separate runs of the simulation program, and 15,826 evaluations of sen-
sitivity coefficients per output station.
IV. Concluding Remarks

Significant progress has been made in the solution of stiff ordinary dif-
ferential equations (ODE's). Many methods and codes have been devel-
oped for stiff ODE's in general, and combustion kinetics equations in
particular. However, several areas need further development. Estimation
of the global error is an important problem because the user wants to know
and control it. The relationship between the global error and the local
error tolerance, EPS, is not clear. In addition, for a given value of EPS,
the global errors obtained with different codes can be significantly different,
reinforcing the need to control the global error. Alternatively, the code
could provide an estimate of the global error.
An issue that has not been addressed concerns use of variable local error
tolerances for combustion kinetics problems. During heat release, the so-
lution changes rapidly, so that small error tolerances are required to track
the solution accurately. 46 However, during equilibration, use of larger
error tolerances can decrease the computational work without incurring
severe error penalities. 4 Successful exploitation of this feature requires
reliable tests to identify the different combustion regimes. These tests
currently are unavailable, although some preliminary work has been
done. 40 ~ 42J01J02 Another method that has been suggested for decreasing
the computational work in the equilibration regime is the use of an ex-
ponential function that is forced to approach the chemical equilibrium
state.40'142 The method requires a priori calculation of the equilibrium state
and has not yet been evaluated.
Because combustion kinetics problems are not stiff in all regimes, com-
bining nonstiff and stiff integration methods may prove beneficial. By
automatically applying the more efficient method in each region, significant
reductions in computational work should be possible. The code LSODA37143
automatically switches between the backward differentiation formula and
the Adams method. LSODA has not been tested systematically on com-
bustion kinetics problems. However, the use of stiff and nonstiff methods
in different regimes has been explored.81 Switching methods was faster,
but lack of reliable regime identification tests prevented automatic switch-
ing. It was also unclear what iteration technique to use in the heat-release
regime.
The preceding techniques use different integration methods in different

regions; however, in each region, all equations are classified as either stiff
or nonstiff. Hence, in each region, the same integration method with the
same order of accuracy is used for all variables. This approach differs from
the procedures used in CHEMEQ and CREK1D. At the beginning of each
step, CHEMEQ partitions the system of ODE's into stiff and nonstiff
subsystems. It uses different integration methods for the different subsys-
tems. CREK1D uses essentially the same method for all variables, but the
order of the approximation is component dependent142; in addition, dif-
ferent iteration techniques are used in different combustion regimes. The
different approaches need to be evaluated carefully, especially because
CREK1D compares favorably with LSODE, despite using inefficient local
error and step length control strategies. Reliable tests to identify different
combustion regimes also need to be developed.
The use of sparse matrix techniques for combustion kinetic rate equations
should be evaluated. The equations do not present a problem if the tem-
perature and density are computed using algebraic equations. However,
if ODE's are used to solve for the temperature and density, the sparseness
of the matrix is reduced significantly, because many nonzero elements are
introduced. An alternative treatment preserving sparseness has been
suggested89 but has not yet been evaluated.
Many local sensitivity analysis methods that can be used to assess quan-
titatively the effects of uncertainties in rate constant parameters and errors
in initial conditions have been developed. Systematic sensitivity analysis
can be accomplished much more efficiently than with the brute force method
of varying each parameter in turn. The methods presented substantially
reduce the computational work and require much less effort by the user.
The decoupled direct method (DDM) was found to be the most efficient
and accurate local sensitivity analysis technique. More significantly, the
DDM is efficient and accurate for nonisothermal combustion kinetics prob-
lems. One question remaining for the DDM concerns the update of the
Jacobian matrix at every step. The tradeoff involves accuracy and com-
putational efficiency. In addition, use of an old Jacobian may require
additional local error tolerances. However, during equilibration, when the
solution changes slowly, an old Jacobian may be adequate.
For further efficiency gains, the use of sparse matrix techniques should
be evaluated. The model and sensitivity equations have the same Jacobian.
Hence, if sparse techniques can be applied successfully to nonisothermal
problems, the savings in computing sensitivities should be significant. Fi-
nally, because the sensitivity equations are independent of one another,
the use of multiprocessors will realize substantial reductions in computer
turnaround time.
Acknowledgments
The author's work was supported by NASA Lewis Research Center
through Grants NAG3-147 and NAG3-294, an NRC-NASA Research As-
sociateship, and Contracts NAS3-24105 and NAS3-25266. The author would
like to thank Drs. J. P. Boris and E. S. Oran for their encouragement and
constructive criticisms, and Drs. D. A. Bittker, C. J. Marek, and D. T.
Pratt for helpful discussions. The encouragement and support of the man-
agement at Sverdrup and NASA Lewis, particularly J. M. Barton and B.
A. Miller and Drs. M-S. Liou, E. J. Mularz, and P.M. Sockol, is gratefully
acknowledged.
References
ârdiner, W. C., Jr. (ed.), Combustion Chemistry, Springer-Verlag, New York,
1984, pp. 1-19.
2
Basevich, V. Y., "Chemical Kinetics in the Combustion Processes: A Detailed
Kinetics Mechanism and its Implementation," Progress in Energy and Combustion
Science, Vol. 13, 1987, pp. 199-248.
3
Edelson, D., "Computer Simulation in Chemical Kinetics," Science, Vol. 214,
1981, pp. 981-986.
4
Radhakrishnan, K., "Comparison of Numerical Techniques for Integration of
Stiff Ordinary Differential Equations Arising in Combustion Chemistry," NASA
TP-2372, 1984.
5
Bittker, D. A. and Scullin, V. J., "GCKP84—General Chemical Kinetics Code
for Gas-Phase Flow and Batch Processes Including Heat Transfer Effects," NASA
TP-2320, 1984.
6
Radhakrishnan, K. and Bittker, D. A., "LSENS—An Efficient General Chem-
ical Kinetics and Sensitivity Analysis Code for Gas-Phase Reactions," NASA RP
(to be published).
7
Pratt, G. L., Gas Kinetics, Wiley, London or New York, 1969.
8
Gordon, S. and McBride, B. J., "A Computer Program for Complex Chemical
Equilibrium Compositions—Incident and Reflected Shocks and Chapman-Jouguet
Detonations," NASA SP-273, 1971.
9
Burcat, A., "Thermochemical Data for Combustion Calculations," Combustion
Chemistry, edited by W. C. Gardiner, Jr., Springer-Verlag, New York, 1984, pp.
455-504.
10
Oran, E. S. and Boris, J. P., Numerical Simulation of Reactive Flow, Elsevier,
New York, 1987.
n
Pratt, D. T., "New Computational Algorithm for Chemical Kinetics," Char-
acterization of High Temperature Vapors and Gases, Vol. II, edited by J. W. Hastie,
National Bureau of Standards, Washington, DC, SP 561, 1979, pp. 1265-1279.
12
Henrici, P., Discrete Variable Methods in Ordinary Differential Equations, Wiley,
New York, 1962.
13
Fox, L. (ed.), Numerical Solution of Ordinary and Partial Differential Equa-
tions, Addison-Wesley, Reading, MA, 1962.
14
Conte, S. D. and de Boor, C., Elementary Numerical Analysis, McGraw-Hill,
New York, 1965.
15
Ralston, A., A First Course in Numerical Analysis, McGraw-Hill, New York,
1965.
16
Acton, F. A., Numerical Methods That Work, Harper & Row, New York,
1970.
17
Gear, C. W., Numerical Initial Value Problems in Ordinary Differential Equa-
tions, Prentice-Hall, Englewood Cliffs, NJ, 1971.
18
Lapidus, L. and Seinfeld, J. H., Numerical Solution of Ordinary Differential
Equations, Academic, New York, 1971.
19
Lambert, J. D., Computational Methods in Ordinary Differential Equations,
Wiley, Chichester, UK, 1973.
20
Dahlquist, G. and Bjorck, A., Numerical Methods, Prentice-Hall, Englewood
Cliffs, NJ, 1974.
21
Shampine, L. F. and Gordon, M. K., Computer Solution of Ordinary Differ-
ential Equations: The Initial Value Problem, W. H. Freeman & Co., San Francisco,
1975.
22
Hall, G. and Watt, J. M. (eds.), Modern Numerical Methods for Ordinary
Differential Equations, Clarendon, Oxford, UK, 1976.
23
Forsythe, G. E., Malcolm, M. A., and Moler, C. B., Computer Methods for
Mathematical Computations, Prentice-Hall, Englewood Cliffs, NJ, 1977.
24
Stoer, J. and Bulirsch, R., Introduction to Numerical Analysis, Springer-Verlag,
New York, 1980.
25
Press, W. H., Flannery, B. P., Teukolsky, D. P., and Vetterling, W. T.,
Numerical Recipes, The Art of Scientific Computing, Cambridge University Press,
Cambridge, UK or New York, 1986.
26
Hull, T. E., Enright, W. H., Fellen, B. M., and Sedgwick, A. E., "Comparing
Numerical Methods for Ordinary Differential Equations,' 1 SI AM Journal on Nu-
merical Analysis, Vol. 9, 1972, pp. 603-637.
27
Shampine, L. F., Watts, H. A., and Davenport, S. M., "Solving Nonstiff
Ordinary Differential Equations—The State of the Art," SIAM Review, Vol. 18,
1976. pp. 376-411.
28
May, R. and Noye, J., "The Numerical Solution of Ordinary Differential Equa-
tions: Initial Value Problems," Computational Techniques for Differential Equa-
tions, edited by J. Noye, North-Holland, New York, 1984, pp. 1-94.
29
Radhakrishnan, K., "A Comparison of the Efficiency of Numerical Methods
for Integrating Chemical Kinetic Rate Equations," Computational Methods, edited
by K. L. Strange, Chemical Propulsion Information Agency, Johns Hopkins Applied
Physics Lab., Laurel, MD, Publ. 401, 1984, pp. 69-82; also, NASA TM-83590,
1984.
30
Curtiss, C. F. and Hirschfelder, J. O., "Integration of Stiff Equations," Pro-
ceedings of the National Academy of Sciences of the United States of America, Vol.
38, 1952, pp. 235-243.
31
Shampine, L. F., "What Is Stiffness?," Stiff Computation, edited by R. C.
Aiken, Oxford University Press, New York, 1985, pp. 1-16.
32
Gear, C. W., "The Automatic Integration of Stiff Ordinary Differential Equa-
tions," Information Processing, edited by A. J. H. Morrell, North-Holland,
Amsterdam, 1969, pp. 187-193.
33
Bui, T. D. and Bui, T. L., "Numerical Methods for Extremely Stiff Systems
of Ordinary Differential Equations," Applied Mathematical Modelling, Vol. 3,
1979, pp. 355-358.
34
Shampine, L. F. and Gear, C. W., "A User's View of Solving Stiff Ordinary
Differential Equations," SIAM Review, Vol. 21, 1979, pp. 1-17.
35
Gelinas, R. J., "Stiff Systems of Kinetic Equations—A Practitioner's View,"
Journal of Computational Physics, Vol. 9, 1972, pp. 222-236.
36
Shampine, L. F., "Stiffness and the Automatic Selection of ODE Codes,"
37
Byrne, G. D. and Hindmarsh, A. C., "Stiff ODE Solvers: A Review of Current
and Coming Attractions," Journal of Computational Physics, Vol. 70, 1987, pp. 1-
62.
38
Hindmarsh, A. C., "LSODE and LSODI, Two New Initial Value Ordinary
Differential Equation Solvers," ACM SIGNUM Newsletter, Vol. 15, 1980, pp. 10-
11.
39
Hindmarsh, A. C., "ODEPACK: A Systematized Collection of ODE Solvers,"
Scientific Computing, edited by R. S. Stepleman, M. Carver, R. Peskin, W. F.
Ames, and R. Vichnevetsky, North-Holland, Amsterdam, 1983, pp. 55-64; also,
Lawrence Livermore Lab., Rept. UCRL-88007, Livermore, CA, 1982.
40
Pratt, D. T., "Exponential-Fitted Methods for Integrating Stiff Systems of
Ordinary Differential Equations: Applications to Homogeneous Gas-Phase Chem-
ical Kinetics," Computational Methods, edited by K. L. Strange, Chemical Pro-

pulsion Information Agency, Johns Hopkins Applied Physics Lab., Laurel, MD,
Pub. 401, 1984, pp. 53-67.
41
Pratt, D. T. and Radhakrishnan, K., "Physical and Numerical Sources of Com-
putational Inefficiency in the Integration of Chemical Kinetic Rate Equations:
Etiology, Treatment and Prognosis," NASA TP-2590, 1986.
42
Radhakrishnan, K. and Pratt, D. T., "Fast Algorithm for Calculating Chemical
Kinetics in Turbulent Reacting Flow," Calculations of Turbulent Reactive Flows,
edited by R. M. C. So, J. H. Whitlaw, and H. C. Mongia, American Society of
Mechanical Engineers, New York, Pub. AMD-Vol. 81, 1986, pp. 313-314; also,
Combustion Science and Technology, Vol. 58, 1988, pp. 155-176.
43
Dold, A. and Eckmann, B. (eds.), Lecture Notes in Mathematics: Conference
on the Numerical Solution of Differential Equations, Vol. 363, edited by G. A.
Watson, Springer-Verlag, Berlin or New York, 1974.
44
Willoughby, R. A. (ed.), Stiff Differential Systems, Plenum, New York, 1974.
45
Lapidus, L. and Schiesser, W. E., (eds.), Numerical Methods for Differential
Systems, Academic, New York, 1976.
46
Desloux, J. and Marti, J. (eds.), Numerical Analysis, Birkhauser-Verlag, Basel,
Switzerland, 1977.
47
Bennet, A. W. and Vichnevetsky, R. (eds.), Numerical Methods for Differential
Equations and Simulation, North-Holland, New York or Amsterdam, 1978.
48
Dold, A. and Eckmann, B. (eds.), Lecture Notes in Mathematics: Numerical
Treatment of Differential Equations, Vol. 631, edited by R. Bulirsch, R. D. Gri-
gorief, and J. Schroder, Springer-Verlag, Berlin or New York, 1978.
49
Finlayson, B. A., Nonlinear Analysis in Chemical Engineering, McGraw-Hill,
New York, 1980.
50
Miranker, W. L., Numerical Methods for Stiff Equations and Singular Pertur-
bation Problems, D. Reidel Publishing, Dordrecht, Holland or Hingham, MA,
1981.
51
Ebert, K. H., Deuflhard, P., and Jager, W. (eds.), Modelling of Chemical
Reaction Systems, Springer-Verlag, Berlin or New York, 1981.
52
Dold, A. and Eckmann, B. (eds.), Lecture Notes in Mathematics: Numerical
Integration of Differential Equations and Large Linear Systems, Vol. 968, edited
by J. Hinze, Springer-Verlag, Berlin or New York, 1982.
53
Aiken, R. C. (ed.), Stiff Computation, Oxford University Press, New York,
1985.
54
Seinfeld, J. H., Lapidus, L., and Hwang, M., "Review of Numerical Integration
Techniques for Stiff Ordinary Differential Equations," Industrial and Engineering
Chemistry Fundamentals, Vol. 9, 1970, pp. 266-275.
55
Johnson, A. I. and Barney, J. R., "Numerical Solution of Large Systems of
Stiff Ordinary Differential Equations in a Modular Simulation Framework," Nu-
merical Methods for Differential Systems, edited by L. Lapidus and W. E. Schiesser,
Academic, New York, 1976, pp. 97-124.
56
Bui, T. D., Oppenheim, A. K., and Pratt, D. T., "Recent Advances in Methods
for Numerical Solution of O.D.E. Initial Value Problems," Journal of Computa-
tional and Applied Mathematics, Vol. 11, 1984, pp. 283-296.
57
"Proceedings of a Symposium on Reaction Mechanisms, Models, and Com-
puters," Journal of Physical Chemistry, Vol. 81, No. 25, 1977, pp. 2309-2586.
58
Lomax, H. and Bailey, H. E., "A Critical Analysis of Various Numerical
Integration Methods for Computing the Flow of a Gas in Chemical Nonequili-
brium," NASA TN D-4109, 1967.
59
Graves, R. A., Jr., Gnoffo, P. A., and Boughner, R. E., "An Implicit Semi-
analytic Numerical Method for the Solution of Nonequilibrium Chemistry Prob-
lems," NASA TM X-71985, 1974.
60
Enright, W. H. and Hull, T. E., "Comparing Numerical Methods for the
Solution of Stiff Systems of ODEs Arising in Chemistry," Numerical Methods for
Differential Systems, edited by L. Lapidus and W. E. Schiesser, Academic, New
York, 1976, pp. 45-66.
61
Warner, D. D., "The Numerical Solution of the Equations of Chemical Ki-
netics," Journal of Physical Chemistry, Vol. 81, 1977, pp. 2329-2334.
62
Dahlquist, G., Edsberg, L., Skollermo, G., and Soderlind, G., "Are the
Numerical Methods and Software Satisfactory for Chemical Kinetics?," Numerical
Methods for Differential Equations and Simulation, edited by A. W. Bennet and
R. Vichnevetsky, North-Holland, Amsterdam, 1978, pp. 149-164.
63
Gottwald, B. A. and Wanner, G., "Comparison of Numerical Methods for
Stiff Differential Equations in Biology and Chemistry," Simulation, Vol. 38, 1982,
pp. 61-66.
64
C6me, G. M., "The Use of Computers in the Analysis and Simulation of
Complex Reactions," Chemical Kinetics: Modern Methods in Kinetics, Vol. 24,
edited by C. H. Bamford and C. F. H. Tipper, Elsevier, Amsterdam or New York,
1983, pp. 249-332.
65
Edsberg, L., "The Development of Software," Stiff Computation, edited by
R. C. Aiken, Oxford University Press, New York, 1985, pp. 203-205.
66
Radhakrishnan, K., "New Integration Techniques for Chemical Kinetic Rate
Equations. I. Efficiency Comparison," Combustion Science and Technology, Vol.
46, 1986, pp. 59-81.
67
Radhakrishnan, K., "New Integration Techniques for Chemical Kinetic Rate
Equations: Part II—Accuracy Comparison," Journal of Engineering for Gas Tur-
bines and Power, Vol. 108, 1986, pp. 348-353; also, NASA TM-86893, 1985, and
ASME Paper 85-GT-30, 1985.
68
Radhakrishnan, K., "A Critical Analysis of the Accuracy of Several Numerical
Techniques for Chemical Kinetic Rate Equations," NASA TP (to be published).
69
Frank, P. M., Introduction to System Sensitivity Theory, Academic, New York,
1978.
70
Rabitz, H., Kramer, M., and Dacol, D., "Sensitivity Analysis in Chemical
Kinetics," Annual Review of Physical Chemistry, Vol. 34, 1983, pp. 419-461.
71
Radhakrishnan, K., "Decoupled Direct Method for Sensitivity Analysis in
Combustion Kinetics," Advances in Computer Methods for Partial Differential
Equations—VI, edited by R. Vichnevetsky and R. S. Stepleman, International
Association for Mathematics and Computers in Simulation, New Brunswick, NJ,
1987, pp. 479-486; also, NASA CR-179636, 1987.
72
Teets, R. E. and Bechtel, J. H., "Sensitivity Analysis of a Model for the Radical
Recombination Region of Hydrocarbon—Air Flames," Eighteenth Symposium
pp. 425-432.
73
Edelson, D., Kaufman, L. C., and Warner, D. D., "The Use of a High-Speed
Vector Processor Machine for Chemical Kinetic Sensitivity Analysis," ACS Sym-
posium Series No. 173, Supercomputers in Chemistry, 1981, pp. 79-88.
74
Coffee, T. P. and Heimerl, J. M., "Sensitivity Analysis for Premixed, Laminar,
Steady State Flames," Combustion and Flame, Vol. 50, 1983, pp. 323-340.
75
Kramer, M. A., Rabitz, H., Calo, J. M., and Kee, R. J., "Sensitivity Analysis
in Chemical Kinetics: Recent Developments and Computational Comparisons,"
International Journal of Chemical Kinetics, Vol. 16, 1984, pp. 559-578.
76
Cukier, R. L, Levine, H. B., andShuler, K. E., "Nonlinear Sensitivity Analysis
of Multiparameter Model Systems," Journal of Computational Physics, Vol. 26,
1978, pp. 1-42.
77
Tilden, J. W., Costanza, V., McRae, G. J., and Seinfeld, J. H., "Sensitivity
Analysis of Chemically Reacting Systems," Modelling of Chemical Reaction Sys-
terns, edited by K. H. Ebert, P. Deuflhard, and W. Jager, Springer-Verlag, Berlin,

1981, pp. 69-91.
78
Hindmarsh, A. C., ''GEAR: Ordinary Differential Equation System Solver,"
Lawrence Livermore Lab., Livermore, CA, Rept. UCID-30001, Rev. 3, 1974.
79
Hindmarsh, A. C. and Byrne, G. D., "EPISODE: An Effective Package for
the Integration of Systems of Ordinary Differential Equations," Lawrence Liver-
more Lab., Livermore, CA, Rept. UCID-30112, Rev. 1, 1977.
80
Ortega, J. and Rheinbolt, W. C., Iterative Solution of Nonlinear Equations in
Several Variables, Academic, New York, 1970.
81
Radhakrishnan, K., "Integrating Combustion Kinetic Rate Equations by Se-
lective Use of Stiff and Nonstiff Methods, AlAA Journal, Vol. 25, 1987, pp. 1449-
1455, also, NASA TM-86923, 1985; also, AIAA Paper 85-0237, 1985.
82
Rosenbrock, H. H., "Some General Implicit Processes for the Numerical Solu-
tion of Differential Equations," Computer Journal, Vol. 5, 1963, pp. 329-330.
83
Prokopakis, G. J. and Seider, W. D., "Adaptive Semi-Implicit Runge-Kutta
Method for Solution of Stiff Ordinary Differential Equations," Industrial and En-
gineering Chemistry Fundamentals, Vol. 20, 1981, pp. 255-266.
84
Kaps, P. and Rentrop, P., "Generalized Runge-Kutta Methods of Order Four
with Stepsize Control for Stiff Ordinary Differential Equations," Numerische Math-
ematik, Vol. 33, 1979, pp. 55-68.
85
Gottwald, B. A. and Wanner, G., "A Reliable Rosenbrock Integrator for Stiff
Differential Equations," Computing, Vol. 26, 1981, pp. 355-360.
86
Deuflhard, P., "Recent Progress in Extrapolation Methods for Ordinary Dif-
ferential Equations," SIAM Review, Vol. 27, 1985, pp. 505-535.
87
Deuflhard, P., Bader, G., and Nowak, U., "LARKIN: A Software Package for
the Numerical Simulation of Large Systems Arising in Chemical Reaction Kinetics,"
Modelling of Chemical Reaction Systems, edited by K. H. Ebert, P. Deuflhard,
and W. Jager, Springer-Verlag, Berlin, 1981, pp. 38-55.
88
Bader, G., Nowak, U., and Deuflhard, P., "An Advanced Simulation Package
for Large Chemical Reaction Systems," Stiff Computation, edited by R. C. Aiken,
Oxford University Press, New York, 1985, pp. 255-264.
89
Deuflhard, P. and Nowak, U., "Efficient Numerical Simulation and Identifi-
cation of Large Chemical Reaction Systems," Berichte der Bunsengesellschaft fuer
Physikalische Chemie, Vol. 90, 1986, pp. 940-946.
90
Bader, G. and Deuflhard, P., "A Semi-Implicit Mid-Point Rule for Stiff Sys-
tems of Ordinary Differential Equations," Numerische Mathematik, Vol. 41, 1983,
pp. 373-398.
91
Deuflhard, P., "Order and Stepsize Control in Extrapolation Methods," Nu-
merische Mathematik, Vol. 41, 1983, pp. 399-422.
92
Keneshea, T. J., "A Technique for Solving the General-Reaction-Rate Equa-
tions in the Atmosphere," AFCRL-67-0221, 1967.
93
Keneshea, T. J. and Swider, W., "Formulation of Diurnal D-Region Models
using a Photochemical Computer Code and Current Reaction Rates," Journal of
Atmospheric and Terrestrial Physics, Vol. 34, 1972, pp. 1607-1613.
94
Young, T. R. and Boris, J. P., "A Numerical Technique for Solving Stiff
Ordinary Differential Equations Associated with Reactive-Flow Problems," Na-
tional Research Lab., Washington, DC, NRL Memo. Rept. 2611, 1973.
95
Ordinary Differential Equations Associated with the Chemical Kinetics of Reactive-
Flow Problems," Journal of Physical Chemistry, Vol. 81, 1977, pp. 2424-2427.
96
Young, T. R., Jr., "CHEMEQ—A Subroutine for Solving Stiff Ordinary Dif-
ferential Equations," National Research Lab., Washington, DC, NRL Memo.
4091, 1980.
97
Pope, D. A., "An Exponential Method of Numerical Integration of Ordinary
Differential Equations," Communications of Association for Computing Machin-
ery, Vol. 6, 1963, pp. 491-493.
98
Liniger, W. and Willoughby, R. A., "Efficient Integration Methods for Stiff
Systems of Ordinary Differential Equations," SI AM Journal on Numerical Anal-
ysis, Vol. 7, 1970, pp. 47-66.
"Brandon, D. M., Jr., "A New Single-Step Implicit Integration Algorithm with
A-Stability and Improved Accuracy," Simulation, Vol. 23, 1974, pp. 17-29.
100
Babcock, P. D., Stutzman, L. F., and Brandon, D. M., Jr., "Improvements
in a Single-Step Integration Algorithm," Simulation, Vol. 33, 1979, pp. 1-10.
101
Pratt, D. T. and Radhakrishnan, K., "CREK1D: A Computer Code for Tran-
sient, Gas-Phase Combustion Kinetics," NASA TM-83806, 1984.
102
Pratt, D. T., "Fast Algorithms for Combustion Kinetics Calculations," Stiff
Computation, edited by R. C. Aiken, Oxford University Press, New York, 1985,
pp. 217-229.
103
Forsythe, G. E. and Moler, C. B., Computer Solution of Linear Algebraic
Systems, Prentice-Hall, Englewood Cliffs, NJ, 1967.
104
Byrne, G. D., Hindmarsh, A. C., Jacson, K. R., and Brown, H. G., "A
Comparison of Two ODE Codes: GEAR and EPISODE," Computational and
Chemical Engineering, Vol. 1, 1977, pp. 133-147.
105
Carver, M. B. and Boyd, A. W., "A Program Package Using Stiff, Sparse
Integration Methods for the Automatic Solution of Mass Action Kinetics Equa-
tions," International Journal of Chemical Kinetics, Vol. 11, 1979, pp. 1097-1108.
106
Radhakrishnan, K., "Fast Algorithms for Combustion Kinetics Calculations:
A Comparison," NASA CP-2309, 1984, pp. 257-267.
107
Burcat, A. and Radhakrishnan, K., "High Temperature Oxidation of Pro-
pene," Combustion and Flame, Vol. 60, 1985, pp. 157-169.
108
Radhakrishnan, K. and Burcat, A., "Kinetics of the Ignition of Fuels in Ar-
tificial Air Mixtures. II: Oxidation of Propyne," Combustion Science and Tech-
nology, Vol. 54, 1987, pp. 85-101.
109
Dagaut, P., Cathonnet, M., Boettner, J. C., and Gaillard, F., "Kinetic Mod-
eling of Propane Oxidation," Combustion Science and Technology, Vol. 56, 1987,
pp. 23-63.
110
Dunker, A. M., "The Decoupled Direct Method for Calculating Sensitivity
Coefficients in Chemical Kinetics," Journal of Chemical Physics, Vol. 81, 1984,
pp. 2385-2393.
1H
Dickinson, R. P. and Gelinas, R. J., "Sensitivity Analysis of Ordinary Dif-
ferential Equation Systems—A Direct Method," Journal of Computational Phys-
ics, Vol. 21, 1976, pp. 123-143.
112
Gelinas, R. J. and Skewes-Cox, P. J., "Tropospheric Photochemical Mech-
anisms," Journal of Physical Chemistry, Vol. 81, 1977, pp. 2468-2479.
113
Atherton, R. W., Schainker, R. B., and Ducot, E. R., "On the Statistical
Sensitivity Analysis of Models for Chemical Kinetics," AIChE Journal, Vol. 21,
1975, pp. 441-448.
114
Hwang, J-T., Dougherty, E. P., Rabitz, S., and Rabitz, H., "The Green's
Function Method of Sensitivity Analysis in Chemical Kinetics," Journal of Chemical
Physics, Vol. 69, 1978, pp. 5180-5191.
115
Dougherty, E. P., Hwang, J-T., and Rabitz, H., "Further Developments and
Applications of the Green's Function Method of Sensitivity Analysis in Chemical
Kinetics," Journal of Chemical Physics, Vol. 71, 1979, pp. 1794-1808.
116
Dougherty, E. P. and Rabitz, H., "A Computational Algorithm for the Green's
Function Method of Sensitivity Analysis in Chemical Kinetics," International Jour-
nal of Chemical Kinetics, Vol. 11, 1979, pp. 1237-1248.
117
Seinfeld, J. H. and Lapidus, L., ''Mathematical Methods in Chemical Engi-
neering," Process Modeling, Estimation, and Identification, Vol. 3, Prentice-Hall,
Englewood Cliffs, NJ, 1974.
118
Kramer, M. A., Calo, J. M., and Rabitz, H., "An Improved Computational
Method for Sensitivity Analysis: the Green's Function Method with k AIM'," Ap-
plied Mathematical Modeling, Vol. 5, 1981, pp. 432-441.
119
Magnus, W., "On the Exponential Solution of Differential Equations for a
Linear Operator," Communications on Pure and Applied Mathematics, Vol. 7,
1954, pp. 649-673.
120
Chan, S., Light, J. C., and Lin, J., "Inelastic Molecular Collisions: Exponential
Solution of Coupled Equations for Vibration-Translation Energy Transfer," Jour-
nal of Chemical Physics, Vol. 49, 1968, pp. 86-97.
121
Kramer, M. A., Calo, J. M., Rabitz, H., and Kee, R. J., "AIM: The Ana-
lytically Integrated Magnus Method for Linear and Second Order Sensitivity Coef-
ficients," Sandia National Labs., Livermore, CA, Rept. SAND82-8231, 1982.
122
Kramer, M. A., Kee, R. J., and Rabitz, H., "CHEMSEN: A Computer Code
for Sensitivity Analysis of Elementary Chemical Reaction Models," Sandia Na-
tional Labs., Livermore, CA, Rept. SAND82-8230, 1982.
123
Hwang, J-T., "Nonlinear Sensitivity Analysis in Chemical Kinetics," Pro-
ceedings of National Science Council Taiwan, Part B: Basic Science, Vol. 6, 1982,
pp. 20-28.
124
Hwang, J-T., "The Scaled Green's Function Method of Sensitivity Analysis
and its Application to Chemical Reaction Systems," Proceedings of National Science
Council Taiwan, Part B: Basic Science, Vol. 6, 1982, pp. 37-44.
125
Hwang, J-T. and Chang, Y-S., "The Scaled Green's Function Method of
Sensitivity Analysis. II. Further Developments and Application," Proceedings of
National Science Council Taiwan, Part B: Basic Science, Vol. 6, 1982, pp. 308-
317.
126
Vemuri, V. and Raefsky, A., "On a New Approach to Parameter Estimation
by the Method of Sensitivity Functions," International Journal of Systems Science,
Vol. 10, 1979, pp. 395-407.
127
Dunker, A. M., "Efficient Calculation of Sensitivity Coefficients for Complex
Atmospheric Models," Atmospheric Environment, Vol. 15, 1981, pp. 1155-1161.
128
Saito, H. and Scriven, L., "Study of Coating Flow by the Finite Element
Method," Journal of Computational Physics, Vol. 42, 1981, pp. 53-76.
129
Lojek, B., "Sensitivity Analysis of Nonlinear Circuits," IEE Proceedings, Vol.
129G, 1982, pp. 85-88.
130
Dunker, A. M., "A Computer Program for Calculating Sensitivity Coefficients
in Chemical Kinetic and Other Stiff Problems by the Decoupled Direct Method,"
General Motors Research Labs., Warren, MI, Research Pub. GMR-4831, ENV.-
192, 1985.
131
Leis, J. R. and Kramer, M. A., "Sensitivity Analysis of Systems of Differential
and Algebraic Equations," Computational and Chemical Engineering, Vol. 9, 1985,
pp. 93-96.
132
Caracotsios, M. and Stewart, W. E., "Sensitivity Analysis of Initial Value
Problems with Mixed ODEs and Algebraic Equations," Computational and Chem-
ical Engineering, Vol. 9, 1985, pp. 359-365.
133
Radhakrishnan, K. and Bittker, D. A., "GCKP86—An Efficient Code for
General Chemical Kinetics and Sensitivity Analysis Computations," Chemical and
Physical Processes in Combustion, Proceedings of the 1986 Fall Technical Meeting
of the Eastern Section of the Combustion Institute, The Combustion Institute, Pitts-
burgh, PA, 1986, pp. 46-1-46-4.
134
Leis, J. R. and Kramer, M. A., 'The Simultaneous Solution and Sensitivity
Analysis of Systems Described by Ordinary Differential Equations," ACM Trans-
actions on Mathematical Software, Vol. 14, 1988, pp. 45-60.
135
Leis, J. R. and Kramer, M. A., "Algorithm 658, ODESSA—An Ordinary
Differential Equation Solver with Explicit Simultaneous Sensitivity Analysis," A CM
Transactions on Mathematical Software, Vol. 14, 1988, pp. 61-67.
136
Lutz, A. E., Kee, R. J., and Miller, J. A., "SENKIN: A Fortran Program
for Predicting Homogeneous Gas Phase Chemical Kinetics with Sensitivity Anal-
ysis," Sandia National Labs., Livermore, CA, Rept. SAND87-8248, UC-4, 1988.
137
Cacuci, D. G., "Sensitivity Theory for Nonlinear Systems. I. Nonlinear Func-
tional Analysis Approach," Journal of Mathematical Physics, Vol. 22, 1981, pp.
2794-2802.
138
Cacuci, D. G., "Sensitivity Theory for Nonlinear Systems. II. Extensions to
Additional Classes of Responses," Journal of Mathematical Physics, Vol. 22, 1981,
pp. 2803-2812.
139
Seigneur, C., Stephanopoulos, G., and Carr, R. W., Jr., "Dynamic Sensitivity
Analysis of Chemical Reaction Systems," Chemical Engineering Science, Vol. 37,
1982. pp. 845-853.
140
Koda, M., Dogru, A. H., and Seinfeld, J. H., "Sensitivity Analysis of Partial
Differential Equations with Application to Reaction and Diffusion Processes,"
141
Hwang, J-T., "Sensitivity Analysis in Chemical Kinetics by the Method of
Polynomial Approximations," International Journal of Chemical Kinetics, Vol. 15,
1983. pp. 959-987.
142
Pratt, D. T., "Exponential-Fitted Methods for Stiff Ordinary Differential
Equations," Proceedings of Eleventh IMACS World Congress on System Simulation
and Scientific Computation, Aug. 1985.
143
Petzold, L., "Automatic Selection of Methods for Solving Stiff and Nonstiff
Systems of Ordinary Differential Equations," SIAM Journal of Scientific and Sta-
tistical Computations, Vol. 4, 1983, pp. 136-148.
Chapter 5
Reduction of Chemical Reaction Models
Michael Frenklach
I. Introduction: Why Reduction?
C HEMICAL reaction models, describing the evolution of a mixture of

chemical species, usually with corresponding energy changes, are
important components of reactive-flow models. 1 In most combustion sys-
tems, the species are not in equilibrium and the precision of the overall
model prediction depends on the accuracy of the representation of the
chemical kinetics, among other factors. More reliable numerical results are
obtained with more detailed reaction mechanisms, usually at the level of
a set of elementary reaction steps. Current interests, such as simulation of
engine knock, 2 " 4 production of NO X , 5 ^ 6 and soot formation, 7 " 10 require
increasingly large reaction mechanisms—hundreds of chemical reactions
and species. However, the computational capabilities for an accurate, full-
scale gasdynamic/chemical-kinetic simulation of these multidimensional
reactive flows are not yet available.11
Two general views of approaching this problem have been expressed.
One is to wait until sufficient computational power and efficient numerical
integration techniques become available to make large-scale modeling fea-
sible. The technological needs and scientific interests will, however, prob-
ably always outgrow the increase in computational capabilities. The other
view, which is the subject of this chapter, is to reduce the reaction mech-
anism. "Reduction" does not merely mean "simplification" of the chem-
istry involved, but rather reduction in the complexity of the mathematical
form that describes the chemical transformations of a given reaction model.
These two views are rooted in two different general philosophies of
chemical kinetic modeling. The first is that the same single-reaction mech-
anism must be used in all reactive-flow models, because different parts of
this reaction set become important at different conditions. For instance,
it is quite common that the rate-limiting step switches from one reaction
to another with the flame height. Omission (or reduction) of certain parts
of the reaction mechanism is viewed, then, as the loss of mechanism uni-
Copyright © 1990 by the American Institute of Aeronautics and Astronautics. All rights
reserved.
129
130 REDUCTION OF CHEMICAL REACTION MODELS
versality. Scientific progress in this view is perceived as building and expanding

a comprehensive reaction set to cover the increasing number of experi-
mental conditions. Obviously, one cannot argue against universality of
nature and, hence, against universality of the reaction mechanism required
to model it, because elementary processes—bond-rearrangement in chem-
ical reactions, energy transfer in molecular collisions, etc.—are the same
regardless whether they occur in hydrogen, octane, or coal flames. How-
ever, current advances in experimental techniques and theoretical methods
rapidly expand the present database of combustion chemistry. In principle,
the kinetic database is infinitely large, and the practical reality for this to
happen may be in the not-too-distant future (similarly to the presently
unlimited thermodynamic database: thermodynamic properties for any
chemical compound can be estimated by group additivity or quantum
mechanical methods). This brings to bankruptcy the philosophy of the
comprehensive kinetic model and leads in a natural way to the following
view.
All conceivable chemical reactions with the associated rate parameters—
available experimentally, calculated theoretically, or simply estimated—
constitute a reaction data bank. Given a specific problem, a subset of the
data bank should be taken to assemble the reaction mechanism. The ques-
tion is how to choose this reaction subset such that it describes faithfully
the dynamic behavior of a reactive-flow system. That is, how does one
reduce an extensive reaction mechanism to a minimum set of reactions
required to address a given problem? Moreover, can this quantitatively
accurate subset be reduced further to a different but much smaller reaction
set or, in general, to a different mathematical form that substantially decreases
the demands on computational capacity? The methodologies required to
answer these questions constitute the now-growing field of mechanism
reduction.
The idea of mechanism reduction is not new. Introduction at the begin-
ning of this century of the pseudo-steady-state approximation for reactive
intermediates led to analytical solution of nonlinear reaction systems that
otherwise could not be solved at that time. 12 Expressing the reaction events
in terms of the outcome probability allowed the analysis and prediction of
the behavior of chain reactions.13 A rigorous mathematical formulation for
the problem of "lumping"—in which chemical species are transformed,
or "lumped," into a few dynamically equivalent "lumped classes"—was
started in the 1960s.14 ~ 18 To rationalize the concept of lumping, Wei and
Kuo17 wrote: "We often consider, for example, all oxygen molecules as
'oxygen' even though the kinetic energies of the individual oxygen mole-
cules are different. Such lumping also gave petroleum processing the . . .
analysis, in which all species are divided into four classes: paraffins, olefins,
naphthalenes, and aromatics." It is important to realize that even the most
detailed chemical reaction models used presently in combustion are lumped
models in the sense that the chemical species of these models are actually
lumped classes of energy-distributed species and the "elementary" rate
coefficients used for the reactions of these lumped classes are certain aver-
ages over the population of these energy-distributed species (see, for exam-
ple, Ref. 19).
M. FRENKLACH 131
As in many other areas of science, terminology is certainly a problem.

The same terms—like mechanism reduction—are used to describe differ-
ent methods and, conversely, different terms—like reduction and lump-
ing—are used by different scientific communities to describe the same
mathematical problem. In this chapter, an attempt is made to bridge the
different terminologies pertaining to reduction of chemical reaction models.
First, the modeling objectives typically encountered in combustion science
and technology are discussed. The following sections deal with definition
and classification, and present the main ideas of the different individual
methods. Several specific methods, those developed by the present author
and his co-workers, are discussed in more detail.
II. Modeling Strategy

Model reduction constitutes an integral part of modeling, and, as such,
the reduction strategy and specific techniques chosen depend on the objec-
tives of the modeling. There is no single definition for the terms modeling
and model. Common usage reflects a multitude of definitions used in dif-
ferent fields by different people in different ways (e.g., Ref. 1, Chap. 1;
Refs. 20-24). Models and modeling can be classified in several ways: by
the mathematical form the equations take (e.g., algebraic, differential, or
integral); by the nature of the model (empirical, physical, or physico-
empirical); by the phenomena described (kinetic, gasdynamic, or a coupled
reactive flow); and so forth.
Perhaps the most important classification, particularly in the context of
the present discussion, is the one based on the objectives of the modeling
undertaking. The objectives typically encountered in the areas of com-
bustion modeling are the following:
1) Quantitative prediction of system responses. These are models devel-
oped for practical applications, such as computer-aided design (CAD) of
combustors or forecast of pollution levels. The principal requirement for
these models is a given prediction accuracy. The model itself can take any
form, as long as it delivers results with the required accuracy. The challenge
is to make it least computer intensive so that it is suitable for CAD appli-
cations.
2) Quantitative or qualitative analysis of certain experimental or theo-
retical relationships, dependencies, and trends with "established" models.
Examples of these could be analysis of the chemical-kinetic/gasdynamic
coupling in a reactive flow, simulation of flame-stretching phenomena,
testing of a sooting correlation, or verification of a concept for modeling
turbulence. The models used for these purposes must contain a detailed
description of the phenomenon in question, so that controlling factors can
be identified. The numerical accuracy of model prediction is not necessarily
at issue here. For instance, a model predicting the amounts of soot formed
to within a factor of 2 will certainly be suitable for the determination of
factors controlling the critical C-to-O ratios for the onset of sooting.
3) Parameter determination for well-established models. The most famil-
iar example here is the determination of rate coefficients in chemical kinetic
studies. It is not only necessary to establish the best-fit values of the
"unknown" parameters, but also to estimate their uncertainties in rigorous

statistical terms, i.e., to determine the joint confidence region, and to
examine the correlations between the parameter values. The models in
this category should include all active parameters—i.e., those rate coef-
ficients, equilibrium constants, transport coefficients, etc.—to which avail-
able experimental data exhibit significant sensitivity.
4) Exploratory modeling with ''unknown" models. This could mean, for
example, identification of possible reaction pathways leading to soot for-
mation. The focus here is on discovering the unknown by extrapolating
the present knowledge. These models, therefore, should include as many
details (i.e., chemical reactions and species) as possible so as not to miss
important features.
These different objectives dictate that different methods and strategies
should be adopted for model reduction. Several methods of model reduc-
tion have been proposed. They are classified according to the general
nature of the method as follows. The techniques included in the first class,
global modeling, are those that transform a complete reaction scheme into
a small number of global reaction steps. The approach in the second class,
response modeling, is to develop direct algebraic relationships between
model responses and model variables. In the third class, detailed reduction,
the objective is to reduce the size of a detailed mechanism by eliminating
noncontributing reactions. The fourth class, chemical lumping, presents a
mathematically rigorous reduction technique that describes the kinetics of
an infinite sequence of polymerization-type reactions by a small number
of differential equations developed for the moments of the polymer dis-
tribution function. Finally, the reduction methods of the fifth class, grouped
under the title of statistical lumping, are focused on systems of interacting
particles.
III. Global Modeling

In this method, the entire reaction scheme is reduced to a single overall
reaction step or a small number of usually unrealistic chemical reactions.
These global reduction strategies can be subdivided further based on the
methodology employed.
Empirical Fitting
The crudest approach to mechanism reduction is the development of
overall formal rate equations by using simple empirical fitting. An example
is given by Hautman et al.,25 who developed a four-step reaction set,
CnH2n+2-*(n/2)C2H4 + H2
C2H4 + O2 -> 2CO + 2H2
CO + i02 -» C02
H2 + K)2 -* H20
M. FRENKLACH 133
to fit flow-reactor and shock-tube measurements on propane oxidation.

The rates of the preceding empirical reactions were fitted into formal rate
expressions
Reaction rate - Ae~Q/T H Cf' (1)
/
where T is the reaction temperature, Ct are the concentrations of reactants
and products, and A, 0, and (3, are fitted parameters. Equations of this
type can be used to develop some practical correlations. Other examples
using this approach include modeling of hydrocarbon-air combustion in a
stirred tank reactor,26 simulation of gaseous detonation,27 a one-step de-
scription of moist CO oxidation,28 and the use of global-model expressions
for correlation of various rate-related properties, such as ignition delays
and maximum rates of soot formation.
It is usually difficult, however, to obtain high-quality fits over large
ranges of parameter values with simple mathematical expressions like Eq.
(1), especially for those properties that have a rather complex behavior.
The main disadvantage of this purely empirical approach is that the cor-
relations developed can be used to predict just one or two properties and
only under the conditions for which these equations were developed.
Reduction by Approximations
Reductions in this category are accomplished by identifying the main
reaction chains and rate-limiting steps of a given reaction system, and
applying simplifying kinetic approximations such as pseudo steady state
and partial equilibrium. This approach has been pursued actively by Peters
and co-workers (Refs. 29 and 30 and references therein). For example,
Peters29 developed a reaction mechanism for modeling methane flames
composed of four global steps:
CH4 + 2H 4 H20 -> CO 4 4H2
CO 4- H2O -> CO2 4 H2
H4H4M-»H 2 4M
O2 4 3H2 -» 2H 4 2H2O (2)
Each step represents the stoichiometry of a certain reaction subset. The
development of the first global reaction in this mechanism is illustrated
subsequently.
The principal (rate-limiting) reaction pathway for the oxidation of CH4
to CO in a methane flame is given by the following set of elementary
reactions:
CH4 4 H -» CH3 4 H2 (3)
CH3 4 O -» CH2O 4 H
CH20 4 H -* CHO 4 H2
CHO 4- M -» CO 4 H 4 M (4)
Adding these chemical equations together, we obtain

CH4 + O -» CO + 2H (5)
In flame environments, O atoms are rapidly converted to H atoms (and
vice versa) by the following reactions:
O + H2 -» H + OH (6)
OH + H2 -» H + H20 (7)
Upon subtraction of chemical equations (6) and (7) from Eq. (5), the global
reaction (2) is obtained. Other global reactions in that mechanism are
obtained in a similar manner. The rates of the global reactions, however,
are not specified by the stoichiometric relationships, as in the case of
elementary reactions, but are determined by further analysis. For instance,
the rate of global reaction (2) is determined primarily by the rate of ele-
mentary reaction (3). Peters warns, however, that the assumptions made
can and should be verified by asymptotic analysis.
This approach, based on a physical analysis, offers formal reaction schemes
with a wider range of applicability than those obtained by the purely empir-
ical fitting, although it is still limited to the specific ranges of conditions
under which the assumptions invoked hold. Thus, for instance, partial
equilibrium assumed for the burned-gases region may not be attained in
the preheat zone of the same flame. Another disadvantage is that each
system and each different condition requires an individual examination.
For example, the reduced reaction scheme developed by using this method
for a methane flame cannot be used for methane ignition, because the
principal oxidation step of methane ignition,
CH3 + O2 -> products
is different from reaction (4), which dominates the oxidation of CH3 in
the flame. The method, however, is well suited for theoretical and ana-
lytical analysis of fundamental aspects of complex dynamic systems. Further
examples of the reduction-by-approximation approach can be found in
Refs. 31-38.
Lumping
As mentioned earlier, the initial ideas of lumping were introduced by
chemical engineers in the early 1960s.14~18 The main concepts of this method
are illustrated next.
Let us consider a system of linear ordinary differential equations (ODE's)
that fully describe the evolution of chemical species,
f-K (8,
where C = {C\, C2, ..., Cn}T is an /^-dimensional column vector of con-

centrations of all of the reacting species in the system, K a matrix of the
corresponding rate coefficients, and t the reaction time. It is possible to
M. FRENKLACH 135
transform the complete set of Eq. (8) into a kinetically equivalent but
reduced set of equations
where c = {c l5 c 2 , ..., cm}T is an ra-dimensional column vector of the lumped

concentrations and m is (much) smaller than n. The lumped concentrations
c are defined as linear combinations of the original concentrations C,
/, 22C2 + ••• + A^Q,
Ci = ViQ + X/,
/ - 1, 2, ..., m, m < ^ (10)
The rate coefficients of the reduced system, &, are then also determined
by certain combinations of the rate coefficients of the original system, K.
In the matrix notation, the transformation specified by Eq. (10) is given
as
where the subscripts indicate the corresponding dimensions. Matrix L =

{\i}} is called the lumping matrix. As an example, consider the relationship
which specifies a transformation of three original concentrations (Cl, C2,

and C3) into two lumped concentrations (c1 and c2) as follows:
Cj = C, c2 = C2 + C3
Another example of a lumping transformation will be given later in this

chapter in the section entitled "Chemical Lumping."
Wei and Kuo17 defined the system described by Eq. (8) to be exactly
lumpable by a matrix L if there exists a matrix k such that the kinetic
behavior of the lumped system can be described by Eq. (9); in other words,
when the transformation from a complete system into a lumped system is
accomplished with full mathematical rigor and, thus, without loss of kinetic
information. Wei and Kuo proved that, for a linear system, i.e., a system
of differential equations describing first-order reaction kinetics, to be exactly
lumpable, it is necessary and sufficient that condition
L x K = k x L (12)
holds. Kuo and Wei18 extended their analysis to first-order kinetics systems
that are not exactly lumpable. The necessary and sufficient conditions of
exact and approximate lumping of these systems were extended by Liu and
Lapidus.39 Lumping analysis of nonlinear systems (e.g., second-order reac-
tion kinetics) are presented in Refs. 40-46.
The underlying principle in all of these lumping methods is the search

for a mathematical transformation without taking into account the specifics
of a given system. In fact, a goal of the current research in this area is to
determine a universal transformation procedure that would allow the devel-
opment of a "push-button" numerical algorithm. Although it is not clear
that such a technique can be found, a question has been raised as to whether
it would be more appropriate to consider the specific "topology" of the
reaction network and, thereby, to develop different lumping methods for
different topologies. For instance, Hutchinson and Luss47 and Luss and
Hutchinson48 suggested a lumping method for a system of parallel reactions.
Bailey49 proposed a lumping analysis suitable for a continuous mixture of
chemical compounds. Taking into account the analogy in chemical structure
of the reacting species, a lumping method for polymerization kinetics was
developed in Ref. 50. The particular approach used in this method is
referred to as chemical lumping, as opposed to the other lumping methods
discussed earlier, which can be referred to as mathematical lumping. Chem-
ical lumping is discussed in detail later in this chapter.
The use of the mathematical lumping offers a higher accuracy and, to a
certain degree, better extrapolation capabilities of the reaction model reduced
in this manner than other methods of global modeling, empirical fitting,
and reduction by approximation. It is pertinent to stress here that the
mechanism reduction following any approach of global modeling presented
in this section always results in differential equations, as contrasted by
algebraic ones obtained using the methods of response modeling discussed
next.
IV. Response Modeling

A mathematical model specifies relationships between model responses
and model variables. In detailed chemical kinetic modeling, the mathe-
matical model takes the form of differential equations. The model responses
are typically concentrations of chemical species. The model variables are
the initial and boundary conditions of the reacting mixture and the model
parameters, such as thermochemical data, rate coefficients, and transport
properties. A complete solution of the differential equations provides rela-
tionships between all of the responses and all of the variables. We will
refer to these relationships as response functional relationships, regardless
of whether they are given in an analytical, tabular, or numerical form. In
most applications, one only needs to know several of these response func-
tional relationships.
One such application was demonstrated by Oran et al.51 They used an
ODE solver to compute ignition delay times of hydrogen-air and methane-
air mixtures using a full set of chemical reactions describing hydrogen
combustion. The computed ignition delays, T, were fitted to a simple alge-
braic equation that represented the computed dependence of the ignition
delay time on pressure P and temperature T. The fitted function T(P,T)
was then used as an input in one- and two-dimensional simulations of gas-
phase detonation phenomena. In other words, some of the main features
of combustion chemistry of interest to the detonation studies were sum-
M.FRENKLACH 137
marized by the relationship T(P,T). This response model was developed

in separate computer simulations with full chemistry, and then used as a
"chemistry input" to a computationally intensive fluid dynamics code. In
another example, Coffee et al.52 53 developed a response functional rela-
tionship based on detailed chemical kinetics computations for the rate of
energy release, and then used this response model in numerical simulations
of several premixed, laminar, steady-state flames.
A general procedure, named solution mapping, was suggested in Refs.
23 and 54. In this method, model responses are expressed as simple alge-
braic functions (usually polynomials) in terms of model variables. These
algebraic functions, called response surfaces, are obtained by using a rel-
atively small number of computer simulations, referred to as computer
experiments, with the complete kinetic model. The computer experiments
are performed at preselected combinations of the values of the model
variables, and the entire set of these combinations is called a design of
computer experiments. The terminology and the methods are taken from
the field of response surface design,55^ developed for optimization of a
poorly understood process by performing the smallest number of experi-
ments possible. Application of this methodology to a computer model has
been demonstrated in such areas as electronic circuit design, controlled
nuclear fusion, plant ecology, and thermal energy storage. 57
Imagine an agricultural study in which the researcher is interested in
finding the effects of different soil treatments on the growth of a grain.
Let us assume that one experiment takes a year to complete and a limited
number of experiments can be performed each year. The knowledge gained
after one round of experiments will depend on the way the individual
experiments were chosen, i.e., on the experimental design. Some designs
may leave the researcher with no information whatsoever, whereas other
designs can illuminate clearly the important features. The objective of the
surface-design techniques is to provide the most economical designs. Choosing
such an optimal design, the agricultural researcher can be confident that
the maximum information possible was obtained for the given constraints
and resources. The effectiveness of the response surface design increases
with the increase in the number of variables. Computer experiments can,
in principle, be treated similarly to physical experiments. In addition, the
fact that computer experiments do not have random errors may lead to
more efficient designs.57'58
When the model responses are assigned to be experimentally determined
entities (e.g., measured species concentrations) and the model variables
are chosen to be parameters whose values are uncertain (e.g., unknown
rate coefficients), the developed response surfaces can be used conveniently
for parameter determination and model optimization.23 59 ~ 62 These response
surfaces represent a reduced model, specifying the essential dependencies
that exist between the variables of concern. The response surfaces take
the form of simple algebraic equations and, thus, can be implemented as
a simple numerical subroutine in a sophisticated optimization algorithm.
In other words, this technique allows one to decouple the optimization
search, which by itself is not a trivial task in problems associated with
chemical kinetics,23 from numerical integration of the differential equa-
tions. Hence, the solution mapping approach offers a large overall numer-
ical efficiency in model optimization, as was demonstrated recently with a
practical-size case of methane combustion.62
The same method of solution mapping, only now with the initial con-
ditions taken as model variables, can be used to "replace" the detailed
description of chemical kinetics by using simple algebraic relationships in
complex fluid dynamic simulations of reactive flows. In principle, this is
similar to the induction-parameter model of Oran et al.51 discussed earlier,
but offers a more systematized approach to the problem. The method was
demonstrated recently using an example of a three-dimensional simulation
of the ozone level in the Los Angeles Air Basin.63 A chemical reaction
mechanism composed of 52 reactions and 24 species was adopted to describe
the chemical kinetics of ozone production. The concentrations of 15 of the
chemical species at a particular reaction time t were chosen as model
responses, and the concentrations of these chosen species at t = 0 were
assumed to be model variables. The concentrations of the remaining nine
species were computed based on the pseudo-steady-state approximations.
Using a factorial design for computer runs with different sets of initial
concentration values, response surfaces for each of the 15 chosen species
concentrations were developed in the following polynomial form:
15 15
lnCt = aQ + 2 ai '"Qo + 2 a
u '"Qo '"Cjfi
/=! i,;=l
where Ci0 is the initial concentration of species i and a the fitted coeffi-
cients. These runs were performed for isothermal isobaric conditions,
otherwise the initial temperature and pressure could also be included as
model variables. The developed polynomials [Eq. (13)] then served as an
input for gasdynamic simulations, which were performed employing the
operator splitting method. The computed ozone concentration profiles were
found to be in a close agreement with the solution obtained using the
detailed-chemistry approach.
The main advantage of this method is that the reduced model takes the
form of algebraic equations, e.g., Eq. (13). It means that the time step for
integration of reactive-flow equations becomes independent of the stiffness
of the kinetic equations, the factor that usually slows down the integration
of coupled chemical-kinetic/gasdynamic systems.1 As with model optimi-
zation, the methodology of solution mapping allows one to decouple the
numerical procedure into two separate routines: integration of chemical
kinetics and solution of gasdynamic equations. Another advantage of this
approach is that the reduction of a dynamic kinetic model into an algebraic
form is accomplished without sacrifice of chemical information contained
in the complete reaction scheme. Further increase in the efficiency of the
technique can be achieved by incorporating some of the ideas of global
modeling discussed in the preceding section, which will lead to the reduc-
tion in the number of reactive species and thus to the reduction in the total
number of algebraic equations. This methodology makes it feasible to
approach simulation of realistic combustion environments even at the pres-
ent state of computational capabilities.
M. FRENKLACH 139
V. Detailed Reduction
In many situations—for instance, in exploratory modeling—one is not
interested in the mathematical simplicity of the model but rather in its
completeness with respect to incorporating important chemical and physical
processes. Initially, such a chemical model should include all potentially
important chemical reactions and species. Not all of them contribute sig-
nificantly to specific responses of interest, and those can be removed safely.
The question then is how to identify these noncontributing reactions, or,
in other words, how to reduce a first-trial large-size reaction mechanism
to a minimum set that is sufficient for a given study. Removing one reaction
at a time or using formal sensitivity analysis may be a tedious and awkward
approach.
An effective reduction strategy is to consider the contribution a given
reaction rate term makes to certain sums of such terms (e.g., those com-
prising the rates of accumulation of given reaction species). One suggested
approach23'54 was to identify noncontributing reactions by comparing the
individual reaction rates with the rate of a chosen reference reaction. The
reference reaction could be the rate-limiting step or the fastest reaction.
It was further proposed64 that the computations used to obtain the required
rate data do not have to be performed with the actual full fluid dynamic/
chemical-kinetic model—the reduction can be achieved with a much sim-
pler computation using a similar, but greatly simplified, geometry and
conditions.
These ideas were demonstrated using convectively perturbed methane
ignition.64 The reduction was based on the results computed with a simple
constant-density model using a full detailed chemical reaction mechanism.
The following criteria were aplied to identify the noncontributing reactions:
|K/I < e*|/?rls| (14)
|tf,A//,j < zQQmax (15)
where Rf is the rate of reaction /, R rls the rate of the rate-limiting step
(which, for the most part, was the reaction H + O2 -> OH + O), A//,
the heat of reaction /, Q max the maximum value among all of the terms
l/^-A/f/l, and ER and e0 chosen parameters considerably smaller than unity.
Reactions whose corresponding rates satisfied inequalities (14) and (15)
were removed from the mechanism. This procedure eliminates a specific
species when all reactions of such a species happen to be removed. Three
reduced reaction mechanisms were developed, based on the assumed values
ZR = ZQ = 0.01, 0.05, 0.1. The larger these values are, the smaller and
less accurate the reduced mechanism becomes. The reduced reaction mech-
anisms were then used in a series of reactive flow simulations with and
without a large-amplitude sinusoidal perturbation applied to a system that
is initially quiescent and whose temperature is high enough to start the
ignition process. The results showed that in every case the trends computed
with the perturbation mimic those without them and the faithfulness of the
numerical description with the reduced mechanisms increases with a decrease
in the values of ZR and EQ.
These results illustrate the postulate64 that a chemical reaction mecha-

nism faithfully describing the dynamics of both thermal and chain reaction
processes when there is no convective transport present will generally describe
the chemical processes to the same degree of accuracy in actual reactive
flows such as flames and detonations. Furthermore, the accuracy of a
reduced model can be set a priori.
VI. Chemical Lumping

In the discussion of lumping methods, it was mentioned that taking into
account the specific reaction network " topology" can be a better strategy
for mathematical method development than a search for a universal lump-
ing algorithm. One such topology is polymerization-type kinetics, where,
beginning with a certain polymer size, the chemical reactions describing
the polymer growth are basically the same and the associated thermo-
chemical and rate parameters remain essentially constant or exhibit only
a weak dependence on the polymer size. An example can be provided by
the mechanism of the production of polycyclic aromatic hydrocarbons (PAH)
in flame environments.7-8 In this reaction system, the PAH growth beyond
a certain size was identified to proceed by a replication-type reaction se-
quence. It appeared, however, that reversibility of chemical reactions in
the replicating reaction sequence is critical for accurate modeling of the
process and, therefore, none of the classical methods developed for free-
radical polymerization kinetics could be applied because they all assume
irreversible growth steps.
In the method of chemical lumping described in this section, the lumping
procedure is guided by similarity in chemical structure or chemical reac-
tivity of the reacting species in the replicating reaction sequence.50-54 65 A
simple example is used here to illustrate the underlying ideas involved.
This example is based on the following reaction system proposed to rep-
resent the fundamental aspects of PAH growth in hydrocarbon systems66:
Ai: + H «± A* + H2 (U)
A* + C2H2 *± AjCHCH* 0,2)
+ C2H2-*A[ + l + H 03)
+±A + * + H2 0+14)
+ C 2 H 2 -> A + 2 + H 0+13)
... etc. 0 + 2, / + 3, ..., <*)
where At denotes an aromatic molecule containing i fused aromatic rings,
A? is an aromatic radical formed by the abstraction of an H atom from
Ah and ^4f-CHCH» is a radical formed by adding C2H2 to A*. Reactions
0,1)-0,3) a°d 0 + 1, !)-(/ + 1,3) are chemically similar, and each set of
the three reactions completes the building cycle of one aromatic ring. It is
assumed that this replicating reaction sequence continues to infinity.
M. FRENKLACH 141
The differential equations describing the population of polymer species

specified by the given reaction sequence are
[H2] (I,/)
(2,7)
- A:2[/l>][C2H2] + fc_2[A,.CHCH«]
= k2(A,.}[C2H2] - k 2(A,CHCH-} (3,7)
- A:3[A,.CHCH«][C2H2]
J = /c3[ACHCH-][C2H2] - UA+.HH] (1,7+1)
(2,7+1)
- *2[A-+i'][C2H2]
(3,7+1)
LU
- £3[A + iCHCH-][C2H2]
... etc. (oc)
where r0 is the rate of formation of Aj by preceding initiation reactions, t
is the reaction time, and k} is the rate coefficient of reaction j assumed to
have the same value in each cycle /. Summing Eqs. (1,/)-(!,oc), we obtain
d[Ai + l] ,
or
dM0
= /•„ (16)
dr
where
M0 = [A,] + [A,-] + [A-CHCH-] + [Ai+1] + -
Equations (16) describes the evolution of the total PAH concentration,
M0.
Multiplying Eqs. (l,/)-(l,°c) by the corresponding molecular mass, i.e.,

multiplying Eq. (I,/) by ra0, where ra0 is the number of carbon atoms
contained in Ah Eq. (2,7) by m0, (3,7) by m() -h 2, (1,7 -hi) by ra0 -f 4,
(2,7 +1) by ra0 -h 4, (3,7+1) by m0 + 6, etc., and adding them together,
we obtain
d[Ai + l]
m }
dr dt
^ , ,.
4) (m0 + 6)
-h
- 2/c3[^/+1CHCH-][C2H2]
or
[/1,-CHCH-
2£3[C2H2 (17)
where
(m0
(m0 (m0 4-
is the total number of carbon atoms accumulated in PAH's or the first

concentration moment of the PAH distribution function. The terms X,{A/«]
and S/[^4/CHCH-] appearing in Eq. (17) are the lumped concentrations.
Let us define the lumped concentrations of this reaction system as
(18a)
c2 = (18b)
[A.CHCH-] (18c)
These lumped concentrations are determined by solving differential equa-

tions
/c3[C2H2]c3 (19a)
(19b)
M. FRENKLACH 143
dr
-£ = k2[C2H2]c2 - k_2c3 - k3[C2H2]c3 (19c
which are obtained in the folio wing way. Summation of Eqs. (1,7), (1,7+ 1),
..., (l,oo) results in Eq. (19a); summation of Eqs. (2,7), (2,7+1), ..., (2,oc)
in Eq. (19b); etc.
Integration of Eqs. (16), (17), and (19), combined with the differential
equations that describe the kinetics of the initiation reactions defining r(),
[H], [H2], and [C2H2], determines M() and Ml at any given reaction time.
The ratio
determines the average number of carbon atoms in a PAH molecule, or

in other words, the mean PAH size, or the first moment of the PAH
distribution function. In a similar manner, equations for higher concen-
tration moments are derived and higher moments of the PAH distribution
are computed, thus specifying such statistical properties as the width, skew-
ness, etc.,67 of the size distribution function.
The lumping represented by Eqs. (18), obtained in a natural way, can
be expressed following the formalism of Eq. (11) as
^3X1 ^3 X 3C L DC X 1
'100100. . A I [ACHCH-]
010010. .. I [Ai+l]
vOOlOOl . .
where the symbols in the top row — c, L, and C — identify the corresponding
entities in reference to Eq. (11). It can be shown by a straightforward
matrix multiplication that the lumping matrix in Eq. (20) satisfies the con-
dition of exactness, given by Eq. (12).
Equations (19) can be represented schematically as
r0 ki[H] te[
/ ru i
k-i[H2\ /
k-2
indicating that the lumping discussed in this section can be called linear
lumping of a nonlinear system. Most of the reactions in the PAH growth
sequence (21) are bimolecular and, thus, represented by nonlinear terms
in the corresponding differential equations. However, the concentrations
of H, H2, and C2H2 are independent of the index of summation, i\ hence,
the summations defining the lumping are linear with respect to the PAH
concentrations. Nonlinear lumping is discussed in the next section.
VII. Statistical Lumping

The reduction method discussed in the previous section deals with a
specific topology of the reaction system: linear kinetics of polymer growth,
i.e., a process when a polymer grows by reaction with "monomer" species
only. In this section, we consider the case when polymer-polymer inter-
actions are of concern. Such processes become of interest to combustion
modeling in areas of soot formation and flame-generated production of
powders such as silica and various metal oxides.
Let us consider a coagulation process as an example. Given an initial
number of particles per unit volume of gas, they begin to coagulate, i.e.,
to collide with each other forming a new particle upon each collision. This
process is represented schematically by
At + Aj-Â^j, i,j= 1,2, ...,« (22)

where At is a particle with the mass mi = /m l 7 and m± is the mass of the
smallest particle. Here (3 /y is the collision frequency coefficient whose value
depends on / and j. The dynamics of the reaction system (22) is described
by the Smoluchowski equations (see, for example, Ref. 68):
—
j, - = -o y
^ 6 -N-N-'-; - y
R/,/-rV ^ B -N-N-
P'^VV i = 2' • • • ' ^ (23
\^j.\)i)
uf L j=\ j=i
where t is the reaction time and Nt the number density of particles of size
i. Equations (23) do not possess a known closed-form solution. Friedlander
and co-workers (see, for example, Ref. 69 and references therein) showed
that Eqs. (23) have an asymptotic solution at t^ oc; that is, at large reaction
times, the normalized distribution function does not change its shape with
time. This solution is referred to as self-preserving size distribution. The
numerical integration of Eqs. (23) when the explicit form of p, 7 is specified
is straightforward, but encounters severe, practically prohibitive demands
on computer time and memory.70 The reason for this is that the particle
population spreads very quickly, involving many species of different sized
particles.
Several mechanism reduction methods have been proposed for this prob-
lem. Seinfeld and co-workers (see, for example, Ref. 71) developed a
discrete-sectional method, in which the continuous aerosol size distribution
is approximated with a finite number of size sections and the properties
within each section are averaged. The computer code developed by these
researchers, known as MAEROS, works successfully for systems without
a strong source term, i.e., when there is no large production of particles
by (homogeneous) nucleation. With such a source present, as is the case
in soot particle inception within the main reaction zone of a flame, the
MAEROS code requires fine size resolution (i.e., a large number of sec-
tions) to achieve satisfactory accuracy and thus becomes computer inten-
sive. Landgrebe and Pratsinis72 developed a modification of the MAEROS
M. FRENKLACH 145
algorithm based on flexible section width and selection of conserved prop-

erties. Although their code is significantly more efficient (a typical run
takes on the order of about an hour on a Macintosh II microcomputer),
it is still too computer intensive to be used in multidimensional reactive-
flow simulations.
The most computer-efficient approach is the method of moments. Equa-
tions (23) describe the evolution of an aerosol by solving for the distribution
function of the particle size, i.e., the population Nf over the entire range
of /. In the method of moments, this evolution is described by the moments
of the particle size distribution function,
fjir - M,/M0, r = 1, 2, ... (24)
where
•x.
M,. = ^ irN, (25)

/=!
is the rth concentration moment. The transformation of Eqs. (23) into the
moment equations,
——- = ~ - 2 Pi y-NiNj (26.0)

At 2/,/=i '
——- = 0 (mass conservation) (26.1)

at
- = 2 iJPijNiNj (26.2)
= 3 2 iffrjNM (26.3)
at jj= i
... etc. (26.r)
is achieved by the summations of Eqs. (23), similar to those described in
the previous section. For example, direct summation of Eqs. (23) results
in Eq. (26.0) for the zeroth moment, summation of Eqs. (23), each of the
equations multiplied by the corresponding /, results in Eq. (26.1) for the
first moment, summation of Eqs. (23) multiplied by i2 results in Eq. (26.2)
for the second moment, etc.
Integration of Eqs. (26), if possible, provides the values of the rth con-
centration moments, Mr, and hence, by Eq. (24), the rth moments of the
particle size distribution function, \Lr. The knowledge of all moments is
equivalent to knowing the distribution function itself. 73 In most practical
applications, however, the properties one considers are fully determined
by the first few moments, and, hence, only a small number of differential
equations in (26) must be solved. Thus, in the method of moments, the
history of individual particles is lost, but this loss is compensated for by a
significant increase in computational speed and a dramatic reduction in
computer memory requirements.
The method of chemical lumping described in the previous section is

also a method of moments. In that case, the summations led to a well-
defined, closed mathematical system: only integer-order moments appear
in the moment equations [Eqs. (16), (17), and (19)]. The main difficulty
encountered in solving Eqs. (26) is a rather complex mathematical form
of the collision coefficient. For instance, in the free-molecular regime,
assuming particles are spherical,68J39 the collision coefficient takes the form
(i^ + y1/3)2 (27)

J
where C is a physical constant,
C =
Here kB is the Boltzmann constant, T the temperature, and p the mass

density of the particles.
Several approaches have been proposed to solve Eqs. (26). The most
common one is to assume a specific functional form for the particle size
distribution function, i.e., the dependence of N, on i, to approximate the
discrete distribution by a continuous one and thus replace the summations
on the right-hand side of Eqs. (26) by equivalent integrals, and then to
solve the integrals (see, for example, Refs. 74-76). The principal limitation
of this approach is that the size distribution function is often unknown.
For instance, there is growing evidence that the distribution obtained with
a strong nucleation source present (as is the case in soot inception) is far
different from the usually assumed log-normal distribution (see, for exam-
ple, Refs. 70 and 72).
Two mathematical methods were proposed to solve Eqs. (26) without
prior assumption of the form of the distribution function. 70 In the first
method, Eqs. (26) are approximated by
(28.0)
(28.1)
dM2
= O>Aff (28.2)
df
~ = 3 <p)M1M2 (28.3)
dt
...etc. (28.r)
where
E Wi (29)
M. FRENKLACH 147
is the average collision coefficient. For (p) = const, i.e., if (p) is assumed
to be independent of the distribution and time, there exists an analytical
solution.77 Otherwise, for (p) ^ const, solutions of Eq. (29) can be found
for some limiting cases. Thus, for the free-molecular regime when p / y is
given by Eq. (27), two such solutions can be obtained:
<3>! = C[5.657 + (0.589a2/|ui?)]|x11/6 for D -> 1 (30)
and
<p> 2 - 6.998Qji{/6 for D =2 (31)
where D is the dispersion, defined as

D — |x2/fJi2
and cr2 = |UL2 - jo,2 is the variance. The first limit [Eq. (30)] corresponds
to a very narrow particle size distribution, which exists, for example, in
the beginning of the coagulation process of a monodispersed powder. Equa-
tion (31) is obtained in the limit of a self-preserving distribution, for which
D = 2. Linear interpolation between the two limits with respect to param-
eter D,
<p> = (2 - DXPX + (D - l)<p> 2
results in the expression
<P) - C(3.138 + 3.108D - 0.589D2)fju}/6 (32)
which should be applicable to the entire range 1 < D < 2. Integration of
Eqs. (28) with the average collision coefficient in Eq. (32) appears to be
very accurate and extremely economical for the case of coagulation without
a source. With a nucleation source present, the distribution quickly widens,
i.e., the value of D exceeds 2, and although it is still possible to develop
approximate expressions similar to Eq. (32), they are not reliable since the
limit of the particle size distribution for D —> *> is not well defined.
The second method of Frenklach and Harris70 uses interpolation to eval-
uate the summations on the right-hand side of Eqs. (26). As an example,
let us consider the summation in Eq. (26.2) for the free-molecular regime,
oc
X (i + y)1/2/1/2/1/2(z1/3 + y1'3)2*,-"/ (33)

'V = i
where particle number densities are expressed in a nondimensional form
Equation (33) is obtained upon substitution of Eq. (27) into Eq. (26.2)
and some simple algebraic rearrangements. Because of the (/ + y)1/2 term,
it is impossible to perform the summation in Eq. (33) directly. Nonetheless,
it can be evaluated in the following manner.
Let us define the following function of r:

oc
fr = S (' + /)r/1/2y'1/2(/1/3 + 71/3)2H,-«/ C34)

which will be referred to as a grid function. Its values for r = 0, 1, 2, ...
can be obtained by direct summation. For example, for r = 1, we obtain
A = 2 (' + J)limjl/2(i1/3 + /"3)2«,«y

oc
The fractional-order moments appearing in this equation are defined by

Eqs. (24) and (25) with the corresponding values of r. Their numerical
values are determined by interpolation between the integer-order moments
(i.e., |x0 = 1, JU4, |x2, ...), whose values are computed upon the integration
of Eqs. (26).
When the grid function is computed at several points—thus establishing
/0, /i,/ 2 , ...—the needed expression (33) is evaluated as its r = 1/2 point
by interpolation between the available integer-point values. For instance,
using Lagrange quadratic interpolation, the second-order approximation
for expression (33) is obtained as
f _ f 3/8f3/4f- 1/8
11/2 —JQJI 12
Other methods of interpolation are also possible (see, for example, Ref.
78).
The grid function [Eq. (34)] was constructed to evaluate the summation
in Eq. (26.2), which now takes the form
dM7 _ _
i , J L/Z (I
Similarly, grid functions for all other summations in Eqs. (26) can be
formulated. Using the appropriate interpolations, the system of the moment
equations (26) can then be solved. A detailed example is presented in Ref.
70.
VIII. Summary: Use of Reduction in Combustion Modeling

Whether they acknowledge it or not, most modelers use mechanism
reduction in daily practice. For instance, composing a chemical reaction
mechanism on the basis of including only those reactions that have been
proposed in the literature and whose rate coefficients are "known" is a
reduction strategy. In this approach, many plausible reactions are ignored
without first estimating their potential effects. Obviously, one cannot ana-
lyze the effect of a chemical reaction without assigning to it a concrete rate
coefficient value. It is the author's opinion that the inclusion of all poten-
M. FRENKLACH 149
tially important reactions with estimated rate parameters should be the

preferred methodology of assembling the first-trial mechanism over the
outright disregard of the "unknown." 23 Many estimation techniques are
available, and the uncertainty associated with the estimated values can and
should be tested by a sensitivity analysis. The tested reactions can possibly
be removed, but only following a rigorous reduction criterion. This approach
to reaction model development—construction of a large reaction set first,
followed by kinetic/sensitivity/reaction-path analysis, and then its reduction
to a suitable mathematical form—assures the correctness and, hence, use-
fulness of the modeling efforts.
Mechanism reduction thus constitutes an integral part of the model-
building process. Even in the case of exploratory modeling, when the
objective is to construct a mechanism, model reduction becomes an impor-
tant, often necessary step. One example of such a situation can be the
search for the reaction pathways leading to the formation and growth of
polycyclic aromatic hydrocarbons (see, for example, Ref. 7). In this mod-
eling study, the reaction set first assembled to test a large number of
molecular growth pathways for a given aromatic-ring size was reduced to
contain only several of the most prominent pathways prior to a similar
analysis for the next ring size.
In choosing a suitable method for model reduction, the modeler must
recognize, from the start, the objective of the overall modeling efforts and
the topology of the reaction network. It is unlikely that an all-purpose
reduction method can be found. Among those that exist, various individual
methods are well suited to certain applications, as discussed in this chapter,
but no single approach can satisfy all of the needs. For instance, the meth-
ods of global modeling can be effective for the dynamics of the main
reaction zone of a flame, but they are inefficient for application to coag-
ulation processes. In fact, many situations, such as soot formation, would
require that several reduction techniques be used concurrently.
Acknowledgments
The methods and ideas presented in this chapter were developed by the
author while working over the years on several projects supported by the
NASA Lewis Research Center under Contract NAS 3-23542 and Grants
NAG 3-477, NAG 3-668, NAG 3-991; the Gas Research Institute under
Contract 5086-260-1320; and the Air Force Office of Scientific Research
under Grant 88-0072.
References
^ran, E. S., and Boris, J. P., Numerical Solution of Reactive Flows, Elsevier,
New York, 1987.
2
Warnatz, J., "Chemistry of High Temperature Combustion of Alkanes up to
Octane," Twentieth Symposium (International) on Combustion, The Combustion
Institute, Pittsburgh, PA, 1985, pp. 845-856.
3
Westbrook, C. K., and Pitz, W. J., "Kinetic Modeling of Autoignition of Higher
Hydrocarbons: n-Heptane, n-Octane, andiso-Octane," Springer Series in Chemical
Physics: Complex Chemical Reaction Systems, Mathematical Modelling and Sim-
ulation, Vol. 47, edited by J. Warantz and W. Jager, Springer-Verlag, Berlin, 1987,
pp. 39-54.
4
Axelsson, E. I., Brezinsky, K., Dryer, F. L., Pitz, W. J., and Westbrook
C. K., "Chemical Kinetic Modeling of the Oxidation of Large Alkane Fuels: n-
Octane and iso-Octane," Twenty-First Symposium (International) on Combustion,
The Combustion Institute, Pittsburgh, PA, 1988, pp. 783-793.
5
Keller, J. O., and Westbrook, C. K., "Response of a Pulse Combustor to
Changes in Fuel Composition," Twenty-First Symposium (International) on Com-
bustion, The Combustion Institute, Pittsburgh, PA, 1988, pp. 547-555.
6
Thorne, L. R., Branch, M. C., Chandler, D. W., Kee, R. J., and Miller,
J. A., "Hydrocarbon/Nitric Oxide Interactions in Low-Pressure Flames," Twenty-
First Symposium (International) on Combustion, The Combustion Institute, Pitts-
burgh, PA, 1988, pp. 965-977.
7
Frenklach, M., Clary, D. W., Gardiner, W. C., Jr., and Stein, S. E., "Detailed
Kinetic Modeling of Soot Formation in Shock-Tube Pyrolysis of Acetylene," Twen-
tieth Symposium (International) on Combustion, The Combustion Institute, Pitts-
burgh, PA, 1985, pp. 887-901.
8
Frenklach, M., and Warnatz, J., "Detailed Modeling of PAH Profiles in a
Sooting Low-Pressure Acetylene Flame," Combustion Science and Technology,
Vol. 51, 1987, pp. 265-283.
9
Harris, S. J., Weiner, A. M., and Blint, R. J., "Formation of Small Aromatic
Molecules in a Sooting Ethylene Flame," Combustion and Flame, Vol. 72, 1988,
pp. 91-109.
10
Frenklach, M,, Yuan, T., and Ramachandra, M. K., "Soot Formation in Binary
Hydrocarbon Mixtures," Energy & Fuels, Vol. 2, 1988, pp. 462-480.
n
Espino, R. L., "Problem Solving by Computer Simulation," Chemical Engi-
neering Progess, Vol. 83, 1987, pp. 20-24.
12
Hill, C. G., Jr., An Introduction to Chemical Engineering Kinetics and Reactor
Design, Wiley, New York, 1977, pp. 92-93.
13
Semenov, N. N., Development of Theory of Chain Reactions and Thermal
Explosion, Znaniye, Moscow, 1969 (in Russian).
14
Wei, J., and Prater, C. D., "The Structure and Analysis of Complex Reaction
Systems," Advances in Catalysis, Vol. 13, 1962, pp. 203-392.
15
Prater, C. D., Silvestri, A. J., and Wei, J., "On the Structure and Analysis
of Complex Systems of First-Order Chemical Reactions Containing Irreversible
Steps. I. General Properties," Chemical Engineering Science, Vol. 22, 1967, pp.
1587-1606.
16
Silvestri, A. J., Prater, C. D., and Wei, J., "On the Structure and Analysis of
Complex Systems of First-Order Chemical Reactions Containing Irreversible Steps.
II. Projection Properties of the Characteristic Vectors," Chemical Engineering
Science, Vol. 22, 1968, pp. 1587-1606.
17
Wei, J., and Kuo, J. C. W., "A Lumping Analysis in Monomolecular Reaction
Systems. Analysis of Exactly Lumpable Systems," Industrial and Engineering
Chemistry Fundamentals, Vol. 8, 1969, pp. 114-123.
18
Kuo. J. C. W., and Wei, J., "A Lumping Analysis in Monomolecular Reaction
Systems. Analysis of Approximately Lumpable Systems," Industrial and Engi-
neering Chemistry Fundamentals, Vol. 8, 1969, pp. 124-133.
19
Dove, J. E., and Raynor, S., "Modelling Studies of Elementary Chemical
Reactions," Springer Series in Chemical Physics: Complex Chemical Reaction Sys-
tems, Mathematical Modelling and Simulation, Vol. 47, edited by J. Warnatz and
W. Jager, Springer-Verlag, Berlin, 1987, pp. 90-99.
20
Box, G. E. P., and Hunter, W. G., "The Experimental Study of Physical
Mechanisms," Technometrics, Vol. 7, 1965, pp. 23-42.
21
Box, G. E. P., Hunter, W. G., and Hunter, J. S., Statistics for Experimenters.
An Introduction to Design, Data Analysis, and Model Building, Wiley, New York,
1978.
M. FRENKLACH 151
22
C6me, G. M., "The Use of Computers in the Analysis and Simulation of
Complex Reactions," Comprehensive Chemical Kinetics: Modern Methods in Kinet-
ics, Vol. 24, edited by C. H. Bamford and C. F. H. Tipper, Elsevier, New York,
1983, pp. 325-332.
23
Frenklach, M., "Modeling," Combustion Chemistry, edited by W. C. Gardiner,
Jr., Springer-Verlag, New York, 1984, Chap. 7.
24
Smarr, L. L., "An Approach to Complexity: Numerical Computations," Sci-
ence, Vol. 228, 1985, pp. 403-408.
25
Hautman, D. J., Dryer, F. L., Schug, K. P., and Glassman, L, "A Multiple-
Step Overall Kinetic Mechanism for the Oxidation of Hydrocarbons," Combustion
Science and Technology, Vol. 25, 1981, pp. 219-235.
26
Duterque, J., Borghi, R., and Tichtinsky, H., "Study of Quasi-Global Schemes
for Hydrocarbon Combustion," Combustion Science and Technology, Vol. 26,
1981, pp. 1-15.
27
Taki, S., and Fujiwara, T., "Numerical Simulation of Triple Shock Behavior
of Gaseous Detonation," Eighteenth Symposium (International) on Combustion,
28
Yetter, R. A., Dryer, F. L., and Rabitz, H., "Complications of One-Step
Kinetics for Moist CO Oxidation," Twenty-First Symposium (International) on
Combustion, The Combustion Institute, Pittsburgh, PA, 1988, pp. 749-760.
29
Peters, N., "Systematic Reduction of Flame Kinetics: Principles and Details,"
Progress in Astronautics and Aeronautics Series: Dynamics of Reactive Systems.
Part I: Flames, Vol. 113, edited by A. L. Kuhl, J. R. Bowen, J.-C. Leyer, and
A. Borisov, AIAA, Washington, DC, 1988, pp. 67-86.
30
Peters, N., and Williams, F. A., "The Systematic Structure of Methane Flames.
Part I: Stoichiometric Flames," Springer Series in Chemical Physics: Complex Chemical
Reaction Systems, Mathematical Modelling and Simulation, Vol. 47, edited by
J. Warnatz and W. Jager, Springer-Verlag, Berlin, 1987, pp. 310-317.
31
Chen, J.-Y., "A General Procedure for Constructing Reduced Reaction Mech-
anisms with Given Independent Relations," Combustion Science and Technology,
Vol. 57, 1988, pp. 89-94.
32
Frenklach, M., and Bornside, D. L., "Shock-Initiated Ignition in Methane-
Propane Mixtures," Combustion and Flame, Vol. 56, 1984, pp. 1-27.
33
Jones, W. P., and Lindstedt, R. P., "Global Reaction Schemes for Hydrocarbon
Combustion," Combustion and Flame, Vol. 73, 1988, pp. 233-249.
34
Law, R., Metghalchi, M., and Keck, J. C., "Rate-Controlled Constrained
Equilibrium Calculation of Ignition Delay Times in Hydrogen-Oxygen Mixtures,"
Twenty-Second Symposium (International) on Combustion, The Combustion In-
35
Paczko, G., Lefdal, P. M., and Peters, N., "Reduced Reaction Schemes
for Methane, Methanol and Propane Flames." Twenty-First Symposium (Interna-
tional) on Combustion, The Combustion Institute, Pittsburgh, PA, 1988, pp.
739-748.
36
Rogg, B., "On Numerical Analysis of Two-Dimensional, Axisymmetric, Lam-
inar Jet Diffusion Flames," Mathematical Modeling in Combustion and Related
Topics, edited by C.-M. Brauner and C. Schmidt-Laine, Dordrecht, Boston, MA,
1988, pp. 551-560; also, "Response and Flamelet Structure of Stretched Premixed
Methane-Air Flames," Combustion and Flame, Vol. 73, 1988, pp. 45-65.
37
Rogg, B., and Williams, F. A., "Structures of Wet CO Flames with Full and
Reduced Kinetic Mechanisms," Twenty-Second Symposium (International) on
38
Tam, R. Y., andLudford, G. S. S., "Kinetic Extinction: A Three-Step Model,"
Combustion and Flame, Vol. 72, 1988, pp. 27-34; also, "The Lean Flammability
Limit: A Four-Step Model," Combustion and Flame, Vol. 72, 1988, pp. 35-43.
39
Liu, Y. A., and Lapidus, L., "Observer Theory for Lumping Analysis of
Monomolecular Reaction Systems," AIChE Journal, Vol. 19, 1973, pp. 467-473.
40
Li, G., "A Lumping Analysis in Mono- or/and Bimolecular Reaction Systems,"
Chemical Engineering Science, Vol. 29, 1984, pp. 1261-1270.
41
Ho, T. C., and Aris, R., "On Apparent Second-Order Kinetics," AICHE
Journal, Vol. 33, 1987, pp. 1050-1051.
42
Astarita, G., and Ocone, R., "Lumping Nonlinear Kinetics," AIChE Journal,
Vol. 34, 1988, pp. 1299-1309.
43
Chou, M. Y., and Ho, T. C., "Continuum Theory for Lumping Nonlinear
Reactions," AlChE Journal, Vol. 34, 1988, pp. 1519-1527.
44
Lam, S. H., and Goussis, D. A., "Basic Theory and Demonstration of Com-
putational Singular Perturbation for Stiff Equations," Proceedings of 12th IMACS
World Congress on Scientific Computations, Vol. 1, edited by R. Vichnevetsky, P.
Borne, and J. Vignes, Gerfidn, Cedex, France, 1988, pp. 187-190.
45
Lam, S. H., and Goussis, D. A., "Understanding Complex Chemical Kinetics
with Computational Singular Perturbation," Twenty-Second Symposium (Inter-
national) on Combustion, The Combustion Institute, Pittsburgh, PA, 1989, pp.
931-941.
46
Li, G., and Rabitz, H., "A General Analysis of Exact Lumping in Chemical
Kinetics," Chemical Engineering Science, Vol. 44, 1989, pp. 1413-1430; also "A
General Analysis of Approximate Lumping in Chemical Kinetics" (to be pub-
lished).
47
Hutchinson, P., and Luss, D., "Lumping of Mixtures with Many Parallel First
Order Reactions," Chemical Engineering Journal, Vol. 1, 1970, pp. 129-135.
48
Luss, D., and Hutchinson, P., "Lumping of Mixtures with Many Parallel Mh
Order Reactions," Chemical Engineering Journal, Vol. 2, 1971, pp. 172-177.
49
Bailey, J. E., "Lumping Analysis of Reactions in Continuous Mixtures," Chem-
ical Engineering Journal, Vol. 3, 1972, pp. 52-61.
50
Frenklach, M., "Computer Modeling of Infinite Reactions Sequences: A Chem-
ical Lumping," Chemical Engineering Science, Vol. 40, 1985, pp. 1843-1849.
51
Oran, E. S., Boris, J. P., Young, T. R., Flanigan, M., Burks, T., and Picone,
M., "Numerical Simulations of Detonations in Hydrogen-Air and Methane-Air
Mixtures," Eighteenth Symposium (International) on Combustion, The Combustion
52
Coffee, T. P., Kotlar, A. J., and Miller, M. S., "The Overall Reaction Concept
in Premixed, Laminar, Steady-State Flames. I. Stoichiometries," Combustion and
Flame, Vol. 54, 1983, pp. 155-169.
53
in Premixed, Laminar, Steady-State Flames. II. Initial Temperatures and Pres-
sures," Combustion and Flame, Vol. 58, 1984, pp. 59-67.
54
Frenklach, M., "Modeling of Large Reaction Systems," Springer Series in
Chemical Physics: Complex Chemical Reaction Systems, Mathematical Modelling
and Simulation, Vol. 47, edited by J. Warnatz and W. Jager, Springer-Verlag,
Berlin, 1987, pp. 2-16.
55
Box, G. E. P., and Draper, N. R., Empirical Model-Building and Response
Surfaces, Wiley, New York, 1987.
56
Myers, R. H., Khuri, A. L, and Carter, W. H., Jr., "Response Surface Meth-
odology: 1966-1988," Technometrics, Vol. 31, 1989, pp. 137-157.
57
Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P., "Design and Analysis
of Computer Experiments," Statistical Science, Vol. 4, 1989, pp. 409-435.
58
Sacks, J., Schiller, S. B., and Welch, W. J., "Designs for Computer Experi-
ments," Technometrics, Vol. 31, 1989, pp. 41-47.
M. FRENKLACH 153
59
Miller, D., and Frenklach, M., "Sensitivity Analysis and Parameter Estimation
in Dynamic Modeling of Chemical Kinetics," International Journal of Chemical
Kinetics, Vol. 15, 1983, pp. 677-696.
60
Frenklach, M., "Systematic Optimization of a Detailed Kinetic Model Using
a Methane Ignition Example," Combustion and Flame, Vol. 58, 1984, pp. 69-72.
61
Frenklach, M., and Miller, D., "Statistically Rigorous Parameter Estimation
in Dynamic Modeling Using Approximate Empirical Models," AIChE Journal,
Vol. 31, 1985, pp. 498-500.
62
Frenklach, M., and Rabinowitz, M. J., "Optimization of Large Reaction Sys-
tems," Proceedings of 12th IMACS World Congress on Scientific Computations,
Vol. 3, edited by R. Vichnevetsky, P. Borne, and J. Vignes, Gerfidn, Cedex,
France, 1988, pp. 602-604.
63
Marsden, A. R., Jr., Frenklach, M., and Reible, D. D., "Increasing the Com-
putational Feasibility of Urban Air Quality Models that Employ Complex Chemical
Mechanisms," Journal of Air Pollution Control Association, Vol. 37, 1987, pp.
370-376.
64
Frenklach, M., Kailasanath, K., and Oran, E. S., "Systematic Development
of Reduced Reaction Mechanisms for Dynamic Modeling," Dynamics of Reactive
Systems Part II: Modeling and Heterogeneous Combustion, Vol. 105, edited by J.
R. Bowen, J.-C. Leyer, and R. I. Soloukhin, Progress in Astronautics and Aer-
onautics Series, AIAA, Washington, DC, 1986, pp. 365-376.
65
Frenklach, M., and Gardiner, W. C., Jr., "Representation of Homogeneous
Polymerization in Detailed Computer Modeling of Chemical Kinetics," Journal of
Physical Chemistry, Vol. 88, 1984, pp. 6263-6266.
66
Frenklach, M., "On the Driving Force of PAH Production," Twenty-Second
Symposium (International) on Combustion, The Combustion Institute, Pittsburgh,
PA, 1989, pp. 1075-1082.
67
Johnson, N. L., and Leone, F. C., Statistics and Experimental Design in En-
gineering and the Physical Sciences, Vol. 1, Wiley, New York, 1977, Chap. 3.
68
Seinfeld, J. H., Atmospheric Chemistry and Physics of Air Pollution, Wiley,
New York, 1986, Chap. 10.
69
Friedlander, S. K., Smoke, Dust and Haze: Fundamentals of Aerosol Behavior,
Wiley, New York, 1977, Chap. 7.
70
Frenklach, M., and Harris, S. J., "Aerosol Dynamics Modeling Using the
Method of Moments, Journal of Colloid and Interface Science, Vol. 118, 1987, pp.
252-261.
71
Warren, D. R., and Seinfeld, J. H., "Simulation of Aerosol Size Distribution
Evolution in Systems with Simultaneous Nucleation, Condensation, and Coagu-
lation," Aerosol Science and Technology, Vol. 4, 1985. pp. 31-43.
72
Landgrebe, J., and Pratsinis, S. E., "Gas-Phase Manufacture of Particulates:
Interplay of Chemical Reaction and Aerosol Coagulation in the Free-Molecular
Regime," Industrial and Engineering Chemistry Research, Vol. 28, 1989, pp. 1474-
1481.
73
Hudson, D. J., Lectures on Elementary Statistics and Probability, CERN, Ge-
neva, 1963, Chap. 1.
74
Dobbins, R. A., and Mulholland, G. W., "Interpretation of Optical Measure-
ments of Flame Generated Particles," Combustion Science and Technology, Vol.
40, 1984, pp. 175-191.
75
Seigneur, C., Hudischewskyj, A. B., Seinfeld, J. H., Whitby, K. H., Whitby,
E. R., Brock, J. R., and Barnes, H. M., "Simulation of Aerosol Dynamics: A
Comparative Review of Mathematical Models," Aerosol Science and Technology,
Vol. 5, 1986, pp. 205-222.
76
Pratsinis, S. E., "Simultaneous Nucleation, Condensation, and Coagulation
in Aerosol Reactors," Journal of Colloid and Interface Science, Vol. 124, 1988,
pp.416-427.
77
Frenklach, M., "Dynamics of Discrete Distribution for Smoluchowski Coag-
ulation Model," Journal of Colloid and Interface Science, Vol. 108, 1985, pp. 237-
242.
78
Deuflhard, P., and Wulkow, M., "Computational Treatment of Polyreaction
Kinetics by Orthogonal Polynomials of a Discrete Variable, Impact of Computing
in Science and Engineering, Vol. 1, 1989, pp. 269-301.
Background
This second section focuses on modeling the dynamics and structure of
premixed and diffusion flames. In subsonic flows with reaction, the mo-
lecular and energy-diffusion processes are as important as any of the fluid
effects in controlling the behavior of the reactions. Having to represent all
of these processes makes flames difficult to model quantitatively, and this
is particularly true as the flow becomes faster, fluid dynamically unstable,
and eventually turbulent.
In the first chapter of this section, Peters summarizes the various regimes
of laminar and turbulent flames. Because there are so many potentially
important physical processes involved, there are a number of time and
length scales that are useful in describing the system. Figure 8 of Peters'
chapter describes the important velocity and length-scale ratios and the
related nondimensional Reynolds, Karlovitz, and Damkohler numbers. His
classification of these provides useful background and perspective on the
next five chapters dealing with flames simulated by using various numerical
methods.
Smooke has created an extensive body of work including new numerical
methods and their application to modeling the structure of laminar, steady-
state diffusion flames with complex chemistry. In the second chapter, Smooke
describes the application of his adaptive boundary-value methods to two
types of steady-state laminar methane diffusion flames: the counterflow
flame and the axisymmetric diffusion flame. He discusses the numerical
solution of the governing equations and compares his predictions to ex-
perimental data for both configurations considered. As he points out, there
is general agreement between the computations and the data. Where there
is disagreement, the reader sees two of the important uses of numerical
simulations: to determine when important physical processes actually are
missing from the mathematical model and to show perhaps unsuspected
problems with experiments. Steady-state problems, such as those described
in this chapter, can often be solved with rather specialized methods that
work because all of the physical and chemical transients are assumed to
be absent.
Kailasanath then presents a short history of modeling laminar flames
using time-dependent methods. He reports highlights of a body of work
that started over ten years ago to develop time-dependent models of lam-
inar flames, and describes selected studies of one-dimensional and two-
dimensional time-dependent laminar premixed flames using detailed hy-
drogen-oxygen chemistries and the full set of diffusive transport properties.
Making the transition from one dimension to two dimensions forced an
important change in the philosophy of solving the fluid-dynamic parts of
the problem: changing the convective algorithms from a Lagrangian to an
Eulerian representation. Initially only a Lagrangian method was thought
to be accurate enough to resolve the physical diffusion effects in flames.
However, Lagrangian methods proved to be extremely difficult to imple-
ment accurately enough in multidimensions. This promoted the develop-
ment of very accurate implicit Eulerian algorithms. This new method,
developed by Patnaik and described briefly in this chapter, is the basis of
the two-dimensional simulations of unstable premixed flames. Finally, in
the context of time-dependent premixed flames, it is important to direct
the reader to the pioneering work by Dixon-Lewis,1 the work of Takano,2
and the recent finite-element simulations by Larrouturou and coworkers.3
The natural progression from here is to consider transitional and low-
Reynolds-number turbulent flames. In these problems, the flowfields are
extremely complex and expensive to compute. The models of chemistry
and diffusion must necessarily be simplified correspondingly. Fortunately,
the fluid dynamics "cooperates" in the sense that convection begins to
become the most important as a macroscopic mixing mechanism, whereas
in the laminar flames described by Smooke and Kailasanath the various
physical diffusion processes were dominant. However, these diffusion proc-
esses still control the small-scale molecular mixing on which the chemical
reactions depend.
This section presents two quite different numerical methods for describ-
ing the flowfield. McMurtry and Givi use a spectral method, in which
spatial variations are described by an expansion in wave number space.
Ghoniem uses vortex dynamics, a Lagrangian method that transports vor-
ticity packets. Both approaches solve for the large-scale time-dependent
coherent structures in the transitional flowfield. McMurtry and Givi give
an excellent summary of the work to date on chemically reacting flows
computed with spectral methods, and then give examples from their own
recent work. Ghoniem's work is a tour de force, showing how a funda-
mentally incompressible Mach-number-zero method can be transformed
into a description of a general compressible flow.
The section concludes with a chapter by Pope on combustion modeling
of flows in which the small-scale structure of the flames and species fields
cannot be resolved. Pope uses Probability Density Functions (PDF) to
convey the statistical aspects of the unresolved structure of the reactive
flow. PDF methods have been applied to a variety of turbulent-flow prob-
lems. The idea here is to derive a transport equation for the important
PDFs in a problem and to solve this equation by a suitable numerical
method. However, the method bears some philosophical points in common
with more standard moment methods for describing the closures required
by the Navier-Stokes equations. In this case, all of the moment information
is contained in the PDFs. Pope summarizes the use of Monte Carlo methods
to solve the PDF transport equations and gives an example of the appli-
cation in a chemically reacting flow.
Missing from this section are methods that use finite-difference or finite-
volume algorithms for simulating subsonic, transitional, premixed and dif-
fusion flames. This obvious exclusion appears unwise: because of their
basic simplicity and flexibility, these types of method are really the most
prevalent for describing such difficult problems. Many examples of appli-
cations of these methods can be found now in the proceedings of various
combustion meetings. To see representative applications of these methods
to subsonic reacting flows, the reader may examine the examples in a
previous book on numerical methods for reacting flows.4 More specifically,
the reader should explore the recent works by Kailasanath and coworkers5
and Jou and coworkers6 on acoustic-chemical-vortex coupling in combus-
tors. In addition, many of the chapters included in Parts 3 and 4 use finite-
difference and finite-volume methods.
îxon-Lewis, G., and Shepherd, I. G., ''Some Aspects of Ignition by Localized Sources
and of Cylindrical and Spherical Flames," Fifteenth Symposium (International) on Combus-
tion, The Combustion Institute, Pittsburgh, PA, 1975, pp. 1483-1491; also, Dixon-Lewis,
G., "Structure and Extinction Limits of Some Strained Premixed Flames," Dynamics of
Reactive Systems, Part I: Flames," edited by A. L. Kuhl, J. R. Bowen, J.-C. Leyer, and A.
Borisov, Vol. 113, Progress in Astronautics and Aeronautics, AIAA, Washington, DC, 1988.
2
Takano, Y., "Flame Propagation Model by Use of Finite-Difference Methods," Dynamics
of Reactive Systems, Part I: Flames, edited by A. L., Kuhl, J. R. Bowen, J.-C. Leyer, and
A. Borisov, Vol. 113, Progress in Astronautics and Aeronautics, AIAA, Washington, DC,
1988, pp. 275-288; also, Takano, Y., "Finite-Difference Simulations of Flame-Induced Flows
in Closed Vessels by Use of Deflagration Front Model," 12th International Colloquium on
the Dynamics of Explosions and Reactive Systems, Ann Arbor, MI, 1989.
3
Benkhaldoun, F., and Larrouturou, B., "Explicit Adaptive Calculations of Wrinkled
Flame Propagation," International Journal of Numerical Methods of Fluids, Vol. 7, 1987, pp.
1147-1158; also, Dervieux, A., Larrouturou, B., and Peyret, R., "On Some Adaptive Nu-
merical Approaches of Thin Flame Propagation Problems, Computational Fluids, Vol. 17,
pp. 39-60; also, Benkhaldoun F., Denet, B., and Larrouturou, B., "Numerical Investigation
of the Extinction Limit of Wrinkled Flames," Combustion Science and Technology, Vol. 64,
1989, pp. 187-198; also, Benkhaldoun, F., and Larrouturou, B., "A Finite-Element Adaptive
Investigation of Curved Stable and Unstable Flame Front," Computational Methods of Ap-
plied Mechanical Engineering, Vol. 76, 1989, pp. 119-134.
4
Oran, E. S., and Boris, J. P., Numerical Simulation of Reacting Flows, Elsevier, New
York, 1987.
5
Kailasanath, K., Gardner, J. H., Boris, J. P., and Oran, E. S., "Acoustic-Vortex Inter-
actions and Low-Frequency Oscillation in Axisymmetric Combustors," Journal of Propulsion
and Power, Vol. 5, 1989, pp. 165-171; also, Kailasanath, K., Gardner, J. H., Oran, E. S.,
and Boris, J. P., "Effects of Energy Release on High-Speed Flows in an Axisymmetric
Combustor," AIAA Paper 89-0385.
6
Menon, S., and Jou, W., "Simulations of Acoustic-Entropy Wave Interactions and Com-
bustion Instability," AIAA Paper 90-0267.
Chapter 6
Length Scales in Laminar and Turbulent Flames
Norbert Peters
I. Introduction
N UMERICAL computations of combustion processes with fine reso-

lution of flame structures have become feasible in recent years. Not
only have larger and faster computers become available, but our knowledge
of the detailed chemistry that governs the combustion process has also
improved, and—most important—the understanding of the interaction of
chemistry with turbulence has progressed. Direct numerical simulations of
chemically reacting flows have contributed considerably to this progress of
understanding. Another major factor is the identification of different re-
gimes of turbulent combustion.
From a more global point of view, the interaction of turbulence and
chemistry may be classified by two criteria: premixed or nonpremixed
combustion, and slow or fast chemistry. The first criterion is relevant with
respect to different applications. Reactive flows in large furnaces and gas
turbines essentially are nonpremixed, whereas combustion in the Otto-type
reciprocating engine or the reaction of pollutants in the atmosphere occurs
in the premixed regime. The second criterion describes the ratio of chemical
to convective and diffusive time scales. Slow chemistry is seldom of practical
interest. There are a few applications, such as combustion of highly diluted
reactants in postflame regions, where chemistry is slow compared with
convection and diffusion. Conversely, reacting flows with fast chemistry
occurs in nearly all types of engines and combustion devices. Therefore,
this regime presents a major challenge to the modeler. Since there is a vast
disparity between the short chemical time scales and the very large fluid
dynamical time scales, chemical activity is confined to thin layers, which
now are called "flamelets." Flamelets in a turbulent flowfield are thin by
definition, their structure is essentially one-dimensional normal to the layer
and may be resolved by well-developed numerical or asymptotic methods.
Therefore, in the flamelet regime, chemistry and turbulence may—in prin-
ciple—be treated separately.
reserved.
155
156 LENGTH SCALES IN LAMINAR AND TURBULENT FLAMES
Among the four regimes covered by the two criteria stated earlier, the
regimes of premixed and nonpremixed reactants with slow chemistry may
be treated as a perturbation of a nonreacting turbulent flow by including
equations for the reacting scalars. One rather successful approach in this
direction relies on the formulation of a transport equation for the joint
probability density function (pdf) of the reacting scalars. In principle, this
pdf could be calculated at each location within the flowfield on the basis
of a pdf transport equation. Although this approach avoids some of the
modeling assumptions used in moment methods and, therefore, should
yield more general results, it still requires modeling some of the most
important terms, particularly the fluctuating pressure gradient term and
the molecular diffusion term. A recent review on pdf methods for turbulent
reactive flow has been given by Pope,1 and he describes the recent progress
in a subsequent chapter of this book.
The third regime, premixed combustion with fast chemistry, is the regime
where the most interaction between turbulence and chemistry occurs. It is
perhaps the most difficult regime to model numerically. The different re-
gimes of premixed turbulent combustion have been discussed in Ref. 2.
Since the flame moves locally normal to itself with a characteristic velocity,
the laminar burning velocity, a velocity scale is defined. The comparison
of this velocity scale with the range of velocities in a turbulent flowfield
leads in the flamelet regime to a particular interaction scale L G , the Gibson
scale. Another interaction scale, the quench scale 8^, which is relevant in
the distributed reaction zones regime, will also be derived.
The fourth regime, nonpremixed combustion with fast chemistry, has
received much attention in the combustion literature. The flamelet ap-
proach has been very successful in describing this regime.3 However, re-
cently Bilger4 has argued that flamelets do not exist in turbulent jet diffusion
flames, particularly in flames of hydrogen in air. Therefore, it seems ap-
propriate to reconsider this important regime by comparing the relevant
time scales of turbulence and combustion. This will lead us to a phase
diagram comparable to the one presented for premixed turbulent com-
bustion in Ref. 2.
II. Length Scales in Laminar Flames

In this chapter, we will present the characteristic scales for premixed
and nonpremixed flames. Before discussing the complex interactions in
turbulent combustion, it will be worthwhile to consider the situation in
laminar flames. We will review characteristic length scales, time scales,
and velocity scales that can be compared with those in turbulent flow.
A. Premixed Flames
An important property of a premixed flame is its ability to propagate
normal to itself. This leads to a characteristic velocity scale—the burning
velocity. Flame propagation is caused by a combination of chemical, dif-
fusive, and convective effects, as illustrated in Fig. 1 by considering the
structure of a plane, steady laminar hydrocarbon flame. It consists of a
chemically inert upstream preheat zone, a thin inner reaction layer, and a
N. PETERS 157
iliU
•II- oxidation layer
0(e)
0 x —*-
Fig. 1 Schematic illustration of the structure of a premixed methane-air flame.
downstream oxidation layer. The details of the structure depend on the

kinetic model used. However, from a recent asymptotic analysis of methane
flames,5 it now seems clear that there always exists a thin reaction zone,
at least in hydrocarbon flames, comprising an inner layer and an oxidation
layer. To obtain a steady configuration, the laminar burning velocity SL
relative to the unburnt gas of the flame illustrated in Fig. 1 must be equal
to the velocity vu of the oncoming flow in the unburnt gas (indexed by
subscript u in the text). The heat generated in the thin reaction zone diffuses
against the flow into the preheat zone. There is a reactive-diffusive balance
in the reaction zone and a convective-diffusive balance in the preheat zone.
The chemical reaction provides a time scale, the chemical time tc, which
is a characteristic measure of the composite time required for the various
elementary reactions to transform the reactants into products within the
reaction zone. The burning velocity and the chemical time scale are linked
to each other by the heat conduction process. Therefore, from dimensional
analysis,
SL = V\/(pcp Q (1)
where p is the density, X the thermal conductivity, and cp the heat capacity
at constant pressure. Equation (1) defines the chemical time /c, given the
burning velocity and the conditions at which the thermodynamical prop-
erties are to be evaluated. Since the temperature varies considerably within
the flame structure, and the mass flow rate pusL through the flame is
constant, a convenient nondimensional coordinate normal to the flame is5
s (*EE dc (2)
Jo X
To define a flame thickness that is consistent with this definition, it is

necessary to evaluate \lcp at a well-defined reference temperature Tref and
p at the unburnt gas temperature Tu. The reference temperature should
be that of the inner reaction layer, which occurs typically between 1600
and 2000 K for atmospheric hydrocarbon flames in air. Then the flame
thickness can be defined by considering a length AJC* equal to unity in the
nondimensional coordinate. Setting the left-hand side of Eq. (2) equal to
unity and cpl\ constant and equal to the reference value, the flame thickness
€ F is
*F ^ (Mc,) re f/Pu SL (3)
A characteristic time for the flame propagation process is the time required
for the flame to traverse its own thickness, called the "flame time," and
equals
tF = tF/sL (4)
In view of Eqs. (1), (3), and (4), the flame time îs equal to tc. Therefore,
it has the same physical significance as the chemical time.
The flame thickness defined by Eq. (3) is a measure of the width of the
preheat zone of a premixed flame. Typical values of (F for atmospheric
stoichiometric methane-air flames lie around 0.175 mm when (X/c p ) ref /p M
is estimated as 7 x 10~5 m2/s and SL as 0.4 m/s. The corresponding flame
times is 0.4375 x 10~3 s. As the pressure p increases, the density increases
linearly with pressure, whereas \/cp is independent of pressure for ideal
gases. However, the burning velocity also changes with pressure. For hy-
drocarbon flames, the burning velocity decreases as p~", where the ex-
ponent n is around 0.5. Therefore, the flame thickness varies with pressure
as pn~l.
For numerical calculations, the reaction zone itself must be resolved.
The zone extends only over a fraction of the premixed flame structure. To
compare the reaction zone thickness with the preheat zone thickness, re-
sults from Ref. 5 can be used. The thickness e of the downstream oxidation
layer was measured in the nondimensional coordinate x* of Eq. (2). Values
of e are shown in Fig. 2 in a phase plot for different pressures and preheat
temperatures. Introducing e = (A**)/? into Eq. (2), one realizes that the
thickness of the oxidation layer £RO = (&x)R in physical space is
<R,O = e£F (5)
However, the oxidation layer is not the thinnest layer that must be resolved.
Reactions involving hydrocarbon species occur essentially in the inner layer
of order 8 within the flame structure, which, in physical space, results in
an inner layer thickness
**,/ = 8€F (6)
Values of 8 from Ref. 5 for the same conditions are also reproduced in
Fig. 2. As a consequence of these estimates, on the order of at least 20
unequally spaced grid points are necessary to resolve accurately the struc-
N. PETERS 159
Tu [K] 700
500
10 15 20 25 30
—— p[atm]
T U [ K ] 700
30
—— plafrn]
Fig. 2 Phase plots for the reaction layer thickness e and 8 of a premixed methane-
air flame as a function of preheat temperature Tu and pressure p.
ture of a laminar premixed flame. This estimate increases by a substantial

factor for increasing pressures as 8 decreases, but decreases slightly for
elevated preheat temperatures.
In addition, premixed flames may extinguish due to flame stretch. Flame
stretch, which was introduced by Karlovitz et al., 6 is the local fractional
increase of flame surface area per unit time. Local extinction of a premixed
flame causes the reactants and the burnt products to interdiffuse and,
thereby, broadens the flame structure locally. The characteristic time tq
for a premixed flame to extinguish due to the stretch, called the "quench
time," may be defined as the inverse of the velocity gradient aq at extinc-

tion, calculated for a plane flame in a divergent flowfield
tq = a,"1 (7)
7
Numerical calculations for aq by Rogg for a stoichiometric methane-air
flame show a velocity gradient of aq = 2275/s at extinction, which corre-
sponds to a quench time tq of 0.44 x 10~3 s. This differs very little from
the previous estimate given for the flame time tF. Therefore, we will ap-
proximate tq by tF for premixed flames.
B. Diffusion Flames
If the reactants — fuel and oxidizer — are fed separately into a combus-
tion system, they diffuse toward each other and burn in a flame structure
called a "diffusion flame." Combustion occurs preferentially at those lo-
cations in the flowfield where mixing is stoichiometric. The global reaction
equation for complete combustion of a hydrocarbon fuel F, written as
v'F F + v'02 02 -» v£02 C02 + v^2o H20 (8)
defines the stoichiometric coefficients v'02 and v'F. The reaction equation
relates the change of mass fractions of oxygen dY0, and fuel dYF that are
consumed as
dY02/v'02 M02 = dYF/v'F MF (9)
where Mi is the molecular weight. This equation may be integrated from
the unburnt state to any later state as
vYF - Y02 = vYF,u - Y02M (10)
where v = v'02 M02/vF MF is the stoichiometric mass ratio. Away from the
stoichiometric location, the mixture is either fuel lean or fuel rich and,
therefore, leaves either some oxygen or the fuel (which may partially be
oxidized to CO and H2) unreacted. Therefore, it cannot produce a high
enough temperature to maintain a reaction rate, since combustion chem-
istry is very temperature sensitive.
To describe the mixture field and to identify the location of the stoi-
chiometric mixture, it is useful to introduce the mixture fraction Z as a
dependent variable. In a system of only two streams, where subscript 1
denotes the fuel stream ml and subscript 2 the oxidizer stream ra2, Z
represents the local mass fraction of the fuel stream in the unburnt mixture,
Z = m<J(rhi + ra2) (11)
Since both the fuel and oxidizer streams may contain inerts such as nitrogen,
the local mass fraction YFu of the fuel is the same fraction as in the original
fuel steam (if effects of differential diffusion are neglected),
YF,U = YF9lZ (12)
N. PETERS 161
where YF1 denotes the mass fraction of fuel in the fuel stream. Similarly,
since 1 — Z represents the local mass fraction of the oxidizer stream in
the unburnt mixture, for the local mass fraction of oxygen one obtains
Y02.u = Y02a (1 - Z) (13)
where Y02 2 represents the mass fraction of oxygen in the oxidizer stream
(y02,2 = 0.232 for air). Introducing Eqs. (12) and (13) into Eq. (10), the
mixture fraction is
Z = (vYF - Y02 + Y02,2)/(vYFA + Y02,2) (14)
For a stoichiometric mixture with vYF = Y02,
Zst = (1 + vYFfl/Y02J-> (15)
where Zst is the stoichiometric mixture fraction.
Laminar diffusion flames may occur in the counterflow boundary layer
of air flowing around a porous cylinder from which fuel is injected into
the airstream.8 Experimental studies in counterflow diffusion flames have
been reviewed by Tsuji. 9 The profiles of concentrations and temperature
in a diffusion flame may be represented in two ways: in the physical space
coordinate y, or in the mixture fraction as a scalar coordinate. This is
demonstrated in Fig. 3 for calculated temperature profiles taken from Ref.
10.
The mixture fraction appears natural as an independent variable for
diffusion flames. Under the condition that equal diffusivities of chemical
species and temperature can be assumed (an assumption that is good for
hydrocarbon flames but much less realistic for hydrogen flames), all Lewis
numbers
Lei = X/cppDh i = 1, 2, ..., n (16)
are unity, such that a common diffusivity coefficient D can be introduced.
The balance equation for Z and the temperature T are
p
2 M-, (18)
Here ht are the specific heats and ra, the chemical production rates of the
reacting species (/ = 1, 2, ..., n). The specific heat capacities cpi are all
assumed constant and equal to cp for simplicity. Equation (17) does not
contain a chemical source term, since Z represents the chemical elements
originally contained in the fuel, and elements are conserved during com-
bustion. We assume the mixture fraction Z to be given in the flowfield as
2500
0.00 0.04 008 0.12 0.16 0.20 0.4 0.6 0.8 1.0
Fig. 3 Temperature profiles of a methane-air diffusion flame in the physical co-

ordinate v and the mixture fraction coordinate Z.
N. PETERS 163
a function of space and time by solution of Eq. (16). Then the surface of
the stoichiometric mixture can be determined from
Z(xa9t) = Zst (19)
Combustion occurs in a thin layer in the vicinity of this surface if the local
mixture fraction gradient is sufficiently high. Let us locally introduce a
coordinate system attached to the surface of a stoichiometric mixture. We
replace the coordinate xl by the mixture fraction Z and define the original
coordinate system such that the coordinate xl does not lie within this
surface. This is a coordinate transformation of the Crocco type. (Crocco
expressed the temperature as a function of another dependent variable,
the velocity, in a flat-plate boundary layer.) Here the temperature Twill
be expressed as a function of the mixture fraction Z. By definition, the
new coordinate, Z, is locally normal to the surface of the stoichiometric
mixture. Using Z2 = x2, Z3 = Jt3, t* = t as the other independent variables,
with the transformation rules we obtain
A=A Â A = îi A
dt ~ dt* dt dZ' dxl ~ dxl dZ
,= 2>3 (20)
dxk dZk dxk dZ'
the temperature equation in the form

f
dT dT dT\ _ d(pD) dT
/ dx2 dZ2
. . _ \2 82T „ dZ a2
-t>D\( — ] —=2 + 2 —
dx3 dZ3 I \dxj dZ dx2 dZdZ2
dZ 32T 82T 7
H- 2 — ——— + —o + ^d = - — X hmf (21)
If the flamelet is thin in the Z direction, an order-of-magnitude analysis

similar to that for a boundary layer shows that the second derivative with
respect to Z is the dominating term on the left-hand side of Eq. (21). This
term must balance the equation term on the right-hand side. The term
containing the time derivative is important only if very rapid changes, such
as extinction, occur. Formally, this can be shown by introducing the stretched
coordinate £ and the fast time scale T
£ = (Z - Zst)/e (22a
T - r/e 2 (22b
where e is a small parameter representing the width of the reaction zone.
If the time derivative term is retained, the flamelet structure is to leading
order described by the one-dimensional time-dependent temperature

equation
2
dT Xd T 1 A .
P— - rP o ^ = T 2, h^ (23)
dt 2 /'
Similar equations may be derived for the chemical species. In Eq. (23),
(24)
dx
is the instantaneous scalar dissipation rate. It has the dimension 1/s and
may be interpreted as the inverse of a characteristic diffusion time. As a
result of the transformation, it implicitly incorporates the influence of
convection and diffusion normal to the surface of the stoichiometric mix-
ture. In the limit x —» 0, the local equilibrium model and the flame-sheet
model are obtained. Local quenching of the flameiet occurs if x exceeds
a critical value x^- 3 For the counterflow geometry, the scalar dissipation
rate at the location where the mixture is stoichiometric may be approxi-
mated, assuming constant density and diffusivity, by
Xst = (fl/ir)exp{-2[erfc- l (2Z s t )] 2 }
~ 4 a Zgt [erfc^ 1 (2 Zst)]2 (25)
where a is the velocity gradient and erfc" l the inverse of the complementary
error function. For example, erfc" 1 (2 Zst) is 1.13 for methane-air flames
with Zst = 0.055 and 1.34 for H2-air flames with Zst = 0.0284. The second
0.025
0.020
0.015
0.010
0.005
0.0 0.2 0.6 0.6 1.0
XSJ [sec]
Fig. 4 Oxidation layer thickness £ in a methane-air diffusion flame as a function

of the inverse of the scalar dissipation rate x-
N. PETERS 165
expression in Eq. (25) is derived using an approximation of the error

function for small values of Zst.
The basis for the universal coordinate transformation is the assumption
that the reaction zone around Zst is asymptotically thin [of ©(e)] in the
mixture fraction coordinate. This may be verified by asymptotic analysis
for real flames or simply by evaluating the reaction rate. An asymptotic
analysis for methane-air diffusion flames has been performed in Ref. 11
and the thickness e of the broadest reaction layer—the H2-CO-oxidation
layer on the lean side of the flame structure—is reproduced in Fig. 4. As
an alternative, the chemical heat-release rate, i.e., the source term in the
temperature equation, of counterflow diffusion flames of CH4 and H2 in
air has been evaluated from a numerical calculation and was plotted in
Fig. 5 as a function of Z. The heat-release rate has been approximated
using a least-squares fit of a Gaussian around the maximum value. The
variance &R of the Gaussian curves, representing the width of the reaction
zone and therefore e, is shown as a function of the scalar dissipation rate
for different pressures in Fig. 6. It is interesting to note that for methane
the width of the reaction zone in the mixture fraction coordinate is con-
siderably smaller than Zst, the width of the outer structure on the lean side
of the flame. The reaction zone width is always much less than 0.3 Zst,
which will be shown subsequently to be a borderline value for turbulent
jet diffusion flames. For hydrogen flames, &R is larger than 0.3 Zst only
for atmospheric flames and if x is larger than 0.1/s.
As in premixed flames, chemistry introduces a characteristic time scale.
When nondimensional equations are derived to describe the flamelet struc-
ture in mixture fraction space, the chemical source term is divided by the
scalar dissipation rate x s r 3 This ratio represents the comparison of the
chemical time to diffusion time. As noted earlier, the extinction of the
diffusion flamelet occurs at a particular value of the imposed scalar dissi-
pation rate, namely \q, which therefore represents a chemical time scale
for extinction. We will define the chemical time scale at this extinction
threshold as
tc = Z 2 (l - Zst)2/X, (26)
This definition is motivated by the expression for \q using one-step large
activation energy asymptotics3 with equal temperatures, T1 and T2, for the
fuel and oxidizer stream
* *-* K'Qt •- I i-> »^ i • • • » - • - «;r 1 I •*—* I //T7\

1
7 \2-
— Z-'ct ) A/1 f (T
IrJ. J- ii \^
\ 4. ct — T" I \\ F
J-s
where the inverse of the right hand side is proportional to the equivalent
chemical time for a stoichiometric premixed flame. 11 Another reasoning
in favor of the definition of rc is based directly on Eq. (25). Since Zst is
numerically small for most fuels, a rescaling of the domain of integration
between Z = 0 and Z = Zst in Eq. (25) would lead to a rescaling of x
with Zirt, as in Eq. (26). The (1 — Zst)2 term is then retained for the sake
of symmetry.
1.0E+09
I" J ] CH4
m3s
8.0E+08- p = 1atm;X=3-17/sec
———— numerical solution
———— Gauss aproximation
6.0E+08-
40E+08- >
2.0E+08-
O.OE+00
0.05 0.1
j 5.0E+06-
r i
.m s. 3
H2
p= 1atm;X=0.00117/sec
4.0E+06- ———— numerical solution
———— Gauss aproximation
3.0E+06-
2.0E<06
1JOE+06
O.OE+00
0 0.05 0.1
Fig. 5 Heat-release rates and their approximation by a Gaussian for methane-air

and hydrogen-air diffusion flames.
The scalar dissipation rate at extinction was estimated in Ref. 3 as \q =

8/s based on aq = 320/s for methane flames, which leads to tc = 0.34 x
10"3 s. This estimate is of the same order of magnitude as tq for premixed
flames. However, by introducing the second expression of Eq. (25) into
Eq. (26), one obtains for small values of Zst
1 (1 - Zst)2
(28)
4[erfc- 1 (2Z st )] 2
Comparing this with Eq. (7), one realizes that, while the extinction time
scales in premixed and diffusion flames are comparable, the velocity gra-
dients aq for extinction are larger by a factor of between 5 and 10 for
N. PETERS 167
premixed flames. This indicates that, in terms of the imposed strain rate,
a diffusion flame extinguishes more easily than a premixed flame. Physi-
cally, this is due to the fact that the inner structure of a diffusion flame
loses heat to both sides, the lean and the rich, while a premixed flame
loses heat only to the preheat zone.
In diffusion flames, in contrast to premixed flames, there is no velocity
scale, such as the burning velocity, from which a characteristic length scale
such as €F could be defined. Nevertheless, the reaction zone must be
resolved in numerical calculations. With an estimate of OR taken from Fig.
6, one needs to transform this width back into physical space. In a flame-
attached coordinate system, in which the x and z coordinates lie locally
within the surface of a stoichiometric mixture, the flame structure extends
normal to this surface in the y direction. The width of the reaction zone
in physical space is related to that in mixture fraction space by
(29)
(dZ/dy)st
Here the stoichiometric value of xst nas been used to represent the scalar
dissipation rate. Bilge r4 has argued that spatial variations of x around xst
should be considered. He performs a regular asymptotic expansion for high
Damkohler numbers for hydrogen flames and derives an expression for
the reaction rate. Then he calculates reaction rate contours on the basis
of experimental data from a nonreacting Freon-air jet. From such ''syn-
thesized" images of the spatial reaction rate structure, he draws extensive
conclusions about the non validity of the flamelet concept. However, a
regular asymptotic expansion becomes incorrect if the chemistry is confined
to thin layers. Then a singular asymptotic analysis as in Ref. 11 must be
performed, and spatial variations of x represent a higher-order correction.
This higher-order effect has been included in Ref. 11 for methane-air flames
and was found to be small. Alternatively, results from numerical calcula-
tions may be considered. In Fig. 7, the maximum temperature in a methane-
air counterflow diffusion flame is plotted over XsT1 for a calculation with
a constant value of x equal to xst and a calculation where x varies with Z.
The differences between these two cases are within the magnitude of a
higher-order correction in a singular asymptotic analysis, as in Ref. 11.
As a consequence of Eq. (29), in numerical calculations of diffusion
flames, the grid must be refined in the vicinity of the flame front to resolve
the local value of the mixture fraction gradient normal to the surface of
the stoichiometric mixture. This also indicates that the reaction zone thick-
ness in physical spaces varies from flame to flame according to variations
of the mixture fraction field and its local gradients. The mixture fraction
variable is independent of chemistry and may, in fact, be calculated from
a convective-diffusive equation that contains no chemical source term. Heat
release by combustion, however, changes the density and, thereby, affects
the mixture fraction field indirectly by expanding the flow in physical space.
Therefore, the numerical grid in a multidimensional calculation cannot be
fixed a priori, but must be adjusted during the calculation.
0.04-I
0.03
0.02
0.01-
0.1 10 100 1000
0.015
CH4
—————————
Hbar]
—————————
5[bar]
0.01- —— . —— .——
10[bar]
0.005-
0.01 100
Fig. 6 Variance of the approximated heat-release rates for hydrogen-air and meth-
ane-air diffusion flames as a function of the inverse of the scalar dissipation rate for
different pressures.
HI. Length Scales in Turbulent Flames

The most characteristic feature of a turbulent flow is the occurrence of
many length scales, ranging from the largest scale €„ the integral scale,
down to the smallest scale, €K, the Kolmogorov scale.
Many numerical calculations of turbulent reacting flows are now being
based on A>e-type models, and Favre averaging is being used to account
for strong density changes caused by combustion. In terms of the Favre-
averaged turbulence kinetic energy k and its dissipation e, the turbulence
intensity v' and the integral length scale may be defined as
V' = = (2k/3) (30a)
l/2
€, = v'3/e (3Gb)
N. PETERS 169
The integral time scale may be derived from these as

(31)
As an intermediate scale, the Taylor length scale may also be considered.
It relates the fluctuating velocity gradient to the turbulence intensity, and
for isotropic turbulence reduces to12
~ (32)
With the definition of the turbulent Reynolds number
Re = (vf €,)/v (33)
the Taylor scale may be expressed as
(34)
Finally, the smallest scale is the Kolmogorov microscale
(35)
also called the "dissipation scale." It represents the lower cutoff of length
scales within a turbulent flowfield.
A. Turbulent Premixed Flames
We have defined the Reynolds number and the Kolmogorov scale with
v = v ref , where vref = /V(\/c p ) ref /p w , and for the subsequent order-of-
magnitude analysis, the Prandtl number is assumed equal to unity. This
assumption leads for premixed flames with Eq. (3) to
sLZF = v (36)
X =Xst
1700
Fig. 7 Maximum temperature in a methane-air counterflow diffusion flame as a

function of the inverse of the scalar dissipation rate \st f°r a constant scalar dissi-
pation rate x = Xst and for a scalar dissipation rate \ = \(Z) that varies with Z.
such that we may define the turbulent Reynolds number

Re = v'tt/sLtF (37)
the turbulent Damkohler number
Da = ttltF = sL€t/v'fF (38)
and the turbulent Karlovitz number
Ka = yfF/sL = tFltK, y = lltK = VtH> (39)
Here -y is the inverse of the Kolmogorov time tK = €K/vK and describes
the straining by the eddies of the Kolmogorov size tK, which have a turn-
over velocity VK = (ve)5. In addition, e is the dissipation of the turbulent
kinetic energy in the unburnt gas. These definitions can be used to derive
the following relations between the ratios V'/SL and ft/fF in terms of the
three nondimensional numbers Re, Da, and Ka as
v'lsL = Re (tt/tp)-1
= Da"* (€,/€F)
= Ka2/3 (€,/€F)1/3 (40)
2 2
as well as the relation Re = Da Ka .
In the following, we will adopt a modified version of Borghi's13 phase
diagram for premixed combustion and plot the logarithm of v'lsL over the
logarithm of ft/€F in Fig. 8. In this diagram, the lines Re = 1, Da = 1,
and Ka = 1 represent boundaries between the different regimes of pre-
mixed turbulent combustion. Another boundary of interest is the line v'l
SL = 1, which separates the wrinkled and corrugated flamelets.
The regime of laminar flames (Re < 1) in the lower-left corner of the
diagram is not of interest in the present context. Among the remaining
four regimes, the wrinkled and corrugated flames belong to the flamelet
regime, which is characterized by the inequalities Re > 1 (turbulence), Da
> 1 (fast chemistry), and Ka < 1 (sufficiently weak flame stretch). Con-
sidering the boundary to the distributed reaction regime given by Ka =
1, where Ka may be written as
Ka = tFltK = e*teK = vllsl (41)
this boundary represents the condition that the flame thickness is equal to
the Kolmogorov scale (the Klimov-Williams criterion). However, in ad-
dition, since viscosity as a molecular transport process relates Kolmogorov
velocity, length, and time scales to each other in the same way as the
velocity, length, and time scales are related for flame propagation, the
flame time is equal to the Kolmogorov time and the flame velocity is equal
to the Kolmogorov velocity. This will be important in the subsequent
discussion.
The distributed reaction regime is characterized by Re > 1, Da > 1, and
Ka > 1, the last inequality indicating that flame stretch is strong and that
N. PETERS 171
10- Da=l
v
Vr
well-stired reactor
Da<l
6-
4 - distnbuted reaction zones

Da>l,Ka>
2-
1 -
Re<l
I I I I I I I T
l/ F
1 2 4 6 8 10
Fig. 8 Phase diagram showing different regimes in premixed turbulent combustion.
the smallest eddies can enter into the flame structure since £K < €F, thereby
broadening the flame structure. The smaller eddies produce the largest
strain rates and may lead to local extinction of the inner reaction layer.
Finally, the well-stirred reactor regime on the upper left of the diagram is
characterized by Re > 1, Ka > 1, but Da < 1, indicating that the chemistry
is slow compared with turbulence.
We will now enter into a more detailed discussion of the various regimes.
The flamelet regime is subdivided into the regimes of wrinkled flamelets
and corrugated flamelets. This boundary is viewed by Williams14 as the
one between single and multiple flame sheets. Clearly, if v' < SL and v'
is interpreted as the turnover velocity of the large eddies, even those eddies
cannot convolute the flame front enough to form multiply connected re-
action sheets. Flame propagation is dominating and there is no strong
interaction between turbulence and combustion in this regime. In the re-
gime of wrinkled flamelets, asymptotic methods using large activation en-
ergy have been a very powerful tool to describe the interaction between
weak turbulence and the flame front. An excellent review on theoretical
as well as experimental results was given by Clavin.15
The regime of corrugated flamelets is much more difficult to analyze

analytically or numerically. In view of Eq. (41), we have with Ka < 1,
V ^ SL > VK (42)
within this regime. Since the velocity of the large eddies is larger than the
burning velocity, these eddies will push the flame front around, causing a
substantial convolution. Conversely, the smallest eddies, having a turnover
velocity less than the burning velocity, will not wrinkle the flame front.
We may construct a discrete sequence of eddies within the inertial range
by defining
€„ = €,/2", €„>€*, n = 0,1,2,... (43)
Then, energy-cascade arguments require that e is independent of n and
dimensional scaling laws lead to a turnover velocity vn of the eddy of size
€„, as
vl = § €„ (44)
indicating that the velocity decreases as the size of the eddy decreases.
To determine the size of the eddy that interact locally with the flame
front, set the turnover velocity vn equal to the burning velocity SL. This
determines the Gibson scale
LC = si/E (45)
The Gibson scale is the size of the burnt pockets that move into the unburnt
mixture. These pockets try to grow there due to the advance of the flame
front normal to itself, but are reduced in size again by newly arriving eddies
of size LG. Therefore, there is an equilibrium mechanism for the formation
of burnt pockets, while unburnt pockets that penetrate into the burnt gas
will be consumed by the flame advancement. It is worth noting that LG
increases with SL if the turbulence properties are kept constant. At suffi-
ciently low turbulence levels, the mean thickness of a turbulent flame
should be influenced by this mechanism and, therefore, also increase with
SL. This is observed in the V-shaped flame by Namazian et al.,16 where
the mean flame thickness increases by a factor of between 2 and 3, as the
equivalence ratio is changed from cf> = 0.6 to 0.8, thereby increasing SL.
Using Eq. (30), one may also write Eq. (45) in the form
LG/f, = (sJv'Y (46)
An illustration of the kinematics of the interaction between a premixed
flame and a turbulent flowfield may be found in Fig. 9 of Ref. 17. In this
numerical study, the characteristic integral length scale ft was kept con-
stant, while the turbulence intensity was increased, showing corrugations of
smaller and smaller size. A similar effect is observed in the two-dimensional
visualizations of the flame front in a square piston engine by Ziegler et
al.18 In these experiments, the engine speed was increased from 500 to
1500 rpm and the measured Gibson length decreased from 0.27 to 0.01
N. PETERS 173
logv n
logln———-
Fig. 9 Graphical illustration of the Gibson scale LG within the inertial range.
mm. For engines, it also may be argued that the integral length scale is
constant and depends only on the geometrical dimensions of the combus-
tion chamber, while the turbulence intensity increases linearly with engine
speed.
A graphical derivation of the Gibson scale LG within the inertial range
is shown in Fig. 9. Here the logarithm of the velocity vn is plotted over
the logarithm of the length scale €„ according to Eq. (44). If one enters
on the vertical axis with the burning velocity SL equal to vn into the diagram,
one obtains LG as the corresponding length scale on the horizontal axis.
This diagram also illustrates the limiting values of LG\ if the burning velocity
is equal to v', LG is equal to the integral length scale €,. This case corre-
sponds to the borderline between corrugated and wrinkled flamelets in
Fig. 8. Conversely, if SL is equal to the Kolmogorov velocity VK, LG is
equal to €K. This corresponds to the line Ka = 1 in Fig. 8. Therefore, LG
may vary between €K and ft in the corrugated flamelet regime.
The next regime of interest in Fig. 7 is the regime of distributed reaction
zones. As noted earlier, the small eddies can enter into the flame structure
and even destroy it, since £K < £F in this regime. Therefore, the notion
of a well-defined flame structure and the burning velocity as a relevant
velocity scale has no meaning in this regime. Another scale, however, the
chemical time scale tc, remains still meaningful since reactions do also occur
independently of flame propagation. In addition, since there is an inner
reaction zone smaller than the flame thickness, the quenching of this inner
zone by flame stretch is the important physical process. From the physical
point of view, the quench time tq is the more relevant time scale but, at
least for stoichiometric methane flames, the preceding estimate has shown
that tq ~ tF, the latter being equal to tc. Analogous to Eq. (45), we may
relate the turnover time tn to the size (n of an eddy with tn = fnfvn as
tl = €2/e (47)
Then, by setting tq = tn, one obtains the quench scale
8, = (e ^)1/2 (48)
This scale may be interpreted as the size of the largest eddy within the
inertial range that is still able to quench a thin reaction zone. Smaller
eddies up to 8^ induce a larger stretch and, thus, will quench the thin
reaction layers within the flowfield more readily and, thereby, try to ho-
mogenize the scalar field locally over a distance up to 8^. Therefore, 8^
may be interpreted as a correlation length for the reactive scalar field or,
in physical terms, as the size of a localized well-stirred reactor. This justifies
the characterization of this regime as the distributed reaction zones regime.
It also has been characterized by Williams19 as the regime of broken re-
action zones, which gives a clear picture of the physical process described
earlier.
In the distributed reaction zones regime, the Damkohler number, which
is based on the integral time scale tt and the chemical time scale tc = tF,
is large. At the order of these scales, chemistry is fast, and thin reaction
zones may be generated locally. These zones are quenched by flame stretch
such that an equilibrium mechanism exists between the generation and the
destruction of thin reaction zones over a region of thickness 8^.
Again, the derivation of 8^ is illustrated in a diagram in Fig. 10, showing
Eq. (47) in a log-log plot of tn over €„. If one enters the time axis at tq =
tn, the quench scale 8^ on the length scale axis is obtained. It should be
noted that all eddies having a size between €K and 8^ have larger stretch
than 8^ and, therefore, are able to quench thin reaction zones locally. If
tq is equal to the Kolmogorov time tK, Fig. 10 shows that 8^ is equal to the
Kolmogorov scale (K. In this case, setting tq = tF, one obtains 8^ = £F at
the border between the distributed reaction zones regime and the corru-
gated flamelet regime. Similarly, from Fig. 10, if the quench time tq is
equal to the integral time tt, the quench scale is equal to the integral length
scale. This would correspond to Da — 1 in Fig. 8 and delineates the
borderline between the distributed reaction zones regime and the well-
stirred reactor regime.
As a final remark related to the distributed reaction zones regime, it
may be noted that turbulence in real systems is not homogeneous and e
is not a local constant but has a distribution. This refinement of Kolmo-
gorov's theory has led to the notion of intermittency, or "spottiness," of
the activity of turbulence in a flowfield.20 This may have important con-
sequences on the physical appearance of turbulent flames at sufficiently
large Reynolds numbers. One may expect that the flame front shows mani-
festations of distributed reaction zones with local quenching events as well
as of regions where corrugated flamelets appear. The regimes discussed
earlier may well overlap each other in an experimentally observed turbulent
flame.
N. PETERS 175
logt n
Fig. 10 Graphical illustration of the quench scale 89 within the inertial range.
In the well-stirred reactor regime, chemistry is slow, as noted earlier.

Now all eddies up to €r can quench the inner reaction zones, since Da <
1. Turbulence homogenizes the scalar field by rapid mixing, leaving the
slow chemistry to be the rate-determining process. No specific interaction
between turbulence and combustion can occur in this regime, and no spe-
cific interaction scale is to be defined.
As far as direct numerical calculations of premixed turbulent combustion
are concerned, it would appear desirable to resolve all scales down to the
smallest scale, which is the flame thickness in the flamelet regime and the
Kolmogorov scale in the distributed reaction zones regime. However, in
the corrugated flamelet regime, the flame thickness does not enter into
the formulation if the burning velocity is specified and its variation due to
flame stretch is ignored. The Gibson scale rather than the Kolmogorov
scale appears as the smallest scale in the resolution of the flame front.
Conversely, in the distributed reaction zones regime, the smallest scales
down to the Kolmogorov scale and the inner reaction zone must be resolved
to capture the interaction between flame stretch and chemistry and the
occurrence of local quenching.
Numerical calculations of premixed turbulent combustion on the basis
of semiempirical turbulence models still suffer from the unresolved closure
problem regarding the chemical source term. New models based on a
second equation for the flame surface area per unit volume 2, a quantity
that has the dimension of the inverse of a length scale, have been developed
recently.21 This seems to be a promising direction, since it combines phys-
ically appealing features such as the creation and destruction of flame

surface area with the advantage of cost-efficiency of semiempirical models.
B. Turbulent Diffusion Flames

It is evident from the analysis in Sec. II.B that the reactants are initially
nonpremixed and that combustion is determined by the rate of mixing in
diffusion flames. This situation is generally referred to as nonpremixed
combustion. As noted earlier, in contrast to premixed combustion, there
is no characteristic length scale that is independent of the flowfield in
nonpremixed combustion. Therefore, the criterion for flamelets defined
by Bilger4 that requires that the reaction zone thickness should be thinner
than the Kolmogorov scale has no physically meaningful basis. The only
quantity that is independent of the flowfield is the width of the reaction
zone a/?, which is measured in mixture fraction space, not in physical space.
According to Eq. (29), the width of reaction zones in physical space will
therefore depend on the local mixture fraction gradient in a turbulent
flowfield. To estimate the mean reaction zone thickness in physical space,
one has to take the average of Eq. (29). For isotropic turbulence, a suitably
defined average mixture fraction gradient (3Z/dx)av is related to the Favre
mean scalar dissipation rate by
(f
The mean scalar dissipation rate x rnay be expressed in terms of the ratio
£/e, which defines the turbulent time according to Eq. (31), and the mixture
fraction variance by an empirical relation
X = cx (§/£) Z^ (50)
It is useful to define the Favre mixture fraction fluctuation as the square

of the mixture fraction variance
Z' = V Z"2 (51)

Then the average reaction zone thickness is
(52)
Using Eq. (32), this may be expressed in terms of the Taylor microscale
as
With the previous assumptions of Pr = I and Le = 1, the term under the

square root is a constant of order unity. Therefore, the average reaction
N. PETERS 177
zone thickness is proportional to the scalar ratio o-^/Z' and the Taylor scale
XT. As noted in Ref. 12, the Taylor scale is not a characteristic length of
the strain-rate field and should be used only in combination with the tur-
bulence intensity v' and the Kolmogorov time scale defined in Eq. (39).
Therefore, the Taylor scale represents the distance over which a fluid
element is convected by a large eddy with a velocity v' during a Kolmogorov
time. According to Eq. (53), the average reaction zone thickness is pro-
portional to this distance multiplied by the ratio vR/Zr. The physical sig-
nificance of the scalar ratio a//Z' will become apparent in the following.
Figure 11, adopted from Fig. 12 in Ref. 3, shows two examples of an
instantaneous mixture fraction profile along an arbitrarily chosen spatial
coordinate jc. Here the value of cr^ on the Z axis was introduced to illustrate
the extent of the reaction zone. The corresponding values of the reaction
zones in physical space are shown on the x axis. It was argued in Ref. 3,
based on the concept of freezing flamelets, that for the example on the
left part of Fig. 11, the reaction zones are separated and individual diffusion
flamelets may exist. Conversely, this is less likely to occur for the small
variance case on the right-hand side of Fig. 11. The reaction zone is con-
nected and a premixed flame may travel through the slightly inhomoge-
neous mixture.
In addition, we want to consider a location in a turbulent diffusion flame
where the Favre mean mixture fraction Z equals Zst. At this location, time
sequences of Z resemble those in Fig. 11 and the maximum fluctuation
around Zst should correspond to the mixture fraction fluctuation. The
conditioned Favre mixture fraction variance
(54)
Fig. 11 Schematic illustration of mixture function profiles in physical space. For

the large scalar variance case on the left-hand side, the reaction zone thickness (Ay)*
corresponding to cr^ is shown. For the small variance case on the right-hand side,
the reaction zones are connected.
extinguishing
Zst(l-Zst) flamelet regime
Xst =
10-
distributed
reaction flamelet regime
zones
•i __
0.1 - Z' s t =a R
connected reaction zones
0.01
0.1 10
Fig. 12 Phase diagram showing different regimes in nonpremixed turbulent com-
bustion.
is expected to be larger than cr^ for the case on the left-hand side of Fig.
11, whereas it would be smaller for the case on the right-hand side. Sep-
arated flamelets will exist for Z^t > o> and connected reaction zones for
Zst < VR-
Based on these considerations, we propose a phase diagram for non-
premixed turbulent combustion as shown in Fig. 12. In this diagram, the
conditioned mixture fraction variance Z^ normalized by Zst (1 - Zst) is
plotted over the ratio of the turbulent time tt to the chemical time tc, which
by analogy corresponds to the turbulent Damkohler number of the previous
section.
The choice of the vertical and horizontal axes in Fig. 12 was motivated
by the criterion for quenching of diffusion flamelets22'23
Xst = X, (55)
which has led to a liftoff criterion for turbulent jet diffusion flames.23 For
an order-of-magnitude analysis, we shall use cx = 3/2 for the empirical
constant in Eq. (50) and write the ratio of \l\q at a location Z = Zst with
Eqs. (26), (50), and (54) as
Xst
(56)
X, tt Z2 (1 - Zst)2
Setting the left-hand side of Eq. (56) equal to unity, one obtains a line
with a slope 1/2 in Fig. 12, representing the liftoff limit.
Analogous to the Klimov-Williams limit in premixed turbulent com-
bustion, an additional criterion separating the flamelet regime from the
distributed reaction zones regime must be considered. Since only time
N. PETERS 179
scales are relevant for nonpremixed combustion, the condition Ka = 1

reduces to tK = tc, requiring that the characteristic time for extinction must
be smaller than the Kolmogorov time for flamelets to exist.
There are four regimes in Fig. 12 that will be discussed in the following.
The extinguishing flamelet regime is defined by xst > Xq and *K > tc. In
this regime, the mean scalar dissipation rate is large due to large mixture
fraction fluctuations, while the chemical time scale tc is smaller than the
Kolmogorov time. Flamelets extinguish in this regime due to large mixture
fraction gradients. The reaction zone in physical space becomes so narrow
that diffusive heat loss will lead to quenching. This is followed by unsteady
interdiffusion of reactants and heat conduction into the rich and lean mix-
ture surrounding the flamelet. This unsteady transition occurs on the dif-
fusive-convective time scale t*, which is long compared with the short time
scale T, Eq. (22), which governs unsteady effects in the reaction zone.
Because of fluctuations of the scalar dissipation rate, there exists a finite
probability for reignition and burning of those flamelets, whose instanta-
neous value of Xst is smaller than x^- 3 Masri and Bilger,24 for instance, have
measured for such a situation the electrical connectedness in burner-sta-
bilized flames and found highly intermittent signals, indicating local ex-
tinction events. For lifted jet diffusion flames, the condition xst > Xq applies
for the region upstream of the stabilization height.
The flamelet regime in Fig. 12 is defined by xst < X?? *K > ^> an d Z'^
> a^. This regime is valid, for instance, for jet diffusion flames downstream
of the stabilization height. Local quenching of flamelets is still possible in
this regime and will reduce the probability of burning and, thereby, the
overall heat-release rate.3 However, since the average scalar dissipation
rate is lower than \q, most flamelets will be burning. Their response to
unsteady fluctuations of the outer field and, thereby, of the scalar dissi-
pation rate will occur on the short time scale T in the reaction zone but on
the long time scale t* in the outer regions.
In the flamelet regime, the chemical time at extinction tc is smaller than
the Kolmogorov time, tK. This indicates that burning is faster than per-
turbations induced by the smallest eddies, which have the shortest turnover
time. This is similar to the condition tF < tK for premixed flamelets. A
different scaling, however, applies for length scales. It follows from Eq.
(53) that since <JR < Z'si = Z', the average reaction zone thickness is smaller
than the Taylor scale in the flamelet regime for nonpremixed combustion.
This differs fundamentally from the condition for premixed flamelets where
the laminar flame thickness is small compared with the Kolmogorov scale.
The next regime in Fig. 12, the regime of connected reaction zones, is
defined by Z'si < <JR and tK > tc. This regime is not likely to occur in jet
flames, as will be shown subsequently. It could occur in closed combustion
devices, if one allows for a decay of mixture fraction fluctuations leading
to partial premixing, before an ignition source is introduced. The subse-
quent burning will resemble a flame propagation process in an inhomo-
geneous mixture. In this respect, it is more closely related to premixed
than nonpremixed combustion.
Finally, the fourth regime of the phase diagram in Fig. 12, the distributed
reaction zones regime, is defined by tc > tK. In this regime, the turnover
time of the small eddies is shorter than the chemical time at extinction.
Similar physical arguments are valid in this regime, as in the corresponding

distributed reaction zones regime for premixed turbulent combustion. In
particular, one may consider an eddy with a turnover time equal to tc and
thereby define a quench scale
8C = (efc3)1/2 (57)
which describes the correlation length of the distributed reaction zones.

This regime is likely to occur in burner-stabilized flames at large turbulent
Reynolds numbers. Different from Bilger's4 arguments, it will occur close
to the nozzle rather than downstream, since the Kolmogorov time is con-
siderably smaller upstream than in the downstream portion of the flame.
It will be interesting to investigate the relative importance of these re-
gimes for practical applications. For jet diffusion flames, the turbulent time
scale and the mixture fraction variance may be estimated easily. For a
round jet, the product v'(t on the centerline is a constant, since v' decreases
with the distance x from the nozzle as x~1 and €, increases linearly. This
product may be related to the nozzle exit velocity u0 and nozzle diameter
das 25
v'tt/u0 d = 1/70 (58)
The turbulent time scale therefore increases as x2 according to
tt = (t/v' = Uo d/10 v'2 (59)
Since Bilger4 has argued that the flamelet regime would not apply in the
downstream part in a turbulent jet diffusion flame, the turbulent time scale
shall be estimated at the flame length, defined by
ZCL = Zst (60)
where ZCL is the Favre mean mixture on the centerline. The turbulence
intensity v' and the mixture fraction fluctuation in the similarity region on
the centerline of a round jet may be estimated as
VCL = 0.3 WCL (61a)
ZCL - 0.3 ZCL (61b)
The centerline decay of the mean mixture fraction and of ucJuo are aP"
proximately equal since the turbulent Schmidt number is of order unity
ZCL = ucJuQ (62)
Introducing Eqs. (60-62) into Eq. (59), one obtains an estimate for the
turbulent time
0.16 d _
tt = (63)
~
N. PETERS 181
Similarly, the Kolmogorov time tK is obtained from Eq. (35) with Eqs.
(30) and (31) and Eqs. (60-62) as
V2
d_
Z2CL \u(> d
Using as an example a nozzle exit time d/u0 of 10 ~ 4 s and a nozzle Reynolds
number u0d/v of 104 for laboratory jet diffusion flames, the turbulent time
is found to be one order of magnitude larger than the Kolmogorov time.
The ratio tjtc is in the similarity region of a jet 15.5 and tKltc = 1.3 for
methane flames at the flame length. Since tt increases as x2 and Z'CL de-
creases as jc"1, conditions on the centerline of a jet flame would follow a
line with the slope — i in Fig. 12.
With respect to the criterion Z,[t > o^, one finds with Eqs. (60) and (61)
at the flame length Z'si = 0.3 Zst and by comparison with Fig. 6 that the
criterion Z^t > &R is satisfied everywhere in turbulent jet diffusion flames
of methane. For hydrogen flames at 1 atm, this is also true in general since
the mean scalar dissipation rate is typically much larger than 0.1/s.
IV. Conclusions
For numerical calculations of turbulent diffusion flames, it may be con-
cluded that combustion is most likely to occur in the flamelet regime.
Flamelet models for turbulent diffusion flames have been discussed in Ref.
3, but more recent additional references are given in Ref. 2. The improve-
ment of flamelet models, both from the chemistry and the turbulence point
of view, appears to be worthwhile and necessary, in particular with respect
to unsteady transition. An example of such improvement is the recent work
by Haworth et al.26
References
Tope, S. B., "PDF Methods for Turbulent Reactive Flows," Progress in Energy
and Combustion Sciences, Vol. 11, 1985, pp. 119-192.
2
Peters, N., "Laminar Flamelet Concepts in Turbulent Combustion," Twenty-
first Symposium (International) on Combustion, The Combustion Institute, Pitts-
burgh, PA, 1986, pp. 1231-1250.
3
Peters, N., "Laminar Diffusion Flamelet Models in Non-Premixed Turbulent
Combustion," Progress in Energy and Combustion Sciences, Vol. 10, 1984, pp.
319-339.
4
Bilger, R. W., "The Structure of Turbulent Nonpremixed Flames," Twenty-
second Symposium (International) on Combustion, The Combustion Institute, Pitts-
burgh, PA, 1988, pp. 475-488.
5
Peters, N., and Williams, F. A., "The Asymptotic Structure of Stoichiometric
6
Karlovitz, B., Denniston, D. K., Knappschaefer, D. H., and Wells, F. E.,
"Studies on Turbulent Flames," Fourth Symposium (International) on Combustion,
The Combustion Institute, 1953, p. 613.
7
Rogg, B., "Response and Flamelet Structure of Stretched Premixed Methane-
Air Flames," Combustion and Flame, Vol. 73, 1988, pp. 45-65.
8
Tsuji, H., and Yamaoka, I., "Structure Analysis of Counterflow Diffusion Flames
in the Forward Stagnation Region of a Porous Cylinder," Thirteenth Symposium
p. 729.
9
Tsuji, H., "Counterflow Diffusion Flames," Progress in Energy and Combustion
Sciences, Vol. 8, 1982, p. 93.
10
Peters, N., and Kee, R. J., "The Computation of Stretched Laminar Methane-
Air Diffusion-Flames Using a Reduced Four-Step Mechanism," Combustion and
Flame, Vol. 68, 1987, p. 17.
H
Seshadri, K., and Peters, N., "Asymptotic Structure and Extinction of Methane-
Air Diffusion Flames," Combustion and Flame, Vol. 73, 1988, p. 23.
12
Tennekes, H., and Lumley, J. L., A First Course in Turbulence, MIT Press,
Cambridge, MA, 1972.
13
Borghi, R., "On the Structure of Turbulent Premixed Flames," Recent Ad-
vances in Aeronautical Science, edited by C. Bruno, and C. Caseci, Pergamon,
New York, 1984.
14
Williams, F. A., Combustion Theory, 2nd ed., Benjamin/Cummings Publishing,
Menlo Park, CA, 1985, p. 412.
15
Clavin, P., "Dynamic Behavior of Premixed Flame Fronts in Laminar and
Turbulent Flows," Progress in Energy and Combustion Sciences, Vol. 11, 1985,
P.I.
16
Namazian, M., Shepherd, I. G., and Talbot, L., "Characterization of Density
Fluctuations in Turbulent V-Shaped Premixed Flames," Combustion and Flame,
Vol. 64, 1986, p. 299.
17
Ashurst, W. T., and Barr, P. K., "Stochastic Calculation of Laminar Wrinkled
Flame Propagation via Vortex Dynamics," Combustion and Science Technology,
Vol. 34, 1983, p. 227.
18
Ziegler, G. F. W., Zettlitz, A., Meinhardt, P., Herweg, R., Maly, R., and
Pfister, W., "Cycle Resolved 2-D Flame Visualization in a Spark Ignition Engine,"
SAE Paper 881634, 1988.
19
Williams, F. A., private communication, 1988.
20
Kolmogorov, A. N., "A Refinement of Previous Hypothesis Concerning the
Local Structure of Turbulence in a Viscous Incompressible Fluid at High Reynolds
Number," Journal of Fluid Mechanics, Vol. 13, 1962 p. 82.
21
Darabiha, N., Giovangigli, V., Trouve, A., Candel, S., and Esposito, E.,
"Coherent Flame Description of Turbulent Premixed Ducted Flames," U.S.-France
Joint Workshop on Turbulent Reactive Flows, 1987.
22
Peters, N., "Local Quenching Due to Flame Stretch and Non-Premixed Tur-
bulent Combustion," Combustion and Science Technology, Vol. 30, 1983.
23
Peters, N., and Williams, F. A., "Liftoff Characteristics of Turbulent Jet Dif-
fusion Flames," AIAA Journal, Vol. 21, 1983, p. 423.
24
Masri, A. R., and Bilger, R. W., "Turbulent Diffusion Flames of Hydrocarbon
Fuels Stabilized on a Bluff Body," Twentieth Symposium (International) on Com-
25
Peters, N., and Donnerhack, S., "Structure and Similarity of Nitric Oxide
Production in Turbulent Diffusion Flames," Eighteenth Symposium (International)
on Combustion, The Combustion Institute, Pittsburgh, PA, 1981, pp. 33-42.
26
Haworth, D. C., Drake, M. C., Pope, S. B., and Blint, R. J., "The Importance
of Time-Dependent Flame Structures in Stretched Laminar Flamelet Models for
Turbulent Jet Diffusion Flames," Twenty-second Symposium (International) on
Chapter 7
Numerical Modeling of Laminar Diffusion Flames
Mitchell D. Smooke
I. Introduction
T HE most common type of flame in practical combustion devices is the

diffusion flame. These flames are important in the interaction of heat
and mass transfer with chemical reactions in gas turbines and commercial
burners. The ability to predict the coupled effects of complex transport
phenomena with detailed chemical kinetics in these systems is critical for
modeling turbulent reacting flows, in improving engine efficiency, and in
understanding the process by which pollutants are formed. In particular,
many commercial power-generating units employ diffusion flames as their
primary type of flame. Since most of the oxides of nitrogen are formed
during combustion when part of the oxygen combines with atmospheric
nitrogen rather than with the fuel, burning hydrocarbon fuels in these units
can produce large quantities of nitrogen dioxide and nitric oxide. Both
compounds are considered toxic, and nitric oxide is related to the formation
of photochemical smog.1 Improving the efficiency of these units combined
with environmental issues dealing with the production of nitrogen-based
pollutants helps to motivate the study of these flames.
Modeling laminar diffusion flames requires the solution of the coupled
equations of mass, momentum, species balance, and energy with detailed
thermodynamic and transport relations and finite-rate chemistry. These
equations are solved for the density, the velocities, the species mass frac-
tions, and the temperature. The interaction of heat and mass transfer with
chemical reactions in practical combustion systems requires a multidimen-
sional study. Three-dimensional models combining both fluid dynamical
effects with finite-rate chemistry are as yet computationally infeasible. As
a result, modeling diffusion flames generally focuses on one- or two-di-
mensional systems. In this chapter we consider two specific laminar dif-
fusion flame configurations—the counterflow flame and the axisymmetric
coflow flame. In the next section, we discuss modeling one-dimensional
counterflow flames in which a fuel stream and an airstream flow against
reserved.
183
184 NUMERICAL MODELING OF LAMINAR DIFFUSION FLAMES
one another (see also Refs. 2-14). In Sec. Ill, we consider a two-dimen-
sional, axisymmetric laminar diffusion flame in which a cylindrical fuel
stream is surrounded by a coflowing oxidizer jet (see also Refs. 15-18).
Unlike some models in which diffusion in the axial direction is neglected
(see, for example, Ref. 19), we treat the fully elliptic problem. In both
configurations, we can study the interaction of fluid flow and chemical
reaction while obtaining a computationally feasible problem.
II. Counterflow Diffusion Flames

Counterflow diffusion flames have played an important role in recent
models of turbulent nonpremixed combustion. The reacting surface in these
models can be viewed as being composed of a number of thin, laminar,
diffusion flamelets (see, for example, Refs. 8, 20-22). Using the flamelet
concept, researchers have been able to include the effects of complex
chemistry and detailed transport in models of turbulent reacting flows.23
In addition to their use in flamelet models, Counterflow diffusion flames
Oxidizer
Flame
Stagnation
( Plane
Fuel
Fig. 1 Schematic of double-jet counterflow diffusion flame.
M. D. SMOOKE 185
\ ^-Boundary Layer Edge
Free Stagnation Point
Flame Zone
Fig. 2 Schematic of porous cylinder counterflow diffusion flame.
have been used by combustion scientists to investigate chemically con-

trolled extinction limits and to study the complex transport and chemical
kinetic interactions that occur in nonpremixed combustion. 2 ~ 4 - 6 - 12 Exper-
imentally, these flames can be produced when a reaction zone is stabilized
near the stagnation point of two infinitely wide coaxial concentric jets, as
shown in Fig. 1. Fuel is emitted from one jet and oxidizer (air) from the
other. In addition to the two-jet system, Tsuji and Yamaoka 2 " 4 have
investigated counterflow diffusion flames of the type shown in Fig. 2, in
which fuel is emitted from a porous cylinder into an oncoming stream of
air2-4,6,7 ^ f ree stagnation line parallel to the cylinder axis forms in front
of the cylinder's porous surface. In both cases, combustion occurs within
a thin flame zone, where the fuel and oxidizer are in stoichiometric pro-
portion.
Our model for counterflow diffusion flames assumes a steady-state, lam-
inar, stagnation-point flow. The governing boundary-layer equations for
mass, momentum, chemical species, and energy can be written in the form
dx
'
dy (1)
du dU dp d i du
pu — + p v — + — - — I L L — -0 (2)
dx dy dx dy \ dy
tr k , ^r T „ . „, _
pa — * + pv —± + - (pYkVkv) - wkWk = 0
dx dy dy
k = 1, 2, ..., K (3)
dT ST d / dT\ £ v 17 dT
Pucp -+pvCp--~ (X -j + £ pYkVkycpk Yy
+ I) wkWkhk = 0 (4)
=
where a represents a geometric factor (a = 0 for Cartesian coordinates;

a = 1 for cylindrical coordinates). The system is closed with the ideal gas
law,
p = pW/RT (5)
In these equations, x and y denote independent spatial coordinates in the
tangential and transverse directions, respectively; T the temperature; Yk
the mass fraction of the A:th species; p the pressure, u and v the tangential
and transverse components of the velocity, respectively; p the mass density;
Wk the molecular weight of the fcth species; W the mean molecular weight
of the mixture; R the universal gas constant; X the thermal conductivity of
the mixture; cp the constant-pressure heat capacity of the mixture; cpk the
constant-pressure heat capacity of the /cth species; wk the molar rate of
production of the A:th species per unit volume; hk the specific enthalpy of
the A:th species; JJL the viscosity of the mixture; and Vky the diffusion velocity
of the kth species in the y direction. In both configurations, the freestream
(tangential) velocity at the edge of the boundary layer is given by u^ =
ax, where a is the strain rate.
We introduce the notation
u = axf (6)
V = pv (7)
where/' is related to the derivative of the modified stream function (see,

for example, Ref. 10). Using these expressions, the boundary-layer equa-
tions can be transformed into a system of ordinary differential equations
valid along the stagnation-point streamline x = 0. For a system in rectan-
gular or cylindrical coordinates, we have
^ + a(l + a)p/' = 0 (8)
-4-(l>YkVk) - V^p + wkWk = 0, k = 1, 2, ..., K (10)

M. D. SMOOKE 187
The boundary conditions for the double-jet configuration at y = — ^

are given by
(12)
(13)
(14)
(15)
(16)
(17)
T = Tx (18)
The mass flux, temperature, and species mass fractions (V_x, T_^, Yk ^) in
the fuel jet are specified quantities, as are the temperature and species
mass fractions (Tx and Ykx) in the oxidizer jet.
At the cylinder wall (y = 0) in the Tsuji configuration, we have
V(0) = Vw (19)
/'(O) = 0 (20)
Y,(0) + {[pY,(0)Vj/VU = e*, k= 1, 2, ..., K (21)
r(0) = Tw (22)
and at y = ^
f = 1 (23)
Yk = Ykx, k = 1,2, . . . , / f (24)
T - T^ (25)
The mass flux, temperature, and the incoming mass flux fractions (Vw, Tw,
and e^) at the wall are specified, as are the mass fractions of the species
and the temperature (Y^ and Tx) at the edge of the boundary layer. To
complete the flame description, detailed thermodynamic and transport
models must be incorporated into the governing equations along with finite-
rate kinetics. The form of the chemical production rates and the thermo-
dynamic and transport relations that we employ can be found in detail in
Refs. 24 and 25.
The complexity of the counterflow diffusion flame model generally pre-
cludes solving the governing conservation equations analytically. As a re-
sult, we employ numerical methods in which the continuous spatial derivatives
are replaced by finite-difference approximations. The differential equations
are now reduced to a system of discrete nonlinear algebraic equations. As
will be described in more detail in the next section, we solve these equations
by employing a variation of Newton's method. Newton's method is guar-
anteed to converge if the starting estimate lies within the domain of con-
vergence of the Kantorovich theorem.26 Despite the outwardly simple form
of the counterflow problem, the determination of such an initial solution
estimate can be difficult. The difficulty is due to the exponential depend-
ence of the chemistry terms on the temperature and to the nonlinear cou-
pling of the fluid and the thermochemistry solution fields. In adiabatic and
nonadiabatic premixed laminar flame problems, the conservation of mass
and momentum reduces to the specification of a constant mass-flow rate
and a constant thermodynamic pressure.27-28 Hence, thermochemical con-
siderations play a more important role in these problems than do fluid
dynamical aspects. This is not the case in counterflow diffusion flames.
There is a strong coupling between the fluid dynamic and the thermo-
chemistry solution fields in these flames. We have found that, although
the solution procedure used in premixed laminar flame problems can work
in selected counterflow cases, it does not provide a sufficiently robust or
efficient starting estimate from which Newton's method will converge. In
addition, the relaxation to steady state (or at least until the solution is
within the convergence domain of Newton's method) is very slow. The
importance of these flames in modeling turbulent reacting flows and in the
determination of chemically controlled extinction limits, however, neces-
sitates the development of an efficient starting procedure. An excellent
starting estimate can be obtained by coupling the appropriate equations
of mass and momentum with a convection-diffusion equation for a con-
served scalar equation to provide an estimate based on a flame-sheet model
for the mass flux in the transverse direction, the similarity function/', the
temperature, and the stable major species in the flame.
Flame-Sheet Model
The burning rate in a diffusion flame is controlled by the rate at which
the fuel and oxidizer are brought together in the proper proportions. This
is in contrast to premixed flames where the burning rate is controlled by
chemical reactions. In the limit of infinitely fast kinetics, the fuel and
oxidizer are separated by a thin exothermic reaction zone. In this zone,
the fuel and oxidizer are in stoichiometric proportion, and the temperature
and products of combustion are maximized. In such an ideal situation, no
oxidizer is present on the fuel side and no fuel is present on the oxidizer
side. The fuel and oxidizer diffuse toward the reaction zone as a result of
concentration gradients in the flow. In diffusion flames of practical interest,
the oxidation of the fuel to form intermediates and products proceeds
through a series of chemical reactions that comprise the detailed kinetics
mechanism (see, for example, Refs. 29 and 30). In these problems, com-
bustion occurs at a finite rate and some fuel and oxidizer coexist on either
side of the reaction zone. Nevertheless, the use of an infinitely fast, thin
flame global reaction model is a natural starting point for the determination
of a "good" initial solution estimate for the finite-rate counterflow model.
The thin flame approximation has a long and useful history in the com-
bustion literature, 15 and similar ideas have been used by Smith et al.31 in
M. D. SMOOKE 189
the solution of premixed flames in a stagnation-point flow and by Mitchell

et al.16-17 in the solution of axisymmetric laminar diffusion flames.
Our starting point is the assumption that the fuel and oxidizer obey a
single overall irreversible reaction of the type
Fuel (F) + Oxidizer (OX) -> Products (P) (26)
in the presence of an inert gas (N). We have
vFF + voxOX^> vPP (27)
where VF, vox, and vp are the stoichiometric coefficients of the fuel, oxi-
dizer, and product, respectively. In addition, we neglect thermal diffusion
and assume that the ordinary mass diffusion velocities can be written in
terms of Pick's law, i.e.,
Vk=- — ^jj£, k=l,2,...,K (28)
where Dk is the diffusion coefficient of the A:th species with respect to the
mixture. We also take cpk = cp to be a constant independent of temper-
ature. With these approximations we can write
^ + fl (l + a)pf = 0 (29)
j i f* j.. ; •• j.. ' "in* HU > j - " (30)
A ' ~ dYnx\ ,,dYr

dy
WPvPw = 0 (33)
dy J dy
,,dY
dy \^» = 0 (34)
dyj ' dy
dy \cp dy dy
WFvFhF + Woxvoxhox - WpV
where
w = -wFlvF = -woxlvox = Wplvp (36)
is the rate of progress of the reaction and where we have made use of the
fact that ££=! YkVk = 0.
If we introduce the heat release per unit mass of the fuel Q, where
Q = hF + (WoxvoxIWFvF) hox - (WPvP/WFvF) hp (37)
and if we assume that the Lewis numbers
LeF = X/pDFcp Leox = \/pDoxcP (38a
LeP = \/pDPcP LeN — \/pDNcP (38b
are all equal to 1 (i.e., the Dk are all equal to a common D = X/pcP), then
each of the Shvab-Zeldovich coupling functions
ZF = YF - F^ + I (T - T,) (39)
W
Zox = Yox - YOXx + I ™JV°* (T ~ T,) (40)
ZN = YN - y^ (42)
satisfies the differential equation
4: (pD ¥*} - V^ = 0, k = F, OX, P, N (43)
with
Z*(-oo) = Z,_ (44)
Zfc(») = 0 (45)
for the double-jet problem and

PD dZ,(0) (46)
Zt(») = 0 (47)
for the Tsuji configuration. In both cases, Zk_x and Zkn. are constant. As
a result, all of the Zk are proportional to each other and to the conserved
scalar S, which, for the double-jet problem, satisfies
_„ „ (48)
dy
M. D. SMOOKE 191
S(-oo) = 1 (49)
5(oo) = 0 (50)
The flame-sheet model can be developed for both experimental configu-
rations. For convenience, in the discussion that follows, we confine our
development to the double jet. A similar analysis can be applied to the
Tsuji configuration.
From Eqs. (43-45) and (48-50), we can write
Zk = Zk_ 5(>0, k = F, OX, P, TV (51)
Equation (48) can be coupled with Eqs. (29) and (30) to obtain profiles
for V, /', and S. To complete the specification of the starting estimate, we
must be able to recover the temperature and the major species profiles
from the conserved scalar. Of critical importance to this procedure is an
estimate of the location of the flame front, yf.
In the Shvab-Zeldovich formulation, fuel and oxidizer cannot coexist.
Hence, yox = 0 on the fuel side of the flame, and YF = 0 on the oxidizer
side. If we denote variables at the flame front with the subscript /, then
from Eq. (51) we can write
ZF, = (ZF_JZOX_J ZOXf (52)
and with YFf = YOXf = YFx = 0 and Eqs. (39-40), we have
1
1 - ^F-
OX-X S )]
W Xl
° "*\ I
F F / J
(53)
If we use Eq. (53) in the relation S(yf) = ZFfIZF_x, we can obtain a value
for the conserved scalar at the flame front. We have
S(yf) = Sf= YoxJ[YOXx + (Woxvox/WFvF) YF_X] (54)
The location of the flame front can be obtained by solving

%/) = Sf (55)
This relation divides the domain into two subdomains. The fuel side is
given by -co < y < yf and the oxidizer side by yf < y < co. Expressions
for the temperature and species can be recovered from the relations in Eq.
(51). On the fuel side, we have
T = T^S + [V, + YOX_ ^^-]d - S) (56)
yF = y f _5 + Yox^ ^^ (5-1) (57)
Y0^ = 0 (58)
YP = Yox - - ( 1 - S) (59)
Woxvox
YN = YNx(l - S) + YN_,S (60)
On the oxidizer side, we have
T = Tx(l - 5) + [(Qlcp) YF_, + T_x] S (61)
YF = 0 (62)
YOX = YOXx(l - S) - YF_, (WoxvoxIWFvF) S (63)
YP = (WpVplWfVp) YF_, S (64)

YN = y^ (1 - S) + YN_ S (65)
We point out that if we have two products, i.e.,
vFF 4- voxOX -> vPl P! + vP2P2 (66)
then, since yp = YPl H- yF2, we can recover yPl and yP2 by forming
YPl = [WPlvPl/(WPlvPl + WP2vP2)] YP (67)
and
YP2 = [WP2vP2/(WP}vPl + WP2vP2)] YP (68)
Combining these ideas, our flame-sheet starting procedure reduces to
the solution of
dV
— + 2apf = 0 (69)
with the boundary conditions at y = —<*> given by

V = V_x (72)
/' = v^; (73
S = 1 (74)
and at y = <* by
/' = 1 (75)
5 = 0 (76)
M. D. SMOOKE 193
For a given profile of the conserved scalar, we solve Eq. (55) for the
location of the flame front. We then utilize the relations in Eqs. (56-65)
to obtain expressions for T, YF, Yox, YP, and YN. The recovered tem-
perature profile is used in the ideal gas law to evaluate the density. The
temperature is also needed in forming the viscosity and the diffusion coef-
ficient. If we introduce the Prandtl number,
Pr = M,C/A (77)
and recall that all of the Lewis numbers are equal to 1, we can write
pD = X/cp = >i/(P/-)ref (78)
where (Pr)ref is a reference Prandtl number. Specifically, we use an ap-
proximate value for air, (/V)ref = 0.75. Hence, determination of pD is
reduced to the specification of a transport relation for the viscosity. We
use the simple power law
JJL = M,O (T/TQY (79)
where r = 0.7, T0 = 298 K, and JJLO = 1.85 x 10~4 g/cm-s is again a
reference value for air.32 The temperature exponent was determined by
fitting Eq. (79) to the mixture viscosity and temperature data of a repre-
sentative finite-rate chemistry calculation. The scaled heat-release param-
eter Q/Cp = AT can be determined from an estimate of the peak temperature
(e.g., from an experiment) or from the heat of combustion of the system
under consideration and a representative heat capacity.
Method of Solution
Solution of the full set of Eqs. (8-18) or the flame-sheet approximation
in Eqs. (69-71) proceeds with an adaptive nonlinear boundary-value method.
Our goal is to obtain a discrete solution of the governing equations on the
mesh M:
M = {-LF = v0 < yi < ... < ym = Lox} (80)
With the continuous differential operators replaced by finite-difference
expressions, we convert the problem of finding an analytic solution of the
governing equations to one of finding an approximation to this solution at
each point of the mesh M. We seek the solution (/* of the nonlinear system
of difference equations
F(U) = 0 (81)
For an initial solution estimate IP that is sufficiently "close" to £/*, the
system of equations in (81) can be solved by Newton's method. We write
J(Uk)(Uk +l
- Uk) = -XkF(Uk), A: = 0, 1, ... (82)
where Uk denotes the kth solution iterate, \k the kih damping parameter33
(0 < X < 1), and J(Uk) = 3F(Uk)/dU the Jacobian matrix. A system of
linear block tridiagonal equations must be solved at each iteration for
corrections to the previous solution vector. In the counterflow diffusion
flame problem, the cost of forming (we use a numerical Jacobian) and
factoring the Jacobian matrix can be a significant part of the cost of the
total calculation. In such problems, we apply a modified Newton method
in which the Jacobian is re-evaluated only periodically.34
Efficient solution of combustion problems, such as the counterflow dif-
fusion flame, requires that the computational mesh be determined adap-
tively. Many of the methods that have been used to determine adaptive
grids for two-point boundary-value problems can be interpreted in terms
of equidistributing a positive weight function over a given interval. 27 35 We
say that a mesh M is equidistributed on the interval [ — LF, Lox] with respect
to the nonnegative function W and the constant C if
p
Wdy = C, ; = 0, 1, m - 1 (83)
Jy/
We determine the mesh by employing a weight function that equidistributes

the difference in the components of the discrete solution and its gradient
between adjacent mesh points. Upon denoting the vector of N = K -f 3
dependent solution components by u = [«19 w 2 , ..., UN\T, we seek a mesh
M such that
f>y+i
dy < max ut — mm
Jyj dy
j = 0, 1, ..., m - 1, / = 1, 2, ..., (84)
and
f
Jy; dy 2 dy < y max
da,
mm
j = 1, 2, ..., m - 1, i = 1, 2, ..., N (85)

where 8 and y are small numbers less than 1, and the maximum and
minimum values of ut and du^/dy are obtained from a converged numerical
solution on a previously determined mesh.
Double-Jet Counterflow Diffusion Flame
As a representative example, we consider the numerical solution of a
double-jet, methane-air, counterflow diffusion flame studied experimen-
tally by Puri et al.36 To solve this problem, we first generate a flame-sheet
solution. The flame-sheet model provides initial solution profiles for the
mass flux in the transverse direction V, the similarity function /', the
temperature 71, and the major species, i.e., CH4, O2, N 2 , CO2, and H2O.
The initial profiles for the minor species in the full kinetics calculation are
approximated by Gaussian profiles that are centered in the reaction zone
and have peak heights of at most a few percent. To guarantee mass con-
servation in the starting estimate at each grid point, the N2 mass fraction
is set equal to 1 minus the sum of the other species. The detailed kinetics
mechanism used in the calculations is listed in Table 1 (see also Ref. 36).
M. D. SMOOKE 195
The boundary conditions for this problem are given by

V = 5.0041 x 10'2 (86)
/' - 1.341 (87)
yCH4 = i.o, y^CH4 = o (88
T = 300.15 K (89)
at z = —LF= —0.6 cm and
V - -8.566 x 10-2 (90)
/' = 1.0 (91)
y02 - 0.233, YN2 = 0.767, *W,N2 - 0 (92)
r - 300.15 K (93)
at z = L0^ = 0.8876 cm. The mass-flow rate is in units of grams per
square centimeter per second, and the densities of the fuel and the oxidizer
mixtures were used in obtaining the value of the similarity function at the
fuel jet. The boundary conditions for the mass-flow rates correspond to a
fuel duct velocity of 76.82 cm/s and an oxidizer duct velocity of -73.39
cm/s. Since the mass conservation equation is first order in space, the
boundary condition for the mass-flow rate of the oxidizer jet overspecifies
the problem. As a result, we calculate the strain rate as an eigenvalue by
introducing the trivial differential equation,
I -«
The calculated value of the strain rate was determined to be a = 54.67 s" .1
<*>
The adaptive two-point boundary-value solver was used to generate the
flame-sheet starting estimate and the detailed kinetics calculations on the
computational domain. The flame-sheet solution was computed on a non-
uniform grid containing 37 points. This solution was then used as the
starting estimate for a detailed kinetics calculation in which the temperature
was set equal to the flame sheet temperature profile. One hundred adaptive
time steps were taken to help bring the solution within the domain of
convergence of Newton's method on the 37-point grid. After the time steps,
Newton's method converged with only one iteration. Once this solution
was obtained, the mesh was refined and a solution was calculated on the
finer grid. This procedure continued until the adaptive mesh criteria were
satisfied. The refined fixed temperature solution was then used as the
starting estimate for the complete fluid dynamic-thermochemistry solution.
Two additional grid refinements were performed to obtain a final solution
on a grid consisting of 95 nonuniform points. On the refined grids, Newton's
method converged with only 10-20 additional time steps. The mesh spacing
was such that 600 equispaced points would have been needed to obtain
comparable accuracy. The total CPU time for the entire procedure was
approximately 2 h on a Multiflow TRACE 14/200.
Table 1 Methane-air reaction mechanism
Reaction A 3 E
1. /^TT _i_ T T _i. f^T-T
1.90E 4- 36 -7.000 9,050
2. CH4 + O, ^± CH, + HO, 7.90E + 13 0.000 56,000
3. CH4 4- H ^± CH, 4- H, 2.20E 4- 04 3.000 8,750
4. CH4 H- O ^± CH, 4- OH 1.60E 4- 06 2.360 7,400
5. CH4 4- OH ^ CH, + H,O 1.60E 4- 06 2.100 2,460
6. CH,O 4 OH ^± HCO 4- H,0 7.53E + 12 0.000 167
7. CH2O 4 H ^± HCO 4- H, 3.31E + 14 0.000 10,500
8. CH,O 4- M ^ HCO + H 4- M 3.31E + 16 0.000 81,000
9. CH2O 4 O ^± HCO 4- OH 1.81E + 13 0.000 3,082
10. HCO 4- OH ^± CO 4- H,O 5. OOE 4- 12 0.000 0
11. HCO + M^H + CO + M 1.60E + 14 0.000 14,700
12. HCO 4- H ^± CO + H, 4.00E + 13 0.000 0
13. HCO 4- O ^ OH + CO I . OOE + 13 0.000 0
14. HCO 4- O2 ^± HO, 4- CO 3.00£ 4- 12 0.000 0
15. CO 4- O 4- M — CO, 4- M 3.20E + 13 0.000 -4,200
16. CO 4- OH ^ CO, 4- H 1.5LE 4- 07 1.300 -758
17. CO 4- O, ^ CO2 4- O 1.60E 4- 13 0.000 41,000
18. CH3 4- O2 ^± CH,O 4- O 1WE + 12 0.000 25,652
19. CH3O 4- M ^ CH2O 4- H 4- M 2.40£ 4- 13 0.000 28,812
20. CH3O 4 H ^± CH,O 4- H, 2.00E + 13 0.000 0
21. CH3O + OH ^± CH,O 4- H,O I WE + 13 0.000 0
22. CH3O + O ^ CH,O 4- OH I WE + 13 0.000 0
23. CH3O 4- O, ^± CH,O 4- HO, 6.30E 4- 10 0.000 2,600
24. CH3 4- O, ^ CH,6 4- OH 5.20E + 13 0.000 34,574
25. CH3 4- O ^± CH,O 4- H 6.80£ 4- 13 0.000 0
26. CH3 4- OH ^ CH,O 4- H, 7.50£ 4- 12 0.000 0
27. CH2 + H ^± CH + H, 4.00£ + 13 0.000 0
28. CH2 4- O ^ CO + H^4 H 5WE 4- 13 0.000 0
29. CH2 + O2 ^ CO, 4- H 4- H 1.30E 4- 13 0.000 1,500
30. CH2 4- CH, ^ C2H4 + H 4WE + 13 0.000 0
31. CH 4- O ^ CO + H 4WE + 13 0.000 0
32. CH + O2 ^± CO + OH 2.00£ 4- 13 0.000 0
33. CH3 + CH3 ^ C,H6 1.70£ 4- 53 -12.000 19,400
34. CH3 4- CH, ^ C2HS + H 8. OOE + 13 0.000 26,500
35. C2H6 4- H ^± C,HS + H, 5.40E + 02 3.500 5,240
36. C,H6 4- O ^± C2HS + OH 3. OOE + 07 2.000 5,100
37. C2H6 4- OH ^± C2H, 4- H,O 6.00E 4- 12 0.000 19,400
38. C2H5 + O2 ^ C2H4 4- HO, 2. OOE 4- 13 0.000 5,000
39. C2H5 — C 2 H 4 + H 2.60E 4- 43 -9.250 52,580
40. C2H4 4- O ^ HCO 4- CH3 1.60E 4- 09 1.200 740
41. C2H4 4- OH ^± C2H3 4- H2O 4.80E 4- 12 0.000 1,230
42. C2H4 + H ^± C2H3 4- H2 1.10E + 14 0.000 8,500
43. C2H3 4- H ^± C2H2 4- H2 6.00E + 12 0.000 0
44. C2H3 4- O2 ^ C2H2 + HO2 1.58E 4- 13 0.000 10,060
45. C2H3 ^ C7H2 4- H 1.60E 4- 32 -5.500 46,200
46. C2H2 + O"^± CH2 4- CO 2.20E 4- 10 1.000 2,583
47. C2H2 + OH ^ CH2CO 4- H 3.20E + 11 0.000 200
48. CH2CO + H ^± CH3 4 CO 7.00E + 12 0.000 3,000
49. CH2CO 4- 0 ^ HCO 4- HCO 2. OOE 4- 13 0.000 2,300
50. CH2CO 4- OH ^± CH2O + HCO l.OOE 4- 13 0.000 0
51. CH2CO 4- M — CH, 4- CO 4- M l.OOE + 16 0.000 59,250
M. D. SMOOKE 197
Table 1 (cont'd) Methane-air reaction mechanism

Reaction A p E
52. C2H2 + 0 ^ HCCO + H 3.56^ + 04 2.700 1,391
53. HCCO + H ^± CH, + CO 3.00^ + 13 0.000 0
54. HCCO + O ^± CO^+ CO + H 1.2QE + 12 0.000 0
55. C2H2 + OH ^ C2H + H,O l.OOE + 13 0.000 7,000
56. C2H + O ^ CO + CH l.OOE + 13 0.000 0
57. C2H + H2 ^± C2H2 + H 3.50E + 12 0.000 2,100
58. C2H + O2 ^± CO + HCO 5. OOE + 13 0.000 1,500
59. H02 + CO — CO, + OH 5.80E + 13 0.000 22,934
60. H2 + O2 ^± 2OH 1.70E + 13 0.000 47,780
61. OH + H2 ^± H,O + H 1.17E + 09 1.300 3,626
62. H + O2 ^± OH + O 2. OOE + 14 0.000 16,800
63. O + H2 ^± OH + H 1.8QE + 10 1.000 8,826
64. H + O2 + M ^ HO, + Ma 2.10E + 18 -1.000 0
65. H 4- O, + O, ^± HO2 + O, 6.70E + 19 -1.420 0
66. H + O2 + N2 ^± HO, + N, 6.70E -H 19 -1.420 0
67. OH + HO2 ^± H,O + O, 5. OOE + 13 0.000 1,000
68. H + HO, ^ 2OH 2.50E + 14 0.000 1,900
69. O + HO2 ^± O2 + OH 4.80E + 13 0.000 1,000
70. 2OH ^± O + H2O 6. OOE + 08 1.300 0
71. H2 + M ^ ± H + H + Mb 2.23E + 12 0.500 92,600
72. O2 + M ^ ± O - h O - h M 1.85E + 11 0.500 95,560
73. H + OH + M ^± H,O + Mc 7.50E + 23 -2.600 0
74. H + HO, ^ H, H- 6, 2.50E + 13 0.000 700
75. HO2 + HO, ^± H2O, + O, 2.00E + 12 0.000 0
76. H2O2 + M ^ OH + OH + M 1.30E + 17 0.000 45,500
77. H2O, + H ^ HO2 + H, 1.60E + 12 0.000 3,800
78. H2O2 + OH ^ H2O + HO, l.OOE + 13 0.000 1,800
a
Third-body efficiencies: kM(HÔ) = 2U64(Ar), £64(H,) = 3.3£64(Ar), kM(N.) = kM(OJ = 0.
b
c
Third-body efficiencies: *71(H2O) = 6fc 71 (Ar), MH) = 2fc 71 (Ar), MH 2 ) - 3*71(Ar).
Third-body efficiency: £73(H26) = 20A:73(Ar).
Reaction mechanism rate coefficients in the form kf = AT^ e\p( — E/RT*). Units are
moles, cubic centimeters, seconds, Kelvins, and calories/mole.
In Figs. 3-6, we compare the flame-sheet temperature profile, the axial

velocity profile, and the major species profiles with the corresponding
detailed chemistry profiles. We observe that the flame-sheet solution pre-
dicts all of the qualitative features of the detailed kinetics solution. It does
a particularly good job on the relative locations of the temperature peaks,
the double-hump velocity profiles in the heat-release region, and the stag-
nation point.
In Fig. 7, we plot the experimental (circles) and the calculated (solid
line) temperature profiles. Upon comparing the results in Figs. 4 and 7,
we observe that the calculated maximum value of the temperature occurs
on the oxidizer side of the stagnation plane, where the axial velocity reaches
a maximum negative value. We also observe that the shape and the peak
2000 -
CQ
1000 -
IN CM
Fig. 3 Comparison between calculated flame-sheet temperature profile (dotted line)
and calculated finite-rate chemistry temperature profile (solid line) in double-jet
problem.
values (maximum experimental = 1950 K and maximum computational

= 1985 K) of the experimental and calculated temperature profiles are in
excellent agreement. The profiles are, however, shifted by approximately
0.6 mm. The discrepancy in the temperature profiles can be attributed
partially to errors associated with measuring the velocity of the reactants
at the exit of the ducts. Hot-wire anemometer mesurements of the radial
profile of the normal component of the velocity near the exit of the duct
showed variations of ±5 cm/s. Numerical calculations have shown that a
5-cm/s variation in the velocity of either jet can shift the location of the
flame by as much as 0.2 mm. In addition, it has been shown previously37
that the flowfield in the experiment consists of two inviscid rotational zones
extending from the stagnation plane to the ducts with a viscous region
between them. In the computational model, however, we have assumed
that the outer flow is inviscid and irrotational and that the tangential com-
ponent of the flow velocity is not zero (i.e., /' =£ 0) at the exit of the
duct. Asymptotic analysis38-39 of the flame structure shows that the outer
flow only affects the flame location and in the inner reaction zone there
is a balance between convective and diffusive processes. Hence, it is rea-
M. D. SMOOKE 199
100 —i———|———i———i———i———i———|———i———i———r
o 50 -
w
in
o
O
O
>
K-q
o -50
-100
-0.5 0.0 0.5
Y IN CM
Fig. 4 Comparison between calculated flame-sheet normal velocity profile (dotted
line) and calculated finite-rate chemistry normal velocity profile (solid line) in double-
jet problem.
sonable to compare the numerical calculations and experimental measure-

ments even though the outer flows are different. Alternatively, one could
reformulate the problem using the full two-dimensional set of governing
equations and replace the strain-rate parameter with a scaled pressure
gradient. By then measuring the velocity gradients near the jet exits, one
could determine the appropriate boundary conditions for /' using the con-
tinuity equation (see also Ref. 40). It is worthwhile to point out that, in
this formulation, one must still solve a set of nonlinear two-point boundary-
value problems. However, the formulation is somewhat modified over the
boundary-layer model described earlier.
In Figs. 8-10, we compare the experimental and computational profiles
of the major stable species in the flame. From Fig. 8, we see that there is
excellent agreement for the CH4 and N2 profiles but that the O2 concen-
tration in the reaction zone is higher than the results of the calculations.
We also see from Fig. 9 that the agreement between the numerical cal-
culations and the experimental measurements for H2, H2O, CO, and CO2
is good. We observe similar results if we compare the acetylene, ethylene,
and ethane profiles in Fig. 10.
10° =F T
ID" 1 -
GO
^
O
E-
10"
10"
-0.5 0.0 0.5
Y IN CM
Fig. 5 Comparison between calculated flame-sheet CH4, O2, N2 major species pro-
files (dotted lines) and calculated finite-rate chemistry major species profiles (solid
lines) in double-jet problem.
In Fig. 11, we illustrate calculated profiles for some of the trace species
and radicals in the flame. Westbrook and Dryer41 and Warnatz 42 have
shown that oxidation of CH4 proceeds through two parallel paths. Methane
can be attacked by the H, O, or OH radical to form CH3. The methyl
radical can also be formed by thermal decomposition of CH4. Subsequently,
CH3 can be oxidized to form methoxy radicals and/or formaldehyde, or
the CH3 radicals can recombine to form C2H6. Attack on the ethane mol-
ecule can then produce C2H4 and the ethylene can then yield C2H2. The
presence of CH2O, CH3, and C2H6 in Figs. 10 and 11 shows that both
paths are available for the oxidation of the methyl radical. Figure 11 also
illustrates that the peak values of the H, O, and OH radicals are observed
on the lean side of the flame and that their concentration is small at the
reaction zone and on the fuel side. This is due to the high affinity of CH4
to the radicals H, O, and OH, which causes their concentration to increase
only after the concentration of CH4 has been reduced significantly. As a
result, the oxidation of CO and H2 occurs predominantly on the lean side
of the reaction zone. This conclusion is supported by the calculated profiles
of CO, H2, and CH4 shown in Figs. 8 and 9. After the methane has
M. D. SMOOKE 201
10° F
-0.5 0.0 0.5

Y IN CM
Fig. 6 Comparison between calculated flame-sheet H2O, CO2 major species profiles
(dotted lines) and calculated finite-rate chemistry major species profiles (solid lines)
in double-jet problem.
disappeared, the radicals H, O, and OH attack CO and H2 to form the

products CO2 and H2O.
III. Axisymmetric Coflow Diffusion Flames

Conclusions derived from studies of laminar flames are important in
characterizing the combustion process occurring in turbulent flames, in
improving engine efficiency, and in understanding the formation of com-
bustion-based pollutants. In particular, current applications of the laminar
flamelet model (see, for example, Ref. 43) to turbulent nonpremixed com-
bustion have been in relatively simple geometric configurations. The ability
to predict reliably the instantaneous local extinction of a turbulent diffusion
flame may require application of the flamelet model to more complex
multidimensional systems. In addition, an understanding of those factors
that affect flame extinction is critical in improving engine efficiency. Com-
bustor goemetry and flowfield patterns may produce flames with sharp
edges. Extinction along these edges results from incomplete combustion
and, hence, produces a decrease in fuel efficiency.
2000 -
co
1000
-0.5 0.0 0.5

Y IN CM
Fig. 7 Comparison between measured (O) and calculated values (solid line) of
temperature profile in double-jet problem.
By studying laminar diffusion flames, we can identify the important

reactions controlling extinction, and we can identify the important species
involved in pollutant formation while providing information on the fluid
mechanics of the flames. One of the simplest two-dimensional flame con-
figurations with practical importance is the axisymmetric coflow diffusion
flame. Although axisymmetric flames are important in combustion appli-
cations, they have received relatively little attention in theoretical flame
studies. Part of this neglect is due to the two-dimensional nature of the
problem coupled with the complexities associated with the combined effects
of transport phenomena and chemical processes. In this section, we con-
sider confined axisymmetric diffusion flames in which a fuel jet discharges
into a laminar airstream. The tubes through which the fuel and the oxidizer
flow are concentric and have radii Rf and R0, respectively. The two gases
make contact at the outlet of the inner tube, and a flame that resembles
a candle results (see Fig. 12). A cylindrical shield surrounds the fuel and
oxidizer tubes. One can also consider unconfined systems by appropriately
modifying the boundary conditions in the radial direction as r becomes
large.
M. D. SMOOKE 203
-0.5 0.0 0.5

Y IN CM
Fig. 8 Comparison between measured CH4 (Q), O2 (O), and N2 (0) profiles and
corresponding calculated values (solid lines) in double-jet problem.
Our model of axisymmetric laminar diffusion flames considers the full

set of two-dimensional governing equations. In primitive variables where
r and z denote the radial and axial coordinates, respectively, the governing
equations can be written in the form
Continuity:
1 ;j ;^rvj» \
= 0 (95)
Radial momentum:
dvr d 1 dvr\ d I dvr

rpvr — + rpvz — 2 — ra — - — ra —
dr " " * ^ dr I dz \ ^ dz
2 d d(rvr) 2 d dvz dVz

+ m— - — + 2a —
dr 3^ ^ dz dr
2 JJL a . . 2 dvz dp
r) - - JJL —- + r — - 0
- - - — v(rvr) (96)
3 r dr 3 ^ dz dr
10° T I I
10 -1
oo
^
a
£ CO,
O
^-2
i
10"
10" _J____I___1_
-0.5 0.0 0.5
Y IN CM
Fig. 9 Comparison between measured H2O O, CO2 (O), H2 ( + ), and CO (0)
profiles and corresponding calculated values (solid lines) in double-jet problem.
Axial momentum:
dv7 dv d dvz\ „ d I dvz

rpv
H rr —- + rpvz — - — rjji — - 2 — I r|x —
dr dz \ dr\^drj dz \ dz
(
\
2 d d(r^ 2 d 3
( 'U dvr
ar j +
3dz ^ dz) "ari1 fa ,
d
± _ rpg = 0
+ r
az
(97)
Species:
f - (rpYkVk)r + —
dr dz
- rWkwk - 0, A: = 1, 2, ..., i (98)
M. D. SMOOKE 205
10C
10"
,
o
p—•
b
<
C£
tn
_o
10
io- -0.5 0 .0 0.5

Y IN CM
Fig. 10 Comparison between measured C2H2 O, C2H4 (0), and C2H6 (O) profiles
and corresponding calculated values (solid lines) in double-jet problem.
Energy:
dT dT\ d I dT\ d ( dT
r 27 — — — \r\ — — — \r\ —
i_ dr dz \ dr\ dr dz V dz
K
(99)
k=\
Equation of state:
p = pW/RT (100)
The system is closed with appropriate boundary conditions on each side
of the computational domain. For the confined flame, we have
Axis of symmetry (r = 0):
= vr = = = k = 1,2, ..., K (101)

dr dr dr dr
icr2 T
iz;
o
u
10- I
-0.5 0.0 0.5
Y IN CM
Fig. 11 Calculated finite-rate chemistry profiles of some of the minor species and
radicals in double-jet flame.
Exit (z
k = 1,2, ..., (102)

dz dz dz
Inlet (z - 0):
P = P/ (103a)
vr - 0 (103b)
v2 - ^ (103c)
* = 1,2, (103d)
r = r7 (103e)
M. D. SMOOKE 207
^V
r^ ^3j -^
1
i R0 i ^
^F Kx
-x" ^**
AIR AIR
—^c~: Fig. 12 Schematic of confined

axisymmetric laminar diffusion
FUEL flame.
< r < Ro
P = Po (104a)
vr = Q (104b)
Vz = V0 (104c)
— y
^ — -*-' (104d)
T= T0 (104e)
Outer zone (r = R0):

v, = vz = 0 (105a)
—— = 0, k = 1,2, ..., K (105b)
T = Twall (105c)
The subscripts / and O refer to the inner and outer jets, respectively, and
P/> Po> v f , v0, Ykn Yko, Tj, T0, and 7wall are specified quantities.
We can reduce the number of equations to be solved by introducing the
vorticity and the stream function. 44 The vorticity is a measure of the coun-
terclockwise rotation in the flow. In particular, formulation of the vorticity
transport equation serves to eliminate the pressure as one of the dependent
variables. We define the vorticity such that
w
= IT
dz ~ IT
dr (106)
The stream function i|j is used to replace the radial and axial components
of the velocity vector by a single function. It is defined in such a way that
the continuity equation is satisfied identically. We have
prvr = - ^ (107)
pruz = ^ (108)
With the definitions in Eqs. (106-108), the governing equations become

Stream function:
a /1 ai|A a /1 d\b\
— —— + — —— + c a - 0 v(109
dz\rp d z j dr \rp dr J '
Vorticity:
f
' ^ ~\ r /
I _ A I r3 A ( f i wW
dr/ dr \r dz/ \ dz \ dz \ r
Species conservation:
d_
K K +
dz\' drl dr \' dz ' '
M. D. SMOOKE 209
+ ^ (rpYkV^) - rWkwk = 0, k = 1, 2, ..., K (111)

oz
Energy:
a
r S hkV/kwk = 0 (112)
k=\
where the components of the iso operator are given by (dp/dz, —dp/dr).
The boundary conditions for the unconfined flame in the stream function-
vorticity formulation are written in the form
Axis of symmetry (r = 0):
r} V rl T
i|i = CD = —^ = — = 0, A: = 1, 2, ..., K (113)

dr dr
Exit (z-» oc):
^ = acoân = ar= ,^,2,...,^ (H4)

v y
az az az az
Inlet (z - 0):
r <R
n = Yk/, k = 1,2, ..., tf

r = r7 (iisd)
/?7 < r < /?0
41 = ip/V//?? + ip0ô(^2 - /??) (H6a)
<o = 0
n = Yko, k = 1,2, ..., /C (116c)
T - r0 (116d)
Outer zone (r = R0):
- /??) (H7a)
—— = 0, k = 1, 2, ..., A: (117c)
^r - rwall
where quantities with the subscript np in the vorticity boundary condition

are evaluated at the grid points next to the shield.
In addition to the variables already defined, T denotes the temperature;
Yk the mass fraction of the /cth species; p the pressure; vr and vz the
velocities of the fluid mixture in the radial and axial directions, respectively;
p the mass density; Wk the molecular weight of the /cth species; W the
mean molecular weight of the mixture; R the universal gas constant; X the
thermal conductivity of the mixture; cp the constant-pressure heat capacity
of the mixture; cpk the constant-pressure heat capacity of the /cth species;
wk the molar rate of production of the /cth species per unit volume; hk the
specific enthalpy of the /cth species; g the gravitational constant; jx the
viscosity of the mixture; and Vkr and Vkz the diffusion velocities of the /cth
species in the radial and axial directions, respectively. The thermodynamic
and transport relations as well as the form of the chemical production rates
are identical to those used in the modeling of one-dimensional counterflow
diffusion flames (see Sec. II).
Flame-Sheet Model
As in the counterflow problem, the governing equations in Eqs. (95-
105) or (109-117) are highly nonlinear and require a starting estimate for
the finite-difference solution method. The use of a thin, infinitely fast,
global reaction model is again a natural starting point (see also Refs. 15-
18). The development of the two-dimensional flame-sheet model parallels
much of the analysis of the one-dimensional system. Starting with a single
overall irreversible reaction of the type
Fuel (F) + Oxidizer (OX) -» Products (P) (118)
in the presence of an inert gas (N) and upon introducing the assumptions
made in Eqs. (28-42), one can show that each of the Shvab-Zeldovich
variables satisfies the partial differential equation
— \ Zk —
di|A a / dv|A]
— — \ Zkk —
d(k
- — rpDk ——
dz
dz \ dr dr\ dz\ dr V dr
k = F, OX, P, N (119)
One can show that all of the Zk are proportional to each other and to a
conserved scalar S that satisfies an equation similar in form to Eq. (119).
We can recover the temperature and the major species profiles from the
conserved scalar by first obtaining an estimate of the location of the flame
front. In the Shvab-Zeldovich formulation, fuel and oxidizer cannot co-
M. D. SMOOKE 211
exist. Hence, Yox = 0 on the fuel side of the flame, and YF = 0 on the
oxidizer side. If we denote variables at the flame front with the subscript
/, then it can be shown (see Sec. II) that, for a fixed value of the axial
coordinate z, the location of the flame front is defined such that
S(rf)\fmd , = Sf= YOXo/[YOXo + (WoxvoxIWFvF) YFl] (120)
The location of the flame front can be obtained by solving Eq. (120) at
each axial coordinate level. If we utilize the proportionality of the Z^'s to
5" along with the Shvab-Zeldovich expressions, we can derive expressions
for the temperature and species on the fuel and oxidizer sides of the flame
much as was done for the counterflow model. In the two-dimensional
model, however, this process is performed at each level of the z coordinate.
Method of Solution
Our goal is to obtain a discrete solution of the governing equations in
two dimensions on the mesh M2, the initial nodes of which are formed by
the intersection of the lines of the mesh Mr,
Mr = {0 = r0 < /-! < ... < rt ... < rMr = R0} (121)
and the mesh Mz,
Mz = {0 = z0 < Zl < ... < z, ... < zMs = Z} (122)
Computationally, we combine a steady-state and a time-dependent solution
method. A time-dependent approach will be used to help obtain a con-
verged numerical solution on an initial coarse grid using the flame-sheet
starting estimate. Grid points will then be inserted adaptively, and the
steady-state solution procedure will be used to complete the problem.
We approximate the spatial operators in the governing partial differential
equations by using finite-difference expressions. Diffusion terms are ap-
proximated by centered differences and convective terms by upwind ap-
proximations. The problem of finding an analytic solution of the equations
is then converted into one of finding an approximation to this solution at
each point (r,,z;) of the mesh in two dimensions. With the difference equa-
tions written in residual form, we again solve for the solution U* of the
system of nonlinear equations
F(U) = 0 (123)
For an initial solution estimate U° that is sufficiently close to U*, the system
of nonlinear equations in Eq. (123) can be solved using Newton's method.
We point out that with the spatial discretizations used in forming Eq.
(123), the Jacobian matrix can be written in block-nine diagonal form. For
problems involving detailed transport and complex chemistry, it is often
more efficient to evaluate the Jacobian matrix numerically vs analytically.
The numerical procedure we implement extends the ideas outlined by
Curtis et al.45 We form several columns of the Jacobian simultaneously
using vector function evaluations and the Jacobian's given sparsity struc-
ture. If to each column of the Jacobian we associate the i and ; values of
the node corresponding to the column's diagonal block, then all columns
of the Jacobian having the same value of the parameter
a = (i + 3/)mod 9 (124)
can be evaluated simultaneously. Ideas along these lines have also been
explored by Newsam and Ramsdell46 and Coleman and More.47 Once the
Jacobian is formed, we solve the Newton equations with a block-line suc-
cessive over relaxation (SOR) method. The Newton iteration continues
until the size of \\Un + 1 - Un\\2 is reduced appropriately.
The solution of the governing equations in the axisymmetric problem
contains regions in each coordinate direction in which the dependent var-
iables have steep fronts and sharp peaks. Efficient solution of these prob-
lems requires that these regions be resolved adaptively. Adaptive mesh
refinement in two dimensions can proceed along several different paths.
The simplest procedure involves determining the grid points of the mesh
M2 by equidistributing positive weight functions over mesh intervals in both
the r and z directions (see, for example, Refs. 48 and 49). Specifically, we
attempt to equidistribute the mesh Mr with respect to the nonnegative
function Wr and constant Cr for each of the Mz + 1 horizontal grid lines.
We write
Cn+ i
Wr dr < Cr, i = 0,1, ..., Mr - 1 (125)
for/ = 0, 1, ..., Mz. Similarly, we attempt to equidistribute the mesh Mz

with respect to the nonnegative function Wz and constant Cz for each of
the Mr + 1 vertical grid lines. We have
1+ 1
Wz dz < Cz, j = 0, 1, ..., Mz - 1 (126)
f o r / - 0, 1, ..., Mr.
In implementing the two-dimensional adaptive grid strategy, first, we
solve the boundary-value problem on an initial coarse grid. Then we test
the inequality in Eq. (125) one r subinterval at a time for all of the / grid
lines and all of the dependent solution components. If the inequality is not
satisfied, a grid point is inserted at the midpoint of the r subinterval in
question for/ = 0, 1, ..., Mz. Once this procedure has been completed in
the r direction, we reverse the process and begin again in the z direction.
This procedure produces an orthogonal tensor product grid in which the
coordinate lines connect opposite boundaries of the computational domain.
The weight functions in the equidistribution procedure are chosen such
that the grid points are placed in regions of high gradient and high curvature
activity of the solution with the goal of reducing the local discretization
error. We use a combination of first and second derivatives of the solution
profiles. The particular combinations of function and slope and the values
of Cr and Cz can be changed to produce a solution to a desired level of
accuracy.
M. D. SMOOKE 213
In our adaptive grid strategy, the equidistribution condition is checked

one mesh interval at a time and grid points are added appropriately. The
coarse grid solution is then interpolated linearly onto the new finer grid.
The interpolated result serves as an initial solution estimate for the iteration
procedure on the finer grid. The process is continued on successively finer
and finer grids until several termination criteria are satisfied.
The formation of the Jacobian and its partial factorization in the block-
line SOR method accounts for a substantial part of the cost of the diffusion
flame calculation. As a result, the use of a modified Newton method in
which the Jacobian is re-evaluated periodically is recommended. The im-
mediate implication of applying the modified Newton method is that the
partial factorization of the Jacobian can be stored and each modified New-
ton iteration can be obtained by performing relatively inexpensive block-
line SOR back substitutions. The problem one faces when applying the
modified method is how to determine whether the rate of convergence is
fast enough. If the rate is too slow, we want to change back to a full Newton
method and make use of new Jacobian information. If the convergence
rate is acceptable, we want to continue performing modified Newton it-
erations.
We anticipate that as the size of the mesh spacing gets smaller, the
interpolated solution should become a better starting estimate for Newton's
method on the next finer grid. For a class of nonlinear boundary-value
problems, Smooke and Mattheij 50 have shown that there exists a critical
mesh spacing such that the interpolated solution lies in the domain of
convergence of Newton's method on the next grid. As a result, the hy-
potheses of the Kantorovich theorem26 are satisfied, and the sequence of
successive modified Newton iterates can be shown to satisfy a recurrence
relation scaled by the first Newton step.30 As a result, if in the course of
a calculation we determine that the size of the (n + l)st modified Newton
step is larger than the value predicted by the theorem, we form a new
Jacobian and restart the iteration count.
The course to fine grid strategy and the flame-sheet starting estimate
help to eliminate many of the convergence difficulties associated with solv-
ing the governing equations directly. Nevertheless, to obtain a starting
estimate on the initial grid that lies in the convergence domain of Newton's
method, we apply a time-dependent iteration to the flame-sheet solution.
We remark that fundamentally there are two mathematical approaches for
solving flame problems—one uses a transient method and the other solves
the steady-state boundary-value problems directly. In general, the transient
methods are robust but computationally inefficient compared with the
boundary-value methods, which are efficient but have less desirable con-
vergence properties. Most of the numerical techniques that have been used
to solve one-dimensional flame problems have employed a time-dependent
method. Variations of this approach have been considered by a variety of
researchers (see, for example, Refs. 51-64). In these methods, the original
nonlinear two-point boundary-value problem is converted into a nonlinear
parabolic mixed initial-boundary-value problem. This is accomplished by
appending the term d( )ldt to the left-hand side of the conservation equa-
tions. This same procedure can be employed in our two-dimensional cal-

culations. We obtain
^ = F(U) (127)
with appropriate initial conditions. If the time derivative is replaced, for

example, by a backward Euler approximation, the governing equations
can be written in the form
®(Un +l} = F(Un + 1) - [(Un +l - U")/T" + I] = 0 (128)
n
where for a function g(t) we define g = g(f) and where the time step
T/* + i _ £« + i _ p ^t eacj1 time i eve i 9 we must solve a system of nonlinear
equations that looks very similar to the nonlinear equations in Eq. (123).
Newton's method can again be used to solve this system. The important
difference between the system in Eqs. (123) and (128) is that the diagonal
of the steady-state Jacobian is weighted by the quantity I/T"+ I . This pro-
duces a better conditioned system and the solution from the nth time step
ordinarily provides an excellent starting guess to the solution at the (n +
l)st time level. The work per time step is similar to that for the modified
Newton iteration, but the timelike continuation of the numerical solution
produces an iteration strategy that will, in general, be less sensitive to the
initial starting estimate than if Newton's method were applied to Eq. (123)
directly. As a result, when we ultimately implement Newton's method on
the steady-state equations directly, we obtain a converged numerical so-
lution with only a few additional iterations. This time-dependent starting
procedure can also be used on grids other than the initial one. The size of
each time step is chosen by monitoring the local truncation error of the
time discretization process (see also Ref. 12).
Confined Axisymmetric Laminar Diffusion Flame

We apply the flame-sheet starting estimate and the computational method
discussed in this section to a confined, methane-air, axisymmetric laminar
diffusion flame studied experimentally by Mitchell.16 We use a 42-reaction,
15-species "Cy subset of the reaction mechanism in Table 1. The exper-
imental configuration is such that the radius of the inner fuel jet Rj =
0.635 cm, the radius of the outer oxidizer jet R0 = 2.54 cm, and the length
of the tubular Pyrex shield Z = 30.0 cm. Fuel is introduced through the
center tube and air through the outer coflow. The boundary conditions at
the inlet are given by
Inlet (z = 0):
r < Rj
T - 298 K (129a)
^CH4 = 1 . 0 , Yk = 0, k ± CH4 (129b)
vr = 0.0 cm/s (129c)
vz = 4.5 cm/s (129d)
M. D. SMOOKE 215
R2 < r < R0
T = 298 K (130a)
y02 = 0.232, yN2 = 0.768, Yk
= 0, k + O2, N2 (130b)
vr = 0.0 cm/s (130c)
vz = 9.88 cm/s (130d)
The appropriate boundary conditions for the stream function-vorticity sys-
tem can be formed using the definition in Eq. (12) along with Eqs. (115)
and (116). The shield temperature is kept constant at 298 K.
The flame-sheet model provided initial solution profiles for the stream
function, the vorticity, the temperature, and the major species, i.e., CH4,
O2, N2, CO2 and H2O. The starting estimates for the minor species in the
full chemistry solution were approximated by Gaussian profiles that were
centered at the location of the flame sheet on each axial level. They had
peak heights of at most a few percent. To conserve mass in the starting
estimate, the N2 mass fraction was reduced accordingly. The flame-sheet
starting estimate required approximately 150 adaptive time steps and five
Newton iterations to converge. Once the flame-sheet estimate was calcu-
lated, we solved the full set of governing equations in a two-step procedure.
First, we determined a solution to the stream function, vorticity, and species
equations [Eqs. (15-17)] based on the flame-sheet temperature profile.
This fixed flame-sheet temperature solution (rOUT) was then used as input
to the full fluid dynamic-thermochemistry model [Eqs. (15-18)] in which
the energy equation was included (T IN ). This procedure helped to reduce
both convergence difficulties and the total CPU time and is similar to the
two-pass solution method used in the solution of adiabatic premixed lam-
inar flames34 and counterflow diffusion flames.14
The flame sheet and the first roUT calculation were performed on a 40
x 28 tensor product grid. One hundred fifty-five adaptive time steps were
required to reduce the norm of the TOUT steady-state residuals below 1.0
x 10~3. This was sufficient to bring the numerical solution within the
convergence domain of Newton's method. After the time steps, Newton's
method converged with only seven iterations. Once this solution was ob-
tained, the mesh was refined and a solution was calculated on a finer grid.
This procedure was continued until a refined 75 x 41 grid was obtained.
The refined fixed-temperature solution was then used as the starting es-
timate for the complete fluid dynamic-thermochemistry solution. As the
computational mesh was refined, Newton's method typically converged
with a smaller number of time steps than on the coarser grids. The mesh
spacing was such that 32,000 equispaced points would have been needed
to obtain comparable accuracy. The total CPU time for the entire proce-
dure was approximately 150 h on an FPS-264, which is consistent with a
single one-dimensional counterflow diffusion flame calculation that re-
quires between 2 and 3 h on the same machine.
Mitchell16 has measured radial temperature profiles at several heights

above the burner using coated (3 mil Pt vs Pt-13% Rh) thermocouples.
The measurements were corrected for radiative losses. Major species were
sampled with quartz microprobes and then analyzed with a gas chromato-
graph. The probe had a maximum outside diameter of 0.75 mm and was
tapered to 0.15 mm at the tip. The inner diameter was approximately 50
(Jim. In Fig. 13, we compare the radial experimental and calculated tem-
perature profiles at a height of 1.2 cm above the burner. There was excellent
agreement between the two profiles from the axis of symmetry to the low-
temperature coflow region. In Figs. 14 and 15, we compare the radial
experimental and computational profiles for the major species in the flame
(CH4, O2, N2, H2O, CO2, CO, and H2) at a height of 1.2 cm above the
burner. The agreement among all of the computational and experimental
profiles is very good. Several features in Figs. 14 and 15 are worth noting,
however. From Fig. 14 it is clear that the methane diffuses to the reaction
zone, where it is consumed completely. The oxygen concentration outside
the flame region is near its inlet value and then drops nearly to zero in the
reaction zone. It then increases slightly as the symmetric axis is approached.
2000 -
1000 -
IN CM
Fig. 13 Experimental (O) and computational (solid line) temperature profiles for
the confined methane-air laminar diffusion flame at 1.2 cm above burner inlet.
M. D. SMOOKE 217
IN CM
Fig. 14 Comparison between measured CH4 (Q), O2 (O), and N2 (0) profiles and
corresponding computational values (solid line) for the confined methane-air laminar
diffusion flame at 1.2 cm above burner inlet.
This type of oxygen profile results near the flame base because the tem-
peratures are low enough to allow some oxygen to penetrate into the central
core. Oxygen is then convected upwards in the flame. In the radial direc-
tion, water and carbon dioxide maximize in the region of the peak tem-
perature. In the lower regions of the flame, the carbon monoxide and
hydrogen profiles first increase with distance from the symmetric axis, then
proceed through a maximum, and finally decrease to zero in the reaction
zone.
A more global picture of this flame can be obtained by plotting the
temperature and species contours vs the independent spatial coordinates.
In Plates 1-3 (see the color section) we illustrate the temperature isotherms
and the methane and oxygen isopleths, respectively, as functions of the
axial and radial coordinates. We note immediately the high-temperature
region extending from the boundary of the fuel and oxidizer jets to the
axis of symmetry. The figure also points out the extremely high temperature
gradients directly above the burner inlet. The temperature rises from 298
K to nearly 2000 K in approximately 0.8 mm. It is in this region that the
fuel and oxidizer first meet in stoichiometric proportion. In particular, it
10° c-
IN CM
Fig. 15 Comparison between measured H2O O, CO2 (O), CO (0), and H2 ( + )
profiles and corresponding computational values (solid line) for the confined meth-
ane-air laminar diffusion flame at 1.2 cm above burner inlet.
is clear from these figures that combustion occurs only in a thin region
above the inlet. The methane and oxygen coexist in only a very small
domain. Most of the methane disappears within 1.0 cm of the fuel jet. The
resulting heat release produces an extremely rapid rise in the temperature.
Moving in the direction of increasing r, we observe that the temperature
rapidly decreases and ultimately approaches its inlet value. In addition, if
we define the flame height as the first axial location where the maximum
temperature occurs on the axis of symmetry, we obtain a flame height of
approximately 7.0 cm.
In Plates 4-8 (see the color section) we illustrate isopleths for H2O, CO,
H2, CO2, and OH, respectively. We observe from these figures that large
quantities of H2O, CO, and H2 are produced soon after the methane has
been consumed. It is in this region that the methane is attacked by O, H,
and OH radicals, and CH3 is formed. In this region, only small amounts
of OH, H, and O exist (see, for example, Plate 8) due to the high affinity
of methane for these radicals. The oxidation of CH2O to HCO and the
subsequent formation of CO occur in regions of high methyl and formal-
dehyde concentrations.
M. D. SMOOKE 219
The oxidation of CO to CO2 proceeds primarily via the reaction

CO + OH -> CO2 + H
Hence, the rate of CO oxidation depends on the availability of OH radicals.
However, as pointed out in Ref. 41, the presence of most hydrocarbon
species inhibits the oxidation of CO. This can be attributed to the fact that
the rate of the reaction
H + O2 -» OH + O
is considerably smaller than the reaction rates of H atoms with hydrocarbon
species, and the rate of the CO oxidation reaction is also smaller than the
reaction rates of hydrocarbon species with OH. As a result, small quantities
of hydrocarbons can effectively restrict the oxidation of CO to CO2. Al-
though carbon monoxide and hydrogen are found during the oxidation of
the hydrocarbon species, it is not until after the hydrocarbons and the
hydrocarbon fragments have been consumed that the OH level rises and
CO2 is formed. In Plate 8 we observe that in the axial direction the OH
radical pool increases after the disappearance of the methane and the
formation of CO. The CO is then oxidized to form CO2. Hence, in Plate
7 we observe that carbon dioxide forms downstream of the regions of high
CO concentration.
From a fluid dynamic viewpoint, we find the velocities along the cen-
terline to be higher than the values one would expect due to the effects of
natural convection. In particular, the stream function isopleths in Plate 9
(see the color section) illustrate two large recirculation cells that are es-
tablished between the hot surface of the flame and the cooler shield. Air
is entrained into the system at the shield outlet to balance the momentum
of the inlet fuel stream and the airstream along with the frictional losses
at the shield wall. The presence of these recirculation cells reduces the
total area available for the flow of the combustion gases, and, hence, the
velocities are increased due to the combined effects of natural convection
and a reduced flow area.
IV. Conclusions
In this chapter, we have discussed the numerical solution of one- and
two-dimensional laminar diffusion flame problems with detailed transport
and finite-rate chemistry by adaptive boundary-value methods. We have
focused on the use of a flame-sheet model to generate an initial starting
estimate for Newton's method. Once a solution has been obtained on a
specified initial grid, we utilized mesh refinement techniques to increase
the resolution and accuracy of our solutions. Both counterflow and coflow
diffusion flames have been studied, and comparisons with experimental
data have been made.
It is worthwhile to point out that while one-dimensional diffusion flame
systems are readily solvable on small workstations, two-dimensional axi-
symmetric problems are solved most effectively using today's largest su-
percomputers. Even on these machines, however, the required CPU time
per calculation is in the 100 h range. This essentially precludes parameter

and continuation studies. Nevertheless, there are situations in which such
computations are important, and there are applications when even three-
dimensional calculations are needed. Although reductions in machine clock
time will improve the situation somewhat, we believe that the largest gains
to be realized over the next few years will come from the use of parallel
algorithms on architectures with large shared memories.
Acknowledgments
This work was supported in part by the U.S. Department of Energy and
the Air Force Office of Scientific Research. Computations supporting this
research were performed on the Cornell National Supercomputer Facility,
which is supported in part by the National Science Foundation, New York
State, and the IBM Corporation. Discussions with Dr. R. E. Mitchell and
Prof. D. E. Keyes are gratefully acknowledged.
References
*Sax, N. R., Dangerous Properties of Industrial Materials, Reinhold, New York,
1968.
2
Tsuji, H. and Yamaoka, I., "The Counterflow Diffusion Flame in the Forward
Stagnation Region of a Porous Cylinder," Eleventh Symposium (International) on
Combustion, Reinhold, New York, 1967, p. 979.
3
Tsuji, H. and Yamaoka, I., "The Structure of Counterflow Diffusion Flames
in the Forward Stagnation Region of a Porous Cylinder," Twelfth Symposium
(International) on Combustion, Reinhold, New York, 1969, p. 997.
4
Tsuji, H. and Yamaoka, I., "Structure Analysis of Counterflow Diffusion Flames
in the Forward Stagnation Region of a Porous Cylinder," Thirteenth Symposium
(International) on Combustion, Reinhold, New York, 1971, p. 723.
5
Hahn, W. A. and Wendt, J. O. L., "NOX Formation in Flat, Laminar, Opposed
Jet Methane Diffusion Flames," Eighteenth Symposium (International) on Com-
bustion, Reinhold, New York, 1981, p. 121.
6
Tsuji, H., "Counterflow Diffusion Flames," Progress in Energy and Combustion
Science, Vol. 8, 1982, p. 93.
7
Tsuji, H. and Yamaoka, L, "Structure and Extinction of Near-Limit Flames in
a Stagnation Flow," Nineteenth Symposium (International) on Combustion, Rein-
hold, New York, 1982, p. 1533.
8
Combustion," Progress in Energy and Combustion Science, Vol. 10, 1984, p. 319.
9
Miller, J. A., Kee, R. J., Smooke, M. D., and Grcar, J. F., "The Computation
of the Structure and Extinction Limit of a Methane-Air Stagnation Point Diffusion
Flame," 1984 Spring Meeting of the Western States Section of the Combustion
Institute, Univ. of Colorado, Boulder, CO, April 1984, Paper WSS/CI 84-10.
10
Dixon-Lewis, G., David, T., Haskell, P. H., Fukutani, S., Jinno, H., Miller,
J. A., Kee, R. J., Smooke, M. D., Peters, N., Effelsberg, E., Warnatz, J., and
Behrendt, F., "Calculation of the Structure and Extinction Limit of a Methane-
Air Counterflow Diffusion Flame in the Forward Stagnation Region of a Porous
Cylinder," Twentieth Symposium (International} on Combustion, Reinhold, New
York, 1982, p. 1893.
H
Puri, I. K. and Seshadri, K., "Extinction of Diffusion Flames Burning Diluted
Methane and Diluted Propane in Diluted Air," Combustion and Flame, Vol. 65,
1986, p. 137.
M. D. SMOOKE 221
12
Smooke, M. D., Miller, J. A., and Kee, R. J., "Solution of Premixed and
Counterflow Diffusion Flame Problems by Adaptive Boundary Value Methods,"
Numerical Boundary Value ODEs, edited by U. M. Ascher and R. D. Russell,
Birkhauser, Boston, MA, 1985, p. 303.
13
Smooke, M. D., Puri, I. K., and Seshadri, K., "A Comparison Between Nu-
merical Calculations and Experimental Measurements of the Structure of a Coun-
terflow Diffusion Flame Burning Diluted Methane in Diluted Air,' 1 Twenty-First
Symposium (International) on Combustion, Reinhold, New York, 1986, p. 1783.
14
Keyes, D. E. and Smooke, M. D., "Flame Sheet Starting Estimates for Coun-
terflow Diffusion Flame Problems," Journal of Computational Physics, Vol. 73,
1987, p. 267.
15
Burke, S. P. and Schumann, T. E. W., "Diffusion Flames," Industrial Engi-
neering Chemistry, Vol. 29, 1928, p. 998.
16
Mitchell, R. E., "Nitrogen Oxide Formation in Laminar Methane-Air Diffusion
Flames," Sc.D Thesis, Massachusetts Institute of Technology, Cambridge, MA,
1975.
17
Mitchell, R. E., Sarofim, A. F., and Clomburg, L. A., "Experimental and
Numerical Investigation of Confined Laminar Diffusion Flames," Combustion and
Flame, Vol. 37, 1980, p. 227.
18
Smooke, M. D., Mitchell, R. E., and Grcar, J. F., "Numerical Solution of a
Confined Laminar Diffusion Flame," Elliptic Problem Solvers II, edited by G.
Birkhoff and A. Schoenstadt, Academic, New York, 1984, p. 557.
19
Miller, J. A. and Kee, R. J., "Chemical Nonequilibrium Effects in Hydrogen-
Air Laminar Jet Diffusion Flames," Journal of Physical Chemistry, Vol. 81, 1977,
p. 2534.
20
Williams, F. A., "Recent Advances in Theoretical Descriptions of Turbulent
Diffusion Flames," Turbulent Mixing in Non-Reactive and Reactive Flows, edited
by S. N. B. Murthy, Plenum, New York, 1975, p. 189.
21
Liew, S. K.,Bray, K. N. C., and Moss, J. B., "A Flamelet Model of Turbulent
Non-Premixed Combustion," Combustion Science and Technology, Vol. 27, 1981,
p. 69.
22
Liew, S. K., Bray, K. N. C., and Moss, J. B., "A Stretched Laminar Flamelet
Model of Turbulent Nonpremixed Combustion," Combustion and Flame, Vol. 56,
1984, p. 199.
23
Rogg, B., Behrendt, F., and Warnatz, J., "Turbulent Non-Premixed Com-
bustion in Partially Premixed Diffusion Flamelets with Detailed Chemistry," Twenty-
First Symposium (International) on Combustion, Reinhold, New York, 1986, p.
1533.
24
Kee, R. J., Miller, J. A., and Jefferson, T. H., "CHEMKIN: A General-
Purpose, Transportable, Fortran Chemical Kinetics Code Package," Sandia Na-
tional Labs., Livermore, CA, Rept. SAND80-8003, 1980.
25
Kee, R. J., Warnatz, J., and Miller, J. A., "A Fortran Computer Code Package
for the Evaluation of Gas-Phase Viscosities, Conductivities, and Diffusion Coef-
ficients," Sandia National Labs., Livermore, CA, SAND83-8209, 1983.
26
Kantorovich, L. V. and Akilov, G. P., Functional Analysis in Normed Spaces,
Pergamon, New York, 1964.
27
Smooke, M. D., "Solution of Burner Stabilized Premixed Laminar Flames by
Boundary Value Methods," Journal of Computational Physics, Vol. 48, 1982, p.
72.
28
Smooke, M. D., Miller, J. A., and Kee, R. J., "Determination of Adiabatic
Flame Speeds by Boundary Value Methods," Combustion Science and Technology,
Vol. 34, 1983, p. 79.
29
Miller, J. A., Mitchell, R. E., Smooke, M. D., and Kee. R. J., "Toward a
Comprehensive Chemical Kinetic Mechanism for the Oxidation of Acetylene: Com-
parison of Model Predictions with Results from Flame and Shock Tube Experi-
ments," Nineteenth Symposium (International) on Combustion, Reinhold, New
York, 1982, p. 181.
30
Miller, J. A., Smooke,M. D., Green, R. M., andKee, R. J., "KineticModeling
of the Oxidation of Ammonia in Flames," Combustion Science and Technology,
Vol. 34, 1983, p. 149.
31
Smith, H. W., Schmitz, R. A., and Ladd, R. G., "Combustion of a Premixed
System in Stagnation Flow—I. Theoretical," Combustion Science and Technology,
Vol. 4, 1971, p. 131.
32
Kanury, A. M., Combustion Phenomena, Gordon & Breach, New York, 1982.
33
Deuflhard, P., "A Modified Newton Method for the Solution of Ill-Conditioned
Systems of Nonlinear Equations with Application to Multiple Shooting," Numer-
ische Mathematik, Vol. 22, 1974, p. 289.
34
Smooke, M. D., "An Error Estimate for the Modified Newton Method with
Applications to the Solution of Nonlinear Two-Point Boundary Value Problems,"
Journal of Optimization Theory and Applications, Vol. 39, 1983, p. 489.
35
Smooke, M. D., "On the Use of Adaptive Grids in Premixed Combustion,"
AIChE Journal, Vol. 32, 1986, p. 1233.
36
Puri, I., Seshadri, K., Smooke, M. D., and Keyes, D. E., "A Comparison
Between Numerical Calculations and Experimental Measurements of the Structure
of a Counterflow Methane-Air Diffusion Flame," Combustion Science and Tech-
nology, Vol. 56, 1987, p. 1.
37
Seshadri, K. and Williams, F. A., "Flow Between Parallel Plates with Injection
of a Reactant at High Reynolds Number," International Journal of Heat and Mass.
Transfer, Vol. 21, 1978, p. 251.
38
Lirian, A., "The Asymptotic Structure of Counterflow Diffusion Flames for
Large Activation Energies," Acta Astronautica, Vol. 1, 1974, p. 1007.
39
Williams, F. A., Combustion Theory, Benjamin/Cummings, Menlo Park, CA,
1985.
40
Smooke, M. D. and Giovangigli, V., "The Structure and Extinction of Tubular
Premixed Laminar Flames," 23rd Symposium (International) on Combustion, 1991
(to be published).
41
Westbrook, C. K. and Dryer, F. L., "Chemical Kinetic Modeling of Hydro-
carbon Combustion," Progress in Energy and Combustion Science, Vol. 10, 1984,
P.I.
42
Warnatz, J., "The Mechanism of High Temperature Combustion of Propane
and Butane," Combustion Science and Technology, Vol. 34, 1983, p. 177.
43
Rogg, B., Behrendt, F., and Warnatz, J., "Turbulent Non-Premixed Com-
bustion in Partially Premixed Diffusion Flamelets with Detailed Chemistry," Twenty-
First Symposium (International) on Combustion, Reinhold, New York, 1988, p.
1533.
44
White, F. M., Fluid Mechanics, McGraw-Hill, New York, 1986, p. 222.
45
Curtis, A. R., Powell, M. J., and Reid, J. K., "On the Estimation of Sparse
Jacobian Matrices," Journal of Institute of Mathematics and Its Applications, Vol.
13, 1974, p. 117.
46
Newsam, G. N. and Ramsdell, J. D., "Estimation of Sparse Jacobian Matrices,"
Harvard Univ., Cambridge, MA, Rept. TR-17-81, 1981.
47
Coleman, T. F. and More, J. J., "Estimation of Sparse Jacobian Matrices and
Graph Coloring Problems," Argonne National Lab., Argonne, IL, Rept. ANL-
81-39, 1981.
48
Kautsky, J. and Nichols, N. K., "Equidistributing Meshes with Constraints,"
SI AM Journal of Scientific and Statistical Computing, Vol. 1, 1980, p. 499.
M. D. SMOOKE 223
49
Russell, R. D., "Mesh Selection Methods," Proceedings of the Conference for
Working Codes for Boundary Value Problems in ODE's, edited by B. Childs et
al., Springer-Verlag, New York, 1979.
50
Smooke, M. D. and Mattheij, R. M. M., "On the Solution of Nonlinear Two-
Point Boundary Value Problems on Successively Refined Grids," Applied Nu-
merical Mathematics, Vol. 1, 1985, p. 463.
51
Spalding, D. B., "The Theory of Flame Phenomena with a Chain Reaction,"
Philosophical Transactions of the Royal Society of London, Vol. 249A, 1956, p. 1.
52
Adams, G. K. and Cook, G. B., "The Effect of Pressure on the Mechanism
and Speed of the Hydrazine Decomposition Flame," Combustion and Flame, Vol.
4, 1960, p. 9.
53
Dixon-Lewis, G., "Flame Structure and Flame Reaction Kinetics. I. Solution
of Conservation Equations and Application to Rich Hydrogen-Oxygen Flames,"
Proceedings of the Royal Society of London, Vol. 298A, 1967, p. 495.
54
Dixon-Lewis, G., "Kinetic Mechanism, Structure, and Properties of Premixed
Flames in Hydrogen-Oxygen-Nitrogen Mixtures," Philosophical Transactions of
the Royal Society of London, Vol. 292, 1979, p. 45.
55
Spalding, D. B., Stephenson, D. L., and Taylor, R. G., "A Calculation Pro-
cedure for the Prediction of Laminar Flame Speeds," Combustion and Flame, Vol.
17, 1971, p. 55.
56
Bledjian, L., "Computation of Time-Dependent Laminar Flame Structure,"
Combustion and Flame, Vol. 20, 1973, p. 5.
57
Margolis, S. B., "Time-Dependent Solution of a Premixed Laminar Flame,"
58
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames I; Flame
Velocity of Freely Propagating Ozone Decomposition Flames," Berichte der Bun-
sengesellschaft fuer Physikalische Chemie, Vol. 82, 1978, p. 193.
59
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames II; Flame
Velocity and Structure of Freely Propagating Hydrogen-Oxygen and Hydrogen-
Air-Flames," Berichte der Bunsengesellschaft fuer Physikalische Chemie, Vol. 82,
1978, p. 643.
60
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames III; Structure
of Burner-Stabilized Hydrogen-Oxygen and Hydrogen-Fluorine Flames," Berichte
der Bunsengesellschaft fuer Physikalische Chemie, Vol. 82, 1978, p. 834.
61
Westbrook, C. K. and Dryer, F. L., "A Comprehensive Mechanism for Meth-
anol Oxidation," Combustion Science and Technology, Vol. 20, 1979, p. 125.
62
Westbrook, C. K. and Dryer, F. L., "Prediction of Laminar Flame Properties
of Methanol Air Mixtures," Combustion and Flame, Vol. 37, 1980, p. 171.
63
Coffee, T. P. and Heimerl, J. M., "The Detailed Modeling of Premixed, Lam-
inar Steady-State Flames. I. Ozone," Combustion and Flame, Vol. 39, 1980, p.
301.
64
Coffee, T. P. and Heimerl, J. M., "Transport Algorithms for Premixed, Lam-
inar, Steady-State Flames," Combustion and Flame, Vol. 43, 1981, p. 273.
Chapter 8
Laminar Flames in Premixed Gases
K. Kailasanath
Nomenclature
a,- = external force per unit mass, cm/s2
cpi = heat capacity of species /, erg/g-K
Dik = binary diffusion coefficient for species / and k, cm2/s
E = total energy density, erg/cm3
g = gravitational acceleration constant, = 980.67 cm/s2
hi = specific enthalpy for species /, erg/g
hi0 = specific heat of formation for species / evaluated at
temperature T0, erg/g
/ = unit tensor (nondimensional)
kf = forward chemical rate constant
kr = reverse chemical rate constant
kB = Boltzmann constant, = 1.3805 x 10~ 1 6 erg/K
Kf = thermal diffusion coefficient of species /, cm' 1
Lt = chemical loss rate of species /, s" 1
m = mass of a species, g
Hi = number density of species i, cm~ 3
TV = total number density, cm" 3
Ns = number of individual species in model
P = scalar pressure, dyne/cm2
P = pressure tensor, dyne/cm 2
q = heat flux, erg/cm 2 -s
qr = radiative heat flux, erg/cm 2 -s
Qi = chemical production rate of species /, cm" 3 ^" 1
R = gas constant, = 8.3144 x 107 erg/deg-mol
t = time, s
T = temperature, K
v = fluid velocity, cm/s
225
226 LAMINAR FLAMES IN PREMIXED GASES
vdi = diffusion velocities of species /, cm/s

7 = ratio of specific heats, = cplcv
€ = specific internal energy, erg/g
K = bulk viscosity coefficient, g/cm-s
X = thermal conductivity coefficient, g-cm/K-s 3
|x = coefficient of shear viscosity, P, g/cm-s
p = mass density, g/cm3
Superscripts
r = pertaining to electromagnetic radiation
T = transpose operation on a matrix
Subscripts
i, j, k, I = individual species
m = quantity defined for a mixture of species
I. Introduction
LAMINAR flames are among the earliest combustion problems to be

studied theoretically that required the simultaneous consideration of
both fluid dynamics and chemical kinetics for its solution. In 1881, Mallard
and LeChatelier published the results of an analysis that gave the propa-
gation velocity of a deflagration wave.1 Since then there have been nu-
merous analytical studies of various flame-related phenomena, and many
of these are discussed by Williams2 and Clavin.3 Whereas these analytical
techniques are extremely informative, they cannot yet quantitatively de-
scribe the detailed structure of flames since they all include very approx-
imate chemical kinetic schemes and transport properties. For this, one
must use numerical methods.
Outline
This chapter deals with the numerical simulation of laminar flames in
premixed gases. First, a brief survey of various efforts to solve the laminar
flame problem is described, highlighting the variety of approaches used to
tackle the laminar flame problem. Then, the basic equations to be modeled
in a comprehensive description of laminar flames, and the physical and
chemical processes represented by these equations are discussed along with
the numerical requirements to model these processes and various ap-
proaches that have been adopted. This discussion also presents the advan-
tages and disadvantages of the various approaches and indicates that the
choice of a method is dependent on the specific problem to be solved.
Two flame models are then discussed in some detail. FLAME ID is a
numerical model developed to study transient phenomena, such as the
ignition, propagation, and extinction of flames in a one-dimensional con-
figuration. FLIC2D is a two-dimensional counterpart. The various input
parameters needed for such detailed models are also discussed. To illustrate
the use of these models and to show how we can use them to learn about
the structure and dynamics of flames, several studies of laminar flames in
premixed gases are presented. These include calculations of steady-state
K. KAILASANATH 227
flames, the determination of burning velocities, and several transient prob-

lems, such as the ignition and quenching of premixed gases and the effects
of curvature on flame propagation. After a brief discussion of flammability
limits, multidimensional flame propagation is examined. Finally, the cur-
rent status of the modeling of flames in premixed gases is assessed, and
the directions for further modeling efforts are discussed.
Modeling of Laminar Flames

Computers have made it possible to calculate the detailed structure and
evolution of flames. One of the first attempts to solve the laminar flame
problem numerically was that of Hirschfelder et al.4 They formulated the
unsteady flame problem as a system of three-dimensional nonlinear partial
differential equations and solved the one-dimensional steady flame as a
two-point boundary-value problem. They used approximate methods to
estimate the mass flow rate, which is the eigenvalue of the two-point bound-
ary-value problem, and then used a numerical shooting method to obtain
the temperature and species profiles. Although the first problem they stud-
ied involved only single-step kinetics, they later applied the same solution
procedure to study flames for which the kinetics involved chain reactions. 5 - 6
In 1956, Spalding7 solved a time-dependent system of nonlinear partial
differential equations by using explicit finite-difference techniques to study
hydrazine flames. He assumed initial profiles for the temperature and
species concentrations, and obtained the steady-state burning velocity by
carrying out the computations for a sufficiently long time. Spalding's method
was used subsequently by Adams and Cook8 to study the effect of pressure
on the reaction mechanism and the speed of hydrazine flames, and by
Dixon-Lewis9 to study rich hydrogen-oxygen flames by using a set of 14
reactions to describe the kinetics. In 1968, Dixon-Lewis10 studied flame
structure by using detailed expressions for the diffusive transport coeffi-
cients. He also used this time-dependent model to study ignition by lo-
calized sources.11 The work of Dixon-Lewis and co-workers proved that
the one-dimensional unsteady flame with complex chemistry and detailed
diffusive transport coefficients could be solved by using numerical methods.
The emphasis then shifted toward the development of more economical
and improved numerical methods.
In 1971, Spalding et al.12 presented an implicit finite-difference method
in which the unsteady flame equations were transformed into a form that
could be solved by a numerical method developed by Patankar and Spalding13
for boundary-layer equations. They used simplified transport properties4
and a four-step chain reaction mechanism to study the propagation of
hydrogen-bromine flames.14 A peculiarity of this method is that it requires
the equations to be solved in a specific order after linearizing the source
terms in a particular manner. The same numerical procedure was used
later by Tsatsaronis15 to study unsteady flame propagation in mixtures of
methane, oxygen, and nitrogen by using very detailed chemical kinetic
mechanisms and transport properties.
The kinetics and propagation of methane-air flames was studied by Smoot
et al.16 by using a mixed explicit-implicit finite-difference technique. The
diffusive transport terms were solved explicitly, and the kinetic terms were
solved by using linearized implicit techniques. Bledjian17 employed a method-

of-lines technique, where the nonlinear parabolic initial boundary-value
problem is reduced to a set of nonlinear first-order initial value problems.
Sophisticated initial value problem integrators were used to solve the re-
sulting ordinary differential equations system. Instead of using finite-dif-
ference approximations to discretize the spatial derivative, Margolis18 used
a spline collocation procedure in his method-of-lines approach. Westbrook
and Dryer19 have deduced a comprehensive reaction mechanism for meth-
anol oxidation involving 84 reactions by using an implicit finite-difference
algorithm20 to solve the unsteady flame equations. They used simplified
expressions for the diffusive transport coefficients that were adjusted to
give good laminar flame speed prediction for methane-air flames.
Dixon-Lewis has also used a composite flux method21 to study a variety
of problems, such as the kinetic mechanism, structure, and propagation of
flames in hydrogen-oxygen-nitrogen flames22 and flame inhibition by or-
ganic halogen compounds.23 He discusses the ranges of applicability of the
partial equilibrium and quasisteady assumptions in relation to the distri-
bution of radical populations in the flame.22 More recently, Dixon-Lewis
has studied spherically symmetric flame propagation in hydrogen-air mix-
tures by using a Lagrangian finite-difference approach and has shown that
the computed results for the flame speed are in good agreement with
experimental observations.24
Warnatz has studied extensively both freely propagating and burner-
stabilized flames in a variety of premixed gases.25""29 He linearizes the
chemical reaction terms and solves the time-dependent equations implicitly
with assumed initial guesses for the temperature and species profiles. He
also uses a simplified transport model25 that agrees well with the complete
multicomponent formulation of diffusion and thermal conduction. In a
detailed study of flames in hydrogen-oxygen-nitrogen mixtures, he con-
cludes that predictions of laminar premixed flame properties should be as
reliable as experimental results.29
Coffee and Heimerl, using a method-of-lines approach,30 examined dif-
ferent methods of approximating multispecies transport phenomena in models
of premixed, laminar, steady-state flames. They concluded that the selec-
tion of the input values for the individual species transport properties is
more important than the selection of the approximation method.31
There have also been a number of approaches that solve a steady-state
formulation of the flame equations. These include the works of Dixon-
Lewis,32 Wilde,33 and Smooke.34'35 The advantages and the difficulties in
using steady-state solution procedures have been discussed by Smooke.34
These methods are good for obtaining burning velocities and steady-state
profiles, but cannot provide information on the ignition and development
of flames. Several more recent approaches to modeling flames are discussed
in detail in the proceedings of a workshop on modeling laminar flame
propagation.36
FLAME1D is a detailed, time-dependent, one-dimensional model de-
veloped to study the initiation, propagation, and quenching of laminar
flames.36'37 This model uses detailed chemical kinetics without any steady-
state or quasiequilibrium assumptions. It uses an asymptotic coupling method
K. KAILASANATH 229
in conjunction with time-step splitting to couple the various physical and

chemical processes. This approach allows the use of entirely different al-
gorithms for the processes represented by the different terms in the con-
servation equations. Thus, it is numerically very efficient and inexpensive.
A Lagrangian formulation is used to maintain steep gradients for a long
period of time. The model has been used for a variety of flame studies,
such as calculations of minimum ignition energies,37-38 effects of curvature
on flame propagation,39 and flammability limits.40
More recently, a two-dimensional time-dependent flame model, called
Flame Implicit Convection (FLIC2D),41 has been developed. In FLAME ID,
a Lagrangian method was chosen for convective transport because elimi-
nating the advection term in effect means eliminating numerical diffusion
from the calculation. Extending this Lagrangian approach to multidimen-
sions is extremely difficult and expensive. Therefore, an Eulerian approach
was chosen for the two-dimensional flame model. However, most Eulerian
methods are either more numerically diffusive than what is acceptable in
a flame model, or they are explicit and, hence, extremely inefficient at the
very low velocities associated with laminar flames. To circumvent these
numerical problems, the Barely Implicit Correction to Flux-Corrected
Transport42 (BIC-FCT) was developed. BIC-FCT combines an explicit,
high-order, nonlinear FCT method43-44 with an implicit correction process.
This combination maintains high-order accuracy and yet removes the time-
step limit imposed by the speed of sound. By using FCT for the explicit
step, BIC-FCT is accurate enough to compute with sharp gradients without
overshoots and undershoots. Thus, spurious numerical oscillations that
would lead to unphysical chemical reactions do not occur. The development
of this new algorithm has made it possible to model multidimensional
flames in detail. This model is also described later in some detail, and then
an application of the model to the study of flame instabilities is discussed
in the applications section of this chapter.
II. Numerical Approaches to Modeling of Flames
Physical and Chemical Processes to be Modeled

There are different levels at which a model for describing a flame can
be constructed. For example, a simple model in which the chemical kinetics
is ignored may be adequate for some application. In this chapter, the
emphasis is on detailed flame models. For purposes of this chapter, a
detailed flame model is defined as one in which, in addition to fluid con-
vection, both the chemical kinetics as well as the diffusive transport of the
various species involved in the chemical reactions are considered. Diffusive
transport processes include thermal conduction, multispecies diffusion,
thermal diffusion, and viscosity. There are many situations in which buoy-
ancy, radiation, heat, and radical losses to the walls must also be consid-
ered. This is not to say that flame models need to include all of these
processes all of the time. There are many applications in which one or
more of these processes can be neglected without significantly affecting
the solutions. So an ideal model should include all of these processes and
provide means to choose selectively among them.
The equations used to model an unsteady laminar flame are the contin-
uum time-dependent equations for conservation of the mass density p, the
individual chemical species number densities {n}, the momentum density
pi;, and the energy density E. These equations may be written as
dp
-£ = -V • (pi;)J (1)J
dt ^ ^
^= - V - (nfv) - V • (nfvdi) + Qt - L,nh i = 1,..., Ns (2)
dP* - V P
dt
and
— = -V • (Ev) - V - (i; • P) - V • (q + qr)
4- v - Z^mî 4- 2^vdl • m.a, (4)
The variables in these equations and others used later are given in the
Nomenclature.
The first term on the right-hand side of each of Eqs. (1-4) describes
the convective fluid dynamics effects. The remaining terms contain the
source, sink, coupling, external force, and transport terms that drive the
fluid dynamics. The pressure tensor P, heat flux q, total number density
N, and energy E used in these equations are defined by
P = P(N,T)1 + (l[Lm - K)(V • v)I - iLm[(Vv) -f (Vi;)7] (5)
q(N,T) - -\mVT 4- 2p,/W// + P^KTVdi (6)
N - 2>,. (7)
and
E = ±pv - v + pe (8)
This discussion considers gas-phase reactive flows at densities where the

thermal equation of state is the ideal gas equation of state,
P = NkBT (9)
The caloric equation of state is
h, = hi0 + II cpi dT (10)

K. KAILASANATH 231
which relates the enthalpies {h,} to the heat capacities, {cpi}. The enthalpies
are defined through the equation
pe = 2p/A/ - P (11)
/
The external forces {ra/a/} in Eqs. (3) and (4) can be different for each
individual species. For gravity, the accelerations are all the same,
«/ - g (12)
where g is the local strength of gravity. If Eq. (4) is rewritten as an equation
for the internal energy, dddt, the external force term drops out (see, for
example, Refs. 2 and 45). The force terms remain, however, in the mo-
mentum equation [Eq. (3)]. Electric and magnetic forces are neglected
here because we consider uncharged gases.
The chemical production terms and loss rates, {Qt} and {L,} in Eq. (2),
can be written as
Qi = £{/*/ + y*"/** + *£/*/«/**«/ (I 3 )
j j,k J-kJ
î = k\ + 2Xv «y + ^kJLjknink (14)

j M
where kf and kr are forward and reverse chemical rate constants that can
be functions of temperature and pressure.
The diffusion velocity of the species /, vdh is defined as
vdi = Vf - v (15)
where v is the fluid velocity averaged over all species and {v,} the velocities
of the fluid species in the same frame of reference as v. The {vdl} are found
by inverting the matrix equation
(vdk - vai) (16)
The source terms {G,} in Eq. (16) are defined as
-S)TP^T
These diffusion velocities are also subject to the constraint
Zp,-»* = 0 (18)
/
There is no net mass flux arising from the relative interspecies diffusion
because the total mass flux is defined as pv .
The three types of terms represented explicitly in Eqs. (1-4) are con-
vection, chemical reactions, and physical diffusion, such as thermal con-
duction and molecular diffusion. In the next section of this chapter, we
describe the solution methods and the input parameters used for each of
these types of terms in the one-dimensional flame model, FLAME1D.
Basic Numerical Requirements

A detailed model of a flame must contain accurate representations of
the convective, diffusive, and chemical processes. The individual impor-
tance of these processes varies from rich to lean flames, and is especially
notable near the flammability limits40 where the exact behavior of these
flames depends on a delicate balance among the processes. In this section,
the basic numerical requirements to solve these processes satisfactorily are
addressed.
Diffusive transport processes play a very important role in flames and
must be resolved properly. Numerically, this means that any numerical
diffusion in the calculation must be considerably less than any important
physical diffusion effect. A Lagrangian method is ideally suited for the
convective transport, because eliminating the advection term in effect means
eliminating numerical diffusion from the calculation. However, extending
Lagrangian methods to multidimensions is both difficult and expensive.
Therefore, Eulerian methods have to be considered for multidimensional
flames. Eulerian methods have been used successfully to simulate flames,
but these methods need finer grids than their Lagrangian counterparts to
minimize numerical diffusion. Minimizing numerical diffusion is also im-
portant for maintaining the sharp gradients present in flames.
An important factor in the simulation of flame propagation is the choice
of the spatial grid size or resolution. Spatial resolution is achieved by having
a large number of grid points in regions where the gradients are steep. In
a detailed flame model, steep gradients in different species usually occur
at different spatial locations, necessitating the use of fine resolution all
across the flame. The simplest gridding approach then would be to estimate
the smallest grid size required to resolve the feature of the flame under
investigation and to use this grid size over the whole computational domain.
The cost of such an approach is usually prohibitive, and, therefore, a
nonuniform grid is used where only a part of the computational domain is
finely gridded (with small grid size). In a time-dependent problem, the
flame may change its spatial location, and then an adaptive gridding ap-
proach in which the finely gridded region moves with some characteristic
feature of the flame is chosen. Some approaches to developing adaptive
grids for the flame problem are discussed in the next section of this chapter.
Since laminar flames are slow (when compared with the speed of sound),
one would like to use large time steps to study the dynamics. However, if
an explicit method is used to advance the solution in time, very small time
steps that are restricted by the speed of sound must be taken to keep the
calculations numerically stable. Explicit methods are those in which the
solution at the current time step depends only on the solutions at the
previous steps. Explicit methods are easier to code, and the cost per step
is lower than implicit methods. Implicit methods are those in which the
solution at the current time step depends also on a previous estimate of
the solution at the current time step. Thus, it requires an iterative procedure
K. KAILASANATH 233
or matrix inversion and is more expensive per time step. However, for
slowly evolving problems, implicit methods are more efficient because they
require fewer but larger time steps to obtain the solution for a given time
interval.
Solving a set of detailed chemical reaction rate equations is usually the
most expensive part of detailed flame models. In general, the numerical
accuracy required of chemistry algorithms is lower in a fluid flow situation
than in a purely chemical kinetics problem. Therefore, the emphasis on
chemistry solvers is more on speed than on higher orders of accuracy.
Again some of the recent developments in faster chemistry solvers are
discussed in the next section.
The last, but most important part of flame models is the coupling between
the various processes. One approach is to solve all of the equations si-
multaneously. This requires some form of iterative methods because the
equations are nonlinear. In general, this is a very expensive approach, but
it has been adopted successfully in many specific situations, such as the
determination of burning velocity and a steady flame structure. Another
approach is to use an operator or time-step split procedure in which the
individual processes are solved using the most efficient methods and then
the solutions are coupled appropriately. Here particular care must be taken
in the coupling procedure. This approach has been adopted successfully
in a wide variety of reactive-flow calculations, 45 and some of these cases
are discussed later in this chapter. A recent paper46 compares an operator-
split scheme and an unsplit scheme for the simulation of one-dimensional
flames, and comes to the conclusion that the split approach is slightly
superior to the unsplit approach in problems involving detailed chemistry.
Various approaches in between the process splitting and the fully coupled
approaches have also been taken and shown to give satisfactory results in
many specific situations.
Different Numerical Approaches

The brief historical overview of flame modeling in the Introduction also
provides an outline of the different approaches adopted. There are different
ways in which the various numerical approaches can be classified. One way
is in terms of explicit or implicit methods. As discussed earlier, because
of the slow rate at which flames propagate (when compared with the speed
of sound), methods that treat the convection implicitly are more efficient
than explicit algorithms. Except for the early works of Spalding,7 Dixon-
Lewis,9 and others, most recent approaches to solving the flame problem
are implicit in the solution of the convection equations. Even Spalding12
and Dixon-Lewis and co-workers32'47 have adopted implicit methods in
their more recent flame problems. There are also many hybrid approaches
that treat some of the terms explicitly. For example, in the calculations of
Smoot et al.16 and Kailasanath et al., 37 the diffusive transport terms are
treated explicitly. The process-splitting approach is a logical choice for
coupling the various processes together when different methods are used
to solve the individual processes.
Another way of classifying numerical approaches is in terms of steady
state and time dependence. To a large extent, this depends on the appli-
cation. Although time-dependent methods are sometimes used to achieve

steady-state solutions, direct steady-state solvers are definitely advanta-
geous if one is interested only in steady-state solutions. There are also
approaches in which time-dependent methods are used to march in false
time to attain the steady state. However, in general, a well-formulated
steady-state approach will be more efficient at attaining steady solutions
than a time-accurate solution procedure.
A different way of looking at numerical methods is in terms of finite
difference (or finite volumes) and finite elements. Most approaches adopted
for flame modeling have been finite-difference approaches, including the
method of lines. It is only more recently that finite-element methods are
being commonly used for fluid flow problems. They have become the
preferred approach to solve flow problems in complex geometries.48 Their
application to reactive-flow problems is even more recent.49 A sizable
drawback of finite-element methods is that they are time-consuming and
require significantly more memory than comparable finite-difference meth-
ods. However, for flames in complex geometries, finite-element methods
may become cost-effective.
III. Descriptions of Two Detailed Flame Models

In this section, two detailed flame models are described. The description
of these models highlights many of the points raised in the discussion in
the previous section. By no means should these models be considered the
final or best ones. These models themselves are continually evolving, and
it is always possible to develop special-purpose methods that outperform
these models in specific situations. The driving needs behind the two models
are also mentioned because this gives one the idea of how models evolve
and the compromises that are sometimes needed to meet the needs.
Two computational models developed at the Naval Research Laboratory
(NRL), FLAME1D and FLIC2D, have been used to study the propagation
and extinction of premixed gaseous flames. FLAME1D is a one-dimen-
sional, time-dependent Lagrangian model that has been tested extensively
and applied to various studies of flame phenomena, including calculations
of burning velocities,37'50 minimum ignition energies and quenching dis-
tances,37'38 effects of curvature and dilution39 on flame propagation, and
flammability limits.40 FLIC2D has been used to study the tendency of
hydrogen-air mixtures to show cellular structure 51 and, more recently, to
study multidimensional flame instabilities and structures.52-53 Many sub-
models in FLIC2D are based on FLAME1D, but, for completeness, both
models are described here.
FLAME1D
This is a time-dependent model37 that solves the compressible, conser-
vation equations for mass, momentum, and energy 2 - 45 in one spatial di-
mension. Since it was developed to study the initiation, propagation, and
quenching of laminar flames, it consists of algorithms for modeling con-
vection, detailed chemical kinetics and energy release, thermal conduction,
molecular and thermal diffusion of the individual species, and various
K. KAILASANATH 235
energy deposition mechanisms. The model has a modular form, and the
algorithms representing the various chemical and physical processes are
combined by using an asymptotic time-step split approach in which the
individual processes are integrated separately and then coupled to-
gether.45-54 The model also permits a variety of initial and boundary con-
ditions.
The convective transport terms in the equations are solved by using the
algorithm ADINC, a Lagrangian convection algorithm that solves implicitly
for the pressures.55 Since ADINC communicates compression and expan-
sion across the system implicitly, it overcomes the Courant time-step limit.
Since it is Lagrangian, it can maintain steep gradients computationally
without numerical diffusion for a long period of time. As discussed earlier,
this is important in flame calculations where the diffusive transport of
material and energy can govern the system evolution and, therefore, must
be calculated accurately. An adaptive regridding algorithm has been de-
veloped to add and delete computational cells. For the flame calculation,
cells are added in the region ahead of the flame so that the front always
propagates into a finely zoned region. Cells are removed behind the flame,
where the gradients are very shallow.
The chemical interactions are described by a set of nonlinear coupled
ordinary differential equations. This set of equations may be stiff when
there are large differences in the time constants associated with different
chemical reactions and are invariably stiff for combustion problems. These
equations are solved by using VSAIM, a fully vectorized version of the
selected asymptotic integration method CHEMEQ. 5657 In VSAIM, the
stiff equations are identified and solved by using a very stable asymptotic
method, while the remaining equations are solved by using a standard
classical method.
The diffusive transport processes considered in this model are molecular
diffusion, thermal conduction, and thermal diffusion. These processes are
crucial to the description of flame phenomena, since they are the mech-
anisms by which heat and reactive species are transported ahead to the
unburned gas. An interative algorithm, DFLUX, is used to obtain the
diffusion velocities without the cost of performing matrix inversions.54-58
This method has been vectorized and is substantially faster than matrix
inversions when four or more species are involved.
Problems can be set up in either planar, cylindrical, or spherical coor-
dinates, as well as variable power-series coordinates that can model one-
dimensional nozzlelike geometries. The model can also be configured with
either an open or closed boundary at one end. The open boundary simulates
an unconfined system.
FLIC2D
FLIC2D is a time-dependent, Eulerian, implicit, compressible, two-di-
mensional flame model.41 In FLAME1D, a Lagrangian method was chosen
for convective transport because eliminating the advection term in effect
means eliminating numerical diffusion from the calculation. Extending this
Lagrangian approach to multidimensions is extremely difficult and expen-
sive. Therefore, when the need arose for developing a two-dimensional
flame model, we felt that we would have more success in multidimensions

with an Eulerian method. However, most Eulerian methods are either
more numerically diffusive than what is acceptable in a flame model, or
they are explicit and, hence, extremely inefficient at the very low velocities
associated with laminar flames. To circumvent these numerical problems,
we developed BIC-FCT,42 which combines an explicit high-order, nonlin-
ear FCT method43'44 with an implicit correction process. This combination
maintains high-order accuracy and yet removes the time-step limit imposed
by the speed of sound. By using FCT for the explicit step, BIC-FCT is
accurate enough to compute with sharp gradients without overshoots and
undershoots. Thus, spurious numerical oscillations that would lead to un-
physical chemical reactions do not occur. The development of this new
algorithm has made it possible to model multidimensional flames in detail.
However, because there is always some residual numerical diffusion, even
in high-order Eulerian algorithms, we need to monitor calculations to en-
sure that numerical diffusion never becomes larger than the physical dif-
fusion processes we need to resolve.
In generalizing the flame model to two dimensions, simplifications were
made in the diffusive transport calculations in order to reduce the cost of
computations. In FLIC2D, thermal conductivity of the individual species
is modeled by a polynomial fit in temperature to existing experimental
data. Individual conductivities are then averaged by using a mixture rule37'59
to get the thermal conductivity coefficient of the gas mixture. A similar
process is used to obtain the mixture viscosity from individual viscosities.
Heat and momentum diffusion are then calculated explicitly by using these
coefficients.
Binary mass-diffusion coefficients are represented by an exponential fit
to experimental data, and the individual species-diffusion coefficients are
obtained by applying mixture rules.37 The individual species-diffusion ve-
locities are determined explicitly by applying Pick's law, followed by a
correction procedure to ensure zero net flux. 59 This procedure is equivalent
to using the iterative algorithm DFLUX58 (used in FLAME1D) to second
order.
The chemical reaction-rate equations are modeled and solved as before
by using a vectorized version of CHEMEQ, an integrator for stiff ordinary
differential equations.57 Because of the complexity of the reaction scheme
and the large number of computational cells in a two-dimensional calcu-
lation, the solution of the chemical rate equations takes a large fraction of
the total computational time. A special version41 of CHEMEQ was de-
veloped to exploit the special hardware features of the Cray X-MP.
As in the one-dimensional code, all of the chemical and physical proc-
esses are solved sequentially and then coupled asymptotically by time-step
splitting.45'54 This modular approach greatly simplifies the model and makes
it easier to test and change the model. Individual modules were tested
against known analytic and other previously verified numerical solutions.
One-dimensional predictions of the complete model were compared with
those from the Lagrangian model FLAME1D, which has been bench-
marked extensively against theory and experiment.
K. KAILASANATH 237
In summary, all of the chemical and diffusive-transport algorithms are

essentially the same in FLIC2D as in FLAME1D. Thus, given the input
data concerning the species, their transport coefficients, and the reactions
among the species, either the one- or two-dimensional model can be used.
The major differences between the two models are in the costs of running
them and the algorithm chosen for convective transport. In general, we
use the results of one-dimensional calculations as initial conditions for the
two-dimensional calculations to reduce the cost of computations. The two-
dimensional flame model is discussed in detail in Ref. 41.
Input Parameters to Detailed Flame Models

As discussed in the previous section, there are several techniques cur-
rently available that can solve a one- or two-dimensional flame problem
adequately, provided accurate input data are provided. The various types
of input data required are 1) chemical kinetics pathways of reactions along
with chemical reaction rates for these reaction steps, 2) thermodynamic
input data such as enthalpies and heats of formation for all of the species
involved, and 3) diffusive transport coefficients such as molecular diffusion
and thermal conduction. Among these classes of data, the thermodynamic
data are the most easily and accurately available. Either the JANNAF
tables60 or the Gordon-McBride data collection61 are usually sufficiently
accurate to solve a large variety of flame problems. Usually these data can
be fit in a polynomial form to be used in computations. Except for hydro-
gen-air mixtures, the chemical kinetics data are not reliably available.
However, the extensive amount of work going on in this area will enable
one to have satisfactory data for methane and other higher hydrocarbon
mixtures. The diffusive transport coefficients generally are not available
to a satisfactory degree of accuracy and not enough effort has gone into
improving the situation because it is difficult and not necessarily rewarding
for the experimentalist. Two sources of information on diffusive transport
coefficients are Refs. 62 and 63. The specific data needed for modeling
hydrogen-oxygen-nitrogen flames is presented in detail in the report on
FLAME1D37 and to some extent also in Ref. 41.
IV. Numerical Studies of Laminar Flames

The purpose of developing detailed flame models is to be able to solve
accurately a variety of fundamental and practical problems. In this section,
a brief description of various applications of flame models is presented.
This summary is not all-inclusive, and the work done at the Naval Research
Laboratory has been emphasized. Over the last few years, several studies
have been made of the propagation and extinction of flames using FLAME1D.
More recently, the two-dimensional model, FLIC2D, has been used to
study the multidimensional structure of flames. Application of this model
to the study of the mechanisms leading to cellular structure and an ongoing
study of flame instabilities has been highlighted. First, let us look at the
determination of burning velocity and the steady-state structure of flames.
Steady-State Flames and the Determination of Burning Velocities

Flat flames are used extensively in combustion experiments to study the
detailed structure of flames and to investigate chemical kinetics processes.
Therefore, it is not surprising that the most common application of flame
models is to determine the steady-state structure and burning velocity of
one-dimensional flames. There are several approaches that provide a sat-
isfactory solution to the one-dimensional steady problem, and many of
these approaches are described in the proceedings of a workshop on laminar
flames.36 In many approaches, the time-dependent equations are solved
with appropriate boundary conditions until convergence to a steady-state
solution is obtained. Such calculations are robust and have provided sat-
isfactory results but are inefficient for computing a purely steady-state
problem. A particularly good approach for determining the characteristics
of steady flames is the technique developed by Smooke3435 and variations
of this technique.64-65 In this technique, the system of algebraic equations
obtained by discretizing the governing partial differential equations is solved
by using a damped, modified Newton iteration procedure. Furthermore,
a successive grid refinement scheme is incorporated into the solution pro-
cedure. After obtaining the initial solutions on a coarse grid, new grid
points are added in regions where the solution or its gradients change
rapidly. This procedure is continued until no new grid points are needed
to resolve the solution to the degree specified by the user. A key point for
the success of this approach is that the solution on one grid lies within the
domain of convergence of Newton's method on the next finer grid. Thus,
even though the cost per iteration is increasing, the number of required
iterations is decreasing.
Initially, this method was applied to the study of burner-stabilized
flames.35'64 An advantage of simulating burner-stabilized flames is that
there is extensive experimental data available to compare with the results
of the computations. However, these flames are nonadiabatic flames since
there is heat transfer to the burner and the results of the computations are
sensitive to the burner boundary conditions used. Smooke et al.64 circum-
vent this difficulty by using the experimentally measured temperature pro-
file to replace the energy equation. The results of such a computation of
a low-pressure (10.6 Torr) 3:l/hydrogen:oxygen flame are shown in Fig.
1. Excellent agreement is observed for the OH species distribution.
More recently, the approach adopted by Smooke has been generalized
to the calculation of adiabatic burning velocities.65 Here, an energy equa-
tion must be included to obtain the solution. A comparison of the computed
burning velocities of hydrogen-air flames to experimental data is shown in
Fig. 2. Again, the general agreement with experimental observations is
good.
There have also been a number of other calculations of the characteristics
of steady flames, such as the extensive studies of the pressure, temperature,
and dilution dependence of burning velocity by Warnatz. 29 There have also
been many studies in which comparisons of calculations of one-dimensional
steady flames to experimental data have been used as a test of kinetic
models (see, for example, Refs. 66 and 67). A novel use of steady-state
K. KAILASANATH 239
0.0 T.O 2.0 3.0 4.0 5.0 6.0 7.0 8.0 9.0 10.0
HEIGHT ABOVE THE BURNER (CM)
Fig. 1 Comparison of experimental and calculated OH profiles for an H2/O2 flame.
(Reproduced from Ref. 64 with permission.)
00 q
Q 9
LJ O
LJ O
Q_ CN
00
Ld q
^ 6
< £
0.0 20.0 40.0 60.0 80.0

PERCENT H2
Fig. 2 Comparison of computed atmospheric pressure hydrogen-air burning ve-
locities with data compilation of Dixon-Lewis. (Reproduced from Ref. 65 with per-
mission.)
flame calculations has been the validation of simplified multispecies trans-

port formulations by Coffee and Heimerl.31
Ignition and Quenching of Premixed Gases

There are many problems, such as the ignition of premixed gases, that
are intrinsically temporal, hence, an accurate time-dependent calculation
procedure is necessary. A series of calculations of the ignition and quench-
1800
1600 —
200
0 0.1 0.2 0.3 0.4 0.5
RADIUS (cm)
Fig. 3 Time history of temperature profile in an H2:O2:N2/2:1:10 mixture.
K. KAILASANATH 241
ing of premixed gases using the FLAME ID code is described subse-

quently.3738 For these calculations, the model was configured in spherical
geometry with an open boundary at one end to simulate a very large system.
The calculation was set up so that initially there were only 50 computational
cells, and the split-merge rezoning procedure that increased or decreased
the resolution was allowed to occur as needed. Then, a predetermined
amount of energy, 4 mJ, was deposited linearly over a period of 10~ 4 s in
a deposition radius of 0.1 cm. Figure 3 shows the spatial distribution of
temperature at a sequence of times. Even after the energy deposition is
stopped, the central temperature continues to rise due to heat released by
chemical reactions. With time, diffusive transport effects become impor-
tant, and the temperature near the center decreases and the temperature
away from the center increases. Finally, a balance is set up between dif-
fusive transport, energy release, and convection, and a steadily propagating
flame forms.
When the deposited energy was reduced to 3 mJ, the temperature dis-
tribution did not develop into a propagating flame. In this case, the gra-
dients in the temperature profile decrease with time and the profile flattens
out, indicating that 3 mJ is below the minimum ignition energy. By re-
peating the computation with varying amounts of energy deposition, a
minimum ignition energy for a particular radius of energy deposition can
be obtained. The calculated minimum ignition energy varies when the
radius (or volume) of energy deposition is changed. A summary of many
such calculations with different values of the radius of energy deposition
is shown in Fig. 4. For example, a propagating flame resulted when 3.8
mJ was deposited in a sphere with a radius of 0.1 cm. However, when the
same amount of energy was deposited in a sphere of smaller radius, the
rate of heat release was insufficient to compensate for the rate of diffusive
heat loss, and, consequently, there was no ignition. For a slightly larger
radius of energy deposition, the minimum ignition energy is constant; for
much larger radii (larger than 0.11 cm), the minimum ignition energy
increases rapidly with increasing radii. Therefore, this radius, 0.1 cm, is
the quench radius for this particular mixture, and the minimum ignition
energy of about 3.7 mJ is the absolute minimum ignition energy for the
idealized system under study.
These observations are qualitatively similar to the minimum ignition
energies and quenching distances usually measured in spark ignition ex-
periments such as those described in Ref. 68. There are many differences
between the idealized simulations of ignitions described earlier and the
phenomena observed in spark ignition experiments that make quantitative
comparisons difficult. There are heat losses to the electrodes and losses
due to radiation and shock waves in the spark ignition experiments, and
these effects have been neglected in the calculations. The time over which
energy is deposited is also usually shorter in the experiments. In the early
stages of energy deposition, ionization and recombination reactions that
have been ignored here are also expected to occur.69 More detailed cal-
culations, which include some of these effects along with a parallel set of
experiments, are needed for a quantitative study and the development of
a predictive capability of ignition and quenching phenomena.
3 —
_ H2 =
O IGNITION
TQ = 1 x 10"4s
X NO IGNITION
I I
0.08 0.09 0.10 0 . 1 1 0.12 0.13
R 0 (cm)
Fig. 4 Minimum ignition energy as a function of the radius of energy deposition.
Burning Velocity of Curved and Unsteady Flames

The calculations described in the previous section were concerned with
simulating the ignition of premixed gases. If these calculations were carried
out for a long enough time and if no extinction mechanisms were allowed
in the computations, the temperature and species distributions would attain
a profile typical of a steadily propagating flame. This approach provides
an alternate method (to the one described earlier in which the steady-state
equations are solved) for determining the laminar flame structure and
burning velocity. In contrast to the steady-state approach, this method will
also tell us if there is any dynamic variation in the structure and velocity
of the flame, such as that which would occur if any processes leading to
flame extinction are included in the model.
K. KAILASANATH 243
To determine the flame speed and burning velocity, first we need criteria
for determining the location of the flame front and the appropriate value
of the fluid velocity. The burning velocity vb then can be written as
vb = vf- vnmd (19)
where tyis the velocity at which the flame moves and v tluid the fluid velocity
ahead of the flame. Once a criterion is established for defining the flame
location, a time history of the flame locations gives the flame velocity.
Determining the appropriate value to use for the fluid velocity ahead of
the flame is sometimes complicated, as can be seen by comparing Figs. 5
and 6, which show the fluid velocities and the temperature profiles in a
planar and spherical flame propagating in a mixture with the same initial
temperature, pressure, and stoichiometry. For the planar flame in Fig. 5,
there is no ambiguity in choosing the fluid velocity, because the fluid
velocity ahead of the flame is constant. However, for the spherical flame
in Fig. 6, the fluid velocity reaches a maximum within the flame, and then
decreases ahead of the flame due to curvature. Therefore, for curved
flames, the fluid velocity used to calculate the burning velocity must also
be stated.
For many of our flame studies, we use two reference fluid velocities for
estimating the burning velocity. The first is the maximum fluid velocity in
the system, vmax, and the second is velocity of the unburned gases corre-
sponding to the first location ahead of the flame with a temperature of 300
K (or the initial temperature). The lower estimate for vh is obtained as
vf — vmax, and the upper estimate is from vf - vm). For a planar flame,
the two estimates are identical because the fluid velocity ahead of the flame
2000
0.4 0.6 0.8 1.0 1.2 1.4

POSITION (mm)
Fig. 5 Flow velocity and temperature profiles in a planar flame in a mixture of

hydrogen, oxygen, and nitrogen in the ratio 2:1:4.
3.0 i——.——•——.——•——.——•——,——•——.——,——,——,——,——, 2400
0.2 0.4 0.6 0.8 1.0 1.2 1.4

POSITION (mm)
Fig. 6 Flow velocity and temperature profiles in a spherical flame in a mixture of

hydrogen, oxygen, and nitrogen in the ratio 2:1:4.
is constant and the same as the maximum fluid velocity. If the flame is
unsteady, the instantaneous burning velocity can be determined by using
the local values of the flame and fluid velocities. Examples of unsteady
flames are discussed next.
Effects of Curvature and Dilution on Flame Propagation

If enough energy is deposited at the center of a large volume of com-
bustible premixed gas, a spherically propagating flame will develop, which
will eventually behave like a planar flame. In the early stages of flame
propagation, when the radius of the flame is not much larger than the
flame thickness, the curvature of the flame plays a significant role and the
burning velocity of the flame is different from the planar burning velocity.
The examples of such curved flames discussed here are taken from a study
of the effects of curvature and dilution on hydrogen-oxygen flames.39 In
that study, a series of calculations was performed of flames in spherical
geometry for which the amount of dilution with nitrogen was varied. In
Fig. 7, we compare the time evolution of the flame velocity and the two
fluid velocities, vmax and v300, for three mixtures H2:O2:N2—2:1:4, 2:1:7,
and 2:1:13.
Consider Fig. 7a for the 2:1:4 case. First, the flame velocity decreases
with time, reaching a minimum value, and finally it increases again. At
large enough radii, vmax and v300 approach the same value and produce an
unambiguous burning velocity equivalent to that found in an equivalent
one-dimensional planar calculation. Figure 7b, which is for a flame in a
K. KAILASANATH 245
120 160 200 240 280 320
200 400 600 800 1000 1200

c)
2.5
2.0
1.5
1.0
0.5 300
0.0
200 600 1000 1400
TIME
Fig. 7 Flame velocity vf, maximum fluid velocity i;vax, and velocity of the fluid at
300 K, i;300, as a function of time in spherically propagating flames in mixtures of
hydrogen, oxygen, and nitrogen in the ratios a) 2:1:4, b) 2:1:7, and c) 2:1:13.
2:1:7 mixture, shows the same trends as in the less dilute case, except that
it now takes longer to reach the minimum value of ty. For a 2:1:10 mixture,
the minimum value is attained still later around 1.5 ms and increases only
slightly from the minimum value.
In Fig. 7c, the mixture was so dilute that it did not support flame prop-
agation in spherical geometry. We were not able to achieve ignition in
spherical geometry by varying the energy input, the radius of energy dep-
osition, or the mode of energy input. The flame always died. However,
we did achieve a steady, propagating flame when the analogous calculation
was done in planar or cylindrical geometry.
These time histories of the flame velocity can be explained by considering
stretch effects due to diverging flow and energy release resulting from
chemical reactions. First, consider the initial deceleration of tyin spherical
geometry. Initially, because the flow diverges, the energy release does not
balance the energy conducted and diffused into the unreacted mixture.
Because of this, the flame velocity and temperature decrease as the flame
expands. The decrease in the flame temperature decreases the energy-
release rate, which leads to a lower flame velocity, and so on. However,
this process does not continue indefinitely because the stretch effect de-
creases with increasing radius. This results in the minimum flame velocity
seen in the calculations. As the radius of the flame increases further, the
energy released is larger than the energy conducted and diffused into the
unburned mixture. Now the flame velocity increases with increasing radius
until a balance is achieved between the energy released locally and the
energy conducted and diffused into the unburned mixture ahead of the
flame. This balance occurs at large radius, where the flame propagates as
if it were planar.
Flammability Limits
The preceding observations on the effects of curvature and dilution also
lead to the conclusion that a flame can be extinguished with less dilution
in one geometry than in another. The effects of geometry on burning
velocity of dilute mixtures are summarized in Table 1. Specifically, a 2:1:137
hydrogen-oxygen-nitrogen mixture does not support a self-sustained spher-
ical flame, although it does support cylindrical and planar flames. This
suggests that the dilution limits of flammability are different in different
geometrical configurations, and curvature might play an important role in
the propagation of near-limit flames. Indeed, flames observed in experi-
ments in such mixtures are unstable and multidimensional, and, hence, a
time-dependent multidimensional model is needed to calculate their de-
tailed behavior. Such calculations will be described later in this chapter.
Flames in fuel-rich hydrogen-air mixtures appear to be stable and nearly
one-dimensional. Therefore, flammability limits can be explored further
by simulating planar flames in mixtures in which the amount of hydrogen
is increased systematically.40 Burning velocities are calculated according to
Eq. (19). Table 2 summarizes the burning velocity and the temperature of
the burned gas, the flame temperature, for the various mixtures. As seen
in the table, the burning velocity and the flame temperature decrease with
K. KAILASANATH 247
Table 1 Effect of geometry on burning velocity
Minimum burning velocity, cm/s

Mixture
H2:02:N2 Planar Cylindrical Spherical
2:1:10 42 30-35
2:1:11 28 6-8
2:1:13 20 15-17 —
2:1:15 13 9-11 —
increasing fuel concentration. There are a number of observations that can

be made based on the data in Table 2.
First, for an 80.8% hydrogen-air mixture (H2:O2:N2/20:1:3.76), there is
a steady burning velocity of about 8 cm/s. This mixture is beyond the
experimentally observed rich-flammability limit under normal gravity con-
ditions. This difference between the numerical and experimental obser-
vations may be due to the effects of gravity, heat losses to the confining
walls in the experiments, or to multidimensional effects, none of which are
included in the calculations.
Second, we note that an 82.2% hydrogen-air mixture (H2:O2:N2/22:1:3.76)
does not support a steadily propagating flame. We tried different ignition
sources; however, in all cases, the burning velocity decreased to zero. The
behavior of the transient flame was qualitatively similar to that seen pre-
viously for a dying flame. The simulations indicate that there is a flam-
mability limit, even in the absence of stretch, buoyancy, and external heat
losses. Furthermore, this limit is outside that observed in the standard
flammability limit tube under normal gravity conditions.
A possible reason for this flammability limit can be seen in the trend of
flame temperatures shown in Table 2. The flame temperature for a 20:1:3.76
mixture is only 970 K. As the fuel concentration increases, the temperature
would decrease even further. At these temperatures, it is possible that the
H atoms are consumed faster in chain terminating reactions than produced
in chain branching reactions. This would result in a depletion of H atoms
and a decrease in temperature. The decrease in temperature eventually
could lead to a situation where the energy released in exothermic reactions
is not sufficient to balance the energy diffused by the transport processes,
Table 2 Near-limit fuel-rich mixtures
Mixture Burning Flame

H2:O2:N2 velocity, cm/s temperature
16:1:3.76 38 1180
17:1:3.76 30 1108
18:1:3.76 23 1046
19:1:3.76 16 1002
20:1:3.76 8 970
22:1:3.76 — <970
and, hence, a steady flame is not possible. The implication is that there
may be a flammability limit due to chemical kinetic considerations alone.
In a real system, even minimal heat losses would ensure that an incipient
flame is extinguished in such a fuel-rich mixture. Therefore, the exact limit
composition and temperature would vary from one experimental setup to
another. Furthermore, buoyancy and multidimensional effects may also be
important. This is an area of flame dynamics that needs further research.
Flame Instabilities and the Cellular Structure of Flames

Many one-dimensional flames are unstable in the sense that a small
perturbation of such flames would lead to a multidimensional flame. An
example of such a flame is a planar flame in a 1.5:l:10/hydrogen-oxygen-
nitrogen mixture. The two-dimensional flame model, FLIC2D, has been
used to simulate this transition from a planar to a multidimensional struc-
ture.51 The simulations were started by first calculating the one-dimensional
structure of a planar flame in a 1.5:l:10/hydrogen-oxygen-nitrogen mixture
and using this as the initial conditions in the two-dimensional calculations.
The results of perturbing the planar flame are shown in Fig. 8 in terms of
temperature and OH concentrations at a sequence of times. In all of the
frames, the unburnt mixture flows in from the left and the burnt mixture
flows out to the right. The two-dimensional computational domain was 2.0
x 4.5 cm, which was resolved by a 56 x 96 variably spaced grid. Fine
zones, 0.36 x 0.15 mm, were clustered around the flame front. The de-
velopment of a multidimensional flame that has convex (to the inflow)
regions of more intense reaction and concave regions of less intense re-
action is apparent from Fig. 8. By 30 ms, the overall features of the flame
are typical of cellular flames, which are characterized by brighter and
darker regions along the flame front and a convex shape toward the unburnt
mixture. At later times, we see the convex flame splitting into two cells,
as expected from experimental observations. The structure of a cellular
flame can be seen more readily in the color plots shown in plate 10 (see
the color section), where the temperature, OH, H, and H2 concentrations
are shown at a typical time. The regions of more and less intense reactions
alluded to earlier are seen in the OH and H concentrations plot.
A similar calculation was also performed for a fuel-rich mixture of hy-
drogen-oxygen-nitrogen in the ratio, 3:1:16. Flames in this mixture are
stable in the experiments of Ref. 70, and also in the calculations. In this
case, an initial perturbation is damped quickly, restoring the flame profile
to its one-dimensional planar shape. Since these calculations demonstrated
qualitative agreement with experimental observations, the simulations were
then used as a "computational-experimental" tool to investigate the mech-
anism that may be responsible for the flame instability.
Hydrodynamic, buoyancy-induced Rayleigh-Taylor, thermodiffusive, and
preferential diffusion effects are all possible mechanisms that can lead to
multidimensional flames. Hydrodynamic instabilities are due to the fluid
expansion across a flame and can be expected to occur in all flames. How-
ever, the stabilizing mechanisms inherent in flames generally damp the
small wavelengths and usually a system more than an order of magnitude
larger than the flame thickness (larger than the one used in the simulations)
K. KAILASANATH 249
TEMPERATURE H 2 :O 2 :N 2 / 1.5:1:10 Vb° = 12 cm/s
0.5 ms 15 m 30 ms 40 ms 60 ms
Fig. 8 Transition from planar flame to cellular flame in an H2:O2:N2/1.5:1:10

mixture.
is needed to observe hydrodynamic instabilities. Rayleigh-Taylor instability

occurs when a heavier fluid is accelerated into a lighter fluid. On Earth,
this acceleration is provided by gravitational attraction. For the purpose
of isolating the mechanism leading to cellular structure, gravity effects are
not included in the calculations discussed in this section of the chapter.
That leaves us with two major possible mechanisms that can lead to cellular
structure. One mechanism is the preferential diffusion of one reactant over
another due to its higher diffusivity. For example, in a hydrogen-oxygen-
nitrogen mixture, hydrogen diffuses much faster than oxygen or nitrogen.
Another mechanism is the differences in the rate at which mass and heat
diffuse (diffusional-thermal theory). A systematic series of calculations has
been performed to determine if the preferential diffusion or the thermo-
diffusive instability mechanism is the dominant mechanism to cause cellular
flames.
In one simulation, the mass diffusivity of hydrogen was set equal to
oxygen. This eliminates the preferential diffusion effect, and means that
the light fuel will not move fast compared with the other reactant. If the
mechanism of instability were preferential diffusion, the lean flame would
be stable. This is, in fact, the result: with this change of relative diffusivities,
the lean flame was stable in the calculation. This result supports both the
preferential diffusion theory and the diffusional-thermal theory. Therefore,
another simulation was performed in which the mass diffusivity of oxygen
was set equal to that of hydrogen. In this case, the flame was unstable,
which disagrees with the prediction of the preferential diffusion theory.
The diffusional-thermal theory does correctly predict that this flame is
unstable. Thus, for the hydrogen flames studied, the results agree with the
predictions of the diffusional-thermal theory and not the preferential dif-
fusion theory. These simulations provide an example of the use of nu-
merical simulations in identifying the controlling instability mechanism
when more than one mechanism is plausible.
Effects of Gravity on Multidimensional Flame Structure

There are many reasons for studying the effects of gravity on flame
structure. On Earth, the physical mechanisms causing the thermodiffusive
instability may interact with buoyancy forces, and the resulting flame struc-
ture can be quite different. It is also well known from flammability studies
in normal (Earth) gravity that a flame propagating upward in a tube prop-
agates in a wider range of mixture stoichiometries and dilutions than a
flame propagating downward. This suggests that the flammability limits
depend on gravity.
We have studied the effects of gravity on flame structure by comparing
simulations of zero-gravity flames to upward- and downward-propagating
flames.52-53 The two-dimensional computational domain and the initial con-
ditions for the calculations are the same as described earlier in the discus-
sion of cellular flame structure. As before, the planar flame is perturbed,
and the subsequent evolution of the flame front in zero gravity is studied.
Similar calculations are also performed with the buoyancy forces in the
direction of flame propagation and against the direction of flame propa-
gation. In Fig. 9, the multidimensional flame is depicted at a series of times
by using OH concentration contours. The evolution of upward-, zero-
gravity, and downward-propagating flames is compared in the figure. The
differences among the three cases are dramatic. The zero-gravity case
shows a cellular structure. At 20 ms, the shape and size of the three flames
are comparable. The zero-gravity flame exhibits a stable cellular structure
with little change from 20 to 100 ms. The upward-propagating flame be-
comes more and more curved, and the central portion of the flame moves
more rapidly than the sides. In this case, the buoyancy-induced and ther-
modiffusive mechanisms are both destabilizing. At 100 ms, the upward-
propagating flame exhibits a bubblelike appearance characteristic of near-
limit upward-propagating flames. 71 In the downward-propagating case,
buoyancy forces tend to stabilize and return a perturbed flame back to its
initial planar configuration. However, at later times, the flame front begins
to oscillate around its initial planar configuration. The fact that early in
the calculations, at 20 ms, all three flames look alike suggests that the
instability process leading to the cellular structure is growing more rapidly
than the process leading to a buoyancy-induced structure. By 100 ms, the
buoyancy-induced instability process has had enough time to interact strongly
with the thermodiffusive instability process and, in the case of the down-
ward-propagating flame, essentially nullify the thermodiffusive effects.
A similar comparison as in Fig. 9 has been made for flames propagating
in a H2:O2:N2/1.5:1:10 mixture. In that case, the difference between the
three cases are small. From such studies, it can be concluded that the effect
of gravity on flame propagation depends strongly on the stoichiometry of
the mixture. Further studies seem to indicate that, in general, the effects
of gravity are significant in mixtures with low burning velocities, which are
typical of flames in mixtures near flammability limits.
V. Current Status and Future Needs

The simulations discussed earlier are examples of what currently can be
computed with detailed flame models. They also show how numerical
K. KAILASANATH 251
100 nns
80ms
60ms
40ms
20ms
1 ms
Upward Downward
Zero Gravity Propagating
Propagating
Fig. 9 Comparison of the evolution of upward, downward, and zero-gravity flames

in an H2:O2:N2/1:1:10 mixture.
simulations can be used to gain a better insight into problems in combus-

tion. For example, the simulations have been used to identify the dominant
mechanism leading to the cellular structure of flames. Other simulations
also indicate that the instability leading to cellular structure grows more
rapidly than the buoyancy-induced Rayleigh-Taylor instability.
The numerical models and input data necessary for time-dependent, two-
dimensional adiabatic flames in a hydrogen-oxygen-nitrogen mixture are
currently available. Further work is necessary to develop appropriate models
to account for external heat losses due to radiation and heat conduction.
The development of a radiation transport model with the same amount of
detail as the diffusive transport model is a real challenge.
In principle, the dynamics of two-dimensional flames in hydrocarbon
mixtures can also be simulated. However, the large number of species and
chemical reactions involved makes such calculations prohibitively expen-
sive. Simpler models, especially for the kinetics, will help to alleviate this
difficulty. Some of the input parameters, such as the diffusive transport

coefficients, are also not available to the same level of accuracy as for
hydrogen flames.
Flames are, in general, unsteady and three-dimensional. Therefore, a
major effort is required for the development of cost-effective three-di-
mensional models. A major development that has made possible the sim-
ulation of unsteady two-dimensional flames was the development of the
implicit flux-corrected transport algorithm.42 As discussed earlier, this al-
gorithm allows the use of time steps determined by the slowly evolving
physics of the problem rather than one determined by the speed of sound.
The extension of this algorithm to three dimensions is straightforward and
currently is under way. Still a major cost of such flame simulations is in
the integration of the detailed diffusive transport and the chemical reactions
between the various species involved. Currently, a detailed two-dimen-
sional simulation of a cellular flame in a lean (10%) hydrogen-air mixture
in a 51-mm channel takes about 20 h on a Cray X-MP 2/4 for simulating
0.25 s of physical time. Out of this time, about 50% is spent in diffusive
transport and about 40% in chemistry and the cost is due primarily to the
large number of species involved. The development of simplified multistep
chemical reaction mechanisms and simplified diffusive transport models,
which involve fewer species and yet adequately describe the structure and
dynamics of flames, is the major impediment in the development and use
of unsteady, three-dimensional flames on a routine basis.
Acknowledgments
The authors would like to acknowledge numerous discussions and re-
search collaborations with Gopal Patnaik, Elaine Oran, and Jay Boris that
have resulted in many of the works discussed in this chapter. Over a number
of years, the Naval Research Laboratory work discussed in this chapter
has been sponsored by NASA in the Microgravity Science and Applications
Program and by the Naval Research Laboratory.
References
Mallard, E., and Le Chatelier, H., "Recherches Experimentales et Theoriques
sur la Combustion des Melanges Gazeux Explosifs: Memoir II, Sur la Vitesse de
Propagation de la Flamme," Annales des Mines, Vol. 4, 1883, pp. 296-395.
2
1985.
3
Clavin, P., "Dynamic Behavior of Premixed Flame Fronts in Laminar and
Turbulent Flows," Progress in Energy Combustion Science, Vol. 11, No. 1, 1985.
4
Hirschfelder, J. O., Curtiss, C. F., Henkel, M. J., Spaulding, W. P., and Hum-
mel, H., "Theory of Propagation of Flames, Parts I, II and III," Third Symposium
on Combustion and Flame and Explosion Phenomena, Williams and Wilkins, Bal-
timore, MD, 1949, pp. 121-139.
5
Hirschf elder, J. O., Curtiss, C. F., and Campbell, D. E., "The Theory of Flames
and Detonations," Fourth Symposium (International) on Combustion, Williams
and Wilkins, Baltimore, MD, 1953, pp. 190-211.
K. KAILASANATH 253
6
Gidding, J. C., and Hirschfelder, J. O., "Flame Properties and the Kinetics of
Chain-Branching Reactions," Sixth Symposium (International) on Combustion,
Reinhold PubL, New York, 1957, pp. 199-213.
7
Spalding, D. B., "The Theory of Flame Phenomena with a Chain Reaction,"
Philosophical Transactions of the Royal Society of London, London, UK, 249A,
1956, pp. 1-25.
8
Adams, G. K., and Cook, G. B., 'The Effect of Pressure on the Mechanism
and Speed of the Hydrazine Decomposition Flame," Combustion and Flame, Vol.
4, 1960, pp. 9-18.
9
Dixon-Lewis, G., "Flame Structure and Flame Reaction Kinetics, I. Solution
of Conservation Equations and Application to Rich Hydrogen-Oxygen Flames,"
Proceedings of the Royal Society of London, London, UK, A298, 1967, pp. 495-
513.
10
Dixon-Lewis, G., "Flame Structure and Flame Reaction Kinetics, II. Transport
Phenomena in Multicomponent Systems," Proceedings of the Royal Society of
London, London, UK, A307, 1968, pp. 111-135.
H
Dixon-Lewis, G., and Shepherd, I. G., "Some Aspects on Ignition by Localized
Sources, and of Cylindrical and Spherical Flames," Fifteenth Symposium (Inter-
1483-1491.
12
Spalding, D. B., Stephenson, P. L., and Taylor, R. G., "A Calculation Pro-
cedure for the Prediction of Laminar Flame Speeds," Combustion and Flame, Vol.
17, 1971, pp. 55-64.
13
Patankar, S. V., and Spalding, D. B., Heat and Mass Transfer in Boundary
Layers, 2nd ed., Intertext Books, London, UK, 1970.
14
Spalding, D. B., and Stephenson, P. L., "Laminar Flame Propagation in Hydro-
gen -h Bromine Mixtures," Proceedings of the Royal Society of London, London,
UK, A324, 1971, pp. 315-337.
15
Tsatsaronis, G., "Prediction of Propagating Laminar Flames in Methane, Oxy-
gen, Nitrogen Mixtures," Combustion and Flame, Vol. 33, 1978, pp. 217-239.
16
Smoot, L. D., Hecker, W. C., and Williams, G. A., "Prediction of Propagating
17
Bledjian, L., "Computation of Time-Dependent Laminar Flame Structure,"
Combustion and Flame, Vol. 20, 1973, pp. 5-17.
18
Margolis, S. B., "Time-Dependent Solution of a Premixed Laminar Flame,"
19
Westbrook, C. K., and Dryer, F. L., "A Comprehensive Mechanism for Meth-
anol Oxidation," Combustion Science and Technology, Vol. 20, 1979, pp. 125-
140.
20
Lund, C. M., "A General Computer Program for Calculating Time-Dependent
Phenomena Involving One-Dimensional Hydrodynamics, Transport and Detailed
Chemical Kinetics," University of California Lawrence Livermore Laboratory Re-
port, UCRL-52504, 1978.
21
Dixon-Lewis, G., Goldsworthy, F. A., and Greenberg, J. B., "Flame Structure
and Flame Reaction Kinetics, IX. Calculation of Properties of Multi-Radical Pre-
mixed Flames," Proceedings of the Royal Society of London, London, UK, A346,
1975, pp. 261-278.
22
Dixon-Lewis, G., "Kinetic Mechanism, Structure and Properties of Premixed
Flames in Hydrogen-Oxygen-Nitrogen Mixtures," Philosophical Transactions of
the Royal Society of London, London, UK, 292, 1979, pp. 45-99.
23
Dixon-Lewis, G., "Mechanism of Inhibition of Hydrogen-Air Flames by Hy-
drogen Bromide and Its Relevance to the General Problem of Flame Inhibition,"
24
Dixon-Lewis, G., "Spherically Symmetric Flame Propagation in Hydrogen-Air
Mixtures," Combustion Science and Technology, Vol. 34, 1983, pp. 1-29.
25
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames I. Flame
Velocity of Freely Propagating Ozone Decomposition Flames," Berichte der Bun-
sengesellschaft fuer Physikalische Chemie, Vol. 82, 1978, pp. 193-200.
26
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames II. Flame
Velocity and Structure of Freely Propagating Hydrogen-Oxygen and Hydrogen-
Air Flames," Berichte der Bunsengesellschaft fuer Physikalische Chemie, Vol. 82,
1978, pp. 643-649.
27
Warnatz, J., "Calculation of the Structure of Laminar Flat Flames III. Structure
of Burner-Stabilized Hydrogen-Oxygen and Hydrogen-Fluorine Flames," Berichte
der Bunsengesellschaft fuer Physikalische Chemie, Vol. 82, 1978, pp. 834-841.
28
Warnatz, J., "The Structure of Freely Propagating and Burner-Stabilized Flames
in the H^-CO-O2 System," Berichte der Bunsengesellschaft fuer Physikalische Chemie,
Vol. 83,^1979, pp. 950-957.
29
Warnatz, J., "Concentration-, Pressure-, and Temperature-Dependence of the
Flame Velocity in Hydrogen-Oxygen-Nitrogen Mixtures," Combustion Science and
Technology, Vol. 26, 1981, pp. 203-213.
3<)
Coffee, T. P., and Heimerl, J. M., "A Method for Computing the Flame Speed
for a Laminar, Premixed, One-Dimensional Flame," Ballistics Research Labs.,
Aberdeen, MD, ARBRL-TR-02212, 1980.
31
Coffee, T. P., and Heimerl, J. M., "Transport Algorithms for Premixed Lam-
inar Steady-State Flames," Combustion and Flame, Vol. 43, 1981, pp. 273-289.
32
Dixon-Lewis, G., "Flame Structure and Flame Reaction Kinetics, V. Inves-
tigation of Reaction Mechanisms in a Rich Hydrogen + Nitrogen + Oxygen Flame
by Solution of Conservation Equations," Proceedings of the Royal Society of Lon-
don, London, UK, A317, 1970, pp. 235-263.
33
Wilde, K. A., "Boundary-Value Solutions of the One-Dimensional Laminar
Flame Propagation Equations," Combustion and Flame, Vol. 18, 1972, pp. 43-52.
34
Smooke, M. D., "Solution of Burner-Stabilized Pre-Mixed Laminar Flames by
Boundary Value Methods," Sandia National Labs., Livermore, CA, SAND SI-
8040, 1982.
35
Smooke, M. D., "Solution of Burner-Stabilized Pre-Mixed Laminar Flames by
Boundary Value Methods," Journal of Computational Physics, Vol. 48, No. 72,
1982.
36
Peters, N., and Warnatz, J., (eds.), Numerical Methods in Laminar Flame
Propagation, Vieweg & Sohn, Wiesbaden, Germany, 1982.
37
Kailasanath, K., Oran, E. S., and Boris, J. P., "A One-Dimensional Time-
Dependent Model for Flame Initiation, Propagation and Quenching/' Naval Re-
search Lab., Washington, DC, Memo. Rept. 4910, 1982.
38
Kailasanath, K., Oran, E. S., and Boris, J. P., "A Theoretical Study of the
Ignition of Premixed Gases," Combustion and Flame, Vol. 47, No. 173, 1982.
39
Kailasanath, K., and Oran, E. S., "Effect of Curvature and Dilution on Un-
steady, Premixed, Laminar Flame Propagation," Dynamics of Reactive Systems,
Part I: Flames and Configurations, edited by J. R. Bowen, J.-C. Leyer, and R. I.
Soloukhin, Vol. 105, Progress in Astronautics and Aeronautics, AIAA, Washing-
ton, DC, 1986, pp. 167-179.
40
Kailasanath, K., and Oran, E. S., "Time-Dependent Simulations of Laminar
Flames in Hydrogen-Air Mixtures," Complex Chemical Reaction Systems, edited
by J. Warnatz and W. Jager, Springer-Verlag, Heidelberg, Germany, 1987, pp.
243-252.
41
Patnaik, G., Laskey, K. J., Kailasanath, K., Oran, E. S., and Brun, T. A.,
"FLIC-A Detailed Two-Dimensional Flame Model," Naval Research Lab., Wash-
ington, DC, Memo. Rept. 6555, 1989.
K. KAILASANATH 255
42
Patnaik, G., Guirguis, R. H., Boris, J. P., and Oran, E. S., "A Barely Implicit
Correction for Flux-Corrected Transort," Journal of Computational Physics, Vol.
71., No. 1, 1987.
43
Boris, J. P., and Book, D. L., "Solution of Convective Equations by the Method
of Flux-Corrected Transport," Methods of Computational Physics, Vol. 16, No.
85, 1976.
44
Boris, J. P., "Flux-Corrected Transport Modules for Solving Generalized Con-
tinuity Equations," Naval Research Lab., Washington, DC, Memo. Rept. 3237,
1976.
45
Oran, E. S., and Boris, J. P., Numerical Simulation of Reactive Flow, Elsevier,
New York, 1987.
46
Goyal, G., Paul, P. J., Mukunda, H. S., and Deshpande, S. M., "Time De-
pendent Operator-Split and Unsplit Schemes for One Dimensional Premixed Flames,"
47
Carter, N. R., Cherian, M. A., and Dixon-Lewis, G., "Flames near Rich
Flammability Limits, with Particular Reference to the Hydrogen-Air and Similar
Systems, Numerical Methods in Laminar Flame Propagation, edited by N. Peters
and J. Warnatz, Vieweg & Sohn, Wiesbaden, Germany, 1982, pp. 182-191.
48
Lohner, R., Morgan, K., Vahdati, M., Boris, J. P., and Book, D. L., "FEM-
FCT: Combining Unstructured Grids with High Resolution," Communications in
Applied Numerical Methods, Vol. 4, 1988, pp. 717-729.
49
Habbal, A., Dervieux, A., Guillard, H., and Larrouturou, B., "Explicit Cal-
culation of Reactive Flows with an Upwind Finite Element Hydrodynamical Code,"
INRIA-Sophia Antipolis, France, INRIA Rept. 690, 1987.
50
Kailasanath, K., Oran, E. S., Boris, J. P., and Young, T. R., "Time-Dependent
Simulation of Flames in Hydrogen-Oxygen-Nitrogen Mixtures," Numerical Meth-
ods in Laminar Flame Propagation, edited by N. Peters and J. Warnatz, Vieweg
& Sohn, Wiesbaden, Germany, 1982, pp. 152-166.
M
Patnaik, G., Kailasanath, K.,Laskey, K. J., and Oran, E. S., "Detailed Numer-
ical Simulations of Cellular Flames," Twenty-Second Symposium (International)
52
Kailasanath, K., Patnaik, G. P., and Oran, E. S., "Effect of Gravity on Multi-
Dimensional Laminar Premixed Flame Structure," 39th IAF Congress, Interna-
tional Astronautical Federation, Paris, France, IAF Paper 88-354, 1988.
S3
Patnaik, G. P., Kailasanath, K., and Oran, E. S., "Effect of Gravity on Flame
Instabilities in Premixed Gases," AIAA Paper 89-0502, 1989.
54
Oran, E. S., and Boris, J. P., "Detailed Modeling of Combustion System,"
Progress in Energy and Combustion Science, Vol. 7, 1981, pp. 1-71.
55
Boris, J. P., "ADINC: An Implicit Lagrangian Hydrodynamics Code," Naval
Research Lab., Washington, DC, Memo. Rept. 4022, 1979.
56
Young, T. R., and Boris, J. P., "A Numerical Technique for Solving Stiff
Ordinary Differential Equations Associated with the Chemical Kinetics of Reactive
Flows," Journal of Physical Chemistry, Vol. 81, 1977, p. 2424.
57
Young, T. R., "CHEMEQ—-Subroutine for Solving Stiff Ordinary Differential
Equations," Naval Research Lab., Washington, DC, Memo. Rept. 4091, 1980.
58
Jones, W. W., and Boris, J. P., "An Algorithm for Multispecies Diffusion
Fluxes," Computers & Chemistry, Vol. 5, 1981, pp. 139-146.
59
Kee, R. J., Dixon-Lewis, G., Warnatz, J., Coltrin, M. E., and Miller, J. A.,
"A Fortran Computer Code Package for the Evaluation of Gas-Phase Multi-
Component Transport Properties," Sandia National Lab., Livermore, CA, SAND86-
8246, 1986.
60
Stull, D. R., and Prophet, H., JANNAF Thermochemical Tables, 2nd ed.,
National Standard Reference Data Series, U. S. National Bureau of Standards,
No. 37, Government Printing Office, Washington, DC, 1971.
61
Gordon, S., and McBride, B. J., "Computer Program for Calculation of Com-
plex Chemical Equilibrium Compositions, Rocket Performance, Incident and Re-
flected Shocks, and Chapman-Jouguet Detonations," NASA SP-273, 1976.
62
Svehla, R. A., "Estimated Viscosities and Thermal Conductivities of Gases at
High Temperatures," NASA TR R-132, 1962.
63
Mason, E. A., and Marrero, T. R., "Gaseous Diffusion Coefficients, "Journal
of Physical Chemistry Reference Data, Vol. 1, No. 3, 1972.
64
Smooke, M. D., Miller, J. A., and Kee, R. J., "Numerical Solution of Burner-
Stabilized Premixed Laminar Flames by an Efficient Boundary Value Method,"
Numerical Methods in Laminar Flame Propagation, edited by N. Peters and
J. Warnatz, Vieweg & Sohn, Wiesbaden, Germany, 1982.
65
Smooke, M. D., Miller J. A., and Kee, R. J., "Determination of Adiabatic
Flame Speeds by Boundary Value Methods," Combustion Science and Technology,
Vol. 34, 1983, pp. 79-90.
66
Westbrook, C. K., and Dryer, F. L., "Simplified Reaction Mechanisms for the
Oxidation of Hydrocarbon Fuels in Flames," Combustion Science and Technology,
Vol. 27, 1981, pp. 31-43.
67
in Premixed Laminar, Steady-State Flames. I. Stoichiometries," Combustion and
Flame, Vol. 54, 1983, pp. 155-169.
68
Lewis, B., and von Elbe, G., Combustion, Flames, and Explosives, Academic,
New York, 1961.
69
Maly, R., "Ignition Model for Spark Discharges and the Early Phase of Flame
Front Growth," Eighteenth Symposium (International) on Combustion, The Com-
bustion Institute, Pittsburgh, PA, 1981, pp. 1747-1754.
70
Mitani, T., and Williams, F. A., "Studies of Cellular Flames in Hydrogen-
Oxygen-Nitrogen Mixtures," Combustion and Flame, Vol. 39, 1980, pp. 169-190.
71
Levy, A., "An Optical Study of Flammability Limits," Proceedings of the Royal
Society of London, London, UK, A283, 1965, p. 134.
Chapter 9
Spectral Simulations of Reacting Turbulent Flows
Patrick A. McMurtry and Peyman Givi
I. Introduction
T HE Navier-Stokes equations, along with appropriate conservation

equations of energy and chemical species, are generally accepted to
provide an "exact" model for most turbulent combustion phenomena of
interest.1 Unfortunately, the complexity of these equations prohibits both
their analytic and numerical solutions except under idealized conditions.
To provide engineering tools to predict the performance of turbulent com-
bustion systems, it is necessary to resort to approximate analyses in which
the effects of turbulence are not treated exactly but are incorporated using
turbulence models.
Over the past several decades, a variety of approaches to modeling
turbulence have been promoted. The most widespread turbulence mod-
eling techniques used in engineering applications for both reacting and
nonreacting flows are various implementations of the "moment method."
This method involves introducing an appropriately defined average (mean)
for the dependent variables, and decomposing these variables into a mean
and fluctuating component. The averaged transport equations are then
solved to give the mean values of the physical variables. The well-known
difficulty with this approach is that, as a result of averaging, additional
terms not related directly to the averaged variables are introduced. These
terms take the form of correlations between the various fluctuating vari-
ables and must be modeled to obtain a closed set of equations. The famous
k-e model2 is the most common closure used in moment methods.
One of the main advantages of the moment method is its relative ease
of use in computational applications. Most models suggested in the liter-
ature can be incorporated directly into computer codes developed for lam-
inar flow by replacing the laminar viscosity with an effective turbulence
viscosity. This allows the engineer to obtain approximations to combustion
problems for a variety of flow conditions. The results obtained in this way
reserved.
257
258 SPECTRAL SIMULATIONS OF REACTING TURBULENT FLOWS
have, in many cases, exhibited very good agreement with experimental

measurements. This has been a major factor in justifying the use of moment
methods for turbulent combustion simulations. It appears that these meth-
ods will remain popular for engineering applications for the near future.
Unfortunately, none of the models developed have exhibited a robust-
ness that allows their application for general use. The models must often
be modified and tuned for different flow configurations. Although a tre-
mendous amount of effort and physical insight have gone into model de-
velopment, most turbulence models reduce to a correlation with experimental
data as a result of the numerical "constants" that must be specified. There-
fore, a model optimized for a particular flow generally is not adequate for
predicting the flow behavior in other configurations. Consequently, by
adopting a particular closure, much of the interesting dynamical behavior
of the flow cannot be predicted accurately since it has, in effect, been
modeled a priori for other operating conditions.
With the rapid progress in supercomputer technology over the past 15
years, a new methodology has emerged for the study and description of
turbulent flows. Termed direct numerical simulation (DNS), this approach
involves solving the exact governing equations for the detailed time de-
velopment of the flowfield. No closure modeling is used and no assumptions
are made concerning the turbulent behavior of the flow. By resolving all
relevant length and time scales, this approach allows one to simulate and
study the fundamental physics of turbulent flows. Unfortunately, finite
computer resources place rather severe limitations on the range of space
and time scales that can be resolved. This limitation restricts the use of
DNS to flows with only moderate variations in spatial and temporal scales.
Parallel developments in numerical methods have helped to promote the
use of DNS in turbulence research. Foremost among these methods is the
development of spectral methods and their application to the numerical
solution of fluid flow problems. The spectral method preserves more of
the important physics than do other commonly used numerical schemes.
Particular benefits of spectral methods include small phase errors (or first
differencing errors), a lack of numerical diffusion, and truncation errors
that decrease faster than algebraically as the resolution is increased. With
the development of fast Fourier transforms to perform the mapping be-
tween the physical and spectral domains,3 spectral methods have become
a viable and valuable tool for the numerical integration of transport equa-
tions describing fluid motion.
More recently, spectral methods have been applied sucessfully to the
direct simulation of chemically reacting turbulent flows. Because of the
physical complexities of general combustion problems, the flows studied
using DNS and spectral techniques necessarily have been limited to those
involving only elementary reaction mechanisms in simple geometries. These
studies have been concerned primarily with describing basic physics of
turbulent mixing and interactions between the fluid dynamics and chem-
istry. Statistical sampling of the simulation results has also provided a
quantitative basis for evaluating the validity and performance of theories
and models used in turbulent combustion. Direct numerical simulations
are now performed regularly for simplified flow and reactor configurations.
P. A. MCMURTRY AND P. GIVI 259
In this chapter, we review the results of some of these works and discuss
what they have revealed about the physics of turbulent mixing and reaction.
The numerous references made to DNS throughout this paper should
not lead one to believe that DNS and spectral methods are synonymous.
Spectral methods are simply a technique that has been used for discretizing
the equations of motion. Other methods, such as finite differences or finite
elements, can and have been used successfully. In this chapter, only the
accomplishments of spectral simulations are discussed.
In the next section, an overview of spectral methods is given, together
with a discussion of their implementation in the numerical solution of fluid
transport. Various classifications of spectral methods and their convergence
properties are described. The "spectral element" method, which consti-
tutes a new category of spectral approximations, is also presented in this
section, and its flexibility in dealing with complex flow geometries is high-
lighted. The discussions presented in this section are necessarily brief and
are only to provide a flavor of the approximations achieved by spectral
methods. For a more comprehensive discussion of the method, the inter-
ested reader is referred to Refs. 4-6. A review of the recent applications
of spectral methods to reacting flow problems is given in Sec. III. In this
section, emphasis is placed primarily on the interpretation of the physical
phenomena captured by spectral simulations of reactive flows. There is
some mention of spectral simulations of the turbulent mixing of a passive
scalar, but only where an explicit connection to chemical reactions has
been made. We conclude this chapter in Sec. IV with an evaluation of the
benefits and limitations of spectral methods, and some speculation on their
contributions for turbulent combustion research in the future.
II. Spectral Method

As mentioned in the Introduction, spectral methods have made signif-
icant contributions to direct simulations of both reacting and nonreacting
turbulent flows. To illustrate the basic properties and implementation of
spectral methods, consider the unsteady equation for the one-dimensional
advection and diffusion of a scalar variable C, in a velocity field U,
dC + dUC 32C (1)
¥ ^T-^ = °
n
with initial conditions,

C(*,0) = C0(x)
and proper boundary conditions.
The fundamental idea of the spectral methods is to represent the solution
of a differential equation as a truncated series expansion:
C(x,t) = ^ c*(r)4>*00 (2)
where N is the finite wave-number truncation cutoff. In this equation, the

{^(jc), k = 0, ..., N} are called the expansion or basis functions and are
chosen to give the "best" representation of the flow. The {ck(t)} are called
the expansion coefficients and are determined in a manner to ensure that
the series expansion [Eq. (2)] approximates the exact solution as closely
as possible. The specification of the basis functions and the subsequent
adjustment of the expansion coefficients constitute the basic elements of
spectral approximations.
The basis functions in Eq. (2) are infinitely differential, and the ensemble
of these functions in the series expansion represents the dependent variable
in a global sense. In this regard, spectral approximations are different from
finite-difference discretizations in which the solution is obtained locally in
terms of its values on discrete grid points. This global representation gen-
erally leads to a more accurate approximation of the solution and is the
main reason for selecting the spectral method as an alternative to schemes
based on localized basis functions. In prescribing the basis functions, {$k(x)},
one must ensure that the accuracy achieved by the spectral representation
is superior to that attainable by other conventional methods utilizing similar
degrees of freedom. This cannot always be guaranteed, especially in prob-
lems with complex geometries, or in those that contain physical discontin-
uties such as shocks or very thin flame fronts. In addition, for the same
accuracy, a spectral routine should be implemented only if it can be shown
to be more efficient than other competing alternatives. Therefore, trans-
form algorithms must exist that allow the efficient mapping of the depen-
dent variables between the physical and spectral domains.
There are two ways by which spectral methods are classified: 1) the
choice of the basis functions, and 2) the mechanisms by which the error
in the approximation is minimized. The selection of appropriate basis func-
tions is determined primarily by the geometrical configuration and the
boundary conditions of the domain within which the flow is to be simulated.
For example, in a flow with periodic boundaries, exponential Fourier series
are the natural choice, whereas cosine (or sine) Fourier transforms are
applicable for flows with even (or odd) symmetries at the boundaries. For
other boundary conditions, the optimum expansion functions are the ei-
genfunctions of a suitable Sturm-Liouville operator that satisfies the same
boundary conditions. In a finite domain x E [a, b], the family of general
Jacobi polynomials constitutes all of the polynomial solutions that satisfy
the Sturm-Liouville operator. This family includes Legendre and Che-
byshev polynomials of the first and second kind. In a semi-infinite domain
(0 < x < °c), generalized Laguerre polynomials are the appropriate eigen-
functions, whereas Hermite polynomials are the natural expansion func-
tions for simulations in an infinite domain (- oc < x < ^). The expansions
based on these eigenfunctions have the property that, under proper con-
ditions, a rapid convergence independent of the boundary conditions is
achieved. The properties of these expansions are explained in detail in
Refs. 4-6.
After the selection of the basis functions, the subsequent determination
of the expansion coefficients is done in such a way to ensure that the
differential equations and the boundary conditions are satisfied as closely
as possible by the truncated series expansion. This is achieved by mini-
mizing the "residual," i.e., the error produced by approximating the so-
lution as a truncated series expansion. The different means by which this
minimization is performed distinguishes the various spectral approxima-

tions.
Presently, the most commonly used spectral representations can be clas-
sified into one of three main categories: 1) Galerkin methods, 2), collo-
cation methods, or 3) tau methods. The Galerkin approach, which is the
original form of the spectral method, is similar to finite-element methods
in that the physical transport equations can be discretized in terms of a
variational principle. A major problem with implementing the Galerkin
approximation, as will be discussed subsequently, is in evaluating nonlinear
convection terms. The effects of nonlinearities appear in the form of con-
volution sums, which are computationally demanding to evaluate.
The scheme based on collocation, referred to as "pseudospectral" in
Ref. 7, remedies the difficulties associated with evaluating the nonlinear
terms by specifying the dependent variables on collocation points in phys-
ical (rather than spectral) space. With this representation, while spectral
accuracy is achieved in evaluating spatial derivatives, the nonlinear terms
are evaluated explicitly in physical space. This results in significant savings
in the computational effort by eliminating the need to evaluate convolution
sums. However, there is a drawback in this representation in that the
collocation approximation introduces numerical errors known as "alias-
ing." As a consequence of these errors, collocation approximations are
susceptible to numerical instabilities. Techniques for minimizing the effects
of aliasing will be discussed in the following.
The approach followed in the tau method is similar to that of the Galerkin
method. The only difference between the two approaches is in how the
boundary conditions are implemented. In the tau method, the minimization
of the residual is utilized to discretize the conditions at the boundaries as
well as the interior domain, whereas in Galerkin methods, the minimization
is done only for the interior, and the boundary condtions are treated ex-
plicitly by the expansion functions. The tau method has been used mostly
for linear problems, as its implementation is very difficult for nonlinear
differential equations.
Most spectral simulations of turbulent reacting flows have used collo-
cation methods with Fourier or Chebyshev expansions. This is due to the
flexibility of collocation methods in treating nonlinear terms (as will be
shown subsequently), and also because of the computational efficiency
inherent in using trigonometric expansions. Neither the spectral tau method
nor the spectral approximations based on expansions in terms of Jacobi
polynomials other than Legendre or Chebyshev expansions have been used
in turbulence simulations.
To describe the basic elements of spectral approximations, we solve Eq.
(1) using both the Galerkin and collocation methods. The advantages and
drawbacks of the collocation representation in dealing with nonlinear trans-
port will be illustrated. A domain with periodic boundary conditions is
assumed, which allows the use of polynomial expansions in terms of Fourier
series. This selection shows both the influences of aliasing interactions and
how to eliminate them in collocation methods. Finally, Eq. (1) is solved
in a nonperiodic domain to highlight the properties of the spectral method
based on Chebyshev expansions. Only the collocation implementation is

presented, because Cheblyshev polynomials have been most popular in
this representation.
A. Spectral Galerkin Method

The original application of spectral methods in full turbulence simula-
tions was the spectral Galerkin technique.8 To illustrate the properties and
implementation of this technique, we will consider Eq. (1) with periodic
boundary conditions:
[7(0,0 = */(2ir,0 (3a)
C(0,f) - C(27i,r) (3b)
Let Un and Cn denote the values of U and C at TV equally spaced physical
space grid points, xn = 2^nlN, n = 0, 1, ..., TV, where N= 2K 4- 1. In
this case, both U and C can be expanded in finite Fourier series,
u
Un = 2 k exp(ikxn) (4a)
\\k\\^K
(4b)
where the norm, || ||, indicates k = - K, - K + 1, ..., K - 1, K. The

Fourier coefficients uk and ck can be obtained from the inverse transform,
j N
ck = T} 2 Q exp(-ikxn) (5)
/v /j = i
Applying the Fourier transform to Eq. (1) gives
dUC
+ vk2ck = 0 (6)
dt dx
where the tilde indicates variables in the Fourier domain, and k represents
the spectral value at the A:th wave number. The evaluation of the nonlinear
term is slightly complicated. Let W be given by
W(xn) = 2 wk exp(ikxn) = UC
irn^K
u
= 2 k exp(ikxn) E ck exp(ikxn) (7)
\_\\k\\<K J L\\k\\<K J
In this equation, wk will then be given by the convolution sum,

", = + 2 upcq (8)
Inserting Eq. (7) into Eq. (6) yields a set of ordinary differential equations
for the expansion coefficients ck,
-p + _2 ikupcq + vk2 ck = 0 (9)
These equations can be integrated in time using standard methods such

as Runge-Kutta or Adams-Bashforth. The physical space variables are
obtained by substituting the coefficients ck back into Eq. (4).
Accuracy
The popularity of spectral methods in turbulence simulations stems from
three main properties: small phase errors, lack of numerical diffusion, and
rapid convergence. Qualitatively, it has been shown by many authors (e.g.,
Ref. 9) that spectral methods are at least twice as accurate as finite-dif-
ference methods for a given grid resolution. The accuracy of the spectral
method is a result of the rapid convergence properties of series expansions
of smooth functions in terms of orthogonal polynomials. For example, for
the expansion based on Fourier series [Eq. (4)], the truncation error de-
creases faster than any inverse power of k (faster than algebraically) as k
—» o°.4 This fundamental convergence property of the Fourier expansions
is referred to as spectral accuracy. In addition, the spectral method benefits
from a self-diagnostic check of accuracy. Unsatisfactory convergence ap-
pears as large contributions to the series expansion in the high-wave-num-
ber components, so that the adequacy of the resolution can be assessed by
observing the behavior of the spectrum. In finite-difference techniques, a
large amount of damping occurs in the high wave numbers, and this damp-
ing can mask resolution errors.
B. Spectral Collocation (Pseudospectral) Method

A major difficulty with the numerical implementation of the spectral
Galerkin method is due to the computationally intensive evaluation of the
convolution sums resulting from the nonlinear product terms [Eq. (8)]. In
reactive flow problems, these nonlinearities are always present in the con-
vective terms and usually so in the reaction rate terms. A remedy to this
difficulty is to use spectral collocation (pseudospectral) numerical schemes.
The pseudospectral technique is the most common implementation of
the spectral method in use today. In practice, this method differs from the
Galerkin method primarily in the way that the nonlinear terms are com-
puted. In the pseudospectral technique, all of the nonlinear terms are
evaluated in physical space, so that there is no need to determine the
computationally expensive convolution sums. Derivatives are computed in
wave-number space, as illustrated in the following.
The method is outlined here by again considering Eq. (1) subject to
periodic boundary conditions. First, the product UC is computed by mul-
tiplying the two functions in physical space. The spatial derivatives can
then be computed by transforming back into wave-number space and dif-
ferentiating the Fourier series for UC term by term. Let W = UC, where
W is computed by multiplying U and C in physical space. For each physical
space grid point, Wps „ = UnCn. The subscript ps is used here to emphasize
that the multiplication is performed in physical space. W can then be

expanded in a finite Fourier series,
where wps ^ are the Fourier expansion coefficients of WpSM, computed from
the inverse transform [Eq. (5)]. Derivatives are then computed by differ-
entiating Eq. (10) term by term.
(11)
With the derivatives in Eq. (1) computed using the pseudospectral method,
the time advancement may be performed either in physical space or wave-
number space, whichever is most convenient, or more efficient, for a par-
ticular problem. As in the full spectral scheme described in Sec. I. A, any
standard time-stepping scheme may be used for the time integration.
Aliasing Interactions
The evaluation of the nonlinear products as described earlier is free of
phase errors, which, in comparison with finite-difference routines, makes
the pseudospectral scheme very attractive for turbulence simulations. Un-
fortunately, the evaluation of nonlinear terms in physical space generates
wave numbers higher than those retained in the original series expansion.
This is referred to as "aliasing" and often requires special treatment in
pseudospectral approximations. To illustrate the aliasing interactions* that
appear in collocation methods, again consider a multiplication of the type
W(x) = U(x)C(x} (12)
The spectral representation of W(x) is given by
(13)
where wk is the exact spectral representation of W and, as in Eq. (8), is

given by
In the pseudospectral representation, first, the products are evaluated

on the physical grids. After this multiplication, Wps „ is represented by the
finite Fourier expansion,
Wpsn = X w^ Gxp(ikxn) (10')
||Ar||<A:
To show the aliasing errors in Eq. (10'), the Fourier coefficients of Wpsn
are evaluated by their inverse Fourier transform,
H>™ /, — —
1£
>,
= T, E !/„(:„ expC-ita,,) (15)

/V „ = ()
Substituting for Un and C,7 in terms of their Fourier expansions, and by
use of the orthogonality relations of the expansion functions, it can be
shown that10
p + q = k,\\p,q\\<K p + q = k + AM|/7, ^||</C
-f X V* (16)
p + q = k-N,\\p,q\\<K
A comparison of Eq. (16) with the exact representation in spectral space

[Eq. (14)] reveals the aliasing interactions represented by the last two terms
in Eq. (16). These aliasing interactions are introduced when two Fourier
modes are multiplied, and results in frequencies higher than those retained
in the finite expression. In the preceding example, for each component in
the Fourier space, we have
W
"W = k + Wk + N + Wk-N (17)
where wk is the alias-free convolution sum obtained by the spectral Galerkin

method. The last two terms in Eq. (17) indicate that no distinction is made
between a wave number k and its aliases k ± N in the pseudospectral
approximations. In general, if the maximum wave number resolved is K,
the contribution of a wave number k ± jN (j = ± 1 , ± 2 , . . . ) will be aliased
to wave number k. It should also be pointed out that these aliasing errors
are also present in finite-difference approximations. However, finite-dif-
ference approximations generally exhibit a high amount of numerical damp-
ing at the large wave numbers so that effects of aliasing are not as apparent.
Removal of Aliasing
As a result of aliasing interactions, the pseudospectral approximation
can be susceptible to numerical instabilities. In most cases where aliasing
is significant, it is possible to implement procedures easily to minimize this
source of error. A number of different methods have been used to remove
aliasing interactions. An effective and simple method is to ensure that the
energy in the high-wave-number modes remains small. This can be accom-
plished by truncating the series expansion in wave-number space prior to
computing the nonlinear product terms. For Fourier pseudospectral meth-
ods, it can be shown from Eq. (16) that if the maximum retained modes
is reduced from K to M = (2K + l)/3, an alias-free result will be assured.
Note that up — 0 and cq = 0 for \k\ ^ K. To achieve an alias-free result,
the last two terms on the right-hand side of Eq. (16) must be zero. To
satisfy this condition, the cutoff M must be chosen such that
M+M^-M+N
3M < N = 2K + 1 (18)
In practice, a much less severe cutoff is usually acceptable for a suffi-

ciently resolved flowfield. Numerical tests are generally necessary to de-
termine the maximum cutoff wavelength needed in a particular simulation.
A second way of removing aliasing interactions in Fourier collocation
methods is to perform the transforms with an appropriate grid shift. 541 - 12
In this case, the dependent variables must be evaluated both on a grid
shifted by iA* (Ajc = 2^ IN} and on the nonshifted grid. The aliasing errors
cancel upon averaging the two solutions. This doubles the amount of work
over the fully aliased evaluation. An approximation to this scheme is to
evaluate wk on the shifted grid every other time step. In this way, no
additional transforms are necessary and the aliasing errors approximately
cancel on alternate evaluations.
For applications in turbulent flow simulations, the high-wave-number
components can be kept small by restricting the magnitude of the Reynolds
and Peclet numbers to moderate levels. This damps the high-wave-number
components, resulting in a turbulence spectrum that is confined to small
wave numbers. In this way, aliasing interactions at large wave numbers
are minimized.1243 In fact, to obtain an accurate solution in a turbulence
simulation, it is necessary that the major portion of the energy be confined
to the resolvable wave numbers. Under these conditions, the energy in the
highest wave numbers will be small, so that aliasing interactions are not
as severe.
Nonperiodic Boundaries (Chebyshev Expansions}

The trigonometric functions discussed earlier are limited to problems
with either periodic boundary conditions, or in which odd or even sym-
metries can be exploited by the use of sine or cosine expansions. Another
set of widely used expansion functions are Chebyshev polynomials. These
are useful because they allow for a variety of boundary conditions. The
Chebyshev polynomial of the first kind is defined in the interval x G [ — 1,1]
and can be expressed as
Tk(x) = cos[k arccos(jc)] (19)
In the Chebyshev collocation approximation, the solution to Eq. (1) is
assumed to be expandable in a truncated Chebyshev series,
Cn = S ckTk(xn) (20)
k =0
The coefficients ck are determined by the inverse transform,
where
n + 0, N
n = 0, TV
The collocation points usually used in conjunction with Chebyshev poly-

nomials are the Gauss-Lobatto points, given by5
xn = cos(Tm/A/), n = 0, 1, ..., N (23)
This concentrates points near the boundaries at x = ± 1. Using the points
defined by Eq. (23) in the Chebyshev expansion [Eqs. (19-21)] gives
(24)
Equation (24) shows that, with the proper choice of collocation points,
fast Fourier cosine transforms can be used for Chebyshev transformations,
allowing their efficient implementation.
The evaluation of the derivatives for Chebyshev expansions is only slightly
more involved than for full Fourier expansions. Many references are avail-
able for details of the derivation of formulas used in the pseudospectral
Chebyshev method. See, for example, Refs. 4, 5, or 14. Briefly, if Cn is
given by Eq. (20), then the rath derivative can be represented as
c m
~T~z Cn — ^j k Tk(xn) (25)
To obtain the new coefficients, c[m), recurrence relations for the Che-
byshev polynomials and their derivatives are used. For the case of ra — 1,
the algorithm for differentiation is
- - C(x) = 2 cFTk(x) (26)

ax k=o
with
c#> = 0 (21 a)
bk-A1^ = ek+2c^ + 2*4" (27b
where
bk = 2 k =0
= 0 k <0
= I k > 1
and
ek = 1 k <N
= 0 k>N
Note that Eq. (27) indicates that a tridiagonal matrix must be solved to
obtain the coefficients cj^.
With the application of Eqs. (20-27) to evaluate the dependent variables
and their derivatives, the Chebyshev collocation method can be applied
exactly as described earlier. Namely, the spectral expansion based on Eq.

(20) is used primarily to compute spatial derivatives. Nonlinear products
are computed in physical space, and the results are transformed back into
the spectral domain. The mapping between the physical and spectral do-
main is accomplished by application of Eqs. (20) and (21). Finally, with
the implementation of an appropriate time integration scheme, the tem-
poral variation of the dependent variables (in either physical or wave-
number space) can be calculated.
In the preceding discusssions, the spectral approximations were imple-
mented within a specified natural domain, i.e., x E. [0,2ir] for Fourier
expansions, and x E [ - 1, -f 1] for Chebyshev polynomials. Problems within
any interval can be mapped into the natural coordinate by an appropriate
one-to-one mapping function, g, given by
x* = g(x)
where the function g is chosen so that it can be implemented efficiently in
a particular simulation.
C. Spectral Element Methods

Despite the success achieved by spectral and pseudospectral methods in
describing the physics of turbulent flow, the implementation of these schemes
has been restricted to flows with simple geometries. This is mainly due to
the nature of the approximation [Eq. (2)] in which the dependent variables
are represented by a single global expansion within the computational
domain. For practical applications, such an expansion may not be possible
because of the geometrical complexity, or may not be efficient because of
resolution requirements, which may vary widely over the region of interest.
Recently, Patera15 has developed a new method, the "spectral element"
method, that removes the limitations of spectral methods associated with
geometric complexity and resolution requirements, yet retains the high
accuracy. The procedure involves dividing the computational domain into
a number of macroelements, similar to those generated in standard finite-
element techniques. Within these elements, the dependent variables are
interpolated spectrally by high-order expansion functions that pass through
the collocation points defined locally within the element. In most appli-
cations, Chebyshev polynomials have been employed for spectral approx-
imations (the interpolating function) within the macroelements. The
influences of the nonlinear contributions are taken into account by the use
of spectral collocation methods. The linear transport terms are handled
implicitly by the application of a variational principle as in standard finite-
element discretizations. As shown subsequently, the procedure combines
the geometrical flexibility of finite-element methods with the rapid con-
vergence properties of spectral methods. This allows an accurate numerical
representation of the solution that is not attainable by using either of the
techniques alone.
To obtain the spectral element solution of Eq. (1), the nonlinear con-
vective flux and the linear diffusive term are treated separately. This can
be accomplished by using any temporal differencing scheme. For example,
by employing a fractional time-stepping procedure, the time-discretized

form of Eq. (1) at time m becomes
c* - c-
dx
•71 + 1 \
(29)
dx2
where the asterisk denotes the intermediate time step and A/ — tm+ l — tm.
To implement the spectral element method, the domain x E [a,b] is
divided into NE number of macroelements, each of size L' and containing
N'j + 1 collocation points. The values of V and N'j need not be the same
for all elements, allowing the flexibility of achieving high resolution at
desired locations. Each element is identified by its local coordinates xl and
is discretized by collocation points. Using Gauss-Lobatto points in a Che-
byshev representation, the local coordinates x' are mapped into x' in such
a way as to yield x' E [-1,4-1]. The local coordinates of the collocation
points are given by
XJ, = cos (ira/M), n = 0,1, ..., N}, / - 1, 2, ..., N£ (30)
where x« represents the spatial coordinates of the nth collocation point
within the ith element. Within each of these elements, the variables C and
U are represented as the Lagrangian interpolant through the N'j 4- 1 col-
location points. In the 7th element, this is written as
Nf!
n =0
where hln are the interpolation functions and are chosen so that they are
identically zero outside of the ith element, i.e.,
h'n(x() = 8., (32)
With the implementation of Gauss-Lobatto points, these interpolation
functions take the form
W) = Jj E -^ TnW,)T,,(x') (33)
where the coefficients an are provided by Eq. (22).

With approximations (31-33), the nonlinear step is implemented as in
the collocation methods. Equations (31) and (33) are used mainly to form
the spectral approximations of the variables in the Chebyshev domain
within each of the macroelements. The spatial derivatives are evaluated
by relations similar to those given in Eqs. (25-27). Consequently, the
nonlinear term UC is evaluated explicitly on each of the collocation points
by multiplication in physical space. Finally, the advection step is completed,
and the solution C* is obtained at the intermediate time step.
To complete the advancement of Eq. (1) to time level m -f 1, the

diffusion step is carried out [Eq. (29)]. It is here that the relationship to
finite-element methods becomes apparent. This step is completed implicitly
by solving
l m + l 1 1
dx2 * (34)
To construct the finite-element discretization of Eq. (34), the variational

principle is employed in a way similar to that used in conventional finite-
element methods.15 According to this principle, the solution that satisfies
Eq. (34) is equivalent to that which maximizes the functional,
i =y " S J - i [~2\^x~J
(35)
where /' represents the contribution of the /th element in this functional.
To obtain an equation that describes the approximation of Eq. (35), the
Lagrangian interpolants represented by the Chebyshev expansion of
Cl'm +l are used as trial functions and are substituted into Eq. (35). Re-
quiring the stationarity of V with respect to the nodal values Cj;w + 1 (i.e.,
requiring that the variation of 7; with respect to nodal values C^m + 1 vanish),
the equation for each element takes the form
AM™** = ffkn [-(i/vAoa*] (36)
1 an l
where C ^* = C*(x«)> d A* and B denote the elemental matrices and
are related to the coefficients of the Chebyshev expansions (see Ref. 15
or 17 for an explicit derivation and presentation of these matrices).
With the elemental equation [Eq. (36)], the system matrix describing
the solution at all of the interface points may be constructed by employing
the direct stiffness method.15-16 This procedure implies that the variation
of / due to the displacement at a boundary node is simply the sum of its
contributions from each of the elements it bounds. Following this proce-
dure, the system equation takes the form
AMC? +l = 5M[-(l/vAOC,*] (37)
where
Apq = f AL (38a)
/'=!
BM = f B'kn (38b)
i=l
and 2 denotes the direct stiffness summation. The unknown vector.

C™ +1 contains the values of the discrete solution at each interface collo-
cation point (q = 1, 2, ..., NE + 1) at time m + 1. A similar notation is

also employed for the vector C*. The solution of the resulting algebraic
relations, including the elemental equation [Eq. (36)] and the system equa-
tion [Eq. (37)], gives the approximate values of the dependent variables
at time level m + 1 at each interior point (Cj;m + 1 , where n = 0, 1, ...,
Wj, i = 1, 2, ..., A^) and at each interface point (C™ + 1, 9 = 1, 2, ...,
7V£ + 1).
The boundary conditions at i = 1 and NE are imposed in conjunction
with Eq. (36). Dirichlet boundary conditions are imposed by matrix con-
densation [the rows and columns of the matrices corresponding to the point
within the boundary element are eliminated from the system matrix, Eq.
(37)]. The zero-derivative conditions are natural boundary conditions and
do not require special treatment, and nonzero Neumann conditions would
be imposed by a modification of the functional used in the variational
representation. In either case, since Eq. (36) is derived from the variational
principle, the procedure outlined earlier has the advantage that no patching
is required across the boundaries of the element to ensure the continuity
of dC/dx.
A number of numerical tests performed by Patera, 15 Basdevant et al.,18
and Givi et al.19 have confirmed the accuracy of spectral element methods
in the solution of one-dimensional, time-dependent convection-diffusion
equations (see Ref. 17 for a review). In the next section, some additional
points concerning the properties of the spectral element method will be
highlighted.
III. Applications
In general, the application of spectral methods has been restricted to
flows within simple geometrical configurations. Namely, homogeneous
turbulent flows and unsteady shear flows have been the subject of most
spectral simulations. This is primarily due to the nature of the approxi-
mation in which the dependent variables are represented by a single global
expansion.
One of the first applications of spectral methods for the direct numerical
simulation of turbulence is due to the work of Orszag and Patterson.8 They
used the spectral Galerkin method to simulate a three-dimensional ho-
mogeneous isotropic flow within a cubic domain. Periodic boundary con-
ditions were used in all three coordinate directions, allowing all dependent
variables to be expressed in terms of truncated Fourier expansions. Using
a computational grid consisting of 32 points in each direction, Orszag and
Patterson were able to simulate the flow at a turbulence Reynolds number
(\u'/v) of approximately 30, where X is the Taylor microscale and u' the
rms velocity. The results were used to study the decay mechanisms of
isotropic turbulence and to assess the validity of some of the turbulence
models that have been developed previously for predicting homogeneous
turbulence. These simulations demonstrated for the first time the feasibility
of performing direct numerical simulations of a turbulent flow.
This work was soon followed by the application of a pseudospectral
simulation by Fox and Orszag.20 This subsequent work showed the com-
putational efficiency of the pseudospectral technique and indicated that,

with a proper dealiasing procedure, the pseudospectral approximation can
be made as accurate as spectral Galerkin methods. With the available
computer resources, Fox and Orszag were able to simulate a two-dimensional
unsteady flow with a maximum of 1282 collocation points.
These early studies by Orszag and co-workers have been followed by
many additional studies of turbulent flow using spectral methods in direct
simulations. The majority of these works have been directed at nonreacting
flows to study the turbulence dynamics and scalar mixing without the added
complication of chemical reactions.
To evaluate the performance of the pseudospectral method applied to
reacting flows, Riley and Metcalfe21 applied the technique to the ''color"
problem, with the additional influences of diffusion and reaction. The color
problem has been used in many tests to demonstrate the capability of the
pseudospectral method, and involves the advection of a scalar variable
within a rotating velocity field.10-20 Riley and Metcalfe found that high
gradient regions of the scalar were resolved very accurately in these tests
when the number of modes retained in the spectral approximation was
sufficiently large. This observation is a consequence of the lack of numerical
diffusion and dispersion errors in spectral approximations. This is partic-
ularly important for reacting flow simulations, where diffusion of heat and
chemical reactants is the primary mechanism for reaction. When the num-
ber of modes retained was insufficient to resolve the high gradient regions,
a large buildup in the high-amplitude modes was observed, and these lead
to significant errors.
Resolution requirements are one of the main difficulties encountered in
the numerical solution of reacting flows. Flame regions can exhibit ex-
tremely high gradients across the thin reaction zones.1 Resolving these
regions with global polynomial expansions can easily require computer
resources significantly exceeding those presently available. Therefore, it is
necessary to treat reactions with only moderate rates. For example,
Damkohler numbers on the order of 10 (based on the large-scale flow
frequency and the chemical time scale) have been found to be acceptable
for most simulations to date. Damkohler number restrictions can be avoided
entirely when treating the infinitely fast reaction case by using a conserved
scalar.22"24 Under these conditions, the spectral method provides an ac-
curate and efficient means for treating constant-density reacting flows.
With the progress in supercomputer technology, simulations of more
complicated flows with better numerical resolution have been made pos-
sible. Presently, three-dimensional spectral simulations with 643 and 1283
Fourier modes are routine. The results of these investigations have been
very useful in providing insight into mechanisms of turbulent transport.
An extensive body of literature is available that provides a survey of the
state of progress in spectral simulations of nonreacting flows.5-25-26 The
results of some recent applications related to direct numerical simulations
of turbulent reacting flows are described subsequently. For additional dis-
cussion, the reader is referred to Refs. 17 and 27.
In the following, the applications of spectral techniques in the study of

reacting turbulent flows are discussed in order, from the most idealized to
the more complex flows. The simplest flows discussed, homogeneous is-
otropic turbulence, contain highly complex physics, and their accurate
description places severe demands on computer resources. Even for this
flow, computer limitations restrict full turbulence simulations to low Rey-
nolds numbers. Furthermore, the more complex flows described subse-
quently (e.g., the spatially developing shear layer) are still an oversimplified
approximation to turbulent combustion encountered in practical engi-
neering applications. The useful application of the simulations discussed
next stems from their ability to provide detailed information about the
statistics and the dynamics of the flows that have not been attainable by
any other means. Since the complex dynamics of mixing and reaction are
not yet well understood, even in homogeneous turbulence, spectral tech-
niques and direct simulations can and have made considerable contributions
to our knowledge of these flows. Our understanding of more general com-
bustion systems relies on a thorough understanding of the physics in these
idealized configurations.
A. Chemical Reactions in Homogeneous Turbulence

In spectral simulations of turbulent flows, collocation schemes have been
more popular than the Galerkin methods. This, as discussed previously,
is due primarily to the efficiency of the collocation method in treating
nonlinear terms that appear in the transport equations.
An early application of spectral methods for simulating reactive flows is
the work of Hill,28 who considered the effects of mixing on a simple chem-
ical reaction in a two-dimensional homogeneous unsteady flow with up to
32 Fourier modes in each direction. A simple model was used to generate
the unsteady velocity field so that the Navier-Stokes equations were not
solved. The species-transport equations were solved using spectral tech-
niques, and the reaction was assumed to be irreversible and to occur be-
tween two nonpremixed reactants, A and B. The reactants were assumed
dilute in the carrier gas, implying that the amount of heat release due to
chemical reactions was negligible. Under these conditions, the flow was
treated as isothermal and incompressible. The main objective of Hill's
calculations was to use direct simulation to examine the theory of O'Brien29
for an irreversible single-species reaction, and the hypothesis of Toor30 31
for the two-species reaction, A + B —» products. Both of these theories
have been used to predict the rate of reaction in homogeneous, incom-
pressible turbulent reactive flows.
To describe the transport of the scalar quantities, let A(x,t) and B(x,t)
define the instantaneous concentrations of the two reactants. These two
reactants are introduced into the flow under stoichiometric conditions and
undergo a single-step binary reaction of the type A + B —» products in a
constant-density turbulent flow. Under these conditions, the value of the
mean reaction rate can be expressed as
<o>> - Da[(A)(B) + (ab)] (39)
where the transport variables have been decomposed into an ensemble

mean, denoted by (A) and {#), and fluctuating component (denoted by
the lowercase letters). The quantity Da is the Damkohler number based
on the chemical frequency and an appropriate flow frequency. The first
term on the right-hand side of Eq. (39) is the homogeneous mean rate,
which can be computed from the average quantities. The second term,
commonly referred to as the unmixedness , is unknown and requires mod-
eling for complete closure of the mean reaction rate. Toor's hypothesis
provides a means of obtaining a closure for the unmixedness parameter by
relating its decay to the limiting case of no chemical reaction (pure mixing).
The unmixedness term in Eq. (39) can be expressed in normalized form
as
*2(0 = (40)
The subscripts 0 and t refer to time = 0 and r, respectively. In a nonreacting

system, turbulent mixing can be characterized by the mean square of one
of the chemical species,
(41)
where the subscript m refers to pure mixing (no reaction). For no reaction,
d2(t) and ty2(i) are identical. Furthermore, Toor was able to show that
d1(f) and ^2(f) are also identical in the limit of infinitely fast chemistry
under stoichiometric conditions, if a Gaussian shape for the probability
density function (pdf) of the mixture fraction J(x,f) [defined by J(x,t) =
A(x,i) - B(x,t)} is assumed. Based on these observations, Toor hypoth-
esized that
d*(i) = V2(t) (42)
for any Da. Equation (42) states that the evolution of the concentration
correlation (unmixedness) for the conditions given earlier is independent
of chemistry.
Additional simulations by Leonard and Hill3233 and Leonard et al.34
were carried out to look at other effects of turbulence on the reaction rate,
as well as to examine further the validity of Toor's hypothesis. In these
efforts, full three-dimensional direct simulations of a homogeneous tur-
bulent flow using various reaction rates were performed with up to 64
Fourier modes retained in each of the coordinate directions. The results
of these numerical experiments indicated that Toor's hypothesis does not,
in general, hold for the prediction of the conversion rate in nonpremixed
reactants unless the magnitude of the local Damkohler number is very
small. In addition, it was shown that the results deviate substantially from
Toor's result for reactants not in stoichiometric proportions.
Other statistical analysis by Leonard and Hill32 revealed a number of
interesting features that are expected to be important considerations in
future modeling efforts for reacting turbulent flows. In particular, their

work indicated that the local rate of strain in the fluid appears to have a
very important influence on the overall reaction rate. The reaction zones
are aligned perpendicular to the compressive strain, and the magnitude of
the reaction is highest in the regions of highest strain. Because of the
localized nature of the turbulence intensity, the reaction rate was also found
to be highly nonuniform within the reaction zones.
Direct numerical simulations of mixing and reaction in homogeneous
turbulence have since been carried out by Givi and McMurtry, 24 Eswaren
and Pope,35 and McMurtry and Givi.36 Each of these simulated turbulent
mixing in a statistically steady (nondecaying) flow with a resolution of 64
Fourier modes in each direction. The turbulent kinetic energy was main-
tained at an approximately constant value by forcing the energy in the
lowest wave-number components.
Reference 24 involved calculations to examine the validity of Toor's
hypothesis from a fundamental point of view. Simulations were performed
for the two limiting cases of zero-rate chemistry, Da = 0, and infinitely
fast chemistry, Da = °°. The purpose of these simulations was not nec-
essarily to prove or disprove the hypothesis, which was already addressed
in the simulations of Ref. 33, but to use the simulation results to understand
when or why the underlying assumptions are justified, and to suggest ap-
propriate modifications. By using simulation results in this manner, it is
not only possible to evaluate proposed theories or models by comparing
with data, but, in addition, specific strengths and weaknesses can be iden-
tified and understood. This leads to a better understanding of mixing and
reaction in turbulent flows—the primary goal of all studies in turbulent
combustion.
Givi and McMurtry 24 considered a problem in which two reacting species,
A and B, were confined initially to separate regions within the computa-
tional domain. The species were distributed uniformly in the x and z di-
rections, and varied in the y direction such that one-half of the domain
contained species A and the other half contained species B. The simulation
was begun at t = 0, and the species field evolved subject to diffusion,
reaction, and turbulent convection as computed by the scalar transport
equations. The subsequent evolution of the species fields was studied by
flow visualization (computer graphics) and statistical analysis. Two-dimen-
sional contour plots, shown in Fig. 1, provide a visualization of one of the
reacting species concentrations. Figures la and Ib correspond to zero-rate
and infinitely fast chemical reactions, respectively, and each illustrates the
effects of three-dimensional turbulent motion on the mixing of the scalar
field. (At t = 0, the contour lines are parallel to one another for the
initially segregated chemical species.) Comparison of Fig. la with Ib shows
that the effects of chemical reactions increase the steepness of the gradients
and reduce the instantaneous values of the reactant concentrations.
In the infinitely fast reaction simulations, the statistical variation of the
conserved scalar variable J = A - B is the most important parameter to
be identified, because all information about the individual species fields
can be obtained from this variable. Therefore, Givi and McMurtry 24 ex-
a) Da = 0 b) Da
Fig. 1 Two-dimensional contour plots of species A from one x-z plane in three-
dimensional homogeneous turbulence simulation at t* = 0.625.36
amined the evolution of the pdf of this variable. The pdf is defined in the
domain of the random variable / (denoted by O) so that
- Probability (4> < / < < & + d<&)
with
0> e [-!, + !] (43)
The temporal variation of the pdf in the (<£> - t) domain obtained from
the simulations is presented in Fig. 2. This figure shows that, at early times
(t ~ 0), the pdf is composed of two delta functions at 4> = ±1, indicating
the two initially separated reactants A (represented by / = 1) and B
(represented by / = - 1). Subsequent turbulent mixing and diffusion cause
the gap between the two delta functions to fill, and the original peaks
diffuse and move closer to each other. Eventually, the pdf is composed of
a single peak at the mixed concentration, <l> = 0. Proceeding further in
time results in a sharper peak at this mixed concentration, and the simu-
lations indicate that the pdf approaches a self-similar profile that does not
resemble that at t = 0. Since the asymptotic pdf is not self-similar to that
at t = 0, Givi and McMurtry concluded that Toor's hypothesis is not valid
for the infinitely fast reacting case. Instead, they showed that Eq. (42)
should be modified to
lim = Cd2(t) (44)
where
C = 2 (45)
Fig. 2 Evolution of the pdf, P(&,t), of a conserved scalar in a homogeneous tur-

bulent flow.24
P* represents a normalized pdf (P* = Per), where a is the standard de-

viation of the conserved scalar. Whereas Eq. (42) is valid at all times in
the nonreacting case, it should be modified by Eq. (44) as Da = ^ at the
temporal asymptote. For finite Da and for intermediate times, the behavior
cannot be computed analytically, and numerical simulations or laboratory
experiments are necessary to determine the actual description.
In Fig. 3, the ratio [^2(f)ld2(f)} obtained from a series of simulations is
plotted as a function of time for various reaction rates (Da = 0, 2, 8, 30,
and °°). For zero-rate chemistry, the ratio, as expected, maintains a constant
value of 1. For the reacting cases, the value of the unmixedness initially
decreases before it relaxes to an asymptotic value of less than 1, indicating
that the decay rate of the unmixedness does depend on the magnitude of
the Damkohler number. This figure also indicates that the ratio of ^2(t)l
d2(f) approaches the value of 0.64 ~ 2/ir for infinitely fast chemistry. This
value corresponds to the case where P*(cJ>*) asymptotically adopts a Gaus-
sian profile [Eq. (45)].
McMurtry and Givi36 performed additional simulations to study the sta-
tistical behavior of chemical species in turbulent flows. The simulations
were performed again for the limiting cases of zero-rate and infinitely fast
chemical reaction rates for the reaction A + B —> products. These results
1.24-
1. 14- m Dfl -** o

O DR - 30.
(D A Dfl » 8.
O -f Dfl - 2.
1.04- X Dfl « 0.
.64 +
.0 .4 1.2 1.6 2.0 2.4 2.8 „ 3. 2 3. 6
Fig. 3 Temporal variation of ^2ld2 for different reaction rates.24
were used to evaluate the performance of various coalescence/dispersion

(C/D) models that have been used previously for the closure of molecular
mixing in single-point pdf formulations. Models evaluated included those
of Dopazo and O'Brien,37 Janicka et al.,38 and the original C/D model of
Curl.39 The calculations using C/D models were performed using a Monte
Carlo numerical method,40 which solved the pdf evolution equation in a
configuration with the same statistical initialization as used in the direct
simulations. By forcing the lowest velocity modes, a quasisteady cascade
of energy to the small scales was achieved. This resulted in a fairly constant
mixing frequency in the simulations. Therefore, a fixed frequency was
employed in the Monte Carlo simulations. An ensemble of 100,000 com-
putational elements was used to represent the pdf in these simulations.
To evaluate the C/D closures, their predictions obtained in the Monte
Carlo simulations were compared with the DNS results. In Fig. 4, the
predictions for the decay of the mean concentration in the limit that Da
—> oo for each of the C/D models are compared with the direct simulation
results. At early times, the results of the direct simulation are closest to
the predictions from the Dopazo-O'Brien closure.37 At later times, none
of the C/D models exactly predict the conversion rate, and the DNS results
fall between the two models of Refs. 37 and 38. This discrepancy results
from the fact, indicated originally by Pope,41 that none of the C/D models
give correct asymptotic behavior for the evolution of the pdf. Each of the
C/D models considered predicted that the fourth-order moment and all
CD CURL
O JRNICKfl ET flL
DOPflZO flND O'BRIEN
ONS DflTfl
1.2 1.6 2.0 2.4 3.6
Fig. 4 Decay of mean concentration for infinitely fast chemical reaction. Compar-
isons are between DNS and various C/D closure models.36
higher even standardized moments (sixth, eighth, etc.) of a conserved scalar

quantity would become infinite as f —» oc, except for the model of Ref. 37,
which predicts an unphysical constant value for those moments.
This behavior is seen in Fig. 5, which shows a comparison between the
kurtosis (normalized fourth moment) values of species A predicted by the
C/D models and those obtained by DNS. For each model, two curves are
plotted, one for the case without reaction and the other for the case of an
infinitely fast reaction. This figure shows that in the Dopazo-CTBrien ap-
proximation, the two curves coincide, whereas the results of Curl's model
show a maximum difference between the two cases. Consistent with the
previous findings of Ref. 24, the magnitude of the kurtosis of the non-
reacting scalar obtained by DNS approaches an asymptotic value of |JL4 =
2.9, corresponding approximately to that of a Gaussian distribution. In
each case, the standardized moments are higher for the reacting scalar than
for the conserved sealer, indicating that a Gaussian distribution is not
attained for the reacting scalar.
Eswaren and Pope35 have also studied the mixing of a passive scalar in
an isotropic homogeneous flow using spectral techniques. A major focus
of their work was to study the effects of different initial scalar and velocity
length scale ratios (l$ll\ This ratio was found to have a strong influence
on the initial decay of the scalar field, with smaller scalar length scales
tending to decay faster. Later in time (about four eddy turnover times),
70.0-
CURL (RERCTING)
CURL (NON-RERCTING)
JRNICKR ET. RL. (REACTING)
JRNICKR ET. RL. (NON-RERCTING)
DOPRZO-OBRIEN
DN3 DRTR (RERCTING)
DNS DRTR (NON-RERCTING)
1.0
1.2 1.6 2.0 2.4 2.8 ^3. 2 3.6
Fig. 5 Temporal variation of jm4 for Da = 0 and Da —> *>, C/D models and simulation
results.24
the scalar decay rate became independent of the initial conditions. The
initial decay was in agreement with the experiments of Warhaft and Lum-
ley,42 but the later independence of initial conditions was not in agreement.
This discrepancy was attributed to the fact that the observations of Warhaft
and Lumley were obtained for a decaying turbulent velocity field, while
the simulations were for a quasi-steady-state velocity field. However, the
evolution of the scalar pdf was independent of the initial length-scale ratio.
The computed pdf evolved from the initial double-delta-function distri-
bution (representing the two initially unmixed fluids) to an asymptotic
distribution well approximated by a Gaussian (as also observed by Givi
and McMurtry 24 ). Eswaren and Pope35 also found the average dissipation
rate to be independent of the average scalar field only for later times.
These observations are clearly important to modeling efforts for the pdf s
of reacting scalars.
The consistency in results of the various works described earlier rein-
forces the validity of using spectral methods for the DNS of reacting flows.
The results have revealed features of turbulent mixing and reaction that
hopefully will facilitate the construction of a mixing model that will predict
more accurately molecular mixing and chemical reaction in turbulent flows.
B. Chemically Reacting Mixing Layers

The homogeneous turbulence simulations discussed in the previous sub-
section have resulted in a better understanding of turbulent mixing and
have been effective in providing data for evaluating some of the more
commonly used turbulence closures. The comparison of simulation results

with model predictions has revealed both strengths and weaknesses of some
of these models. The homogeneous turbulent velocity field is, however,
an idealized configuration that is not often achieved in practical combustion
applications. Although providing valuable information on the physics of
turbulent mixing and providing data for model evaluation, the homoge-
neous simulations cannot represent many important aspects of mixing that
must be understood in nonhomogeneous systems.
The extension of the pseudospectral technique for the direct simulation
of spatially inhomogeneous reacting flows was initiated by Riley and Met-
calfe.21 Two- and three-dimensional mixing layers were simulated to study
the effects of turbulence on a simple isothermal reaction of the type A +
B —» products. In contrast to the simulations discussed in Sec. III.A., the
small-scale homogeneous mixing in the mixing layer is accompanied by
mixing imposed by large-scale coherent motions. A typical mixing layer is
illustrated in Fig. 6. This flow is characterized by the development of large-
scale coherent structures as the flow evolves in the downstream direction.
The downstream region has steep gradients, and the thickness of the mixing
layer grows with distance. The increase of the shear-layer thickness results
directly in the enhanced mixing, which, in turn, influences the overall
chemical conversion rate. This mixing region provides an ideal configu-
ration for studying the physical processes in the mixing and reaction zones
that exist in real diffusion-controlled combustors. Furthermore, mixing
layers exhibit organized large-scale structures more clearly than any other
flows under the same circumstances, 45 making this an ideal flow for studying
effects of coherent motions on chemical reaction rates.
Temporally Evolving Mixing Layers

The simulations reported in Ref. 21 were of a temporally developing
mixing layer, in which periodic boundary conditions are imposed in the
streamwise direction. This is not the same flow as the spatially evolving
layer that is usually studied in the laboratory, but is an approximation if
one follows a small region of the spatially evolving layer at the mean
Fig. 6 Schematic diagram of a spatially developing mixing layer.

velocity. By studying a temporally growing layer, the requirements of spec-

ifying inflow-outflow boundary conditions, which are difficult to implement
correctly for the spatial layer, are avoided. Periodic boundary conditions
can be applied in the streamwise and spanwise directions, allowing the
construction of a more accurate and efficient numerical routine. This also
allows the efficient implementation of the pseudospectral method. To use
Fourier expansions in the transverse direction, Riley et al.23 assumed a
free-slip condition at the upper and lower boundaries. With this assump-
tion, the expansion of the transport variables in the form of odd/even
Fourier sine/cosine expansions was possible.
Although there are many similar dynamical features between the spa-
tially and temporally developing shear layers (e.g., vortex rollup and pair-
ing process),44 important differences exist. The main disadvantage of a
temporal simulation is that no distinction can be made between the fluids
from the two streams with different velocities. In the temporal simulations,
the amount of fluid entrained into the layer from the two streams is equal.
Therefore, it is impossible to describe some of the interesting entrainment
and asymmetric mixing phenomena that occur in spatially evolving flows.45
Moreover, in a spatially developing layer it is possible for events that occur
downstream to induce changes in flow upstream, whereas in the temporally
evolving flow obviously no current event can affect the flow at previous
times. More of the similarities and differences between the two flows have
been pointed out in Ref. 46.
Riley and Metcalfe21 and Riley et al.23 performed simulations for both
two- and three-dimensional flows. In all cases, the mean streamwise ve-
locity was approximated by a hyperbolic tangent profile, and the initially
unmixed reactants were carried separately in the two feed streams. The
two-dimensional simulations showed the development of the layer in re-
sponse to coherent perturbations corresponding to the most unstable mode
of the hyperbolic tangent profile47 and its first subharmonic. In the three-
dimensional simulations, the velocity field was initialized by random fluc-
tuations with a specified turbulence energy spectrum. Flow visualization
and statistical analysis were performed on the computer-generated data to
study physical mechanisms of mixing and to evaluate the performance of
selected turbulence models.
Figure 7 represents a typical plot of one of the reactants from a three-
dimensional simulation in one particular x-y plane. (In a 643 simulation
there are 64 such planes, each different due to three-dimensional variation.)
The development of the layer was found to be very sensitive to the imposed
initial perturbations, as is the case in laboratory experiments. The pertur-
bations associated with the most unstable mode grew until the initial vortex
rollup was formed within the layer. The subsequent development in time
allows perturbations corresponding to the first subharmonic mode of the
unstable mode to grow, causing the neighboring vortices to pair, thus
generating a second vortex rollup. This rollup and pairing process con-
tributes significantly to the growth and widening of the mixing layer as
first reported by Winant and Browand. 48 The pairing process would con-
tinue if the computational domain were large enough to contain the wave-
length of the higher subharmonic modes, and if additional subharmonics
a)f = 6 b) t = 18
Fig. 7 Contour lines of species A at one particular span wise location. Single rollup
case, three-dimensional simulations.21
were included. The vortex rollup and the pairing of the neighboring vortices
enhance mixing by bringing unreacted species from the two streams into
the chemical reaction zone. As discussed by Riley et al., 23 enhancement
of product formation is a result of both lengthening the reaction front and
the effects of strain associated with the vortex rollup.
The data calculated by Riley et al.23 were in good agreement with both
similarity theory and experiments. The total product formation was com-
pared with that obtained experimentally by Mungal. 49 Although differences
existed between the Reynolds number and entrainment ratios, the scaled
data compared well (Fig. 8). Further statistical sampling of the data was
performed to compute the moments and correlations of the reacting spe-
0.5
0.4 -
_ 0.3 -
G
Ac
0.2 -
0.1 -
0.0
-10 -5
z/z M
Fig. 8 Similarity plots of product concentration. Simulation data of Riley et al.23

compared with experimental measurements of Mungal49 (Ref. 23; courtesy of Physics
of Fluids).
cies, and to determine their importance in turbulent transport and reaction.

In particular, the unmixedness term, (ab) in Eq. (39), was the same order
of magnitude as the homogeneous term, (A) (B} (Fig. 9), emphasizing the
importance of accurately accounting for its contributions.
The data obtained by Riley and Metcalfe21 were compared with various
closure approximations suggested for reacting turbulent flows. In partic-
ular, they looked at the eddy breakup model of Mason and Spalding,50
Toor's hypothesis (see Sec. III.A), and a closure approximation for higher-
order moment equations suggested by Donaldson and Hilst.51 The sug-
gested approach by Donaldson and Hilst was to neglect triple scalar cor-
relations that appear in the high-order closures. Riley and Metcalfe found
that the magnitude of some of these correlations was the same as those of
lower-order correlations, indicating that neglecting these terms would not
result in a good estimate. The model of Mason and Spalding did not predict
the reactant conversion rate accurately. The hypothesis of Toor, however,
which closes the reaction term by estimating the unmixedness based on its
nonreacting value, provided a reasonable approximation to their simula-
tions.
Further numerical studies of three-dimensional instabilities in constant-
density mixing layers have been carried out by Metcalfe et al. 52 Similar to
the work of Ref. 23, pseudospectral direct numerical simulations were used
to simulate a temporally developing mixing layer. The results of these
simulations showed that, after the initial stages of primary growth and the
formation of dominant coherent vortices, the development of three-dimen-
sional (secondary) instabilities became apparent. At high amplitudes, the
0.06
0.04 -
0.02 -
-0.00
-0.02 -
-0.04
-15 -10
z/z
M,
Fig. 9 Statistics of the mean reaction rate plotted across transverse coordinate of
mixing layer. Average product of the concentration, product of the average con-
centrations, and the unmixedness (Ref. 23; courtesy of Physics of Fluids).
secondary instabilities were characterized by streamwise-oriented, counter-

rotating vortices that formed in the braids of the large-scale structures (Fig.
10). An important effect of these vortex structures is the enhanced mixing
they promote by pumping fluid up from the two streams through the vortex
braids. Figure 11 shows the development of mushroom-shaped puffs of the
fluid that result from the vortex-induced velocities. The effects of these
structures simulated by Metcalfe et al.52 are very similar to what has been
observed in the laboratory. 53
Variable-Density Simulations
The simulations discussed earlier have been useful for understanding
many mechanisms of turbulent mixing. However, because a constant-den-
sity flow was assumed, any influence of heat release from chemical reactions
on the fluid dynamics could not be addressed. McMurtry et al.54-55 contin-
ued direct numerical simulations of two- and three-dimensional temporally
developing reacting mixing layers, but accounted for the influences of
density variation resulting from exothermic chemical reactions. In these
simulations, a low-Mach-number approximation to the exact governing
equations was used, which resulted in the removal of acoustic waves. This
Mean Flow
into Page
0)
CO
o3
CO
03
©
Mean Flow
Out of Page
y Spanwise
Fig. 10 Contour plots of stream wise vorticity. Dashed lines indicate negative vor-
ticity, solid lines positive vorticity. View looking into the mixing layer as the contour
is taken from a two-dimensional, stream wise cut through the braid of the vortex
structure (Ref. 52; courtesy of Journal of Fluid Mechanics).
0
Mean Flow
into Page
<D
i
(/>
c3
CO
©
Mean Flow
Out of Page
y Spanwise
Fig. 11 Contour plot of one species concentration at same physical location as Fig.
10 (Ref. 52; courtesy of Journal of Fluid Mechanics).
is similar to methods developed previously and implemented earlier by

Rehm and Baum56 and Sivashinsky.57 With this approximation, the need
for constructing a fully compressible code was not necessary. Therefore,
although the formulation allows for the development of density variations
by the combustion heat release, the computational time-stepping con-
straints required for numerical stability in tracking the high-frequency acoustic
waves were eliminated.
The results of the numerical experiments in Refs. 54 and 55 showed that
heat release has some profound effects on the fluid flow. One of the most
apparent influences observed when heat release accompanies the chemical
reaction was a significant decrease in the total amount of product generated
compared with the constant-density reaction (Fig. 12). Along with this
decrease in product formation, the shear-layer thickness (velocity half-
width or vorticity thickness) grew at a slower rate. This is consistent with
the experimental observations of Refs. 58 and 59.
McMurtry et al.55 studied the development of the flow from a number
of different viewpoints to understand the mechanisms responsible for the
previously mentioned phenomena. This included a study of the turbulent
shear-stress distribution, the turbulent kinetic-energy balance, vorticity dy-
namics, and stability theory. A study of the flow in terms of the vorticity
dynamics was found to be particularly useful because it allowed a descrip-
tion of the flow in terms of the large-scale vortical structures.
Run I (no heat release)

Run 2 (heat release)
Fig. 12 Total product formation. Three-dimensional simulations, with and without

heat release. Curve plotted is the volume-averaged product mass fraction (Ref. 55;
courtesy of Journal of Fluid Mechanics}.
In a general three-dimensional flow, the vorticity equation can be written

in the following form:
+ vV2co (46)
Four different mechanisms can be identified that alter the vorticity: vortex
stretching, thermal expansion, baroclinic torques, and viscous diffusion.
For low-Mach-number flows without heat release, the expansion and bar-
oclinic torque terms are very small. When density changes due to com-
bustion heat release occur, these terms can play an important role in the
development of the vorticity field.
A comparison of the spanwise component of vorticity for simulations
with and without heat release is given in Fig. 13. One of the most apparent
differences is that the maximum amplitude in vorticity is decreased sub-
stantially in the simulations that included chemical heat release. In addi-
tion, with energy release, the vorticity is no longer as concentrated in the
center of the vortex cores as it is in the constant-density case.
The lower growth rate observed in both two- and three-dimensional
simulations was primarily attributed to the contributions of the thermal
expansion and the baroclinic torque. In an expanding flow, V • U is positive.
Thus, the effects of thermal expansion directly result in a decrease in the
magnitude of vorticity. The baroclinic torque describes the effects of dif-
ferential fluid accelerations resulting from nonaligned pressure and density
b) 2KPT
Fig. 13 Contour plots of span wise component of vorticity. Three-dimensional sim-

ulations: a) no heat release, contour interval from — 1.3 to 0.0; b) with heat release,
contour interval from — 0.81 to 0.0 (Ref. 55; courtesy of Journal of Fluid Mechanics}.
2n j 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 M 1 1 1 1 1 1 1 1 1 1 M 1 1 1 M 1 1 1 i M 1 1 1 1 1 1 1 1 1 1 1 1 1 1 L
Fig. 14 Contour plot of spanwise component of baroclinic torque at one spanwise

location. Contour interval from -0.41 to 0.61 (Ref. 55; courtesy of Journal of Fluid
Mechanics).
gradients. Visualization of the spanwise component of this term shows a

sign that alternates across the reaction surface (Fig. 14). This can be under-
stood by realizing that while the density gradient changes sign across the
reaction surface, the pressure gradient always points approximately radially
outward from the center of the vortex cores. With rollup of the unsteady
shear layer, the vorticity generated by the heat release is engulfed into the
vortex core with primarily an opposite sign of the mean shear-layer vor-
ticity. A combination of the effects of the baroclinic torque and the thermal
expansion results in a more diffuse vorticity field and a lower rotational
velocity within the core. This results in the lower growth rate of the layer,
which leads to a reduction in area of the reaction surface and, therefore,
decreases the rate of chemical product formation within the mixing zone.
Calculations of the energy in the individual modes also showed a marked
effect of heat release. The most unstable mode in the constant-density case
stabilized quickly when heat release occurred, while the subharmonic re-
mained unstable but grew at a slower rate (Fig. 15). This is consistent with
the preceding observations, which show that less intense vortex structures
develop when chemical reactions with heat release occur.
In the three-dimensional simulations, it was shown that the secondary
vortex structures are also inhibited as a result of heat release. This is
illustrated by contours of one of the reacting species fields taken at a
streamwise cut through the core of one of the coherent structures. A
comparison of Figs. 16a and 16b shows that the development of the mush-
room-shaped structures is inhibited severely by the heat release. Mecha-
nisms similar to those discussed earlier also act on the streamwise vorticity.
Futhermore, the magnitude of the streamwise component of vortex stretch-
ing also decreased. This further reduces the growth of the streamwise
vorticity, as vortex stretching is the primary mechanism for intensifying
vorticity.
Chen60 used the simulation results generated by McMurtry 61 to further
evaluate methods for modeling turbulent flows. In particular, Chen carried
out a detailed comparison of the results predicted both by Favre averaging
techniques and by Reynolds averaging techniques applied to the conser-
vation equations of chemical species and momentum. In the Favre aver-
aging technique, all dependent variables except pressure are weighted by
the density prior to averaging, then divided by the mean density.62 All
variables appearing in the transport equations for the averaged equations,
and in the transport equations of the scalar fluxes, Reynolds stresses, and
turbulent kinetic energy were computed. The performance of the standard
A>e model2 was studied for each of the averaging methods. Chen compared
models for cases in which the maximum density change resulting from the
heat release was a 50% decrease from the freestream value. His calculations
showed that, for this case, the individual average quantities ({£/), (O) and
the fluctuation correlations ((uv), (uc)} did not vary significantly from their
Favre-averaged counterpart. However, terms such as (p)(U)(V) and (p)(U)(C)
were different from their Favre-averaged counterparts. The density fluc-
tuation terms were important in the various correlations and cannot be
ignored. The k-z model worked fairly well for both averaging techniques
when the mixing layer was fully developed and no countergradient trans-
a)
Run 1 (no heat release)
10
4
10
c
10 - 5
UJ
10
10 ~
10 20 30 40 50 60 70 80 90 100
Time
i i
b)
Run 1 (no heat release)
10
10
\________\
0 10 20 30 40 50 60 70 80 90 100
Time
Fig. 15 Comparison of energy in a) fundamental mode for constant-density case
and chemical reaction case with heat release. Note stabilization effect of the heat
release, b) Energy in subharmonic mode for same two cases (Ref. 55; courtesy of
journal of Fluid Mechanics).
a) Constant density b) Heat-release case
Fig. 16 Contour plot of species concentration at one streamwise location. (Plane

perpendicular to mean flow.) Observer is looking into flow (Ref. 55; courtesy of
Journal of Fluid Mechanics).
port of momentum or scalar was apparent. Under these conditions, the

Favre averaging technique was the better choice, mainly because of the
simplifications it allows in the governing equations.
Nonequilibrium Chemistry
Although the simulations performed by McMurtry et al.54-55 accounted
for effects of energy release, they were unable to describe some of the
important nonequilibrium effects resulting from complex reaction mech-
anisms. To describe the transport of the species due to reaction in a more
realistic manner, Givi et al.63 performed pseudospectrial simulations using
a more general reaction mechanism. Similar to the simulations of Riley et
al.,23 the flow was assumed incompressible, but the reaction rate included
a temperature dependency described by the Arrhenius law,
co = Da A B exp (-EIR T) (47)
where E, R, and Tare the activation energy, gas constant, and temperature,
respectively. Using this formulation, Givi et al.63 were able to study aspects
of chemical nonequilibrium related to local flame extinction phenomena.
The development of the flow they studied is represented in Fig. 17, which
shows contours of a conserved scalar defined by Sz = (A — B + B^)l
(Ax + Bx). Here, A and B are the local species concentrations and the
subscript °° indicates the freestream values. This figure again illustrates the
development of the layer into large coherent vortex structures and their
subsequent pairing. As discussed earlier, this rollup and pairing process
stretches and wrinkles the reaction surface, leading to an enhanced overall
mixing rate. For an infinitely fast stoichiometric chemical reaction, the
a. t*=6 b. t*=12
a) /* = 6 b) /* = 12
Fig. 17 Contour plots of conserved scalar variable. Two-dimensional simulations.63
reaction occurs along the stoichiometric surface, Sz = 0.5, and the rollup
and pairing process results in an increase of the overall chemical conversion
rate.
For the finite-rate, temperature-dependent chemistry applied by Givi et
al.,63 the increased mixing resulting from the vortex rollup did not always
result in an increase in the product formation, shown in Fig. 18. This figure
shows the reaction rate going to zero in the braids of the large-scale struc-
tures, even though mixed fluid exists there. The rollup process results in
high strain rates in the braids, and when this strain rate is sufficiently large,
the reaction rate goes to zero. This is in agreement with the theory and
previous asymptotic studies of flame extinction by Peters64 and Peters and
a) t* = 6 b) I* = 12
Fig. 18 Contour plots of instantaneous reaction rate.63
Williams.65 Qualitatively, this can be explained by realizing that the high

gradients experienced by the scalar variables in the braids result in high
heat-transfer rates. If the dissipation of heat is not at least balanced by the
generation of heat due to reaction, the temperature can drop below a
critical value and the flame will go out, even if the reactants are well mixed.
A logical extension of the work described here would be to include more
complex, multistep kinetic schemes. This will be necessary for a more
complete understanding of extinction and liftoff phenomena in laboratory
flames, where important physics may be linked inherently to multiple-step
reactions.66 For example, there is experimental and numerical evidence67"70
that there can be a preferential "leakage" of reactants through hydrocarbon
diffusion flames. This behavior cannot be described by single-step kinetics
(see Refs. 17 and 66 for an elaborate discussion.)
Spatially Evolving Mixing Layers

The simulation results for temporally evolving shear layers are consistent
with both laboratory experiments and analytical theories. However, there
are issues related to the spatial development of the flame that cannot be
addressed within the temporal framework. These issues, which include
differential entrainment rates and feedback mechanisms, are important for
investigating some of the physical phenomena observed in laboratory re-
acting flows.
The simulation of a spatially developing mixing layer requires a correct
description of the boundary conditions at the inflow and outflow. Because
of the restrictions of the pseudospectral methods in treating general bound-
ary conditions, the numerical schemes used in the works discussed here
cannot be used to account for spatial evolution. The first spectral simu-
lations of spatial layers were hybrid schemes that used finite-difference
techniques in the direction of streamwise development and pseudospectral
methods in the other spatial coordinates (see, for example, Refs. 71-73).
These works were capable of addressing the asymmetric mixing within the
shear layer. Consistent with laboratory experiments, 43 the simulations re-
vealed a preference for more fluid from the high-speed stream to be drawn
in the layer than from the low-speed stream.
To achieve the accuracy of the spectral method and the generality needed
to account for more complex geometries and boundary conditions, Givi et
al.19 and Givi and Jou74 applied the spectral element technique to the spatial
mixing layer (Fig. 6). The two reactants were unmixed initially, with one
reactant entering the computational domain in the upper, high-speed stream,
while the other reactant entered in the lower low-speed side. The subse-
quent reaction between the two species was taken to be single step, irre-
versible, and to obey the temperature-dependent Arrhenius rate equation
[Eq. (47)].
Both two- and three-dimensional simulations were performed. In both
cases, impermeable, free-slip boundary conditions were used at the upper
and lower computational boundaries in the cross-stream direction. The
spectral element method was applied only in the direction of spatial de-
velopment. In the three-dimensional simulations, periodic boundaries were
used in the spanwise direction, allowing pseudospectral methods with Four-

ier transforms, similar to those employed in temporal simulations, to be
used in both the transverse and spanwise directions. The stream wise di-
rection was discretized by NE macroelements, where each element consists
of Nj + 1 collocation points. The results of the two-dimensional simulations
reported in Ref. 74 were obtained by forcing the layer at the inflow by
imposing perturbations corresponding to the most unstable mode and its
subharmonics. At the outflow, a weak condition of zero second derivatives
was applied for all dependent variables. In the two-dimensional simula-
tions, 64 Fourier modes were retained in the cross-stream direction and
42 macroelements were used for the streamwise discretization. Within each
element, a fifth-order Chebyshev polynomial was used to approximate the
variables in the streamwise direction. This results in a total of 13,504 grid
points within the computational domain. Even if spectral accuracy was not
to be considered (which was achieved), this discretization would be equiv-
alent to a fifth-order-accurate finite-difference technique.
Figure 19 illustrates the streamwise development of the spatial layer. In
this figure, the contours of the conserved scalar variable, 52, in the spatial
flow are depicted. The growth of the layer in response to perturbations
applied at the inflow was similar to that observed in previous simulations
of spatially evolving flows (e.g., Refs. 71 and 73) and analogous to that
observed in temporally developing flows. The initial vortex rollup near the
splitter plate is again due to the development of the most unstable fre-
quency. The growth of lower subharmonics results in additional rollups
and pairings. A zero second-derivative condition applied at the outflow
allowed the vortex structures to travel out of the computational domain.
Extensive numerical tests indicated that the errors associated with this
boundary condition tend to be confined within the last computational ele-
ment.74
By applying a finite-rate Arrhenius chemistry, the effects of unsteady
vortex dynamics on the local quenching of the flame could be studied.
Qualitative information about these effects are revealed in a graph of the
instantaneous reaction rate (Fig. 20). This figure shows that the reaction
rate is fairly uniform along the mixing layer near the inlet where the reac-
tants are first brought into contact. Further downstream, where the local
magnitude of the instantaneous dissipation can become large, the reaction
rate approaches zero, and the flame is quenched locally. This is the mech-
anism of flame extinction previously studied in temporal simulations by
Givi et al.63 and also in the experimental observations reviewed by Tsuji75
Fig. 19 Contour plot of conserved scalar variable showing spatial development of

mixing layer. Two-dimensional simulation.74
Fig. 20 Instantaneous reaction rate at same time as Fig. 19.74
and Peters.64 As noted earlier, at the regions of high strain, the local
dissipation of the heat can be greater than that supplied by the chemical
reaction. Under these conditions, the temperature decreases below a crit-
ical value and the flame cannot be sustained.
The extension to three-dimensional mixing-layer simulations using spec-
tral element techniques has been initiated recently by Givi.76 In this initial
effort, a low-resolution grid with only 32 Fourier modes was employed in
the spanwise direction with 13,504 collocation points in the other two
directions as before. In addition to the two-dimensional perturbations, a
three-dimensional disturbance also was added to initiate a three-dimen-
sional evolution. The streamwise development of the flame sheet is illus-
trated in Fig. 21 for an infinitely fast reaction. In this figure, the effects of
TI SO =-0.5000r PLOT 3D.C Z

FIELD - CZr ISTEP= 36, TIME- 7.2
LEVEL= 0.5000, E T E = t -25, -75, 100)
Fig. 21 Convolution of flame sheet in three-dimensional simulations.1

flow instabilities on the convolution of the flame surface and the influences
of both primary and secondary structures of the layer on the distortion of
the flame sheet are displayed. It is shown that, after the initial stages of
primary growth, the secondary streamwise structures play significant roles
in the enhancement of the reaction and the convolution of the flame sur-
face.
Figure 22 illustrates the effects of the increasing strain on the flame with
finite-rate chemistry. This figure shows the instantaneous values of the
temperature across the mixing layer as a function of the mixture fraction.
Figure 22a was generated by sampling data at a location where the strain
rates are low. The magnitude of the temperature rises as the mixture
fraction approaches its stoichiometric value (0.5). At this value, the tem-
perature reaches a maximum, and then decreases to the freestream tem-
perature as the mixture fraction increases to 1 (no mixed fluid). Figure
22b is the same as Fig. 22a except the data are sampled further downstream
where the strain rates are higher. This case shows that, even at the stoichio-
metric mixture fraction, the temperature can, in places, be near the free-
stream value, indicating local extinction.
It would be informative to simulate the spatial development of the flow
under the influence of random three-dimensional perturbations (with a
specified turbulence spectrum) at the inlet of the mixing layer, as did
McMurtry et al.55 and Riley et al.23 in temporal simulations. In this way,
the response of the shear layer to the random forcing would resemble more
closely that of a laboratory experiment under the influence of random
upstream turbulence.

As discussed in the Introduction, the application of spectral methods
has had its most important recent impact in the area of direct numerical
simulation (DNS). This is a result of the exceptional convergence properties
of the spectral method for properly described problems, and also its lack
of phase errors, which make it particularly useful for turbulence simula-
tions. The main disadvantage of spectral methods over other commonly
used discretization methods is that the choice of expansion functions is
determined primarily by the boundary conditions of the problem. In com-
plex geometries with arbitrary boundary conditions, it is often difficult, or
impossible, even with mapping techniques, to represent the flowfield ac-
curately. The nonlocal properties of the spectral approximation are one
reason for its high accuracy. However, these nonlocal properties also make
it difficult to capture physical discontinuities such as those that occur across
thin flames or strong shocks. In addition, the use of spectral methods
involves repeated summations of the expansion functions to transform
variables from physical to spectral space. This is done efficiently only if
fast transform algorithms exist.
Many of the difficulties associated with complex geometries and bound-
ary conditions can be overcome with the spectral element technique. By
using Chebyshev polynomials as the expansion functions within each ele-
ment, spectral accuracy (defined in Sec. II.A) can be achieved. The use
TWOD.PRT1.3601-6801
. 10 .20 .30 .40 .50 .60

MIXTURE FRflCTION
b) *
TWOD.PRT3.3601-6801
.00 .30 .40 .50 .60

M I X T U R E FRflCTION
Fig. 22 Instantaneous temperature as a function of the mixture fraction, a) At a
location where dissipation rates are low. b) At a location where local regions of high
dissipation exist.19
297
of this method in simulations of turbulent reacting flows is in its infancy,

but its limited applications to date have been encouraging. This appears
to be the most promising scheme to be employed in numerical studies of
flame stability, and turbulence-combustion interactions in complex geo-
metries. These types of studies are necessary for a thorough understanding
of the dynamics of reacting turbulent flows. The spectral element technique
can also lead to improvements in treating very high gradient regions be-
cause smaller, more densely packed elements can be inserted in the regions
of high activity.
A fundamental limitation with using spectral or other methods for the
direct numerical simulation of turbulence is the amount of computer re-
sources required. Typical three-dimensional codes for a 643 grid and in-
cluding species transport and energy release written by the authors take
on the order of 10 s per time to run on a CRAY X-MP. Since the flowfields
are highly unsteady, numerous time steps are carried out, resulting in
computer runs of several hours for full simulations. Current computational
limitations restrict the Reynolds number to moderate levels, generally less
than several hundred, to keep the range of length and time scales within
acceptable limits. Many flows of practical engineering application have
Reynolds numbers greater than 105. Because the resolution requirements
in a three-dimensional flow are proportional to Re9/4, it is obvious that the
direct simulation of a turbulent reacting flow at realistic conditions is not
a practical consideration.
Despite these limitations, the use of pseudospectral techniques to per-
form DNS of more idealized flows in simple geometries has lead to im-
portant understanding of mixing and reaction in turbulent flows (Sec. III).
This comes from the extreme detail in which the numerically generated
flowfield can be studied. Since the entire flowfield is known at every grid
point and at every timestep, any statistical quantities of interest can be
computed, including many of which are difficult or impossible to measure
experimentally. The accuracy of the computed statistics is, of course, re-
lated to the number of grid points used in the simulation.
In Sec. Ill, numerous comparisons between simulation results and ex-
perimental observations were described. Although the differences between
physical and numerical conditions generally allowed only a qualitative com-
parison, the agreement between the simulation data and experimental re-
sults has been encouraging. The detail in which the numerical data could
be studied has, in many instances, provided us with the information to
identify specific physical mechanisms of turbulence-chemistry interactions.
Continued contributions in this area are expected.
Higher resolution simulations than currently attainable—256 3 or even
larger—will not allow simulations with significantly higher Reynolds num-
bers, but will allow an accurate description of more complicated flows with
increasingly complex physical phenomena. As such, DNS (the main ap-
plication of spectral schemes) will remain limited to research applications.
Nevertheless, within its domain of realistic application, DNS is expected
to continue to enhance our understanding of turbulent mixing and reaction.
This is accomplished by the ability of the direct simulation to provide
specific information concerning the global and detailed structures of the
flow, and by providing a quantitative basis for evaluating the performance
of turbulence closures.
An important research area in which spectral methods can be expected

to play an important role is in large-eddy simulation (LES). The basic idea
of LES is to solve for the large-scale structure of the flow explicitly, as in
a direct numerical simulation, and model the effects of the small scales
using "subgrid" models.77 For high-Reynolds-number flow, the dynamics
of the small scales do not depend significantly on problem geometry. The
amount of energy dissipated at the small scales is determined by the energy
input to the large-scale motion, which is controlled by the boundary and
operating conditions of the flow. This aspect of the large-scale dynamics
would be computed explicitly in the LES. The action of vortex stretching
over a wide range of wave numbers leads to an isotropic, more universal
flow structure at the small scales of motion. This suggests that simpler
models, with more generality, can be applied to describe the small-scale
dynamics.
For reacting flows, the situation becomes more complex. In addition to
the subgrid turbulent scalar transport, the mean reaction rate must also be
modeled to predict the chemical conversion rate accurately. Unfortunately,
the dynamics of the small-scale chemical behavior are not well understood.
As a result, no subgrid models for the reaction rate have been promoted
with much enthusiasm. It is anticipated that data generated from direct
simulations will help to stimulate modeling efforts for LES. DNS has al-
ready provided us with many clues to relate reaction rate statistics with
the hydrodynamic flowfield. For example, the interactions among strain
rates, vorticity, scalar dissipation, and reaction rates have been described
through many of the works discussed in Sec. III. Any proposed models
would need to be evaluated in a simple configuration of mixing and re-
action. A direct simulation utilizing spectral methods could provide the
data necessary for evaluating proposed models. Such efforts aimed at eval-
uating the performance of subgrid models have been used previously in
nonreacting large-eddy simulations78 and are recommended for reactive
flow LES.
Finally, the application of numerical methods in combustion research
will undoubtedly continue to expand as computing resources and capabil-
ities continue to grow. Significant computational advances are expected to
come from new developments in parallel architecture. This has been shown
to be the most economical way to combine massive memory with massive
processing capabilities. Computer codes utilizing spectral methods are par-
ticularly amenable to parallelization without substantial efforts. Fluid dy-
namics codes that we have developed and utilized are all highly parallel,
and user-friendly hardware and software are becoming more readily avail-
able to take advantage of this. Future work in this area will involve the
optimization of parallel processing procedures in conjunction with direct
numerical simulations of turbulent reactive flows.
Acknowledgments
The work of the first author was supported by the Advanced Combustion
Engineering Research Center. Funds for this center are received from the
National Science Foundation, the State of Utah, 23 industrial participants,
and the U.S. Department of Energy. The work of the second author is
being supported by the National Science Foundation under Grant CTS-
9057460, by NASA Lewis Research Center under Grant NAG 3-1011, and
by American Chemical Society, Petroleum Research Funds under Grant
22227-G6. Computational resources have been provided by NCSA at the
University of Illinois, and the authors are currently supported jointly by
NSF under Grant CTS-9012832.
References
:
Oran, E. S., and Boris, J. P., "Detailed Modeling of Combustion Systems,"
Progress in Energy and Combustion Science, Vol. 7, 1981, p. 1.
2
Launder, B. E., and Spalding, D. B., "The Numerical Computation of Tur-
bulent Flows," Lectures in Mathematical Modeling of Turbulence, Academic, Lon-
don, 1972.
3
Cooley, J. W., and Tukey, J. W., "An Algorithm for the Machine Calculation
of Complex Fourier Series," Mathematical Computations, Vol. 19, 1965, p. 297.
4
Gottlieb, D., and Orszag, S. A., Numerical Analysis of Spectral Methods: Theory
and Applications, SIAM, Philadelphia, PA, 1977.
5
Canuto, C., Hussaini, M. Y., Quarteroni, A., and Zang, T. A., Spectral Methods
in Fluid Dynamics, Springer-Verlag, New York, 1987.
6
Givi, P., "Spectral Methods in Turbulent Combustion," Numerical Modeling
in Combustion, edited by T. J. Chung, Hemisphere, Washington, DC (to be pub-
lished).
7
Orszag, S. A., "Numerical Simulation of Incompressible Flows Within Simple
Boundaries. 1. Galerkin (Spectral) Representation," Studies in Applied Mathe-
matics, Vol. L, No. 4, 1971, p. 293.
8
Orszag, S. A., and Patterson, G. S., "Numerical Simulation of Turbulence,"
Lecture Notes in Physics Statistical Models and Turbulence, Springer-Verlag, New
York, 1972, p. 127.
9
Herring, J. R., Orszag, S. A., Kraichnan, R. H., and Fox, D. H., "Decay of
Two-Dimensional Homogeneous Turbulence." Journal of Fluid Mechanics, Vol.
66, 1974, p. 417.
10
Orszag, S. A., Numerical Simulation of Incompressible Flows Within Simple
Boundaries: Accuracy," Journal of Fluid Mechanics, Vol. 49, 1971, p. 75.
n
Patterson, G. H., and Orszag, S. A., "Spectral Calculations of Isotropic Tur-
bulence: Efficient Removal of Aliasing Interaction," Physics of Fluids, Vol. 14,
1971, p. 2538.
12
Rogallo, R. S., "Numerical Experiments in Homogeneous Turbulence," NASA
TM-81315, 1981.
13
Ferziger, J. H., "Higher Level Simulations of Turbulent Flows," Computational
Methods for Turbulent, Transonic and Viscous Flows, edited by J. A. Essers,
Hemisphere, Washington, DC, 1983, p. 93.
14
Schumann, U., Grotzbach, G., and Kleiser, L., "Direct Numerical Simulations
of Turbulence," Predictive Methods for Turbulent Flows, edited by W. Kollmann,
Hemisphere, Washington, DC, 1980, p. 120.
15
Patera, A. T., "A Spectral Element Method for Fluid Dynamics—Laminar
Flow in Channel Expansion," Journal of Computational Physics, Vol. 54, 1984, p.
468.
16
Baker, A. J., Fianite Element Computational Fluid Mechanics, Hemisphere,
Washington, DC, 1983.
17
Givi, P., "Model Free Simulations of Turbulent Reactive Flows," Progress in
Energy and Combustion Science, Vol. 15, 1989, pp. 1-107.
18
Basdevant, D., Deville, M., Haldenwang, P., Lacroix, J. M., Ouazzani, J.,
Peyret, R., Orlandi, P., and Patera, A. T., "Spectral and Finite Difference So-
lutions of the Burgers Equation," Computers and Fluids, Vol. 14, No. 1, 1986, p. 23.
19
Givi, P., Jou, W.-H., McMurtry, P. A., and Metcalfe, R. W., "Direct Nu-
merical Simulations of a Non-Premixed Turbulent Jet Flow," Final Report to
AFOSR; also, Flow Research Co., Kent, WA, TR-440, 1988.
20
Fox, D. G., and Orszag, S. A., "Pseudospectral Approximations to Two-
Dimensional Turbulent Flows," Journal of Computational Physics, Vol. 11, 1973,
p. 612.
21
Riley, J. J., and Metcalfe, R. W., "Direct Numerical Simulation of Chemically
Reacting Turbulent Mixing Layers," NASA CR-174640, 1984.
22
Williams, F. A., Combustion Theory, Benjamin/Cummings, Inc., Menlo Park,
CA, 1985.
23
Riley, J. J., Metcalfe, R. W., and Orszag, S. A., "Direct Numerical Simulation
of Chemically Reacting Mixing Layers," Physics of Fluids, Vol. 29, No. 2, 1986,
p. 406.
24
Givi, P., and McMurtry, P. A., "Non-Premixed Reaction in Homogeneous
Turbulence: Direct Numerical Simulations," AlChE Journal, Vol. 34, No. 6, 1988,
p. 1039.
25
Rogallo, R. S., and Moin, P., "Numerical Simulation of Turbulent Flows,"
Annual Review of Fluid Mechanics, Vol. 16, 1984, p. 99.
26
Hussaini, M. Y., and Zang, T. A., "Spectral Methods in Fluid Dynamics,"
Annual Review of Fluid Mechanics, Vol. 19, 1987, p. 339.
27
Jou, W.-H., and Riley, J. J., "Progress in Direct Numerical Simulations of
Turbulent Reacting Flows," AIAA Paper 87-1324; also, AIAA Journal, Vol. 27,
No. 11, 1989, pp. 1543-1556.
28
Hill, J. C., "Simulation of Chemical Reaction in a Turbulent Flow," Proceed-
ings of Second R. F. Ruth Chemical Engineering Research Symposium, 1979.
29
O'Brien, E. E., "Postulate of Statistical Independence of Decaying Reactants
in Homogeneous Turbulence," Physics of Fluids, Vol. 12, 1969, p. 1999.
30
Toor, H. L., "Turbulent Mixing of Two Species With and Without Chemical
Reactions," Industrial and Engineering Chemistry Fundamentals, Vol. 8, 1969, p.
655.
31
Toor, H. L., "The Non-Premixed Reaction: A + B —> Products," Turbulence
in Mixing Operations, edited by R. S. Brodkey, Academic, New York, 1975.
32
Leonard, A. D., and Hill, J. C, "Direct Simulation of Turbulent Mixing with
Irreversible Chemical Reactions," Third World Congress of Chemical Engineering,
1986.
33
Leonard, A. D., and Hill, J. C., "A Simple Chemical Reaction in Numerically
Simulated Homogeneous Turbulence," AIAA Paper 87-0134, 1987.
34
Leonard, A. D., Hill, J. C., Mahalingham, S., and Ferziger, J. H., "Analysis
of Homogeneous Turbulent Reacting Flows," Proceedings of Summer Program,
Stanford Center for Turbulence Research, Palo Alto, CA, 1988.
35
Eswaran, V., and Pope, S. B., "Direct Numerical Simulations of the Turbulent
Mixing of a Passive Scalar," Physics of Fluids, Vol. 31, No. 3, 1988, p. 506.
36
McMurtry, P. A., and Givi, P., "Direct Numerical Simulations of Mixing and
Reaction in Non-Premixed Homogeneous Turbulent Flows," Combustion and Flame,
Vol. 77, 1989, pp. 171-185.
37
Dopazo, D., and O'Brien, E. E., "Statistical Treatment of Non-Isothermal
Chemical Reactions in Turbulence," Combustion Science and Technology, Vol. 13,
1976, p. 99.
38
Janicka, J., Kolbe, W., and Kollmann, W., "Closure of the Transport Equation
for the Probability Density Function of Turbulent Scalar Fields," Journal of Non-
equilibrium Thermodynamics, Vol. 4, 1979, p. 47.
39
Curl, R. L., "Dispersed Phase Mixing: 1. Theory and Effects in Simple Re-
actors," AIChE Journal, Vol. 9, 1963, p. 175.
40
Pope, S. B., "A Monte-Carlo Method for the PDF Equations of Turbulent
Reactive Flows," Combustion Science and Technology, Vol. 25, 1981, p. 159.
41
Pope, S. B., "An Improved Turbulent Mixing Model," Combustion Science
and Technology, Vol. 28, 1982, p. 131.
42
Warhaft, Z., and Lumley, J. L., "An Experimental Study of the Decay of
Temperature Fluctuations in Grid Generated Turbulence," Journal of Fluid Me-
chanics, Vol. 88, 1978, p. 659.
43
Brown, G. L., and Roshko, A., "On Density Effects and Large Structures in
Turbulent Mixing Layers," Journal of Fluid Mechanics, Vol. 64, 1974, p. 775.
44
Ho, C. M., and Huerre, P., "Perturbed Free Shear Layers," Annual Review
of Fluid Mechanics, Vol. 16, 1984, p. 365.
45
Koochesfahani, M. M., and Dimotakis, P. E., "Mixing and Reaction in a
Turbulent Liquid Mixing Layer," Journal of Fluid Mechanics, Vol. 170, 1986, p. 83.
46
Corcos, G. M., and Sherman, F. S., "Vorticity Concentration and the Dynamics
of Unstable Free Shear Layers," Journal of Fluid Mechanics, Vol. 139, 1984, p. 29.
47
Michalke, A., "On the Inviscid Instability of the Hyperbolic Tangent Velocity
Profile," Journal of Fluid Mechanics, Vol. 19, 1964, p. 543.
48
Winant, C. D., and Browand, F. K., "Vortex Pairing, the Mechanisms of
Turbulent Mixing Layer Growth at Moderate Reynolds Numbers," Journal of Fluid
Mechanics, Vol. 63, 1974, p. 237.
49
Mungal, N. G., "Mixing and Combustion with Low Heat Release in a Turbulent
Shear Layer," Ph.D. Thesis, California Institute of Technology, Pasadena, CA,
1983.
50
Mason, N. B., and Spalding, B. D., "Prediction of Reaction Rates in Turbulent
Premixed Boundary Layer Flows," Proceedings of First Symposium (European)
on Combustion, Academic, New York, 1973, p. 601.
51
Donaldson, C., and Hilst, G. R., "Effects of Inhomogeneous Mixing on At-
mospheric Photochemical Reactions," Environmental Science and Technology, Vol.
8, 1972, p. 812.
52
Metcalfe, R. W., Orszag, S. A., Brachet, M. E., Menon, S., and Riley, J. J.,
"Secondary Instabilities of a Temporally Growing Mixing Layer," Journal of Fluid
Mechanics, Vol. 184, 1987, p. 207.
53
Bernal, L. P., and Roshko, A., "Streamwise Vortex Structures in Plane Mixing
Layers," Journal of Fluid Mechanics, Vol. 170, 1986, p. 499.
54
McMurtry, P. A., Jou, W.-H., Riley, J. J., and Metcalf, R. W., "Direct Nu-
merical Simulations of a Reacting Mixing Layer with Chemical Heat Release,"
AIAA Journal, Vol. 24, 1986, p. 962.
55
McMurtry, P. A., Riley, J. J., and Metcalfe, R. W., "Effects of Heat Release
on the Large-Scale Structure in Turbulent Mixing Layers," Journal of Fluid Me-
chanics, Vol. 199, 1989, p. 297.
56
Rehm, R. G., and Baum, H. R., "The Equations of Motion for Thermally
Driven Buoyant Flows," Journal of Research of the National Bureau of Standards,
Vol. 83, 1978, p. 297.
57
Sivashinsky, G. J., "Hydrodynamic Theory of Flame Propagation in an En-
closed Volume," Acta Astronautica, Vol. 6, 1979, p. 631.
58
Wallace, A. K., "Experimental Investigation on the Effects of Chemical Heat
Release on Shear Layer Growth and Entrainment," Ph.D. Thesis, Univ. of Ade-
laide, Australia, 1981.
59
Hermanson, J. C., "Heat Release Effects in a Turbulent Shear Layer," Ph.D.
Thesis, California Institute of Technology, Pasadena, CA, 1985.
60
Chen, C., "An Investigation of Favre Averaging in Variable Density Turbulent
Flows," M.S. Thesis, Univ. of Washington, Seattle, WA, 1988.
61
McMurtry, P. A., "Direct Numerical Simulations of a Reacting Mixing Layer
with Chemical Heat Release," Ph.D. Thesis, Mechanical Engineering Dept., Univ.
of Washington, Seattle, WA, 1987.
62
Libby, P. A., and Williams, F. A., Topics in Applied Physics: Turbulent Re-
acting Flow, Springer-Verlag, New York, 1980.
63
Givi, P., Jou, W.-H., and Metcalfe, R. W., "Flame Extinction in a Temporally
Developing Mixing Layer," Proceedings of Twenty-First Symposium (International)
on Combustion, The Combustion Institute, Pittsburgh, PA, 1987, p. 1251.
64
Combustion," Progress in Energy Combustion Science, Vol. 10, 1984, p. 319.
65
Peters, N., and Williams, F. A., "Liftoff Characteristics of Turbulent Jet Dif-
fusion Flames," AIAA Journal, Vol. 21, 1983, p. 423.
66
Williams, F. A., "Structure of Flamelets in Turbulent Reacting Flows and
Influences of Combustion on Turbulence Fields," Turbulent Reactive Flows, edited
by R. Borghi and S. N. B. Murthy, Springer-Verlag, London, pp. 195-212.
67
Isizuka, S., and Tsuji, H., "An Experimental Study of Effect of Inert Gasses
on Extinction of Laminar Diffusion Flames," Proceedings of Eighteenth Symposium
p. 695.
68
Puri, I. K., and Seshardi, K., "Extinction of Diffusion Flames Burning Diluted
Methane and Diluted Propane in Diluted Air," Combustion and Flame, Vol. 65,
1986, p. 137.
69
Smooke, M. D., Puri, I. K., and Seshardi, K., "A Comparison Between Nu-
merical Calculations and Experimental Measurements on the Structure of a Coun-
terflow Diffusion Flame Burning Diluted Methane in Diluted Air," Proceedings
of Twenty-First Symposium (International) on Combustion, The Combustion In-
stitute, Pittsburgh, PA, 1987.
70
Dixon-Lewis, G., David, T., Gaskell, P. H., Fukutani, H., Jinno, H., Miller,
J. A., Kee, R. T., Smooke, M. D., Peters, N., Effelsburg, E., Warantz, J., and
Behrendt, F., "Calculation of the Structure and Extinction Limit of a Methane-
Air Counterflow Diffusion Flame in the Forward Stagnation Region of a Porous
Cylinder," Proceedings of Twentieth Symposium (International) on Combustion,
The Combustion Institute, Pittsburgh, PA, 1984, p. 1893.
71
Lowery, P. S., and Reynolds, W. C., "Numerical Simulation of a Spatially
Developing, Forced Plane Mixing Layer," Dept. of Mechanical Engineering, Stan-
ford Univ., Stanford, CA, Rept. TF-26, 1986.
72
Sandham, N. D., and Reynolds, W. C., "Some Inlet Plane Effects on the
Numerically Simulated Spatially Developing Mixing Layer," Turbulent Shear Flows
6, edited by J. C. Andre, J. Cousteix, F. Durst, B. E. Launder, F. W. Schmidt,
and J. H. Whitelaw, Springer-Verlag, New York/Berlin, pp. 441-454.
73
Givi, P., and Jou, W.-H., "Flame Extinction in a Temporally Developing
Mixing Layer," Journal of Nonequilibrium Thermodynamics, Vol. 13, No. 4, 1988,
p. 355.
74
Givi, P., and Jou, W.-H., "Direct Numerical Simulations of a Two-Dimensional
Reacting, Spatially Developing Mixing Layer by a Spectral Element Method,"
Proceedings of Twenty-Second Symposium (International} on Combustion, The
Combustion Institute, Pittsburgh, PA, 1988, pp. 635-643.
75
Tsuji, H., "Counterflow Diffusion Flames," Progress in Energy and Combus-
tion Science, Vol. 8, 1982, p. 93.
76
Givi, P., "Flame Quenching in a Turbulent Mixing Layer: Pseudospectral-
Spectral Element Simulations," Finite Element Analysis in Fluids, edited by T. J.
Chung and G. R. Karr, UAH Press, 1989, p. 1428.
77
Ferziger, J. H., "Large Eddy Numerical Simulations of Turbulent Flows,"
AIAA Journal, Vol. 15, No. 9, 1977, p. 1261.
78
Clark, R. A., Ferziger, J. H., and Reynolds, W. C., "Evaluation of Sub-Grid
Models Using an Accurately Simulated Turbulent Flow," Journal of Fluid Me-
chanics, Vol. 91, No. 1, 1979, p. 1.
Chapter 10
Vortex Simulation of Reacting Shear Flow
Ahmed F. Ghoniem
I. Background
C ONSIDER the class of unsteady, high-Reynolds-number, reacting shear

flows, including shear layers, jets, wakes, and recirculating flows. These
flows are characterized by a rapid growth of successive natural instabilities,
which lead to substantial changes in the geometry of the streamlines. The
simulation of these flows—i.e., the solution of the unsteady, unaveraged
governing equations—requires computational schemes that accommodate
these geometrical changes, and in which numerical diffusion is maintained
below physical diffusion without sacrificing numerical stability. One way
to achieve this is to utilize schemes based on the Lagrangian formulation
of the conservation equations.
In Lagrangian methods, grid points, where the flow variables are com-
puted, are transported along particle trajectories. These methods offer
natural ways to 1) accommodate the severe distortion experienced by high-
Reynolds-number flows as they evolve through different states; 2) reduce
the artificial diffusion encountered in numerical schemes in which spatial
derivatives are discretized; and 3) account, in a balanced way, for physical
diffusion in regions where the two modes of transport—convection and
diffusion—are of the same order of magnitude. A brief discussion of these
issues is presented.
A. Numerical Issues
A challenge often encountered in numerical simulation of reacting shear
flow is how to capture the severe distortion of the flow map that results
from the nonlinear growth of natural flow instabilities. The saturation of
these instabilities in the primary flow, which can normally be characterized
by almost parallel streamlines, results in a secondary flow with strongly
curved streamlines. The latter often possesses secondary instabilities, which
evolve into tertiary flows. To capture these changes accurately using a
numerical simulation, a very large number of fixed grid points, or a moving
Copyright © 1990 by Ahmed F. Ghoniem. Published by the American Institute of Aero-

nautics and Astronautics with permission.
305
306 VORTEX SIMULATION OF REACTING SHEAR FLOW
grid in which mesh points follow the distortion of the flowfield, is needed.
The latter class belongs to Lagrangian schemes.
Convection is the dominant mechanism of transport at high Reynolds
numbers. Convection in the cross-stream direction, perpendicular to the
streamwise direction, is increased substantially by the growth of natural
flow instabilities and their saturation into fully developed flows. Mean-
while, diffusive transport cannot be ignored. Molecular diffusion is respon-
sible for the transport of vorticity from solid walls into the interior in
boundary layers, the mixing of species across material surfaces, and the
transport of heat and species across laminar flames. Accurate simulation
of diffusion is crucial particularly in reacting flows where mixing, most
often, determines the burning rate. Conversely, excessive numerical dif-
fusion in an algorithm can artificially stabilize the flow. Numerical diffusion
acts as a fictitious source of molecular diffusivity that reduces the effective
Reynolds number of the computed flow. In combustion calculations, exces-
sive mixing due to numerical diffusion tends to increase the burning rate.
To capture the growth of flow instabilities while limiting diffusion to prac-
tically interesting values, one needs to simulate flows at Reynolds number
0(103-105).
B. Classification
One class of Lagrangian methods that has been used successfully in
gas-dynamics utilizes grids to discretize flow derivatives, e.g., Lagrangian
finite-difference methods. In this method, grid points are transported along
particle trajectories. Shear flow can lead to strong distortion of the grid
and a concomitant loss of discretization accuracy. To maintain accuracy a
long time after the action of strong shear, mesh regularization and remesh-
ing become important. 1 " 3
Conversely, grid-free, Lagrangian field methods of the type described
in this chapter do not use approximations to spatial derivatives on a non-
uniform mesh. Instead, computational elements are used to transport finite
values of the spatial gradients of the variables, specifically vorticity and
scalar gradients (scalars are temperature and species concentrations). Prim-
itive variables, such as velocity and scalar concentrations, are obtained by
integration over the strength of the transport elements. These methods are
labeled "field methods" since each element induces a field of both its
strength—e.g., the vorticity of an element extends over a finite area—
and of the primitive variable it is transporting—e.g., the velocity induced
by a vortex element extends to an infinitely large distance.
The goals of this chapter are to review Lagrangian field methods that
have been developed for the simulation of compressible reacting flow and
to describe the mechanisms of shear flow-combustion interaction that have
been revealed using these methods. These goals are achieved simultane-
ously by introducing progressively more complicated models, describing
the necessary numerical algorithms, and presenting their results in a form
most relevant to the study of flow-combustion interaction. The models are
expanded gradually until they include many of the physically interesting
processes in combustion.
A. F. GHONIEM 307
C. Organization
Vortex methods, a particular class of grid-free Lagrangian field methods,
have been used to obtain solutions of the momentum equation. These
methods are based on the discretization of the vorticity among elements
of finite area and the transport of these elements along particle trajectories.
The fact that vorticity is conserved along the particle trajectory in a two-
dimensional, uniform-density flow makes these methods particularly sim-
ple. However, in Sec. II, we show that maintaining accuracy requires the
application of elaborate vorticity-updating schemes as vortex elements are
moved along particle trajectories when shear, or a strong strain field is
present. The extension and application of vortex methods to three-dimen-
sional flows, where the conservation of vorticity along particle trajectories
is not satisfied, also require the careful application of schemes to implement
the effect of vortex stretch on the strength of the elements, as discussed
in Sec. III. Solutions using the two- and three-dimensional methods are
discussed to illustrate some of the most common instabilities encountered
in nonreacting and reacting shear flows and to reveal the mechanisms by
which the maturation of these instabilities enhance mixing and, hence,
burning in a reacting flow.
The application of vortex methods to nonuniform-density reacting and
compressible flows requires compatible Lagrangian, grid-free field schemes
to compute the transport of scalars. For this purpose, we developed the
transport element method to solve the convective-diffusive scalar transport
equation. Even in a nonuniform-density incompressible flow, the transport
of a scalar is required for the computations of the density field that affects
the flow dynamically via the generation of baroclinic vorticity. The for-
mulation of this method is described in Sec. IV, and the results of its
application to compute scalar mixing in a shear layer are reviewed. The
transport element method is then combined with the vortex method to
solve the problem of nonuniform-density shear flow in Sec. V. Baroclinic
vorticity generation is one of the important mechanisms by which com-
bustion affects the fluid dynamics, and, thus, the computational results on
a nonuniform-density shear layer are reviewed in some detail in this section.
Another important mechanism of flow-combustion interaction—namely,
reaction extinction due to the formation of localized regions of strong
strains as instabilities grow into their nonlinear range—is revealed by the
results of incompressible reacting flow models. These results are discussed
in Sec. VI, after reviewing the formulation of the low-Mach-number com-
bustion model and the extension of the transport element method to accom-
modate the reaction terms in the conservation equations. Three-dimen-
sional reacting shear-layer simulations are presented to illustrate the
complexity of the mixing pattern encountered in these flows and how they
affect the structure of the burning zone. Compressible reacting flows exhibit
another mechanism of combustion-flow interaction, namely, the volumetric
expansion associated with energy release within the reaction zone. Results
of computations of combustion in heterogeneous and homogeneous com-
pressible shear layers, using the transport element method, are reviewed
to illustrate the origin and role of this mechanism.
The bulk of this chapter is devoted to a brief description of Lagrangian

methods for the simulation of reacting flows in free shear layers, and a
review of the most important mechanisms of flow-combustion interactions
revealed by the application of these methods. Section VII briefly describes
some extension of vortex methods to confined shear flow. We close the
chapter with some thoughts on the needs for future development in the
methodology and some areas of application that have not been explored.
II. Vortex Element Method in Two Dimensions

The vorticity transport equation of a two-dimensional, incompressible,
inviscid flow is
— + u • Vo> - 0 (1)
dt ^ '
where Vxu = co and V • u = 0. In Eq. (1), u = (u,v) is the velocity, co
the vorticity, x = (x,y), t the time, and V - (d/dx, d/dy). If \(X, 0 describes
a particle path, where X is the Lagrangian coordinate of \ so that \(X,0)
= Jf,thenEq. (1) states that <i>[x(Xj), t\ = o>(^,0); i.e., vorticity is constant
along the particle trajectory. Thus, if the vorticity field at time t = 0 is
divided among elements that move along particle trajectories, the strength,
i.e., the vorticity of each element, will remain the same. Moreover, it can
be shown that u(x, t) = f K(x - x')<»(x')dx' , K(x) = -^r\-y,x) and
r2 = x2 + y 2 , which is the Biot-Savart law. This Lagrangian formulation
of vorticity transport is the basis of vortex methods. 4 5
A. Numerical Scheme
In vortex methods, the vorticity field is discretized into a number of
vortex elements of finite and overlapping cores:
where 00, is the vorticity of an element, TV the total number of vortex

elements, h the average distance between the centers of neighboring ele-
ments in two principal directions (h2 = hxhv), 8 the core radius of a vortex
element, and /8 = l/82/(r/8) the core function describing the distribution
of vorticity associated with an element.
The velocity field of a distribution of vortex elements is obtained by
substituting Eq. (2) into the Biot-Savart law and performing the integra-
tion. The result is
«(*»') = 4 <*th2K*\* ~ x/(*/,0] (3)

where
H^/
),?] (4)
A. F. GHONIEM 309
while
and
K(r) = 2'
Equation (2) is equivalent to expanding a function co(jc,r) in terms of a

number, TV, of kernel functions, /8, located at Xt and with weights u^/i2.
The accuracy of the discretization depends on the choice of /, the initial
distribution of the particles, the evaluation of the values of CD,-, / = 1,2,
..., TV, and the ratio of §/h.6J The selection of a core function to achieve
a given accuracy was discussed extensively in a number of theoretical
analyses.8^11 For the evaluation of the initial values of co,, we found that
collocation on a uniform grid provides the best long-time accuracy. (Col-
location on a nonuniform grid may be a better choice when the initial
vorticity field is highly nonuniform.) We also found, using extensive numer-
ical experimentation, that accurate discretization and long-time accuracy
of the computed flowfield require that 8 = 1.1-1.5/z. This choice of S//z
allows for strong overlap between the fields of the vortex elements. Thus,
the local value of vorticity is determined by the contributions of many
surrounding elements.
A strong strain, associated with the growth of perturbations into the
nonlinear stages of the underlying instability, increases the distance between
neighboring elements, 8x, beyond the desired maximum value of h. Thus,
the accuracy of the spatial discretization, which is governed by 8//*, is
affected negatively, and unorganized random motion on the scale of h is
encountered. This deterioration of accuracy is connected with the failure
of vortex elements of fixed cores to capture accurately the distortion of
the vorticity field locally. To circumvent this problem, more elements are
introduced in areas where 8x > P/z, where p ~ 1.5. The circulation of the
original elements is redistributed locally among the newly introduced ele-
ments, and the local value of vorticity is kept constant.
In the redistribution algorithm, conservation of vorticity is satisfied by
dividing h2 and 82 such that 8//z remains constant; i.e., the core radius of
an element is decreased as the element is exposed to strong positive strain.
An alternative way to capture the effect of strain on the geometry of the
computational elements is to replace a circular element by an elliptical
element, while preserving its area, with its major axis aligned with the
direction of maximum positive strain. If this was done, a circular vortex
element eventually would become a vortex sheet with a velocity jump across
its length equal to the local value of the vorticity 00,. For computational
convenience, however, we replace a strained circular element by several
circular elements, aligned along the major axis of strain, but with smaller
core radii than the original element. 6
B. Primary Shear Flow Instabilities
We used shear layers as test cases for the validation of the numerical
methods and as generic problems for the study of flow-combustion inter-
actions. We start with a two-dimensional, incompressible, nonreacting flow,

and build up to three-dimensional flow, variable-density flow, and com-
pressible reacting flow.
The growth of a sinusoidal perturbation on an infinite shear layer is
shown in Fig. la in terms of all the vortex elements used in the simulation
and their velocity vectors. Figure la illustrates the adaptivity of the scheme;
more elements are introduced to capture the straining layers. The growth
of small perturbations in a spatially developing shear layer in which the
flow was assumed to be semi-infinite is depicted in Fig. 2 in terms of the
vortex element distribution. In both the temporal and spatial shear layers,
numerical results for the growth rate of small perturbations as a function
of the perturbation wavelength were found to agree with the results of the
linear stability theory. The nonlinear regime of these Kelvin-Helmholtz
instabilities is characterized by the formation of large-scale structures due
6.6 6.6
X-DISTRNCE X-DISTflNCE
t - 5.5 t - 11.0
t.n o"
Q
6.6 13.2 6.6

X-DISTflNCE X-DISTflNCE
t - 16.5 t - 22.0
Fig. 1 (continued on next page.)

A. F. GHONIEM 311
b)
t - 5.5 t - 11.0
16.5 22.0
Fig. 1 Rollup of an incompressible, uniform-density, nonreacting, temporally growing

shear layer in which the top and bottom streams move at the same speed but in
opposite directions and the initial vorticity distribution is Gaussian. The initial
amplitude of the sine-wave perturbation is 1% of its wavelength, which is taken as
the most unstable wavelength for this flow. The results are shown in terms of a)
vortex elements and their velocity vectors, and b) isoscalar lines (s = 0 in the top
stream and s = I in the bottom stream, with an error function distribution in
between).
to the rollup of the vorticity layer, as shown in Figs. 1 and 2. Results in

this regime are checked for convergence and against experimental data.
Convergence is tested by establishing the grid independence of the solu-
tion, and determining whether the solution satisfies certain differential and
integral constraints, such as the conservation of vorticity along a particle
path, dco/dr = 0.
Figures 1 and 2 show that, beyond the linear range of the instability,
where the growth is exponential, the mean growth rate of the shear layer
reaches a constant value, in agreement with experimental data. Within this
range, the computed results were used to evaluate the flow statistics. Figure
3 shows the close agreement obtained between the computed results and
the experimental measurements. We note that mean velocity profiles reach
self-similarity earlier than mean fluctuations. The response of the layer to
time-dependent boundary conditions was analyzed by modulating the inlet
flow at frequencies different from the natural shedding frequency. Samples
of the results, in which the flow is forced at the fundamental alone, and
i«.TM^«rM^nrrJinS-igr iTjft
b)
c)
Fig. 2 Development of an incompressible, nonreacting, spatially developing shear

layer between a fast top stream and a slow bottom stream. The velocity ratio across
the layer is 2, and the vorticity distribution at the left-hand side of the domain,
where the splitter plate ends, is Gaussian. The figures show all of the vortex elements
used in the computation and their velocity vectors (measured with respect to the
mean velocity of the flow). The confining walls are slip boundaries. Three cases are
shown: a) the layer is unforced; b) the layer is forced at the most unstable wavelength
and the amplitude of the perturbation is 1 % of the wavelength; c) the layer is forced
at the most unstable wavelength and its first subharmonic, and the two amplitudes
are equal to 1% of the fundamental wavelength.
at the fundamental and subharmonic simultaneously, are shown in Figs.

2b and 2c, respectively. The shear-layer growth, and the accompanying
rate of mixing, are enhanced by inducing an earlier shedding, and pairing
in the second case, through the application of external forcing.7 Multiple
pairing can be achieved by lower frequency forcing, as discussed in Ref. 13.
In a two-dimensional flow, the source of fluctuation is the formation,
growth, and pairing of the large-scale vortex structures due to the natural
flow instability. Two instabilities are observed in Fig. 2: rollup instability,
which leads to the formation of large eddies, and pairing instability, which
is responsible for the amalgamation of these eddies downstream. We call
these the primary instabilities since they grow in the primary, streamwise
direction. Forcing, which can be used either to promote or suppress these
instabilities, was shown to have a direct impact on the values and signs of
these fluctuations, suggesting that, by employing carefully designed forcing
functions, one can control the interactions between the mean flow and the
shear-layer flow.13 Computations of the initial stages of a two-dimensional
turbulent jet were presented in Ref. 14.
A. F. GHONIEM 313
-0.08 - 0 . 0 4 0.00 0 . 0 4
CY-Y0)/(X-X0)
b)
0.30
0.25
- 0.20
Z3
a\ 0.15
w
c
^0.10
0.05
0.00
-0.08 -0.04
(Y-Y0)/(X-X0)
Fig. 3 a) Time-average streamwise velocity profiles for the case shown in Fig. 2a,
computed at sections x = 3, 3.5, 4, 4.5, and 5. The variable x is measured from
the left-hand side of the domain, x = 0, and is normalized with respect to the channel
height, while y0 = 0. Uh and Ul are the high- and low-speed stream velocities.
Computational results are shown as solid lines; experimental measurements of Masu-
tani and Bowman29 are shown as open symbols, b) Time-average streamwise velocity
fluctuations for the same case as in part a, computed at the same sections and plotted
against experimental measurements from the same reference.
III. Vortex Element Method in Three Dimensions

The vorticity transport equation in an incompressible, three-dimen-
sional, inviscid flow is
dco
Vco = co • Vw (5)
dt
where GO = V x u and V = (d/djt, d/dy, d/dz). In this case, u = (u,v,w) and

x = (x,y,z). The Lagrangian form of Eq. (5) is dco[\(AT,r), t]/dt = <&[\(X,t),
t] - Vw, where Vu is the strain tensor duJdXj. An equivalent expression that
can be used to determine the vorticity directly is the Helmholtz theorem:

<o[x(A:,0, t] = V\(Xj) • to(^,0), where V\ is the Jacobian of the flow
map, dXi/dXj. Moreover, it can be shown that u(x,i) - / K(x - x') x
to(jc')dji:', where K(x) - -fir x/r3, which is the Biot-Savart law in three
dimensions. This Lagrangian formulation is the basis for the construction
of three-dimensional vortex methods.15-16
A. Numerical Scheme
Vorticity is discretized among volume elements, of side h, initially cen-
tered around Xh by collocation. The vortex elements are then moved along
particle trajectories, \t(Xhf), while their vorticity is changed according to
the right-hand side of Eq. (5). Thus,
where, in this case, /8(jc) = 1/83 /(r/8) and the rest of the parameters are
defined as before. Note that here one defines strongly overlapping vortex
balls of diameter 8, and that the core function is spherically symmetric,
while the vorticity vector associated with an element is co,.
The total vorticity vector of an element, co/A3, is expressed more naturally
in terms of F, and 8/,, where F, = co,//2 is the circulation of the element
that remains constant along a particle path in an inviscid flow (Kelvin's
theorem), and 8/, is the length of the material element along the vortex
line, 8/, = Ayco/o),-, that changes as the material lines stretch (Helmholtz
theorem). The formulation in terms of (87, F) offers a natural regridding
method, which is used when 8/ > (S/z due to stretch. In this case, a vortex
element is divided into two elements along the vector 8/, while preserving
the value of F. The conditions necessary for the accurate discretization of
the initial vorticity in two dimensions are valid in three dimensions, e.g.,
8 > h. This condition must be satisfied at all times.
The velocity field induced by the discrete vorticity distribution is obtained
by substituting Eq. (6) into the Biot-Savart law and integrating. The results
can be written as
i= 1
8[* - x/(*/, 0] (7)
and
^ = u[Xi(Xn /), r] (8)
s/,(0 = 3[x,+i(*,+1,0 - x,--i(*/-i,0] (9)

where
A. F. GHONIEM 315
and
K(r) - 4ir
To reduce the computations, we utilize the fact that vortex lines are also
material lines in an inviscid flow. Application of Eq. (9) requires main-
taining data on the immediate neighbors in the direction of vorticity. Thus,
one-dimensional Lagrangian grids are employed to describe individual vor-
tex lines as arrays of vortex elements arranged along the vortex line. The
condition that a vorticity field in a three-dimensional free space be sole-
noidal, V • o> = 0, is implicitly satisfied in Eq. (9). Equations (7-9) describe
the vortex filament method. 17
B. Propagation and Instability of Vortex Rings

Before describing their properties, we mention that vortex rings, in addi-
tion to providing a test problem for the numerical scheme, play an impor-
tant role in combustion theory and practice. We refer the reader to the
experimental literature for studies on flame-vortex-ring interactions 18 and
the large-scale structure of jet diffusion flames.19
It has been shown analytically that the self-induced velocity of a thin
vortex ring, a « R, where a is the core radius and R the ring radius, is
a function of <j/R and the vorticity distribution within the core, fl(r/cr). 20
Numerically, using 8 = <r, the dependence of the self-induced velocity on
v/R was computed accurately when overlap was maintained between neigh-
boring vortex elements along the ring axis.21 The long-wave azimuthal
instability of a thin vortex ring, with a wavelength X » <j, was observed
when the ring was perturbed along its axis with a number of waves, n =
2irR/X. The computed unstable wave number, k* = 2^/n*, and growth
rate of these waves within the linear range agreed with the prediction of
the corresponding linear theory. 22 The growing standing waves at A:*, con-
trary to the spinning stable waves at all other wave numbers, expend the
flow energy in stretching the waves in the direction opposite to the ring
self-induced velocity.
In the nonlinear range, the growing waves form almost closed loops of
vorticity behind the original ring, as shown in Fig. 4. These loops are
connected to the original ring via very narrow necks that can be pinched
off by the action of viscosity. Each loop is formed of two vortex rings of
opposite signs of vorticity separated by a very small distance. The sepa-
ration of these loops from the parent ring may lead to the formation of
offspring vortex rings with a smaller diameter than that of the original ring.
This would result in an interesting cascading to smaller scales, a faster
decay of the original ring, and the reduction of the circulation of the parent
ring.23
The study was extended to investigate the growth of short-wave insta-
bilities, X ~ cr, within the core of the ring. In this case, one must allow
the core of the ring to deform under the action of the growing perturbation.
Analysis of variations within the core requires adequate resolution of its
vorticity field by utilizing vortex elements with a core radius 8 < a; i.e.,
10 15 20 0 5 10 15 20
X
0 5 10 15 20 0 5 10 15 20
10 15 20 5 10 15 20
X
15 20 IS 20
Fig. 4 Deformation of a thin vortex ring, represented computationally by a single

vortex filament, when excited at an unstable wave number (number of waves around
the ring) of 6. The radius of the ring core is 0.25 of the radius of the ring. The plots
are obtained by projecting the ring on two planes normal (on the left-hand side)
and parallel (on the right-hand side) to the direction of propagation of the ring
under its own self-induced velocity.
316
A. F. GHONIEM 317
several elements must be used to accurately describe the vorticity field

within the core. The computed value k*, shown in Fig. 5, is in close
agreement with the prediction of the linear theory for short-wavelength
instability in deformable vortex rings. The value of k* depends on the
vorticity distribution within the ring core, H(r/a). Figure 5 shows that the
value of A;*, predicted from the short-wave analysis, is closer to the exper-
imental data than that predicted by the previous long wavelength.
Spectral analysis of the field of an unstable ring shows that, as the
fundamental instability reaches saturation, its harmonic becomes unstable
and starts to grow. The mechanism of excitation of this frequency, displayed
in Fig. 6, is associated with the formation of hairpin vortices at half of the
wavelength of the original perturbation. Examining the vorticity field after
the saturation of the first instability reveals the presence of a strong stream-
wise vorticity component with alternating signs as one moves along the
axis of the original ring.24 This component is the result of the tilting of the
original vortex lines into the streamwise direction. The tilting occurs as the
azimuthal instability grows. Within each azimuthal wave, two vortices of
opposite signs are formed.
The scheme was also used to study the three-dimensional instability of
the initial stages of an axisymmetric jet. 17 It was found that the jet rolls
up into vortex rings that experience a similar instability to that of an isolated
vortex ring.
ID
14 r A -
12 X X- D o
+ D o
10 44 DO o
*t- Q A 0
c 8 x 4- a o o
X -WD 0 A o
6 X + DO A 0
+ a o
4- o
o
2 0
0 —————————————————————————————————————
1.5 2.0 2.5 3.0 3.5 4.0 4
Fig. 5 Wave number of the most unstable mode, /i*, of a vortex ring plotted against
its normalized self-induced velocity, V = 4Tr/?V/F, where R and F are the radius
and circulation of the ring, respectively. The figure shows a comparison between
the experimental results, x ; the analytical results of the long-wavelength instability,
o; the numerical results of the long-wavelength instability, A; the analytical results
of the short-wavelength instability, a for constant vorticity distribution within the
core and + for quadratic vorticity distribution; and the computed results of the
short-wavelength instability, 0 for coarse numerical discretization and * for fine
discretization.
Fig. 6 Perspective views, taken from the point of view of an observer standing
ahead of the ring and looking at ah angle 60 deg with respect to the direction of
propagation, of a vortex ring whose core radius is 0.275 of the ring radius. The ring
is initially excited using 12 waves around. All of the filaments used to discretize the
ring are shown. The ring is propagating upward.
C. Secondary Instabilities of Shear Flow

Beyond the primary instability, shear flows develop secondary instabil-
ities that lead to the generation of streamwise vorticity (the vorticity vector
points in the streamwise direction); an important mechanism of mixing
enhancement.24"26 To simulate this process, the shear layer is initially
perturbed in the streamwise and span wise directions, and the vortex scheme
is modified to accommodate the strong strain field that develops in the
plane normal to the initial vorticity. This is accomplished by redistributing
the vorticity field among a larger number of elements arranged in the
direction of maximum strain.
In the transport element scheme, each vortex element is defined by its
circulation and two material vectors: one in the direction of vorticity and
the other in the direction of maximum strain. This is equivalent to using
two-dimensional Lagrangian grids to describe planes of constant circula-
tion. The stretch of the sides of the grid can be used to update the vorticity
associated with each element in the plane. Alternatively, a grid-free stretch-
ing scheme can be constructed based on Eq. (5). In this scheme, the term
co • Vu is computed for each element by analytically differentiating the
velocity expression in Eq. (7). The computational effort is almost the same
in both schemes, however, there is more bookkeeping in the first scheme.26
The evolution of a temporal shear layer is displayed in Figs. 7-9. Figure
7 depicts the distortion of the material surface initially aligned along the
midsection of the layer, where u — 0, in terms of the two-dimensional grid
used to discretize its vorticity (this is one of several such surfaces used to
A. F. GHONIEM 319
t = 8.0 t = 16.0
12.0 18.0
Fig. 7 Three-dimensional perspective views of the isoscalar surface s = 0, initially

coinciding with an x-y plane located in the middle of the domain shown in the figure.
The shear layer is periodic in two directions. At time t = 0, the layer is perturbed
in the streamwise direction, Jt, at the most unstable wavelength, and in the spanwise
direction, y, at one-half of this wavelength. The amplitudes of the perturbations are
equal to 2% of the streamwise wavelength. The initial vorticity follows a Gaussian
distribution in the spanwise direction, z, as in the two-dimensional case.
discretize the vorticity). Figure 8 shows the distribution of the stream wise
vorticity contours on a y-z streamwise plane (a streamwise plane is normal
to the streamwise direction) that cuts through the core of the spanwise
structure at the middle of the domain. Figure 9 depicts the streamwise
vorticity on a y-z plane that cuts through the braids of the spanwise struc-
tures (the principal axis of this structure coincides with the spanwise direc-
tion). The distribution of the streamwise vorticity indicates that, during
the growth of the streamwise perturbation, the growth of the spanwise
perturbation is suppressed, and the flow remains almost two dimensional.
The growth of the streamwise instability (not shown) matches that of a
a) b)
t = 8.0 t = 8.0
t - 16.0 t = Ih.O
Fig. 8 a) Streamwise vorticity; b) scalar distribution in the three-dimensional sim-

ulation of the periodic shear layer shown in Fig. 7. Contours are shown in a y-z
plane, which cuts through the core of the spanwise eddy at the middle of the domain
of Fig. 7. Positive vorticity is indicated by broken lines, while negative vorticity is
depicted by continuous lines. For the scalar, s > 0 is shown in continuous lines and
s < 0 is shown in broken lines, —0.5 < s < 0.5.
A. F. GHONIEM 321
two-dimensional flow until the streamwise instability saturates, t = 8. The

saturation is accompanied by the formation of a spanwise, cylindrically
shaped vortex core whose axis is perpendicular to the mean flow direction.
The formation of this core is accompanied by a secondary flow whose
streamlines are almost circular.
Beyond the saturation of the streamwise instability, t > 8, the growth
rate of the spanwise instability increases and the vortex core starts to
deform. This deformation, and the concomitant tilting of the vorticity
vectors from the spanwise direction into the streamwise direction, lead to
the establishment of a streamwise vorticity component with an alternating
b)
t = 8.0
t = I6.0 t -•-- 16.0
Fig. 9 a) Streamwise vorticity; b) scalar distribution in the three-dimensional sim-

ulation of a periodic shear layer shown in Fig. 7. Contours are shown in the y-z
plane, which cuts through the braids at a distance 15% of the wavelength into the
domain from the left-hand side. See Fig. 8 for convention.
sign along the axis of the spanwise cylindrical structures. The deformation
of the spanwise cylindrical cores is known as the "translative instability."
The streamwise vorticity associated with this instability is shown in Fig. 8.
The fact that the cores are covered from both sides by two rows of stream-
wise vortex rods is explained next.
The other source of streamwise vorticity is the deformation of the vortex
lines of the braids in the cross-stream direction, as shown in Fig. 7. The
vorticity component in the cross-stream direction is transformed into a
streamwise component by the action of the velocity gradient, du/dz, where
u is the velocity component in the x direction. Along the spanwise direction,
streamwise vorticity changes its sign every half-wavelength. This configu-
ration is unstable, and each half-wavelength rolls up to form a vortex rod
aligned with the local direction of the braids, as shown in Fig. 9. The
vorticity within these rods is amplified constantly as the strain field in the
streamwise direction, generated by the large spanwise cores, strains the
flow along the direction of the braids. With elongation, these rods wrap
around the spanwise cores and produce the distribution shown in Fig. 8.
The deformation of the flowfield due to the evolution of the two- and
three-dimensional instabilities leads to a substantial enhancement of mixing
within the large structures by generating strong entrainment currents of
fluids from both sides of the shear layer toward the vortex center. The
mixing enhancement and its effect on the burning rate are discussed in
detail in the next sections.
IV. Transport Element Method

Given that s is a passive, nondiffusive scalar, the conservation equations
for s, and g = Vs, in an incompressible flow are
I =°
and
^ = -g • Vu - g x o> (11)
Thus, s remains constant along a particle path, while g changes due to the
straining and rotation of the material lines by the local strain field and
vorticity. If the material is exposed to a strong strain in the direction normal
to the gradient, the value of g must increase by the same amount as the
stretch in the material element. This can be seen by deriving an explicit
equation that relates the changes in g = \g\ to the variation of material
elements, or the distortion of the flow map. This is done by expanding
Eq. (11) in terms of gn and implementing kinematical relations that de-
scribe the variations of n = gig, where n is the unit vector normal to an
isoscalar line, i.e., a line along which 5- is constant. After some lengthy
manipulations, we get
n = -g{n • Vii + in x co - [/ • (n • Vus)]l} (12)

A. F. GHONIEM 323
where Vus is the symmetric part of the strain tensor VM, and / is the unit
vector normal to n. Moreover, g — (ds/dn)n ~ (§s/§n)n, where bs is the
variation of s across a small material line 8/t. In two dimensions and for
an incompressible flow, the variation of a material vector element 8/ can
be shown to be governed by an equation similar to Eq. (12).
A. Numerical Scheme
From the preceding discussion on the relationship between scalar gra-
dients and the deformation of the flow map, it follows that g/8/ = const
along a particle path, and that the scalar gradient can be computed from
the following relations7:
- X,(^0] (13)
/=!
where
gi(f) = [Ss&M/h^nM (14)
and \t(Xhf) is, as before, a particle path. Equation (13) is based on the
expansion of g in terms of the core function /8, similar to Eq. (2). Since
an isoscalar line is a material line in a nondiffusive field, 8/, can be updated
as 8/f-(0 ™ (x/ + i ~~ X/-i)/2, while n( - /, = 0. Thus, it suffices to move the
centers of the transport elements while remembering the near neighbors
at t = 0. As in the vortex method, when the distance between neighboring
elements in the direction of maximum strain exceeds a certain maximum,
an element is inserted between two neighboring elements. The total of S//s
for the two original elements is distributed among the three elements, while
h2 and 82 are adjusted so that the total material area is conserved, keeping
8s,- the same.
For a variable-density flow, the preceding analysis is modified to reflect
the fact that, in this case, an equation similar to Eq. (12) can be derived
with g replaced by p8/, and the expression of g, in Eq. (13) changes to 12 - 27
gi(t) = {[5*,.8/,(OP,(0]/P,(0)/*2K(0 (15)
The value of p is computed using the relation pT = const in the low-Mach-

number approximation (see Sec. V). Given the location and strength of
the transport elements, the scalar concentrations are computed by direct
integration over the fields of the transport elements
N
- x/(*/,0] (16)
where, in two dimensions, VG8(jc) = (jc, y)!2^r2 K(r/8), and in three dimen-
sions, VG8(;t) = (jc, y, z)/4irr 3 K(r/8). This formulation is fully compatible
with the vortex method since all of the information needed to compute the
scalar transport is already part of the vortex computations, including all
of the expressions of the Green functions. For extended derivations, see
Refs. 24 and 27.
The effect of molecular diffusion in the vortex and transport element

methods can be modeled by expanding the cores of the elements according
to the relation b2(t + Af) = S 2 (f) + 2aAr, where A/ is the time step and
a the molecular diffusivity. 6 This relation is obtained by direct substitution
of Eq. (11) into the diffusion equation. A limit should be imposed on the
maximum allowable value of 8 to maintain the spatial accuracy of the
calculations. Beyond 8 max , an element should be subdivided into a number
of smaller elements. Another scheme for implementing the effect of dif-
fusion without expanding the cores was proposed in Ref. 28. This scheme
is used in conjunction with the three-dimensional transport element method
in Sec. IV.C.
B. Scalar Mixing in Shear Layer

The transport element method was used to study the transport of species
in a two-dimensional, heterogeneous shear layer. 12 Figure Ib shows the
isoscalar lines of the shear layer of Fig. la when the diffusivity is zero.
Numerically, we found that, provided the field initially is discretized accu-
rately, the condition ds/dt = 0 is satisfied if the core radii of the elements
are allowed to decrease at the rate described in Sec. II, i.e., if //2/82 remains
constant as the elements are deformed. This also guarantees that the scheme
can capture, without introducing numerical diffusion, the large scalar gra-
dients that arise from the strong deformations in the flow that accompany
the evolution of the instability.
Figure 10 shows a comparison between the computed and measured
values of the mean concentration and the root-mean-square (rms) of the
concentration fluctuations in the two-dimensional spatially developing mix-
ing layer of Fig. 2a. 12 - 29 The computational results were obtained for a
range of Peclet numbers between 103 and 105, so that the dominant trans-
port process was convection and not diffusion. However, the effect of
species diffusion was incorporated to study mixing. The effect of diffusion
on the mean scalar distribution is very small since the overall concentration
field is established by the convective field, also called the "entrainment
currents." As a result of vorticity-layer rollup, fluid from both sides of the
shear layer is engulfed into the large structure and, on the average, mean
values, between the two extremes, can be encountered.
The rms of the concentration fluctuations exhibits strong dependence
on the Peclet numbers. Its maximum value, 0.5, can be achieved only at
very high Peclet numbers, where strong unmixedness is present inside the
eddy cores. As the effect of molecular diffusion increases, it homogenizes
the cores, and the fluctuations drop below 0.5. In this case, the profiles
show a mixed region of constant value of fluctuation. For all Peclet number
values, the fluctuations never reached zero inside the cores, indicating that
the fluid does not reach complete homogeneity. Another interesting feature
of these mixing flows is the presence of more high-speed fluid than low-
speed fluid inside the cores.12 This mixing asymmetry, which is a direct
consequence of the unequal velocities across the interface of the layer, will
be shown to play an important role in the distribution of products.
A. F. GHONIEM 325
C. Entrainment Enhancement due to Three-Dimensional Instabilities

The three-dimensional transport element method is based on the rela-
tionship between the distortion of the flow map and the scalar gradients,
as in the two-dimensional method. In three dimensions, the distortion of
the flow is represented by the changes in the magnitude and direction of
a material area element initially aligned with an isoscalar surface 3A. This
relationship was derived in Refs. 24 and 30 for a variable-density flow:
(17)
Equation (17) indicates that g/(p§A) = const along a particle path. To

implement this relationship in the calculations of the scalar gradients, one
0.00 0.05 0. 10 0. L5
(Y-Y0)/(X-X0)
Fig. 10 a) Time-average scalar concentration for the flow in Fig. 2a, computed at
x = 5 for a range of Peclet number Pe. At x = 0, the concentration is s — 0 in the
upper stream and s = 1 in the lower stream. Solid lines show the results of the
computations; open symbols depict the experimental measurements of Masutani and
Bowman.29 b) Time-average concentration fluctuations for the same flow as in part a.
3.5
3.0
UJ
2 2.5
CO
§ 2.0
Ld
1.5
10 12 14 16 18
TIME
Fig. 11 Eddy size in the two-dimensional (v) and three-dimensional (x) computa-
tions of Figs. 1 and 7, respectively. The eddy size is defined by the cross-stream
distance between the contours 5 = 0.03 and 0.97.
must keep track of the area of the elements associated with computational
points moving along particle trajectories. Similar schemes are used in the
vortex element method to monitor the effect of the strain field on the
evolution of the vorticity field. Thus, the computation of the scalar trans-
port is a natural extension of the vortex element method, since it does not
require additional computational effort.
The entrainment due to the development of three-dimensional instabil-
ities in shear flow is depicted in Figs. 8 and 9, where we show the scalar
contours on the same planes where the streamwise vorticity contours are
displayed. The extra entrainment, over what is observed in two-dimen-
sional calculations, is induced by the action of the spanwise deformation
of the vortex core and by development of the streamwise vortex rods within
the braids, as shown in Figs. 8 and 9, respectively. The total entrained
fluid within the large eddies is measured by the size of the eddy in the
cross-stream direction, as shown in Fig. 11 for both the two- and three-
dimensional calculations. 24 - 26 The effect of this extra entrainment on the
rate of burning will be discussed in Sec. VLB.
V. Vortex Methods for Variable-Density Flows

In combustion phenomena of practical interest, the flow is density strat-
ified, i.e., finite- and large-density gradients exist between material layers,
due to heat release and the variation in molecular weight of reacting species.
The effect of density gradients is to introduce a baroclinic source term in
the vorticity transport equation. 33 In this section, we show how this term
is included in the simulation and describe its effect on shear layers. Con-
sidering an incompressible, inviscid, variable density, the conservation
equations are
dp (18)
dt
A. F. GHONIEM 327
= (l/p 2 )Vp X Vp + to - Vu (19)
where Vp = -pdu/dt and Vp is computed using the transport element

method as described earlier. Extensions to gravity driven flows at finite
Froude numbers are considered in Refs. 27 and 31. To take the extra term
in Eq. (19) into account, the computational algorithm proceeds in fractional
steps: 1) the vorticity is transported without change along the particle
trajectory, Eq. (8); 2) the material acceleration along the particle path is
computed to find the pressure gradient; 3) the change in the density gra-
dient along the particle path is evaluated using the transport element method,
Eq. (14); and 4) the vorticity is updated according to the discrete analog
of Eq. (19),
(20
TT
dt - - ~
p, * (T)
\dtji *? »
Computations of a variable-density shear layer were performed to inves-
tigate the effect of baroclinic vorticity generation on the growth rate of the
Kelvin-Helmholtz instability, the rate of entrainment, and the rate of
expansion of a shear layer. Figure 12 shows a sample of a two-dimensional
simulation in which the fast top stream is five times heavier than the slow
bottom stream. This configuration resembles the experimental setup used
to study a shear layer between premixed reactants moving in the top stream
and products of combustion moving in the bottom stream. 34 The computed
initial growth rate and phase velocity of the growing waves compared well
with the results of the linear stability theory. 12 3S
In the nonlinear range, the computational results indicate that density
variation 1) induces a net convective motion on the eddy in the direction
of the heavy stream; 2) enforces entrainment asymmetry on the growing
eddy that results in the presence of more light fluid than heavy fluid, by
volume, within the core; and 3) destabilizes the flow leading to the for-
mation of small-scale structures within the large-scale structure. It was also
observed that baroclinic vorticity enhances entrainment over that of the
uniform-density case. Using a Galilean transformation between the tem-
poral and spatial shear layers — i.e., d§/djt = l/Uc(d§/dt), where Uc is the
finite convective velocity — one can show that a shear layer in which the
heavy stream moves faster than the light stream will grow more slowly in
the streamwise direction than a uniform-density layer, in agreement with
experimental observations.36
Similar changes were found in the results of a three-dimensional simu-
lation of a variable-density flow. Figure 13 shows a sample of these results
for the case of a top, fast stream that is twice as heavy as the bottom, slow
stream.30 The results are represented by the grid used to describe the
material surface initially aligned with the midsection of the shear layer, as
in Fig. 7. The motion of the developing eddy, its asymmetry with respect
to the top and bottom streams, and some added asymmetry in the spanwise
direction are seen clearly in the figure. Secondary instabilities, leading to
the deformation of the spanwise core and the formation of streamwise

vortex rods within the braids, are similar to those encountered in the two-
dimensional simulations.
The dynamic origin of these phenomena is the generation of vorticity
due to the baroclinic torque. The misalignment of the density and pressure
gradients within the growing eddy leads to the generation of additional
vorticity with two opposite signs on the two sides of the shear layer, as
shown in Fig. 12. In the case considered here, vorticity of a sign similar
to that of the shear layer is generated on the light, slow fluid side, while
vorticity of the opposite sign is generated on the fast, heavy fluid side.
This redistribution of vorticity within the growing eddy leads to more
entrainment from the lower, light fluid side of the layer and a net stream-
wise motion of the large eddy in the direction of the heavy fluid. The
computed entrainment ratio and finite eddy velocity were found to agree
with experimental results. 32 In Sec. VI.D, we show how baroclinic vorticity
generation affects combustion in the premixed shear layer.
The effect of baroclinic vorticity on the instability of a jet was presented
in Ref. 14, without and with the effect of the Froude number. The role of
baroclinic vorticity in determining the structure of low-speed jet diffusion
flames is analyzed in Ref. 27. The extension of the calculation of a variable-
density flow to a premixed shear layer, and the corresponding role of
baroclinic vorticity are discussed in Sec. VI.D.
t = 5.5 11.0
t - 16.5 t - 22.0
Fig. 12 (continued on next page)

A. F. GHONIEM 329
5.5 11.0
16.5 22.0
Fig. 12 Rollup of an incompressible, nonuniform-density, nonreacting, temporally

growing shear layer in which the top, heavy stream moves to the right and the
bottom, light stream moves to the left at the same speed. The initial conditions are
the same as in Fig. 1, except for the density distribution across the layer, which is
taken as an error function, 1 < p < 5. The density ratio across the layer is 5. The
results are shown in terms of a) isoscalar lines (5 = 0 and in the top stream; s =
I in the bottom stream), and b) vorticity contours. Broken lines indicate positive
vorticity; solid lines indicate negative vorticity.
VI. Vortex Methods for Reacting Flows

The transport element method was extended to reacting flow. 7 - 35 - 37 In
combustion systems, heat release at low Mach numbers leads to the gen-
eration of an irrotational velocity field, V<|>, which when superimposed on
the existing vorticity-induced rotational velocity, w w , represents the total
velocity in a reacting flow. The low-Mach-number approximation is employed
to filter out the pressure waves and to render the pressure independent of
the density in unconfined flows.38 Employing the velocity decomposition,
the governing equations of combustion can be written as follows:
u — u, Vc|> (21)
(22)
~ = — Vp x Vp - <o(V • « ) + ( & > • V)u + — V2co (23)
= V 2 r + QW (24)
pT = const (26)
where T is the temperature, c the species concentration, W the rate of
creation/destruction of species, Q the enthalpy of reaction, and Re, Pe,
and Le the Reynolds, Peclet, and Lewis numbers, respectively. Here, we
assume that 1) the transport properties are constant except for mass dif-
fusivity, which is replaced by a density-averaged value, and 2) molecular
weights of reactants and products are the same. The pressure gradient is
obtained from the momentum equation: Vp/p = — a + l/(pRe)V2u, where
a = du/dt. Both velocity components, UM and Vcf>, must satisfy the velocity
boundary condition in the direction normal to the boundaries of the domain.
Using the principle of vortex decomposition, i.e., discretizing the source
term in Eq. (22) in a form similar to Eq. (2), Green's function solution of
Eq. (22) can be written as7
- x/(*/,0] (27)
The energy equation and the species transport equations, Eqs. (24) and
(25), respectively, are solved using the transport element method in three
fractional steps: convection, diffusion, and reaction. The reaction fractional
step is implemented by changing the strength of the transport elements
according to the following expression7:
k 1
d ^ dW
where k is the total number of species and sk + l = T.

In the following, we review four solutions of reacting shear layers, com-
puted using the transport element methods. The first three solutions are
obtained for the reaction between two streams of fuel and oxidizer, and
the fourth is for a reaction between a stream of premixed reactants and a
stream of products. In the first two cases, an incompressible flow model
was used to eliminate the complexity of the effect of heat release on the
A. F. GHONIEM 331
t = 4.0 t = 12.0
t = 8.0 t = 16.0
Fig. 13 Three-dimensional perspective views of the isoscalar surface 5 = 0, initially
coinciding with an x-y plane in the middle of the domain. The flow is incompressible
but with variable density. Initially, the stream wise velocity and density vary only
in the cross-stream z direction according to similar error functions with different
boundary conditions: — 1 < u < 1 and 1 < p < 2. The shear layer is periodic in
the jc and y directions. All other conditions are similar to the case presented in
Fig. 7.
flow dynamics. In the third and fourth cases, a compressible flow model
was used. In all cases, we focus on the mechanisms of shear flow-com-
bustion interactions under different physical conditions.
A. Reacting Shear Layer

Computation of a reacting shear layer between a fuel stream, F, and an
oxidizer stream, (9, in a spatially developing flow was described in Refs. 39
and 40. The incompressible flow model, p = const, was used to investi-
gate how the flow that evolves due to the rollup instability affects the com-
bustion process. The chemistry is modeled by one-step reaction kinetics,
F + O —> P, and the reaction rate is taken as W = AfCFc0 exp(— 777",,),
where Af is the frequency factor and Ta the activation energy normalized
with respect to the gas constant. A sample of the results, obtained for
TJTR = 10, Af = 8000, and Q = 1, is shown in Fig. 14, revealing a strong
similarity between the temperature, product concentration, and the vor-
ticity contours, being highest within the cores and reaching very small
values within the braids. A large eddy, defined by a closed set of vorticity
contours, Fig. 14c, can also be described by a set of closed product con-
centration contours, around the same origin and of the same shape as the
vorticity contours, Fig. 14b. This similarity confirms the important role of
advection, which is governed by vorticity in this case, in determining the
local concentration field, the rate of mixing, and the rate of burning.
The effects of the shear-layer rollup, the Damkohler number, the Rey-
nolds number, and the reactants ratio across the layer are shown in Fig.
15. These results were obtained for a temperature-independent reaction
in which W = DacFcP, where Da is the Damkohler number. The rollup
of the shear layer enhances product formation over the laminar layer by
inducing strong entrainment currents into the cores. A drop in the total
product formation in the early stages is due to the thinning of the reaction
zone by the strain field. Product formation depends strongly on the
Damkohler number. However, this dependence vanishes, and the total
amount of product formed within the layer reaches an asymptotic value at
a Damkohler number on the order of 20, based on the length scale of the
large structures. This is the fast chemistry limit at which the rate of product
formation depends only on the mixing rate.
At high Reynolds numbers, of an order of 10,000 based on the length
scale of the large eddies, product formation is a weak function of the
Reynolds number. This indicates that, while mixing is strongly enhanced
by convective entrainment, it is also globally limited by it. The mechanism
of this apparent double role is explained as follows. When the Reynolds
number is high, diffusion between the reactants occurs across a highly
convoluted interfacial layer embedded within the large structures at a much
faster rate than the entrainment of the reactants into the large eddies.
Thus, changing the diffusion rate by varying the Reynolds number does
not affect the product formation since the process is entrainment limited.
The phenomenon of mixing asymmetry, which arises due to the finite
velocity difference between the two streams, 7 is shown in the product
distribution in Fig. 14b, and is manifested in the dependence of the product
formation of the reactants' ratio in Fig. 15c. The reactants' ratio is defined
333
Fig. 14 Results of a computation of a spatially developing, reacting shear layer

using an incompressible flow model and Arrhenius chemical kinetics. Figure 2a
shows the vortex elements and their velocity vectors for this case (since the flow is
incompressible, the reaction does not affect the dynamics). The figure shows contours
of a) the top stream reactant's concentration, b) product concentration, c) vorticity,
and d) temperature.
as the ratio between the two reactants" concentrations across the layer.
For equal concentrations of reactants in the two streams, the average
product concentration in the lower, slower stream is persistently higher
than in the upper, faster stream. This is because the large structures entrain
more high-speed than low-speed fluid. The bias toward the entrainment
of the high-speed stream results in a faster chemical reaction in the lower
section of the layer than in the upper section. Increasing the concentration
of either reactant enhances the rate of product formation since the chemical
reaction becomes faster. However, to improve the overall rate of burning,
it is more effective to increase the slow-stream concentration to compensate
for the entrainment asymmetry.
B. Three-Dimensional Reacting Shear Layers

We have shown in Sec. III.C that three-dimensional secondary insta-
bilities follow the saturation of the two-dimensional instabilities. The com-
H
3 °S "S 'S ^
il J SI**"
o 5r a s* o- M a
* a O5 <t £• I' S PRODUCT INTEGRAL / 1% THICKNESS Ap^ ^
fi*l§i?f
65 ST asS
3 N
T G1- O
^ > 2 i i r S * ' ' S «S3 (S) «S) (S) tS) (Si
Tfc ^ II *^ "** aa ^ "" ^ ^ •** ^
^ C" 65 ^
O 3 CO &5
B 3 O* S" 3 £**
Ô 5T ff1
f
O ft
ti '
M f t
e " £. •? "O w
S Ss» c o 2
= * 2% = a
B; ll * 8. ft f
^ K. § 1 2 &
"
S. S S » 1 a
O TJ 3 «j M S"
a rt cr c ^ 45
S 51s- 3
A. F. GHONIEM 335
plicated mixing patterns, which arise due to the evolution of three-dimen-

sional instabilities in shear layers, are revealed by the results of a reacting
flow simulation between a stream of fuel and a stream of oxidizer. 30 Figure
16a shows product concentration and reaction rate contours on a spanwise
x-z section that passes through the midsection of the domain of Fig. 7, and
Figs. 16b and 16c show the two streamwise y-z sections used to depict
vorticity in Figs. 8 and 9. An incompressible, temporal, periodic flow model
and a reaction rate W = Dac0cF are used. The physical parameters are
Da = 10, Pe — 1000, and Le — 1. We show only the late stages, when
three-dimensional effects become important.
These figures show an unexpected inhomogeneity of mixing, and com-
bustion inside the cores in the streamwise and the spanwise directions,
even at the late stages of instability development. As mentioned earlier,
these flows are entrainment dominated. The mixing between the two streams
is initiated along convoluted interfaces produced by the uneven stretch
within the large structures. The stretch is a consequence of the vorticity
instability. Diffusion, a process much slower than convection at high Rey-
nolds number, occurs within these thin, convoluted zones. The chemical
reaction at Da = 10 converts mixed reactants into products almost imme-
diately. Thus, products are formed within thin convoluted zones embedded
inside the large structures.
In two-dimensional simulations, products are found inside the large cores.
In three-dimensional simulations, products are found within the large cores
and within the braids due to the formation of the rods. In both cases, zones
of maximum strain in the spanwise and streamwise planes are depleted of
products by mechanisms that will be discussed in the next section. Thus,
three-dimensional simulation confirms our earlier observation that product
concentration is maximum where vorticity concentration is highest, and
that at zones of high strain and small vorticity, product concentration is
lowest. This correlation holds at all times and across all sections. This
correlation has been observed in the spanwise plane in several two-dimen-
sional computations. 40 In the streamwise planes, the correlation between
vorticity and product concentration is supported by Figs. 16b and 16c, and
Figs. 8 and 9.
The dynamic origin of this correlation is revealed by inspecting the reac-
tion rate contours shown in Fig. 17. At the early stages, products form at
the center of the eddies, where mixing occurs due to the entrainment field.
Later, reaction is maximum at the outer edges of the structure. Thus, at
this high value of Da, product formation zones move from the center of
the core to the outer edges of the eddies. However, at all stages, the
entrainment field draws the products toward zones of maximum vorticity
that exist inside the cores and rods. Thus, although products may form
away from the zone of maximum vorticity, they are constantly brought
there by the convective field.
The reaction zone does not exhibit a similar correlation with the vorticity.
Thus, it has been suggested that vorticity be used as a primary variable in
turbulent combustion models.41
c)
Fig. 16 Results of a three-dimensional simulation of a doubly periodic reacting

shear layer between two reacting streams. The dynamics of this flow is the same as
that in Fig. 7 since the flow is incompressible, a) Product concentration (left) and
reaction rate (right) on an x-z spanwise plane located at the middle of the domain
in Fig. 7. b) Product concentration (left) and reaction rate (right) across the section
described in Fig. 8. c) Product concentration (left) and reaction rate (right) across
the section described in Fig. 9. Results are shown in gray scale in which the maximum
value is indicated by black and the minimum value is shown in white. The gray
scale for the product concentration is fixed, 0 < cp < 1, while the gray scale for the
reaction rate is floating, Wmin < W < Wmax.
336
A. F. GHONIEM 337
a)
Fig. 17 Product distribution within the initial stages in a reaction jet using a
compressible flow model with Arrhenius chemical kinetics. The model is periodic
in the stream wise direction. Results are shown at three time steps for a) Af = 1 5 0
and b) Af = 750.
C. Reacting Jet
The effect of exothermic energy on the flow dynamics and the structure
of the reaction zone was investigated in the case of a reacting jet of fuel
issuing in an atmosphere of oxidizer. 27 42 The flow is two dimensional and
planar, and the jet forms two vorticity layers with two opposite signs on
both sides of its centerline. We used a compressible flow model [Eqs. (21 —
26)] and a single-step Arrhenius reaction, W = AfcFc0 exp( - TJT). Figure
17 depicts the product distributions for Af = 150 and 750. Both were
obtained at Pe = 1000, Le = 1, TrJTR = 4, Q = 3, and TJTR = 10. The
initial perturbation amplitude is e = 0.05X, where X is the wavelength in
the streamwise direction. Figure 18 shows the reaction rate contours for
the two cases.
Product Distribution
At high Damkohler numbers, product concentration is higher due to the
faster chemical reaction, and the eddies are larger due to the extra heat
release, as shown in Fig. 16. At both values of Da, product concentration
within the shear layers changes from a uniform distribution in the stream-
wise direction to a highly concentrated distribution within the large eddies
as the flow evolves. Product concentration within the braids, however,
decreases continuously. The thinning of the braids, due to the formation
of a strong strain field as the eddies roll up, reduces the zones of overlap
between the jet and oxidizer fluids and the zone of chemical activity.
Moreover, the large relative velocity between the jet and oxidizer streams
a) I ' ' ' '
Fig. 18 Distribution of the reaction rate within one side of the centerline of the
reacting jet of Fig. 17, shown using a gray scale for a) Af = 150 and b) Af = 750.
Results are shown in gray scale in which the maximum value is indicated by black
and the minimum value is shown in white.
within the braids also reduces the time available for mixing within this
region. Finally, the entrainment current convects products forming within
the braids into the cores leading to continuous cooling of the region between
neighboring eddies. This is the physical mechanism of reaction extinction
within the braids. Clearly, it is driven by convection.
The similarity between product distribution and vorticity, observed in
Sec. VI.A, persists in the case of a reacting jet, both at low and high
Damkohler numbers. Both product concentration and vorticity exhibit high
values within the eddies and fall continuously along the braids. Thus,
vorticity still controls product distribution within the eddy. At high Damkohler
numbers, vorticity produces a swirling field that enlarges the surface of
contact between the two reactants, diffusion mixes them across the stretched
surface, and products form. The swirling convective field then entrains
these products into the eddy core. At low Damkohler numbers, reactants
are entrained into the cores and then react. The difference between the
two mechanisms is illustrated by examining the reaction rate contours for
both cases.
A. F. GHONIEM 339
Structure of Reaction Zone

At low Damkohler numbers, reaction rate contours presented in Fig. 17
show that the zone of chemical activity, initially uniform in the streamwise
direction, is enlarged during the entrainment phase of the eddy. The mixing
of jet and ambient fluids, induced by the rotational field within the eddy
core, extends the area of combustible mixture. Moreover, the reaction rate
inside the eddies increases with time, triggered by the acceleration of the
chemical reaction as the temperature rises due to the formation of more
products. Thus, large eddies act as exothermic centers that support com-
bustion. At this low Damkohler number, an eddy acts as a "mixture prep-
aration zone" before combustion proceeds. This is consistent with the fact
that the rotational flow within the eddy occurs at a smaller time scale than
the chemical reaction.
At high Damkohler numbers, the reaction zone differs substantially from
that at low Damkohler numbers. At early stages, the reaction rate is max-
imum inside the eddy where combustion occurs over a distributed zone.
In this case, both reactants are drawn into the mixing zone within the eddy
core before they combust. At late stages, combustion occurs on the outer
edges of the eddies within a thin reaction zone. Now, reactants coexist
only around the outer edges of each eddy, where they react and form
products, which are then entrained into the eddy core. Note that with
finite-rate kinetics, and with low-temperature reactants at the early stages,
there is an ignition delay that keeps the kinetic rate lower than the mixing
rate. As the temperature rises, kinetic rates exceed mixing rates and the
reaction region becomes a thin zone between the jet and ambient fluids.
At low and high Damkohler numbers, the reaction rate within the braids
drops sharply as the eddies form due to the entrainment of products into
the eddy cores. (Note that since the Lewis number is unity, temperature
and product concentration are similar.) Product formation is inhibited within
the braids following their cooling by convection.
Effect of Heat Release

In a reacting flow with finite-rate kinetics, we found that heat release
reduces the rate of growth of the instability 27 when the initial perturbation
is small, e = 0.01X. Suppression of the instability becomes stronger as the
Damkohler number is increased. Similar suppression of instability is observed
at infinite-rate kinetics at a range of initial perturbation e = 0.01-0.02A. 42
In both cases, eddy growth in the direction normal to the jet axis is reduced
as the Damkohler number is increased. The reason is as follows: at infinite-
rate kinetics, exothermic energy and volumetric expansion start early, within
the linear stages of instability growth, forming a sublayer of low-density
products within the vorticity layer. Within these stages, it has been shown
that the presence of a low-density sublayer within the vorticity layer leads
to substantial reduction of the instability growth rate. 43 The delay in rollup
results in the formation of a weak, less-coherent eddy. Even at finite-rate
kinetics with small initial perturbation, an appreciable amount of products
forms within the vorticity layer before the instability amplitude is large
enough to initiate rollup. In this case, the early formation of a low-density
sublayer of products, embedded within the vorticity layer, precedes the

rollup and leads to an overall reduction in the instability growth rate.
At finite-rate kinetics with large initial amplitude, e - 0.05X, instability
suppression is negligible, as seen in Fig. 17. Thus, one can overcome the
stabilizing effect of volumetric expansion on shear-layer growth by forcing
the jet at large amplitudes. Initiating rollup, before measurable volumetric
expansion has occurred, causes eddy formation and the onset of mixing
enhancement during the ignition delay time. Conversely, very small initial
perturbations with high exothermic energy result in instability suppression.
D. Premixed Shear Layer

Computations of a shear layer growing between two streams of premixed
reactants and products were presented in Refs. 7 and 35. In this case, the
reaction rate, based on one-step reaction kinetics R —> P, is expressed as
W = AfCR exp(- 77 TJ. Results for a range of Damkohler numbers and
initial amplitude of perturbations were presented. A sample of the results
is shown in Fig. 19 in terms of the rate of reaction, the product concen-
tration, and the vorticity for the case of TJTR = 10, temperature ratio
TPITR = 5, Q = 4, Af = 1, Pe = 1000, and Le = 1. Figure 20 shows the
total mass of products formed within the layer for the same physical param-
eters but for several values of the Damkohler number. The length of the
line of maximum reaction rate and the total mass of product for a repre-
sentative case are shown in Fig. 21. A careful inspection of these results
reveals several interesting observations.
Figures 19 and 20 indicate that one can divide the combustion in a
premixed shear layer into four phases: 1) a laminar flame, 2) a strained
laminar flame, 3) vortex-driven combustion, and 4) a free-propagating
flame. In the second phase, the reaction zone in Fig. 19 is thinner than
that of a laminar flame due to the strain, and the total amount of products
formed, MP = / pcp dydx, is less for the shear layer than for the laminar
flame, MP — S W X, where X is the perturbation wavelength. In the third
phase, the entrainment associated with the formation of a coherent vortex
leads to the swelling of the core and the establishment of a reaction zone
inside the eddy core. Figure 20 shows that the formation of a vortex core
enhances the rate of burning. As the flame leaves the burning vortex, it
returns to the state of a laminar flame. Since these phases result from the
interaction between the flow and the chemical reaction, their starting time
and duration are strongly dependent on the Damkohler number.
Figure 21 shows that the wrinkled flame model, which states that the
rate of product formation Mp = SltLf, where Slt is the laminar burning
velocity and Lf the total flame length, can be used to approximate the
burning rate during the initial growth phase of the eddy provided that Lf
is measured along the line of maximum reaction rate. However, during
the later stages, the value of 5M, as defined earlier, is found to decrease
below the value of the unstrained flame. During the later stages, after most
of the eddy core has burnt, it is difficult to define a flame front due to the
convolution of the streamlines, and further studies of the detail structure
of the reaction zone are needed.
a)
t = 5.5 t= 11.0
t=16.5 t = 22.0
b)
5.5 t=11.0
t=16.5 t = 22.0
Fig. 19 Rollup of a compressible, temporally growing, reacting shear layer between
cold premixed reactants in the top stream and hot products in the bottom stream.
Results are shown in terms of a) the reaction rate (shown in gray scale) and b) the
product concentration contours. The temperature ratio across the layer is 5.
341
0*
0.0 10. 15. 25.
Fig. 20 Total mass of products formed within the premixed reacting shear layer
shown in Fig. 19 compared with the same quantity for a laminar flame propagating
through the same mixture. Results are shown for three different values of Af. Straight
lines show the total mass of products formed in the corresponding laminar flame.
Baroclinic vorticity generation, associated with the interaction between

the hydrodynamic pressure gradient and density field, contributes to the
dynamics of the shear layer in the same way as in the nonreacting, density-
stratified shear layer described earlier: the large eddies move in the direc-
tion of the cold reactants; entrainment asymmetry biases the composition
of the large eddies toward the hot product stream; and local spottiness
appears within the large eddy. One difference is clear by comparing Figs.
12 and 19: heat release slows down the streamwise motion of the large
eddies. As a result of the volumetric dilatation associated with heat release,
the local concentration of vorticity decreases and the induced field on the
vortex weakens. Thus, heat release weakens the instability through the
mechanism of volumetric expansion and not vorticity generation.
Heat release at high Damkohler numbers reduces the growth rate of the
instability at the initial stages and inhibits the rollup process if the initial
perturbation is small, e = 0.01A, similar to the reacting jet case. The
mechanism of instability suppression in this case is different from that found
in the reacting shear layer. In a premixed shear layer, baroclinic vorticity
by itself enhances the instability growth rates, as shown earlier. Thus,
volumetric expansion is necessary for instability suppression in the pre-
mixed shear-layer flow.
The effect of Damkohler number on the structure of the reaction zone
is similar to the reacting jet; at low Damkohler numbers, reaction occurs
A. F. GHONIEM 343
a
o-i
a a
CD- to-
to
t—
CD
§
ZD
CD-
to
C£
Q_
LO
LO
cn
a
a •
a a
a _ a
Q.Q 5.0 10.0 15.0 20.0 25.0
TIME
Fig. 21 Total mass of products and the length of the line of maximum reaction
rate in a premixed reacting shear layer similar to the one shown in Fig. 19 and in
the corresponding laminar flame. The temperature ratio across the layer is 3, Pe
= 200 and Ta = 5.
within the eddy core, and at high Damkohler numbers, it occurs primarily
on the outer edges of the eddy. As in the reacting shear layer, product
concentration is always higher at the center of the eddy. (Premixed shear
layers are, however, more subtle, since product forming during the devel-
opment of the shear layer and products existing at the initial state are not
easily distinguishable.) As shown in Fig. 20, the effect of the shear layer
on the burning rate is stronger at low Damkohler numbers.
VII. Extensions
Work reviewed so far focuses on combustion in free shear flows. A brief
summary of the numerical methodology and its application to nonreacting
and reacting shear layers has been presented. Analysis of the computational
results has been aimed at testing the accuracy of the numerical schemes
and describing some of the general properties of these flows. In the reacting
flow, mechanisms of shear flow-combustion interactions were analyzed in
light of the numerical solutions.
The application of vortex methods to internal, wall bounded flows in
which the growth of boundary layers along solid walls and their separation
at sharp edges play a dominant role in the dynamics of the flow, has been
based largely on the random vortex method. In this method, the effect of
molecular viscosity is taken into account by adding a Gaussian random

component to the convective motion of the vortex elements. 44 - 45 Extensive
work on the validation of the method 46 - 47 shows that solutions for steady,
low-Reynolds-number flows and unsteady, high-Reynolds-number flows
converge to appropriate limits as numerical parameters are refined. Low-
Reynolds-number results were in agreement with experimental measure-
ments on velocity distributions within the flow. At high Reynolds numbers,
results were shown to converge to oscillatory flows that can be characterized
by time-dependent clusters of large-scale vortices. The dependence of the
shedding frequency and growth rate of these structures on the geometry
was investigated in Ref. 48. The interactions between jet flow, recirculating
flow, and annulus flow, encountered in a bluff-body flame burner, have
been revealed in vortex simulation. 49
The random vortex method was also applied to study reacting recircu-
lating flows of premixed gases at high Reynolds numbers 50 - 51 utilizing the
thin-flame approximation 52 to model the combustion process. Results were
used to study vorticity-flame-pressure interactions in a semiconfined, recir-
culation-stabilized premixed flame. Analysis of these results shows how
the volumetric expansion associated with burning can reduce the amplifi-
cation of flow oscillation in recirculating flows when the pressure is kept
constant, and how these oscillations can be amplified leading to large flame
oscillation if the pressure field within the system is coupled with the flow
processes.51-53 These results address the problem of turbulent flame sta-
bilization in a premixed stream and the associated combustor instabilities
observed in such systems.
VIII. Conclusion
Numerical simulation, using accurate schemes to integrate the unsteady
equations governing reacting flow, can be applied to investigate important
mechanisms of shear flow-combustion interaction in different systems and
within a wide range of physical parameters. Mechanisms of flow-combus-
tion interaction are different in nonpremixed and premixed shear layers.
Entrainment, associated with the formation of structures that develop due
to natural shear flow instability, is the primary mechanism by which the
flow enhances the rate of burning in both cases. Burning enhancement is
substantial, especially at low Damkohler numbers, where entrainment
changes the reaction region from a thin front into a distributed zone. Strain
may have some effect on the burning mechanisms, especially at low
Damkohler numbers, where local extinction has been observed at regions
of high strain rate in the nonpremixed shear layer. In the premixed shear
layer, the reduction in the rate of burning due to strain is greater at high
Damkohler numbers.
Heat release establishes zones of density gradients within the vorticity
layer. Baroclinic vorticity, generated from the interaction between these
density gradients and material acceleration, reduces the growth rate of the
instability in the nonpremixed case while enhancing it in the premixed case.
The latter occurs even in the nonreacting case. Volumetric expansion,
however, suppresses the instability in both cases. The effect is more pro-
A. F. GHONIEM 345
nounced at high Damkohler numbers. Volumetric dilatation reduces the

vorticity and weakens the large structures. Forcing at high amplitudes in
the initial stages can be used to overcome this instability suppression.
Work is under way to extend the numerical methodology in different
directions; for example, 1) formulation of a transport element method for
confined flows in which the interaction between the interior flow and the
wall thermal boundary layer is important; 2) application of the transport
element method to flows in which pressure-density interaction plays a
significant dynamic role54; 3) implementation of fast solvers to reduce the
computational effort from ©(N 2 ) to €>(N) 55 ; 4) implementation of a meth-
odology to allow the application of the scheme to extended, multistep
chemical kinetic models; and 5) formulation of Lagrangian schemes for
multiphase flow such as fuel droplets in gas streams.
Acknowledgments
This work was supported by the U.S. Air Force Office of Scientific
Research under Grant AFOSR 84-0356, the National Science Foundation
under Grant CBT-8709465, and the U.S. Department of Energy under
Grant DE-FG04-87AL44875. Computer support was provided by a grant
from the John von Neumann Computer Center. Results shown in this
chapter were obtained by doctoral candidates O. Knio, A. Krishnan, and
G. Heidarinejad. Computations were performed at the John von Neumann
Computer Center.
References
*Clark, R. A., "Compressible Lagrangian Hydrodynamics Without Lagrangian
Cells," Numerical Methods for Fluid Dynamics II, edited by K. W. Morton and
M. J. Baines, Clarendon Press, Oxford,'England, 1986, pp. 255-272.
2
Fritts, M. J., and Boris, J. P., "The Lagrangian Solution of Transient Problems
in Hydrodynamics Using Triangular Mesh," Journal of Computational Physics,
Vol. 31, 1979, pp. 173-215.
3
Zabusky, N. J., and Overman, E. D., II, "Regularization of Contour Dynamical
Algorithms. 1. Tangential Regularization, Journal of Computational Physics, Vol.
52, 1983, pp. 351-373.
4
Chorin, A. J., and Bernard, P., "Discretization of a Vortex Sheet with an
Example of Roll-Up," Journal of Computational Physics, Vol. 13, 1973, pp. 423-
428.
5
Leonard, A., "Vortex Methods for Flow Simulation," Journal of Computational
Physics, Vol. 37, 1980, pp. 289-335.
6
Ghoniem, A. F., Heidarinejad, G., and Krishnan, A., "Numerical Simulation
of a Thermally-Stratified Shear Layer Using the Vortex Element Method," Journal
of Computational Physics, Vol. 79, 1988, pp. 135-166.
7
Ghoniem, A. F., Heidarinejad, G., and Krishnan, A., "Turbulence-Combustion
Interactions in a Reacting Shear Layer," Lecture Notes in Engineering: Turbulent
Reactive Flows, Vol. 40, edited by R. Borghi and S. N. B. Murthy, Springer-
Verlag, New York, 1989, pp. 638-671.
8
Chorin, A. J., Hughes, T. J. R., McCracken, M. F., and Marsden, J. E.,
"Product Formulas and Numerical Algorithms," Communications on Pure and
Applied Mathematics, Vol. 31, 1978, pp. 205-256.
9
Hald, O., "Convergence of Vortex Methods for Euler's Equations," SI AM
Journal of Numerical Analysis, Vol. 16, 1979, pp. 726-755.
10
Beale, J. T., and Majda, A., "Higher Order Vortex Methods with Explicit
Velocity Kernels " Journal of Computational Physics, Vol. 58, 1985, pp. 188-209.
H
Anderson, C., and Greengard, C., "On Vortex Methods, 11 SI AM Journal of
Numerical Analysis, Vol. 22, 1985, pp. 413-440.
12
Ghoniem, A. F., Heidarinejad, G., and Krishnan, A., "On Mixing, Baroclin-
icity and the Effect of Strain in a Chemically Reacting Shear Layer, 11 AIAA Paper
88-0729, Jan. 1988.
13
Ghoniem, A. F., and Ng, K. K., "Numerical Study of a Forced Shear Layer, 11
Physics of Fluids, Vol. 30, 1987, pp. 706-721.
14
Krishnan, A., and Ghoniem, A. F., "Numerical Simulation of the Structure
of a Heated Jet in a Cold Environment, 11 AIAA Paper 89-0485, Jan. 1989.
13
Chorin, A. J., "Vortex Models and Boundary Layer Instability, 11 SIAM Journal
of Scientific and Statistical Computations, Vol. 1, 1980, pp. 1-21.
16
Leonard, A., "Computing Three Dimensional Incompressible Flows with Vor-
tex Elements, 11 Annual Review of Fluid Mechanics, Vol. 15, 1985, pp. 532-559.
17
Ghoniem, A. F., Aly, H. M., and Knio, O. M., "Three Dimensional Vortex
Simulations with Application to Axisymmetric Shear Layers, 11 AIAA Paper 87-
0379, Jan. 1987.
18
Cattolica, R. J., "Combustion-Torch Ignition: Fluorescence Imaging of NO 2 ,"
Twenty-First Symposium (International) on Combustion, The Combustion Institute,
Pittsburgh, PA, 1986, pp. 1551-1559.
19
Gutmark, E., Parr, T. P., Parr, D. M., and Schadow, K. C., "Evolution of
Vortical Structures in Flames,11 Twenty-Second Symposium (International) on Com-
20
Saffman, P. G., "The Velocity of Viscous Vortex Rings, 11 Studies in Applied
Mathematics, Vol. 49, 1970, pp. 371-380.
21
Knio, O. M., and Ghoniem, A. F., "Numerical Study of a Three-Dimensional
Vortex Method, 11 Journal of Computational Physics, Vol. 86, Jan. 1990, pp. 75-
106.
22
Widnall, S. E., and Tsai, C-Y., "The Instability of the Thin Vortex Ring of
Constant Vorticity, 11 Proceedings of the Royal Society of London, Series A: Math-
ematical and Physical Sciences, Vol. 287, 1977, pp. 273-305.
23
Maxworthy, T., "Some Experimental Studies on Vortex Rings, 11 Journal of
Fluid Mechanics, Vol. 81, 1977, pp. 465-495.
24
Knio, O. M., and Ghoniem, A. F., "Three Dimensional Vortex Simulation of
Entrainment Augmentation due to Streamwise Structures, 11 AIAA Paper 89-0574,
Jan. 1989.
25
Ashurst, W. T., and Meigburg, E., "Three Dimensional Shear Layers via
Vortex Dynamics," Journal of Fluid Mechanics, Vol. 189, 1988, pp. 87-116.
26
Knio, O. M., and Ghoniem, A. F., "Three Dimensional Vortex Simulation of
the Rollup and Entrainment in a Shear Layer, 11 Journal of Computational Physics,
1990 (to be published).
27
Krishnan, A.,'"Vorticity-Combustion Interaction in Reacting Shear Flow,11
Sc.D. Thesis, Dept. of Mechanical Engineering, Massachusetts Institute of Tech-
nology, Cambridge, MA, 1989.
28
Raviart, P. A., "Particle Numerical Models in Fluid Dynamics," Numerical
Methods for Fluid Dynamics II, edited by K. W. Morton and M. J. Baines, Cla-
rendon, Oxford, England, 1986, pp. 231-254.
29
Masutani, S. M., and Bowman, C. T., "The Structure of a Chemically Reacting
Plane Mixing Layer," Journal of Fluid Mechanics, Vol. 172, 1986, pp. 93-126.
30
Knio, O. M., "Three Dimensional Lagrangian Simulation of Reacting Shear
A. F. GHONIEM 347
Flow," Ph.D. Thesis, Dept. of Mechanical Engineering, Massachusetts Institute

of Technology, Cambridge, MA, 1990.
31
Krishnan, A., and Ghoniem, A. F., "Rollup and Mixing in Rayleigh-Taylor
Flow," Journal of Computational Physics, 1991 (to be published).
32
Ghoniem, A. F., and Krishnan, A., "Mixing Patterns and the Generation of
Vorticity in a Density-Stratified Shear Layer," Lecture Notes in Engineering: Pro-
ceedings of the Workshop on Physics of Compressible Turbulent Mixing, Springer-
Verlag, New York, Oct. 1988.
33
Ghoniem, A. F., and Krishnan, A., "Baroclinic Effects in Stratified Flows,"
Physics of Fluids, 1989 (submitted for publication).
34
Keller, J., and Daily, J. W., "The Effect of Highly Exothermic Chemical
Reaction on a Two-Dimensional Mixing Layer," AIAA Journal, Vol. 23, 1985,
pp. 1937-1945.
35
Krishnan, A., and Ghoniem, A. F., "Origin and Manifestation of Flow-Com-
bustion Interaction in a Premixed Shear Layer," Proceedings of Twenty-Second
PA, 1988, pp. 665-675.
36
Dimotakis, P. E., "Two Dimensional Shear Layer Entrainment," AIAA Jour-
nal, Vol. 24, 1986, pp. 1791-1796.
37
Hiedarinejad, G., and Ghoniem, A. F., "Vortex Simulation of the Reacting
Shear Layer; Effect of Reynolds Number and Damkohler Number," AIAA Paper
89-0573, Jan. 1989; also, Combustion Science and Technology, Vol. 72, 1990, pp.
79-99.
38
Majda, A., and Sethian, J. A., "The Derivation and Numerical Solution of
Zero Mach Number Combustion," Combustion Science and Technology, Vol. 42,
1985, p. 185.
39
Heidarinejad, G., "Numerical Simulation of the Reacting Shear Layer Using
the Transport Element Method," Ph.D. Thesis, Dept. of Mechanical Engineering,
Massachusetts Institute of Technology, Cambridge, MA, 1989.
40
Ghoniem, A. F., and Heidarinejad, G., "Effect of Damkohler Number on the
Reactive Zone Structure in a Shear Layer," Combustion and Flame, 1990.
41
Knio, O. M., and Ghoniem, A. F., "Three-Dimensional Lagrangian Simulation
of a Reacting Shear Layer," AIAA Paper 90-0150, Jan. 1990; also, AIAA Journal,
1991, (to be published).
42
Krishnan, A., and Ghoniem, A. F., "Vorticity-Combustion Interactions in a
Reacting Jet," Proceedings of Twelfth International Colloquium on the Dynamics
of Explosion and Reacting Systems, July 1988 (to be published by AIAA).
43
Riley, J. J., and McMurtry, P. A., "The Use of Direct Numerical Simulation
in the Study of Turbulent, Chemically Reacting Flows," Lecture Notes in Engi-
neering: Turbulent Reactive Flows, Vol. 40, edited by R. Borghi and S. N. B.
Murthy, Springer-Verlag, New York, 1989, pp. 486-514.
44
Chorin, A. J., "Numerical Study of Slightly Viscous Flow," Journal of Fluid
Mechanics, Vol. 57, 1973, pp. 785-794.
45
Chorin, A. J., "Vortex Sheet Approximation of Boundary Layers," Journal
of Computational Physics, Vol. 27, 1987, pp. 423-442.
46
Ghoniem, A. F., and Gagnon, Y., "Vortex Simulation of Laminar Recircu-
lating Flow," Journal of Computational Physics, Vol. 68, 1987, pp. 346-377.
47
Sethian, J. A., and Ghoniem, A. F., "Validation Study of Vortex Methods,"
48
Najm, H., and Ghoniem, A. F., "Numerical Simulation of the Convective
Instability in a Dump," AIAA Paper 87-1874, 1987; AIAA Journal (to be pub-
lished).
49
Martins, L. F., "Vortex Computations of Axisymmetric High Reynold Number
Flows in Complex Domains," Ph.D. Thesis, Dept. of Mechanical Engineering,

Massachusetts Institute of Technology, Cambridge, MA, 1990.
50
Ghoniem, A. F., Chorin, A. J., and Oppenheim, A. K., "Numerical Modelling
of Turbulent Flow in a Combustion Tunnel," Philosophical Transactions of the
Royal Society of London, Series A: Mathematical and Physical Sciences, Vol. 304,
1982, pp. 303-325.
51
Ghoniem, A. F., and Najm, H., "Numerical Simulation of the Coupling Between
the Vorticity and Pressure Oscillations in Combustor Instability," AIAA Paper 89-
2665, 1989.
52
Ghoniem, A. F., and Knio, O. M., "Numerical Simulation of Flame Propa-
gation in Constant Volume Chambers," Proceedings of Twenty-First Symposium
pp. 1313-1320.
53
Najm, H., "Numerical Investigation of the Instability of Premixed Dump Com-
bustor," Ph.D. Thesis, Dept. of Mechanical Engineering, Massachusetts Institute
of Technology, Cambridge, MA, 1989.
54
Ghoniem, A. F., "Vortex Methods in Two and Three Dimensions, a Review
and Some Extensions," Proceedings of First AIAA/ASME/SIAM/APS National
Fluid Dynamics Congress, AIAA, Washington, DC, July 1988, p. 658.
55
Ambrosiano, J., Greengard, L., and Rockhin, V., "The Fast Multipole Method
for Gridless Particle Simulations," Yale Univ., New Haven, CT, Research Rept.
YALEU/DCS/RR-565, Sept. 1987.
Chapter 11
Combustion Modeling Using

Probability Density Function Methods
S. B. Pope
I. Introduction
P ROBABILITY density function (pdf) methods have been applied to

a variety of turbulent flows both with and without combustion. In
general, single-phase, low-Mach-number flows have been considered, in
which radiation is not a major factor. For such flows, the fundamental
dependent variables are the velocities U(x,t) and the compositions c)>(jc,r)
(e.g., the species mass fractions and enthalpy). Different probabilistic ap-
proaches to modeling turbulent flows can be categorized according to the
statistics of U(x,t) and <|>(jt,0 that are considered. For example, in a mean-
flow closure, (U(x,t')) and (<j>(jt,/)) are the primary dependent variables; in
second-order closures, the variances (M/H,), (<t>a<t>p) an ^ covariancesX^/cfO
are also included. Angled brackets denote means (i.e., mathematical ex-
pectations), and Uj(x,t) = Ui(x,i) - (Uf(x,t)) and <b'a(xj) = <j>a(x,t) -
(<$>a(x,t)) are the fluctuating components of Uf and cj>a, respectively.
In pdf methods, the dependent variable is a pdf, or joint pdf of U(xj)
and <]>(#,£). The pdf contains information equivalent to all the moments;
hence, in this sense, pdf methods are more comprehensive than moment
closures (e.g., second-order closures). The methods that have proved most
successful are based on one-point, one-time pdf s, which contain infor-
mation at each point in the flow separately, but no joint information at
two or more distinct points.
In the last 15 years, pdf methods have advanced from being only of
theoretical interest to a small group of specialists, to being a practical
approach for calculating the properties of turbulent reactive flows. In ad-
dition to having been applied to idealized flames and simple laboratory
flames, as the subsequent review indicates, the methods have been applied
to flames requiring multistep chemical kinetics (e.g., Refs. 1 and 2) and
to computationally difficult flows (e.g., that in the cylinder of a spark-
ignition engine 3 - 4 ).
reserved.
349
350 COMBUSTION MODELING USING PDF METHODS
This chapter reviews the work on pdf methods with some emphasis on
the numerical issues and on the applications to turbulent combustion. In
the next section, the different pdf methods are described, along with the
modeling they entail. Monte Carlo methods have proved to be the most
successful means of solving pdf transport equations. The essence of these
solution techniques is described in Sec. III. Sections II and III briefly
describe the principal features with no attempt at rigor. The theoretical
foundations of pdf methods (including the modeling and Monte Carlo
solution algorithms) are described comprehensively by Pope.5 Section IV
reviews the applications of pdf methods to turbulent diffusion flames and
premixed flames. Recent applications to constant-density inert flows have
been reviewed in Ref. 6. In Sec. V, some of the outstanding problems and
future directions are assessed.
II. PDF Methods

A. Definitions and Properties
Let c(> denote the value of a composition variable (the mass fraction of
oxygen, for example) at a particular location x(} and time t() in a turbulent
reactive flow. For ease of exposition, we suppose that the flow can be
realized any number of times, and the time t is measured from the initiation
of the flow. Thus, from each realization we obtain a value of 4>; given the
nature of turbulence, these values are, in all probability, different. In other
words, c|) is a random variable. It is not possible to predetermine the value
of ()) that will be obtained in a given realization. However, it is possible to
ascribe probabilities to its value being in a given interval: this can be done
through the pdf.
For every random variable, we introduce an independent (sample-space)
variable: in particular, i|/ is the sample-space variable corresponding to $.
The cumulative distribution function (cdf), F^(\\i), is then defined as the
probability that cj) is less than i|>:
Fîl/) - Probjci) < i|/} (1)
And the pdf of 4>, /^(ip), is defined to be
/*(«io - now (2)

Whereas F^(v|/) is a probability function, /^(i)/) is a probability density
function. That is, /^(i)/) is the probability per unit \\t of the event c|> - \\i.
Equivalently, /^(fy) dijj is the probability of the ()> being in the range ty <
cj) < i|i + dij/.
The three fundamental properties of the pdf [in addition to Eq. (2)]
follow:
AWO ^ ° (3)
S. B. POPE 351
since probabilities are nonnegative;
ft(ty) d»|» = l (4)

since Prob{4> < =c} = 1 and Prob{4> < —x} = 0; and, for any (nonpath-
ological) function Q(<$>),
(0) = f J — =C
/*(*) QW d^J (5)
Equation (5) shows that, if the pdf is known, the mean (or mathematical
expectation) of any function of the random variable can be calculated. In
particular, the mean (cj>) and the rath central moment {<(>''") (ra > 1) can
be determined (if it exists).
For a general turbulent reactive flow, we need to consider a set of cr >
1 composition variables t|> = {^>1, (J>2, . • - , <|>(J}. Accordingly, the a sample-
space variables vji = {ijij, v^2, • • • , V|JCT} are introduced, and the joint pdf of
<|>, /^(i|i) is defined to be the probability density of the compound event
tj> - i|/ (i.e., cj>i - i)/!, 4>2 - i|/2, ..., c()(J = iKT).
Clearly, the joint pdf defined at the particular location x0 and time t(}
can be defined at any (x,t). We denote by f^(\\f,x,t) the joint pdf of c|>(jc,/).
It is important to realize that this is a one-point, one-time joint pdf: it
contains no joint information between <|> at two or more positions or times.
The pdf method described in the next subsection is based on f^(\\f-,x,t),
which is called the composition joint pdf.
Another pdf method described (in Sec. II. C) is based on the velocity -
composition joint pdf, f(y,ify\x,i). Here V = {Vl, V2, V3} are the three
independent velocity variables, and /is the probability density of the com-
pound event {U(x,t) = V, $(x,t) = i|/}.
In the treatment of variable-density flows, two other probability func-
tions prove useful, and are now defined. By assumption (see Ref. 5), the
set of composition variables is sufficient to determine the fluid density.
Thus, if the composition is c{>, the density is given by the function p a (<J>),
which can be determined from a thermodynamic calculation. Conse-
quently, at (x,t) the fluid density is
) (6)
and the mean density [evaluated using Eq. (5)] is
where integration is over the whole of the composition space. Having made
the distinction between the different functions p(je,£) and p(T(()>), we now
follow conventional (if imprecise) notation and denote both by p.
Favre, or density- weighted, pdf's are defined by, for example,
^(i!/) - P(i|/)/(,(il/)/(P) (7)
It then follows that density-weighted means are given by
Q - ^ = /_ e(*)/(*) di|i (8)
[cf. Eq. (5)]. The mass density function ^ is defined by

;jr:,0 (9)
The use of these functions is made apparent in the next two subsections.
B. Composition Joint PDF Equation

Dopazo and O'Brien 7 were the first to consider the transport equation
for /^(i|i;ar,r). Since then, a number of derivations have been given.5 - 8 ~ n
Here we state the result, and refer the reader to Ref. 5 for a detailed
derivation.
The composition $J(x,i) evolves according to the conservation equation
. (10)
Dt p dXf
where/" is the (molecular) diffusive flux of c)>a, and 5tt — a known function
of ((> — is the rate of creation of c|>a due to chemical reaction. The pdf
transport equation corresponding to Eq. (10) is
(11)
On the left-hand side, the first two terms represent the rate of change
following the Favre-averaged mean flow. The third term is—in composi-
tion space—the divergence of the flux of probability due to reaction. The
form of this term gives this pdf method its advantage over other statistical
approaches. Since S(i|/) is known,/^ is the subject of the equation, and v|/a
is an independent variable, the term contains no unknowns. Thus, however
complicated and nonlinear the reaction scheme, in the composition joint
pdf equation the effect of chemical reaction is in closed form, requiring
no modeling.
In contrast, the terms on the right-hand side require modeling. The
quantity (M"|I|J) is the conditional mean of the Favre velocity fluctuation
(u' = U - U)—conditional, that is, upon the event <J> = \\f. The term in
(u"\\\f) represents the transport of /^ in physical space by the fluctuating
velocity. Although there have been other suggestions, this term generally
is modeled by gradient diffusion:
(12)
S. B. POPE 353
where TT is a turbulent diffusivity. Such gradient transport models are, of

course, subject to many objections, especially when applied to variable-
density reactive flows. The final term in Eq. (11) represents the effect of
molecular mixing. It is generally treated by a stochastic mixing model (see,
for example, Refs. 5, 12, and 13). Although some aspects of this modeling
are discussed later, the cited references should be consulted for a full
account.
The composition joint pdf equation [Eq. (11)] is not a self-contained
model. Mean momentum equations must be solved for £7, and a turbulence
model (A>e, say) is needed to determine both F r (~/c 2 /€) and the mixing
rate (~e/fc) used in the stochastic mixing model.
C. Velocity-Composition Joint PDF Equation

Two shortcomings of the composition pdf approach are that turbulent
transport ({w-'|i|i)) has to be modeled, and that the velocity and turbulence
fields have to be treated separately. Both of these shortcomings are over-
come in the velocity-composition joint pdf approach.
The instantaneous momentum equation is
dr. dp
-
where T7> is the stress tensor, p the pressure, and g the gravitational ac-
celeration.
From this equation [and that for cj>, Eq. (10)], the following equation
can be derived5 for the mass density function <ZF(V,ty,x',t) [Eq. (9)]:
BX,
None of the terms on the left-hand side requires modeling. In order, the
terms represent the following: rate of change with time; transport in po-
sition space (by both mean and fluctuating components of velocity); trans-
port in velocity space (by gravity and the mean pressure gradient); and,
as before, transport in composition space due to reaction.
The terms requiring modeling [on the right-hand side of Eq. (14)] are
means conditional on the compound event {U(x,i) = V, fy(x,t) = i|/}. The
term involving Ja—as in the composition pdf equation—represents mo-
lecular mixing. The remaining term involving T,y and/?' represents transport
in velocity space due to molecular stresses and the fluctuating pressure
gradient. A discussion of how the term can be modeled is deferred to the
next subsection.
It may be seen, then, that the velocity-composition joint pdf method

retains the advantage of treating reaction without approximation, and, in
addition, treats transport in physical space (turbulent convection) exactly,
thus avoiding gradient-diffusion assumptions. It also provides a more com-
plete closure: the mean velocity U(x,i), the Reynolds stresses, and indeed
all one-point velocity-composition statistics can be calculated from 3?. The
model equation for 3F is not quite self-contained because the modeled terms
require a knowledge of the turbulent time scale (/c/e) that cannot be de-
duced from 2F.
D. Lagrangian Viewpoint
Thus far, the Eulerian view has been adopted: we have considered func-
tions [e.g., 3£(V,i|/,jt;r)] at a fixed position x. It proves extremely helpful,
both to the modeling and to the numerical solution technique, to take the
alternative Lagrangian viewpoint also.
Let x+(t), U+(t), and <$> + (t) denote the position, velocity, and compo-
sition of the fluid particle that was at a reference point JCG at a reference
time t0. These particle properties evolve according to
!/+(,) = i/(jf+[,],, ,) (15)
since, by definition, a fluid particle moves with the local fluid velocity;
AA>= 8i P (d>
+ +
[from Eq. (13)]; and
^ -"•-***>-{£},.
[from Eq. (10)].
The connection between these equations for the properties of a fluid
particle and the equation for the mass density function 3F [Eq. (14)] is
immediately apparent. Equation (14) can be written
f + ^[W/NH-ÎW y |>] + ^»<ej>]-o (is)

where the expectations are conditional on the compound event {x + (t) =
x, U+(t) = V, fy + (t) = i|i}. Furthermore, it may be noticed that the terms
in braces in Eqs. (16) and (17) appear on the right-hand side of Eq. (14) —
that is, they need to be modeled — whereas all other terms appear on the
left-hand side and are treated exactly.
E. Stochastic Models
The standard approach to turbulence modeling is to construct consti-
tutive relations for the unknown correlations (see, for example, Ref. 14).
S. B. POPE 355
In the context of the mass density function, this approach is to model the
unknown conditional expectations on the right-hand side of Eq. (14) in
terms of known quantities, i.e., functions or functionals of 2£(V,iJj ,*;£).
However, the Lagrangian viewpoint offers a different approach to mod-
eling, namely, to use stochastic processes to simulate unknown contribu-
tions to U+(t) and cf> + (r) [i.e., the terms in braces in Eqs. (16) and (17)].
To illustrate this approach, we consider U*(t) — a stochastic model for
U+(t). If the model is accurate, then U*(t) is (statistically) an accurate
approximation to U+(t). In general, the time series U* is not differentiate.
Consequently, we express the models in terms of the infinitesimal incre-
ment
dJ7*(0 = U*(t 4- dr) - t/*(0 (19)
rather than in terms of the derivative dU*/dt. Note that for a deterministic,
differentiable process [e.g., U + ( t ) ] , the infinitesimal increment is nonran-
dom (i.e., zero variance) and is of order dr.
In view of the equation for U+(t) [Eq. (16)], the increment dU* can be
written
where (similar to U*) x* and c|>* are models of x+ and <(> + . The stochastic
increment dUs models the effects of the fluctuating pressure gradient and
viscous stresses, whereas the term in dr is an exact expression for the effect
of gravity and the mean pressure gradient.
Two types of models for stochastic increments such as dUs have been
used. The first type — of which the stochastic mixing model is an example —
is called the particle interaction model. In the terminology of stochastic
processes, these are point processes. According to these models, the infin-
itesimal increment dUs is nearly always zero. But with probability of order
dr, the increment is of order unity. Thus, the time series is a piecewise
constant, with a finite number of jumps per unit time.
The second type of model uses diffusion processes in which dUs is a
random variable with (conditional) mean and variance both of order dr.
Note that this implies that the rms is of order dr172, and hence the process —
though continuous — is not differentiable. The different variants of the
Langevin model are diffusion processes (see, for example, Refs. 5, 15-17).
For more information on this general modeling approach, the reader is
referred to Ref. 5, whereas the current status of the Langevin model is
described in Ref. 17.
III. Numerical Solution Algorithms

The velocity-composition joint pdf/(V,i|i;jt,f) is a single function defined
in a multidimensional space. In general, /depends on the three velocity
variables, cr composition variables, three spatial variables, and time—(7
+ cr) independent variables in all. In many cases, the dimensionality may
be less, but still large. For example, in a statistically stationary and two-
dimensional flow with a single composition variable, f(V,\\}l\xl,x2) depends
on six independent variables. The composition joint pdf f^(ty\x,t) in general
depends on (4 + or) variables; however, for the simpler flow cited earlier,
/^(iK;*!,^) is a function of just three variables.
Given the large dimensionality of joint pdf s, it is clear that conventional
grid-based numerical methods (e.g., finite differences) are impractical for
all but the simplest cases. Just to provide an accurate representation of a
function of six independent variables is a major task. Consequently, al-
though one or two finite-difference solutions have been obtained for
/cjXîîr*^) (e.g., Refs. 9 and 10), all investigators currently use Monte
Carlo methods instead.
In the next subsection, the general Monte Carlo method devised by Pope5
to solve for the velocity-composition joint pdf is outlined. Then, in Sec.
III.B, Monte Carlo solution algorithms for the composition joint pdf are
reviewed.
A. Monte Carlo Method for the Velocity-Composition Joint PDF
The Monte Carlo method to solve the modeled equation for the velocity-
composition joint pdf is conceptually simple and natural. Rather than dis-
cretizing the space, we discretize the mass of fluid into a large number N
of representative or stochastic particles. At a given time t, let M be the
total mass of fluid within the solution domain. Then each stochastic particle
represents a mass Am = MIN of fluid. The nth particle has position x(n\t),
velocity C/ (w) (0> and composition cj>('7)(0-
Starting from appropriate initial conditions, the particle properties are
advanced in time by the increments
dx^(t) = lfl*\t) dt (21)
1 s
dU^(t) = [g - p(^)- V(p)] dt + dU (22)
w) 5
d<J>< (0 = S($W) dt + d<J> (23)
where dU and dcf> are the stochastic increments that simulate molecular
s 5
processes and the fluctuating pressure gradient. At symmetry boundaries,

particles are reflected; at inflow boundaries, particles are added with ap-
propriate properties; and, at outflow boundaries, particles are discarded.
Although wall boundaries have been treated,18 a comprehensive account
of this treatment is not available in the literature.
The correspondence between the ensemble of stochastic particles and
the joint pdf has been established by Pope.5 The main results follow:
1) The expected density of the stochastic particles in physical space
N
f
Am 2 <8[
L «=i
is equal to the fluid density (p(x,t)).
2) The joint pdf of the stochastic particle properties U(n\t), 4> (/l) (f) is
the density-weighted joint
S. B. POPE 357
3) From particle properties, expectations [e.g., U(x,t)] can be approx-

imated as ensemble averages, with a statistical error of order N'112.
The accurate numerical determination of means [such as U(x,t)} from
the particle properties [x^n\f)9 U(n\i)} is far from trivial. The straightfor-
ward method is to bin the particles in physical space, and then to approx-
imate U (at bin centers) as the ensemble average formed from the particles
within the bin. Although correct in principle, this method produces large
statistical errors. An accurate alternative, based on statistical techniques,
is to estimate means in terms of cubic splines.5-19
Several implementations of the algorithm based on Eqs. (21-23), and
variants of it, have been reported. For example, the turbulent jet diffusion
flame calculations reported in Sec. IV are performed using a "boundary-
layer" variant.5 Haworth and Pope20 report a variant of the algorithm
designed specifically for self-similar shear flows. From a numerical stand-
point, this work is of particular interest, because the convergence of the
method (as N~1/2) is demonstrated. The basic algorithm has been imple-
mented and demonstrated for statistically two-dimensional recirculating
flows by Anand et al.18 Haworth and El Tahry3 4 report calculations of the
three-dimensional time-dependent flow in the cylinder of a spark-ignition
engine. In these calculations, the pdf algorithm is coupled to a conventional
finite-volume algorithm that is used to calculate the mean pressure field
and the turbulent time-scale field.
B. Monte Carlo Algorithms for the Composition Joint PDF

Two different algorithms have been proposed to solve the modeled trans-
port equation for the composition joint pdf.
The algorithm proposed by Pope5 is similar to that described earlier for
the velocity-composition joint pdf. Again, it is a grid-free algorithm in
which the mass of fluid is discretized into N stochastic particles, the nth
of these having position x(n\t) and composition cj> (w) (r). In each time step,
the composition is incremented according to Eq. (23), while the position
is incremented by
[t]j) At + dxs (24)
where the stochastic component dxs causes a random walk to simulate

gradient diffusion. No implementations of this algorithm have been re-
ported in the literature.
A different Monte Carlo algorithm for the composition joint pdf that
has been used extensively was devised by Pope.21 In this method, there is
a finite-difference grid in physical space. At each grid node, the compo-
sition joint pdf is represented by N particles, the nth having composition
^n\t). Reaction and mixing are performed according to Eq. (23), while
particles are moved from node to node to simulate convection and turbulent
diffusion. This algorithm is used in the premixed flame calculation of Refs.
21 and 22, and in the diffusion flame calculations of Refs. 2, 23, and 24.
IV. Turbulent Flame Calculations

A. Turbulent Diffusion Flames
Some of the first pdf calculations are of turbulent diffusion flames.9-10-23-25-26
The calculations reported by Nguyen and Pope23 are the first use of the
Monte Carlo method for jet flames. The results include demonstrations of
convergence of the solutions as N~112 tends to zero.
In the calculations cited earlier, the thermochemistry is handled in a
simple manner — by assuming chemical equilibrium, for example. This re-
duces the number of composition variables to one, namely, the mixture
fraction £. Finite-rate, multistep kinetics have been used by Pope and
Correa27 (see also Ref. 28), Jones and Kollmann, 24 and Chen and Koll-
mann. 2 - 29 A computational challenge is to implement the integration of
the rate equation, i.e.,
f = S«W (25)
in an efficient manner. This equation has to be integrated for every particle

on every time step to determine the change in composition due to chemical
reaction. All investigators have used table look-up algorithms. In the con-
ceptually simplest algorithm, the change in composition Ac)) over the time
step Ar is tabulated as a function of the composition (j> at the start of the
step.
Considerable attention has been paid to the CO/H2-air turbulent dif-
fusion flame studied experimentally by Drake et al.30 Using the velocity-
composition joint pdf approach, Pope and Correa27 and Correa et al.28
report calculations based on a partial equilibrium assumption. This reduces
the number of composition variables to two: the mixture fraction ^l = %
and a reaction progress variable 4>2 — TJ (for the radical recombination
reactions). Consequently, the general set of coupled ordinary differential
equations [Eq. (25)] reduces to the single equation
= S2(4>i,4>2) (26)
since the mixture fraction is conserved (i.e., Sl = 0). The numerical in-
tegration of Eq. (26) is a simple task. As an example of the joint pdf
calculations made for this CO/H2-air turbulent diffusion flame, Fig. 1 shows
the profiles of the mean mole fractions of the major species at an axial
location of 25 nozzle diameters. It may be seen that there is good agreement
between the measurements and the calculations. Further comparisons can
be found in the original works.27 28
Again using the velocity-composition joint pdf approach, Haworth et
al 31,32 |iave mac|e calculations of the CO/H2-air flame using a flamelet
model. In this approach, the instantaneous composition of the fluid is
assumed to be a unique function of the mixture fraction £ and of its dis-
sipation rate \- Consequently, the determination of statistics of composi-
tion — the mean density, for example — requires a knowledge of the joint
pdf of £ and x- The pdf of £ is determined from the velocity-mixture fraction
S. B. POPE 359
pdf equation, whereas different assumptions are made about the distri-
bution of x-
B. Turbulent Premixed Flames

The composition joint pdf approach using the Monte Carlo method has
been applied to premixed flames by Pope1 and McNutt. 22 The former
calculation demonstrated the ability of the pdf method to handle nonlinear
reaction kinetics. A three-variable kinetics scheme was used to calculate
the oxidation of CO and the formation of NO in a propane-air flame
stabilized behind a perforated plate.
0.25 i-
co
o 0.15
03
CD
0.05 -
0.00 4.0 6.0 8.0 10.0 12.0 14.0

Radial Position, r/a
Fig. 1 Comparison of velocity-composition joint pdf calculations (lines) with ex-
perimental data (symbols) for a turbulent syngas diffusion flame.27 Mean mole
fractions against radial distance (normalized by the nozzle radius a) at an axial
distance of 25 nozzle diameters. Symbols: o, H2; A, CO; D, CO2; 0, O2.
The works of McNutt,22 Pope and Anand,33 and Anand and Pope34 are
concerned with the idealized case of a statistically steady, one-dimensional
turbulent premixed flame. In Ref. 34, the velocity-composition joint pdf
method is used, and the effects of variable density are studied. It is shown
that, similar to the Bray-Moss-Libby model,35 the pdf method is capable
of accounting for countergradient transport and large turbulence energy
production due to heat release. The application of the method to a spark-
ignited turbulent flame ball is described by Pope and Cheng.36
Turbulent premixed combustion usually occurs in the flamelet regime.37
This fact presents a challenge to any statistical approach, since the small
scales of the composition fields are no longer governed by the turbulent
straining motions, rather they are determined by reaction and diffusion
occurring in thin flame sheets. Pope and Anand33 present and demonstrate
a model applicable to the flamelet regime. However, as discussed by Pope,5-37
this model is not entirely satisfactory. An alternative approach to treating
flamelet combustion is the stochastic flamelet model of Pope and Cheng.38
This can be viewed as a pdf approach in which a modeled pdf equation is
solved by a Monte Carlo method. In this case, however, the pdf is not that
of fluid properties (i.e., velocity and composition), but is rather the pdf
of flamelet properties (i.e., position, area, and orientation of flamelets).
V. Discussion and Conclusion

The works reviewed briefly in the previous section demonstrate that pdf
methods provide a practicable means of calculating the properties of tur-
bulent reactive flows. Calculations have been made with thermochemical
schemes involving up to three composition variables with finite-rate ki-
netics.1'2 The Monte Carlo method used to solve the pdf equations has
been implemented for a variety of flows, including two-dimensional recir-
culating flows18 and the three-dimensional transient flow in a spark-ignition
engine.34
The most advanced method considered here is the velocity-composition
joint pdf approach. This approach has the advantage, compared with mo-
ment closures, that chemical reaction can be treated exactly, without ap-
proximation. Compared with the composition joint pdf approach, it has
the advantages that turbulent transport is treated exactly, and that a sep-
arate turbulence model is not needed to determine the Reynolds stresses.
A shortcoming of the velocity-composition joint pdf approach is that it
does not provide a completely self-contained model, in that the turbulence
frequency <co) = <e)/fc must be determined by separate means. For example,
in some calculations of simple free shear layers, it has been assumed that
(w) is constant across the flow, and scales with the mean-flow velocity and
length scales.17-27 In more complex flows, another approach is to solve the
standard model equation for (e) (e.g., Ref. 3), or, similarly, to solve a
modeled equation for {co} deduced from those for k and (e).39
A natural extension is to consider/(V,i|/,£;jt,/)—the joint pdf of velocity,
composition, and dissipation. This is the probability density function of
the compound event {U(x,i) = V, <$>(xj) = i|/, e(jc,f) = £}, where e(jc,f)
is the instantaneous mechanical dissipation. Following some preliminary
S. B. POPE 361
investigations,2-40'43 a satisfactory model equation for/fT,^,^;*,/1) has been

developed.41-42 The incorporation of dissipation within the pdf allows more
realistic and accurate modeling. More important, however, the single equa-
tion for /(V,i|/,£;jt,r) provides a completely self-contained model for tur-
bulent reactive flows.
There are three major areas in which progress can be expected in the
next five years. The first area concerns turbulent mixing models. As dis-
cussed in Sec. II, the stochastic mixing models used in pdf methods lead
to discontinuous composition time series — this is clearly contrary to the
physics of the problem. Nevertheless, the stochastic models have many
advantages over alternative suggestions in spite of their lack of physical
appeal. For inert mixing, their performance is generally acceptable; for
reactive flows, especially in the flamelet regime, their performance is highly
suspect. We expect that stochastic models will be improved and refined to
account better for the microstructure of the composition fields, and also
to allow mixing and reaction to proceed simultaneously at finite rates.
The second area of expected progress is in the computational imple-
mentation of complex kinetics. When the Monte Carlo method is used to
solve the joint pdf equation for an inert flow involving a compositions, the
computer time and storage increase nearly linearly with a. In a naive
implementation with complex reaction kinetics, it is necessary to solve the
coupled set of cr ordinary differential equations
= $„(<!>„ <(,2, ..., cJO, a = 1, 2, ..., a (27)
for each particle, on each time step. The right-hand side (which is a com-
bination of reaction rates) is computationally expensive to evaluate scaling
roughly as cr2, and, as is well known, the set of equations is likely to be
stiff. Hence, such a naive implementation is impractiable for all but the
lowest values of a.
As mentioned in Sec. IV. A, the more efficient alternative approach
followed by all investigators is to implement Eq. (27) through a table look-
up scheme.24'27 To date, this has been done on an ad hoc basis, although
progress toward a general methodology can be expected.
The third area of expected progress is in the determination of the mean
pressure field (p(x,t)) using the Monte Carlo algorithm. For thin shear
flows, the mean pressure is determined readily by invoking the boundary-
layer approximations. For statistically stationary, constant-density, two-
dimensional recirculating flow, an algorithm to determine (p) has been
developed and demonstrated.17 However, for the general case, a compu-
tationally efficient and robust algorithm needs to be developed. (In the
three-dimensional transient calculations of Refs. 3 and 4, the Monte Carlo
method is coupled to a finite-volume code that determines (p).)
Acknowledgments
This work is supported in part by National Science Foundation Grant
CBT-8814655, and in part by the U.S. Air Force Wright Aeronautical
Laboratory, Wright-Patterson AFB, under Contract F33615-87-C-2821.
References
]
Pope, S. B., "Monte Carlo Calculations of Premixed Turbulent Flames," Eight-
eenth Symposium (International) on Combustion, The Combustion Institute, Pitts-
burgh, PA, 1981, pp. 1001-1010.
2
Chen, J.-Y., and Kollmann, W., "PDF Modeling of Chemical Nonequilibrium
Effects in Turbulent Nonpremixed Hydrocarbon Flames," Twenty-Second Sym-
posium (International) on Combustion, The Combustion Institute, Pittsburgh, PA,
1988, pp. 645-653.
3
Haworth, D. C., and El Tahry, S. H., "Application of a PDF/Monte Carlo
Approach to In-Cylinder Flows in Reciprocating Engines," Bulletin of the American
Physics Society, Vol. 33, 1988, p. 2281.
4
Haworth, D. C., and El Tahry, S. H., "Application of a PDF Method to In-
Cylinder Flows in Reciprocating Engines," Seventh Symposium on Turbulent Shear
Flows, 1988, Paper 13-1, submitted for publication.
5
Pope, S. B., "PDF Methods for Turbulent Reactive Flows," Progress in Energy
and Combustion Science, Vol. 11, 1985, p. 119.
6
Pope, S. B., "Turbulent Flow Computations Using PDF Methods," Recent
Advances in Computational Fluid Dynamics, Lecture Notes in Engineering,
Springer-Verlag, Vol. 43, 1989.
7
Dopazo, C., and O'Brien, E. E., "An Approach to the Autoignition of a
Turbulent Mixture," Acta Astronautica, Vol. 1, 1974, pp. 1239-1266.
8
Pope, S. B., "The Probability Approach to the Modeling of Turbulent Reacting
Flows,", Combustion and Flame, Vol. 27, No. 3, 1976, pp. 299-312.
9
Janicka, J., Kolbe, W., and Kollmann, W., "Closure of the Transport Equation
for the Probability Density Function of Turbulent Scalar Fields," Journal on Non-
Equilibrium Thermodynamics, Vol. 4, 1979, pp. 47-66.
10
Janicka, J., Kolbe, W., and Kollmann, W., "The Solution of a PDF-Transport
Equation for Turbulent Diffusion Flames," Proceedings of the Heat Transfer and
Fluid Mechanics Institute, Stanford Univ. Press, Stanford, CA, p. 296.
H
O'Brien, E. E., "The Probability Density Function (PDF) Approach to Re-
acting Turbulent Flows," Turbulent Reactive Flows, edited by P. A. Libby and
F. A. Williams, Springer-Verlag, 1980, pp. 185-218.
12
Flagan, R. C., and Appleton, J. P., Combustion and Flame, Vol. 23, 1975,
p. 249.
13
Pope, S. B., "A Monte-Carlo Method for the PDF Equations of Turbulent
Reactive Flows," Combustion Science and Technology, Vol. 28, 1982, p. 131.
14
Lumley, J. L., "Computational Modeling of Turbulent Flows," Advances in
Applied Mechanics, Vol. 18, 1978, pp. 124-174.
15
Pope, S. B., "A Lagrangian Two-Time Probability Density Function Equation
for Inhomogeneous Turbulent Flows," Physics in Fluids, Vol. 26, 1983, pp. 3448-
3450.
16
Haworth, D. C., and Pope, S. B., "A Generalized Langevin Model for Tur-
bulent Flows," Physics in Fluids, Vol. 29, 1986, pp. 387-405.
17
Haworth, D. C., and Pope, S. B., "A PDF Modeling Study of Self-Similar
Turbulent Free Shear Flows," Physics in Fluids, Vol. 30, 1987, pp. 1026-1044.
18
Anand, M. S., Pope, S. B., and Mongia, H. C., "A PDF Method for Turbulent
Recirculating Flows," Turbulent Reactive Flows II: Structure and Predictions, Lec-
ture series in Engineering, Springer-Verlag, Vol. 40, 1988, pp. 672-693.
19
Pope, S. B., and Gadh, R., "Fitting Noise Data Using Cross-Validated Cubic
Smoothing Salines," Comm. Statistics, Vol. 17, 1988, pp. 349-376.
20
Haworth, D. C., and Pope, S. B., "Monte Carlo Solutions of a Joint PDF
Equation for Turbulent Flows in General Orthogonal Coordinates," Journal of
Computational Physics, Vol. 72, No. 2, 1987, pp. 311-346.
S. B. POPE 363
21
Pope, S. B., "A Monte Carlo Method for the PDF Equations of Turbulent
Reactive Flows," Combustion Science and Technology, Vol. 25, 1981, pp. 159-174.
22
McNutt, D. G., "A Study of Premixed Turbulent Flames," M. S. Thesis, MIT,
Cambridge, MA, 1981.
23
Nguyen, T. V., and Pope, S. B., ''Monte Carlo Calculations of Turbulent
Diffusion Flames," Combustion Science and Technology, Vol. 42, 1984, pp. 13-
45.
24
Jones, W. P., and Kollmann, W., "Multi-Scalar PDF Transport Equations for
Turbulent Diffusion Flames," Turbulent Shear Flows 5, edited by F. Durst et al.,
Springer-Verlag, 1987, pp. 294-309.
25
Frost, V. A., "Model of a Turbulent, Diffusion-Controlled Flame Jet," Fluid
Mechanics-Soviet Research, Vol. 4, No. 2, 1975, pp. 124-133.
26
Bywater, R. J., "Numerical Solutions of a Reduced PDF Model for Turbulent
Diffusion Flames," AlAA Journal, Vol. 20, No. 6, 1982, pp. 824-830.
27
Pope, S. B., and Correa, S. M., "Joint PDF Calculations of a Non-Equilibrium
Turbulent Diffusion Flame," Twenty-First Symposium (International) on Combus-
tion, The Combustion Institute, Pittsburgh, PA, 1986, pp. 1341-1348.
28
Correa, S. M., Gulati, A., and Pope, S. B., "Assessment of a Partial-Equilibrium/
Monte Carlo Model for Turbulent Syngas Flames," Combustion and Flame, Vol.
72, No. 2, 1988, pp. 159-173.
29
Chen, J.-Y., and Kollmann, W., "Chemical Models for PDF Modeling of the
Hydrogen-Air Nonpremixed Turbulent Flames," Combustion and Flame, Vol. 79,
1990, pp. 75-99.
30
Drake, M. C., Pitz, R. W., Correa, S. M., and Lapp, M., "Prediction and
Measurement of a Non-Equilibrium Turbulent Diffusion Flame," Twentieth Sym-
1984, pp. 337-343.
31
Haworth, D. C., Drake, M. C., and Blint, R. J., "Stretched Laminar Flamelet
Modeling of a Turbulent Jet Diffusion Flame," Combustion Science and Technol-
ogy, Vol. 60, 1988, p. 287.
32
Haworth, D. C., Drake, M. C., Pope, S. B., and Blint, R. J., "The Importance
of Time-Dependent Flame Structures in Stretched Laminar Flamelet Models for
Turbulent Jet Diffusion Flames," Twenty-Second Symposium (International) on
33
Pope, S. B., and Anand, M. S., "Flamelet and Distributed Combustion in
Premixed Turbulent Flames," Twentieth Symposium (International) on Combus-
34
Anand, M. S., and Pope, S. B., "Calculations of Premixed Turbulent Flames
by PDF Methods," Combustion and Flame, Vol. 67, No. 2, 1987, pp. 127-142.
35
Bray, K. N. C., Libby, P. A., and Moss, J. B., "Unified Modeling Approach
for Premixed Turbulent Combustion—Part I: General Formulation," Combustion
and Flame, Vol. 61, No. 1, 1985, pp. 87-102.
36
Pope, S. B., and Cheng, W. K., "Statistical Calculations of Spherical Turbulent
Flames," Twenty-First Symposium (International) on Combustion, The Combustion
37
Pope, S. B., "Turbulent Premixed Flames," Annual Review of Fluid Mechanics,
Vol. 19, 1987, pp. 237-270.
38
Pope, S. B., and Cheng, W. K., "The Stochastic Flamelet Model of Turbulent
Premixed Combustion," Twenty-Second Symposium (International) on Combus-
39
Anand, M. S., Pope, S. B., andMongia, H. C., "Calculations of Axisymmetric
Turbulent Jets by the PDF Method," Seventh Symposium on Turbulent Shear
Flows, 1989, Paper 3-3, submitted for publication.
40
Pope, S. B., and Haworth, D. C, "The Mixing Layer Between Turbulent
Fields of Different Scales," Turbulent Shear Flows 5, edited by F. Durst et al.,
Springer-Verlag, 1986, pp. 44-53.
41
Pope, S.B., "Stochastic Model for Lagrangian Dissipation," Cornell University,
Ithaca, NY, Rept. FDA-88-07, 1988.
42
Pope, S. B., "Stochastic Model of Lagrangian Velocity Accounting for Internal
Intermittency," Cornell University, Ithaca, NY, Rept. FDA-88-11, 1988.
43
Pope, S. B., and Chen, Y.-L., "The Velocity Dissipation Probability Density
Function Model for Turbulent Flows," Physics in Fluids, Vol. 2, 1990, pp. 1437-
1449.
Background
When the energy release or fluid flow is very fast, shock waves can develop,
fluid-dynamic effects can dominate diffusive phenomena, compressibility
effects in turbulence are important, and detonation waves may appear.
The chapters in this section present methods and resulting computations
of several very different kinds of high-speed and supersonic reactive flow.
Though they give a picture that is far from all-inclusive, they do represent
current attempts to solve the more basic types of problems.
Supersonic reacting flows, described here in the chapter by Drummond,
have again come to the forefront of technical consideration because of their
importance to the design of propulsion systems for supersonic transport
vehicles. The crucial issue is how to obtain mixing of the fuel stream with
the supersonic air flow that is fast enough to allow efficient combustion in
a finite length without resulting in unacceptable flow losses by slowing the
air down excessively. Drummond describes a number of computational
efforts that address this question, beginning with a description of numerical
methods for high-speed flows, and then treating the application of these
methods to increasingly difficult combustion problems. As motivation,
Drummond provides some background on the history of hypersonic ve-
hicles and their propulsion systems, relating these to computational diffi-
culties. In addition to this comprehensive article by Drummond, the reader
may wish to explore the many papers on supersonic mixing in reactive
flows that have been presented at various American Institute of Aeronau-
tics and Astronautics meetings, particularly the Aerospace Sciences Meet-
ings and the Fluid Dynamics Meetings.
From the discussion of supersonic shear flows, Part III proceeds to det-
onations. The first of four chapters on this broad subject describes research
on numerical simulations of propagating detonations. This is an area in
which numerical simulations have provided important information about
basic mechanisms of ignition, detonation structure and propagation, and
extinction. A number of different research efforts, referenced in this chap-
ter, have tackled the simulation of detonations using models with a wide
range of physical and chemical sophistication, spatial resolution, and di-
mensionality. The chapter first describes one-dimensional reaction-wave
computations using full transient kinetics models that show the transition
of shocks to detonations. In addition to showing the details of transition
physics, these models are well suited to calibrating the simplified chemical
models used in multidimensional simulations of detonation structure. The
chapter then proceeds to discuss some aspects of the multidimensional
structure of propagating detonation waves as uncovered by simulation.
Readers should also be aware of efforts to simulate detonations in two and
three dimensions by Fujiwara and coworkers1 and others referenced in the
chapter.
Next, Sichel begins his chapter on numerical modeling of heterogeneous
detonations with the alarming statement, "Almost all heterogeneous fuel-
oxidizer mixtures will explode under the proper conditions . . . " This state-
ment explains the importance of understanding the properties of hetero-
geneous mixtures and developing ways of safeguarding against coal-mine,
grain-dust, and industrial-dust explosions. Sichel then discusses detonations
in which the solid- or liquid-volume fraction is negligible, so that particle
collisions are neglected, an assumption satisfied for most dust and spray
detonations. Sichel's chapter provides an explanation of the physical proc-
ess, an exposition of what has been computed to date, and an excellent
introduction to where the new research should begin.
Deflagration-to-detonation transition in solid, granular explosives is dis-
cussed in the chapter by Baer, Nunziato, and Embid. This extremely com-
plex problem rivels or exceeds in computational difficulty the closely related
multiphase combustion topics discussed in Section IV. In solid, granular
explosives, the initial combustion is usually slow and propagates by heat
conduction within and between the grains. The speed of the combustion
wave increases as the hot gas products push forward into the pores of the
unreacted material and heat the grains. Under the right conditions the
speed of the burning front can then transition to a detonation. The authors
describe this process through a series of time-dependent, two-dimensional
numerical computations (based on a postulated set of conservation equa-
tions and an extremely complex input data set) to describe the explosives
HMX and CP.
The final chapter in this section deviates somewhat from the pattern of
the preceding ones, and in some ways would also fit well into the more
fundamental and microscopic considerations of the first section of this
book. In this chapter, Peyrard and Odiot discuss the rather ambitious
molecular-dynamics approach to describing the initiation and propagation
of detonation waves in crystal lattices. Molecular dynamics—an approach
in which individual atoms and molecules are placed in a domain, potentials
between them defined, and the system allowed to evolve dynamically—is
well known to physicists. For many years this computational approach has
been used only to study the approach of groups of particles to equilibrium
and to investigate phase transition and the temperature dependence of
transport coefficients.
Other more recent applications, encouraged by the rapid growth of com-
putational capability, treat nonequilibrium problems which reach into the
realms of complex biological molecules and reactive flows, including com-
bustion (recall the chapter by Brown in Section I), and, here, to the prop-
agation of a shock or detonation in a well-defined lattice structure. The
computational representation is particle not fluid dynamics, and different
types of algorithmic developments are necessary. The objective is to try
to combine what the chemists have learned about energy-transfer processes
on the microscale with the macroscopic propagation process. The difficulty
is still the very wide range of time and space scales separating the micro-
scopic atomic and chemical processes and the macroscopic scales of the
fluid. Peyrard and Odiot summarize some of the recent work in this area.
An important field of fundamental research not covered specifically in
any of these chapters is compressible turbulence. A recent conference has
treated this subject but the applications to combustion were specifically
downplayed.2 As compressible turbulence is a relatively new subject, the
numerical methodology is as yet poorly developed.
!
Fujiwara, T., and Reddy, K.V., "Propagation Mechanism of Detonation: Three Dimen-
sional Phenomena," 12th International Colloquium on the Dynamics of Explosions and Re-
active Systems, Ann Arbor, MI, 1989.
2
Proceedings of the International Workshop on Compressible Turbulent Mixing, Springer-
Verlag, 1990.
Chapter 12
Supersonic Reacting Internal Flowfields
J. Philip Drummond
Nomenclature
Aj = reaction rate constant for yth reaction
b( = body force of species i
Ct = concentration of species i
€,- = time rate of change of C,
cp = specific heat at constant pressure
DfJ = binary diffusion coefficient
DT = thermal diffusion coefficient
E = total internal energy; activation energy
E,F,G — flux vectors in x, y, and z coordinate directions
/• = mass fraction of species /
gj = Gibbs energy of species /
GR = Gibbs energy of reaction
H = source vector
h = height of channel or duct
hf = enthalpy of species /
HI = reference enthalpy of species i
Kb = backward rate constant
Keq = equilibrium constant
Kf = forward rate constant
M — Mach number
Mf — molecular weight of species /
ni = moles of species /
nr = number of chemical reactions
ns = number of chemical species
p — pressure
q — heat flux

365
366 SUPERSONIC REACTING INTERNAL FLOWFIELDS
R° = universal gas constant

T = temperature
t - time
TR = reference temperature, =298 K
re = effective temperature
U = dependent variable vector
u - streamwise velocity
u{ = streamwise diffusion velocity of species i
v = transverse velocity
Vj — diffusion velocity vector of species /
vt = transverse diffusion velocity of species /
Wj = species production rate of species i
x = streamwise coordinate
Xt = mole fraction of species /
y - transverse coordinate
Ar = time step
7 = ratio of specific heats
yfj = stoichiometric coefficient; species /, reaction/
8 = Kronecker delta function
r\ = computational transverse coordinate
X = second viscosity coefficient
JJL — mixture laminar viscosity
(x/ = laminar viscosity of species /
£ = computational streamwise coordinate
p = density
cr = normal stress
<jjj = effective collision diameter
T = shear stress
O^ = diffusion collision integral
I. Introduction
Ta renewed interest in the modeling of supersonic reacting flows, both

HE national program to develop a transatmospheric vehicle has kindled
in the United States and abroad. A supersonic combustion ramjet, or

scramjet, has been proposed to provide the propulsion system for this
vehicle. Work has been under way for the past 25 years to develop and
optimize a scramjet propulsion system for a variety of purposes, but the
program has not reached the level of intensity that it currently enjoys during
most of this period. With the maturing scramjet development program,
techniques to model the engine flowfield have also progressed significantly.
This has been due to both advances in numerical methods for computing
reacting flowfields as well as an appreciable growth in computer speed and
storage. The improvements in both algorithms and computer power have
also led to a gradually improved understanding of the physics of reacting
flowfields, which is very important if computation is to have a truly sig-
nificant impact on scramjet design. This chapter will deal principally with
the development of computational techniques for modeling supersonic re-
J. P. DRUMMOND 367
acting flows, and the application of these techniques to an increasingly

difficult set of combustion problems. Since the scramjet development pro-
gram has been largely responsible for motivating this computational work,
we will begin with a brief history of hypersonic vehicles and their propulsion
systems. This will be followed by a discussion of some early modeling
efforts applied to high-speed reacting flows. We will then discuss current
activities to develop accurate and efficient algorithms and improved phys-
ical models for modeling supersonic combustion. Following that discussion,
some new problems where computer codes based on these algorithms and
models are being applied will be described. We will then look forward,
drawing some conclusions concerning future needs and directions for mod-
eling supersonic reacting flows, and, hopefully, challenge the reader to
tackle one of these exciting problems that we will face in the future.
II. Background
Research to develop a supersonic combustion ramjet, or scramjet, pro-
pulsion system was under way in the late 1950s. Work had also begun to
develop computational techniques for solving the equations governing the
flow through a scramjet engine. Scramjet technology and the computational
methods to assist in its evolution would remain apart for another decade,
however. The principal barrier was the lack of a high-speed computer
technology for solving the discrete equations provided by the numerical
methods. Even today, computer resources remain a major pacing item in
overcoming this barrier. Significant advances have been made over the
past 30 years, however, in modeling the supersonic chemically reacting
flow in a scramjet combustor. To see how the two fields finally merged,
we briefly trace the evolution of the technology in both areas.
Following pioneering efforts of Ferri1 and Digger2 in the 1950s, there
was a significant increase in the research to develop scramjet engine con-
cepts in the 1960s. In 1965, the NASA Langley Research Center initiated
the Hypersonic Research Engine (HRE) Project to develop a high-speed
air-breathing technology for application in the propulsion systems of hy-
personic cruise vehicles.3 The goal of the HRE Project was to flight test
a regeneratively cooled, hydrogen-fueled, pylon-mounted scramjet on the
X-15 research airplane and demonstrate design performance levels. The
HRE did not reach the flight demonstration stage due to cancellation of
the X-15 program, but the ground-based program did continue and resulted
in the development and construction of two variable-geometry engine models.
Work with these models significantly increased the scramjet technology
base to be applied in more advanced configurations.
Following completion of the HRE Project, attention moved to propul-
sion concepts that would provide high performance when installed on a
vehicle. The original concept, a pylon-mounted HRE, would have resulted
in excessive levels of external drag, so the pylon was removed and work
began to highly integrate the engine with the airframe of candidate vehicles.
In addition, engine weight was reduced by moving from a variable to a
fixed geometry, which reduced the engine structure. As a result of this
activity, the Langley airframe integrated scramjet engine concept was con-
ceived. This program has continued to the present day, and has resulted
in the successful demonstration of the concept to produce net thrust in
subscale hardware. A detailed review of this program is given in Ref. 3.
In addition to the NASA scramjet research and development program,
other government activities included a Navy-sponsored scramjet program
at the Applied Physics Laboratory of the Johns Hopkins University (JHU/
APL)4'5 This work also increased in the 1960s and was directed toward the
development of an air-breathing shipboard missile using a scramjet pro-
pulsion system. Development of this concept continued until 1977, when
concern over the storage of the highly reactive and toxic fuels to be used
forced a change to more conventional but safer fuels. This change resulted
in the development of an integral-rocket/dual-combustor ramjet concept
that used a fuel-rich gas generator to preburn the fuel for a main supersonic
combustor, thus allowing the use of hydrocarbon fuels. 6
The U.S. Air Force also sponsored scramjet research and development
during the 1960s.4 They continued the support of several programs that
were initially funded by the HRE program. In 1964, a program was started
at the General Applied Science Laboratory to continue development of a
low-speed fixed-geometry scramjet engine. A dual-mode scramjet program
was continued with the Marquardt Company at the same time. Soon there-
after, in 1965, the U.S. Air Force began an effort with the United Aircraft
Research Laboratory to continue development of a water-cooled variable-
geometry scramjet design. These three efforts ended in 1968, and only the
NASA and JHU/APL programs continued into the 1970s.
During the 1970s, computational techniques were first applied to study
the supersonic reacting flow found in a scramjet combustor. A detailed
review of those activities is given in Ref. 4. A summary of that discussion
and additional work is now provided. Some of the earliest work to model
supersonic reacting flows was by Ferri7 and his colleagues, Moretti,8 Edel-
man and Weilerstein,9 and Dash.10-11 They employed an explicit viscous
characteristics method that split the governing equations into hyperbolic
and parabolic parts, followed by a coupled numerical solution of each part
at each integration step. Modeling multistep finite-rate chemistry was also
included in their solution strategy. Spalding and his colleagues then took
Ferri's splitting-based approach and improved its efficiency by developing
a fully implicit solution procedure for solving the governing equations. 12
Spalding then developed several implicit parabolized Navier-Stokes (PNS)
programs for modeling scramjet combustor flowfields. These codes in-
cluded the CHARNAL two-dimensional axisymmetric program 13 and the
SHIP three-dimensional program. 14 Both programs used the well-known
SIMPLE solution procedure for spatially marching the governing equations
in the parabolized direction while employing a tridiagonal matrix solution
procedure to perform repetitive sweeps for solution of the equations in the
cross plane(s).15 These programs assumed that a state of chemical equilib-
rium always existed, but they were later modified by Evans and Schexnayder16
to include the effects of finite-rate chemical reactions. The modified pro-
grams are still being used today for studies of mixing and reaction in
combustor configurations.
J. P. DRUMMOND 369
The works of Ferri and Spalding were then adapted by Dash to develop
the SCORCH program, which used a hybrid explicit-implicit procedure
for modeling supersonic reacting flows. The method again split the gov-
erning equations into hyperbolic and parabolic parts. The hyperbolic part
was solved using a viscous characteristics approach that employed an up-
wind finite-difference procedure. The parabolic part was solved by using
an implicit finite-difference procedure. 17 Work on this program and its
application to supersonic combustion problems has continued to the present
day.
While Ferri, his colleagues, and Spalding were developing analysis tech-
niques for direct application to the supersonic reacting flow problem in a
scramjet, other algorithm development was under way, directed primarily
at solving high-speed external flow problems. These techniques ultimately
found their way, however, into the internal reacting flow arena. The first
of these algorithms was the MacCormack explicit, unsplit predictor-cor-
rector method initially developed to model the hypervelocity impact cra-
tering problem.18 The MacCormack method is a robust variation of the
Lax-Wendroff second-order-accurate scheme that is easily applied to com-
plex geometries. Because of these qualities, the algorithm was readily
adopted and used to study a wide class of internal flow problems. In fact,
due to its general applicability, MacCormack's unsplit algorithm is still
used today as one option in several codes that are applied extensively to
the modeling of scramjet flowfields. Implicit algorithms were also devel-
oped for external flow problems in the 1970s, motivated by the need to
resolve the high gradients present in wall boundary layers. The resolution
of boundary layers requires fine computational grids, resulting in a severe
stability constraint on the marching time step size of an explicit method.
Where only a steady-state solution is required and time accuracy is not
necessary, implicit methods converge much more rapidly. Early work to
develop implicit solution techniques for the Navier-Stokes equations was
carried out by Briley and McDonald19 and Beam and Warming. 20 Both
approaches used a spatial factoring procedure that reduced the multidi-
mensional problem to one of sequentially solving a set of one-dimensional
spatial implicit operators. Using this computationally efficient procedure,
convergence rates one to two orders of magnitude faster than the explicit
method were achieved for steady-state problems on highly stretched grids.
Although the application of implicit methods was generally limited to
scramjet inlet flowfields through the late 1970s and early 1980s, explicit
methods were applied extensively in studies of combustor flowfields. In
1977, Drummond developed the two-dimensional TWODLE combustion
program, based on the MacCormack method, to model internal scramjet
combustor flowfields. The code used an equilibrium chemistry scheme to
model H2-air reaction and several algebraic eddy-viscosity methods to model
the turbulence field. The program was applied to several scramjet com-
bustor component problems. Particular emphasis was given to the scramjet
fuel injector problem in an attempt to better understand the complex
flowfield in this region of the engine. 21 - 22 Development on the program
continued into the early 1980s, when the program was used to carry out
the first simulation of a scram jet flowfield using a two-dimensional model

engine module.24 Detailed studies to optimize the configuration of the
scramjet fuel injectors were also completed during this period.23-25
An explicit solution procedure was also employed by Schetz et al.28
during the early 1980s to model the APL dual-combustion ramjet described
earlier. They employed a modular approach to carry out the analysis. The
mixing and burning of the center jet from the fuel-rich gas generator was
calculated with a jet mixing code26-27 that was modified to include a tur-
bulent kinetic energy mixing model, a chemistry model, and other im-
provements. Because of the high static pressures and temperatures that
were present in the device, a local diffusion-controlled, equilibrium chem-
istry model was used to model reaction in the combustor. Schetz's pro-
cedure for modeling combustor flows was ultimately combined with an
inlet analysis procedure to compute performance estimates for the dual-
combustion ramjet. 29
While numerical methods for modeling scramjet flowfields were devel-
oping through the 1960s, 1970s, and early 1980s, there was a parallel growth
in computer hardware upon which these methods could be applied. Many
of the early calculations were carried out on IBM 7090- and CDC 6600-
class machines. Hardware improvements, which allowed the consideration
of more realistic problems, came in the late 1960s with the arrival of the
CDC 7600 computer. The most significant hardware improvement came
in the mid- to late 1970s, however, when vector processing supercomputers
became available to the computational community. These machines in-
cluded the CDC Star-100 and the Cray 1, followed in the early 1980s by
the Cyber 205 and the Cray X-MP, which gave performance capabilities
several orders of magnitude greater than the previous scalar machines.4
Until this time, the state of computer resources had resulted in a major
barrier to advancing the state-of-the-art in modeling supersonic reacting
flows. W^h the Cyber 205 and Cray X-MP, however, the researcher was
now in a position to begin dealing with the detailed physics contained in
these complex flows. The burden now returned at least partially to the
state of numerical algorithms used to model supersonic combustion.
Both numerical algorithms and computer technology are now pacing our
ability to formulate an improved understanding of supersonic reacting flows.
For the remainder of this chapter, we will concentrate on the numerical
challenge, i.e., what is needed to advance the computational state-of-the-
art and result in an improved understanding of supersonic reacting flows.
We will then explore how this improved understanding can be used to solve
practical problems associated with modeling and designing a scramjet com-
bustor. There is a critical need today for creditable methods for modeling
flows typical of those found in a supersonic combustor. The National Aero-
Space Plane (NASP) will operate at Mach numbers as high as 25. To
develop a successful design for the NASP propulsion system, extensive use
will be made of ground-based facilities to create flowfields consistent with
those that the engine will injest over its operating envelope. Unfortunately,
ground-based facilities are able to create only continuous-flow conditions
up to a flight Mach number of about 8. Beyond Mach 8, the only options
J. P. DRUMMOND 371
available to the experimentalist are pulse facilities that create flight con-
ditions for only a short period of time, providing a data collection window
of only a few milliseconds. Numerical methods provide an alternative to
the Mach 8 barrier, but only when they are applied properly to the problem.
To examine the challenge that this poses for those who are applying com-
putational methods, we will proceed along the following path. First, we
will review the equations that govern the supersonic reacting flow problem
and the modeling that these equations require. Next, we will explore a
number of promising numerical methods, both old and new, for accurately
solving these governing equations. Several solutions for reacting flow prob-
lems using some of these methods will then be presented to assess the
capabilities of the techniques and the computers that provided their results.
We will then be in a position to evaluate where we are with these methods
and where we need to go.
III. Theory
Governing Equations
The Navier-Stokes, energy, and species continuity equations governing
multiple species undergoing chemical reaction have been derived by Wil-
liams.30 The terms used in these and subsequent equations are defined in
the Nomenclature. The governing equations are given by
Continuity:
^ + v - ( P V) = 0 (1)
Momentum:
^P + V - (pW) = V • T + p f /A- (2)

dt /=!
Energy:
Species continuity:
dt - + v - (pv/;.) = w,: - v - (p/M) (4)

where
ii du,-\ ^ du
and
q = -kVT + p £ hfyt + R°T (V, - Vy) (6)

1= 1 1 = 1 ; = 1 \M.iUijJ
Radiation heat transfer is not included in Eq. (6). Also,

ns 3 2
T
4- I cp.dT, i = 1, 2, ..., ns (8)
The diffusion velocities are found by solving
(Vj - V,) + (/, - X,) ^
P J=l
, - »,) + E 7=1 Pîy V /;
- // / ^
do)
Note that if there are ns chemical species, then / = 1, 2, ..., (ns - 1) and
(ns — 1) equations must be solved for the species/. The final species mass
fraction fns can then be found by conservation of mass since
E /, = i
Thermodynamics Model
To calculate the required thermodynamic quantities, the specific heat
for each species is first defined by a fourth-order polynomial in tempera-
ture:
^ - Aj + B{T -f dT2 + AT3 + EfT4 (11)

R
The coefficients are found by a curve fit of the data tabulated in Ref. 31.
Knowing the specific heat of each species, the enthalpy of each species is
then found from Eq. (8), and the total internal energy is computed from
Eq. (7).
To determine the equilibrium constant (required in the next section) for
each chemical reaction being considered, first, the Gibbs energy of each
species must be found. For a constant-pressure process, first, cpIT from
Eq. (11) is integrated over temperature to define the entropy of the species,
J. P. DRUMMOND 373
and then the resulting expression is integrated again over temperature to

obtain a fifth-order polynomial in temperature for the Gibbs energy of
each species.
gl!R = At(T - T f»T) + (5/2)7* + (C/6)r3
4- (0/12)7* + (£/20)r5 + Ff- - G,T (12)
The coefficients Ff and G, are again defined in Ref. 31. The Gibbs energy
of reaction is then calculated as the difference between the Gibbs energy
of product and reactant species:
AGR, = § ifo - f ^,, / = 1, 2, .... nr (13)

/= 1 /= 1
The equilibrium constant for each reaction can then be found from 32
Kcq. = (l/R°T)*" exp[-AG« y ./(/?°r)] (14)
where A« is the change in the number of moles when going from reactants
to products.
Chemistry Models
The rate of chemical reactions is often defined by using the Arrhenius
law. A modified form of the Arrhenius law usually is employed when
modeling supersonic combustion. It is given by
Kfj = yL/r"'exp[-E/(/?°r)] (15)
The values of the pre-exponential constant A, power constant N, and the

activation energy E have been determined for a number of reaction schemes.
Unfortunately, there is a great deal of uncertainty for many chemical re-
actions. One of the best understood mechanisms, however, is the hydrogen-
air reaction system. This is not the reason that hydrogen fuel was chosen
for several scramjet concepts, but it has proven convenient for its com-
bustor analysts! Values for A, TV, and E for a typical hydrogen-air mech-
anism are given in Table 1. Knowing the forward rate, the reverse rate is
then given by
Kbl = Kfi/Kcqi (16)
Once the forward and reverse reaction rates have been determined, the
production rates of the species are found from the law of mass action. For
the general chemical reaction,
™ Kj) ™
S 7//C,. E ^Q, j - 1, 2, ..., nr (17)
Table 1 Finite-rate chemistry model and arrhenius rate coefficients

for each reaction
Reaction
no. Reaction A N E, kJ/g-mole
1 H2 + O,=OH + OH 0.1700^+14 0.00 201.5
2 H + O2=OH + O 0.1420E+15 0.00 68.6
3 OH + H,=H,O + H 0.3160E + 08 1.80 12.78
4 O + H,=OH + H 0.2070E+15 0.00 57.5
5 OH + OH=H,O + O 0.5500E+14 0.00 29.3
6 H + OH=H,O + M 0.2210£ + 23 -2.00 0.0
7 H + H=H2 + M 0.6530E+18 -1.00 0.0
8 H + O,=HO, + M 0.3200E+19 -1.00 0.0
9 HO2 + OH=H,O + O, 0.5000^+14 0.00 4.2
10 HO, + H=H, + O, 0.2530E+14 0.00 2.9
11 HO, + H=OH + OH 0.1990£+15 0.00 7.5
12 HO, + O=OH + O, 0.5000^+14 0.00 4.2
13 HCX + HO,=:H,O, + O, 0.1990^+13 0.00 0.0
14 HO, + H,=H,O, + H 0.3010E+12 0.00 78.2
15 H2O2 + OH=HO2 + FUO 0.1020^+14 0.00 7.9
16 H,O, + H=OH + H,O 0.5000^+15 0.00 41.9
17 HÔ2 + O==OH + HO, 0.1990E+14 0.00 24.7
18 M + H2O2==OH + OH 0.1210^+18 0.00 190.4
For the single-step reaction 0.5510^+ 15 0.00 30.2
the law of mass action states that the rate of change of concentration of
species / by reaction j is given by30
(Q, - (Hi - y'jM L Kf/ /ft= i CT/' - Kbi /fl= i C7

/ = 1,2, ..., ns (18)
The net rate of change in concentration of species / by reaction j is then
found by summing the contributions from each reaction:
(19)
7=1
Finally, the production of species / can be found by multiplying its rate of

change of concentration by its molecular weight:
vv, - C,M, (20)
The source terms in Eq. (4) are now determined as a function of the
dependent variables.
Molecular Diffusion Models

The coefficients governing the molecular diffusion of momentum, en-
ergy, and mass are determined from models based on kinetic theory. The
J. P. DRUMMOND 375
set of models that is often used is now described. Individual species vis-
cosities are computed from Sutherland's law:
P
T + S
where JJLO and T0 are reference values and S is Sutherland's constant. These
constants are tabulated for many species in Refs. 33 and 34. Once the
viscosity of each species has been determined, the mixture viscosity is found
from Wilke's law35:
ns
^m = ^J (22)
~ / 1 ^
1 + 2
V X
Ajj=i,j±i ">
where
/ \ 0.5
0.5 / \ 0.25-
i + fefiA (*L
U,W \M, — (23)
Species thermal conductivities are also computed from Sutherland's law:
^"frf <24>
with different values of the reference values k0 and T'Q and Sutherland's
constant S'. These values are also tabulated for a number of species in
Refs. 33 and 34. The mixture thermal conductivity is computed using con-
ductivity values for the individual species and Wassilewa's formula36:
ns
km = y\ -—————'•—————- (25)
where ^ = 1.065<|)iy, and (f>/y is taken from Eq. (18).

For dilute gases, Chapman and Cowling used kinetic theory to derive
the following 33
expression for the binary diffusion coefficient Dfj between
species / and/ :
D,y = {0.001858r-5 [(Mt + M^MiM,]™}!^ SID) (26)
Here, the diffusion collision integral flD is approximated by
flD = f -0.145 + (f + o.5) ~2 (27)
where f = T/TeiJ. Values of the effective temperature Te and the effective

collision diameter cr are taken to be averages of the separate molecular
properties of each species, giving
try = 0.5(a,. + ay) (28)
and
TBi. = (reiTe.)°-5 (29)
For most molecules, the thermal diffusion coefficient is generally small
when compared with the binary diffusion coefficient, and, therefore, the
thermal diffusion coefficient can be neglected. This is a fortunate fact,
since values of the thermal diffusion coefficient are generally not known
for most species. For low-molecular-weight molecules such as hydrogen,
though, the thermal diffusion coefficient can be important. A set of rela-
tionships for the thermal diffusion coefficient of species having a molecular
weight less than 5 has been developed by Kee et al.37 The reader is referred
to Ref . 37 for further information and some numerical details for computing
thermal diffusion coefficients of light molecules.
Once the binary and thermal diffusion coefficients for all species com-
binations are known, the diffusion velocities of each species can be com-
puted from Eq. (10). The diffusion velocity is the velocity induced upon
each species by all diffusion processes that are present in the flow. The
solution of Eq. (10) requires solving a simultaneous equation system, with
the number of equations equivalent to the number of species present for
each component of the diffusion velocity. It should be noted that for i
species, however, the system of / equations defined by Eq. (10) is not
linearly independent. One of the equations must be replaced by the con-
straint
to make the system linearly independent. The resulting simultaneous sys-

tem of equations must then be solved for the diffusion velocities.
The process of solving for the diffusion velocities can be computationally
quite expensive. A coupled system of equations must be solved for each
of the three components of the diffusion velocity at each computational
grid point. This process can require as much time as solving the Navier-
Stokes equations for the three components of the convection velocities.
Alternately, for hydrogen-air chemistry where large amounts of nitrogen
are present, it is sometimes assumed that each species is present as a "trace"
in a mixture with N2.28 Then each species is assumed to diffuse only into
N 2 , with that process defined by its binary diffusion coefficient with N2.
Finally, for engineering calculations, it is often further assumed that the
diffusivities of each chemical species present in the flow are the same. Then
the diffusion of each species into the remaining species varies only with its
respective concentration gradient. The diffusion velocities then decouple,
J. P. DRUMMOND 377
and Eq. (10) reduces to
VLJ = ~ 7//• —
°xj (3°)
where VLJ is the diffusion velocity vector of the /th species in the /th co-
ordinate direction (/ = |*,.y,z|) and D is the binary diffusion coefficient.
If the binary diffusion with N2 is not used, the value of D is determined
by choosing an appropriate value of the Schmidt number Sc, since D =
|x/(pSc). The mixture viscosity JJL is determined as before from Wilke's law.
When the binary diffusion assumption is invoked, it is often further as-
sumed that the mixture thermal conductivity can be defined by k = (€P\L)/
Pr after an appropriate value of the Prandtl number Pr has been chosen.
Turbulent Diffusion Models

Although the techniques for defining the molecular diffusion of mo-
mentum, heat, and mass are reasonably well established in a supersonic
reacting flow, the same statement cannot be made for our ability to describe
the turbulent diffusion of these quantities. Work to develop methods for
modeling turbulent supersonic combustion is now in its early stages. Con-
ventional approaches have included the use of algebraic eddy-viscosity
models or differential transport models. Several eddy-viscosity models have
been used, in particular, the Cebeci-Smith model38 and the Baldwin-Lomax
model.39 The differential transport models include the k/e turbulent kinetic
energy model and its variants, 40 a modified k/e model that included a
supersonic flow compressibility correction,41-43 and a multiple dissipation
length scale (A:/multiple e) model with a compressibility correction41-42 that
addressed the existence of multiple dissipation length scales that exist in
the energy cascade of a turbulent flow. In addition to these differential
transport models, the algebraic Reynolds stress models of Rodi44 and Sindir45
have also been considered for use in modeling turbulent supersonic reacting
flows. In Ref. 45, Sindir reviews all of these models. In addition, he crit-
ically compares the models against several nonreacting flow experiments
prior to using the models for studying flows with reaction. He concludes
that forms of the algebraic Reynolds stress model that he considered pro-
duced the best agreement with nonreacting data. He also found that the
multiple dissipation length scale model did not offer any advantage over
the basic k/e model.
All of the turbulence models described earlier have a major disadvantage
when applied to reacting flowfields: they fail to account for the important
coupling between the fluid mechanics and the chemistry. Turbulent fluc-
tuations in the fluid mechanic variables have a direct effect on the species
production rates. The coupling between these two fields occurs through
the Arrhenius rate expression [Eq. (15)] and the law of mass action [Eq.
(18)]. The Reynolds averaging process applied to the governing equations
eliminates the direct effect of temperature and species fluctuations on
species production rates. For example, a positive temperature fluctuation
would cause a decrease in the size of the exponential argument of the
Arrhenius rate expression, with a corresponding increase in the forward

kinetic rate of a particular reaction. This would, in turn, produce an in-
crease in the time rate of change of the products of that reaction. More
importantly, if the reaction were at a critical point, where perhaps a small
increase in temperature would cause a reaction to enter an ignition stage,
the entire species distribution of the flowfield downstream could be changed.
Two promising ways for accounting for the effects of fluid and species
fluctuations on chemical reaction would be through probability density
functions or direct numerical simulation. The application of the probability
density function approach to a reacting flow has been covered by Pope in
a companion chapter in this book; therefore, this subject will not be ad-
dressed further here. Direct numerical simulation offers another attractive
approach for modeling a turbulent reacting flow. The method has been
used for several years to accurately model lower speed reacting flows.46"
48
With this approach, the Navier-Stokes and species continuity equations
are resolved down to the smallest scale features of the flowfield. The size
of those scales goes inversely with the Reynolds number of the flowfield.
Clearly, for the high Reynolds numbers that occur in typical supersonic
reacting flows, the smallest scales can become quite small, necessitating a
very fine computational grid to resolve them. Also, when high-speed flow
undergoes chemical reaction, additional scales are introduced by the com-
bustion process. Herein lies the principal difficulty of applying direct sim-
ulation to a high-speed flow. The difficulty is not so much one of numerical
algorithms as it is of computer power. Highly accurate numerical algorithms
are required, but appropriate high-order finite-difference methods, finite-
volume methods, or spectral methods have been developed that satisfy
that requirement. The large number of computational grid points required
to resolve the smallest scales in the flow requires large computer storage,
and, therefore, meaningful calculations can be carried out only on large
memory machines. Currently, direct numerical simulations have been made
for nonreacting flows with Reynolds numbers up to about 10,000 on a Cray
2 computer.49 Work is proceeding to directly simulate a chemically reacting
flow of a similar Reynolds number, and those activities will be discussed
with other applications later in this chapter.
As an alternative to direct numerical simulation with its intensive mem-
ory requirements, it is possible to model rather than compute the smallest
scales. In this approach, termed large-eddy simulation, the larger scales
above a chosen wavelength are still computed. The smaller scales below
the cutoff wavelength are modeled, however, using a subgrid scale model.
Large-eddy simulation is an attractive alternative because only the larger-
scale effects are computed, lessening the computer memory requirements
for higher-Reynolds-number flows. Subgrid scale models must be con-
structed, though, that give an accurate rendering of the physics of small-
scale phenomena. This is a difficult task. Work is under way to develop
subgrid scale models for nonreacting flow, e.g., the early work of Schumann50
and later work described by Speziale et al.51 Large-eddy simulation is an
attractive technique for modeling high-speed reacting flows. Little has been
done so far with this technique, but it warrants serious attention in the
future.
J. P. DRUMMOND 379
Discretization of the Governing Equations

Once the governing equations and required modeling are attained, the
numerical method of choice can be applied to discretize the governing
equations in space and time. The numericist has three basic options for
discretizing the equations in time. He may express the equations explicitly,
implicitly, or in a partially implicit manner. The merits of the first two
approaches were discussed in the Introduction. The latter approach is
attractive when the time scales for chemical reaction are quite small as
compared with the prevailing fluid dynamic time scales. In this case, the
governing equations become stiff, and a significant advantage can be gained
in convergence of the equations to steady state by casting only the chemical
source term in the equations implicitly.52-53 Before discretization, it is con-
venient to express the governing equations [Eq. (1)] in vector form. In
that form, they become
3U + dE + d_F + d_G = H
J
dt dx dy dz ^
where U is the vector of dependent variables; £", F, and G are flux vectors
containing convective and diffusive terms; and H is the source term con-
taining body forces and the chemistry production terms. The temporally
discrete form of Eq. (31), written explicitly, is then given by
—— + —— + —— - / v(32)
;
dx dy dz J
where n is the previous time level and n + 1 is the new time level. Written
implicitly, Eq. (31) becomes
+1
—
dx
+ - —
dy
+ —
dz
- ir v (33)
}
A partial implicit statement of Eq. (31) is obtained when only the source
term is written implicitly, i.e., Hn +1, and the remaining terms are written
explicitly.
Once the temporal discretization of Eq. (31) has been chosen, the spatial
derivatives must also be discretized. There are many choices available. In
the next section, we will examine a number of those choices, using earlier
techniques as well as some newer ones. We will then review some appli-
cations of these methods to practical supersonic combustion problems that
we are able to simulate numerically today.
IV. Numerical Algorithms

A number of numerical algorithms have been used over the past 20 years
to solve the equations that govern a supersonic reacting flowfield. The
earliest of those approaches were reviewed briefly in the Introduction. We
will now discuss, in somewhat more detail, several of those algorithms that
are still in use today. New algorithms that have appeared fairly recently
will then be considered, and their merits and the advantages that they offer
over the earlier approaches will be discussed. Accurate methods used in

other fields and recently borrowed to model combustion problems will also
be reviewed. With the review concluded, in the next section we will proceed
to the application of these algorithms to several practical combustion prob-
lems. We will then be in a position to assess where we are in our ability
to model these practical combustion problems and what is needed to extend
our capabilities further.
Conventional Approaches
The algorithms developed by Spalding, Dash, MacCormack, and their
colleagues today continue to be popular tools for modeling supersonic
reacting flows typical of those found in scramjet combustors. Each of these
algorithms was described in the Introduction, and references were given
to provide more details. The Spalding three-dimensional PNS code, SHIP,14
as modified by Evans and Schexnayder, 16 is still being used to carry out
engineering design studies of scramjet configurations as well as basic high-
speed fuel-air mixing studies. The two-dimensional PNS code, SCORCH,
of Dash et al.17 recently has seen considerable use to perform analyses of
the NASP propulsion system. In addition, the SCORCH code has also
been used to carry out several fundamental studies of experiments being
used to design that propulsion system. The MacCormack algorithm was
employed by Drummond in the TWODLE code21 24 to solve the two-
dimensional Navier-Stokes equations describing a scramjet flowfield. We
will examine some interesting computations from these conventional al-
gorithms in the "Applications" section of this chapter.
A number of other extensions of the MacCormack algorithm were made
following the work that was just described. Drummond et al. extended the
TWODLE code54 to include detailed models for finite-rate chemistry and
kinetic-theory-based models for the molecular diffusion of momentum,
heat, and species. They also added an option for treating the chemical
source term implicitly, as suggested by Bussing and Murman, 52 to allow
for a more efficient treatment of stiff kinetic source terms. In that form,
the new code (SPARK) then solved the complete governing set of equations
[Eqs. (1-29)] in two dimensions without simplification. Options were also
provided, however, to simplify the diffusion modeling using the approach
indicated by Eq. (30). Following its development, the SPARK code was
also applied to NASP configurations. Perhaps more importantly, it was
applied to a number of basic high-speed reacting flow problems to seek an
improved understanding of important physical processes that occur in these
flows, ultimately changing the performance levels that can be achieved by
the propulsion system.
As scramjet technology evolved, a critical need developed for a three-
dimensional analysis tool for modeling high-speed combustor flowfields.
Uenishi and Rogers55 extended the three-dimensional nonreacting Navier-
Stokes inlet code (NASCRIN) developed by Kumar56-57 to include multiple
species, but initially they did not include chemical reaction. The program
again used the unsplit MacCormack method18 to integrate the governing
equations [Eqs. (1-10)]. Thermodynamic properties were also defined
using Eq. (11) and the procedure was described in the discussion following
J. P. DRUMMOND 381
that equation. Molecular diffusion of momentum, energy, and species was

modeled with Sutherland's and Wilke's law, the Reynolds analogy, and
Pick's law, respectively. Turbulent diffusion was modeled with the Baldwin-
Lomax turbulence model39 along with proper values of the turbulent Prandtl
and Schmidt numbers. A clever storage scheme was also employed with
MacCormack's predictor-corrector scheme that saved predictor values lo-
cally only until the corrected values could be computed. 57 The scheme
allowed computations to be carried out with three chemical species on
computational grids with up to 500,000 points on a Cyber 205 having 32
million words of storage. When the code was completed, a number of
studies were conducted to model supersonic fuel-air mixing in scramjetlike
flows.55 Of particular interest were the calculations of the downstream
mixing of a transverse hydrogen fuel jet injected across a supersonic air-
flow. Uenishi's calculations were the first simulations of the three-dimen-
sional near-field fuel injector problem in a scramjet engine. At that time,
a clear understanding of the flowfield that existed near the fuel injectors
was critical to achieving a successful engine design. Work to optimize fuel
injector design based on these and other simulations is continuing today.
Uenishi et al.58 then extended their program to include finite-rate chem-
ical reactions. Motivated by the need to model hydrogen-air combustion
in a scramjet, they chose a two-step hydrogen-air reaction model developed
by Rogers and Chinitz.59 The model considered five species (H2, O2, OH,
H2O, and N2 [inert]) participating in the following chemical reactions:
H2 -f 02 ^ 20H (34a
H2 + 2OH ^± 2H2O (34b

The approach defined by Eqs. (15-20) was then used with this chemistry
model to determine values of the chemistry source terms. The source terms
were sometimes found to be quite stiff numerically, and so the numerical
method was also modified to include implicit source terms. Otherwise, the
numerical method and the physical modeling remained unchanged from
the previous code. Uenishi again applied his extended program to the
transverse fuel jet problem, but in this case with chemical reaction. En-
couraged by those results, he then proceeded to model an actual combustor
configuration. Some results from those calculations will be included in the
next section of this chapter.
Because of the need for a more fundamental modeling capability in three
dimensions, Carpenter and Kamath 61 then extended the SPARK combus-
tion code to three dimensions. The three-dimensional code retained all of
the basic modeling of the two-dimensional code; it solved the system de-
fined by Eqs. (1-30). In addition, Carpenter added a generalized equilib-
rium chemistry model and a generalized finite-rate chemistry model that
allowed consideration of any fuel-air system with any number of reaction
paths.62 Typically, either a seven-species, eight-reaction model or a nine-
species, eighteen-reaction model was used to represent hydrogen-air chem-
istry. The eighteen-reaction model is given in Table 1. Turbulence was
modeled with either a Cebeci-Smith or Baldwin-Lomax eddy-viscosity model,
or with a two-equation k/z turbulent kinetic energy model. Calculations

were then carried out to validate the code by modeling several of the
problems considered by Uenishi. Following agreement with Uenishi's re-
sults and experimental data, the code was used to study a model scramjet
combustor. Those results will also be presented in the "Applications"
section.
Three-dimensional parabolized Navier-Stokes programs were also de-
veloped to model supersonic combustor flowfields. These programs often
provided a more efficient solution procedure if the flowfield contained no
subsonic regions. The flowfield in the neighborhood of the fuel injectors
in a scramjet combustor contains subsonic separated regions necessitating
a solution of the full (spatially elliptic) Navier-Stokes equations. Somewhat
downstream of this region, however, the flow takes on the principal super-
sonic flow direction, allowing solution of the parabolized equations and
the application of parabolized codes. In response to this need, Chitsom-
boon and Northam developed a three-dimensional PNS program63 by ex-
tending a two-dimensional PNS program that Chitsomboon had developed
earlier.64'65 He solved the conventional PNS equations together with a set
of species continuity equations given vectorally by
^ + ^ + *= H (35)
dx dy dz
where E, F, G, and H have the same definitions as those given in Eq. (31).
The dependent variable vector q = [p, pu, pv, pw, T, p,, ...] was chosen
nonconventionally. The energy equation was written in terms of temper-
ature rather than total enthalpy or total internal energy. Equation (35) was
then discretized using the Vigneron method, 66 and the nonlinear vectors
F, G, and H were linearized with respect to the dependent variable vector
q. To ensure that the numerical scheme was then stable, the flux vector E
was linearized using the Schiff-Steger approach.67 Thermodynamics and
chemistry were handled identically to the approach used in the Uenishi
code.60 Following initial construction of the code, it was compared with
the Uenishi program, and fairly good agreement was achieved. Currently,
the code is undergoing further development.
During this same period, Gielda and McRae developed a three-dimen-
sional explicit PNS program68 using the MacCormack explicit algorithm.18
The code was vectorized to run efficiently on vector supercomputers such
as the Cray 2 and the Cyber 205. Gielda and McRae found that their
explicit scheme was quite competitive with implicit algorithms for problems
at high Mach number or with surface discontinuities. They were able to
resolve the problem of decoding the axial flux vector [E in Eq. (35)], which
had earlier limited the application of explicit PNS codes, by also employing
the Vigneron procedure of splitting the axial pressure gradient. Following
completion of this nonreacting program, Gielda and coworkers extended
the code of Gielda and McRae (then named the SSCPNS code) by adding
the parabolized species continuity equations to the governing equation
system.69 They also incorporated both an equilibrium and a global one-
step H2-air finite-rate scheme into the program. The extended program
J. P. DRUMMOND 383
was then validated against several experimental cases, with generally ex-
cellent agreement being obtained between data and computation. The code
was then applied to a three-dimensional generic inlet-combustor scramjet
configuration that included gaseous hydrogen fuel injection and reaction.
Some of these interesting results will be presented in the following section.
Carpenter and Kamath then employed the Gielda algorithm to develop
a parabolized version of the three-dimensional SPARK combustion code.61
They generalized the coordinate transformation to allow the streamwise
coordinate to be orientated in the most supersonic direction. They also
used the generalized equilibrium and finite-rate chemistry schemes devel-
oped by Carpenter,62 so that any multistep reaction scheme could be con-
sidered with the algorithm. The extended code was validated next with
several test cases, including one also utilized by Gielda, and it gave gen-
erally good agreement with the data. Work on this program is continuing.
Alternate Approaches
Several numerical algorithms have been developed to solve the equations
governing high-speed reacting flowfields, but these algorithms have not
been applied to the scramjet combustor problem. Most of these methods
have been developed to model supersonic or hypersonic flow with inter-
acting air chemistry moving externally about the configurations. The re-
maining algorithms have been developed to study basic phenomena associated
with high-speed reacting flows. All of these approaches fall into the general
class of monotone methods; i.e., they employ flux-correcting or flux-lim-
iting procedures to preserve high numerical resolution without the nu-
merical oscillations associated with higher accuracy. Included in this class
of algorithms are flux-corrected transport (FCT) methods, total variation
diminishing (TVD) methods, and upwind methods that exhibit TVD be-
havior. These algorithms would offer the modeler advantages over con-
ventional methods when studying scramjet problems, and they should be
considered seriously for future work. It is for that reason that we discuss
them here.
The first monotone method applied to chemically reacting flows was the
FCT algorithm developed by Boris,70 Boris and Book,71 and Oran and
Boris.72 In this method, a small amount of artificial diffusion is added to
the governing equations in smooth regions of the flow to stabilize the
solution. In regions where high gradients exist, larger amounts of diffusion
are added to maintain monotonicity. The diffusion is added in such a
manner, however, that the overall dissipation is held below that of the
conventional algorithms because most of the diffusion is removed subse-
quently. One proceeds as follows. Starting with Eq. (32), the flux terms
are discretized in space, and then a diffusive term is added to ensure
positivity as the equations are integrated in time. That integration is then
performed to determine a first value of the dependent variable vector at
the next time step. An "antidiffusive" correction using the first values of
the dependent variables is then applied to reduce the numerical diffusion
added in the first step. Care must be taken when applying this step, how-
ever, because the antidiffusive correction can degrade the monotonicity of
the method. Therefore, the antidiffusive terms are limited by a "flux cor-
rection" procedure such that no new maxima or minima are introduced
into the solution. The initial fluxes in Eq. (32) are then replaced by the
corrected fluxes, and the equation is again advanced over the same time
step to arrive at the final value of the dependent variables at the new time.
A more detailed but very readable description of the FCT method is given
in Ref. 72.
Following development of the FCT method, a more general approach
was suggested by Zalesak.73 His approach allowed the method to be readily
incorporated into existing algorithms that did not provide monotone be-
havior. In addition, the method could be generalized more easily to two
and three spatial dimensions. Zalesak viewed FCT as a hybridization of a
low-order and a high-order method. The antidiffusive flux was then found
as the difference between the fluxes determined by the high- and low-order
methods. Once found, the antidiffusive flux was then limited as before by
flux correction, and the solution was advanced using the corrected fluxes.
A more detailed discussion of Zalesak's approach is also given in Ref. 72.
Most of the current work to model high-speed reacting flows has occurred
over the last 5 years. It was motivated primarily by the need to model
hypersonic flows about vehicles, including hypersonic aircraft such as the
NASP, and re-entry vehicles. Therefore, the methods were developed to
model high-speed strongly shocked flows undergoing air chemistry. To
compute flows of this type, MacCormack and Candler developed an im-
plicit flux-split scheme, as an extension to MacCormack's explicit predictor-
corrector finite-difference method,18 to solve the Navier-Stokes equations.
MacCormack initially developed the implicit algorithm to consider only
nonreacting flows.74 A finite-volume approach was used to discretize the
flux terms. In addition, Steger-Warming75 flux vector splitting was intro-
duced to account more properly for the propagation of information through
the flowfield. To split the fluxes, first, the eigenvalues of the governing
equations were determined. The fluxes were then differenced numerically
according to the sign of the eigenvalues. Finally, to relax the constraint on
the time step imposed by the Courant condition in his explicit method, the
flux terms were written implicitly and then linearized. This procedure re-
sulted in a coupled set of algebraic equations to be solved at each time
step once the spatial operators had been applied to the flux terms. The
equation system was then solved iteratively using either a line Gauss-Seidel
procedure or Newton iteration.
Following development of the basic algorithm, Candler and Mac-
Cormack extended the method to consider high-speed airflows that were
ionized and in thermodynamic and chemical nonequilibrium. 76 77 To model
such flows, equations describing species continuity, vibrational energy of
each diatomic molecule, and electron energy were appended to the Navier-
Stokes equations. The fully coupled system of equations was again solved
by using Gauss-Seidel line relaxation along with an implicit, flux-split treat-
ment of the flux terms. Air chemistry was modeled with a seven-species,
six-reaction model that included N2, O2, NO, NO + , N, O, and e~. For
these species, four vibrational temperatures and an electron temperature
were computed. The chemical source terms associated with the seven spe-
J. P. DRUMMOND 385
cies and the thermal source terms were also computed implicitly due to
the small kinetic time scales, relative to fluid scales, that were introduced
by the chemistry. When the authors completed the extension of their al-
gorithm to include chemical reaction, they applied it to several high-speed
external flow problems in both two and three dimensions, as described in
Refs. 76 and 77. Although these examples considered only external flows
with air chemistry, it appeared that the algorithms could readily be modified
to consider internal flows with combustion chemistry, and, therefore, serve
as a means for modeling scramjet combustor flowfields.
Flux-splitting methods were also employed by Grossman, Walters, and
Cinnella to model high-speed chemically reacting flow problems. Grossman
and Walters78 initially developed their algorithm to solve the Euler equa-
tions for nonreacting flows, but included real gas effects. Three forms of
flux splitting were considered, including Steger-Warming flux vector split-
ting,75 van Leer flux vector splitting,79 and Roe flux difference splitting.80
Each of these splitting methods was derived originally to be applied to
ideal gas flows. They were rederived in Ref. 78 to allow their application
to problems with real gas effects. The new derivations are based on the
concept of an equivalent y to replace the ratio of specific heats for an ideal
gas. Each approach is derived with sufficient detail in Ref. 78 to allow the
reader to incorporate the modifications into existing flux-split codes written
to model ideal gas flows. The flux-split equations were solved using a two-
step predictor-corrector method that was second-order accurate in space
and time. Spatial differences were formed using the monotone upstream-
centered scheme for conservation laws (MUSCL) differencing procedure
and flux limiting, as described in Ref. 81. Following the successful appli-
cation of the algorithm to a one-dimensional shock-tube problem, the real
gas splitting was incorporated into a two-dimensional, implicit, finite-vol-
ume code that originally utilized van Leer splitting and Gauss-Seidel line
relaxation to solve the equations governing ideal gas flows.82 This modified
code was then applied to a two-dimensional inlet flowfield exhibiting equi-
librium real gas behavior.
Grossman and Cinnella then extended the algorithms to include vibra-
tional and chemical nonequilibrium. 83 84 They again began with the one-
dimensional Euler equations, but appended species continuity equations
to account for each chemical species present in the reacting flow and vi-
brational energy conservation equations to account for those species in
vibrational nonequilibrium. The authors then redeveloped the relationships
described previously that were required to implement Steger-Warming,
van Leer, and Roe flux splitting. Once these splitting approaches had been
implemented, a finite-volume scheme was used along with either an explicit
Runge-Kutta time integration or an implicit Euler time integration to solve
the governing equations. The shock-tube problem was again revisited, but
nonequilibrium effects were modeled with a five-species, five-reaction model
that included N2, O2, NO, N, and O. Finally, a supersonic nozzle flow
undergoing H2-air chemistry was modeled using a five-species, two-reaction
model involving H2, O2, OH, H2O, and N 2 . The calculations exhibited
excellent agreement with those provided by another program utilizing a
conventional method.83 Extensions of the algorithm to two and three di-
mensions are currently under way, and once in place, this approach should
also offer an attractive option for modeling scramjet combustor flowfields.
A class of more exact flux-split algorithms for nonequilibrium chemistry
was developed recently by Liu and Vinokur.85 Steger-Warming, van Leer,
and Roe flux splitting were again generalized for the nonequilibrium case.
No simplifying assumptions were made, however. The most general ther-
mal and chemical nonequilibrium flow of an arbitrary gas was considered.
The results have not yet been incorporated into a computer program, but
one should become available in the future.
Additional interesting work using flux splitting has also been completed
recently by Liou, van Leer, and Shuen.86 This work has only recently been
submitted for publication; therefore, details cannot be given here. The
authors again employed van Leer flux vector splitting or Roe flux difference
splitting and derived real gas versions of these approaches. The derivations
were begun by assuming a general equation of state for a real gas in
equilibrium. Approaches similar to those discussed previously were then
used to modify the splitting, but the number of assumptions employed were
kept to a minimum. The modified splitting was then incorporated into an
available TVD algorithm87 and used to model several problems described
by the one-dimensional Euler equations. Excellent agreement with the
exact solution for a classical shock-tube problem was obtained.86 Work on
the method is continuing.
A considerable amount of work has also been carried out over the last
several years to develop new TVD schemes for chemically reacting real
gas flows. Beginning in 1985, Yee88 developed a symmetric TVD scheme
that could be employed in the context of either explicit or implicit numerical
integration procedures. The approach was later generalized to consider
chemically reacting flows.89 Yee noted that her approach could readily be
added to existing algorithms that did not exhibit TVD behavior, e.g., the
1969 MacCormack method, resulting in a more robust method with better
shock-capturing qualities. New explicit, semi-implicit, and implicit algo-
rithms employing the symmetric TVD method were then developed and
discussed.89 An explicit multistep TVD scheme was constructed using the
1969 MacCormack method18 for the first two (predictor-corrector) steps,
followed by the addition of a conservative dissipation term as a third step,
such that the overall scheme was TVD. The dissipative term was made up
of products of eigenvectors of Jacobians of the governing equation system
and their associated eigenvalues, an entropy correction, and a limiter func-
tion. Details regarding the construction of the dissipative term and the
determination of its magnitude are given in Ref. 90. For situations de-
scribed earlier, where stiff kinetics results in small chemistry time scales,
the explicit procedure was altered to include an implicit source term while
retaining explicit flux derivatives and the same dissipative terms. Finally,
a fully implicit TVD method was developed including both an implicit
source and flux terms for situations where both chemistry and fluid scales
are small. Calculations using both the explicit and semi-implicit schemes
were also given in Ref. 89. The explicit scheme was used to model a shock
interacting with an object located in a two-dimensional flow. The shock
was captured quite crisply, and more subtle flow discontinuities, including
J. P. DRUMMOND 387
sliplines that were present, were also captured. The semi-implicit scheme
was generalized to three dimensions and incorporated into the Uenishi
semi-implicit code58 described earlier.91 That program also contained the
two-step Rogers-Chinitz H2-air chemistry model discussed earlier. This
three-dimensional TVD code was then used to model the reaction of a
stoichiometric H2-air mixture flowing supersonically in a channel.91 The
flow was ignited by a shock produced by a wedge along the lower wall of
the channel. The resulting fluid and chemistry profiles predicted by the
program in the neighborhood of the shock did not exhibit the overshoots
and undershoots typical of those observed with classical shock-capturing
methods.
When stiff terms are present in the governing equation system and the
fully implicit TVD procedure of Ref. 89 is used to solve two- or three-
dimensional problems, the procedure employed to integrate temporally the
governing equations must be chosen carefully. When alternating direction
implicit (ADI) procedures are used, the factorization error that results
when the implicit operator is spatially factored often cannot be neglected.89
Iterative procedures such as line Gauss-Seidel or Newton iteration, applied
in the approaches discussed earlier, are attractive options. An alternate
procedure developed by Gnoffo et al. employed point-implicit relaxa-
tion.92'93 Gnoffo and coworkers used this procedure in their three-dimen-
sional finite-volume code with a symmetric TVD upwind discretization of
the governing Navier-Stokes, species continuity, vibrational, and electron
energy equations. Pseudo-time relaxation was used to drive the solution
to a steady state. A point-implicit procedure implies that, during a relax-
ation step, the equations at each cell in the computational plane are dis-
cretized by using the latest available data from neighbor cells and by implicitly
updating only those terms that are functions of variables at the cell center.
This procedure has proven to be very efficient on vector computers. Two
options for coupling the governing fluid and chemistry equations, strong
and weak implicit coupling, were also utilized. With strong implicit cou-
pling, the complete equation set was solved as a unit, an approach typical
of those described earlier. Weak implicit coupling involved splitting the
fluid and chemistry equations into two groups, and applying the point-
implicit method to each group separately during the relaxation process.
The former approach was more physically exact, better accounting for
complex wave interactions and fluid-kinetic coupling. The latter approach
allowed for the relaxation strategy and time stepping to be tailored to the
needs of the equation set.93 Air chemistry was modeled in the program
using an eleven-species scheme that included N, O, N 2 , O2, NO, N + , O + ,
N2, O2, NO + , and e~. Further details on the chemistry model and other
physical modeling are given in Ref. 94. The program of Gnoffo et al. was
validated against several experiments and has performed well. The code
was then used successfully to model the high-speed external flow about
several configurations. Because of the code structure, it should be relatively
straightforward to add a H2-air combustion model. Therefore, the program
of Gnoffo et al., utilizing a symmetric TVD upwind discretization with
point relaxation, appears to be an attractive candidate for modeling scramjet
combustor flowfields as well.
Another attractive option to an ADI integration scheme for solving the

spatially discretized governing equations is an LU scheme that approxi-
mately splits the implicit operator into upper and lower operators that are
independent of the dimensionality of the problem. Shuen and Yoon95 de-
veloped a scheme for solving the two-dimensional Navier-Stokes and spe-
cies continuity equations governing chemically reacting flows that employed
an implicit finite-volume time-marching LU method. Details of the deri-
vation of the LU scheme are given in Ref. 95. The approach is quite
attractive because, even though the method is fully implicit, it requires
only scalar diagonal inversion for solution of the flow equations and di-
agonal block inversion of the species equations. The authors stated that,
as a result, the scheme exhibits a fast convergence rate while requiring
only about the same amount of work as an explicit method. 95 This advan-
tage can be particularly important when problems with a large number of
chemical species are being solved. Following development of the LU code
RPLUS, an eight-species, fourteen-reaction chemistry model and the Bald-
win-Lomax algebraic turbulence model was added to the program. The
code was then compared with experimental data and calculations from
other combustion programs, showing very good agreement.95 Encouraged
by their success, Yu and Shuen96 then extended the LU code to three
dimensions (RPLUS3D). A finite-rate chemistry model and a turbulence
model currently are being added to the program. Once completed, the
three-dimensional LU code will provide another valuable tool for modeling
scramjet combustor flowfields.
The alternative methods described earlier for modeling reacting flows
(with the exception of FCT) have generally exhibited second-order nu-
merical accuracy in both space and time. Two high-order-accurate methods
recently have been developed and applied to high-speed combustion prob-
lems. These methods offer improved accuracy as compared with lower-
order methods on a given computational grid and reduced phase error.
One method was developed by Carpenter and Kamath using a fourth-order
compact finite-difference scheme.61 The scheme was developed initially by
Abarbanel and Kumar97 to solve accurately the Euler equations in two and
three dimensions. Carpenter extended these ideas to the Navier-Stokes
equations and used them to alter the 1969 MacCormack method, producing
a fourth-order "compact MacCormack" scheme. The modifications did not
change the basic structure of the MacCormack scheme, allowing it to be
incorporated easily into existing codes using the 1969 algorithm. The mod-
ification significantly improved the accuracy of the algorithm, while mark-
edly reducing the phase error. As a result, the improved scheme was able
to capture strong shocks crisply with very little of the pre- and postshock
oscillations present in the earlier scheme. In fact, the algorithm exhibited
a TVD-like behavior when capturing waves. Because of its attractiveness,
the scheme was added to the two- and three-dimensional SPARK codes
and applied to several supersonic chemically reacting flow problems.61 Very
good agreement with available experimental data was achieved, and the
three-dimensional program was then applied successfully to a three-di-
mensional scramjet combustor. Results from that study will be described
later in this chapter.
J. P. DRUMMOND 389
High-order-accurate spectral methods have also been applied to super-

sonic reacting flows. Drummond and coworkers extended a Chebyshev
spectral method developed to study transitioning flows98 " to include finite-
rate chemical reactions.100 Spectral methods are based on the represen-
tation of the solution of a problem by a finite series of global functions,
in this case, Chebyshev polynomials. To apply this method to the Navier-
Stokes and species continuity equations, the flux terms in these equations
were expanded in terms of the Chebyshev series, and then the required
spatial derivatives were taken. The resulting ordinary differential equations
were then integrated with respect to time using a Runge-Kutta time-step-
ping scheme. Drummond et al. initially developed this technique for the
one-dimensional Euler equations and species continuity equations. 100 The
method was then extended to two dimensions in Ref. 101, where a hybrid
spectral/finite-difference algorithm was used to model two-dimensional
supersonic reacting flows.
Now that both conventional and new alternate methods for modeling
supersonic reacting flows have been examined, we will explore the results
from some of the problems where these methods have been applied. This
should clarify how far we have been able to carry these techniques to solve
some of the difficult problems that we face in the course of designing and
developing a scramjet engine.
V. Applications
Many of the computer programs described earlier in this chapter have
been applied to study supersonic reacting flowfields. In this section, we
will review some of those applications. Space constraints disallow a review
of each effort. Rather, a representative sample of the research will be
provided to show the progress that has been made. We will begin with
some basic two-dimensional studies performed in the late 1970s. We will
then examine some of the first attempts to model simple scramjet flowfields
and see how those studies affected later design activities. The early scramjet
studies raised several basic issues regarding mechanisms controlling the
mixing and chemical reaction of fuels and air at supersonic speeds. These
issues spawned several research efforts undertaken to understand high-
speed combustion physics better and to use this understanding to improve
engine performance. Therefore, next we will examine the results of some
of those efforts. With advancing computer resources in the mid-1980s,
scramjet calculations moved ahead to consider three-dimensional reacting
flows. Those calculations began with simple three-dimensional configura-
tions that served as model problems for the more complicated configura-
tions to be considered later. We will complete our review of the applications
by tracking the progress of these three-dimensional calculations, ending
with a complex combustor configuration not substantially different from
those being considered for advanced propulsion systems to be employed
early in the next century.
Early supersonic reacting flow simulations centered on configurations
defined by experiments that were being performed at that time. Evans et
al.102 used the CHARNAL program as modified by Evans and Schexnayder16
to model a number of experiments available in the late 1970s and compared

their simulations with the available data. One case, the Burrows and Kur-
kov experiment,103 is shown in Fig. 1. Hydrogen was injected at Mach 1.0
through a slot in the lower wall of the configuration. Hot air passed by
the slot at Mach 2.44. The air temperature was sufficient to ignite the
hydrogen gas. Figure 2 shows a comparison between the computed results
and the data, including pitot pressure and reactant and product mass flow
35.6 cm downstream of the slot. The chemistry calculations were made
using either a complete (infinitely fast) or a finite-rate (eddy-breakup)
model.102 The agreement between the data and calculation is quite good
for both pitot pressure and chemical species distributions. Encouraged by
their success, the authors went on to examine several other experimental
cases, each of which is discussed in Ref. 102.
While basic combustion studies continued, some of the effort was shifted
to begin considering scramjet combustor flowfields. The first efforts con-
sidered only a part of the combustor, and they were still limited to two
dimensions. One example is given in Fig. 3, which shows a sketch of a slot
injection experiment undertaken to simulate the transverse injection of
fuel into a scramjet combustor. In this case, helium was injected at Mach
1.0 into a Mach 2.9 air cross stream in a duct. Helium was used to represent
gaseous hydrogen fuel because the test facility in which the experiment
was performed could not support combustion. Weidner and Drummond
modeled the flowfield in the channel using the TWODLE code24 and ob-
tained the results shown in Fig. 3.25 Comparisons of the calculations with
measured static pressures were fair, and the comparison with helium mass
fraction was quite good. Interest in the slot injector case continued through
the 1980s. Shuen and Yoon95 revisited this case in 1988 using their new
LU scheme in the RPLUS code and showed improved agreement with the
static pressure data. They also went on to examine a reacting hydrogen-
air transverse jet case.
Following the nonreacting study, Weidner and Drummond went on to
consider several two-dimensional transverse injector configurations, in-
cluding the ''staged" injectors in Fig. 4.25 Staged injection of hydrogen fuel
provides a means for producing a large region of separated flow between
the injectors, thereby providing improved flameholding in the combustor.
//////////////////////////////////////
0.1048 m
h = 0.004 m
Fig. 1 Sketch of Burrows-Kurkov reacting wall jet experiment.103
J. P. DRUMMOND 391
1.2
-—- Complete reaction O Experiment *

1.0 - - - No reaction
f .6
O Experiment
12
O Experiment /
puy, kg/ms
.2
6 8 4 6 8
y/h
Fig. 2 Profiles of composition, pitot pressure, and mass flow rate vs nondimensional
duct height (y/h) at survey station (x = 35.6 cm) from Burrows-Kurkov experi-
ment.103
The final configuration resulting from the study is shown in Fig. 5. A large
region of separation was produced between the injectors, and significant
reaction of hydrogen fuel and air took place in that region.
Following completion of the fuel injector studies, scramjet calculations
were extended in 1981 to consider a model problem for a complete engine
module, including both inlet and combustor. The calculations were still
constrained to two dimensions due to available computer resources at that
time. Drummond and Weidner24 used the TWODLE program to simulate
the flow through the module shown in Fig. 6. The left portion of the module
provided inlet compression, and the combustor made up the right portion
of the module. A single fuel injection strut was positioned between the
two module walls. Gaseous hydrogen fuel (indicated by the arrows) was
injected at Mach 1.1 from the strut walls as well as from the engine si-
dewalls. That fuel reacted with air that entered the inlet at Mach 5.0. The
O Data
— Calculation
3r
Pref = 66323.0 Pa
_ O
- 5 - 4 - 3 - 2 - 1 0 1 2 3 4 5
x, distance from injector, cm
Fig. 3a Comparison of experimental and computed wall pressures \s streamwise
distance x along channel.
1.0
.8
.6
y/h O Data
.4 —— Calculation
Pref = 2.1 x 106Pa
.2
x = 3.81 cm
.02 .04 .06 .08 1.0
Fig. 3b Comparison of experimental and computed static pressure profiles \s non-

dimensional channel height (y/h).
1.0 O Data
—— Calculation
.8
h = 7.62 cm
x = 3.81 cm
y/h .6
.4
.2
.08 .12 .24 .32
Fig. 3c Comparison of experimental and computed helium mass fraction profiles

\s nondimensional channel height (y/h).
J. P. DRUMMOND 393
M00
r-Jet shock
Upstream / r Downstream
separation separation
^
Reattachment
Stage 1 Stage 2
injection injection
Fig. 4 Staged injection flowfield schematic.
resulting velocity, static pressure and temperature contours, and water

contours are also shown in Fig. 6. The air entering the inlet was turned
by shocks from the inlet leading edges. Shocks are indicated by a coales-
cence of pressure contour lines in the figure. These shocks struck near the
strut leading edge and coalesced with shocks produced by the strut. The
resulting shocks then had sufficient strength to separate the boundary layer
when they reached the engine sidewalls, as indicated by a reversal of the
velocity vectors.
The transverse hydrogen fuel injectors located downstream of the engine
cross-sectional minimum can also be seen in the figures along with their
associated flow separations leading and trailing the injectors. The flow
becomes subsonic near the fuel injectors due to airflow blockage and heat
release from chemical reaction. Some reaction takes place in the separated
regions ahead of the injectors. Significant reaction occurs downstream of
the injectors. The temperature rise and water production associated with
the reaction can also be seen in the figures. The results of the simulation
show that the injectors are not strong enough to penetrate across the flow
at the point of injection. The mixing layers do not meet until around 11
20.0
A A
^-80.0
Fig. 5a Hydrogen mass fraction contours with staged injection.
Fig. 5b Velocity vector field with staged injection.

Fig. 6a Sketch of scramjet engine problem.
Fig. 6b Computed velocity vector field in engine.
Fig. 6c Computed pressure contours in engine.
Fig. 6d Computed temperature contours in engine.
f
H2o;
Fig. 6e Computed water mass fraction contours in engine.

J. P. DRUMMOND 395
cm downstream of the injectors. In an attempt to improve the level of fuel-

air mixing, the injectors were moved a small distance upstream of their
previous location. It was hoped that the change would improve the amount
of injector interaction and enhance mixing. The resulting velocity field is
shown in Fig. 7. An inlet-combustor interaction and choking then occurred,
resulting in large regions of separated flow in the inlet and, clearly, an
unacceptable design. To examine the source of choking, the calculation
was repeated without chemical reaction. The flow then proceeded normally
without any significant separation. Therefore, it was concluded that thermal
choking produced by chemical reaction and subsequent heat release near
the injectors produced the inlet-combustor interaction.
While performed for a two-dimensional engine model problem, these
scramjet simulations did demonstrate in 1981 the potential of numerical
modeling for studying and better understanding engine flowfields. As com-
puter resources improved, calculations of this type in three dimensions
were also attempted. The calculations also raised some basic questions
regarding the very complex physics of mixing and reaction at supersonic
speeds in a scramjet combustor. These questions resulted in several re-
search efforts aimed at achieving an improved understanding of these phe-
nomena. We will now move on to describe some of these basic studies and
then complete this section by examining some three-dimensional simula-
tions of scramjet combustor flowfields.
The supersonic velocities that must exist in a scramjet combustor produce
a perplexing but interesting problem for the engine designer. At such high
speeds, the mixing of the hydrogen fuel injected into the combustor and
air from the inlet is reduced relative to the mixing achieved in lower-speed
operation. As a consequence of the reduced mixing, the degree of reaction
that can occur and the overall combustor efficiency are also suppressed.
The phenomenon of reduced mixing in nonreacting mixing layer flows at
supersonic speeds had been observed experimentally by Brown and Roshko104
as early as 1974, and the effect was studied further by Papamoschou and
Roshko105 in 1986. Numerical studies of the nonreacting problem by Oh106
in 1974 and Hussaini et al.107 in 1986 were performed to better understand
this phenomenon. Oh suggested that the reduced mixing was related to
weak shocks that formed when the local Mach number exceeded 1 about
vortical structures that developed in an evolving mixing layer. He argued
that other vortices would then interact with these shocks and produce
fluctuations in the flowfield that ultimately could reduce the turbulent
kinetic energy and the degree of mixing. Hussaini et al. examined the
detailed interaction of a vortex convecting subsonically relative to a locally
supersonic flow. They showed that a transient shock structure, which they
Fig. 7 Computed velocity field in engine with choked flow.

termed an eddy shocklet, formed as the eddy accelerated, and the shocklet
tended to deform the eddy. Finally, due to this interaction, a vortex of
opposite circulation formed, and the length scale of the original vortex was
reduced. Based on those results, the authors then concluded that eddy
shocklets reduced turbulent mixing through both the production of coun-
terfluctuating vorticity and a reduction of turbulence scale.
The effects of chemical reaction on mixing in a supersonic flow have
also been considered. Beginning in 1986, Drummond and coworkers stud-
ied the supersonic reacting mixing layer that formed between coflowing
streams of hydrogen gas and air at moderate Mach numbers in the range
of Mach 2.54'101-108 Their simulations using the SPARK code indicated that
supersonic reacting flows exhibited some of the same features observed
for subsonic reacting and nonreacting flows. Vortical structure, noted in
much of the subsonic nonreacting flow literature, was shown to be quite
predominant. In agreement with the earlier reacting subsonic literature,
the vortical structure had a marked effect on chemical reaction in super-
sonic flow. Significant burning took place in the eddies on the edges of
the mixing layer, broadening the reaction zone relative to the layer thick-
ness defined by the velocity gradient. In addition, the vortical flow resulted
in the rollup of unburned reactants inside a layer of partially or fully burned
products. This phenomenon, often termed unmixedness in subsonic flows,
prohibited the reaction of captured reactants and reduced the overall ef-
ficiency of the combustion process. There was also some indication that
heat release resulting from the chemical reaction also reduced the amount
of mixing and reaction further downstream in a mixing layer.
Work is continuing to understand better the phenomena that produce
reduced combustion efficiency in supersonic combustors. Related efforts
also began in 1987 to develop techniques for enhancing the degree of fuel-
air mixing and combustion that occurred in high-Mach-number flow. Guir-
guis et al.110 used the flux-corrected transport algorithm of Boris to study
convective mixing in high-speed nonreacting flows. Guirguis and coworkers
again chose the mixing layer as a model problem for mixing in a scramjet
combustor and solved the two-dimensional Euler equations to examine the
effects of imposed pressure gradients on the spatial development of the
layer. The upper stream of the mixing layer, made up of "species 1,"
entered a confined channel at Mach 4.5 and a pressure of 4 atm. The lower
stream of "species 2" entered at Mach 1.5 and a pressure of 2 atm. The
stream temperatures were matched. In addition, two other cases were
considered. In the second case, the pressures of the two streams were
matched and the mixing layer was unconfined; in the third case, the lower
stream entered at a pressure of 2.1 times that of the upper stream. Other
minor differences between the cases were discussed in Ref. 110. One result
from the study by Guirguis is shown in Fig. 8. In that figure, mass fraction
contours of species 1 from 0.01 to 0.99 are plotted for the three cases,
respectively. The matched pressure case showed little development with
increasing streamwise coordinates, whereas the first and third cases with
the imposed transverse gradient showed significant spread of the layer.
Therefore, Guirguis et al. concluded that, in a channel of given length,
differences in pressure across the layer enhanced mixing. They also sug-
J. P. DRUMMOND 397
Case 1
y cm
y cm
y cm
10 15 20
xcm
Fig. 8 Mass fraction contours from 0.01 to 0.99 for mixing layer cases 1-3.
gested means whereby pressure differences could be induced by the ge-

ometry of the combustor. Guirguis and colleagues then continued their
work to find a means of enhancing mixing where the two stream pressures
were matched.111 They concluded that mixing could be enhanced if a bluff
body was placed at the trailing edge of the splitter plate separating the two
matched pressure streams. One result from the bluff-body case is shown
in Fig. 9. The mass fraction contours given in the figure again show that
a significant degree of enhancement is provided by the presence of the
bluff body.
Kumar et al.112 also examined the problem of mixing in flows undergoing
supersonic combustion using a two-dimensional version of the NASCRIN
code. Several techniques for enhancing turbulence and mixing were sug-
gested, and one enhancement technique that employed an oscillating shock
was studied numerically. In this case, a premixed stoichiometric hydrogen-
air flow was processed through a spatially and temporally oscillating shock
wave, and the resulting flow was studied with and without chemical re-
action. Reaction was modeled using an equilibrium chemistry model. One
reacting flow case that was considered is shown schematically in Fig. 10.
Here, the premixed fuel-air flow was introduced into a two-dimensional
2.4
y cm
10 15 20
x cm
Fig. 9 Mass fraction contours from 0.01 to 0.99 for bluff-body-enhanced mixing
layer case.
15'
Fig. 10 Geometry for oscillating shock enhancement study.
channel at Mach 3, a pressure of 1 atm, and a temperature of 1500 K, and

the mixture was passed through a 10-deg shock produced by a lower wall
compression. A periodic oscillation was imposed on the flow near the lower
wall, and this resulted in a periodic oscillation of the shock wave. The
resulting pressure field over one period of the imposed oscillation is shown
in Fig. 11. A wave propagates along the shock. The oscillating shock
increases the level of turbulence in the flowfield, and the degree of tur-
bulence enhancement increases with a decreasing frequency of shock os-
cillation. Chemical reaction as defined by the equilibrium model has little
effect relative to the nonreacting results.
Several techniques for enhancing mixing and reaction in a model scramjet
combustor were studied in 1988 by Drummond and Mukunda. 109 They
considered the shock and expansion structure that existed in scramjet com-
bustors, and sought to redirect this existing wave structure to enhance
mixing. Planar and curved shocks were created by structure and fuel in-
jection in scramjet engines, and so both shock shapes were considered as
candidates for mixing enhancement. Three cases were studied to determine
the efficiency of shock shapes as shown in Fig. 12. Case 1 shows a two-
dimensional spatially developing supersonic mixing layer that served as a
baseline to indicate the degree of mixing and reaction that can be achieved
without enhancement. Here, gaseous hydrogen and fuel and air are sep-
arated initially by a thin splitter plate. Hydrogen and air both enter at
Mach 2, a pressure of 1 atm, and a temperature of 2000 K. The hydrogen
and air begin mixing at the trailing edge of the splitter plate, ignition occurs
at some small distance downstream, and combustion continues from that
point. Two enhancement studies, cases 2 and 3, are also shown in Fig. 12.
In case 2, the mixing layer is processed through two 10-deg planar shocks
produced by wedges in the fuel and air streams. In case 3, a small inter-
ference body that produces a curved bow shock is placed at the center of
the mixing layer. The flowfield in each case was then simulated using the
J. P. DRUMMOND 399
Fig. 11 Pressure contours from shock enhancement study showing shock motion
over a period with reaction.
SPARK code. Chemistry was modeled using either a four-species, one-

reaction or a nine-species, eighteen-reaction H2-air finite-rate chemistry
model. The resulting instantaneous mixing efficiencies shown as a function
of the streamwise coordinate are given in Fig. 13.109 The degree of mixing
achieved in the mixing layer without enhancement is shown by the result
from case 1. Very little additional mixing results from passing the layer
through planar shocks in case 2. Case 3 exhibits a significantly higher degree
of mixing. When the high-velocity gradient across the mixing layer is proc-
essed by the curved bow shock, vorticity is produced, and the mixing
subsequently is enhanced downstream of the shock.
With a background of planar and curved shock mixing enhancement in
hand, fuel injection configurations typical of those used in a scramjet engine
were considered. A typical design is described in Fig. 14a, which shows
the trailing edge of a NASA Langley fuel injection strut. Inlet air crosses
the strut as shown. Gaseous hydrogen fuel is injected parallel to inlet air
Air
M=2
T = 2000 K
Fig. 12a Sketch of supersonic reacting mixing layer in case 1.
M=2
T = 2000
Air
M=2
T = 2000 K
Fig. 12b Sketch of supersonic reacting mixing layer interacting with two shocks
in case 2 mixing enhancement study.
M=2
T = 2000K |
Air
M=2
T = 2000 K
Fig. 12c Sketch of supersonic reacting mixing layer interacting with curved shock
in case 3 mixing enhancement study.
from the base of the strut and transverse to the inlet air behind a small
rearward-facing step in the surface of the strut. The step caused a small
recirculation region to form that captured a small amount of fuel and air.
Combustion then occurred in that region providing a flameholding sight
in the engine combustor.
The results of the previous analysis suggested a simple change in the
fuel injector configuration that might provide improved fuel-air mixing and
J. P. DRUMMOND 401
.25
—— Case 1
.20 — - - Case 2
— - CaseS
.15
Mixing
efficiency
.10
.05
.02 .04 .06 .08 .10

Axial location, M
Fig. 13 Mixing efficiency vs streamwise coordinate from mixing enhancement study
for cases 1-3.
enhanced combustion efficiency. If the parallel injector from the strut base
were moved to the step face as shown in Fig. 14a, then the high-velocity
gradient of that jet would interact with the curved bow shock that formed
ahead of the transverse injector. The mixing enhancement modification
was then examined by again simulating the two-dimensional flowfields.
Chemical reaction was described by the nine-species, eighteen-reaction
finite-rate model. The resulting water mass fraction contours downstream
of the modified strut for only the parallel injector and for both the parallel
and transverse injector are given in Figs. 14b and 14c, respectively. The
multiple-jet configuration clearly gave significantly higher levels of jet de-
velopment, mixing, and reaction. A quantitative assessment of mixing and
reaction is given in Fig. 15. In that figure, the combustion efficiency for
both cases is plotted against the streamwise coordinate. Jet interaction
produced a significantly faster increase in mixing and combustion. Instan-
taneous combustion efficiencies of about 80% were obtained for the en-
hanced case within the first 40% of the solution length, whereas the
unenhanced case required about 75% of the solution length to achieve the
same level of reaction. Therefore, enhancement by such approaches could
result in shorter combustor lengths for a given level of performance.
While the basic studies we have discussed were under way, several in-
teresting three-dimensional scramjet combustor simulations were under-
taken in 1987 and 1988. Uenishi et al.58-60 used their three-dimensional
combustor code, described earlier, to study several generic combustor con-
figurations that utilized wall fuel injection. One of the interesting cases is
shown in Fig. 16. This combustor model was also being used in an exper-
imental program at NASA Langley when the study was undertaken, al-
though no data were yet available. Vitiated hot air, produced by the
combustion of hydrogen, oxygen, and air in a facility heater, entered the
model combustor from the left. Gaseous hydrogen fuel was injected initially
a small distance downstream from this point through a row of secondary
Air
Old
strut
Air
H 2 +N 2
Modified
strut
Fig. 14a Conventional and modified

strut configurations for improved fuel-air
mixing.
Fig. 14b Water mass fraction contours downstream of strut with only parallel fuel
injection.
Fig. 14c Water mass fraction contours downstream of strut with interacting parallel
and transverse fuel injection.
J. P. DRUMMOND 403
Parallel
1.0 - - - Parallel and transverse
.8 A..;,
Combustion .6
efficiency
.4
.2
.02 .04 .06 .08 .10

Axial location, M
Fig. 15 Combustion efficiency vs streamwise coordinate for fuel injection strut.
orifice fuel injectors located on opposite walls of the combustor. These

injectors were intended to pilot a pair of primary hydrogen fuel injectors
again located on opposite walls some distance downstream of rearward-
facing steps in the walls. Hydrogen from the secondary orifices was injected
at Mach 1.0, a temperature of 250 K, and a pressure of 2.1 atm. Hydrogen
from the primary orifices was injected at Mach 1.0, a temperature of 250
K, and a pressure of 12.0 atm. The flow in the combustor was modeled
using the Uenishi code.60 Features of the combustor flowfield resulting
H2 upstream injectors
H2 Primary fuel
Vitiated injector
hot air
Side view Top view

t 1 lh
H-.X2 WX2
1
——I f h t
JE——JC————————Jt 4 J
L L
l 2 L L LL.
l 2 3
Injector diameters: c^ = ,132 cm, d2 = ,304 cm
Step height: h = ,381 cm
H! = 5h LL = 3,41h L2 = 12,33h l_3 = 9,67h
La = 5,77 h S = 3,33h W = 6,66h
Fig. 16 Sketch of model three-dimensional combustor showing perspective, side
view, and top view with dimensions.
from that study are shown in Fig. 17. Figure 11 a shows the distribution of
the injected fuel given as the mass fraction of total H2 in the streamwise
x-z plane of the computation along the duct centerline. The total H2 dis-
tribution represented the sum at each point of the hydrogen atoms in any
form, i.e., H2, H2O, or OH, so that it indicated the amount of injected
H2 fuel. The fuel-rich regions in the vicinity of the injectors are evident
as is the penetration of the fuel across the duct. The contours of total H2
in Fig. 17a show that the upstream fuel injectors result in relatively well
mixed fuel and air in the downstream region. This mixture then augments
the fuel mixing and combustion process produced by the downstream in-
jectors.60
The three dimensionality of the flowfield is shown by the distributions
of total H2 and static temperature given in Fig. 17b at several cross planes
along the combustor length. The static temperature was normalized by the
initial static temperature. The locations of the y-z cross planes are indicated
in Fig. 17a. The distribution of total H2 in each cross plane indicates the
mixing of injected fuel. At location A, the fuel jets from the upstream
injectors have not merged, and the cold fuel has not mixed with the hot-
air stream to an extent sufficient to result in any significant heat release,
as is indicated by the absence of any appreciable temperature rise at this
location. At location B, however, the fuel has mixed and reacted with the
hot-air stream to an extent that the temperature is about 1.7 times the
airstream. Further downstream, at locations C and D, the 0.5% contour
line becomes more uniform and a large area of the flow contains reacting
fuel as also indicated by the elevated temperatures. Results from location
E indicate that some fuel from the downstream injector is entrained up-
stream by the recirculating flow behind the step. This upstream entrainment
of H2 from the downstream injector is also indicated by the regions of high
and low temperature at location E. The fuel-rich regions have a lower
temperature because the H2 is cold and because the fuel-air mixture is too
rich to react. At locations F and G, which are downstream of the primary
fuel injector, large regions of unreacted fuel can be seen. With the present
flow conditions, the H2 from the primary injectors penetrates very well,
passing beyond the combustor centerline.60
Calculations are continuing with the Uenishi code to model complex
three-dimensional combustor flowfields. A major effort is now also under
way to compare the code with additional experimental data that have
recently become available. Work is also currently being completed to cou-
ple the spatially elliptic Uenishi code to several PNS programs, including
the Chitsomboon and Kamath PNS codes, that were described earlier in
this chapter. PNS codes can be more efficient for simulating the combustor
far field well away from the fuel injectors where the flow is spatially elliptic
due to flow separation. Therefore, there is an incentive, based on com-
putational efficiency, to switch to PNS programs wherever possible in the
combustor.
A number of interesting combustor calculations have been made by
Gielda et al. using their three-dimensional PNS program, SSCPNS, also
described earlier in this chapter. One configuration that they analyzed is
given in Fig. 18, which shows a sketch of a generic hypersonic propulsion
injector
-12,3 h H2 14 h
injector
9,7h
Fig. 17a Total H2 mass fraction in jc-z plane along center of combustor.
Total H2 Normalized
(mass percent) static temperature
Z ,x
Flow Flow
Fig. 17b Total H2 mass fraction and static temperature in y-z planes.
405
Fuel injector pattern
injection region
Fig. 18 Sketch of three-dimensional generic hypersonic propulsion system including

an inlet, combustor, and nozzle.
system including an inlet, combustor, and nozzle.69 Conditions entering

the inlet were determined by assuming that the hypersonic vehicle was
cruising at Mach 16 and at an altitude resulting in a freestream temperature
and pressure of 452°R and 0.004 atm, respectively. Gaseous H2 fuel was
injected through five injectors into the engine, beginning at a location xl
L = 0.625 and continuing for a distance of xlL = 0.24 along the combustor,
as indicated by Fig. 18. This injection schedule resulted in a fuel equivalence
ratio of approximately 3. The fuel was injected at an angle of 25 deg to
the engine cowl defined in the figure by the x axis.
Once the engine geometry and fuel injection configuration had been
defined, the entire engine flowfield was simulated using the SSCPNS code.
The computed flowfield was highly three dimensional. The three dimen-
sionality can be seen in Fig. 19, which shows Mach number contours at a
number of cross planes along the length of the inlet, combustor, and en-
closed portion of the nozzle. The study of Gielda et al. also showed that
there were significant losses associated with the fuel injection process in
the generic engine. The losses resulted from the formation of strong oblique
shocks as the inlet air interacted with the injected fuel. The induced shock
waves can be seen in Fig. 20, which shows the computed pressure contours
in the engine. The entropy increase that resulted when the engine flow
passed through these oblique shocks caused a significant loss in the average
stagnation pressure of the flow. This loss in stagnation pressure ultimately
would result in a loss in the overall thrust that could be achieved. The fuel
injector configuration did, however, result in good fuel-air mixing and
combustion. The degree of combustion is indicated by the computed water
mass fraction contours shown in Fig. 21 at several cross planes along the
combustor and nozzle lengths. Water mass fractions of up to 0.23 resulted
across a large extent of the cross planes, and an oxygen utilization of
approximately 0.80 was achieved. Normalized static temperature contours
in the engine resulting from the simulation are also shown in Fig. 21.69
To improve the performance of the engine, Gielda and colleagues further
investigated the fuel injection process in an attempt to lower the associated
shock losses. Three numerical experiments were performed that included
a study of the engine flowfield without injection or reaction, with injection,
and with injection and reaction. By comparing the three cases, Gielda et
al. were able to assess the losses associated with the fuel injection process
and with the heat addition resulting from the combustion process.69 They
J. P. DRUMMOND 407
Fig. 19a Mach number contours at 13 stations along the inlet, combustor, and
nozzle.
Fig. 19b Mach number contours at 13 stations along the inlet, combustor, and
nozzle.
concluded that further investigation into the injection process was required,
and that such a study would require simulation of the flowfield near the
fuel injectors using the spatially elliptic form of the Navier-Stokes equa-
tions. Such a study has been undertaken recently.
Carpenter and Kamath used the SPARK3D spatially elliptic Navier-
Stokes program, described earlier, to model the generic scramjet com-
bustor sketched in Fig. 22.61 Flow processed by an inlet enters the com-
bustor from the left in the figure. The flow exits the inlet at a velocity of
1500 m/s, a temperature of 1000 K, and a pressure of 0.5 atm. That flow
then passes over rearward-facing steps on the walls, experiencing a mild
area expansion. Gaseous H2 fuel is injected at Mach 1.0, a temperature
of 250 K, and a pressure of 2.0 atm from a line of four orifice fuel injectors
that lie along the face of the step on each of the four engine walls. The
fuel and inlet air begin mixing just downstream of the steps. The flow for
this configuration was simulated by Carpenter and Kamath using the
SPARK3D code.61 Because the combustor was symmetric about its center,
only the lower quadrant, shown in Fig. 22, was modeled. In addition, even
J. P. DRUMMOND 409
H H
Fig. 19c Mach number contours at 13 stations along the inlet, combustor, and
nozzle.
though the program contained a generalized finite-rate chemistry model,

only fuel-air mixing was considered at this stage of the study. The resulting
H2 mass fraction distributions at 2 and 5 cm downstream of the step are
given in Fig. 23. At the first station, recircuiation behind the step and
diffusion spread the hydrogen toward the walls, but there is little penetra-
tion of the fuel into the core flow in the center of the combustor. The 0.2
mass fraction contour does not extend beyond the height of the step. Fuel-
rich regions having contours with a mass fraction greater than 0.2 were
suppressed for clarity. Further downstream at the 5-cm station, hydrogen
mixes to a greater extent with the air, and the mixture extends further into
the core flow. The extent of mixing is still not sufficient to provide an
acceptable level of mixing efficiency, or, for that matter, combustion ef-
ficiency if reaction had been allowed. As a result of these findings, mixing
enhancement strategies similar to those discussed earlier currently are being
Fig. 20 Nondimensional static pressure contours in the combustor and nozzle.
applied to the generic combustor configuration to achieve an acceptable

level of combustion efficiency.
Computational methods clearly have been established as valuable tools
for simulating combustor flows and providing an improved understanding
of their complex physics. With expected improvements in algorithms, com-
puter resources, and creativity, the situation should only improve, and
computational methods should take on an even greater role in studying
and designing high-speed propulsion systems.
J. P. DRUMMOND 411
Fig. 21a Nondimensional static temperature contours in the combustor and nozzle.
VI. Concluding Remarks

Significant progress has been made in the past 20 years to develop tech-
niques for modeling supersonic reacting flows. The progress has been due
both to advances in numerical methods for computing reacting flowfields
as well as to an appreciable growth in computer speed and storage. The
improvements in both algorithms and computer power have also led to a
Fig. 21b Water mass fraction contours in the combustor and nozzle.
gradually improved understanding of the physics of reacting flows, which

is so very important if computation is to have a significant impact on
scram jet combustor design. The national program to develop a transat-
mospheric vehicle has been quite effective in generating a renewed interest
in supersonic combustion, and it has spawned several projects aimed at
J. P. DRUMMOND 413
Fig. 22a Sketch of three-dimensional generic scramjet combustor.
L = 20.0, W = 10.0 cm
S = 2.0cm, L = 1.5 cm, D = 3.5 mm
Fig. 22b Lower quadrant of generic scramjet combustor with dimensions.
Fig. 23a H2 mass fraction 2 cm downstream of combustor step.
Fig. 23b H2 mass fraction 5 cm downstream of combustor step.
better understanding its complexities. While these programs have begun

to unravel some of the mysteries, they have also shown us that supersonic
reacting flows still contain many unsolved questions. These questions pro-
vide exciting opportunities for future research.
The efficient mixing and combustion of fuel and air in a scram jet com-
bustor currently remains a critical issue. We are just beginning to under-
stand the phenomena that suppress mixing in these high-Mach-number
devices, but a great deal of work remains to be done. Once the mixing
question is answered, many opportunities will still remain to employ this
improved physical understanding to produce an efficient mixing strategy.
Ground-based experimental facilities will be used to study and design scramjet
combustors by providing flows consistent with those that the engine will
ingest over its operating envelope. Ground-based facilities are able to
create only continuous-flow conditions up to a flight Mach number of about
J. P. DRUMMOND 415
8, however. Beyond Mach 8, the only options available to the experimen-

talist are pulse facilities that create flight conditions for only a short period
of time, or the very vehicle that the national program seeks to build.
Numerical methods provide an alternative for studying flows beyond the
Mach 8 barrier, but only if they are applied properly to the problem. For
proper application of these methods, a fundamental understanding of the
physical phenomena present in high-speed reacting flows must be achieved.
Even with this understanding, we still will be required to simulate flowfields
in complex geometric configurations, necessitating highly efficient numer-
ical algorithms and significant advances in computer technology. There is
certainly work here for everyone. The problems are difficult ones, but they
can be solved if we continue to tackle them in a careful and dedicated way.
The United States committed itself to landing men on the moon in 1961,
and it succeeded through a dedicated national commitment to do so in only
eight short years. The national program to build and fly a transatmospheric
vehicle is not a more difficult goal to achieve. It simply requires another
national commitment to succeed, and a willingness on our part to accept
the exciting challenges provided by such a program.
References
Terri, A., "Possible Directions of Future Research in Air Breathing Engines,'1
Fourth AGARD Colloquium on Combustion and Propulsion, edited by A. L.
Jaumotte, A. H. Lefebre, and A. M. Rothrock, AGARD, Brussels, Belgium,
1960.
2
Digger, G. L., "Comparison of Hypersonic Ramjet Engines with Subsonic and
Supersonic Combustion," Fourth AGARD Colloquium on Combustion and Pro-
pulsion, edited by A. L. Jaumotte, A. H. Lefebre, and A. M. Rothrock, AGARD,
Brussels, Belgium, 1960.
3
Northam, G. B. and Anderson, G. Y., "Supersonic Combustion Ramjet Re-
search at Langley," AIAA Paper 86-0159, Jan. 1986.
4
White, M. E., Drummond, J. P., and Kumar, A., "Evolution and Application
of CFD Techniques for Scramjet Engine Applications," Journal of Propulsion and
Power, Vol. 3, No. 5, Sept.-Oct. 1987, pp. 423-439.
5
Walthrup, P. J. and Billig, F. S., "Liquid Fueled Supersonic Combustion Ramjets:
A Research Perspective of the Past, Present, and Future," AIAA Paper 86-0158,
Jan. 1986.
6
Billig, F. S., Waltrup, P. J., and Stockbridge, R. D., "Integral-Rocket Dual-
Combustion Ramjets: A New Propulsion Concept," Journal of Spacecraft and
Rockets, Vol. 17, Sept.-Oct. 1980, pp. 416-424.
7
Ferri, A., "Mixing Controlled Supersonic Combustion," Annual Review of Fluid
Mechanics, Vol. 5, 1973, pp. 301-338.
8
Moretti, G., "Analysis of Two-Dimensional Problems of Supersonic Combus-
tion Controlled by Mixing," AIAA Journal, Vol. 3, Feb. 1965, pp. 223-229.
9
Edelman, R. and Weilerstein, G., "A Solution of the Inviscid-Viscid Equations
with Applications to Bounded and Unbounded Multicomponent Reacting Flows,"
AIAA Paper 69-83, Jan. 1969.
10
Dash, S. M., "An Analysis of Internal Supersonic Flows with Diffusion, Dis-
sipation, and Hydrogen-Air Combustion," NASA CR-111783, May 1970.
H
Dash, S. M. and DelGuidice, P. D., "Analysis of Supersonic Combustion
Flowfields with Embedded Subsonic Regions," NASA CR-112223, Nov. 1972.
12
Elghobashi, S. E. and Spalding, D. B., "Equilibrium Chemical Reaction of
Supersonic Hydrogen-Air Jets," NASA CR-2725, 1977.
13
Spalding, D. B., Launder, B. E., Morse, A. P., and Maples, G., "Combustion
of Hydrogen-Air Jets in Local Chemical Equilibrium," NASA CR-2407, 1974.
14
Markatos, N. C., Spalding, D. B., and Tatchell, D. G., "Combustion of Hy-
drogen Injected into a Supersonic Airstream," NASA CR-2802, 1977.
15
Patanker, S. V. and Spalding, D. B., "A Calculation Procedure for Heat,
Mass, and Momentum Transfer in Three-Dimensional Parabolic Flows," Inter-
national Journal of Heat and Mass Transfer, Vol. 8, No. 15, 1972, pp. 1787-1806.
16
Evans, J. S. and Schexnayder, C. J., "Critical Influence of Finite Rate Chem-
istry and Unmixedness on Ignition and Combustion of Supersonic H2-Air Streams,"
17
Dash, S. M., Sinha, N., and York, B. J., "Implicit/Explicit Analysis of Inter-
active Phenomena in Supersonic Chemically Reacting Mixing and Boundary Layer
Problems," AIAA Paper 85-1717, July 1985.
18
MacCormack, R. W., "The Effect of Viscosity on Hypervelocity Impact Cra-
tering," AIAA Paper 69-354, April 1969.
19
Briley, W. R. and McDonald, H., "Solution to the Multi-Dimensional Com-
pressible Navier-Stokes Equations," Journal of Computational Physics, Vol. 24,
1977, pp. 372-397.
20
Beam, R. and Warming, R. F., "An Implicit Factored Scheme for the Com-
pressible Navier-Stokes Equations," AIAA Journal, Vol. 16, April 1978, pp. 393-
402.
21
Drummond, J. P., "Numerical Solution for Perpendicular Sonic Hydrogen
Injection into a Ducted Supersonic Airstream," AIAA Journal, Vol. 17, May 1979,
pp. 531-533.
22
Drummond, J. P., "Numerical Investigation of the Perpendicular Injector Flow
Field in a Hydrogen Fueled Scramjet," AIAA Paper 79-1482, June 1979.
23
Drummond, J. P. and Weidner, E. H., "A Numerical Study of Candidate
Transverse Fuel Injector Configurations in the Langley Scramjet Engine," Pro-
ceedings of Seventeenth JANNAF Combustion Meeting, Vol. 1, Chemical Propul-
sion Information Agency, Laurel, MD, 1980, pp. 561-586.
24
Drummond, J. P. and Weidner, E. H., "Numerical Study of a Scramjet Engine
Flow Field," AIAA Journal, Vol. 20, Sept. 1982, pp. 1182-1187.
25
Weidner, E. H. and Drummond, J. P., "Numerical Study of Staged Fuel
Injection for Supersonic Combustion," AIAA Journal, Vol. 20, Oct. 1982, pp.
1426-1431.
26
Schetz, J. A., "Turbulent Mixing of a Jet in a Co-Flowing Stream," AIAA
Journal, Vol. 6, Oct. 1968, pp. 2008-2010.
27
Schetz, J. A., Billig, F. S., and Favin, S., "Analysis of Mixing and Combustion
in a Scramjet Combustor with a Coaxial Fuel Jet," AIAA Paper 80-1256, June
1980.
28
Schetz, J. A., Billig, F. S., and Favin, S., "Flowfield Analysis of a Scramjet
Combustor with a Coaxial Fuel Jet," AIAA Journal, Vol. 20, Sept. 1982, pp. 1268-
1274.
29
Griffin, M. D., Billig, F. S., and White, M. E., "Applications of Computational
Techniques in the Design of Ramjet Engines," 6th International Symposium on
Air Breathing Engines, AIAA, New York, 1983, pp. 215-228.
30
Williams, F. A., Combustion Theory, Addison-Wesley, Reading, MA, 1965,
pp. 358-429.
31
McBride, B. J., Heimel, S., Ehlers, J. G., and Gordon, S., Thermodynamic
Properties to 6000 K for 210 Substances Involving the First 18 Elements, NASA
SP-3001, 1963.
J. P. DRUMMOND 417
32
Kanury, A. M., Introduction to Combustion Phenomena, Gordon & Breach,
New York, 1982, pp. 363-371.
33
White, F. M., Viscous Fluid Flow, McGraw-Hill, New York, 1974, pp. 28-36.
34
Suehla, R. A., "Estimated Viscosities and Thermal Conductivities of Gases at
High Temperature," NASA TR R-132, 1962.
35
Wilke, C. R., "A Viscosity Equation for Gas Mixtures," Journal of Chemistry
and Physics, Vol. 18, No. 4, 1950, pp. 517-519.
36
Berman, H. A., Anderson, J. D., and Drummond, J. P., "Supersonic Flow
Over a Rearward Facing Step with Transverse Nonreacting Hydrogen Injection,"
AlAA Journal, Vol. 21, Dec. 1983, pp. 1701-1713.
37
Kee, R. J., Warnatz, J., and Miller, J. A., "A Fortran Computer Code Package
for the Evaluation of Gas-Phase Viscosities, Conductivities, and Diffusion Coef-
ficients," Sandia National Labs., Livermore, CA, Rept. SAND83-8209, March
1983.
38
Cebeci, T. and Smith, A. M. O., Analysis of Turbulent Boundary Layers,
Academic, New York, 1974.
39
Baldwin, B. S. and Lomax, H., "Thin Layer Approximations and Algebraic
Model for Separated Turbulent Flows," AIAA Paper 78-257, Jan. 1978.
40
Jones, W. P. and Launder, B. E., "The Prediction of Laminarization with a
Two-Equation Model of Turbulence," International Journal of Heat and Mass
Transfer, Vol. 15, 1972, p. 301.
41
Fabris, G., Harsha, P. T., and Edelman, R. B., "Multiple-Scale Turbulence
Modeling of Boundary Layer Flows for Scramjet Applications," NASA CR-3433,
1981.
42
Hanjalic, K. and Launder, B. E., "Sensitizing the Dissipation Equation to
Irrotational Strains," Journal of Fluids Engineering, Transactions of ASME, Vol.
102, 1980, pp. 34-40.
43
Hanjalic, K., Launder, B. E., and Schiestel, R., "Multiple-Time Scale Concepts
in Turbulent Transport Modeling," Second Symposium on Turbulent Shear Flows,
Imperial College, London, England, 1979, pp. 10.31-10.36.
44
Rodi, W., "The Prediction of Free Turbulent Boundary Layers by Use of a
Two-Equation Model of Turbulence," Ph.D. Thesis, University of London,
England, 1972.
45
Sindir, M. M., "Numerical Study of Turbulent Flows in Backward-Facing Step
Geometries; Comparison of Four Models of Turbulence," Ph.D. Thesis, Univ. of
California, Davis, June 1982.
46
Riley, J. J., Metcalfe, R. W., and Orszag, S. A., "Direct Numerical Simulations
of Chemically Reacting Turbulent Mixing Layers," Physics of Fluids, Vol. 29, No.
2, 1986, pp. 406-422.
47
Riley, J. J. and Metcalfe, R. W., "Direct Numerical Simulations of Chemically
Reacting Turbulent Mixing Layers," AIAA Paper 85-0321, Jan. 1985.
48
McMurtry, P. A., Jou, W.-H., Riley, J. J., and Metcalfe, R. W., "Direct
Numerical Simulations of a Reacting Mixing Layer with Chemical Heat Release,"
AIAA Journal, Vol. 24, June 1986, pp. 962-970.
49
Erlebacher, G. and Hussaini, M. Y., "Stability and Transition in Supersonic
Boundary Layers," AIAA Paper 87-1416, June 1987.
50
Schumann, U., "Subgrid Scale Model for Finite Difference Simulations of
Turbulent Flows in Plane Channels and Annuli," Journal of Computational Physics,
Vol. 18, 1975, pp. 376-404.
51
Speziale, C. C., Erlebacher, G., Zang, T. A., and Hussaini, M. Y., "On the
Subgrid-Scale Modeling of Compressible Turbulence," NASA CR-178420, 1987;
also ICASE Rept. 87-73, 1987.
52
Bussing, T. R. A. and Murman, E. M., "A Finite Volume Method for Cal-
culation of Compressed Chemically Reacting Flows," AIAA Paper 85-0311, Jan.
1985.
53
Widhopf, G. F. and Victoria, K. J., "On the Solution of the Unsteady Navier-
Stokes Equations Including Multicomponent Finite Rate Chemistry," Computers
and Fluids, Vol. 1, 1973, pp. 159-184.
54
Drummond, J. P., Rogers, R. C., and Hussaini, M. Y., "A Detailed Numerical
Model of a Supersonic Reacting Mixing Layer," AIAA Paper 86-1427, June 1986.
55
Uenishi, K. and Rogers, R. C., "Three-Dimensional Computation of Mixing
of Transverse Injectors in a Ducted Supersonic Airstream," AIAA Paper 86-1423,
June 1986.
56
Kumar, A., "Numerical Analysis of the Scramjet Inlet Flow Field Using the
Three-Dimensional Navier-Stokes Equations," Chemical Propulsion Information
Agency, CPIA Publ. 373, 1983, pp. 25-40.
57
Kumar, A., "Numerical Simulation of Flow Through Scramjet Inlet Using a
Three-Dimensional Navier-Stokes Code," AIAA Paper 85-1664, July 1985.
58
Uenishi, K., Rogers, R. C., and Northam, G. B., "Three-Dimensional Com-
putations of Transverse Hydrogen Jet Combustion in a Supersonic Airstream,"
59
Rogers, R. C. and Chinitz, W., "Using a Global Hydrogen-Air Combustion
Model in Turbulent Reacting Flow Calculations," AIAA Journal, Vol. 21, April
1983, pp. 586-592.
60
Uenishi, K., Rogers, R. C., and Northam, G. B., "Three-Dimensional Nu-
merical Predictions of the Flow Behind a Rearward-Facing Step in a Supersonic
Combustor," AIAA Paper 87-1962, June 1987.
61
Carpenter, M. H. and Kamath, H., "Three-Dimensional Extensions to the
SPARK Combustion Code," NASA-Langley, Hampton, VA, NASP CP-5029,
Paper 15, Oct. 1988, pp. 107-134.
62
Carpenter, M. H., "A Generalized Chemistry Version of SPARK," NASA
CR-4196, 1988.
63
Chitsomboon, T. and Northam, G. B., "A 3-D PNS Computer Code for the
Calculation of Supersonic Combusting Flows," AIAA Paper 88-0438, Jan. 1988.
64
Chitsomboon, T., "Numerical Study of Hydrogen-Air Supersonic Combustion
by Using Elliptic and Parabolized Equations," Ph.D. Dissertation, Old Dominion
Univ., Norfolk, VA, May 1986.
65
Chitsomboon, T., Kumar, A., and Tiwari, S. N., "Numerical Study of Finite-
Rate Supersonic Combustion Using Parabolized Equations," AIAA Paper 87-0088,
Jan. 1987.
66
Vigneron, Y. C., Rakich, J. V., and Tannehill, J. C., "Calculation of Supersonic
Viscous Flow Over Delta Wings with Sharp Subsonic Leading Edges," AIAA Paper
78-1137, July 1978.
67
Schiff, L. B. and Steger, J. L., "Numerical Simulation of Steady Supersonic
Viscous Flow," AIAA Paper 79-0130, Jan. 1979.
68
Gielda, T. and McRae, D., "An Accurate, Stable, Explicit, Parabolized Navier-
Stokes Solver for High Speed Flows," AIAA Paper 86-1116, May 1986.
69
Gielda, T. P., Hunter, L. G., and Chawner, J. R., "Efficient Parabolized
Navier-Stokes Solutions of Three-Dimensional, Chemically Reacting Scramjet Flow
Fields," AIAA Paper 88-0096, Jan. 1988.
70
Boris, J. P., "A Fluid Transport Algorithm That Works," Computing as a
Language of Physics, International Atomic Energy Agency, Vienna, Austria, 1971,
pp. 171-189.
71
Boris, J. P. and Book, D. L., "Flux-Corrected Transport I: SHASTA—A
Fluid Transport Algorithm That Works," Journal of Computational Physics, Vol.
11, 1973, pp. 38-69.
J. P. DRUMMOND 419
72
Oran, E. S. and Boris, J. P., Numerical Simulation of Reactive Flow, Elsevier,
New York, 1987, pp. 264-298.
73
Zalesak, S. T., "Fully Multidimensional Flux-Corrected Transport Algorithms
for Fluids," Journal of Computational Physics, Vol. 31, 1979, pp. 335-362.
74
MacCormack, R. W., "Current Status of Numerical Solutions of the Navier-
Stokes Equations," AIAA Paper 85-0032, Jan. 1985.
75
Steger, J. L. and Warming, R. F., "Flux Vector Splitting of the Inviscid Gas-
dynamic Equations with Applications to Finite-Difference Methods," Journal of
Computational Physics, Vol. 40, 1981, pp. 263-293.
76
Candler, G. V. and MacCormack, R. W., "Hypersonic Flow Past 3-D Con-
figurations," AIAA Paper 87-0480, Jan. 1987.
77
Candler, G. V. and MacCormack, R. W., "The Computation of Hypersonic
Ionized Flows in Chemical and Thermal Nonequilibrium," AIAA Paper 88-0511,
Jan. 1988.
78
Grossman, B. and Walters, R. W., "An Analysis of Flux-Split Algorithms for
Euler's Equations with Real Gases," AIAA Paper 87-1117, June 1988.
79
van Leer, B., "Flux-Vector Splitting for the Euler Equations," Lecture Notes
in Physics, Vol. 170, 1982, pp. 507-512.
80
Roe, P. L., "Characteristic-Based Schemes for the Euler Equations," Annual
Review in Fluid Mechanics, Vol. 18, 1986, pp. 337-365.
81
Anderson, W. K., Thomas, J. L., and van Leer, B., "A Comparison of Finite-
Volume Flux Vector Splittings for the Euler Equations," AIAA Paper 85-0122,
Jan. 1985.
82
Walters, R. W. and Dwoyer, D. L., "An Efficient Iteration Strategy Based
on Upwind/Relaxation Schemes for the Euler Equations," AIAA Paper 85-1529-
CP, July 1985.
83
Grossman, B. and Cinnella, P., "The Development of Flux-Split Algorithms
for Flows with Nonequilibrium Thermodynamics and Chemical Reactions," AIAA
Paper 88-3596, July 1988.
84
Grossman, B. and Cinnella, P., "Flux-Split Algorithms for Flows with Non-
equilibrium Chemistry and Vibrational Relaxation," Virginia Polytechnic Inst. and
State Univ., Blacksburg, VA, ICAM Rept. 88-08-03, Aug. 1988.
85
Liu, Y. and Vinokur, M., "An Analysis of Numerical Formulations of Con-
servation Laws," NASA CR-177489, June 1988.
86
Liou, M. S., van Leer, B., and Shuen, J. S., "Splitting of Inviscid Fluxes for
Real Gases," Journal of Computational Physics, Vol. 87, March 1990, pp. 1-24.
87
Liou, M. S., "A Generalized Procedure for Constructing an Upwind-Based
TVD Scheme," AIAA Paper 87-0355, Jan. 1987.
88
Yee, H. C., "Construction of Explicit and Implicit Symmetric TVD Schemes
and Their Applications," NASA TM-86775, July 1985; also, Journal of Compu-
tational Physics, Vol. 68, 1987, pp. 151-179.
89
Yee, H. C. and Shinn, J. L., "Semi-Implicit and Fully Implicit Shock Capturing
Methods for Hyperbolic Conservation Laws with Stiff Source Terms," AIAA Paper
87-1116-CP, June 1987.
90
Yee, H. C., "Upwind and Symmetric Shock-Capturing Schemes," NASA TM-
89464, May 1987.
91
Shinn, J. L. and Yee, H. C., "Extension of a Semi-Implicit Shock-Capturing
Algorithm for 3-D Fully Coupled Chemically Reacting Flows in Generalized Co-
ordinates," AIAA Paper 87-1577, June 1987.
92
Gnoffo, P. A., McCandless, R. S., and Yee, H. C., "Enhancements to Program
LAURA for Computation of Three-Dimensional Hypersonic Flow," AIAA Paper
87-0280, Jan. 1987.
93
Gnoffo, P. A. and Green, F. A., "A Computational Study of the Flow Field
Surrounding the Aeroassist Flight Experiment Vehicle," AIAA Paper 87-1575,
June 1987.
94
Gnoffo, P. A., Gupta, R. N., and Shinn, J. L., "Conservation Equations and
Physical Models for Hypersonic Air Flows in Thermal and Chemical Nonequilib-
rium," NASA TP-2867, 1988.
95
Shuen, J. S. and Yoon, S., "Numerical Study of Chemically Reacting Flows
Using an LU Scheme," AIAA Paper 88-0436, Jan. 1988.
96
Yu, S. T. and Shuen, J. S., "Three-Dimensional Simulation of an Underex-
panding Jet Interacting with a Supersonic Cross Flow," AIAA Paper 88-3181, July
1988.
97
Abarbanel, S. and Kumar, A., "Compact High Order Schemes for the Euler
Equations," NASA CR-181625, Feb. 1988.
98
Hussaini,M. Y., Salas, M. D., andZang,T. A., "Spectral Methods for Inviscid,
Compressible Flows," Advances in Computational Transonics, edited by W. G.
Habashi, Pineridge Press, Swansea, England, 1983.
"Gottlieb, D. and Orszag, S. A., Numerical Analysis of Spectral Methods, The-
ory, and Applications, CBMS-NSF Regional Conference Series in Applied Math-
ematics, No. 26, Society of Industrial and Applied Mathematics, Philadelphia, PA,
1977.
100
Drummond, J. P., Hussaini, M. Y., and Zang, T. A., "Spectral Methods for
Modeling Supersonic Chemically Reacting Flowfields," AIAA Journal, Vol. 24,
Sept. 1986, pp. 1461-1467.
101
Drummond, J. P. and Hussaini, M. Y., "Numerical Simulation of a Supersonic
Reacting Mixing Layer," AIAA Paper 87-1325, June 1987.
102
Evans, J. S., Schexnayder, C. J., and Beach, H. L., "Application of a Two-
Dimensional Parabolic Computer Program to Prediction of Turbulent Reacting
Flows," NASA TP-1169, 1978.
103
Burrows, M. C. and Kurkov, A. P., "Analytical and Experimental Study of
Supersonic Combustion of Hydrogen in a Vitiated Airstream," NASA TM X-2828,
1973.
104
Brown, G. L. and Roshko, A., "On Density Effects and Large Structure in
Turbulent Mixing Layers," Journal of Fluid Mechanics, Vol. 64, Pt. 4, 1974, pp.
775-816.
105
Papamoschou, D. and Roshko, A., "Observations of Supersonic Free Shear
Layers," AIAA Paper 86-0162, Jan. 1986.
106
Oh, Y. H., "Analysis of Two-Dimensional Free Turbulent Mixing," AIAA
Paper 74-594, June 1974.
107
Hussaini, M. Y., Collier, F., and Bushnell, D. M., "Turbulence Alteration
Due to Shock Motion," Turbulent Shear-LayerlShock-Wave Interactions, edited by
J. Delery, Springer-Verlag, Berlin, New York, 1986, pp. 371-381.
108
Drummond, J. P., "A Two-Dimensional Numerical Simulation of a Super-
sonic, Chemically Reacting Mixing Layer," NASA TM-4055, 1988.
109
Drummond, J. P. and Mukunda, H. S., "A Numerical Study of Mixing En-
hancement in Supersonic Reacting Flow Fields," AIAA Paper 88-3260, July 1988.
110
Guirguis, R. H., Grinstein, F. F., Young, T. R., Oran, E. S., Kailasanath,
K., and Boris, J. P., "Mixing Enhancement in Supersonic Shear Layers," AIAA
Paper 87-0373, Jan. 1987.
111
Guirguis, R. H., "Mixing Enhancement in Supersonic Shear Layers: III. Effect
of Convective Mach Number," AIAA Paper 88-0701, Jan. 1988.
112
Kumar, A., Bushnell, D. M., and Hussaini, M. Y., "A Mixing Augmentation
Technique for Hypervelocity Scramjets," AIAA Paper 87-1882, June 1987.
Chapter 13
Studies of Detonation Initiation, Propagation,

and Quenching
E. S. Oran, J. P. Boris, and K. Kailasanath
Introduction
I N this chapter, we briefly summarize recent applications of numerical

simulation to the study of the initiation and propagation of gas-phase
detonations. These numerical studies of detonations are relatively recent
because they depend on the availability of high-speed computers and spe-
cialized numerical algorithms to solve the governing equations.
Numerical simulation can be used at a number of levels to study deto-
nations. These range from extremely basic studies of the fundamental
structure and properties of detonation waves to engineering design appli-
cations. For example, in more basic applications, a simulation might deter-
mine whether a particular energy deposited during a certain period of time
could trigger a detonation, or perhaps determine the structure of the trans-
verse waves at the detonation front and how they interact with an approach-
ing obstacle. In the more applied applications, a simulation could predict
the overpressures that accompany a detonation and whether a structure
could withstand such pressures.
A major attraction of using numerical simulations is their use as a vehicle
for looking at fundamental interactions in a systematic way. Thus, we can
isolate the major controlling processes, and study their interactions in
idealized environments. Once this is done, additional complexities can be
added. This summary considers some types of calculations that have been
done to help clarify the fundamental interactions between the fluid dynamic
and chemical processes occurring in a gas-phase detonation and shows how
simulations have been used to isolate and study the basic mechanisms.
Numerical simulations of gas-phase detonations are based on computer-
generated solutions of the compressible, time-dependent conservation
equations for total mass density p, momentum pi?, and energy E,
^ = - V - pPv v(1)
dt '
This paper is a work of the U.S. Government and is not subject to copyright protection
in the United States.
421
422 DETONATION INITIATION, PROPAGATION, AND QUENCHING
= - V • (pvv) - VP (2)
dE
— = - V • (Ev) - V • (vP) (3)
where v is the fluid velocity and P the pressure. Since we are dealing with
a multispecies fluid, in which chemical reactions affect transformations
among species, we also need individual species number densities {«,-}:
5 = - v'n'v + Qi- Ln> l

= i>-< N (4)
where {Q,} and {L,} are chemical production and loss terms, respectively,
for species /. There is a constraint that defines the total number density
N,
N=^nf (5)
i= 1
where Ns is the number of species present.

For purely gas-phase applications, we need the additional relations
E = 5 pv • v + pe (6)
where e is the specific internal energy. The equations that make up the
thermal and caloric equations of state are
P = NkT = pRT (7)
E pA- - P = ph - P (8)
T
+ t cpi dT (9)
J T()
Here k is Boltzmann's constant, {h^ the enthalpies of each species /, {hl()}

the heats of formation, and {cpl} the specific heats.
There are several additional relations:
R = Ro2-
i <*>,- (10)
where
y ^ p / p = ^-m/p (11)
and
co - (/?o/*) m, = NW (12)
where 7?0 is the gas constant and N0 is Avogadro's number.
Equations (1-12) are the multidimensional reactive gasdynamic equa-
tions. They describe not only the basic properties of detonations, but any
E. S. ORAN ET AL. 423
phenomena primarily controlled by interactions of convection and chem-

istry. Each particular situation is specified by a set of initial conditions,
boundary conditions, and input constants. With a sufficiently large chemical
energy release at a sufficiently fast rate, these equations describe a deto-
nation. (Note that viscous effects have been explicitly omitted, so that
these are the Euler equations and not the Navier-Stokes equations.)
The fundamental assumptions made in the computations presented in
this chapter are as follows:
1) The two basic physical processes are gas-phase convection and chem-
ical reactions with energy release. These are the only two processes needed
to describe basic detonation phenomena. The chemistry is treated in vary-
ing levels of detail.
2) We consider only ideal equations of state. Thus, the gas is thermally
perfect, but calorically imperfect.
3) Here we describe only one- and two-dimensional flows. Some recent
three-dimensional computations will be referenced later.
4) We omit discussion of other processes such as boundary layers or
small-scale inhomogeneities in the flow. In addition, a number of physical
effects usually important in combustion have been omitted. These include
molecular diffusion, thermal conduction, external forces, and electromag-
netic radiation. This has been done to simplify the discussion and concen-
trate on the minimal fundamental processes that can produce a detonation.
The other effects may be added later (see, for example, Ref. 1). Possible
effects of these processes are discussed in the last section.
Solving the preceding set of equations on a computer requires addressing
some difficult numerical and conceptual problems. For example, the order
of accuracy to which the convective parts of these equations are solved is
limited by the computer speed and memory available, and this affects the
representation of the important structures (such as shock fronts) in the
problem. There are also major limitations to solving the chemistry parts
of these equations, which involves solving sets of coupled ordinary differ-
ential equations representing the elementary chemical reactions. There are
always questions about the chemistry model based on limited knowledge
of the input quantities: pertinent chemical reactions, associated rate con-
stants of these, specific heats and heats of formation of constituent species,
and even which particular species are involved. The numerical problems
are discussed next, and the problems with physical and chemical inputs are
referred to throughout this chapter.
Numerical Considerations
Equations (1-12) are continuous, fluid representations of the convection
and chemistry processes. To solve these equations, they must be trans-
formed into a form suitable for solution on a digital computer. The most
common approach is to set up a discretized grid in time and space. Space
is divided into computational zones of finite size, and time is divided into
finite time intervals. Equations (1-12) are then transformed into a set of
coupled, algebraic equations. The exact form of these equations depends
on the particular algorithms chosen to represent the various physical pro-
cesses. This approach, called finite differences or finite volumes, is the most
versatile and efficient numerical approach and is, therefore, used most
frequently.
Solving for the convective flow requires solving compressible conser-
vation equations with a pressure source term. Solving for the chemical
kinetics and the resulting energy release requires solving sets of coupled,
nonlinear ordinary differential equations. There is an extensive body of
work on the numerical analysis of each type of problem individually. We
will review some of the relevant aspects of numerical solutions of each type
of process. Our emphasis throughout is on those methods best suited for
numerical simulations of detonations. Next we discuss coupling of these
solutions.
Coupling Convection and Chemistry

There are several distinct approaches to coupling the solution of sets of
coupled, time-dependent, nonlinear, partial differential equations. These
are the global implicit approach, the fractional-step or time-step splitting
method, and other general classes of techniques such as the method of
lines and finite-element methods. Each of these methods has a different
philosophy, although they do have common elements.
Time-step splitting is the approach most commonly used in time-
dependent combustion problems. This approach, described in formal detail
in Ref. 2 and in some detail in the context of reactive flows in Ref. 1,
requires independently solving individual collections of terms in the gov-
erning equations and then coupling these together. The various processes
and the interactions among them may be solved individually by very dif-
ferent methods. For example, time-step splitting can be used to couple the
asymptotic method described subsequently for chemical reactions to a mon-
otone method for convection. Often time-step splitting is used to calculate
the convection in different directions. The exact way in which the processes
are coupled depends on the individual properties of the different algorithms
used as well as the type of flow being modeled.
A major advantage of time-step splitting is that it results in modular
simulation models. Each type of term can be programmed using the best
available technique for that process. If a new algorithm becomes available,
it can be incorporated without changing the entire structure of the computer
program. Thus, convection, diffusion, equation-of-state calculations, and
chemical kinetics are all in separate packages. It is particularly important,
however, to test the coupling.
Convection Algorithms
There is a large existing body of literature and a great deal of practical
experience in solving idealized forms of the conservation equations. The
basic equation we need to solve has the form
f + V.F = 0 (13)
In the simplest case, i|/ is a scalar that could, for example, represent the
density, p. If we consider the components of the vector v|/ as density,
momentum, and energy, we have the coupled continuity equations describ-

ing compressible gasdynamics. The numerical solutions of these equations
are discussed in some detail in Refs. 3-5, but since these books were
written, there has been substantial progress in this area of computational
methods. Through the years, many techniques have been proposed for
solving continuity equations. Regardless of their accuracy, they are all in
common use even though the newer nonlinear monotone methods produce
superior solutions.
The fundamental problem in solving continuity equations numerically
becomes apparent when applying standard error analysis procedures to the
various algorithms proposed. This involves choosing an idealized problem
with an exact, analytic answer, choosing an algorithm, expanding the solu-
tions in a finite Fourier series, and comparing the analytic and numerical
predictions (see, for example, Ref. 5). This type of analysis shows that
algorithms for continuity equations have the possibility of three major types
of error:
1) Numerical diffusion or, its inverse, numerical instability. Numerical
diffusion tends to reduce nonphysically the amplitudes of the various Four-
ier modes. Numerical instability tends to increase them improperly.
2) Numerical dispersion. This changes the phases of the modes.
3) Gibbs errors. These are a result of using a finite representation to
describe continuous profiles. They produce spurious oscillation near a steep
gradient such as a shock or detonation.
These errors can be reduced by decreasing the computational cell sizes,
decreasing the computational time step, and choosing a method with a
high order of accuracy.
Figure 1 shows how several different methods reproduce the solution of
the continuity equation for the particularly simplified problem of the con-
vection of a one-dimensional square wave at constant velocity. 1 The exact
solution to the equation uniformly moves the profile along in time, as
shown by the solid lines.
The top panels show the results of using the one-sided donor-cell method.
This solution is first-order accurate and dominated by numerical diffusion.
It degrades in time, although the area under the curve is conserved. A
simulation of a detonation must be able to describe the propagation of a
discontinuity, such as a square wave or a shock front, as accurately as
possible. Thus, the donor-cell algorithm would not give acceptable results.
If enough computational cells were used, a donor-cell calculation would
be adequate for awhile, but the solution degrades in time.
The second and third sets of panels show two square-wave calculations
performed by using the Lax-Wendroff and leapfrog algorithms, respec-
tively. In these cases, the numerical diffusion is of second order, whereas
it was of first order in the donor-cell calculation of the top panel. Here
the amplitude errors resulting from numerical diffusion are less severe than
dispersion errors. Characteristic ripples are seen near the top and bottom
of discontinuities. Initially, they appear at the shortest wavelength and
look similar to Gibbs errors. The pronounced undershoots and overshoots
pose problems for coupled reactive-flow simulations. The phase- and Gibbs-
error problems are as serious as those caused by excess numerical diffusion.
t=20 cycle =100 t=J60 cycle = 800
ONE-
SIDED • P
• I .-
.•'
-IP ^
' ' l\^
(cell no.) ——- I0
° ° X (cell no.) 100
LAX-
WEND. * P
0 100
° X (cell no.) 100
X (cell no,)
LEAP-
FROG
I0 100
X (cell no.) ——- ° ° X (cell no.) ——-
FCT
0 IUU u 100
X ( c e l l no.)—— Xfcellno.) —
Fig. 1 Linear convection of a square wave using four different algorithms. Solid
lines are exact solutions; dotted lines are computed solutions.
To describe any combustion flow, we need to include the effects of energy

released into the fluid by chemical reactions. Because of the exponential
form these energy-release terms usually take, overshoots or undershoots
in the background fluid conditions, derived from the solutions to the con-
vection equations, cause severe problems when these solutions are coupled
to the chemistry. Small overshoots in the temperature can result in a large
energy release, which, in turn, increases the temperature. The results can
be a runaway effect in the calculations.
These problems are avoided to a large extent by using one of the non-
linear monotone methods. The last set of panels in the Fig. 1 calculation
uses the flux-corrected transport (FCT) algorithm.6 7 This particular vari-
ation of FCT was based on a modified Lax-Wendroff method with fourth-
order phase accuracy. Figure 2 shows how FCT solutions are generally
improved by increasing the phase accuracy. We have generally found that
fourth order is both necessary and sufficient to obtain a good solution for
shock and detonation problems.
2nd ORDER FCT

3.0 I—————————————I 2.0
0.0
4th ORDER FCT

3.0P————————————I 2.0
0.0
16th ORDER FCT

3.0—————————————I 2.0
0.0
Fig. 2 Linear convection of a square wave, a Gaussian, and a half-dome, performed

by FCT algorithms with three different phase accuracies1 (courtesy of S. Zalesak).
Another approach commonly used is to enforce the monotonicity prop-

erty by combining a low-order and a high-order method.8 One applica-
tion of this approach was carried out by Sigamura et al., 9 who used a
MacCormack method as the high-order method. Other approaches to mon-
otonicity are discussed in Ref. 1. The important property that any method
must have to simulate detonations is that it must resolve steep gradients
well throughout the course of the calculation. Numerical errors must be
minimized to predict accurately the correct detonation velocity and deto-
nation structure. The convection algorithm must provide the correct back-
ground conditions for the chemistry.
Solving the Chemistry Equations

As our knowledge of elementary chemical reaction rates has improved,
it has become possible to combine these to form detailed chemical mech-
anisms describing complex intermediate reaction pathways of a global reac-
tion. For example, consider
CrL 207 CO7 + 2H7O
This expresses the global process of oxidation of methane to produce car-

bon dioxide and water. However, this process may be decomposed into a
number of elementary reactions, describing the removal of a hydrogen
from the CH4 to produce methyl radical, the reaction of methyl radical
with molecular oxygen, and, finally, the production of CO2 and H2O.
The equations describing mechanisms for a set of chemical reactions,
such as methane oxidation or hydrogen oxidation, can be put into the form
^ = &: - LM, i = 1, ...,NS (14)
which are extracted from Eqs. (1-12). As discussed in an earlier chapter

of this book by Radhakrishnan, these equations model the time evolution
of species densities in a region of gas that is assumed to be homogeneous
and premixed. As they stand, these equations may be solved with stringent
physical restrictions, e.g., constant temperature, constant pressure, or con-
stant volume. From the solution of Eq. (14), a number of quantities can
be derived that may be compared directly with data. These include chemical
induction times, emission intensities from excited states of reacting species,
and bulk quantities such as overall activation energy and total energy
released.
The ordinary differential equations describing chemical kinetics are usu-
ally stiff equations. The property of stiffness has a number of definitions,
but the most practical is the easiest to understand: an equation is stiff if
solving it accurately requires time steps that are just too small to be afford-
able. There are a number of methods that can be used to solve stiff equa-
tions. For example, the implicit Runge-Kutta and backward-differentiation
methods are good, conservative methods with variable orders of accuracy,
and they are stable for large time steps. Most of the methods, however,
have several major computational drawbacks. First, they require inverting
matrices, which can be costly when there are many chemical species and
reactions among them. Second, the higher-order solutions require storing
information from previous time steps, which can produce major data-
storage problems in a multidimensional reactive-flow simulation. These
two problems complicate the use of these sophisticated high-order methods
for describing the chemical kinetics in reactive-flow problems, but they are
often excellent for the solution of Eq. (14) alone with no convective flow.
For simulating detonations, however, we are dealing with systems of
coupled chemical reaction equations in the more difficult context of the
full set of reactive-flow equations. This is really a rather different problem
from solving the chemical reaction equations alone. There are several
important points to note.
1) The reaction rate equations must be integrated for many grid points,
perhaps tens of thousands, for relatively short periods of time between
hydrodynamic time steps. Thus, we require self-starting techniques with
low overhead for stopping and starting the calculations.
2) The large number of computational cells in fluid calculations usually
precludes storing auxiliary information for the spatial gbrid points between
time steps. Since much of the computer time would be spent starting a
high-order method, using low-order methods is easier and faster.
3) Fluid dynamic errors are rarely smaller than a few percent at best,
therefore, it may not be necessary to solve the chemical reaction equations
to very high accuracy. Reaction rates are often poorly known, so that low-
order methods again are the logical choice over complicated high-order
approaches.
4) Reactive-flow codes are usually changed continually. A general inte-
gration technique is preferable to a complicated, highly specialized method.
5) The amount of computation involved in multidimensional reactive-
flow calculations is usually so vast that any evaluation of transcendental
functions should be avoided, no matter how good the integration algorithms
that require these functions. Whenever analytic solutions are to be approx-
imated, a polynomial or rational-function approximation should be con-
sidered instead.
These requirements and restrictions motivated our development of the
selected asymptotic integration method (SAIM). This method is actually
best used within the framework of reactive-flow simulations, although it
has been used extensively for integrating Eq. (14) alone. SAIM consists
of a hybrid integrator, CHEMEQ,10 n which solves nonstiff equations by
a classical method and solves the stiff equations by an asymptotic method.
An important part of the method is determining criteria both for the initial
step size and for determining which equations to treat as stiff. The method
is efficient, accurate, self-starting, stable, relatively low order, and does
not require matrix inversions. However, it is not inherently conservative,
so that how well it conserves becomes a measure of the accuracy of the
answer. Conservation is controlled by monitoring the solution and properly
choosing the step size.
Solving the Temperature Equation
Another equation, usually solved in conjunction with Eqs. (1-12), describes
the time rate of change of the temperature due to the chemical energy
released. This equation does not appear explicitly in the conservation equa-
tions, but can be derived from parts of the energy equation. Alternatively,
we can use the algebraic expression
8 - H(T,{n$) - P (15)
where H is the enthalpy of the system. Knowing the value of the enthalpy,
we can find an expression for dTldt. Since the expression for the temper-
ature derivative involves all of the rates of change of the densities, {drij/dt},
as well as expressions of the form {dhj/dt}, which may be expressed as
powers of T, solving for T may be very expensive computationally.
As is clear from Eq. (15), the enthalpy may also involve complicated
sums over excited states of the molecule. However, when each of the
species is in local thermodynamic equilibrium, the individual {ht(T}} can
be evaluated as a function of temperature and fit to a polynomial expansion.
This has been done for the JANNAF tables12 and in Ref. 13.
By using tabulated values of {hf(T)} and {hi0(T}}, the tedious temperature
integration can often be avoided completely.1 If, during each chemical time
step, the total internal energy of the system does not change but may be
redistributed among chemical states, we may then solve Eq. (15) iteratively
for a new temperature that is consistent with the new number densities
calculated. A simple Newton-Raphson iteration technique is usually suf-
ficient. When this is used in integration programs, it converges to five
figures of accuracy in one or two iterations.
Phenomenologies for Energy Release

The most sophisticated models of multispecies, gas-phase chemically
reacting systems involve solving sets of equations such as Eq. (14). The
production term Pt and the loss term L, are functions of the species densities
{n}} and the temperature. The ideal way to do the problem is to couple a
solution of these equations, consisting of elementary chemical reactions,
to solutions of the convective transport equations at each time step for
each computational zone. In fact, this has been done for many of the
reactive shock and detonation calculations described in the next section.
However, integrating the equations representing the chemical kinetics
is by far the most time-consuming and core-consuming parts of the cal-
culation. For example, there may be 9 or even 25 coupled equations for
H2 or CH4 combustion, respectively, and any number of production and
loss terms in the equations, each nominally involving the evaluation of an
exponential. Since this integration has to be performed at each time step
in each computational cell, the time to conduct the reactive-flow calculation
can be longer than the time for an unreacting calculation by factors of 4
or 20. There are new methods using table lookups to evaluate the expo-
nential factors,14 but these require large amounts of computer memory.
The cost of complete reactive-flow calculations has led us to develop
methods for representing the energy release and temperature change due
to chemical reactions without having to integrate the full set of differential
equations describing the elementary chemical reactions. There have now
been a number of approaches to developing such phenomenological models,
as discussed in an earlier chapter by Frenklach. One approach is to use a
model that describes energy release and transition of fuel and oxidizer to
products. An example of this is the induction parameter model.15 Another
approach is to use simplified chemistry models. For example, Sigamura et
al.9 use a two-step reaction model in which the first reaction is a ther-
moneutral induction reaction and the second step is an exothermic energy-
release reaction. Presumably, these models could become significantly more
complicated and yet still not be as complicated as a full reaction mechanism
consisting of elementary reactions. The basic idea in these approaches is
to create a phenomenology that puts energy into the flow with a prescription
that appears similar to chemical energy release.
Variable and Adaptive Gridding

Simulations of detonations require that we model relatively long systems,
perhaps on the order of meters, while we simultaneously maintain high
accuracy around the steep gradients at the detonation front. To put the
resolution in a calculation where it is needed and not waste computational
cells in regions where there is relatively little action, rather sophisticated
variable and adaptive gridding methods have been developed.
Numerical simulations of detonations require both variable and moving

grids. Figure 3 schematically shows the grid used in a one-dimensional
model of a shock-tube experiment.16 There are two finely gridded regions:
one surrounding the shock wave and the other surrounding the contact
surface. The fine-zoned region around the contact surface moves with the
contact surface at the fluid velocity. The region around the shock front
moves with the front. Each of these finely gridded regions may have a
different minimum computational cell size. The computational cells in the
regions ahead of the shock wave and behind the contact surface change
exponentially in size from the smallest cells near the shock wave or the
contact surface to the largest cells at the walls.
Figure 4 shows the computational grid used in two-dimensional calcu-
lations of a propagating detonation. 15 In this case, bands of fine zones
move with the detonation front, so that the front is always well resolved.
The change in zone size from the smallest to the largest computational
zone is about an order of magnitude. However, it is important in these
calculations to make sure that the size of the zones changes smoothly from
small to large.
Variable and adaptive gridding is a forefront research area in compu-
tational physics. It is currently receiving a lot of attention, and advances
are close at hand. One particularly attractive approach is a finite-element
method based on triangular or tetrahedral meshes. As they have been
developed to date, these methods allow variable and time-evolving com-
putational grids that cluster computational cells around the regions of steep
FEM-FCT gradients.17 An example of this approach, the Finite-Element
Method FCT (FEM-FCT) developed by Lohner et al., is shown in Fig. 5.
DC
ID
Q_
SHOCK
RAREFACTION CONTACT SURFACE FRONT AMBIENT GAS
EXPONENTIAL FINE FINE EXPONENTIAL

SPACING ZONES ZONES SPACING
POSITION
Fig. 3 Schematic representation of adaptive grid used for numerical solution of an
incident shock in a shock tube. Chemical reactions initiate in the high-temperature
region near the contact surface, and a reaction wave propagates down to meet the
shock front and forms a detonation.
9 11!III
STEP 200 300 400 500 600
b) COMPUTATIONAL GRID
X
LOCATION OF FRONT
Fig. 4 a) Pressure contours showing the development and propagation (left to right)
of a two-dimensional detonation.29'30 b) Possible adaptive grid for this type of cal-
culation, showing approximately every eighth grid line.
Weak and Strong Ignition Behind Reflected Shocks

Shock-induced ignition of detonations occurs in one of two modes,
depending on the stoichiometry and thermodynamic conditions of the shocked
gas. At lower temperatures, the ignition is weak or mild. Many locations
where chemical reactions begin appear behind the shock and gradually
develop into an explosion. At higher temperatures, ignition is strong or
sharp. A single reaction center causes the abrupt appearance of a secondary
shock induced by the explosive reaction. These modes have been studied
extensively by Strehlow, Oppenheim, Edwards, Soloukhin, and their col-
laborators. Their experiments were performed in shock tubes in which they
studied ignition behind incident and reflected shocks. Typical diagnostics
included pressure measurements and schlieren photography. Here we sum-
marize selected results of numerical simulations that provide some insight
into the differences of the behavior of these modes.
The behavior of shocks and detonations in reactive mixtures of hydrogen
and oxygen was computed by using a numerical model that combined FCT
with CHEMEQ to integrate the set of reactive-flow conservation equations
with a complete model of the chemical kinetics.18'19 These results were
Fig. 5 Finite-element FCT calculation of a shock over two irregularly shaped obsta-
cles (courtesy R. Lohner).
then compared to shock-tube experiments in regimes where there is strong

ignition.20 The problem simulated was an incident shock moving into an
unreacted mixture of hydrogen, oxygen, and argon. The passage of the
shock raises the temperature and pressure of the reactive gas. When the
shock hits the wall at the end of the tube, it reflects and raises the tem-
perature and pressure of the gas behind the shock even higher. After the
chemical induction time has elapsed, the gas ignites first at the reflecting
wall because the mixture there is heated for the longest time. Sometime
after the gas has ignited, a reaction wave is established that may eventually
meet with the reflected shock. In the idealized picture of this problem, the
reflected shock brings the gas behind it to rest. The material behind this
shock, at the reflecting wall, which has been heated the longest, should be
the first to react after a suitable period equal to the chemical induction
time of the gas.
Results from calculations are shown in Fig. 6 along with a schlieren

photograph from the Ballistics Research Laboratory (BRL) experiment.
To compare the simulations most effectively with the experiment, we show
the time-dependent behavior of two quantities: the maximum pressure in
the system and the fluid velocity at the reflecting wall. The maximum
pressure at each time step is constant until the incident shock is reflected;
then it jumps to a higher value. It increases again when chemical energy
is released and the reaction wave starts, and it reaches a maximum when
the reaction wave and reflected shock merge. In an ideal problem, the
b)
107
-15- -109-
UJ
oc
106
UJ
oc I04g
a.
o
UJ
x a
10 100
TIME (Msec)-
Fig. 6 Comparisons of calculation and experiment for the strong ignition case,
a) Schlieren photograph.18"20 b) Calculated time-dependent behavior of two
quantities: maximum pressure in the system modeled and the fluid velocity at the
reflecting wall.
fluid velocity behind the reflected shock would be zero. Just after the
incident shock is reflected from the wall in the simulations, the fluid velocity
oscillates wildly and then settles quickly to a very small value. Eventually,
the reaction wave is ignited, at which time the velocity jumps up again,
quickly decays, and then rises to a substantial value when the pressure
pulse generated from the merging of the shock and reaction wave propa-
gates back into the wall.
Comparing the times marked on the schlieren photograph and the cal-
culation in Fig. 5, we see that the time calculated for the formation of the
reaction wave is in good agreement with the experiment. There is apparent
disagreement among the calculations and experiment in the individual times
calculated for merging of the shock and reaction wave and the reflection
of this wave from the wall. Part of this could be attributed to nonideal
effects at the wall in the experiment. However, the sum of these numbers
is very close to that determined experimentally. In fact, there is some
ambiguity in determing these quantities from the schlieren photograph.
Figure 7 shows the calculated position of the reflected shock front, reaction
wave, transmitted detonation, and contact discontinuity as a function of
time. In addition to the information about the macroscopic flow properties,
the calculation also produces number densities of all of the individual
species as a function of position and time.
Now consider the weak ignition case. Here we would not expect the
agreement between the experimental results and a one-dimensional cal-
culation to be as close as shown earlier. Experimentally, the distinct reac-
15.0
0.0
100 150 200
Time (jus)
Fig. 7 Calculated position of reflected shock front, reaction wave, transmitted de-
tonation, and contact discontinuity as a function of time for the strong ignition case.
tion wave seen in the strong ignition case does not appear at reproducible
times at the reflecting wall of the shock tube. The time of ignition is blurry
and indistinct. Compared with the strong ignition experiment, the time
scales are much larger as a result of the lower temperatures and consid-
erably longer induction time. However, it is instructive to see just how
"wrong" the one-dimensional model is in this case.
Based on the results of the strong ignition simulation, we expect the
one-dimensional simulation to show ignition at the wall at about 1550 JJLS,
the chemical induction time of the gas at the conditions behind the reflected
shock. At that time, a reaction wave would form and eventually meet with
the shock front. The schlieren photograph for this case shows that ignition
occurred at about 600 fjus after shock reflection, about a factor of 2 too
soon. Note that the change of induction time with temperature is a much
steeper curve in weak ignition than it is for strong ignition. Therefore, the
induction time is a much more sensitive function of temperature. Figure
8 shows the positions of the various fronts and discontinuities in the problem
as a function of time. The reaction wave is defined as that location where
the ratio of OH to Ar becomes significant and energy has just started to
be released. We first note that the incident shock reflection occurs at about
100 JJLS into the calculation. The first indication of energy release occurs
at 750 JULS after shock reflection. By 1120 JJLS, a detonation has formed in
the region behind the shock.
80
Transmitted
Detonation '
60
Contact
o 40 Discontinuity
"53
o — Reflected Shock
0.
20
500 1000 1500

Time (jus)
Fig. 8 Calculated position of reflected shock front, reaction wave,
transmitted detonation, and contact discontinuity as a function of time
for the weak ignition case.
The curious feature here is that the time to ignition is reduced from the
chemical induction time in the one-dimensional calculation, just as it is in
the experiment. Resolution tests of the calculation, in which the compu-
tational cell sizes near the reflecting wall are increased, also give results
that show much shorter ignition times than would be expected. A close
inspection of the temperature and velocity profiles from the computations18-19
show that ignition has not started at the wall in the calculation, but that a
hot spot forms at some distance from the wall.
The explanation for the results shown in the calculations lies in both the
physical and numerical deviations of the calculations from idealized con-
ditions. First, the calculations are fully compressible solutions of the con-
servation equations and so include the presence of sound waves. Although
the gas behind the reflected shock comes fully to rest in the idealized shock-
tube problem, there are always acoustic waves present behind the reflected
shock in a compressible medium. There are also numerical effects that
generate small fluctuations. The reactive mixture is very sensitive to acous-
tic disturbances,19 so that small fluctuations in temperature or pressure
greatly affect the induction time. This strong sensitivity, the increased ease
by which a system ignites due to fluctuations, is at the heart of what causes
weak ignition. Both experiments22 and extensions of these calculations that
focus on weak ignition behind incident shocks21 have shown that a number
of hot spots may form before ignition occurs.
Multidimensional Detonation Structure

It has become well known in the last 20 years that the front of a self-
sustained propagating detonation in gases is not uniform and planar. Its
structure is complex and multidimensional, involviong interactions between
incident shocks, Mach stems, transverse waves, and boundaries of the
regions through which the detonation is moving.23 24 The triple points formed
at the intersection of the transverse wave with the Mach stem and the
incident shock trace out patterns that are called detonation cells. These
structures may be obscured for certain material conditions or if the deto-
nation is heavily overdriven, but they seem to occur consistently in self-
sustained gas-phase detonations. Extensive experimental data show that
the size and regularity of this cell structure are characteristic of the par-
ticular combination of initial material conditions, such as composition,
density, and pressure. In particular, it has been shown25 that, in the hydro-
gen-oxgen system, fuel-lean mixtures and mixtures diluted with greater
than 40% argon show a very regular structure. As the mixture becomes
less dilute, the structure becomes more irregular. When the diluent is
nitrogen instead of argon, the structures are also irregular.
Numerical simulations have been used to help explain detonation struc-
ture, to look at those features of the structure that determine the cell size,
to study the interaction of the structure with the geometry, and to study
those factors that can cause irregularity in cell structures (see, for example,
Refs. 26-30). Most of this work has been for H2-O2 mixtures highly diluted
with argon at initial pressures substantially below 1 atm. More recent work
has included studies of the structure of detonations in liquid nitromethane.
Planar detonation fronts may be unstable to perturbations in the trans-

verse directions. Given an initially planar front, it tends to depart from a
one-dimensional configuration. The upper diagram in Fig. 9 contains a
series of pressure contours from a calculation of the early effects of per-
turbing a planar detonation front. 28 The calculation was initialized by plac-
ing an elliptical pocket of hot, unburned gas behind a planar detonation.
The pocket, a device for initiating the perturbation, burns slowly and sends
out pressure disturbances in all directions. These waves interact with the
incident shock front, causing it to curve outward. The pressure waves also
reflect from the sidewalls of the channel and move transverse to the incident
shock front, as seen in frame 2 of the figure. These pressure waves are
strengthened due to collisions with each other and further increase the
curvature of the incident shock front, as can be seen in frame 7. After a
short time, a portion of the incident shock reflects from the sidewalls of
the channel. Frame 8 shows that a Mach reflection occurs since a pair of
triple points are seen. The reflected shock waves, which are the transverse
waves, are initially weak but are later strengthened when the incident shock
accelerates due to the increased rate of chemical reactions caused by the
transverse waves.
The further evolution of the pair of triple points is shown in the lower
diagram in Fig. 9. The path and the direction of movement of the triple
o
q
in
25.8 X(cm) 42.4
1 3 4 5 6
64.5 X(cm) 78.9

Fig. 9 Upper figure: pressure contours at lO-fxs intervals showing the effects of
perturbing a two-dimensional planar detonation. Lower figure: later-time pressure
contours.
points are indicated by the lines with arrows. In the first frame, the trans-
verse waves are moving toward each other and away from the wall. By
the fourth frame, they have collided with each other and are moving away
from the center of the channel. In frame 7, they are again moving toward
the center after colliding with the walls. Frames 7 and 8 are similar to
frames 1 and 2 and show that a configuration with a repeating pattern has
been established.
Detonation Cells and Unreacted Pockets

The first numerical simulations of the evolution to the triple-point struc-
ture in detonations was done for low-pressure, stoichiometric H2-O2 mix-
tures heavily diluted with argon.30 These calculations demonstrated that
computer simulations can reproduce the types of shock and reaction-front
structures characterizing detonations. Figure 10 shows patterns of triple
points at the front of detonations in dilute H2-O2 mixtures for three dif-
ferent channel heights.28 These figures show that, as the height of the
channel increases, a structure typical of a detonation cell is seen.
One initially unexpected but potentially important phenomenon found
in the experiments and simulations is the formation of unreacted pockets
of material behind the detonation front. 28 31 Experiments first showed such
pockets, but their implications were unexplored until they were seen in
numerical simulations. Plate 11 (see the color section) presents a series of
contours from a calculation similar to that shown earlier, but in a very
72 82 132 137
Fig. 10 Comparison of triple-point paths for the same mixture in tubes of

increasing height.
narrow channel. The figures on the left show the extent of reaction. The
detonation propagates to the right into the unreacted gas mixture, and
fully reacted gas is on the left-hand side. The various shades in between
represent different degrees of reactedness and together comprise the det-
onation front. In the third frame, we see a detached pocket of unburned
gas behind the detonation front. The figures on the right show the tem-
perature of the gases in the various regions. The figures show a cold,
unreacted pocket cut off by interacting shock waves.
The formation of the unburned pockets can be traced directly to the
curvature of the transverse shock waves. When two transverse waves collide
or one hits a wall, the interaction can cut off a portion of unreacted, cold
material. If the material in the pockets burns slowly enough, the process
effectively draws energy out of the detonation and can provide a mechanism
for detonation extinction. If the pockets burn rapidly, they can generate
new pressure pulses that perturb the system and cause new cellular struc-
tures to form at the detonation front. This is one possible mechanism for
generating the irregular cellular pattern observed in most gases and liquids.
These findings are curious because they show how an initially homo-
geneous material can develop an extremely inhomogeneous structure as
shocks move through it. These findings are important because the existence
of unreacted pockets and their properties can determine whether a deto-
nation lives or dies.
Irregularity in the Structure of Detonation Waves

Most of the experimental results showing detonation cells do not show
the regular structure we have seen in the calculations given earlier, but
show irregular structure in which the size of a cell is not clearly determined.
In the experiments, the regular cell patterns appear in the highly diluted
argon mixtures of hydrogen and oxygen, as we have predicted in the cal-
culations. If the mixture is less dilute, or argon is replaced by molecular
nitrogen, the structure again becomes irregular.
Some insight into the mechanism causing the irregularity of the structures
was provided in Refs. 27 and 29, which study detonation cell structure in
liquid nitromethane. In these studies the parameters for the induction
parameter model were taken from experimental data, as opposed to being
derived from a detailed chemical reaction mechanism. Guirguis et al.27-29
defined two quantities, the induction parameter, defined in the form T, =
A1 exp (E'/RT), and an energy release time, which also had an Arrhenius
form, Tr = Ar exp(Er/RT). The initial calculations were performed with
the experimentally derived values of Tr and T,-, and typical results are shown
in Fig. 11. The outstanding feature of this figure is that the detonation
structure appears very irregular. The solid horizontal lines drawn through
the pressure contours outline the movement of the stronger triple points,
and the dashed lines indicate weak triple points.
Numerical tests in which the time step and computational mesh spacing
were varied proved that these results are a physically correct property of
the calculation, and not numerically induced. The next question is, What
is the origin and what controls this structure? Figure 12 shows the results
of increasing the value of El in the expression for the induction time. This
STEP 2500 2900 3300 3600 4000 4600 5000 5500

TIME (ns) 20.13 23.16 26.51 28.93 32.19 37.12 40.28 44.34
0.06
0.12 0.15 0.20 0.25 0.30

STEP 5600 6000 6600 7100 7400 7800 8200
TIME (RS) 45.14 48.37 53.34 57.45 59.81 62.94 66.17
0.06
J
>.
0.0^
I I I————L_
0.27 0.30 0.35 0.45 0.44
x (mm)
Fig. 11 Pressure contours of a detonation were propagating in a channel filled with
liquid nitromethane. Solid traces are loci of main triple points; dashed traces are
loci of secondary triple points. Induction time and energy-release parameters derived
from the experimental data.27'29
STEP 1700 2100 2400 2600 2900 3300 3600 3900 4200 4500
TIME (ns) 14.03 17.45 19.88 21.49 23.93 27.31 29.8032.37 35.04 37.70
0.05
E
E
0.0
0.08 0.10 0.15 0.20 0.25
x (mm)
Fig. 12 Pressure contours of a detonation wave propagating in a channel filled with
liquid nitromethane with modification to the temperature dependence of the exper-
imentally derived induction time.27'29
has the effect of increasing the difference of the size of the induction zones
behind the Mach stems and reflected shocks. The resulting structure is
now quite regular and qualitatively similar to the dilute hydrogen-oxygen
calculations described earlier. An interesting feature of this calculation is
the formation of large unreacted pockets behind the propagating detona-
tion front, which was not really seen in the calculations represented in
Plate 11.
Similar studies were carried out that varied the energy-release param-
eters. From these, we found that detonation structure is also affected by
the energy-release parameters, Ar and Tr. Instantaneous energy release
leads to a one-dimensional structure. Very fast energy release results in
less regular structures, and very slow energy release results in large pockets
and highly curved fronts. In the latter case, the detonation may die out.
The obvious question now is, What can we say about irregularity in the
gas phase? To answer these questions, we have been doing calculations of
detonations in hydrogen-oxygen diluted with argon, as earlier, and com-
paring these to calculations of hydrogen-oxygen diluted with nitrogen,
which we know gives irregular structures.27 29 The calculations indicate that
the nitrogen dilution has a tendency to give irregular structures. This result
indicates that the multidimensional structure of a detonation depends on
the differences of the thermodynamic properties in the induction zones
behind the Mach stem and the incident shock. Whereas the chemical in-
duction times for equivalent nitrogen and argon dilutions are the same,
the rate and amount of energy release are different, and so the relative
sizes of the induction zones behind the Mach stem and incident shock are
different for argon and nitrogen dilution.
Summary
The computations described in this paper have been used as tools for
determining some of the physical mechanisms controlling shock-to-deto-
nation transitions, weak and strong ignition, and detonation cell structure.
When the one-dimensional computations first were done, they were state-
of-the-art. Now they can be done rather matter-of-factly on personal com-
puters. The use of such relatively simple, fundamental computations for
exploring basic physical-chemical mechanisms, however, should not be
underestimated. There are still many computations of this sort that will
contribute to understanding fundamental mechanisms of detonation igni-
tion, transition, and even propagation.
Now the two-dimensional computations, with the model energy release,
can be done rather routinely on large, but not the largest, computers. If
detailed chemical reaction mechanisms are included, however, the com-
putations require a great deal of computer time on the largest supercom-
puters. The types of two-dimensional computations described here have
taught us much about fundamental detonation structure, propagation, and
extinction. From these, we conclude the following:
1. The multidimensional structure of a detonation depends on the dif-
ferences of the thermodynamic properties in the induction zones behind
the Mach stem and the incident shock.
2. The formation of unreacted pockets behind the detonation front de-

pends on the inclination of the transverse waves and the curvature of the
shock fronts. Highly curved shocks may result in large pockets.
3. The temperature dependence of the induction time is a major factor
in the regularity of detonation structure.
4. Detonation structure is affected by the energy-release parameters.
Instantaneous energy release leads to one-dimensional structures. Fast energy
release results in less regular structures. Very slow energy release results
in large pockets and highly curved shocks, and the detonation may die out.
There are certainly many questions that remain unanswered, and these
simulations have, in fact, actually brought up as many new questions as
they have answered old ones. Now the community is in a position to do
three-dimensional simulations of propagating detonations as has already
been started by Fujiwara and Reddy32 and Schoffel.33 Here we have the
opportunity to study the intriguing physics of spinning and galloping det-
onations and so approach simulating actual experiments.
There are certain fundamental questions that have to be addressed
regarding the accuracy required to model a propagating detonation and
exactly what we expect to see from these computations. Preliminary com-
putations by Kailasanath and Oran34 have shown that it is possible to obtain
a converged solution of a propagating two-dimensional cellular detonation
structure. However, using a fourth-order phase-accurate flux-corrected-
transport algorithm, the computations require about 30-50 computational
cells per detonation cell. Jones et al.,35 however, have shown that is pos-
sible to obtain good solutions of larger-scale propagating detonations in
which the cell size is very much smaller than the main shock structures.
Extending Jones' computations, in particular, to three dimensions, is now
straightforward.
Acknowledgments
This work was done in collaboration with T. R. Young, J. M. Picone,
C. Oswald, and R. H. Guirguis, and was sponsored by the Naval Research
Laboratory through the Office of Naval Research.
References
E. S. and Boris, J. P., Numerical Simulation of Reactive Flow, Elsevier,
New York, 1987.
2
Yanenko, N. N., The Method of Fractional Steps, Springer- Verlag, New York,
1971.
3
Richtmyer, R. D. and Morton, K. W., Difference Methods for Initial-Value
Problems, Interscience, New York, 1967.
4
Roache, P. J., Computational Fluid Dynamics, Hermosa, Albuquerque, NM,
1972.
5
Potter, D., Computational Physics, Wiley, New York, 1973.
6
Boris, J. P., "A Fluid Transport Algorithm That Works," Computing as a
Language of Physics, International Atomic Energy Agency, Vienna, 1971, pp.
171-189.
7
Boris, J. P. and Book, D. L., "Solution of the Continuity Equation by the
Method of Flux-Corrected Transport," Methods in Computational Physics, Vol.
16, Academic, New York, 1976, p. 85-129.
8
Zalesak, S., "Fully Multidimensional Flux-Corrected Transport Algorithms for
Fluids," Journal of Computational Physics, Vol. 31, 1979, p. 335.
9
Sigamura, T., Taki, S., and Fujiwara T., Propagation of Two-Dimensional
Nonsteady Detonation in a Channel with Backward-Facing Step, Shock Tubes and
Waves, edited by C. Treanor and J. G. Hall, State University of New York Press,
Albany, NY, 1981.
10
Flow Problems," Journal of Physical Chemistry, Vol. 81, 1977, pp. 2424-2427.
H
Young, T. R., "CHEMEQ: Subroutine for Solving Stiff Ordinary Differential
Equations," Naval Research Lab., Washington, DC, Memo. Rept. 4091, 1979.
12
Stull, D. R. and Prophet, H., JANNAF Thermo chemical Tables, 2nd ed.,
National Standard Reference Data Series, No. 37, U.S. National Bureau of Stand-
ards, Gaithersburg, MD, 1971.
13
Gordon, S. and McBride B. J., "Computer Program for Calculation of Complex
Chemical Equilibrium Compositions, Rocket Performance, Incident and Reflected
Shocks, and Chapman-Jouguet Detonations," NASA SP-273, 1976.
14
Young, T. R., "Table Look-Up: An Effective Tool on Vector Computers,"
Proceedings of 10th International Meetings on Numerical Simulation of Plasmas,
San Diego, CA, 1982.
15
Oran, E. S., Boris, J. P., Young, T., Flanigan, M., Burks, T., and Picone,
M., "Numerical Simulations of Detonations in Hydrogen-Air and Methane-Air
Mixtures," Eighteenth Symposium (International) on Combustion, The Combustion
16
Kailasanath, K. and Oran, E. S., "Ignition of Flamelets Behind Incident Shock
Waves and the Transition to Detonation," Combustion Science and Technology,
Vol. 34, 1983, p. 345.
17
L6hner, R., Morgan, K., and Zienkiewicz, O. C., "An Adaptive Finite Element
Procedure for Compressible High Speed Flows," Computational Methods in Ap-
plied Mechanical Engineering, Vol. 51, 1985, pp. 441-465.
18
Oran, E. S., Young, T. R., Boris, J. P., and Cohen, A., "Weak and Strong
Ignition. I. Numerical Simulations of Shock Tube Experiments," Combustion and
Flame, Vol. 48, 1982, pp. 135-148.
19
Oran, E. S. and Boris, J. P., "Weak and Strong Ignition. II. Sensitivity of the
Hydrogen-Oxygen System," Combustion and Flame, Vol. 48, 1982, pp. 149-161.
20
Cohen, A. and Larsen, J., "Explosive Mechanism of the H2-O2 Reaction Near
the Second Ignition Limit," Ballistics Research Labs., Aberdeen, MD, BRL Rept.
1386, 1967.
21
Kailasanath, K. and Oran, E. S. "Ignition of Flamelets Behind Incident Shock
Waves and the Transition to Detonation", Combustion Science and Technology,
Vol. 34, 1983, p. 345.
22
Edwards, D. H., Thomas, G. O., and Williams, T. L., "Initiation of Detonation
by Unsteady Planar Incident Shock Waves," Combustion and Flame, Vol. 43, 1981,
p. 187.
23
Fickett, W. and Davis, W. C., Detonation, University of California Press,
Berkeley, CA, 1979.
24
Strehlow, R. A., Combustion Fundamentals, McGraw-Hill, New York, 1984.
25
Strehlow, R. A., "The Nature of Transverse Waves in Detonations," Ada
Astronautica, Vol. 14, 1969, pp. 539-548.
26
Taki, S. and Fujiwara, T., "Numerical Simulation of Triple Shock Behavior
of Gaseous Detonation," Eighteenth Symposium (International) on Combustion,
27
Guirguis, R., Oran, E. S., and Kailasanath, K., "Numerical Simulations of the
Cellular Structure of Detonations in Liquid Nitromethane—Regularity of the
Structure," Combustion and Flame, Vol. 65, 1986, pp. 339-366.
28
Kailasanath, K., Oran, E. S., Boris, J. P., and Young, T. R., "Determination
of Detonation Cell Size and the Role of Transverse Waves in Two-Dimensional
Detonations," Combustion and Flame, Vol. 61, 1985, pp. 199-209.
29
Guirguis, R., Oran, E. S., and Kailasanath, K., "The Effect of Energy Release
on the Regularity of Detonation Cells in Liquid Nitromethane, Proceedings of 21st
PA, 1987, pp. 1659-1668.
30
Taki, S. and Fujiwara, T., "Numerical Analysis of Two-Dimensional Nonsteady
Detonations," AlAA Journal, Vol. 16, 1978, pp. 73-77.
31
Oran, E. S., Young, T. R., Boris, J. P., Picone, J. M., and Edwards, D. H.,
"A Study of Detonation Structure: The Formation of Unreacted Gas Pockets,"
Nineteenth Symposium (International) on Combustion, The Combustion Institute,
Pittsburgh, PA, 1982, pp. 573-582.
32
Fujiwara, T. and Reddy, K. V., "Propagation Mechanism of Detonation: Three-
Dimensional Phenomena," Proceedings of 12th International Colloquium on the
Dynamics of Explosions and Reactive Systems, Progress in Astronautics and Aer-
onautics series, (to be published).
33
Schoffel, S. U., "The Mechanism of Spinning Detonation—Numerical Study
for a Rectangular Cross-Section Tube," Proceedings of 12th International Collo-
quium on the Dynamics of Explosions and Reactive Systems (to be published).
34
Kailasanath K., and Oran, E. S., "Detailed Structure of Two-Dimensional
Detonations, Proceedings of 12th International Colloquium on the Dynamics of
Explosions and Reactive Systems, Progress in Astronautics and Aeronautics, (to
be published).
35
Jones, D. A., Sichel, M., and Oran, E. S., "Numerical Simulation of Layered
Detonations," Proceedings of 12th International Colloquium on the Dynamics of
Explosions and Reactive Systems, Progress in Astronautics and Aeronautics, (to
be published).
Chapter 14
Numerical Modeling of Heterogeneous Detonations
Martin Sichel
Nomenclature
CD = particle drag coefficient
Cf = wall skin friction coefficient
Dh = hydraulic diameter
E = activation energy
h = specific enthalpy of gas phase
hp = specific enthalpy of solid or
liquid phase
hpQ = specific enthalpy of solid or
liquid phase upstream of shock
h0 = specific enthalpy of gas phase
upstream of shock
Hp = total enthalpy of a single particle
n = particle number density
p = pressure
p0 = pressure upstream of shock
Qr = radiative heat transfer to the wall
Qs = viscous dissipation due to wall
motion in shock-fixed coordi-
nates
Qw = convective heat transfer to the
wall
<20 = heat of reaction per unit mass of
solid or liquid phase
rp = particle radius
R = universal gas constant
Re = Reynolds number
t = time
tc = ignition delay time
T = gas-phase temperature
reserved.
447
448 NUMERICAL MODELING OF HETEROGENEOUS DETONATIONS
7} = gas recovery temperature

Tw = wall temperature
u = velocity of gas phase
up = particle or droplet velocity
UQ = velocity of two-phase mixture
upstream of shock
W = molecular weight of gas phase
WL = momentum loss due to viscous
shear at the wall
Y = fraction of the induction distance
y = ratio of specific heats
r\ = convective heat-transfer coeffi-
cient
r\r = radiative heat-transfer coeffi-
cient
p = density of gas phase
pp = solid or liquid density
Po = gas density upstream of shock
crp = mass of dust particles or droplets
per unit volume of mixture
crp0 = particle or droplet mass per unit
volume of mixture upstream of
shock
cr^ = microspray per unit volume of
mixture
TW = viscous shear stress at the wall
I. Introduction
A LMOST all heterogeneous fuel-oxidizer mixtures will explode under

the proper conditions, which include a rich enough mixture and a
suitable ignition source. Coal mine, grain dust, and industrial dust explo-
sions, which are responsible for significant loss of life and property every
year, are important examples and the reason for considerable interest in
the explosion of heterogeneous fuel-oxidizer mixtures. In some cases, the
ignition process leads to the formation of a detonation, which then results
in maximum damage and a worst-case scenario. As a result, there has been
considerable experimental and theoretical research on heterogeneous det-
onations in recent years, as discussed, for example, in the reviews of Bor-
isov and Gelfand1 and Sichel.2-3 In the last two decades, there have been
significant advances in numerical methods and computational power. The
application of numerical methods to the problem of heterogeneous deto-
nation is the focus of this chapter.
Heterogeneous detonations and combustion processes are of fundamen-
tal interest aside from their very important role in accidental explosions.
Analysis of such detonations requires the combined study of unsteady
compressible and sometimes viscous flow, chemistry, two-phase flow7, and
M. SICHEL 449
the physics of droplet breakup or particle combustion. Realistic simulation

of these interacting processes represents a severe challenge, and the lit-
erature on numerical simulation of this complex process is, therefore, still
relatively limited.
A more precise definition of the term heterogeneous as used in this
chapter is required. Detonations will be considered in which the solid or
liquid volume fraction is negligible so that particle-particle collisions can
be neglected. This assumption is satisfied for most dust and spray deto-
nations of interest. Herein the discussion is limited mainly to dusts or sprays
distributed evenly throughout the gaseous medium. In addition, there are
a number of situations in which the solid or liquid starts out as a thin dust
layer or liquid film so that dispersal becomes part of the detonation process.
This class of heterogeneous detonations, which also is responsible for many
accidents, will be discussed briefly. Here again, the volume fraction of the
solid or liquid material is generally small after dispersal. Solid explosives
with gas inclusions or solid explosive phenomena that include heteroge-
neous flow are discussed by Baer and Nunziato in another chapter of this
book and so are not considered here.
In order of increasing complexity, the problems of detonation theory
are as follows: 1) thermodynamic calculation of the propagation velocity
of an unconfined lossless Chapman-Jouguet (C-J) detonation, 2) deter-
mination of the structure of the reaction zone, 3) simulation of the process
of direct initiation of a detonation by a sudden release of energy such as
by a high explosive charge, and 4) simulation of deflagration to detonation
transition. All of these problems require the application of numerical meth-
ods. Even the relatively simple problem of determining the propagation
velocity of a lossless C-J detonation requires the application of numerical
methods to compute the equilibrium composition behind the wave (e.g.,
the Gordon-McBride code4-5).
To determine the effect of wall losses on the propagation of detonations
in ducts and to establish the minimum energy required to initiate a deto-
nation, a knowledge of the detailed structure of the reaction zone is required.
Transverse waves or detonation cells are dominant factors in the structure
of a steadily propagating gaseous detonation, and the numerical simulation
of this complex process has been the subject of a number of studies (e.g.,
Refs. 6 and 7). So far it has not been established what role, if any, transverse
waves play in heterogeneous detonations. Most studies of heterogeneous
detonation structure are, therefore, based on a simple one-dimensional
Zeldovich-Neumann-Doering (ZND)-type model, which typically requires
the integration of a system of ordinary differential equations. Although
this problem does not sound too demanding from a numerical point of
view, there are several computational difficulties. There is often a saddle-
like singularity at the end of the reaction zone at the C-J plane that must
be considered in determining the solution. In dust detonations, the unsteady
temperature variations in the interior of the dust particles can have a
significant effect on the ignition and the combustion process, and this
requires that a solution of the unsteady heat conduction equation for the
particle interior be coupled to the other computations. Some spray deto-
nation models are based on the observation that the fuel droplets explode
at the end of an induction period and so introduce a sequence of exploding
droplets or blast waves behind the leading shock front. 8 - 9 Computation of
the "steady-state" structure then involves the appropriate combination of
a series of inherently unsteady blast wave solutions. Perhaps the most
important role of studies of steady-state structure is to provide a means of
examining the physical mechanisms governing heterogeneous detonations
in greater detail than is possible in the unsteady simulations because of
computational time limitations. Since the steady-state structure should rep-
resent the limit of a properly executed unsteady simulation, knowledge of
the steady-state reaction zone structure provides an important means of
validating unsteady simulations.
Direct initiation of plane, cylindrical, or spherical gaseous detonations
by the rapid release of energy by a high explosive charge has been the
subject of numerous experimental and analytical studies.10 Analysis of the
corresponding heterogeneous problem requires the solution of the partial
differential equations governing one-dimensional, unsteady two-phase flow
and is by now relatively amenable to numerical simulation using, for exam-
ple, the Godunov or flux-corrected transport (FCT) algorithms. There is
a limited set of experimental data available for evaluating the results.
Therefore, it is not surprising that most of the numerical studies to date
have concentrated on various aspects of the direct or shock initiation
problem.
The transition from deflagration to detonation (DDT), which remains
perhaps the most intractable problem in detonation theory even in the case
of purely gaseous detonations, has received relatively little attention. The
physics of the process by which laminar flames accelerate and induce shock
waves in the flow that ultimately lead to the formation of a detonation is
not clearly understood even in the case of gasous mixtures, but appears
to involve interaction among boundary-layer turbulence, inhomogeneities
in the combustible mixture, and the presence of obstacles in the flow.
Although there has been some numerical modeling in the heterogeneous
case, the results are sparse; consequently, DDT will not be considered in
this chapter.
We will review studies based on a numerical solution of the basic gov-
erning equations. Numerical simulation of direct initiation and the com-
putation of steady detonation structure satisfy this criterion. The chapter
begins with a discussion of heterogeneous detonation structure in order to
set the stage for consideration of the direct initiation of heterogeneous
detonations. Numerical simulations of the direct initiation of detonations
in sprays and dusts are then considered in detail. The chapter ends with a
discussion of a series of important unsolved problems in heterogeneous
detonation theory whose treatment will be possible only by the judicious
application of numerical simulation.
This review of heterogeneous detonations is not all-inclusive and, of
necessity, reflects the prejudices and interests of the author. The objective
is to provide the reader with at least an entree to this fascinating and
expanding subject.
M. SICHEL 451
II. Structure of Heterogeneous Detonations

A. Formulation
Most simulations of the structure of heterogeneous detonations are based
on the ZND model, in which a leading shock wave is followed by an
induction and reaction zone, and terminated at the C-J plane where the
velocity becomes sonic. The structure of a plane heterogeneous detonation
based on this model, propagating through a duct with velocity w 0 , is shown
schematically in Fig. 1 in a coordinate system fixed to the shock wave. The
shock is stationary in this coordinate system while the wall moves with
velocity UQ. Particles or droplets and the gas approach the leading shock
wave with the initial gas and particle velocities u0 and up(), which are usually
equal to each other upstream of the leading shock. The gas velocity u
decreases almost discontinuously across the shock, while the particle veloc-
ity up remains at up() and then approaches the gas velocity u in an extended
relaxation zone behind the shock. For shock Mach numbers in the deto-
nation regime that are on the order of five or so, the particles initially
move with supersonic velocity relative to the gas. Thus, each particle will,
at least at the beginning of the relaxation zone, be preceded by a bow
shock, as shown in the figure.
The structure of spray detonations has been considered on the basis of
the preceding model by Borisov et al., 11 Gubin et al., 12 and Gubin and
Sichel13; these authors provide citations to earlier experimental and mod-
eling studies of this process. Dust detonation structure has been investi-
gated in Refs. 14-16. The main objective of these studies was determi-
nation of the effects of the physical properties of the spray or dust on the
propagation velocity as a function of the mixture ratio or the amount of
dust present.
I:nduction Reaction Zone

Zone / \
U / I \ f
p0 > °) -\
^ o) 0)
o o
I o
0 o o) °^ -^
——^ °
o o o) °) G
>^
0
0 >
Shock / ^ \ C-J Plane

Front x \.
Ignition Plane
Fig. 1 Schematic diagram of the structure of a heterogeneous detonation in shock-
fixed coordinates.
The conservation equations used to describe the two-phase flow behind

the leading shock wave are essentially the same in all of these studies and
have been discussed in detail in Ref. 17. In integrated form, the equations
for the conservation of mass, momentum, and energy can be expressed as
pu + vpup + cr^w - p0M0 + ap0u0 (1)
pu2 + vpu2p + aX + p + WL = p0w§ + (jp0w§ + Po (2)
- + h } + vpup ( -f + hp + <2o ) + <V ( - + hp + (

/
\
(3)
Here p is the density of the gas phase, <JP the mass of dust particles or
droplets per unit volume of mixture, p the pressure, and h and hp the
specific gas and particle enthalpies, and Q0 the heat of reaction per unit
mass of liquid or solid phase.
In the case of detonations propagating through ducts, the terms WL,
Qw, Qr, and Qs are loss and dissipation terms representing the interaction
between the wall and the heterogeneous reactive mixture within the dust.
The quantity WL is the momentum loss due to viscous shear at the wall,
Qw and Qr the convective and radiative heat transfer to the wall, and Qs
the viscous dissipation that arises because the wall, which exerts a shear
stress TW on the fluid within the tube, moves with velocity u0 in shock-fixed
coordinates. In the case of spray detonations, droplet shattering is accom-
panied by the formation of a fine micromist of liquid fuel that may ignite
rapidly after a certain ignition delay period, and o^ is the mass per unit
volume of mixture of this fine mist. These quantities are discussed in greater
detail in Refs. 13 and 16, and all symbols are defined in the Nomenclature.
It is usually assumed that the gas satisfies the perfect gas equation of
state, so that
p = p(RIW)T (4)
where R is the universal gas constant and W the molecular weight.
Here the gas-particle mixture is treated as a single fluid as opposed to
the two-fluid model typically used in the formulation of the direct initiation
problem, which is discussed in Sec. III. However, the particle velocity and
temperature are different from those of the gas, and the interaction between
the gas and the particles enters through the equations of particle motion
and heating. The particles are decelerated by the aerodynamic drag force
due to the relative velocity between the gas and the particles. It is usually
reasonable to assume that the volume fraction of the particles is negligible.
The particle velocity up is determined from the particle equation of motion17:
3 ^p^~df =
2^
M. SICHEL 453
Boiko et al.18 have shown that, in the case of solid or liquid particles
distributed within a gaseous medium, the aerodynamic drag on the right-
hand side of this equation is the dominant force acting on the particles.
Other forces, which include that due to the pressure gradient, the virtual
mass and Basset forces associated with the unsteady motion of the particles,
and gravitational forces, are negligible in the case of shock acceleration.
If Hp is the total enthalpy of a single particle of radius rp, as opposed
to the enthalpy hp per unit particle mass, the rate of particle heating or
cooling is described by the following equation:
- T(rp,t)} + T\M[Tw ~ T(rp,t)]} (6)
In this equation, r\(t) is the convective heat-transfer coefficient, r\r(t) the

radiative heat-transfer coefficient, T(rp, f) the temperature of the particle
surface, Tf(f) the recovery temperature of the gas at the particle surface,
and Tw the temperature of the wall for detonations confined in a tube.
Equations (5) and (6) implicitly assume that the particles are spherical;
although this is not an unreasonable assumption in the case of liquid drop-
lets, dust particle shapes are usually quite complex. The relationship between
Hp and hp is discussed in more detail subsequently.
Although these algebraic equations may appear simple enough, it will
be seen from the following discussion that there are computational diffi-
culties in solving them, and some of these difficulties remain to be resolved.
Even though the equations are quite general, they involve a number of
implicit assumptions. In the case of detonations propagating through ducts,
the core flow is treated as a uniform slug flow with very thin boundary
layers at the wall where all of the losses occur, an approach also proposed
by Zeldovich and Kompaneets.19 As written, the equations imply that the
dust or spray is monodisperse, which is rarely the case. In general, each
term due to the solid or liquid medium is the sum of contributions from
each group of particles of the same size class.20 For liquid droplets, eval-
uation of the flux <jû of micromist requires a suitable numerical or empir-
ical model of the droplet shattering process, whereas the variation of the
dust particle flux vpup is determined by the heterogeneous reaction rate.
The combustion heat release drives the gas velocity u toward the sonic
value, and, in the case of gases, the propagation velocity of a C-J detonation
is determined by the requirement that the gas velocity reach the sonic value
as the reacting mixture reaches equilibrium. The C-J condition or point is
associated with a singularity in the equations described earlier that can
cause difficulties in their solution. This problem was discussed, for example,
in Refs. 21 and 22. A similar singularity, which creates difficulties in numer-
ical simulations, exists in the case of heterogeneous detonations.
B. Detonations in Sprays, ZND Models

Spray detonations have a reaction zone on the order of centimeters. This
is in contrast to a fraction of a millimeter, which is typically found for
gaseous detonations at atmospheric conditions. For droplets greater than
10 (xm in diameter, spray detonations propagate at velocities significantly
below the ideal value. The earliest attempt to simulate spray detonation
structure is due to Williams,23 who assumed that droplet vaporization gov-
erns the rate at which the droplets react with the gaseous oxidizer. As also
shown by Borisov et al.,11 the reaction zones computed on this basis were
much longer than those observed for droplets greater than 20 jjum in diam-
eter. Ragland et al.24 ignored two-phase effects in the equations by assum-
ing that the reduction in propagation velocity was due only to the wall
losses, WL, Qw, and Qs. Even then, an important problem in dealing with
Eqs. (1-3) is that the loss terms and the reaction zone parameters are
coupled to each other. Thus, for example, the viscous momentum loss term
WL is given by
where x is the distance from the shock, Dh the hydraulic diameter of the
duct, and Cf the friction coefficient. To evaluate this and the other loss
terms, it is necessary to determine the variation of the velocity u within
the reaction zone, which, in turn, depends on the wall losses. Ragland et
al.24 used average properties to evaluate the loss terms. Experimental data
were used to estimate the length LR of the reaction zone, the empirical
value Cf = 0.074 Re~l/5 was used for the friction coefficient, and Reynolds
analogy between heat transfer and skin friction was used to evaluate the
heat loss Qw. The condition that u = a at the end of the reaction zone
was imposed rather than dealing with the singularity at the C-J plane.
Determination of the propagation Mach number Ms from Eqs. (1-3) was
then reduced to a solution of a fourth-order algebraic equation for Ms.
The results of this analysis were in reasonable agreement with experiment
in a limited number of cases.
As indicated earlier, the structure of spray detonations is influenced
strongly by the drop shattering process. Gubin and Sichel13 noted that the
reaction zone lengths used in the wall loss model of Ref . 24 to explain the
experimentally observed reductions in propagation velocity were approx-
imately twice the experimentally observed droplet breakup distances. Hence,
they proposed a completely different model that neglected the wall loss
terms in Eqs. (1-3) but took account of the droplet breakup process. The
key assumption was that only the heat release from the droplet micromist
formed during a certain ignition delay period contributed to the propa-
gation velocity of the spray detonation; therefore, the reduction in prop-
agation velocity was charged to incomplete combustion of the spray.
Equations (1-3) were supplemented by a particle conservation equation,
nup = const (8)
where n is the particle number density. An empirical relation for the particle
stripping rate was used to determine the micromist flux cr^w. Instead of
solving the particle equation of motion [Eq. (5)], Gubin and Sichel13 used
an empirical expression for up. The empirical expression
tc = lQ-lop-1 Qxp[-E/RT] (9)
M. SICHEL 455
proposed by Mullins25 was used for the ignition delay time tc with E =
40,000 cal/mole. This relation is typical of empirical ignition delay expres-
sions used in detonation modeling. A difficulty is that the temperature T
varies throughout the induction zone. Therefore, the ignition time delay
tign was calculated by requiring that
2(Af/rcy) = 1 (10)
where tcj is the time delay at step/ as determined from Eq. (9). This relation
is empirical, but it is of interest that a similar formulation has been used
to determine the ignition delay time in the case of gaseous detonations
with the summation replaced by an integral. 26 It is assumed that the time
for evaporation and mixing of the oxidizer and micromist is negligible
compared with droplet breakup time, and that the micromist and oxidizer
react instantaneously at the end of the induction period determined by Eq.
(10) so that o-^ drops to zero.
The system of Eqs. (1-4) and (8-10) is solved iteratively starting with
an assumed value of the propagation velocity «0. With «0 given, it is possible
to determine the variation of/?, T, p, u, up, the ignition delay time r ign ,
and the droplet mass m. The parameters at the end of the induction zone
can then be determined, and the system of governing algebraic equations
can be solved for the density resulting is an expression of the form
p = A(l ± VB) (11)
where A and B are complex algebraic expressions. These two solutions,
which correspond to strong and weak detonations, merge for a C-J deto-
nation requiring that 5 = 0, and this condition results in an expression
for the C-J propagation velocity (WO)GJ- The iteration was repeated until
the assumed value and (w 0 )cj were in agreement. This iterative procedure,
which again sidesteps the C-J singularity, is closely related to that used to
determine the C-J velocity of lossless gaseous detonations, and this velocity
is independent of the structure of the reaction front. However, the prop-
agation velocity is closely coupled to the structure in the Gubin-Sichel
model through its dependence on the ignition delay, which, in turn, deter-
mines the combustion heat release. The variation of dimensionless values
of/?, r, p, u, up, and particle mass m for a kerosene-O2 detonation com-
puted using this model are shown in Fig. 2 as a function of time t after
passage of the leading shock. The symbol Wis used for the particle velocity
in this figure. The variables are normalized using the values of p, T, p, m,
and W immediately behind the leading shock front and w 0 , the propagation
velocity. The sudden increase in the temperature and pressure immediately
behind the leading shock wave is a two-phase effect caused by the decelera-
tion of the droplets.
The model was used to compute the variation of detonation velocity
with droplet size, and the result is shown together with measured values
in Fig. 3. The model appears to predict the reduction in propagation veloc-
ity with increased particle size for the limited available experimental data
even though the underlying mechanism is completely different from that
hypothesized in Ref. 24. As indicated earlier, the loss-based model of Ref.
r~
12
0.8
0.6 Kerosene - 02, a «1 ^^. \

3
&o * 0.3 kg/m , d o « 0 5 m m .
O.4 U0 - 1720 m/sec. fV 1 arm, T0- 298*K
hgn_
8 16 24 32 4O
Fig. 2 Variation of induction zone properties of a detonation propagating through

a kerosene sprayj(from Ref. 13) (m = normalized droplet mass, W = normalized
droplet velocity, U = normalized gas velocity).
- d0 • [41
- d0 • 171
6 23OO
A - d0 • 94O/A [4J
O - d0 -1300^ [71
3
u 20OO
A - d0 -260OM (4J
o P0 -lotm.
1500 o-0 • 0 284 kg/m3, a • 1
o
UJ
Q
I
10 10O 10OO 3000
DROPLET SIZE (MICRONS)
Fig. 3 Variation of the velocity of a detonation propagating through a kerosene

spray with droplet diameter (from Ref. 13).
M. SICHEL 457
24 also predicted the effect of particle size on (w0)CJ but overpredicted the
reaction zone length. The Gubin-Sichel13 model assumes that the reaction
zone is governed mainly by droplet breakup and imposes a physically
realistic value for this length by the introduction of empirical droplet breakup
relations. The agreement between the droplet breakup model and exper-
iment, while encouraging, may well be fortuitous, but may also be a reflec-
tion of the crucial role played by the droplet breakup process. The result
that two-phase effects are absent for particles less than about 10 jjim in
diameter appears to be well confirmed by experiment.
C. Dust Detonations, ZND Structure

The physics of dust detonations is quite different since it does not involve
droplet shattering. As before, the dust particles find themselves in a high-
speed relative flow at the instant after passing through the leading shock
wave. Typically, a certain ignition delay period elapses during which the
particles are heated by the convective flow to a temperature high enough
for ignition to occur at the particle surface. The subsequent combustion
reactions may occur at the particle surface or in the gas phase for particles
containing a large fraction of volatile material. The ignition delay can have
an important influence on the propagation of dust detonations and on the
detonability of dust-oxidizer mixtures, and proper modeling represents an
important and difficult aspect of computing dust detonation structure. Sev-
eral combined experimental-theoretical studies of the ignition delay of dust
particles following the passage of a high-speed shock wave have been
conducted18-27 ~ 29 ; these will be described as an introduction to the dis-
cussion of dust detonations.
Ural et al.27 and Sichel et al.28 considered the ignition of single particles
of coal, graphite, grain dust, diamond, oats dust, and RDX dust behind
an incident shock wave. Because the relative velocity between the particles
and the gas is supersonic immediately after passage of the shock, the
convective heat transfer to the particle surface is initially large. Thus, in
the case of organic materials considered earlier, the Biot number (Bi =
hrpl\p] is of 0(1), and it is necessary to consider the temperature distribution
within the particle. This is not the case for the metal particles considered
in Ref. 18. With Bi ~ ©(1), ignition occurs at the surface where the tem-
perature within the particle has its highest value. Although the flow around
the particles and the particle shapes are extremely complex, the analysis
treats the particles as spheres and the temperature field within the particle
is assumed to be spherically symmetric. These assumptions are analogous
to starting the study of automotive aerodynamics by assuming that a car
can be treated as a sphere. The heat conduction equation governs the
temperature distribution within the particle and becomes
dt r2dr\ dr/ f
PpC
(12)
Here u'"(r, f) is a reactive source term that accounts for the heat released
by heterogeneous reactions on the particle surface and in the porous inte-
rior of the particle, and was taken as29
u'" = QppSlPo2A exp[-E/RT] (13)
where Q is the heat of combustion per unit particle mass, St the internal
surface area per unit particle mass, p02 the partial pressure of oxygen in
the oxidizing gas, A a pre-exponential factor, and E the activation energy
of the heterogeneous reaction. The convective boundary condition at the
particle surface is then
dT
= h(t)[Tf(t) - T(rp,t)] (14)
and
at the particle center. Radiative heat transfer has not been included in Eq.
(14) and is usually negligible for the short ignition delays (order of 100 JJLS)
behind strong shock waves. References 18, 27, and 28 considered isolated
particles and did not have to consider the interaction between the particles
and the gas flow behind the incident shock. The more complete analysis
of Ref . 29 took this interaction into account and is discussed in more detail
subsequently.
The analysis of Ref. 29 required simultaneous solution of Eqs. (1-3)
including the wall loss terms, the trajectory and particle heating [Eqs. (5)
and (6)], and the reactive heat conduction [Eq. (12)] with associated bound-
ary and initial conditions. The particle temperature and gas flow are cou-
pled through the film conductance T](£), which, in turn, depends on the
Reynolds number of the relative particle-gas flow. Equation (12) was solved
by using finite differences with boundary conditions adjusted at each time
step to match conditions in the also varying gas flow. Since the particle
temperature is variable, the total particle enthalpy Hp, which appears in
Eq. (6), must now be determined from
Hp = Q 47rr2CT(r) dr (15)
As the particles decelerate, the relative velocity varies from supersonic to

Stokes flow. To cover this wide range of flow regimes, it was necessary to
use empirical expressions for the Nusselt number Nu and the particle drag
coefficient CD. The trajectory and particle heating equations were inte-
grated using a fourth-order Runge-Kutta method. The ignition delay prob-
lem thus requires the coupled solution of the reactive heating problem for
the individual particles and of the flow in the two-phase relaxation zone
immediately behind the shock.
M. SICHEL 459
As the particle surface temperature increases, the heterogeneous reac-

tion rate increases, and the result is thermal runaway at the particle surface.
The ignition delay is taken as the time when this occurs. The variation of
the temperature distribution with time in a 37-jxm RDX particle is shown
in Fig. 4. It can be seen that the temperature is a maximum at the surface
and the last uppermost curve displays the start of thermal runaway. The
ignition delay depends on both the particle size and the loading ratio, i.e.,
the ratio of particle mass to gas mass per unit volume of mixture. The
particle number density n and, hence, the total particle surface area increase
for a given particle loading as the particle diameter decreases. For a high
enough loading ratio, the increased surface area can result in such rapid
cooling that the gas temperature and, hence, the particle surface temper-
ature actually may decrease during a part of the ignition delay period. As
2000. T
1300,
1600, 37 H1CROH
1400.
1200.
£ 800.
a 600.
OS
<
o,
400*
200*
0.30 0.85 0*90 0.95 1.00
r/R
Fig. 4 Time history of the temperature distribution inside a 37-jxm RDX particle
(from Ref. 29).
a consequence, counter to intuition, the particle ignition delay time may

actually increase with decreasing particle diameter under some conditions.
This situation is illustrated in Fig. 5, which shows the ignition delay as a
function of particle diameter for different values of the loading ratio com-
puted by simultaneous solution of the particle and two-phase flow equa-
tions. The increased ignition delay for the smaller particles is supported
by the experimental observation that, for a given loading ratio, RDX dust
clouds with very fine particles are more difficult to detonate than coarser
particle clouds.
The simulation of the complete structure of dust detonations based on
the ZND model has been considered in Refs. 14-16. An experimentally
determined value of the ignition delay time was used in Refs. 14 and 15,
and then the analysis was based on a solution of Eqs. (1-3) supplemented
by the trajectory and heating equations [Eqs. (5) and (6)]. Wolanski et
al.14 used an iterative method similar to that used by Gubin and Sichel13
to deal with the C-J singularity. Lee and Sichel,15 while still using an
empirical ignition delay time, developed a more accurate treatment of the
LOADING RATIO-e.2
4. 8. 12. 16. 30. 24. 28. 32. 36. 40,

PARTICLE SIZE (PIICROHETER)
Fig. 5 Ignition delay vs particle size of RDX in air/O2 (88/12) mixture behind a
Mach 5.1 shock wave (from Ref. 29).
M. SICHEL 461
C-J singularity. Fan and Sichel16 coupled analysis of the ignition delay to
the succeeding reaction zone and C-J singularity, and their simulation is
described next.
The treatment of the ignition delay zone used by Fan and Sichel16 was
the same as that of Lee et al.29 and involved numerical solution of the heat
conduction equation [Eq. (12)] using the Crank-Nicolson scheme together
with simultaneous solution of the conservation, trajectory, and heating
equations. Now, however, the rate-limiting process changes after ignition.
Prior to ignition, the particle temperature is relatively low, so that the
oxidizer can diffuse deeply into the particle before reacting with the surfaces
of the internal pores. Thus, the gas concentration is almost uniform within
the particle. After ignition, the particle temperature is so high that oxygen
can no longer penetrate into the particle before being consumed, and the
burning rate is controlled by the rate of diffusion across the particle surface,
which is determined by a solution of the particle diffusion equation. The
reactive heat release <2C, which can be expressed as
Qc = (VpQUpQ ~ VpUp)Qo
is now determined from the integral

=
Qc 47T<20 nr2Rs dr dx (16)
Jo Jo
where Rs is the heterogeneous reaction rate within the particle. The heat
conduction equation is still used in the post-ignition zone to determine the
particle surface and interior temperature, taking into account the decrease
in particle radius due to chemical reactions.
The reaction zone is terminated at the C-J plane, where the gas velocity
becomes sonic and the governing equations become singular. The character
of this singularity becomes evident when Eqs. (1-3) and (8) are rearranged
to form
Here A, B, and C are complex algebraic functions of the variables appear-

ing in the conservation equations. The solution is continuous at the C-J
plane only if the numerator vanishes when the Mach number M reaches
the value of unity. As before, an iterative method was used to determine
the solution starting with an assumed value of the propagation velocity.
Now, however, the iteration was based on Eq. (17) with the criterion that
the numerator and denominator of Eq. (17) must vanish simultaneously
for the correct choice of the propagation velocity u0.
Determination of dust detonation structure now requires solution of the
particle heat conduction equation simultaneously with the algebraic con-
servation equations, a change in reaction mechanism at the end of the
induction period, integration over the particles to determine the combus-
tion heat release, and, finally, iteration to determine the detonation veloc-
ity. The gas and particle surface temperature and the internal temperature
distribution computed for a wheat dust-air mixture with 50-jjim particles

and a concentration of 0.31 kg/m2 are shown in Figs. 6 and 7. Figure 6
shows that the particle temperature rapidly rises to a maximum after an
initial induction period of about 15 JJLS. The particle radius and elapsed
time are indicated on the temperature contours shown in Fig. 7. The tem-
perature becomes uniform after the initial short induction period; however,
the variation of particle temperature plays a crucial role in determining
the ignition delay.
D. Blast Wave Models of Detonation Structure

The analyses or simulations described earlier are based on the steady-
state ZND model. However, in many cases, it has been observed that the
detonation of sprays involves secondary explosions of the micromist entrained
in the droplet wakes during the droplet breakup process.24 30 The propa-
gation of the detonation, which is then inherently unsteady, occurs through
the interaction of the blast waves generated by these explosions with the
leading shock front. Several unsteady models to fit these observations have
been proposed. The simplest is due to Cherepanov,31 who computed the
propagation velocity by assuming that combustion occurs through a series
of periodic explosions at the C-J plane. Half of the energy of combustion
is carried downstream beyond the C-J plane by this mechanism, so that
the ratio us/u0 of the observed value to the premixed gaseous detonation
TP-TBFEIWTURE OF THE PARTICLE SURFACE

g i ^_/ >. T -TENPEROTURE OF THE BULK FLUID
C-J PLflNE
en
85-
Q_
LU
100.00 200.00 300.00 400.00 500.00 600.00

TIME flFTER SHOCK FRONTJMS)
Fig. 6 Variation of particle surface temperature and gas temperature with time
after shock passage (from Ref. 16).
M. SICHEL 463
0.20 0.40 1.00

0.60 0. 80
RflDIRL POSITION IN PflRTICLE(rYR)
Fig. 7 Temperature profiles within a shrinking particle in the reaction zone (from
Ref. 16).
velocity is equal to 1/V2 = 0.71. This result is not too far from a number
of experimental observations and, as in the Gubin-Sichel model,13 incom-
plete combustion of the fuel between the shock and C-J plane is responsible
for the reduction in propagation velocity. The Cherepanov model is inher-
ently unsteady and does not require any consideration of the C-J condition
and its associated singularity.
A detailed simulation of shock-driven propagation based on Cherepa-
nov's idea was developed by Pierce.8 The spray was assumed to be mon-
odisperse and arranged in a regular array consisting of rows of droplets or
droplet sheets. The droplet wakes in each sheet explode after an induction
time tt from the passage of the leading shock, and the resultant individual
droplet blast waves are assumed to coalesce to form two plane blast waves
(one moving downstream and one moving upstream) that reinforce the
leading shock, as shown schematically in Fig. 8. After ignition, the remain-
der of the droplets is assumed to burn at a uniform rate. The major features
of the model are shown in the x-t diagram in Fig. 9, which is centered on
the explosion center at the ignition distance xf behind the leading shock
wave. It was shown that blast coalescence occurs so rapidly that the details
of this process can be ignored. The computation then starts from the plane
blast wave immediately after the wake of a given droplet row explodes.
Approximate initial blast profiles satisfying mass, momentum, and energy
conservation are used to start the calculations of blast propagation. The
subsequent development of the flow is determined by using the Mac-
Cormack scheme (see, for example, Ref. 32) to integrate the unsteady
one-dimensional Euler equations8:
dp du dp
— — —u— (18)
dt dx dx
dli _ dl4 _! d£
(19)
dt dx p dx
dp
(20)
dt dx dx
Leading Shock
Forward-Mo v ing
Shock
/ s s s / • / 7777 X X X ///// SSSSSS/A
*JJL Rearward- (
ing Shock
/ / / / / s / / / s / / / A
Fig. 8 Droplet-generated blast waves behind the leading shock of a spray detonation
(from Ref. 8).
M. SICHEL 465
Reflected
Expansion
Shock
Interaction
Leading
Shock
Position
End of
-•-Reaction
Zone
Particle\
Path y
Dropletx
Path V
Fig. 9 Distance-time diagram centered on the explosion center for the blast model
of spray detonations (from Ref. 8).
A new blast wave is introduced after the ignition delay time /,- elapses,
and its interaction with the preceding blast waves is taken into account.
The computed oscillatory pressure distribution behind the leading shock
generated by the explosion of several droplet rows is shown in Fig. 10.
The results of this numerical simulation have been used to determine the
type of pressure profiles that would be observed by a pressure transducer
located at the wall of the detonation tube, and the resultant profiles are
at least qualitatively similar to experimental observation. The continual
interactions between the blast waves and the shock result in oscillations in
the leading shock front similar to those observed. It should be noted that
this model is completely different from the ZND model, involves a sub-
structure within the reaction zone somewhat similar to the transverse wave
structure of gaseous detonations, and necessitates extensive numerical work
to integrate the governing equations. The chemistry enters the formulation
through the ignition delay time th and there is no need to deal with the
problems introduced by the C-J condition and C-J singularity.
1200. -|
XHAT CIN3
Fig. 10 Pressure distribution development behind the leading shock wave computed
using the Pierce blast wave model of spray detonations (from Ref. 8) (XHAT =
distance from the leading shock front).
As will be seen subsequently, blast waves appear naturally in the sim-

ulation of the direct initiation of both spray and dust detonations.
All of the simulations or models described earlier consist of submodels
of the various physical processes that determine the structure of spray and
dust detonations. These submodels are then coupled together. The indi-
vidual processes themselves tend to be quite complex, and considerable
uncertainty is generally associated with the physical constants, such as
reaction rates, drag coefficients, and droplet breakup rates. Hence, various
levels of empiricism must be introduced in the course of the analysis. It
would be desirable to have a single overall simulation that takes everything
into account, obviating the need for all of the patching and iteration described
earlier. It is the author's opinion that that remains a very long range goal.
Even the simulation of the shattering and ignition of a single droplet behind
a high-speed shock wave is probably beyond current computing capacity.
The simulations or models described earlier provide considerable insight
into the complex physics of heterogeneous detonations, and a guide and
means of evaluating the simulations of direct initiation, which are discussed
in the next section.
III. Direct Initiation of Heterogeneous Detonations

A. Formulation
To date, simulations of the direct initiation of detonations in droplet or
dust clouds have been restricted to plane, cylindrical, or spherical clouds
M. SICHEL 467
with a uniform particle distribution. In contrast to the preceding discussion,

a two-fluid formulation typically is used in the direct initiation problem.
The separate conservation equations for the gaseous and solid or liquid
phases are coupled through solid-gas interaction terms on the right side.
A general formulation is presented, for example, in Refs. 33 and 34. The
equations for the gaseous phase can be written in the general conservative
form as
*- —— = Cil v(21)
dt dx '
For plane waves,
/ Pi \
Qi = PI"I I
V e1 I
5p2
2~~ Pz
"2
8p2U
PiCY - 1) 2
where subscripts 1 and 2 refer to the gaseous and solid phases, respectively.
q is the heat of combustion, 8p2 represents the mass exchange between the
solid or liquid phase and the gas, el is the energy of the gas phase per unit
volume, and M represents the momentum exchange due to the aerody-
namic drag of the particles. The particle equations are
+ 2 c 22
¥ f -=
dt dx <>
with
„ _ / P2 \ „ _ /P2«2\ r = I -§P2
The extension to cylindrical and spherical geometry is straightforward

and can be found in the references discussed subsequently. The preceding
equations are supplemented by the trajectory equation [Eq. (5)], and then
the particle heating equation when this effect is taken into account. As in
the modeling of the steady detonation structure, the physics of drop shat-
tering and of the chemistry of dust particle combustion that determine the
mass exchange rates are major sources of uncertainty. Certain assumptions,
typical of dilute two-phase systems, are implicit in the preceding formu-
lation. The contribution of the particles to the hydrodynamic pressure is
neglected. Viscous effects enter only through the aerodynamic particle drag
and through wall loss terms when these are considered. Particle-particle
interactions or collisions usually are neglected in the formulation of the 8
mass loss terms. The gas phase is invariably described by the perfect gas
equation of state.
B. Shock-Fitting Methods
One of the earliest, and perhaps canonical, studies of the heterogeneous
direct initiation problem is by Zhdan,35 who used essentially the formu-
lation presented earlier. A polydisperse spray with the Nukiyama-Tanasawa-
size distribution was considered, and the spray was divided into several
size classes with an equation of the form of Eq. (22) for each class. The
droplet stripping model of Ref. 36 was used to describe mass exchange
between the droplets and the surrounding gas. A key assumption, also
used in many other studies, was that the microspray reacts with the gaseous
oxidizer instantaneously after it is stripped from the droplet surface. The
empirical values of the drag coefficient CD used in the trajectory equation
were taken as
CD = 27 Re'°-8\ when Re < 80
2
- 0.27 Re°- \ when 80 < Re < 104
- 2, when Re > 104 (23)
Heat transfer between the gas and the droplets, i.e., Eq. (6), was not
considered so that the droplet energy remained constant.
The initial profiles of velocity, pressure, and density were taken from
the solution of Ref. 37 for a strong point explosion with energy E0. The
boundary conditions for the gas equations were that ul = 0 when r — 0,
where r is the radial coordinate. At the shock front propagating into the
droplet cloud where r = rs, the gas is required to satisfy the shock jump
conditions, which are given in dimensionless form by
,2 _
2[Cy0 - Dp -•* + !]
p = ud + 1 (24c)
where d is a dimensionless shock velocity, and u, p, and p are the dimen-

sionless velocity, pressure, and density, as defined in Refs. 35 and 20.
Because of inertia, there is no change in particle velocity immediately
downstream of the shock so that the particles satisfy
u2 = 0 n = n0 (25)
immediately behind the shock.

The gas equations were reduced to characteristic form and solved using
an implicit finite-difference scheme suggested by Landau et al.38 The dif-
ference equations that were used are given by Zhdan.35 A uniform spatial
step size h was used in the calculations; however, the time increments were
variable and chosen so that btn = hldn, where dn is the shock velocity. In
M. S1CHEL 469
this way, the shock was always located at a grid node. Zhdan states that
the error is 6(A^,/z2).
Calculations were carried out for a spray in air with pQ = 1 atm and p()
= 1.293 kg/m2. The fuel heating value was taken as 3.55 x 107 J/kg, and
the ratio of the oxygen in the air to the fuel by weight was 3.4. The spray
was divided into five size classes with diameters of 100, 80, 60, 40, and 20
juim. The point explosion energy E0 was taken as 3.35 x 105 J. The com-
puted variation of the dimensionless pressure/?, gas velocity u, and density
p are shown in Fig. 11 and are typical of the results obtained from this
type of calculation. During the first stage, the initiating blast wave decays
and the combined reaction and shock front is essentially an overdriven
detonation. After this initial decay, a C-J surface is formed where u -f c
= d, as indicated by the vertical lines on the figure. Here c and d are the
dimensionless sound and wave velocity. The reaction zone immediately
behind the shock becomes isolated from disturbances behind it once the
C-J surface has been established. This portion of the reaction zone structure
then becomes isolated from any disturbances behind the C-J plane. With
increasing time, the flow behind the C-J plane approaches the self-similar
Zeldovich and Kompaneets 19 solution for a gaseous detonation with instan-
taneous energy release. As discussed in more detail later, the propagation
velocity decreases to a value below the C-J value before reaccelerating to
the final C-J value. It is interesting to note that similar results were obtained
by Bach et al.39 in their numerical analysis of the direct initiation of gaseous
detonations. From the numerical results, it also is possible to follow the
detailed evolution of the spray as it vaporizes within the reaction zone.
Zhdan40 also applied the technique described earlier to larger droplets
using more complex droplet breakup models.
Mitrofanov et al.20 applied the calculation method developed by Zhdan35
to the study of the direct initiation of cylindrical spray detonations with
the same parameters as those used in the experiments of Ref. 41. First,
Mitrofanov et al.20 computed the steady-state spray detonation structure
for three different spray breakup and stripping models using the conser-
vation equations described in Sec. II. They considered three models: single
droplet breakup combined with boundary-layer stripping, multiple break-
ups with stripping, and a model with an accelerated breakup mechanism.
They present the detonation structure corresponding to each of these models.
As mentioned earlier, the choice of the physical mechanism by which the
solid phase is brought into contact with the oxidizer so that it can react is
of crucial importance in all heterogeneous detonation studies.
On the basis of comparison with experiments, Mitrofanov et al.20 chose
to use the first of these models for the unsteady calculations. The com-
putational technique was essentially the same as that used by Zhdan.35 The
self-similar blast wave solution of Sedov42 was used as the initial condition
for the calculations. Only monodisperse sprays were considered. The accu-
racy of the solution was checked on the basis of the constancy of integrals
of the total mass and energy, and conservation was observed to be within
1.5%. Conditions for the calculations were chosen to correspond to those
used in the experimental studies of Ref. 41. The computed pressure profiles
were similar to those reported by Zhdan and showed the development of
0.023 0.048 0.072 0.089 P'120 °-143 .°'188
r/n
r/ri
Fig. 11 Profiles of dimensionless pressure/?, gas velocity w, and density p at different
values of dimensionless time t. The C-J surface is denoted by vertical lines (from
Ref. 35).
a C-J plane after an initial blast decay period. The variation of the shock
front velocity with distance is shown in Fig. 12 for several different values
of the initiation energy E0. As already indicated, the curves pass through
a minimum below the C-J value, after which they gradually accelerate to
the C-J velocity. With decreasing E0, the minimum wave velocity drops to
M. SICHEL 471
C0= 1 1 - 5 x 105
RO = 192 |im 2-10 6
= 1.3069 3 - 2 x 106
4 - 4 x 106
4r
Fig. 12 Variation of shock front velocity with distance for different values of ini-
tiation energy EQ (from Ref. 20).
values considerably below the C-J propagation velocity DC]; however, even
then a C-J detonation ultimately is established.
In all of the calculations described so far, it has been assumed that the
chemical time tc is much less than the droplet breakup time tb, so that
reaction is assumed to occur at the instant of atomization and vaporization.
This condition may no longer be valid as the wave velocity drops to lower
and lower minimum values with decreasing E0. To account for this effect,
Mitrofanov et al.20 introduced the following differential equation for an
induction distance parameter *P:
1
t = T" exp[E//?r] (26)
At
with W = 1 at the shock and 0 at the point where reaction starts. The
induction time tc is evaluated at local conditions, and so varies throughout
the induction zone. Note that this relation is very closely related to Eq.
(10), which was used by Gubin and Sichel13 in the analysis of spray det-
onation structure. Figure 13 shows the variation of shock front velocity
with distance for various values of E0 when ignition delay is taken into
account. The change is drastic, because below a certain minimum value of
£"0 the wave decays and no detonation is established, a behavior in accord
with experimental observation.
The method of Godunov and Zabrodin43 was used to integrate the gas-
phase equations for the explosion products of the initiating charge and for
the spray in more recent studies.44-45 Voronin and Zhdan45 considered a
detonation tube rather than a high-explosive initiator, and since such ini-
tiators are used frequently in heterogeneous detonation experiments, their
study is of particular interest. Their goal was to simulate the experimental
spray detonation results of Lu et al.46
x 10-6
1 = 1.1827
RO = 192 Jim 2 = 2.1025
Yl = 1.3069 3 = 2.3736
e = 34cc/mole 4 =2.661
5 = 3.2852
Fig. 13 Variation of the shock front velocity with distance for different values of
the initiation energy E0 with ignition delay taken into account (from Ref. 20).
The initiator is a tube of length L0 filled with an H2-O2 mixture attached

to a tube filled with a spray air or spray oxygen mixture. There is no flow,
i.e., zero velocity, at the upstream or left end of the detonation tube, and
pressure p1 and velocity ul on the two sides of the contact surface between
the detonation products and the spray are equal. The Zeldovich self-similar
solution was used to describe the properties of the detonation products in
the detonation tube at the start of the calculation.19 Spray equations [Eqs.
(21) and (22)] were supplemented by the ignition delay relation [Eq. (26)]
written in the following form:
I
(27)
dt dr
A one-dimensional geometry was considered, and the governing gas-phase

equation [Eq. (21)] was supplemented by a wall loss term similar to those
in Eqs. (1-3) to account for wall shear and heat transfer. By using what
the authors call the mobile net method, it was possible to track both the
leading shock, where Eq. (24) applies, and the interface between the ini-
tiator products and the spray.
A nonspray calculation was used to validate the computer code. Cal-
culations were carried out for a stoichiometric mixture of a monodisperse
1400-jULm heptane spray in air and in oxygen and a detonator tube length
of L0 = 13 cm to correspond to the experimental conditions used by Lu
et al.46 As in the simulations described earlier, the wave velocity decays
to a minimum value. In the case of O2, the wave reaccelerates to a steady
C-J wave with a velocity below the C-J value because of the wall losses.
The wave continues to decay in the case of air so that no detonation is
established, in accord with experiment. Typical pressure profiles at differ-
M. SICHEL 473
ent times are shown in Fig. 14 in the case of the O2 atmosphere. Note, in
particular, the development of a secondary wave that overtakes the leading
shock resulting in the reacceleration of this wave. This phenomenon may
be related to the sequential blast wave model of Pierce8 discussed in
Sec. II.
C. Shock-Capturing Methods
The previously described calculations essentially use shock fitting to
describe the motion of the leading shock wave through the introduction
of the shock jump conditions given by Eq. (24). A different approach
based on the FCT algorithm47 was used by Eidelman and Burcat 48 ~ 50 to
solve the direct initiation problem. With the FCT algorithm, it is no longer
necessary to fit the shock to the solution. Instead, the shock is captured
or appears naturally in the appropriate location in the course of the cal-
culations. Of the various methods, the FCT algorithm appears to be one
of the most effective in suppressing extraneous oscillations.49 50
The FCT algorithm was used to solve Eqs. (21) and (22). Again, heat
transfer to the droplets was neglected so that the droplet energy equation
could be replaced by an equation for the conservation of the number of
drops. A combined vaporization and Engel36 shattering model was used
to describe droplet breakup. The drag coefficients given by Eq. (23) were
used in the trajectory equation. At time / = 0, it is assumed that for
r < r0 the fluid dynamic properties correspond to those for a strong blast
wave with energy E0 as described by the Sedov42 solution, but no spray or
dust is present in this region. The region r > r0 contains undisturbed
monodisperse spray or dust particles. The effects of chemical ignition delay
are not included.
A typical sequence of computed velocity and pressure profiles for a
decane cloud with 300-juim droplets and an initiating energy of EQ = 1.56
x 105 J is shown in Fig. 15. The leading shock wave now appears as a
result of the initial and boundary conditions rather than being inserted by
introducing appropriate jump conditions. The wave decays and then reac-
celerates, as is typical of direct initiation. Perhaps of greatest interest is
the appearance of secondary compression or blast waves, denoted by S in
the figure. The appearance of these waves is associated with sprays with
r.m
Fig. 14 Pressure profile evolution for
heptane spray in O2 with detonation
tube initiator (from Ref. 45).
0.30 0.41 LSI

RADIUS Rlfl]
Fig. 15 Variation of gas pressure and velocity with radius for a decane-oxygen
mixture. 5 denotes secondary wave (from Ref. 51).
E(J) *RZ F/A mixture
2x10 6 11cm 0.284 decane-oxygen
io6 10cm 0.158 decane-oxygen

106 13cm 0.412 wheat -air
IO6 13cm 0.94 RDX-air
IO6 12cm 0.27 wheat- air
R[m]
Fig. 16 Pressure behind the leading shock front for various combustible mixtures
and initiation energies (from Ref. 51).
M. SICHEL 475
larger droplets and the resultant delay in atomization and combustion heat
release, and lends further credence to the sequential blast model of Pierce.8
The variation of the pressure behind the leading shock wave with shock
radius is shown in Fig. 16 and includes several curves calculated for wheat
dust and RDX dust-air mixtures. As in the results reported earlier, it can
be seen that the pressure reaches a minimum and then increases to a final
C-J value. Chemical ignition delays were not considered in these calcula-
tions, so that the phenomenon of a critical initiation energy below which
detonation fails is not observed.
The transition from a blast wave to a fully established detonation has
been investigated by Eidelman and Sichel51 using the FCT algorithm. A
typical example of the structure of a detonation computed in this way,
including a secondary wave, is shown in Fig. 17 for 300-jjim decane droplets
with respect to laboratory coordinates, and includes the variation of the
Mach number, the particle velocity, the gas velocity, and the speed of
sound. In a certain sense, these results are an example of an overall sim-
ulation of the structure of a heterogeneous detonation, discussed in Sec.
II, in that such phenomena as secondary blast waves do not have to be
introduced ad hoc, but simply appeared during the course of the numerical
calculation. The results of this simulation also have been used to determine
the ground and dynamic impulse of a heterogeneous detonation. 52
1000
C-J
plane
500
100
0.5 0.6 0.7 0.8 0.9 R[m]
Fig. 17 Structure of a detonation

wave with a secondary shock near the
leading shock front (from Ref. 51).
IV. Discussion and Conclusions

The physics of spray atomization or dust ignition and combustion is a
major source of difficulty and uncertainty in the meaningful simulation of
heterogeneous detonations. Each material has its own characteristics and
may involve a completely new set of parameters. Thus, for example, the
process of breakup and mixing of highly volatile fuels will be different from
that of nonvolatile fuels. The situation is complicated further by the fact
that the velocity relative to the particles ranges from supersonic to the
Stokes flow regime. Large-droplet breakup may be accompanied by wake
explosions, whereas small particles may just vaporize with negligible strip-
ping or droplet breakup. The detailed numerical simulation of these indi-
vidual particle breakup and reaction processes is a major challenge, and
is essential for continued progress in the understanding of heterogeneous
detonations.
The numerical simulation of heterogeneous detonations is still at a begin-
ning stage. There has been some success in the simulation of the steady-
state structure and of direct initiation. Most of the structure studies are
based on the one-dimensional Zeldovich-Neumann-Doering structure, and
all of the simulations considered here treat only one-dimensional flow.
Transverse waves and the resulting detonation cells are known to play an
important role in the case of gaseous detonations; for heterogeneous det-
onations with a low enough dust or spray density, it would appear reason-
able that such structures might play a role. It would also appear that
localized explosions of droplets, as observed in some spray detonation
experiments, need to be taken into account in some cases. However, the
proper simulation of such effects requires a two-dimensional treatment at
the very least and, thus, would involve two-dimensional, unsteady two-
phase flow.
With few exceptions, most simulation studies consider monodisperse
uniform particle clouds. However, the particle size distributions, and par-
ticularly nonuniformities in the particle concentration, play an important
role in the initiation and propagation of heterogeneous detonations and
must be incorporated into future simulations. There are, of course, also
many nonsymmetrical cloud configurations for which explosive and damage
characteristics are of interest. Although there has been considerable prog-
ress in simulating complex geometries in the case of gaseous explosives,
much remains to be done in the detonation of complex heterogeneous
cloud configurations.
This chapter is by no means comprehensive; however, it is hoped that
an introduction to the subject of heterogeneous detonations and their
numerical simulation has been provided. There has been considerable prog-
ress, and it is anticipated that a careful combination of simulation and
experiment will lead to an enhanced understanding of heterogeneous
detonations.
Acknowledgment
Partial support for the preparation of this chapter was provided by the
Department of Health and Human Services, National Institute of Occu-
pational Safety and Health, under Grant OHO1122-06.
M. SICHEL 477
References
ôrisov, A. A., and Gelfand, B. E., "Review of Papers on Detonation of Two
Phase Systems, "Arch. Termodynamik, Spalaia, Vol. 7, No. 2, 1976, pp. 273-287.
2
Sichel, M., "Modeling Gaseous and Heterogeneous Detonation Phenomena,"
Transactions of Twenty-Seventh Conference of Army Mathematicians, Army Research
Office, Raleigh/Durham, NC, Rept. 82-1, 1982.
3
Sichel, M., "The Detonation of Sprays: Recent Results," Fuel Air Explosions,
University of Waterloo Press, Waterloo, Ontario, Canada, 1982, pp. 265-302.
4
Gordon, S., and McBride, B. J., Computer Program for Calculation of Complex
Chemical Equilibrium Compositions, Rocket Performance, Incident and Reflected
Shocks, and Chapman-Jouguet Detonations, NASA SP-273, 1971.
5
Gordon, S., McBride, B. J., and Zeleznik, F. J., "Computer Program for
Calculation of Complex Chemical Equilibrium and Applications," NASA-CET86,
1987.
6
Oran, E. S., Kailasanath, K., and Guriguis, R. H., "Numerical Simulations
and the Development and Structure of Detonations," Dynamics of Explosions,
Vol. 114, edited by M. Summerfield, Progress in Astronautics and Aeronautics
Series, AIAA, Washington, DC, 1988, pp. 155-169.
7
Taki, S., and Fujiwara, T., "Numerical Simulations on the Establishment of
Gaseous Detonation," Dynamics of Shock Waves Explosions and Detonations, Vol.
94, edited by M. Summerfield, Progress in Astronautics and Aeronautics Series,
AIAA, New York, 1984, pp. 186-200.
8
Pierce, T. H., "Experimental and Theoretical Study of the Structure of Two-
Phase Detonations in Sprays," Ph.D. Thesis, Univ. of Michigan, Ann Arbor, MI,
1972.
9
Dabora, E. K., "A Model for Spray Detonation," Acta Astronautica, Vol. 6,
1979, p. 269.
10
Lee, J. H., "Initiation of Gaseous Detonation," Annual Review of Physical
Chemistry, Vol. 28, 1977, p. 75.
n
Borisov, A. A., Gelfand, B. E., Gubin, S. A., Kogarko, S. M., and Podgre-
benkov, A. L., "The Reaction Zone of Two Phase Detonations," Astronautica
Acta, Vol. 15, 1970, pp. 411-418.
12
Gubin, S. A., Borisov, A. A., Gelfand, B. E., and Gubanov, A. V., "Deto-
nation Rate in Mixture of Fuel and Gaseous Oxidizing Agents," Combustion of
Explosion and Shock Waves, Vol. 14, 1972, p. 71.
13
Gubin, S. A., and Sichel, M., "Calculation of the Detonation Velocity of a
Mixture of Liquid Fuel Droplets and a Gaseous Oxidizer," Combustion Science
and Technology, Vol. 17, 1977, pp. 109-117.
14
Wolanski, P., Lee, D., and Sichel, M., "The Structure of Dust Detonations,"
Dynamics of Shock Waves Explosions and Detonations, Vol. 94, edited by M.
Summerfield, Progress in Astronautics and Aeronautics Series, AIAA, New York,
1984, pp. 241-263.
15
Lee, D., and Sichel, M., "The Chapman-Jouguet Condition and Structure of
Detonations in Dust-Oxidizer Mixtures," Dynamics of Explosions, Vol. 106, edited
by M. Summerfield, Progress in Astronautics and Aeronautics Series, AIAA, New
York, 1986, pp. 505-521.
16
Fan, B. C., and Sichel, M., "A Comprehensive Model for the Structure of
Dust Detonations," Proceedings of Twenty-Second Symposium (International) on
17
Rudinger, G., "Relaxation in Gas Particle Flow," Non-Equilibrium Flow,
Marcel-Dekker, New York, 1970, p. 119.
18
Boiko, V. M., Fedorov, A. V., Fomin, V. M., Papyrin, A. N., and Soloukhin,
R. L, "Ignition of Small Particles Behind Shock Waves," Shock Waves Explosions
and Detonations, Vol. 87, edited by M. Summerfield, Progress in Astronautics and

Aeronautics Series, AIAA, New York, 1983, pp. 71-87.
19
Zeldovich, I. B., and Kompaneets, A. S., Theory of Detonation, Academic,
New York, 1960.
20
Mitrofanov, V. V., Pinaev, A. V., and Zhdan, S. A., "Calculations of Det-
onation Waves in Gas Droplet Systems," Ada Astronautica, Vol. 6, 1979, pp. 281-
294.
21
Williams, F. A., Combustion Theory, 2nd ed., Benjamin/Cummings, Menlo
Park, CA, 1985.
22
Hayes, W. D., 'The Basic Theory of Gasdynamic Discontinuities," Princeton
Series in High Speed Aerodynamics and Jet Propulsion: Fundamentals of Gas
Dynamics, Vol. Ill, edited by H. Emmons, Princeton Univ. Press, Princeton, NJ,
1958, pp. 416-482.
23
Williams, F. A., "Structure of Detonations in Dilute Sprays," Physics of Fluids,
Vol. 4, 1961, pp. 1434-1443.
24
Ragland, K. W., Dabora, E. K., and Nicholls, J. A., "Observed Structure of
Spray Detonation," Physics of Fluids, Vol. 11, 1968, pp. 2377-2388.
25
Mullins, F. B., "Studies on the Spontaneous Ignition of Fuels Injected into a
Hot Air Stream Ill-Effect of Chemical Factors upon the Ignition Delay of Kerosine-
Air Mixtures," Fuel, Vol. 32, 1953, pp. 327-379.
26
Oran, E. S., Boris, J. P., Young, T. R., Flanigan, M., Burks, T., and Picone,
M., "Numerical Simulations of Detonations in Hydrogen Air and Methane Air
Mixtures," Proceedings of Eighteenth Symposium (International) on Combustion,
27
Ural, E. A., Sichel, M., and Kauffman, C. W., "Shock Wave Ignition of
Pulverized Coal," Proceedings of 13th International Symposium on Shock Tubes
and Waves, State University of New York Press, Buffalo, NY, 1982, pp. 809-817.
28
Sichel, M., Baek, S. W., Kauffman, C. W., Maker, B., and Nicholls, J. A.,
"The Shock Wave Ignition of Dusts," AIAA Journal, Vol. 23, 1985, pp. 1375-
1380.
29
Lee, F. P., Kauffman, C. W., Sichel, M., and Nicholls, J. A., "Detonability
of RDX Dust in Air/Oxygen Mixtures," AIAA Journal, Vol. 24, Nov. 1986, pp.
1811-1816.
30
Bar Or, R., Sichel, M., and Nicholls, J. A., "The Propagation of Cylindrical
Detonations in Monodisperse Sprays," Proceedings of Eighteenth Symposium
(International) on Combusion, The Combustion Institute, Pittsburgh, PA, 1981,
pp. 1599-1606.
31
Cherepanov, G. P., "The Theory of Detonations in Heterogeneous Systems,"
PMTF, Vol. 4, Moscow.
32
MacCormack, R. W., "Numerical Solution of the Interaction of Shock Wave
with a Laminar Boundary Layer," Proceedings of Second International Conference
on Numerical Methods in Fluid Dynamics, Lecture Notes in Physics, Vol. 8, edited
by M. Holt, Springer-Verlag, New York, 1971, pp. 151-163.
33
Nigmatulin, R. L, Fundamentals of the Mechanics of Heterogeneous Media,
Nauka, Moscow, 1978.
34
Wallis, G. B., One Dimensional Two Phase Flow, McGraw-Hill, New York,
1970.
35
Zhdan, S. A., "Calculation of a Spherical Heterogeneous Detonation," trans-
lated from Fizika Goreniya i Vzryva, Vol. 12, No. 4, 1976, pp. 586-594.
36
Engel, O. G., "Fragmentation of Waterdrops in the Zone Behind an Air
Shock," Journal of Research of the National Bureau of Standards, No. 3, 1958, pp.
245-280.
37
Kestenboim, K. S., Roslyakov, G. S., and Chudov, L. A., Point Explosion,
Nauka, Moscow, 1974.
M. SICHEL 479
38
Landau, L. D., Meyman, N. N., and Khalatnikov, J. M., "Numerical Methods
of Integration of Partial Equations by Net Method," Proceedings of All-Union
Math Congress, Vol. 3, Izd AN SSR, Moscow, 1958, pp. 92-100.
39
Bach, G. G., Knystautas, R., and Lee, J. H., "Initiation Criteria for Diverging
Gaseous Detonations," Proceedings of Thirteenth Symposium (International) on
40
Zhdan, S. A., "Calculation of Heterogeneous Detonation Taking Into Account
Deformation and Breakdown of Fuel Droplets," translated from Fizika Goreniya
i Vzryva, Vol. 13, No. 2, 1977, pp. 258-263.
41
Nicholls, J. A., Sichel, M., Fry, R., and Glass, D. R., "Theoretical and Exper-
imental Study of Cylindrical Shock and Heterogeneous Detonation Waves," Ada
Astronautica, Vol. 1, 1974, pp. 385-404.
42
Sedov, L. I., "Propagation of Strong Blast Waves," Applied Mathematics and
Mechanics, Vol. 10, 1946, p. 242.
43
Godunov, S. K., and Zabrodin, A. V., Numerical Solution of Multi-Dimen-
sional Problems of Gas-Dynamics, Nauka, Moscow, 1976.
44
Zhdan, S. A., "Calculation of the Initiation of a Heterogeneous Detonation
with a Charge of Condensed Explosive," Fizika Goreniya i Vzryva, Vol. 17, No.
6, 1981, pp. 674-679.
45
Voronin, D. V., and Zhdan, S. A., "Calculation of Heterogeneous Detonation
Initiation for a Hydrogen-Oxygen Mixture in an Explosion Tube," Fizika Goreniya
i Vzryva, Vol. 20, 1984, pp. 112-117.
46
Lu, P. L., Slagg, N., and Fishburn, B. D., "Relation of Chemical and Physical
Processes in Two Phase Detonations," Acta Astronautica, Vol. 6, 1979, p. 815.
47
Boris, J. P., and Book, D. L., "Flux Corrected Transport III: Minimal Error
FCT Algorithms," Journal of Computational Physics, Vol. 20, 1976, p. 397.
48
Eidelman, S., and Burcat, A., "Evolution of a Detonation Wave in a Cloud
of Fuel Droplets: Part I. Influence of Igniting Explosion," AIAA Journal, Vol. 18,
1980. pp. 1103-1109.
49
Eidelman, S., and Burcat, A., "Numerical Solution of a Non-Steady Blast
Wave Propagation in Two Phase ("Separated Flow") Reactive Medium," Journal
of Computational Physics, Vol. 39, 1981, p. 456.
50
Eidelman, S., and Burcat, A., "The Mechanism of a Detonation Wave
Enhancement in a Two-Phase Combustible Medium," Proceedings of Eighteenth
PA, 1981, pp. 1661-1670.
51
Eidelman, S., and Sichel, M., "The Transitional Structure of Detonation Waves
in Multi-Phase Reactive Media," Combustion Science and Technology, Vol. 26,
1981. pp. 215-224.
52
Eidelman, S., and Sichel, M., "Static and Dynamic Impulses Generated by
Two Phase Detonations," Physics of Fluids, Vol. 25, 1982, pp. 38-44.
Chapter 15
Deflagration-to-Detonation Transition in Reactive

Granular Materials
Melvin R. Baer, Jace W. Nunziato, and Pedro F. Embid
I. Introduction
T HE modes of flame spread and the deflagration-to-detonation transi-

tion (DDT) in gas-permeable, reactive granular materials have been
the subject of extensive study. Prior to the 1950s, much of the research
was conducted in the USSR and the United Kingdom, and the results are
collected in a monograph by A. F. Belyaev et al. 1 An overview of current
studies in this field, including the American works relevant to propellant
hazards, is found in the state-of-the-art review article by Bernecker. 2 De-
spite the wealth of previous studies, many questions remain concerning
the fundamental mechanisms that control the initiation and the acceleration
of the combustion process.
In general, combustion in granular materials begins with the ignition of
a few grains by a thermal or mechanical source. Initially, the combustion
process is slow and dominated by heat conduction within and between
grains. Andreev3 first postulated that the hot product gases generated
during the early stage penetrate into the pores of the unreacted material
and, by preheating the grains, augment the rate of flame spread by several
orders of magnitude above the deflagration rate driven by thermal con-
duction alone. This mode of flame spread can be self-accelerating and,
under strong confinement, high gas pressures are produced, which lead to
conditions favorable for detonation.
Several experimental studies have been insightful in revealing the phys-
ical processes during DDT. The pioneering work of Griffiths and Groocock4
confirmed the existence of a convectively driven flame front during flame
spread in a column of HMX (high melting explosive, a polycrystalline
organic material with the chemical structure C4H8N8O8) and showed that
the onset of detonation was well removed from the location at which
combustion originated. Furthemore, the predetonation column length was
481
482 DOT IN REACTIVE GRANULAR MATERIALS
shown to depend on the permeability of the granular material. In later

studies, Bernecker et al.5 and Sandusky and Bernecker 6 investigated the
effects of high-pressure gas generation, which causes mechanical load trans-
fer to the granular reactant. Their studies confirmed that considerable bed
compaction occurs in the region ahead of the flame front, whereby the
permeability of the granular bed is reduced, leading to significant pressure
buildup behind the flame front. The granular bed is then compressed, and
combustion is accelerated until a shock wave has developed. (Shock for-
mation is a key process in the initiation of solid explosives and has been
described in the models of Macek,7 Tarver et al., 8 and Forest and Mader. 9
They treat DDT as a process in which the gas generated due to combustion
acts strictly as a "driving piston" on the material and does not permeate
into the explosive. In this case, the onset of detonation occurs far in front
of the flame front and is due to the coalescence of the pressure disturbances
that results from the rapid gas generation behind the flame front.) Once
a shock wave is formed, local "hot spots" release additional energy, which
enhances wave growth to detonation. Various mechanisms for this localized
heating have been proposed, including microscale shock focusing during
pore collapse, plastic work heating, and intergranular friction at grain
boundaries or in regions of high shear stress.1(U1
Since the mechanical response of the granular material appears to play
a key role in the DDT process, recent experimental studies have focused
on the effects of compaction induced by projectile or piston impact. Studies
by Green et al.12 demonstrated that compaction alone can trigger a com-
bustion event that readily accelerates to detonation, provided that a certain
level of porosity is attained. In these experiments, the energetic material
fractures after projectile impact and wave reflections with recompaction
produce conditions favoring accelerated combustion. In other studies,
Sandusky and Liddiard13 used piston impacts to produce dynamic com-
paction of a column of granular material embedded with tracers that were
followed with flash x rays. Their studies showed that compaction proceeds
as a shocklike wave that travels at a speed much less than the material
sound speed. In related experiments, McAfee and Campbell 14 studied gran-
ular HMX combustion within thin-wall steel tubes and showed that multiple
compaction waves form as a result of the interplay of compaction, com-
bustion, and the loss of wall confinement. In all of these studies, low-
velocity impacts produce low-amplitude compressive waves, which are in-
sufficient to cause direct shock initiation 15 ; yet, DDT was observed. In
view of these general experimental observations, it is now recognized that
DDT in a granular energetic material essentially consists of four regimes:
conductive burning, convective burning, compaction-induced combustion,
and detonation.
It is clear that the coupled thermal/mechanical processes associated with
the combustion of a granular explosive must be included in a complete
flowfield model. However, the formulation of the conservation laws that
govern these complex processes is currently a controversial subject in the
multiphase flow literature (see, for example, Ref. 16). The controversy
lies in the derivation of these laws and in their final form. No procedure
M. R. BAER ET AL. 483
is accepted universally and several different approaches have been em-

ployed. Most of the current modeling utilizes a continuum approach that
describes the field equations assuming the coexistence of the phases at
every resolved point in the flowfield.
In the continuum mixture theory approach, conservation laws are de-
termined for each phase, and these laws account for the exchange of mass,
momentum, and energy between phases (i.e., drag, etc.). In addition,
conservation of mass, momentum, and energy is required for the overall
mixture, resulting in certain constraints imposed on the interactions be-
tween phases. Previous studies, relevant to DDT modeling, include Refs.
17-25. In considering a complete treatment of reactive two-phase flow in
total nonequilibrium, a closure problem exists. This problem results from
including the volume fraction of each phase as an independent kinematic
variable (see Ref. 26 for a detailed discussion). A means for model closure
proposed in Refs. 27 and 28 uses a rate equation for the volume fraction
that permits compressibility of each phase and describes compaction of the
granular material. This evolutionary equation, analogous to the pore-
collapse model developed by Carroll and Holt, 29 is driven by pressure
differences between the solid and gas phases, and is resisted by the internal
frictional effects of the granular bed. This approach, formulated to be
consistent with themodynamics, was used herein.
A two-phase reactive-flow model has been developed in the context of
the continuum theory of multiphase mixtures. Although this model has
been applied successfully to a variety of reactive energetic materials, this
work will primarily examine the granular explosive HMX. In the sections
that follow, we provide a brief overview of the theory, including appro-
priate choices for constitutive models. In treating multiphase combustion,
particularly by using numerical simulation, a complete specification of the
model that is consistent with the equations of motion and thermodynamics
for every possible choice of parameters and field values is an absolute
prerequisite. Then, we examine the mathematical character of the multi-
phase flow equations. In later sections, we discuss several numerical meth-
ods in one and two dimensions, and show demonstrative calculations that
are compared with experimental observations. Finally, we conclude with
a discussion of unresolved numerical and technical issues, and suggest
possible areas for further study.
II. Theory for Reactive Two-Phase Flows

A. Equations of Motion and Thermodynamics
Chemically reacting mixtures, such as porous or granular explosives with-
out binders, will be assumed to consist of two phases: solid granular reactant
(a = s) and interstitial gas products (a = g). On some appropriate length
scale, each phase may be viewed as occupying every spatial point jc in the
body. Physically, however, this is not the case, since each phase occupies
a volume distinct from the others. Thus, we assign to each spatial point jc
a phase density ya and a volume fraction 4> w . The phase density represents
the mass of phase a per unit volume of phase a, and the volume fraction
represents the fraction of space occupied by phase a at the point x. Since

the entire volume of the mixture is occupied,
S <!>* = fc + <!>* = 1 (1)
and the density of the mixture is the sum of the partial densities p(l:
P = 2 Pa> Pa = <baya (2)
The motion of each phase is characterized by the velocity

va = va (x,t) (3a)
va = \va\ (3b)
The general equations of motion and energy for each of the phases can be
expressed as
P* = - P«v ' va + cl (4)
pal>fl = V • <rfl + p tt 6 fl + ml ~ clva (5)
PA = va ' ^va 4- para + el - ml • va - cl (ea - V2al2) (6)
where the plus superscript denotes a phase interaction quantity and the
overdot denotes the time derivative following the motion of phase a:
L = f + va • V/a (7)
In Eqs. (4-7), V denotes the spatial gradient, cl is the mass exchange due
to chemical reactions, ba the external body force (e.g., gravity), ml the
momentum exchange resulting from the forces (such as drag) acting at
the interfaces between phases, ea the internal energy (per unit mass), ra
the external heat source, and el the energy exchange due to the local heat
transfer between phases and the work done at the internal phase bound-
aries. In addition, a fl is the symmetric stress tensor; it will prove useful to
express crfl in terms of the phase pressure pa and the shear stress Trt:
aa = -$apal + Tfl (8a)
tr(Tj = 0 (8b)
If we add together the equations of motion and energy for each phase,
then we should recover the well-known equations of motion and energy
for a single material. Indeed, this requirement imposes restrictions on the
interaction terms:
2 cl = 0 (9a)
2>J = 0 (9b)
S el = 0 (9c)
The second law of thermodynamics serves to establish restrictions on

forms of the constitutive equations appropriate for reactive mixtures. In
terms of the Helmholtz free energy
tya = ea ~ Tasa
where Ta is the absolute temperature and sa the entropy, the second law
can be expressed as the thermodynamic inequality
(4 - ea + sata + ra) - sac* <0 (10)

a J
It is important to note that, in writing the first and second laws of

thermodynamics [Eqs. (6) and (10)], we have interpreted the phase tem-
perature Ta as the local mean temperature. Heat conduction effects at
length scales large in comparison to particle size are assumed to be neg-
ligible. This is not to imply, however, that heat conduction in individual
grains is not important. Indeed, we include the effect of conduction in
grains due to the presence of hot-gas convection through a surface tem-
perature function introduced later.
B. Constitutive Models and Closure

Given the equations of motion and energy, it remains to consider the
specific constitutive models that are appropriate for granular energetic
materials without binders. The individual phases are clearly separated phys-
ically, and, thus, it is plausible to require that the thermodynamic variables
for a given phase depend only on the state variables (ya, v(l, <)>„, Ta) cor-
responding to that phase. In particular, for the solid reactant, a nonlinear
thermoelastic equation-of-state description is an appropriate model choice30:
p,
P =
Shear stresses are neglected in this description and the constants are obtained
from Hugoniot and thermophysical data.
The Jones-Wilkins-Lee equation of the state is chosen for the gas phase
because it describes the very dense thermodynamic states encountered at
detonation as well as ideal gas behavior at low pressures and temperatures31:
^
The constants for this model are best fit to hydrodynamic calculations of
explosively driven, cylinder expansion experiments and to the Chapman-
Jouguet detonation properties estimated from chemical equilibrium cal-

culations for a mixture of gaseous reaction products.
By also requiring the equations of state to satisfy the thermodynamic
inequality (10); it can be shown (see Ref. 26) that the Helmholtz free
energy, tya = tya(ya, Ta, <|>fl), is related to the phase pressure pa and the
entropy sa by familiar thermodynamic identities,
P. = ^ (13a)
Oâ
and the phase interactions must satisfy the dissipation inequality

1
I
Ta Ia
1
~ (P« " Pa)$a \ - 0 (14)
\ 'Y« ^ / J
where
Pa = ch/y, ^ (15)
The configuration pressure p,, is a function of the solid volume fraction

and reflects the contact forces between grains, evaluated from quasistatic
compaction data, whereas the phase pressure ps represents the average
hydrostatic stress within the solid grains. We assume $g = 0 without loss
in generality. Specific expressions for the phase interactions cj, mj, and
el are formulated on the basis of microstructural models for the detailed
chemical kinetics, the local drag, and the local heat transfer, and must be
consistent with Eqs. (1) and (9). As one would expect, it is these models
that are most crucial in the prediction of initiation and flame spread to
detonation in condensed-phase energetic mixtures.
It is important to note that specification of the equations of state and
expressions for the phase interactions are not sufficient to provide closure;
i.e., if the response of the mixture does indeed depend on the eight fields
"yfl, Ta, c|v and va (a = s,g), then the set of six conservation equations
[Eqs. (4-6)] is incomplete. This difficulty is well known in the theory of
two-phase flows, and various solutions have been proposed. Herein, we
propose an additional evolutionary equation for the solid volume fraction 26 :
AA
M"c
(P,- P,
This equation is motivated by the dissipation inequality equation [Eq. (14)]
and permits us to account for both the compressibility of the constituents
and the compaction of the granular materials by recognizing the volume

fraction as an independent, internal degree of freedom within the mixture
analogous to the internal degrees of freedom that arise in real gases. In
essence, Eq. (16) represents two contributions to changes in the solid
volume fraction. The first term on the right-hand side asserts that the
compaction of the granular mixture has a rate that depends on the com-
paction vicosity JJLC, and the mixture can exhibit hysteresis upon quasistatic
loading and unloading. The second term on the right-hand side of Eq. (16)
accounts for changes in the solid volume fraction due to combustion. Clearly,
Eqs. (1) and (16) provide closure.
To describe the momentum and energy interactions between the phases,
we assume constitutive equations of the form 26
/HJ = -HI* = pgV<k - 2) + cjvs (17)
~ (PS - P,) (Ps -ps-M + W- Escl (18)
where 9) = (D 4- c|/2)(i?5 - vg) and Vt = h(Ts - Tg) are, respectively,

the interphase drag and heat transfer. For the interphase drag, an exchange
coefficient, D, is defined by using experimental data of the permeability
of the granular material. Similarly, interphase heat transfer is represented
by using existing convective heat-transfer correlations for the exchange
coefficient, h, appropriate for packed beds. The effects of high-speed flows
are included as Reynolds number corrections to these exchange coeffi-
cients.
In terms of multiphase mixtures, the corresponding reaction rate model
is an expression for the mass exchange c*. In the current work, we shall
express this mass exchange by combining the effects of compressive com-
bustion and grain burning into a relationship similar to that posed by Lee
and Tarver32:
T
~ = ~
where ys is the reactant material density, ^ the undisturbed solid volume

fraction, TH a "hot-spot" reaction time characterizing compaction-induced
combustion, (S/V) the specific surface area of the granular reactant, and
apng the pressure-dependent granular surface burn rate that is activated by
the Heaviside function, H(T{ - T*). The Heaviside function has a value
of 1 when the granular surface temperature Ti exceeds a critical "ignition"
temperature T*, otherwise it is zero. The ignition temperature is inter-
preted as the temperature where rapid decomposition occurs. For the gran-
ular explosive HMX, this temperature is close to its melt temperature.
The first term of this expression applies only in compacted regions (i.e.,
4>5 > c|>°) and reflects the effect of compaction-induced combustion resulting
from the formation of hot spots due to grain distortion. Following the
experimental observations of Sandusky and Bernecker, 6 this reaction time

scale, T//, is proprotional to the inverse of the square of mixture pressure:
T*1 - bp2m (20)
The coefficient b is determined to best fit available data on combustion-

assisted compaction wave speed.
The second term of Eq. (19) represents the effect of grain burning. A
grain recession rate of a = 0.001 m/s is representative of HMX at a pressure
of 100 MPa, as taken from strand data of Ref. 33, and the rate law exponent
is n ~ 1. The specific surface area of the granular reactant reflects modi-
fications due to grain particle recession and the effect of reduced surface
area during compaction; thus,
d
Vf P
where dp is the unreacted surface mean particle diameter.

During grain burning, the surface layers of the granular reactant are first
ignited during flame spread. The grain surface temperature Tt is the tem-
perature associated with the granular particle surfaces separating the solid
and gas phases. This temperature field is determined by using a relationship
that approximates the heat conduction thermal field within particles sub-
jected to convective heat transfer from the permeating hot decomposition
gases. Thermal conduction effects within individual solid particles are re-
solved by using an integral method with assumed thermal distributions,
and granular surface temperatures are determined along solid-phase
streamlines. The resulting thermal fields are defined by an evolutionary
equation, in terms of a surface temperature function £, given as
t = ~ + vs • V£ = ^ (7; - T,) (22)
where Tg is the local bulk gas temperature, a, the solid thermal diffusivity,
and Bl the Biot modulus based on the local gas film coefficient hg, solid
thermal conductivity ks, and the particle diameter dp:
Bt = (hgdp)/2ks (23)
The interface temperature is a composite of solutions valid as Bf —> 0 and

BJ —> oc and is given by
_ rj + B,Tg
' (5 + B,)(l + £,-)
B,(T,
(Z
'
and Ts is the local bulk temperature of the solid phase. Detailed devel-
opment of this evolutionary equation and comparison to exact solutions
are discussed in Ref. 34.
III. One-Dimensional Initial Value Problems

A. Field Equations and Boundary Conditions
In one dimension, the conservation equations can be written in con-
servative form as
Conservation of mass:
Conservation of momentum:
- (pava) 4- — (pavl + $apa) - p? —^ - m* (26)

ot uX oX
Conservation of energy:
wv
• \ -j ra _ ;£ ÔT\
LX
3t dX ~" a a a f> s ^
where Ea is the total energy,

Ea - pa(efl + i vl) (28)
As discussed later, it is convenient to separate the volume fraction gradient
terms from the phase interactions; thus, the phase interactions ra* and
e* are related to the phase interactions m* and el by
m* = -m* = m+ - pK —— (29)
The description of two-phase combustion is completed with the saturation

constraint given by Eq. (1).
For the purpose of numerical computation, an equation for the solid-
phase material density is written in conservation form by using Eqs. (2),
(4), and (16) combined to produce the following equation:
where
- -^-M (Ps - Pg - PJ ( 32 )
Thus, all of the equations governing combustion in two-phase systems [Eqs.

(25-27) and (31)] are now in conservation form.
In formulating initial boundary-value problems for two-phase reactive
flows, it remains only to specify the initial and boundary conditions. In
these studies, we consider flow in a finite column of length L. Furthermore,
initially we consider the region to be undisturbed and uniform. For rigid
confinements, reflection boundary conditions are employed that yield van-
ishing gradients of the dependent variables: phase density, phase temper-
ature, and volume fraction. Moreover, the velocity of the gas and the solid
vanish at these endpoints, which is equivalent to the second derivative of
the mass flux being zero. Thus, for t > 0,
^ = *I* = ^ = 0 (33)
dx dX dx
(34)
At x = 0 and L, note the adiabatic boundaries also imply that
^= ^= Oi ,> o (35)
dX dx
In implementing impact boundary conditions, a coordinate transfor-

mation is introduced:
x
T] = ~ x°® T, e (0,1) (36)
-
where x0(t) and xL(t) are, respectively, the right and left locations of the
boundaries relative to a fixed space. With this transformation, boundary
motion appears explicitly in the transport equations and the domain is
transformed into a fixed space, with boundary conditions imposed by using
Eqs. (34-35).
B. Mathematical Structure of Reactive Multiphase Flow Equations

In this section, we briefly describe the mathematical character of the
multiphase flow equations and provide the necessary ingredients to recast
the description into characteristic form. Of primary interest is the dem-
onstration that these field equations are well-posed mathematically.
Following single-phase hydrodynamic theory, it is convenient to choose
a solution vector, U = (ys, vs, ss, cj>5, yg, vg, sg)T, and represent the field
equations in matrix form as
(37,
M. R. BAER ET AL 491
where A(U) is the 7 x 7 matrix,

Vs Is 0 0 0 0 0
2
CS l
s TT
s s 0 0 0
Js 4w,
0 0 vs 0 0 0
0 0 0 0 0 0 (38)
0 0 0 («-,-.,) vg i* 0
0 0 0 0 VK r gTX
yK
1 1
0 0 0 0 vg_ 0 0
Z(U) is a 7 x 1 matrix of algebraic source terms, and a Griineisen coef-
ficient Fa is defined for each phase as
ra = - I -^ I (39
In addition, the frozen sound speed of each phase, ca, is defined by
(40)
1) Eigenvalues of A (U) are determined from det[A (U} — X/] = 0, which

can be shown to yield
X = (vs - cs, vs, vs + cs, vg - cg, vg, vg + cg}T (41)
Hence, A has real eigenvalues, as required for well-posedness, and they
closely parallel those known for single-phase flow.
2) Right eigenvectors ofA(U) are obtained following tedious matrix al-
gebra. The final forms are given as a matrix R = (rl, ..., r7)7xl whose
columns are the right eigenvectors of A:
1 1 1 0 0 0 0
_£i 0 £i 0 0 0 0
0
c2s
0
(/>,-/>* + FA) 0 0 0
ysTsTs <bsysrsTs
0 0 0 i 0 0 0 (42)
yg(vs - vgy
0 0 0 1 1 1
2
c g(vs-vg)
0 0 0 0
$ [(vs - vg)2 - c2g] yg
%?
c2
0 0 0 0 0 - 0
ygrgT}
The system is said to be hyperbolic if the matrix R is nonsingular. R is

generally nonsingular; however, certain elements become infinite when
vs — vg = ±Cg, which corresponds to a local choked-flow state. At this
condition, two eigenvalues are coincident, and the system is said to possess
a parabolic degeneracy. Special treatment of the description is required at
this condition, as discussed in Ref. 35.
3) Left eigenvectors of A(U). Given the matrix R whose columns are
right eigenvectors of A, the rows of the matrix L = R~l are the left
eigenvectors of A, normalized such that /,ry = 8/y, /,/ = 1,..., 7. The resulting
matrix is L — (/ 1? ..., / 7 )fx? :
0 0 0
2 2cs
0 0 0 0 0
1 ^_
0 0 0
2 2^
0 0 0 0 0 (43)
0 0
0 0
0 0
4) Characteristic form of the reactive multiphase equations. Finally, the

reactive multiphase equations can be written in characteristic form after
multiplication of the left eigenvector /, associated with the eigenvalue X, as
idU dU\ idU dU\ 1 , 7^

L\
l — H- A— = /, — 4- X,—
l - //Z((7), = (44)
\dt dx \dt dxj ' v J
Having established this result, generalized Riemann invariants, discontin-
uous states, and simple waves can then be determined (see Ref. 35). In
addition, this provides the necessary framework for characteristic-based
numerical methods.
C. Numerical Methods in One Dimension

Typical of most reactive-flow models, the coupling of transport of mass,
momentum, and energy with the effects of phase interactions, such as
chemical reactions, leads to flow descriptions that exhibit disparate length
and time scales. In these transient flows, the effects of transport and phase
interaction have dominating influences at different times, and the com-
bustion processes may produce regions of high gradients and shocks. An
accurate numerical solution of these flows requires a method that can
resolve the relevant chemical and physical aspects of the multiphase flow
M. R, BAER ET AL. 493
for all appropriate length scales and at all times. In the brief discussions
to follow, we review several methods explored in numerical solution of the
Eulerian reactive multiphase equations. Methods for Lagrangian formu-
lations, applied to multiphase flows, are still in development and are not
discussed in this chapter.
Method of Lines
The method of lines (MOL), a popular numerical approach to the so-
lution of combustion problems, can efficiently resolve the temporal stiffness
of a model by using specialized time-integration algorithms. In one di-
mension, numerical solutions of several thousand stiff, ordinary differential
equations (ODE's) can be obtained using an acceptable amount of com-
putational time and storage.
In our studies, using the method of lines, we have implemented the
recommendations of Hyman36 for a general system of the conservative
equations given as
* = -* + Z (45)
dt dx
where S is the seven-dimensional vector whose elements are the dependent

variables, G the flux vector that can be a nonlinear function of S, and Z
the source vector. A discrete grid is then defined consisting of N 4- 1
points, denoted by an index /, with the left-hand boundary i = 1 and the
right-hand one at i = N + 1. At each node, the partial differential equa-
tions are evaluated:
z
?
dt = -dxf -
Space discretization of the flux vector G, is then approximated, for ex-

ample, by using centered differences. To second order [0(A*)2], these are
represented as
3
dx 2A*
where AJC = LIN.
A frequent problem inherent in numerical solutions of hyperbolic con-
servation equations is the occurrence of nonphysical high-frequency waves
or dispersive errors. The traditional method of reducing these effects is to
introduce artificial dissipation in the equations. 37 In addition to the dissi-
pation of truncation error, a minimal amount of artificial viscosity is used
to provide sufficient local dissipation at shock fronts to satisfy Rankine-
Hugoniot conditions. The artificial viscosity used in our studies is of Rusinov
form in which a second-derivative dissipation term is included in all of the
field equations:
<48>
The artificial viscosity coefficient v is chosen to be proportional to the grid

spacing, and X" = \va\ + ca is the maximum characteristic of the multiphase
equations. The boundary conditions at both ends of the column are in-
corporated into the differenced flux, and the system of N 4- 1 ordinary
differential equations is advanced with time, t. For example, the system
of 1(N + 1) ODE's is solved by using an implicit solver (DEBDF) de-
veloped by Shampine and Watts.38 This integrator uses backward difference
formulas appropriate for the solution of stiff problems. In addition, a
Jacobian matrix is computed numerically, and in one dimension the for-
mulation has a simple block-diagonal structure. Specifically, with second-
order spatial differencing, the half-width of the required Jacobian consists
of 13 diagonal members.
Adaptive Finite-Element Method

Similar to the finite-difference method, the finite-element method is an
established procedure (see, for example, Ref. 39) for discretizing systems
of partial differential equations. In one dimension, the finite-element method
is implemented roughly as that described for MOL, except that evaluations
are carried out as Gauss quadrature points rather than grid points.
Traditionally, a finite-element or finite-difference mesh, once con-
structed, is used unaltered throughout a simulation. However, a fixed mesh
is either inadequate to resolve a detonation wave or so fine as to be in-
efficient because most nodes are wasted in regions with little structure in
the solution. An attractive criterion for generating and adapting a mesh is
minimizing the discretization error, which is equivalent to distributing the
total error uniformly among the finite elements. 40 Often nodal movement
algorithms are aimed at equidistribution of error. Alternatively, nodal
refinement methods are designed to keep errors within specified bounds
and, at present, are more robust than their nodal movement counterparts.
In our studies, we have used this adaptive meshing to enhance the accuracy
of the numerical solutions primarily for the purpose of minimizing the
effects of artificial dissipation near extreme gradients. As a static rezone
approach, mesh refinement occurs after each time step. First, relative error
estimates are computed for each state variable on each element by using
error estimates based on interpolation error of the finite-element solution.
A cumulative error estimate for each element is obtained by using a weighted
sum of the component errors on that element. Elements are candidates
for refinement (node insertion) if their error exceeds a preset tolerance,
cr0. Elements are candidates for combination with a neighboring element
(node removal) if their error is less than another tolerance (e.g., O.lcr () ).
A smoothness restriction is imposed: no element can be more than twice
as large or less than half as small as its nearest neighbors. Additional mesh
refinement is imposed or a combination of elements is canceled where
needed to meet this criterion. Maximum and minimum nodal spacings are
imposed, between which mesh spacings are halved or doubled as needed.
With the adaptive finite elements, multiple interacting fronts can evolve
and be followed accurately. In addition, the effects of artificial dissipation
are minimized by the nodal refinement. Further details on this approach
for the solution of DDT problems are given in Ref. 41.
Operator-Splitting Method
In many of the standard numerical techniques, the smallest time scale
in the problem is used to advance the solution in an effort to resolve all
of the physics, a costly practice when the characteristic time scales of
individual terms vary by several orders of magnitude. As an alternative to
these approaches, we have investigated a time-splitting, finite-difference
numerical technique for the solution of the reactive two-phase flow equa-
tions.42 The general method of time splitting is well established in the
computational literature (see, for example, Ref. 43) and is often referred
to as the method of fractional steps or operator splitting. Previous appli-
cations of time-splitting methods to chemically reacting flows (see, for
example, Ref. 44) have illustrated the robustness of the technique. More-
over, the method appears to be extended easily to multidimensions while
maintaining reasonable execution times and computer storage.
For the multiphase combustion problems of interest here, two operators
can be defined. One operator advances the solution in time while consid-
ering only slowly varying transport terms and ignoring the phase interaction
effects. The second operator evaluates the rapidly changing phase inter-
actions by using many smaller time steps without the presence of the trans-
port terms. To be more specific, a spatially dependent operator, ££r, is
separated from the algebraic phase interaction operator, t£P, in accord with
the vector equation
£P(S) (49)
where the S is the seven-dimensional solution vector S = (pA., pg, psvs, pKvg,
(P
dx ^
—— ( 2 ^£
s g
dx dx
4-
(50a)
= c*
= m*
(sob)
The numerical solution is advanced from a time Mo a new time t

using a sequential application of each operator, i.e.,
(51)
which implies that each operator acts twice during the entire time step.
First, a step is taken to transport the field variables to a half-time level
followed by a corrective step of phase interaction. The phase interaction
operator is continued over the entire time interval, followed by a corrective
transport step covering the remaining half-time interval. At the end of this
procedure, the solution for the field variable includes the effects of both
transport and phase interaction. There are several reasons for using the
order of operations indicated in Eq. (51). First, a symmetric time splitting
can often improve accuracy and efficiency of the algorithm since vastly
different time-stepping methods may be used for the transport and for the
phase interaction. Second, a significant computational saving is realized
by combining the two 2?P operators into what can be considered as one
step. This is particularly true if ODE solvers for stiff equations are used.
Such solvers normally must be restarted every time an ^£P operation is
performed, and a substantial portion of the computer time in the solver is
associated with the startup algorithms. Thus, the number of calls to the
ODE solver is halved by using the symmetric operator. In addition, it has
been noted in Ref . 44 that proceeding from t to t + 2Ar without interrupting
the integrator permits a more efficient use of the variable step-size feature
of the solver. The only requirement is that the overall time step be restricted
by the Courant condition needed for stability. This prevents the implicit
ODE solver, resolving the stiff phase interactions, from selecting time steps
larger than those imposed by the Courant condition.
In treating the nonstiff problem of transport, an explicit flux-corrected
transport (FCT) finite-difference method is used, as outlined in Refs. 45
and 46. The technique was devised to resolve accurately hyperbolic systems
of conservation equations, including shock flows, by maintaining the prop-
erties of positivity and monotonicity. This is accomplished by using a high-
order, two-step, time-centered scheme that is numerically stabilized by
nonlinear numerical diffusion. An antidiffusion, of a similar nature, is
formed and proportionally added to the transport equations in such a
fashion to assure that no new local minima or maxima appear in the final
solution. The critical step is known as flux limiting, which determines how
much antidiffusion must be added to assure positivity. In the current study,
the strong-limit formulas resulting in sixth-order dispersive errors are used.46 47
For the solution of the stiff ODE's describing phase interaction, the
DEBDF solver is retained. The required Jacobian matrix is compact and
strictly diagonally banded. In addition, because the mixture constraints
given in Eq. (9) are asymmetric and the conservation equations for the
mixture are purely transportive in nature, only the interaction equations
for a single constituent need be considered. Thus, the Jacobian forms a
diagonal matrix of half-bandwidth of only four. In this case, the linear
algebra is reduced greatly and compact storage allows a large number of
spatial grid points.
IV. DDT in Granular Explosive HMX

In this section, we provide several applications of the reactive multiphase
flow theory that describe various modes of flame spread in a granular
column of HMX. Although the intent here is to demonstrate current pre-
dictive capabilities, model verification is also demonstrated with compar-
isons to experimental observation.
First, we show a calculation of multiphase combustion that is thermally
initiated by a hot, high-pressure gas point source whereby compaction and
mechanical load transfer to the solid reactant during deflagration lead to
shock formation and accelerated combustion produces a detonation wave.
In these one-dimensional calculations, a 10-cm-long column of HMX packed
to a 73% initial density with uniform particles of a surface-mean diameter
of 100 jjim is considered. For numerical solution, MOL with a grid consisting
of 201 computational nodes distributed evenly throughout the column is
used. (Several numerical experiments have been performed with two- and
fourfold increases in the number of computational nodes, and only slight
changes in the combustion wave characteristics were observed. This assured
that adequate grid resolution was attained and that artificial viscosity effects
were small.) Figure 1 displays a numerical-experimental comparison of the
trajectory of the burn front with time. The experimental data, obtained
by Price and Bernecker,48 are determined as the combustion front triggers
ionization pins embedded in the granular bed (the experimental data are
denoted by the open circles). As seen in this comparison, burn duration,
deflagration wave speed, detonation wave speed, and the distance to det-
onation were reproduced well.
To display the development of compaction during combustion, the vari-
ations of solid volume fraction are shown in Fig. 2. During the early stages
of combustion, significant compaction is seen ahead of the burn front
(combustion is initiated at the location of the crest of the compaction wave).
Compaction grows in extent and amplitude to 85% of total void closure.
With increased flow resistance, rapid pressurization accelerates combustion
that consumes compacted material. During this pressurization, disparate
pressure waves evolve and quickly the stress of the solid phase overtakes
the gas-phase pressure wave. Typical pressure wave profiles are shown in
Fig. 3. At a higher initial density, compaction effects are more pronounced.
HMX
(6.9 mm/fis) /
20.0
CJ
Price and Ocrncckcr data /
(1978) P
10.0
V Calculation
20 40
Time (fj,s)
Fig. 1 Comparison of calculated and experimental
distance-time burn-front trajectory in 73% dense
HMX.
Figure 4 illustrates a typical calculation of the pressure wave development

during DDT for the case of 95% dense HMX. In this case, rapid combustion
forms a completely compacted plug (total void closure) and a pressure
disturbance reflects back into the reacted two-phase combustion region.
Very rapid pressurization eventually produces a detonation wave structure.
Next, we apply the multiphase model to the Sandusky and Liddiard 13
impact experiments undertaken as a study of the compaction behavior of
granular materials. First, we consider impacts that are insufficient to cause
combustion. Since the load transfer to the unreacted granular material is
due solely to the action of the piston (as opposed to loading by the gas
pressurization of combustion), the focus of a numerical-experimental com-
parison centers on simulation of the mechanical response of the granular
material. In these one-dimensional calculations, the granular column was
assumed to be composed of uniform-size Class D HMX particles (870 jam
and <$ = 0.73). Figure 5 shows the time and space evolution of mixture
pressure and solid volume fraction after impact with a piston velocity of
76 m/s. An overlay of the wave profiles along the time coordinate clearly
shows that a compaction wave with shocklike character forms and moves
into the granular bed with a wave speed exceeding that of the piston. After
an early time transient, steady wave motion evolves. Figure 6 compares
predicted compaction wave speed with those experimentally observed for
3
CTQ
SO
VOLUME FRACTION
PRESSURE (KILOBARS) 0.0 0.2 0.4 0.6
5 10 15 20
O
rt
S"
CfQ
&:
2 M
I
CL Cn b
CD 1
CD
C/3 n
e M
3 ^,,
Q.
SOLID PHASE PRESSURE
Traing edge of
compaction wave
Compaction wave
Fig. 4 Evolution of pressure wave development during combustion of 95% HMX.
Class D HMX. The experimental data at various piston velocities are

denoted by filled circles. Clearly, the mixture model predictions correspond
well to the experimental observations. Details and additional compaction
studies can be found in Ref. 49.
To investigate compressive reaction induced by compactions, Sandusky
and Bernecker6 conducted piston-impact experiments in which the columns
of Class D HMX were subjected to long-duration (>200 JJLS), low-level
shock loading (<0.18 GPa). In this series of experiments, there was visual
evidence that reaction commenced with the generation of hot spots during
the compaction process; however, there was a delay time between impact
of the column and the detection of reaction. To simulate these experiments,
we used the adaptive finite-element method. The effect of delayed reaction
is included in the model by activating the compressive combustion when
an induction time / has been reached. This effect is expressed mathemat-
ically by the following evolutionary equation:
dl
= •— + v. VI = (52)
where the parameter c - 2.4 s~ 1 -MPa~ 2 is based on experimental time-

to-reaction data given in Ref. 6. Thus, compressive burning begins within
MIXTURE PRESSURE
Piston trajectory
Stress Wave
SOLID PHASE VOLUME FRACTION

Piston trajectory
Compaction wave
Fig. 5 Mixture pressure and solid volume fraction waveforms showing development
and growth of a compaction wave in a granular column of HMX impacted at 76 m/s.
500.0
HMX, <f>s = 0.73
I/T 400.0 Sandusky and Liddiard Data
300.0
Q-
Cfl
0)
Calculations
200.0
co
o
cd
a 100.0
o.o
0.0 20.0 40.0 60.0
80.0 100.0
Piston Velocity [m/s]
Fig. 6 Comparison of numerical and experimental results for com-
paction wave speed at various conditions of piston impact (filled
circles represent experimental data).
the granular reactant bed for conditions / > 1. Depending on the impact
velocity, the confinement, and the run distance, various modes of com-
bustion occur. For example, with a 100-m/s impact condition, a delayed
luminous reaction front is generated behind the compaction front. This
reaction front overtakes the compaction wave to produce a steady, low-
velocity combustion wave that propagates down through the remaining
length of the column. These wave features can be seen clearly in Fig. 7.
The adaptive finite-element meshing is shown at every 15-|jis interval, and
the clustering of grid nodes indicates where the numerical computations
are resolving the wave fronts. Superimposed on the computational node
locations are the experimental trajectories of the compaction and reaction
fronts.
Figures 8 and 9 show calculations of the evolution of the multiphase flow
when compressive combustion is induced. As seen in Fig. 8, a compaction
wave first appears at the origin, and after a delay of — 100 JJLS energy release
in the gas phase produces a secondary wave (supported by combustion)
that steepens and accelerates toward the compaction front. Initially, the
inert compaction wave travels into the bed with a wave speed of 430 m/s.
When the secondary combustion-driven pressure wave reaches the initial
compaction wave, the wave speed increases to —600 m/s. Figure 9 shows
the evolution of the solid volume fraction, and it is seen that very little
solid-phase decomposition is required to support the combustion-assisted
compaction. During the duration of the computations, grain surface tern-
COMPUTATIONAL NODE LOCATIONS

Sandusky arid Bernecker (1985)
14.0 experimental trajectories -
o o o o o c o ~
n
12.0 Combined wave 600 m/s
(Combustion-assisted compaction,)
„ ~ n O O O O ^ ~ '
10.0
8.0
Piston-induced compaction
u 430 m/s
O 6.0
4.0
Yeak compressive reaction

750 m/s
0.0
0.0 / 59.0 118.0 177.0 236.0 295.0
Piston/bed interface TIME (MICROSEC)
125 777 /8
Fig. 7 Calculation of impacted granular HMX whereby combustion is delayed by

— 100 |xs. The overlay of experimental trajectories and computational node locations
demonstrates adequate capture of various wave features.
peratures are well below the melt temperature of HMX and, hence, grain
burning has not been initiated; thus, rapid acceleration to detonation was
not observed. At higher impact velocities, grain burning is initiated and
the combustion readily accelerates to detonation. 11 - 50
V. Two-Dimensional Computations of Flame Propagation

The principle features of flame acceleration and the growth to detonation
in multiphase mixtures can clearly be predicted within the context of the
one-dimensional theory presented here. However, it is well known that
combustion waves can exhibit significant multidimensional effects in com-
plex geometries. To investigate this type of phenomenon in multiphase
flows, we consider the generalization of the theory outlined in Sec. II and
the FCT/time-splitting numerical algorithm described in Sec. III. To the
authors' knowledge, multidimensional, multiphase flows have not been
Ignition locus
Piston
Fig. 8 Solid-phase pressure wave profiles for impacted HMX with delayed com-
bustion. (Note that the onset of combustion produces ramp waves that grow and
accelerate to the compaction front.)
studied in the past. In this section, we will present some results for the
case when the effects of shear strength are negligible.
In two dimensions, each phase is allowed to move independently and is
assigned two components of velocity: va, the longitudinal velocity, and ua,
the transverse velocity. Conservation of mass, momentum, and energy for
each phase a (a = s,g) are as follows:
Mass:
(53)
Momentum:
d_
dt i< - (54)
_d_ d_
—
dt (PA
dx dy dy "*
Combustion-assisted compaction
Compressive combustion
Piston
Fig. 9 Overlay of solid volume fraction profiles that clearly show evolution of a
compaction wave and the weak trailing effects of compaction-induced combustion.
Energy:
.% 771 .-> .-,
-^ + - [(Ea - [(£„
s (56)
* dx dy
The phase interactions are straightforward generalizations of Eqs. (25

27) and (31):
c; = (57)
cj (58)
(59)
- ^J - Escj + m (60)
where 3)x and 2)-v signify the longitudinal and transverse components of the
drag force. In addition to these equations, we use the evolutionary equation
for solid volume fraction:
^
dt
The model is completed by utilizing the material parameters of the one-
dimensional analysis.
In solving the partial differential equations [Eqs. (53-61)] for two-
dimensional flow in a column, we incorporate the natural extension of the
initial and boundary conditions used in the one-dimensional analysis. The
time-splitting method based on FCT and ODE solvers has been extended
to two dimensions and fully vectorized for use on a Cray computer. In the
following demonstrative computation, a fixed mesh of uniformly spaced
cells is used and an appropriate time step A/ is chosen based on the Courant
condition:
+c):
^.nax/MjtO!.
Ax
('" 2
t±y 2
where c is the local sound speed and b a constant whose value is less than
0.5.
As a sample two-dimensional calculation, consider the evolution of a
two-phase reactive planar wave as it spreads into a T-shaped region of
granular CP. (CP is a highly reactive inorganic granular explosive that will
transition to detonation in a run length of a few millimeters and is consid-
erably more sensitive than HMX. Material characterization and model
inputs are given in Ref. 51.) Plate 12 (see the color section) shows contours
of solid volume fraction (left) and normalized pressures (right) for three
different times as the wave expands around the corners of the granular
bed. In the solid volume fraction figures, yellow shaded regions correspond
to undisturbed material, red shading indicates compacted material, and
the white areas correspond to consumed reactants. In the adjacent figures,
pressure contours are shown indicating high to low pressures by red to blue
shading. As seen in these calculations, release waves form at the corner
edges, and the resulting rarefaction impedes the flame spread. Although
the wave moves into the corner region as a detonation, wave expansion at
the corners causes the reactive wave to revert back to a deflagration and
the acceleration processes then rebuild the reactive wave into a detonation.
VI. Unresolved Technical Issues

Current reactive multiphase models are capable of predicting DDT in
granular energetic materials. Indeed, the interactive use of numerical models
and experimental data illustrated here both calibrates the models quali-
tatively and elucidates the operative mechanisms in the experiments.
Nevertheless, these models are far from complete due to the complex
nature of the thermal-mechanical-chemical processes involved, and there
are many unresolved physical and computational issues that warrant con-
tinued study.
The existing models are well-founded from thermodynamic and me-
chanical points of view, and the concept of hot-spot formation in granular
energetic materials is well recognized. However, little progress has been
made in determining the precise microscopic physical mechanisms that
control the nucleation, growth, and coalescence of hot spots. Moreover,
the description of the chemistry is phenomenological, at best, and certainly
does not reflect a fundamental knowledge of the reaction mechanisms.
This is readily apparent from observations that, in general, the parameters
describing the thermal decomposition of the explosive have no bearing on
those governing the thermal decomposition of hot spots at high pressures.
In current studies, a set of parameters has been established for both the
hot-spot decomposition and the grain burning, which suffice to replicate
the available experimental data within the context of a multiphase mixture
theory. Unfortunately, this set of parameters may not be unique and the
reaction mechanisms assumed to be underlying may not even be appro-
priate. For example, it has been recognized that the grains fracture during
shock loading and, thus, it is possible to envision chemical bonds being
broken as the grains are cleaved, resulting in free radicals or ions at the
grain surfaces. Such phenomena would have a profound influence on the
model chosen to represent the surface reactions. Chemical bonds may also
be broken in zones of high shear within a grain. Indeed, it has been sug-
gested that it is these scission reactions that cause the rapid rollover to
detonation observed in the late stages of shock-wave growth. The current
phenomenological models have difficulty reproducing this result. Clearly,
there is a need for more detailed investigations of the fundamental chem-
istry of explosives at high dynamic pressure. Efforts to identify intermediate
species that evolve during the loading conditions of interest could prove
to be extremely insightful. Recent advancements in optical diagnostics,
including spectroscopy and quantum chemistry, may shed new light on
which species are present at various stages of wave growth and on the
major reaction paths. 52
Beyond the chemical issues, there are other combustion aspects that are
not well understood. Models that apply in the study of ignition or quenching
at the pore level have yet to be addressed adequately. In addition, mul-
tidimensional modeling must include material strength effects (e.g., in the
study of shear-induced combustion behavior) and treat the effects of con-
finement.
In other reactive multiphase flows, such as pyrotechnic or thermite com-
bustion, high pressures can be generated at low Mach numbers (without
shock formation). Models that remove acoustic wave phenomena are rel-
evant for these descriptions, and much of the relevant combustion physics
for these flows remains to be explored.53 In these slow multiphase flows,
the reactive field can be of a three-phase nature due to the presence of a
melt component. Models that treat three-phase systems are currently under
study, and the thermodynamics suggests complex admissible interactions
with many degrees of freedom (e.g., momentum exchange for a three-
phase system is represented by a drag relationship with three terms de-
scribing the interaction of gas-solid, gas-liquid, liquid-solid flows). Hence,
model closure again becomes a serious issue, and much experimental guid-
ance is necessary in defining appropriate constitutive laws.
Beyond upgrading model descriptions, many computational issues re-
quire additional study. Having provided mathematical foundations to the
reactive multiphase flow equations, additional numerical studies are needed
in the development of multiphase Lagrangian methods. Many of the mod-
ern developments in numerical computations as applied to single-phase
compressible flow have yet to be expanded to multiphase flows. For ex-
ample, characteristic information could be incorporated into numerical
solutions, as is done in high-order Godunov-like methods.16
Modeling full three-dimensional reactive multiphase flows is an ultimate
goal of this effort but the limits of computational memory and speed are
easily reached due to the complexity of the models. Better algorithms that
avoid these constraints are imperative. For example, adaptive meshing in
multidimensions may greatly enhance calculations that would otherwise be
limited to coarse gridding. Perhaps, in multiphase modeling where greatly
disparate reactive waves evolve in each phase, it is appropriate to consider
separate adaptive meshes for each phase. To enhance speedup, parallel
processing algorithms have shown great progress over their serial coun-
terparts. Gustafson et al.54 demonstrated that speedups by factors of 1000
are possible by multitasking numerical calculations using a multiple-instruc-
tion multiple-data 1024 processor hypercube. The FCT algorithm, similar
to that used in the two-dimensional DDT simulations, was shown to be an
ideal candidate for parallel processing because it uses an explicit method.
Time splitting is well suited for parallel processing because a compact,
diagonally banded Jacobian can be used in solution of the stiff phase inter-
actions. Single processors can be assigned tasks of evaluating equation-of-
state information and resolving phase interactions without interprocessor
communication. However, the effects of multiple materials, moving bound-
aries, and complex geometries on computational parallelism have yet to
be addressed.
VII. Summary
In this study, we have provided an overview of modeling reactive mul-
tiphase compressible flows and have demonstrated that such models can
describe the complex combustion processes associated with convective
burning, compaction-induced combustion, and the growth to detonation
in granular explosives. The formulation of this thermodynamically con-
sistent model treats each phase as compressible and in complete thermo-
dynamic nonequilibrium. Therefore, it is well suited for describing the
various thermal and mechanical processes leading to detonation. We have
also provided a brief review of the mathematical character of these mul-
tiphase flow descriptions and described several numerical methods in one
and two dimensions. As demonstrative applications of these methods, we

showed one-dimensional calculations for the flame spread in granular HMX
and two-dimensional calculations illustrating the flame spread in granular
CP. We then concluded with a discussion of unresolved issues and suggested
areas of futher study. Although our studies have focused on the application
of reactive multiphase flow to granular energetic materials, the framework
of the model may be applicable to other multiphase flows where none-
quilibrium effects are suspected to be of major importance.
Acknowledgments
This work was performed at Sandia National Laboratories (SNL) and
supported by the U.S. Department of Energy under Contract DE-AC04-
76DP00789. During the course of our investigations, we have benefited
from the many fruitful discussions with D. B. Hayes, J. E. Kennedy, P.
Stanton, S. Passman (all at SNL), H. Krier (University of Illinois), P. B.
Butler (University of Iowa), and D. Kooker (Ballistics Research Labo-
ratory). We also thank R. E. Benner (SNL) for his assistance with the
adaptive gridding used in this work and R. J. Gross (SNL) for his help in
the numerical analysis.
References
ifielyaev, A. F., Bobolev, V. K., Korotkov, A. L, Sulimov, A. A., and Chuiko,
S. V., Transition from Deflagration to Detonation in Condensed Phases, IzdateFstvo
"Nauko," Moscow, through Isreal Program for Scientific Translations, 1975.
2
Bernecker, R. R., "The Deflagration-to-Detonation Transition for High Energy
Propellants—A Review," AIAA Journal, Vol. 24, Jan. 1986, pp. 82-91.
3
Andreev, K. K., "The Problem of the Mechanism of the Transition from Burning
to Detonation in Explosives, "Journal of Physical Chemistry, Vol. 17, 1944, p. 533.
4
Griffiths, N., and Groocock, J. M., "Burning to Detonation of Solid Explo-
sives," Journal of the Chemical Society of London, Vol. 814, 1960, p. 4154.
5
Bernecker, R. R., Sandusky, H. W., and Clairmont, A. R., "Deflagration-to-
Detonation Transition Studies of Porous Explosive Charges in Plastic Tubes,"
Seventh Symposium (International) on Detonation, Naval Surface Weapons Center,
Annapolis, MD, NSWC MP 82-334, 1981, p. 119.
6
Sandusky, H. W., and Bernecker, R. R., "Compressive Reaction in Porous
Beds of Energetic Materials," Eighth Symposium (International) on Detonation,
Naval Surface Weapons Center, Annapolis, MD, NSWC MP 86-194, July 1985,
pp. 881-891.
7
Macek, A., "Transition from Deflagration to Detonation in Cast Explosives,"
8
Tarver, C. M., Goodale, T. C., Shaw, K., and Cowperthwaite, M., "Defla-
gration to Detonation Transition Studies for Two Cast Primary Explosives," Sixth
Symposium (International) on Detonation, Office of Naval Research, Annapolis,
MD, ACR-221, 1976, p. 321.
9
Forest, C. A., and Mader, C. L., "Numerical Modeling of the Deflagration to
Detonation Transition," Seventeenth Symposium (International) on Combustion,
10
Nunziato, J. W., "Initiation and Growth-to-Detonation in Reactive Mixtures,"
Shock Waves in Condensed Matter—1983, edited by J. R. Asay, R. A. Graham,
and G. K. Straub, Elsevier, New York, 1984, p. 581.
H
Nunziato, J. W., and Baer, M. R., "A Microscopic Approach to the Initiation
and Detonation of Condensed-Phase Energetic Mixtures, 1 ' Journal de Physique,
Colloque C4, No. 9, Tome 48, Sept. 1987, pp. 67-83.
12
Green, L. G., James, E., Lee, E. L., Chambers, E. S., Tarver, C. M., West-
moreland, C., Weston, A., and Brown, B., "Delayed Detonation in Propellants
from Low Velocity Impact," Seventh Symposium (International) on Detonation,
Naval Surface Weapons Center, Annapolis, MD, NSWC MP 82-334, June 1981,
pp. 256-264.
13
Sandusky, H. W., and Liddiard, T. P., '"Dynamic Compaction of Porous Beds,"
Naval Surface Weapons Center, Annapolis, MD, NSWC TR-83-246, Dec. 1985.
14
McAfee, J. M., and Campbell, A. W., "An Experimental Study of the Defla-
gration to Detonation Transition in Heavily Confined HMX," 1986 JANNAF Pro-
pulsion Systems Hazards Meeting, CPIA Pub. 446, Vol. 1, March 1986, pp. 163-
186.
15
Dick, J. J., "Measurement of the Shock Initiation Sensitivity of Low Density
HMX," Combustion and Flame, Vol. 53, 1983, pp. 121-129.
16
New York, 1987, pp. 462-517.
17
Kuo, K. K., and Summerfield, M., "High Speed Combustion of Mobile Gran-
ular Solid Propellants: Wave Structure and the Equivalent Rankine-Hugoniot Re-
lation," Fifteenth Symposium (International) on Combustion, The Combustion
18
Van Tassel, W. F., and Krier, H., "Combustion and Flame Spreading Phe-
nomena in Gas-Permeable Explosive Materials," International Journal of Heat and
Mass Transfer, Vol. 18, 1975, p. 1377.
19
Beckstead, M.W., Peterson, N. L., Pilcher, D. T., Hopkins, B. D., and Krier,
H., "Convective Combustion Modeling Applied to Deflagration-to-Detonation
Transition of HMX," Combustion and Flame, Vol. 30, 1977, p. 231.
20
Krier, H., and Gokhale, S. S., "Modeling of Convective Mode Combustion
Through Granulated Propellant to Predict Detonation Transition," AIAA Journal,
Vol. 16, No. 2, 1978, p. 177.
21
Kooker, D. E., and Anderson, R. D., "A Mechanism for the Burning Rate
of High Density, Porous Energetic Materials," Seventh Symposium (International)
on Detonation, Naval Surface Weapons Center, Annapolis, MD, NSWC MP 82-
334, 1981, p. 198.
22
Butler, P. B., Lembeck, M. F., and Krier, H., "Modeling of Shock Devel-
opment and Transition to Detonation Initiated by Burning in Porous Propellant
Beds," Combustion and Flame, Vol. 46, 1982, p. 75.
23
Kim, K., "Numerical Simulation of Convective Combustion of Ball Propellants
in Strong Confinement," AIAA Journal, Vol. 22, June 1984, pp. 793-796.
24
Price, C., and Boggs, T. L., "Modeling the Deflagration to Detonation Tran-
sition in Porous Beds of Propellant," Eighth Symposium (International) on Deto-
nation, Naval Surface Weapons Center, Annapolis, MD, NSWC MP 86-194, July
1985, pp. 934-942.
25
Weston, A., and Lee, E., "Modeling 1-D Deflagration to Detonation Tran-
sition (DDT) in Porous Explosives," Eighth Symposium (International) on Deto-
nation, Naval Surface Weapons Center, Annapolis, MD, NSWC MP 86-194, July
1985, pp. 914-925.
26
Baer, M. R., and Nunziato, J. W., "A Two-Phase Mixture Theory for Defla-
gration-to-Detonation Transition in Reactive Granular Materials," International
Journal of Multiphase Flow, Vol. 12, 1986, pp. 861-889.
27
Nunziato, J. W., and E. K. Walsh, "On Ideal Multiphase Mixtures with Chem-
ical Reactions and Diffusion," Archive for Rational Mechanics and Analysis, Vol.
73, 1980, p. 285.
28
Passman, S. L., Nunziato, J. W., and Walsh, E. K., "A Theory of Multiphase
Mixtures," Rational Thermodynamics, edited by C. Truesdell, Springer-Verlag,
New York, 1984.
29
Carroll, M. M., and Holt, A., "Static and Dynamic Pore-Collapse for Ductile
Porous Materials," Journal of Applied Physics, Vol. 43, 1972, p. 1627.
30
Sheffield, S. A., Mitchell, D. E., and Hayes, D. B., "An Equation of State
and Chemical Kinetics for Hexanitrostilbene (HNS) Explosive," Sixth Symposium
(International) on Detonation, Office of Naval Research, Annapolis, MD, ACR-
211 1977, p. 748.
31
Lee, E., Hornig, H. C., and Kury, J. W., "Adiabatic Expansion of High
Explosive Detonation Products," Lawrence Livermore National Laboratory, Liv-
ermore, CA, LNLL UCRL-50422, 1968.
32
Lee, E., and Tarver, C. M., "Phenomenological Model of Shock Initiation
with Density Discontinuities," Physics of Fluids, Vol. 23, 1980, p. 2362.
33
Boggs, T. L., "The Thermal Behavior of Cyclotrimethylenetrinitramine (RDX)
and Cyclotetramethylenetetranitramine (HMX)," Fundamentals of Solid-Propel-
lant Combustion, Vol. 90, edited by K. Kou and M. Summerfield, AIAA, New
York, Progress in Astronautics and Aeronautics: 1984, pp. 121-175.
34
Baer, M. R., "A Model for Interface Temperature Induced by Convective
Heat Transfer in Porous Materials," Sandia National Labs., Albuquerque, NM,
Rept. SAND88-1073, 1989.
35
Embid, P. F., and Baer, M. R., "Mathematical Analysis of a Two-Phase Model
for Reactive Granular Material," Sandia National Labs., Albuquerque, NM, Rept.
SAND88-3302, 1989.
36
Hyman, J. M., "A Method of Lines Approach to the Numerical Solution of
Conservation Laws," Los Alamos Labs., Los Alamos, NM, Rept. LA-VR 79-837,
1979.
37
von Neumann, J., and Richtmyer, R. D., "A Method for the Numerical Calcu-
lations of Hydrodynamic Shocks," Journal of Applied Physics, Vol. 21, 1950, p. 232.
38
Shampine, L. F., and Watts, H. A., "DEPAC—Design of a User Oriented
Package of ODE Solvers," Sandia National Labs., Albuquerque, NM, Rept.
SAND79-2374, 1980.
39
Strang, G., and Fix, G. J., An Analysis of the Finite Element Method, Prentice-
Hall, Englewood Cliffs, NJ, 1973.
40
Pereyra, V., and Sewell, E. G., "Mesh Selection for Discrete Solution of
Boundary Value Problems in Ordinary Differential Equations," Numerical Math-
ematics, Vol. 23, 1975, p. 261.
41
Baer, M. R., Benner, R. E., Gross, R. J., and Nunziato, J. W., "Modeling
and Computation of Deflagration-to-Detonation Transition (DDT) in Reactive
Granular Materials," Lectures in Applied Mathematics, American Mathematical
Society, Providence, RI, Vol. 24, 1986, pp. 479-498.
42
Gross, R. J., and Baer, M. R., "A Study of Numerical Solution Methods for
Two-Phase Flows," Sandia National Labs, Albuquerque, NM, Rept. SAND84-
1633, 1986.
43
Yanenko, N. N., The Method of Fractional Steps, translated by M. Holt,
Springer-Verlag, New York, 1971.
44
Kee, R. J., and Miller, J. A., "A Split-Operator, Finite-Difference Solution
for Axisymmetric Laminar-Jet Diffusion Flames," AIAA Journal, Vol. 16, No. 2,
1978, p. 169.
45
Boris, J. P., and Book, D. L., "Solution of Continuity Equations by the Method
of Flux-Corrected Transport," Methods in Computational Physics—Vol. 16, edited
by J. Kelleen, Academic, New York, 1976.
46
Book, D. L., and Fry, M. A., "Airblast Simulations Using Flux-Corrected
Transport Codes," Naval Research Lab., Washington, DC, NRL Memo. Rept.
5334, 1984.
47
Baer, M. R., and Gross, R. J., "A Two-Dimensional Flux-Corrected Transport
Solver for Convectively Dominated Flows," Sandia National Labs., Albuquerque,
NM, Rept. SAND85-0613, 1986.
48
Price, D., and Bernecker, R. R., "Effect of Wax on the Deflagration to Det-
onation Transition of Porous Explosives," Behavior of Dense Media Under High
Dynamic Pressure, Commissariat a 1'Energie Atomique, Paris, France, 1978, p. 149.
49
Baer, M. R., "Numerical Studies of Dynamic Compaction of Inert and En-
ergetic Granular Materials," Journal of Applied Mechanics, Vol. 55, March 1988,
pp. 36-43.
50
Baer, M. R., and Nunziato, J. W., "A Multiphase Flow Model for the Initiation
and Combustion of Granular Explosives," AIAAIASMEISIAMIAPS 1st National
Fluid Dynamics Congress, AIAA, Washington, DC, 1988, pp. 861-889.
51
Baer, M. R., Gross, R. J., Nunziato, J. W., and Igel, E. A., "An Experimental
and Theoretical Study of Deflagration-to-Detonation Transition (DDT) in the
Granular Explosive, CP," Combustion and Flame, Vol. 65, 1986, p. 15.
52
Trott, W. M., and Renlund, A., "Time-Resolved Spectroscopic Studies of
Detonating Heterogeneous Explosives," Eighth Symposium (International) on Det-
onation, Naval Surface Weapons Center, Annapolis, MD, NSWC MP 86-194, July
1985, pp. 691-700.
53
Baer, M. R., and Shepherd, J. E., "A Thin Flame Model for Reactive Flow
in Porous Materials," Sandia National Labs, Albuquerque, NM, Rept. SAND83-
2576, Feb. 1984.
54
Gustafson, J. L., Montry, G. R., and Benner, R. E., "Development of Parallel
Methods for a 1024-Processor Hypercube," SI AM Journal on Scientific and Sta-
tistical Computing, Vol. 9, No. 4, July 1988, pp. 609-638.
Chapter 16
Toward a Microscopic Theory of Detonations in

Energetic Crystals
Michel Peyrard and Simone Odiot
I. Introduction
O UR present understanding of propagating detonations is based exten-

sively on the Zeldovitch-von Neumann-Doering theory 1 formulated
in 1940-1943.2 The flux of material mass, momentum, and energy is de-
scribed by hydrodynamic equations of motion, and the properties of the
reacted and unreacted materials are modeled by equilibrium equations of
state. Although such macroscopic approaches can be rather successful for
diluted gases, they fail to provide a proper description of shock-induced
detonations in condensed phases, and particularly in solids.3 Among the
unexplained features are the correlations found experimentally between
detonation properties of different energetic materials and their crystal
structures,4 large anisotropies in detonation speed, and the directional
sensitivity of monocrystals to shocks.5 Moreover, recent experiments in
fluids6 suggest that the width of a shock front in a condensed phase is of
the order of a few atomic distances, so that the assumption of equilibrium
in the wake of the shock is questionable. These experimental results can
be explained only by a microscopic theory of the propagation of a deto-
nation wave, but such a theory does not exist presently. Some attempts
have been made to explain the initiation of a shock-induced detonation at
the molecular level,7 but the microscopic structure of a steady detonation
wave and the mechanisms underlying its propagation are not yet under-
stood.
Before a microscopic theory can be proposed, additional experimental
results are necessary to improve our understanding of detonations at the
molecular level. However, these results are extremely difficult to obtain
because they require very high spatial resolution and very small time scales
to resolve the very fast propagation of a detonation. Only indirect obser-
vations can be made, such as the spectroscopic analysis of the decompo-
reserved.
513
514 TOWARD A MICROSCOPIC THEORY OF DETONATIONS
sition products8 or investigations of ultrasonic attenuation in solids, to

obtain some information on the energy exchange between acoustic waves
and atomic vibrations that may lead to dissociation.9 These very small
space and time scales that hinder experiments correspond exactly to scales
that can be investigated by molecular dynamics calculations. This is the
reason why numerical experiments are such useful tools to investigate con-
densed-phase detonations at a molecular level. They can be used to observe
the effect of a shock on an energetic crystal, record the atomic displace-
ments, determine the structure of the shock front, etc. The resolution of
such numerical "experiments" is very high, since they can, for instance,
resolve the motion of a single atom. However, they cannot avoid an "un-
certainty principle": this high accuracy is not obtained on a real explosive
but on a model; the results reflect the model, and their validity may depend
on the physical choices underlying that model. In addition, there are severe
restrictions on the space and time scales that can be investigated.
In addition to their use for interpreting experiments on energetic crystals,
investigations of the microscopic structure of detonation waves are useful
for extending our basic understanding of the solid state. In a detonation
wave, a crystal cell can be compressed to one-half of its equilibrium size.
As a result, detonations probe regions of the atom-atom interaction po-
tential curves that can hardly be investigated by any other means.
In this chapter, we describe the first investigations of energetic materials
after discussing briefly the molecular dynamics techniques themselves and
presenting their application to shock waves in solids. We then focus on
two particular topics in which molecular dynamics has brought new insights
to the propagation of a detonation wave in a crystal, the role of the crystal
structure (Sec. Ill), and the effects of the different steps in the chemistry
(Sec. IV). Section V presents a new approach that combines a model for
the chemistry with standard molecular dynamics techniques, an approach
that extends the domain of investigation of the numerical simulations and
provides a step toward a microscopic theory of the propagation of a det-
onation wave. Section VI discusses the results and the future of these
approaches.
II. First Molecular Dynamics Investigations of Energetic Materials

A. Molecular Dynamics Investigations of Shock Waves in Crystals
The basic idea of the molecular dynamics approach is to use Newton's
law of motion to determine the behavior in time of interacting atoms and
molecules, given the forces between them and initial and boundary con-
ditions. Thus, a standard molecular dynamics program solves the 3N clas-
sical equations of motion of the N atoms in the computational cell. For
this purpose, the integration time is discretized into steps Af that must be
much shorter than the characteristic times of the atomic motions in a solid.
A typical value of Ar is 10"15 s or less. Present high-speed computers can
treat 1000-20,000 particles for 104-106 time steps, thus defining the space
and time scales that can be investigated currently by molecular dynamics.
The space scale is at most on the order of a few hundred lattice spacings,
whereas the time scale ranges from picoseconds to nanoseconds. These
M. PEYRARD AND S. ODIOT 515
scales are suitable for the study of the steady-state propagation of a det-
onation wave (for instance, a wave moving at 7000 m/s will propagate along
700 A in the 10 ps investigated in a typical calculation). However, molecular
dynamics cannot represent accurately the early phase of a shock-induced
detonation, the initiation that occurs on the time scale of a few microsec-
onds. Despite these limitations, molecular dynamics is a powerful tool for
the theoretical understanding of the properties of matter, because it can
reach regimes that an experiment cannot attain and it is free from pertur-
bations due to crystal defects or the experimental setting.
An interesting illustration of the validity of the molecular dynamics ap-
proach is provided by the study of the second sound in a solid excited by
a strong heat pulse by Tsai and MacDonald.10 These authors were not only
able to observe the second sound in a molecular dynamics simulation, but,
due to the high resolution of the method, they could also show that the
experimentally observed second-sound wave is, in fact, made of several
components generated because the longitudinal and transverse stress waves
are not in thermal equilibrium. This provided a clue for an explanation of
the "anomalous" temperature dependence of the second-sound velocity
observed experimentally and also pointed out how the assumption of ther-
mal equilibrium included in many theories can be incorrect for a highly
dynamical system.
For shock-induced detonations, numerical simulations must describe the
chemical processes together with the lattice dynamics. For this reason, the
first molecular dynamics models were restricted to the investigations of
the propagation of simple shock waves. Some early investigations consid-
ered a one-dimensional lattice made of only one type of atom. Because
the displacements in a strong shock wave can be very large, the nonlinearity
of the interatomic potential was taken into account by introducing cubic
and quadratic terms. The simulations and the analytical analysis11 show
that a shock front is not steady in such a lattice. The initial excitation
generated by an impact on one side of the lattice decomposes into several
solitary waves with different amplitudes and speeds. As a result, the dis-
tance between the fastest and the slowest components keeps growing, and
the profile of the disturbance changes as it propagates. This result em-
phasized the important role that could be played by nonlinear excitations
in detonations. Similar findings were obtained with a three-dimensional
model by Tsai and Beckett12 and Batteh and Powell.13 The conclusion of
these simulations was that the Rankine-Hugoniot jump conditions could
not be used to analyze the data from planar impact experiments because
they depend on the existence of a steady wave in addition to the conser-
vation of mass, momentum, and energy. In contrast to this result, Paskin
and Dienes14 reported a series of molecular dynamics calculations on three-
dimensional models where only steady waves were observed.
Some clues for the resolution of these disparate findings were provided
by the simulations performed by Holian and Straub, 15 showing that the
structure of the asymptotic shock wave changes as a function of the shock
strength. A steady state is observed for strong shocks due to the partial
relaxation of compressive shear stress behind the shock front as transverse
motions arise. Theoretical analysis of the properties of solitary waves in
discrete lattices later confirmed the peculiarity of one-dimensional mon-

atomic lattices16 and showed that, for polyatomic lattices or higher dimen-
sions, the solitary waves may become unstable for high shock loading. In
addition to pointing out the crucial role of the model used for molecular
dynamics simulations, these results gave important information on the
structure of a shock wave in a crystal. Whatever the model, current sim-
ulations exhibit very narrow shock fronts on the molecular scale. Although
the strong anharmonic coupling induces rather fast thermal equilibration
in the wake of the shock for the steady shock front under heavy loading,
there is always a region behind the shock that is not in thermal equilibrium.
This region is important for the theory of detonations because the excitation
of the molecules, leading to molecular breakup, starts inside this domain.
B. First Investigations of Energetic Materials

The first molecular dynamics simulation of the propagation of a deto-
nation in a crystal was performed by Karo et al. 17 They considered a two-
dimensional monatomic lattice of atoms coupled by a predissociative exo-
thermal potential. A compressive force applied to one end generates a
compressive wave that propagates along the lattice. When the wave reaches
the other end of the solid, it moves an atom far from its neighbor and a
bond breaks. The breaking of interatomic bonds is then reflected back into
the bulk of the solid. This model is not consistent with real shock-induced
detonations since, in reality, the shock initiates reactions in the region of
the initial interaction. As discussed in the next section, this is due to the
existence of different atomic species and interactions in a real crystal and,
therefore, could not be seen with the simple model investigated. However,
this work demonstrated that the propagating shock stays sharp on an atomic
scale and that its energy is not degraded into random thermal motion. The
region behind the shock is not in thermal equilibrium until long after the
shock has passed. In recent molecular dynamics studies, Karo et al. have
considered diatomic molecules embedded in a monatomic host lattice.18
They show that shock fronts, because of their sharpness on an atomic scale,
can transmit large amounts of translational energy to the internal motions
of molecular groups. This coupling is extremely rapid (~10~ 12 or 10~13 s)
and the time scales are consistent with the propagation of a detonation
wave along a few unit cells. This suggests that, in a solid-phase detonation,
molecular dissociation can occur close to the shock front, in regions that
are not in thermal equilibrium.
A molecular dynamics study of a detonation initiated by local heating
or mechanical impact in a dense system has been presented by Tsai and
Trevino.19'20 Their model is a hypothetical body centered cubic (bcc) crystal
in which the atoms (all identical) interact through a compound Morse
potential of the form Vi — V2. When the distance between two neighboring
atoms is smaller than a threshold, they form a strong bond through the
potential V2. Once formed, this bond is "saturated," which means that
other neighbors of the atomic pair interact with it through the weaker
potential, Vl. When the distance between two atoms is larger than the
threshold, the atoms interact through V1 and the transition from the V2
interaction to the Vl interaction releases some energy. The model crystal

is then heated and submitted to a mechanical shock on one end. This
initiates dissociations that cause further reactions to occur, and the accom-
panying expansion of the bonds drives a strong shock wave into the ma-
terial.
The model and the boundary conditions are such that, for sufficient
heating, the explosive reaction would always start. This model has been
used to study the structure of the detonation wave, but investigations of
the role of the crystal structure have not been performed because the
structure is difficult to control in such a model. The simulations show that
the energy transfer between the translational, rotational, and vibrational
degrees of freedom is inefficient, so that there is no thermal equilibrium
in the shock-compressed region. These numerical results question the va-
lidity of the macroscopic theory of detonations for crystal, as do the results
of Karo et al.17
III. Role of the Crystal Structure

The simulations summarized in the preceding section show that molec-
ular dynamics can simulate at the molecular level the basic deterministic
mechanical processes involved in solid-state detonations, but that some
important features were missing. In the monatomic lattices considered, all
of the atoms were classically linked by the same predissociative potential
and the crystal structure could not be controlled. Indeed, an energetic
material must be such that some chemical bonds can be broken with a
large exothermic effect, however, the properties of these exothermic bonds
alone are insufficient to determine the characteristics of the detonation
wave. The site and orientation of the molecules in the crystal and how
these molecules are linked by nonexothermic bonds are also essential.
Consequently, a microscopic model for molecular dynamics simulations
must include these features. However, it is beyond present computing
capabilities to model a real explosive that may have several tens of atoms
in a unit cell. This would also be meaningless. Although we know atomic
positions to high accuracy, we do not know atom-atom interaction forces
to the same accuracy, particularly for the very large lattice distortions that
occur in a detonation. Thus, we cannot expect at the current state-of-the-
art a molecular dynamics calculation to give quantitative results for a real
explosive, but it can provide a fundamental understanding of the phenom-
ena. For this purpose, we can use simple generic models that do not attempt
to mimic exactly a specific crystal but contain the essential features of a
real energetic crystal. In this section, we show that even a one-dimensional
model can give useful information, and then we discuss two-dimensional
models that allow a more complete investigation of the role of crystal
structure and bonding.
A. One-Dimensional Model
Despite its simplicity, the one-dimensional model that we present here
predicts many of the experimental properties of detonations and allows us
to determine certain microscopic conditions that a solid must have to sustain
a detonation wave.21 The motivation for this model results from obser-
vations of correlations between electronic properties of the molecules,22
crystal structures, and detonation characteristics of energetic materials. In
a series of molecules containing the same type of C-NO 2 bond, it has been
shown that an excited electronic state was involved in the dissociation
process. In the particular case of nitromethane, this state is a predissociative
exothermic state.23 Moreover, Bertrand and Mutin 4 had investigated the
role of the crystal structure on the detonation properties of this same series
of energetic materials. Because the different compounds in the series in-
cluded the same type of bond, the chemical properties of the individual
molecules were expected to be very similar. Instead, some compounds are
explosives and others are not. Their analysis showed that there is a strong
correlation between the explosive character and the presence in the struc-
ture of chains of closed packed NO 2 groups.
This result can be understood by simulations using the lattice model
shown schematically in Fig. 1. The model consists of a one-dimensional
array of two-component N-C molecules. The two parts of the molecule
are henceforth denoted as atoms C and N, although, in fact, they can be
more complicated groups. For example, N and C might represent NO2 and
CH3, respectively, to model nitromethane. Only the displacements of at-
oms C and N along the direction of the chain are considered. They are
denoted by yn(f) and zn(t), where subscript n is the index of the unit cell.
Three types of interatomic coupling are introduced.
1) The N-C molecules are coupled between nearest neighbors with
Morse-type potentials, which connect atoms of the same type:
V(yn +l ~ yn) = D(e I) 2 (1)
where D and K are equal to Dc, Kc or DN, KN for the C and N atoms,
respectively.
2) The two components of an N-C molecule are connected with a pre-
dissociative exothermic intramolecular potential W(yn - zn), which has
the shape as plotted in Fig. 2. For computational convenience, this potential
is obtained by difference between two Morse potentials:
W(yn - zn) = n - zn) - V2(y,, ~ z,,)
(2)
e-W"-** - I)2 - - I) 2
with D2 > and K2 <
Fig. 1 Schematic of the one-dimensional model consisting of a chain of "diatomic"

molecules. The atomic displacements are restricted to the direction of the chain.
p
con
—i— —I— —I— —i— —I

1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.7 2.9 3.1 3.3 3.5
N—C distance (Angsroms)
Fig. 2 Predissociative intramolecular N-C potential. Dotted line: potential energy
of the predissociative excited state of the nitromethane molecule calculated theoret-
ically. Solid line: intramolecular potential used in the model.
If the relative displacement yn — zn of the two components of a molecule

is smaller than the distance d for which W is maximum, the molecule is
stable. If yn - zn becomes larger than d, the two atoms repel each other
and the molecule tends to dissociate with a dissociation energy D2 - Dl.
A detonation wave is then the transmission of the molecular dissociations
along the lattice assisted by the energy release that accompanies each
dissociation.
Each time a molecule is dissociated, the potential energy D2 ~ Dl is
available for conversion into kinetic energy. Thus, in the absence of any
added energy-absorption mechanism in the model, the kinetic energy of
the system, and hence its temperature, would grow continuously. This
nonphysical result is an artifact of the one-dimensional character of the
model. In a real three-dimensional crystal, some energy would be spread
laterally into transverse modes and would no longer be available to sustain
the detonation wave. Moreover, in a real crystal, the motion of dislocations
dissipates energy and, because the ground state of an N-C molecule may
not be the required electronic state for dissociation, some energy may also
be absorbed to bring the molecule into its predissociative state. All of these
energy sinks that exist in a real detonation wave are introduced in the
model by phenomenological damping terms, proportional to the atom ve-
locities with a coefficient y. These terms are added to the equations of
motion derived from the Hamiltonian obtained by summing the atom ki-
netic energies and their potential interaction energies.
A typical set of parameters for the simulations was obtained by using

solid nitromethane as a sample crystal. Interaction potentials between CH3
and NO2 groups were obtained by adding together atom-atom potentials
of the Buckingham type, and the resulting potential functions were then
fitted by the Morse potentials defined by Eq. (1) to determine Dc, DN,
Kc, and KN. The parameters for the predissociative intramolecular poten-
tial were chosen so that Eq. (2) gives the best fit for the excited NO2-CH3
potential derived from ab initio studies of the potential energy surfaces of
the isolated nitromethane molecule.23 Table 1 lists this typical parameter
set.
The detonation wave is generated in the model by applying a constant
force F to one end of the lattice during a time t{. Then this end of the
lattice is left free to expand. The intensity of the excitation is determined
mainly by the impulse P = F x t±. To initiate the detonation in a time
compatible with molecular dynamics simulations, it is necessary to apply
an excitation that is much stronger than actually used in shock-induced
detonation experiments. Although this method of excitation allows a qual-
itative analysis of the initiation process, it cannot give quantitative results
in agreement with experiments for the early phase of the detonation.
Simulations using this simple model have provided the conditions at
which a detonation wave can exist as well as information about the dynamics
of this wave. These results are summarized here.
1) Conditions that allow a detonation wave. There are conditions that
the excitation and the model crystal itself must fulfill to produce a stable
detonation. First, consider the effects of varying the amplitude of the
excitation. In agreement with the experiments, the simulations exhibit two
different excitation thresholds, a dissociation threshold P1 and a detonation
threshold P2 > PI (Fig. 3). When P < Pl, no chemical dissociation occurs
and an ordinary shock wave propagates into the lattice with a speed slightly
larger than the speed of sound. When Pl < P < P2, the excitation breaks
some N-C molecules, but the dissociation does not propagate along the
Table 1 Typical set of parameters for

one-dimensional model
Atom C: Masse mc = 16 amu

Atom N: Masse mN = 48 amu
C-C intermolecular potential:
Kc = 1.4 A- 1
Dc = 0.0158 eV
N-N intermolecular potential:
KN = 2.1 A- 1
DN = 0.0638 eV
C-N intramolecular potential:
K, = 2.335 A- 1
K2 = 1.950 A- 1
Dl = 14.559 eV
D2 = 15.159 eV
a) 20.
150
26.
150
b)22,
450
28.
0.\--\-
n 450
Fig. 3 Aspect of a detonation wave in the one-dimensional model plotted at different

propagation times. The upper part of each figure (a or b) shows the absolute dis-
placements yn of atoms N and the lower part shows the stretching yn - zn of the N-
C bond as a function of n. The dotted horizontal line on the lower parts indicates
the value of the N-C bond stretching over which the interaction becomes repulsive
(broken bond), a) Initial impulse between dissociation and detonation thresholds.
Breaking of the N-C bonds stops propagating after some time and only a shock
wave subsists, b) Initial impulse larger than the detonation threshold. The detonation
wave continues propagating along the lattice and reaches a stationary state.
lattice and finally only a shock wave exists. However, the number of dis-
sociated molecules increases drastically as P approaches P2 and diverges
for P = P2. When P > P2, there is a steady detonation wave consisting
of a leading shock wave followed by the breaking of the N-C molecules.
After the dissociation, the repulsion between the two components of a
molecule assists the propagation of the leading shock, but this process is
efficient in sustaining a steady detonation only if the model crystal has
some specific characteristics. The two sublattices that comprise the lattice
must be very different. One of them must be soft or made of light atoms
to allow the large displacements that are necessary for the bonds to break.
On the contrary, the second sublattice must be very rigid or massive to
act as a support for the repulsive force. For instance, if DN - Dc instead
of the values listed in Table 1, the model no longer can sustain a steady
detonation, even for very large excitations. The detonation threshold is
infinitely large because no differential acceleration on the N and C sites is
felt. This is in agreement with the analysis of Bertrand and Mutin. 4 The
explosive compounds are the compounds with chains of closely packed
NO2 groups, because these chains provide the rigid sublattice necessary to
sustain the detonation.
2) Dynamics of detonation waves. The one-dimensional model presented
in this section does not exhibit the peculiarities of the monatomic one-
dimensional models discussed in Sec. II. Conversely, it shows features
found in three-dimensional models. At low shock loading, for P < P1?
some solitary waves are found and the shock front is not steady21; at the
high shock-loading level required to initiate the detonation, the shock front
shape becomes steady, as shown in Fig. 3b. This is because, despite its
one-dimensional aspect, the model contains some of the features essential
to a qualitatively correct structure for the shock and detonation waves—
namely, it contains two different atomic species16 and a mechanism for
energy dissipation.
One advantage of the one-dimensional model is that it can be used on
a small computer. A typical run requires a few CPU hours on a PDP-11
when only the part of the lattice that has been reached by the detonation
is included in the calculation to save computer time. Thus, the investigated
domain grows as the detonation propagates. The steady state is achieved
after a few hundred cells.
In agreement with the experiments, the simulations show that the det-
onation speed Vf is determined by the characteristics of the model crystal
and not of the initial excitation, provided that it exceeds the threshold P2.
The detonation reaches a steady state when the energy released by the
molecular dissociations is balanced exactly by the energy dissipation due
to damping. For a given intramolecular potential, there is a range of 7
values that can lead to stable detonation waves. When y is too large, the
detonation slows down and finally dies out. Conversely, if 7 is too small,
the dissipation cannot balance the energy released by the molecular dis-
sociations. Interestingly, for realistic model parameters, ty varies between
two and five times the speed of sound, a result that agrees roughly with
the range of experimentally observed detonation speeds. Among the pa-
rameters of the intramolecular potential, the dissociation energy D2 — Dl
is very important in determining the final speed, whereas the height of the
barrier that has to be passed before dissociation has little effect on vf but
affects the detonation threshold. Thus, the sensitivity to shocks and the
detonation speed are related to different features of the intramolecular
potential, and this can perhaps explain why the attempts to correlate these
two parameters in experimental studies are not conclusive. As opposed to
y and the intramolecular potential, the masses mc and mN of the atoms
or the parameters of the intermolecular Morse potentials have little effect
on the detonation speed.
B. Two-Dimensional Model
The major limitation of the model presented in the last section is that
it is restricted to one dimension. Thus, it cannot describe the energy trans-
ferred from the longitudinal shock wave into transverse motions, but in-
stead uses a phenomenological damping term. Moreover, in this model,
the crystal structure enters only through the relative strength of the inter-
molecular potentials, but not through the geometrical positions of the
molecules. The two-dimensional model presented in this section has been
introduced to overcome these limitations.24 The investigations of a two-
dimensional model instead of a three-dimensional model take transverse
motions into account without incurring the cost of three-dimensional com-
putations. This allows the simulation of larger systems for longer periods
of time and introduces complexity in a progressive way.
Following the ideas used in the one-dimensional model, the new model,
shown in Fig. 4, consists of an array of two-component N-C molecules
N- -N- -N- -N-
J+1
IM- -N -N
N- -N- -N
J-1
N- -N- -N- -N-

1-1 1+ 1 X
Fig. 4 Schematic of the two-dimensional model defining the lattice parameters.

linked by Morse intermolecular potentials, which connect nearest and next-

nearest neighboring molecules. The equilibrium lengths of the bonds be-
tween nearest neighbors are chosen such that each bond is in equilibrium
when the atoms are in the desired crystal structure; these short-distance
bonds impose the lattice structure. The intramolecular bond is defined by
the same predissociative potential as in the previous model [Eq. (2)]. The
atoms have two degrees of freedom along the X and Y axes of the lattice.
The X direction is the direction along which the detonation propagates.
A major difference from the one-dimensional case is that the two-dimen-
sional model does not include "ad hoc" damping terms to model the energy
transfer into transverse modes, which is now an intrinsic part of the model.
The lattice consists of 64 or 128 cells in the X direction and between
four and eight cells in the Y direction along which periodic boundary
conditions are used. As in the one-dimensional model, the shock is initiated
by an impulse on the left boundary, which is left free to expand after the
end of the excitation impulse. To model the propagation of the detonation
for a long distance, the previously-described lattice is treated as part of an
infinite lattice that would be seen through a window that moves with the
detonation so that the front can be followed as it propagates along several
hundred lattice cells. A typical calculation requires 5-20 min of Cray X-
MP computer time to achieve a steady state. Contrary to fluid-phase sim-
ulations in which a significant part of the computation time is used in finding
the neighbors of a given atom, the simulation of a solid phase is much
faster since the neighbors are known a priori. The regular organization of
the atoms in a crystal can even be used for an efficient vectorization of
the simulation program. The drawback of using a method for detonations
that does not dynamically check the neighbors is that the fluid phase behind
the detonation is not described accurately. However, the errors are min-
imized by the large number of neighbors taken into account and because
the most distorted part of the fluid phase is dropped from the computation
as the calculation window moves with the detonation front.
The two-dimensional simulations show that many of the microscopic
features demonstrated in the one-dimensional calculations are still valid.
These features are the following:
1) The wave front is sharp on an atomic scale. Only about 10-40 cells
separate the solid phase in front of the shock from the gaslike phase in its
wake. This distance may be even shorter when the intermolecular bonds
are particularly weak. As a result of this sharpness, the atomic motions in
the shock front are violent and thermal equilibrium is not reached until
long after the shock passes.
2) To sustain a detonation, a solid must have both a strongly bonded
sublattice, which is responsible for the coherence of the propagation, and
a weakly bonded sublattice, which allows large-amplitude motions leading
to chemical dissociations.
3) The excitation thresholds for the dissociations of the first molecules
or the generation of a stable detonation are different.
In addition to providing confirmation and better understanding of the
one-dimensional results, the two-dimensional model shows the important
role of transverse motions in determining the detonation speed and even
the possibility for a solid to sustain a detonation. This is illustrated in Fig.

5, showing the position of the shock and detonation as a function of time
in a lattice initially at zero temperature. Although the initial temperature
is indeed not zero in a real experiment, this case is interesting because it
shows more clearly the role of the transverse modes. Figure 5 shows that
there are two propagation regimes with different speeds. A high-speed
regime immediately follows the excitation. An analysis of the motions of
the atoms in this regime shows that the detonation wave propagates as a
perfect plane wave. All atoms lying in the same plane orthogonal to X
have exactly the same motions, which are almost purely longitudinal. This
high-speed regime is very similar to the detonation waves found in the
previous one-dimensional studies.21 Since the two-dimensional model does
not include a dissipation term, and because the energy transferred to trans-
verse motions is very small, almost all of the energy of the initial impulse
stays in the detonation wave front. The result is that in the high-speed
regime, the detonation speed depends on the amplitude of the excitation.
This is contrary to both experiment and the results of macroscopic fluid
models of detonations.1 The conclusion is that this particular high-speed
regime does not exist in a real system. It is observed in the particular case
of zero initial temperature because the lattice is initially at rest and the
excitation is purely longitudinal, but it is not stable.
o
o -i
C\2
O
CO -
fi \
QJ
T-
—l———i———i———i———i
0 50 100 150 200 250 300 350
Time
Fig. 5 Position of the shock front (solid line) and of the last broken bond (crosses)
as a function of time when a detonation propagates in the two-dimensional lattice
at zero initial temperature.
a) N atoms a) C atoms
b) N atoms
c) N—C bonds stretching
d)
Fig. 6 Atomic displacements when a detonation wave propagates in a low-speed

regime in a two-dimensional lattice, a) Longitudinal and b) transverse displacements
of the N and C atoms are shown in the vertical Z direction on a lattice grid, c)
Stretching of the N-C bonds. In each graph, the position of the shock front along
the X axis is indicated by a vertical arrow, d) Region in the lattice that includes the
shock front, showing the atomic positions. To represent enough cells, the scale has
been shrunk along the X axis. The actual cell parameters are a = 5 A and b = 3
A. The heavy lines show the N-C bonds that are not broken.
The knee in Fig. 5 corresponds to the breakdown of the high-speed

regime, which occurs when rather larger transverse motions appear due to
the geometrical distribution of the forces in the lattice. The new state is
much less regular than the previous one because the detonation no longer
propagates as a perfect plane wave. However, it is important to notice that
a well-defined structure persists along the propagation. This is shown clearly
in Fig. 6, where the shock front is followed by positive N longitudinal
displacements (i.e., a compression of the N sublattice) and negative lon-
gitudinal C displacements (i.e., a rarefaction wave in the C sublattice).
These two motions stretch the N-C bond as shown in Fig. 6c and cause
the molecular dissociations that sustain the detonation wave, as in the one-
dimensional model. However, a significant part of the energy released in
the dissociations is transferred into the rather large transverse motions that
can be seen in Fig. 6b. They do not contribute to motions in the direction
of the propagation. As a result, the detonation speed in the second regime
is much smaller and simulations show that, in the low-speed regime, the
detonation speed is independent of the intensity of the excitation and is a
characteristic of the material.
For a given type of molecule, which determines the amount of energy
released in dissociation, the detonation speed results from the way this
energy is partitioned between transverse and longitudinal motions. As a
consequence, the detonation speed is very sensitive to the crystal structure
that governs the geometrical distribution of the interatomic forces. The
parameter that plays the larger role in determining the speed is the ratio
alb of the cell dimensions along X and Y. Figure 7 shows the variation of
the detonation speed vs alb. The speed increases as alb increases, because
a larger value of a means less energy transferred into transverse motions
(in the limit a —» o°? the motions tend to be one-dimensional because all
forces lie along X). More importantly, Fig. 7 shows that alb = 1.2 is the
smallest value for which a stable detonation has been found. When alb <
1.2, the high-speed regime of the propagation of the detonation can be
obtained. However, when the transverse motions appear, they drain too
much energy from the longitudinal motions and the detonation dies, no
matter how strong the initial excitation is. Thus, the model predicts that
the crystal structure can determine if a given chemical compound is an
explosive or not.
To the authors' knowledge, there are no experimental investigations of
energetic compounds that can exist in different phases in the solid state.
They could be very interesting for the understanding of the microscopic
mechanisms of the propagation of detonations. There is, however, some
experimental evidence of the role of transverse modes in the propagation
of detonations. Very large anisotropies in shock-initiation sensitivity of
pentaerythritol tetranitrate have been reported by Dick.5 In his paper,
Dick proposes a semimicroscopic interpretation of the experiments in-
volving the plasticity of the crystal. It is interesting to notice that he finds
that the direction in which the detonation is not initiated (or propagates
at very low speed) is the direction in which the initial shock induces the
largest transverse motions.
Another parameter in the structure is the angle a between the N-C
molecules and the X axis. It has little influence on the detonation speed
en
I
o
LLJ
ffi
C/5
1 1.5 2
(a/b)
Fig. 7 Variation of the detonation speed as a function of the ratio alb of the cell
parameters. No stable detonation has been found when alb < 1.2 (shaded area).
but significantly changes the detonation threshold, as shown in Fig. 8.

Although variation of a is not important when a > 90 deg, reducing a
below 90 deg drastically increases the detonation threshold and no stable
detonation has been observed for a < 30 deg.
Although we have presented results on detonations propagating in lat-
tices initially completely at rest, the main conclusions hold when the sim-
ulations are carried on lattices at T =£ 0. The only major change then is
that the high-speed regime is no longer observed, because it is destroyed
by the thermal fluctuations. In contrast, the low-speed regime with a speed
characteristic of the material is still observed. The detonation wave reaches
this regime after a gradual slowdown, during which the memory of the
amplitude of the initial impulse is lost.
Experiments show that, for most of the detonations in fluids, a purely
planar detonation front perpendicular to the direction in which the deto-
nation propagates is not stable.1-25 Transverse waves appear and give rise
to the well-known "cellular structure" of the detonation wave. Some det-
O
</3
LU
QC
4
z
o
O 3
LU
Q
30 60 90 120 150 180
Fig. 8 Variation of the detonation threshold as a function of the angle a between

the N-C molecules and the X axis. The insets show schematically the atomic dis-
placements when the detonation propagates in the low-speed regime. The detonation
threshold is indicated by the ratio of the speed of the N atoms at the end of the
excitation time to the speed of sound in the lattice.
onation cells have also been observed in solids.26 Their size is much smaller
than in gases and their presence has not been detected in all cases. The
preceding simulations could not detect any tendency for the wave to form
such structures due to the limited size of the system in the transverse
directions and the periodic boundary conditions that impose some restric-
tion on the wave pattern. Some simulations have been carried out on much
larger systems (20-40 cells in the Y direction) to try to remove these
restrictions.27 The results show that, at the microscopic scale as well, a
planar detonation wave orthogonal to the X direction may not be stable.
Its final stable state depends strongly on the crystal structure and on the
distribution of the interatomic forces introduced in the model. 28 Figure 9
shows a typical result for a lattice with 20 cells along the Y direction and
a distribution of interatomic forces chosen to be slightly asymmetric with
b)
Fig. 9 Structure of a detonation wave in a lattice with 20 cells in the Y direction,

a) Schematic map of the lattice showing the different regions that comprise the
detonation. The white cells have not yet been reached by the shock and correspond
to the unreacted part of the material. The darker shaded cells have been reached
by the shock, but the N-C bond is not yet broken (induction zone). The lighter
shaded cells are the cells in which the N-C bond is broken (reaction zone). Note
the oblique patterns, b) Longitudinal displacements of the N atoms plotted in the
vertical direction on a lattice grid.
respect to the X axis, which could occur due to distortions of the electronic
shells in a crystal structure. This geometrical structure promotes transverse
motions that result in a detonation front making a 45-deg angle with the
X axis. This oblique detonation wave is the stable state of the detonation.
It is reached after a transient during which the detonation stays planar,
orthogonal to X. The model shows that the final state depends on the
crystal structure and force distribution, but it is independent of the initiation
process. This is illustrated in Fig. 10, which shows a detonation initiated
in two separate points. The two detonations combine and eventually gen-
erate the same oblique detonation front as for a homogeneous initiation.
It is interesting to notice that, for an inhomogeneous excitation, the det-
onation threshold is lower than for a homogeneous excitation by about
30%. This supports some suggestions made in the macroscopic theory of
detonations,25 stating that the shear stress generated when transverse waves
merge in the cellular structure promotes molecular dissociations.
Molecular dynamics simulations cannot observe any real cellular struc-
ture because, even in solids, the cell size is several orders of magnitude
larger than the typical length scales in a microscopic simulation. Thus, it

is not clear whether the transverse waves generating the oblique detonation
fronts observed in some simulations would evolve into a cellular pattern
in a larger system. Their presence is very sensitive to the symmetry of the
crystal structure or force laws, and they are not observed in structures with
high symmetry (see Sec. IV). It is interesting to compare these results with
some experiments made on different forms of solid TNT.26 Cast TNT,
which has some internal orientation, shows a more pronounced cellular
structure than the pressed powders in which the microcrystals have random
orientation, resulting in a higher average symmetry. The numerical sim-
ulations show that the formation of transverse waves, which are a more
coherent form of the transverse motions responsible for the appearance of
the low-speed regime, is important because they drain even more energy
from the longitudinal motions and thus are more likely to cause a deto-
nation to die.
IV. Effects of the Chemistry

Defining a proper model to represent the chemistry in the microscopic
investigations is one of the major problems in the molecular dynamics
simulations of detonations. Standard molecular dynamics is well adapted
to study motions of well-defined particles, but chemistry, in addition to its
microscopic quantum aspects, causes changes in the particle type and num-
ber due to dissociations or recombinations. Molecular dynamics techniques
have to be extended to represent these features by adding certain k t ad hoc"
assumptions about the atom-atom interactions. These extensions include
rules that can select one interaction potential or another according to a
particle internal state or bonding with its neighbors,19 or choosing inter-
action potentials that provide an approximate description of chemical re-
actions.
Using the predissociative potential (Fig. 2) to represent the N-C inter-
actions in the simulations described earlier amounts to condensing a se-
quence of events into a single process, the molecular dissociation. However,
the chemistry of a real detonation involves many chemical steps. The first
step is generally an endothermic molecular dissociation, whereas exothermic
recombinations of some fragments occur later. One important factor in
the physics of the detonation wave is the location where the energy is
released: if it is released far from the shock front, its transport to the front
to assist the propagation of the shock becomes a crucial factor. To inves-
tigate this effect, the models described in Sec. Ill have been generalized
to introduce a two-step chemistry27 that separates the molecular dissocia-
tion from a second step when the energy is released. This model, however,
does not actually describe recombination processes. Just their exothermic
effect is introduced because it is expected to play a more important role
in the physics of the detonation than the changes in particle type or number.
The exothermic reactions are characterized by the power q that they can
transmit to an N or C fragment. This power depends on the energy released
by an individual reaction and on the number of reactions per time unit,
i.e., the reaction rate. Since the exothermic reactions heat the reaction
b)
•Q
) D>
-8.5
Fig. 10 Different steps in the evolution of a detonation wave excited by an inhomo-

geneous impulse in a lattice with 40 cells in the Y direction. The representation is
made as in Fig. 9. a) Shortly after the initiation, the two separated initiated regions
are clearly visible, b) At a later time, the two detonation waves combine to form a
single planar detonation wave. This wave is unstable and evolves into an oblique
wave c) that corresponds to the steady state.
C)
Fig. 10 (continued)
products, q is used in the simulations to increase the kinetic energy of the

fragments after some delay following the dissociation, corresponding to a
chemical induction time. During a time step Ar, the kinetic energy of a
fragment is increased by
d[^m(v2x + = q x (3)
where vx and vy are the components of the velocity of the fragment. As-
suming that equipartition of energy holds in the reaction zone, which is
rather far from the shock front, the variation of one of the components of
the fragment speed (for instance, vx) is obtained by changing vx to
q&t/m (4)
Consequently, the integration algorithm is modified to make these velocity

changes after the standard step in which the positions and velocities are
updated. For a given fragment, this procedure starts after the dissociation
when an induction time has passed and it is repeated for the time tf during
which this fragment is in the calculation window, so that the mean energy
gained by the two components of an N-C molecule is
G = 2qtf (5)
Since this approach models the exothermic effect of the recombinations,
the molecular dissociation no longer needs to be exothermic and various
cases—exothermic, thermoneutral, and endothermic—have been treated.
This model has been used to investigate the role of the power released
by the exothermic recombinations and the role of the potential barrier for
molecular dissociation.27
1) Role of q. As one would expect, the simulations show that a larger
energy release increases the detonation speed. A quantitative analysis brings,
however, a more subtle insight into the mechanisms by which the deto-
nation propagates. It shows that a simple energy balance by no means
determines the detonation speed. The speed is not determined merely by
the amount of energy release, but also by the way it is released and trans-
ported in the system. A rather small increase in the energy exchanged
when the molecules dissociate can produce a velocity increase that only a
very large increase in q can generate. This is because the dissociations
occur close enough to the shock front to affect the leading shock strongly
and in a coherent way, whereas a large part of the energy dissipated in the
reaction zone results simply in random thermal fluctuations. Unless strong
confinement prevents energy losses to the sides and back, the energy of
the recombination is not efficient in assisting the shock propagation. The
calculations show that even if Q is raised to values as high as 300 kcal/
mole, the detonation speed does not exceed roughly eight times the sound
speed in the crystal. This may explain why the detonation speeds measured
for various solids vary by much smaller relative amounts than the chemical
energies released by these compounds.
2) Role of the potential energy barrier for molecular dissociation. A sys-
tematic investigation of various thermoneutral N-C potentials with dif-
ferent potential-energy barriers shows that a change in the barrier height
does not change significantly the detonation speed27 in agreement with the
results of the one-dimensional model discussed previously,21 although it
has some effect on the speed. This is again an indication that a simple
energy balance is not enough to determine the detonation speed.
These two results provide some elements for a theoretical analysis of
the detonation speed in energetic crystals because they point out some of
the factors that have to be included, such as the energy transport from the
reaction zone to the shock front or the mechanism of bond breaking that
is related to the dissociation barrier.
The two-step chemistry model has been used to solve a question raised
by the one-dimensional simulations. When the potential parameters of the
one-dimensional model are chosen to correspond to nitromethane, 21 the
model sustains detonation waves. This does not appear surprising since
liquid nitromethane is known to be an explosive. However, the first ex-
periments of shock-induced detonation in nitromethane crystals29 made to
test the molecular dynamics calculations failed to produce a detonation,
even for very strong excitations. Although explanations related to impur-
ities that trigger off the detonation in liquid nitromethane cannot be ex-
cluded completely, the two-dimensional molecular dynamics calculations
suggest that the discrepancy could come from large transverse motions that
arise in the nitromethane lattice.
To check this conjecture, the two-dimensional model has been extended
to provide an approximate description of the nitromethane structure.27 The
new model shown in Fig. 11 is a projection on the (a,b) plane of the three-
c}________(c\
i-1 i U1
Fig. 11 Extended two-dimensional model that corresponds to the nitromethane

structure in the (a,b) plane. The heavy lines show the shortest N-N bonds that form
zig-zag chains in the crystal.
dimensional nitromethane structure in which the internal motions of the

CH3 and NO2 groups have been neglected so that these groups can be
represented by particles C and N, as discussed earlier. The (0,6) plane has
been chosen because the strongest bonds in the crystal (shortest NO 2 -NO 2
bonds) form zig-zag chains that are almost parallel to this plane. A unit
cell in the model contains two N-C molecular units. The simulations do
show the same trend as the experiments: a steady detonation is very difficult
to generate in such a lattice. First, unless the duration of the excitation is
doubled with respect to the simulations with the simpler lattice of Fig. 4,
even for a large excitation amplitude or large q, the detonation eventually
dies out. Second, the energy released by recombination has to be raised
to more than Q = 107 kcal/mole, a value that is much bigger than the
estimated value for nitromethane (—54 kcal/mole). Consequently, the two-
dimensional simulations conclude that a shock-induced detonation cannot
persist in a nitromethane crystal, whereas the one-dimensional calculations
concluded the opposite. The reason is that the compressed zig-zag chains
fold similar to an accordion, creating large transverse motions that absorb
a large amount of energy. Because these chains do not exist in a liquid,
this phenomenon does not occur in the fluid phase and liquid nitromethane
can explode. Whereas this result provides a possible explanation for the
puzzling difference in explosive properties of solid and liquid nitromethane,
it also demonstrates how much care must be taken when extrapolating the
results of molecular dynamics simulations to real systems, since the results
often depend on the details of the model.
The simulations performed with the model shown in Fig. 11 also give
additional information on the stability of a planar detonation front. In this
model, the crystal structure has a symmetry plane parallel to the X axis
(glide plane with translation a/2) and the detonation front stays planar,
orthogonal to X. This tends to confirm that the evolution toward an oblique
detonation front is stimulated by asymmetry in the model.
V. New Approach: Microscopic Modeling

Representing the chemistry of a detonation by a pure mechanical ap-
proach in which appropriate potentials are chosen to describe the disso-
ciations, as in the models discussed in the previous sections, has two main
drawbacks. First, because the intramolecular potentials describe internal
modes that have a much higher frequency than the external modes asso-
ciated with the intermolecular bonds, the simulation requires very small
time steps to resolve accurately the internal modes. This slows down the
calculations significantly. From the computational standpoint, it would be
desirable to separate the two time scales for internal and external motions.
Second, reducing the complex process of chemical dissociation to a simple
classical mechanics description may be a crude approach and, for instance,
introducing some quantum effects may become necessary.
The approach outlined in this section is a step toward the resolution of
these drawbacks. We call it microscopic modeling, because it embeds a
model for the chemistry of the dissociation into a molecular dynamics
calculation. It has been made possible by the Adaptive Constraint Algo-
rithm developed by Lambrakos et al.30 The first implementation of this
method for shock-induced detonations31 has been restricted to a simple
model for the chemical dissociation, but the approach is flexible enough
that more complete models can be introduced.
In the simplest version of the present model, the strong intramolecular
bond is replaced by a rigid bond before dissociation. Thus, the algorithm
has to maintain a constraint as long as the bond is not broken. While this
constraint is maintained, the effects of the intermolecular forces on the N-
C bond are monitored and used to update a "state function" S(t), which
describes the internal state of the molecule. When the state function reaches
a preassigned value Sdis, the constraint forces are set to zero and an in-
termolecular N-C Lennard-Jones potential is switched on. Because the
constraints are turned off or on depending on the internal state of the
molecule, they are called adaptive constraints. The impulse A/ along the
intramolecular bond due to intermolecular interactions during a time step
Ar is used to establish a criterion for assigning particular values to S(t).
When A/is positive and its modulus exceeds a given threshold Fth, A5ex is
added to S(t). This corresponds to an excitation of the N-C molecule
under a high compression. If A/is below the threshold Fth or corresponds

to an expansion, spontaneous de-excitation of the molecule is allowed by
subtracting A5dex from S(t). Thus, A5ex is the change in S(t) per time step
for energy transfer from intermolecular to intramolecular modes, and ASdex
is the change in S(t) per time step for energy transfer from intramolecular
to intermolecular modes. The quantities ASex and A5dex have been made
integral multiples of each other. Thus, the function S assumes discrete and
equally spaced values between 0 and 5dis. The threshold parameter Fth
defines the sensitivity of the coupling between the internal and external
states. A small value of Fih allows a weak external impulse to increase the
value of the internal state function, and thus allows the molecules to dis-
sociate in a weak external field.
In this model, a change in the intramolecular state occurs only when A/
> 0, which means that compression causes the electronic transitions leading
to dissociation. This condition is physically realistic for molecules in the
induction zone, the region between the shock front and the reaction front.
However, by allowing state changes when A/ < 0, the effects of expansion
on the molecule could also be considered.
Because S(t) can increase at most by ASex during a time step Ar, the
minimum time tmin for a molecule to dissociate is
r min = Ar(Sdis/ASex) (6)
However, since S(f) may not increase at each time step if the molecule is
not compressed sufficiently, and because spontaneous de-excitation is al-
lowed, the actual time for a molecule to reach the dissociation threshold
after the shock passes is generally longer than tm[n. The combination of the
parameters Fth, A5ex, A5dex, and Sdis determines the rate of energy transfer
between the intramolecular and intermolecular states and the minimum
time delay preceding molecular dissociation. The bond-breaking time should
be at least on the order of one period of oscillation for intramolecular
vibrational modes, e.g., 1.5-3 x 10~14 s. Its value is important because
it determines the distance between the region where bonds break (i.e., the
reaction front in this model) and the shock front. Because the variation of
S(t) corresponds to excitation or de-excitation of the N-C molecule, the
model includes the corresponding energy transfer by adding or subtracting
energy from the intermolecular motions. This is done simply in the current
model by adding or subtracting kinetic energy from the N-C system.
In addition to the embedded model for chemistry, the simulation pro-
gram also includes a fast algorithm for tracking particle positions, the
Monotonic Lagrangian Grid (MLG) algorithm,32 33 so that the program is
fast enough to allow the investigation of three-dimensional lattices on a
rather long time scale. In an MLG, adjacent particles in space have grid
indices that are also adjacent in computer memory. The MLG indexing is
also used to index velocities and other attributes of particles, such as par-
ticule-type labels or flags. The MLG indexing is compact in the sense that
the memory locations required for indexing particle positions equal the
number of particles in the system, and it is designed to provide an efficient
vectorization of the program. Typical calculations contain 896-3600 lattice
cells (1792-7200 particles). The systems simulated are periodic in the X

and Y directions. The detonation always propagates in the Z direction.
Lattice cells may be subtracted and added at the left and right boundaries
in the Z direction while keeping the number of particles in the system
constant. As in the two-dimensional model, this defines a moving window
that follows the detonation front. A typical calculation for one set of pa-
rameters requires 20-60 min of Cray X-MP time.
Although this model is still under development, it has already produced
some interesting results.31 First, although it is fundamentally different from
the models discussed in the previous sections, it also shows a sharp shock
front, the important role of the crystal structure, and that no artificial
damping is required to obtain a detonation speed characteristic of the model
crystal and not of the initial impulse. This suggests that these results are
not model dependent and validates the new approach.
The new model has also shown new results. As expected, the transverse
motions play a larger role in three dimensions than in two dimensions.
Although the periodic boundary conditions and the limited size of the
system in the transverse directions tend to stabilize a planar detonation
front perpendicular to the direction of the propagation, the front becomes
oblique by forming transverse waves for much smaller systems in three
dimensions than in two dimensions. In addition, the three-dimensional
simulations do not show the same initial high-speed regime, almost without
transverse motions, that appeared in two dimensions before the transition
into the steady detonation. By separating intramolecular states from in-
termolecular motions, the model allows investigations of the role of the
minimum delay time for molecular breaking and energy release. The results
show that the qualitative structure of the detonation wave is preserved
when the minimum delay increases by more than one order of magnitude.
However, this variation affects the detonation by changing the detonation
velocity and increasing the size of the induction zone. This effect, known
also from hydrodynamic simulations of detonations in fluids, 3 is not in-
cluded in the standard theory of detonations.1
VI. Discussion and Conclusion: Toward a Microscopic Theory

Although the first molecular dynamics simulations of shock-induced det-
onations appeared in 1978, most of the results have been published in the
last five years so that application of molecular dynamics to detonations is
still rather recent. However, it has already introduced important new ideas
in the field. Some of these ideas are still the source of controversy between
the ''supporters" of the macroscopic or microscopic approaches.
One major result of the molecular dynamics calculations is the sharpness
of the front in the microscopic scale. It is now supported by recent exper-
iments in the condensed phase.6 As a result of this sharpness, the motions
of the molecules in the shock front are very violent and dissociations due
to mechanical effects rather than thermal ones can occur. It is interesting
to notice that the same conclusion has been reached by all of the molecular
dynamics calculations despite the fact that they use rather different meth-
ods. Although not a complete proof of the validity of this result because
errors due to the need to initiate the detonation in a time scale compatible
with molecular dynamics cannot be excluded completely, this universality
is a strong indication that the conclusion is correct. It does imply very fast
molecular dissociations that are compatible with recent analysis of the
lifetime of some excited internal vibrational modes in hydrocarbon chains,
which demonstrate that nonlinear resonances can cause extremely fast
energy transfer. 34 This energy transfer occurs in times as small as 50 femto
seconds, times comparable with the time necessary for mechanical breaking
in the molecular dynamics calculations. This idea of mechanical breaking
led Walker to propose an alternative35 to the standard "starvation kinetic"
model to explain why the high-temperature decomposition of many mol-
ecules have nearly the same reaction rate even though the low-temperature
rates may be very different from each other. This "physical kinetics" model
allowed Walker to propose a new approach36 for the calculation of deto-
nation velocities. Although these ideas have to face further tests, their
confirmation would be a very interesting contribution of the molecular
dynamics simulations.
There are additional areas, such as the role of transverse motions on the
detonation speed or the role of the crystal structure in determining if a
crystal can sustain a detonation, in which molecular dynamics has brought
new insights. However, the full potential of molecular dynamics will be
exploited further by improving the analysis of the results. For instance,
Tsai and Trevino19-20 have investigated the energy transfer between trans-
lational or rotational modes, an investigation that is very difficult to do
experimentally, even with advanced spectroscopic techniques. Similarly,
additional analyses should be performed to determine how the energy from
the shock front flows into internal modes and leads to dissociation. A
detailed statistical analysis of the numerical experiments could also provide
a basis for a quantitative comparison between the microscopic numerical
experiments and the macroscopic experimental and theoretical investiga-
tions.
However, the aim of the molecular dynamics simulations is not only to
make observations but to help define a microscopic theory of detonations.
One interesting aspect of the numerical approach is that it allows division
of the complicated problem of detonation into simpler problems that can
be investigated separately. These include shock-molecule energy transfer,
dynamics of the chemical reactions in the wake of the front, energy transfer
inside the induction zone, etc. Each of these simpler problems can be
investigated in great detail with simulations. For instance, calculations using
three-body interactions to achieve a better representation of chemical re-
actions are currently under way.37 In addition, some theoretical investi-
gations of the energy exchange mechanism between the shock front and
the internal modes of the molecules could take advantage of the recent
progress in the dynamics of nonlinear excitations.
However, despite the rapid increase in computer capacity, it is unlikely
that these complicated methods can be incorporated in a full-scale simu-
lation in the very near future. They can be used to design and test models
that give a correct representation of the various phenomena involved in a
detonation. Introducing these models one at a time, as done in the cal-
culations discussed in Sec. V, allows their validation in the difficult con-

ditions of a detonation. Then the pieces of the "jigsaw puzzle" will be
combined to achieve a microscopic description of detonation waves better
than the one curently available. In this microscopic modeling approach,
analytical models can also be used as intermediate models so that bringing
the pieces together will finally achieve a microscopic theory of detonation.
However, this goal is still far away, owing to the large variety of physical
and chemical phenomena involved in detonations. Moreover, all of the
investigations presented here deal with monocrystals. This is certainly a
severe restriction for practical application, although it is also a necessary
step for the basic understanding of solid-state detonations. However, in-
vestigations of the role of defects or inhomogeneities with molecular dy-
namics are in progress.
90 TO
Acknowledgments
The authors would like to acknowledge many helpful discussions with
S. Lambrakos, E. Oran, J. Boris, and D. Tsai. Part of the work presented
in this chapter has been supported by the Direction des Recherches Etudes
et Techniques (France), the Physics Division of the U.S. Office of Naval
Research, and the Naval Research Laboratory.
References
Tickett, W., and Davis, W. C., Detonation, Univ. of California Press, Berkeley,
CA, 1979.
2
Zeldowitch, Y. B., "On the Theory of the Propagation of Detonation in Gaseous
Systems," NACA TM 1261 1960; also, Von Neumann, J., "Theory of Detonation
Waves," John von Neumann Collected Works, Vol. 6, Macmillan, NY, 1942, and
Doering, W., "On Detonation Processes in Gases," Annals of Physics, Vol. 43,
1943, p. 421.
3
Approches Microscopique et Macroscopique des Detonations, edited by S. Odiot,
Journal de Physique, Vol. C4, 1987.
4
Bertrand, G., and Mutin, J. C., unpublished; also, Rudel, P., Odiot, S., Mutin,
J. C., and Peyrard, M., "Struture cristalline et caractere detonique des cristaux
moleculaires nitres," Journal de Chimie Physique, Vol. 87, 1990, p. 1307.
5
Dick, J., "Effect of Crystal Orientation on Shock Initiation Sensibility of Pen-
taerythritol Tetranitrate Explosive," Applied Physics Letters, Vol. 44,1984, p. 859;
also Koch, H. W., and Barras, J. C., "Mesure optique de la vitesse de detonation
des monocristaux d'explosifs suivant leurs axes cristallographiques. I. Hexogene
et pentrite." Institut Franco Allemand de Recherches de Saint Louis, France, Rept.
28/71, 1971.
6
Presles, H. N., Hallouin, M., and Harris, P., "Etude du comportement de 1'eau
dans le front d'une onde de choc," Approches Microscopique et Macroscopique
des Detonations, edited by S. Odiot, Journal de Physique, Vol. C4, 1987, p. 127.
7
Zerilli, F. J., and Toton, E. T., "Shock Induced Molecular Excitations in Solids,"
Physical Review B: Solid State, Vol. 29, 1984, p. 5891.
8
Redlung, A., and Trott, W. M., "Spectroscopic Studies of Initiation and Det-
onation Chemistry," Approches Microscopique et Macroscopique des Detonations,
edited by S. Odiot, Journal de Physique, Vol. C4, 1987.
9
Perrin, B., and Zarembowitch, A., "Contribution des approches ultrasonores
a la comprehension des phenomenes d'initiation de la detonation dans des cristaux
moleculaires energetiques." Approches Microscopique et Macroscopique des Det-

onations, edited by S. Odiot, Journal de Physique, Vol. C4, 1987.
10
Tsai, D. H., and MacDonald, R. A., "Molecular-Dynamical Study of Second
Sound in a Solid Excited by a Strong Heat Pulse," Physical Review B: Solid State,
Vol. 14, 1976, p. 4714.
H
Tasi, J., "Initial-Value Problem for Nonlinear Diatomic Chains," Physical
Review B: Solid State, Vol. 14, 1976, p. 2358; also, "Evolution of Shock-Waves in
a One-Dimensional Lattice," Journal of Applied Physics, Vol. 51, p. 5804; also,
"A Second Order Korteweg-de-Vries Equation for a Lattice," Journal of Applied
Physics, Vol. 51, 1980, p. 5816.
12
Tsai, D. H., and Beckett, C. W., "Shock Wave Propagation in Cubic Lattices,"
Journal of Geophysical Research, Vol. 71, 1966, p. 2601.
13
Batteh, J. H., and Powell, J. D., "Solitary-Wave Propagation in a Three Di-
mensional Lattice," Physical Review B: Solid State, Vol. 20, 1979, p. 1398.
14
Paskin, A., and Dienes, G. J., "A Model for Shock Waves in Solids and
Evidence for a Thermal Catastrophe," Journal of Applied Physics, Vol. 43, 1972,
p. 1605.
15
Holian, B. L., and Straub, G. K., "Molecular Dynamics of Shock Waves in
Three-Dimensional Solids: A Transition From Nonsteady to Steady Waves in Per-
fect Crystals and Implications for the Rankine-Hugoniot Conditions," Physical
Review Letters, Vol. 43, 1979, p. 1598.
16
Peyrard, ]yj pnevmatikos, S. and Flytzanis, N., "Discreteness Effects on Non-
Topological Kink Soliton Dynamics in Nonlinear Lattices," Physica, Vol. D19,
1986, p. 268.
17
Karo, A. M., Hardy, J. R., and Walker, F. E., "Theoretical Studies of Shock-
Initiated Detonations," Acta Astraunautica, Vol. 5, 1978, p. 1041.
18
Karo, A. M., and Hardy, J. R., "Energy Transfer Processes in Condensed
Molecular Solids. Detonation Theory in a Crystal: Limits," Approches Micros-
copique et Macroscopique des Detonations, edited by S. Odiot, Journal de Physique,
Vol. C4, 1987, p. 235.
19
Tsai, D. H., and Trevino, S. F., "Simulation of the Initiation of Detonation
in an Energetic Molecular Crystal," Journal of Chemical Physics, Vol. 81, 1984,
p. 5636.
20
Trevino, S. F., and Tsai, D. H., Proceedings of the Eighth Symposium (Inter-
national) on Detonation, edited by J. M. Short, Naval Surface Weapons Center,
White Oak, Silver Spring, MD, 1986, p. 870.
21
Peyrard, M., Odiot, S., and Lavenir, E., "Interpretation microscopique par
un mecanisme cooperatif des detonations induites par choc dans les cristaux mo-
leculaires. Application au nitromethane," Comptes Rendus Hebdomadaires des
Seances de VAcademic des Sciences, Vol. 299, Serie II, 1984, p. 917; Peyrard, M.,
Odiot, S., Lavenir, E., and Schnur, J. M., "Molecular Model for Cooperative
Propagation of Shock-Induced Detonations in Energetic Solids and its Application
to Nitromethane," Journal of Applied Physics, Vol. 57, 1985, p. 2626.
22
Delpuech, A., and Cherville, J., "Relation Between the Electronic Structure
and the Shock Sensitivity of Secondary Nitrated Explosives—Criterion of Molec-
ular Sensitivity. I. Nitroaromatics and Nitramines," Propellants and Explosives,
Vol. 3, 1978, p. 169; also, "Relation Between Shock Sensitiveness of Secondary
Explosives and their Molecular Structure. III. Influence of Crystal Environment,"
Propellants and Explosives, Vol. 4, 1979, p. 61, and Delpuech, A., PhD Disser-
tation, Univ. of Bordeaux, France, 1980.
23
Mijoule, C., Odiot, S., Fliszar, S., and Schnur, J. M., "Etude ab-inito du
caractere dissociatif des premiers etats singulets excites du nitromethane," Theo-
chem, Elsevier, The Netherlands, Vol. 149, 1987, p. 311.
24
Peyrard, M., Odiot, S., Oran, E., Boris, J., and Schnur, J., "Microscopic
Model for Propagation of Shock-Induced Detonations in Energetic Solids," Phys-
ical Review B: Solid State, Vol. 33, 1986, p. 2350.
25
Dupre, G., Knystautas, R., and Lee, J. H., "The Propagation of Turbulent
Flames and Detonation in Tubes," Dynamics of Explosions, Vol. 106, edited by
J. R. Bowen, J. C. Leyer, and R. I. Soloukin, Progress in Astronautic and Aero-
nautics, AIAA New York, 1986, p. 244.
26
Howe, P., Frey, R., and Melani, G., "Observations Concerning Transverse
Waves in Solid Explosives," Combustion Science and Technology, Vol. 14, 1976,
p. 63.
27
Peyrard, M., Odiot, S., and Blain, M., "Etude par dynamique moleculaire de
la propagation des detonations induites par chocs dans les cristaux energetiques.
III. Influence des mouvements transversaux, de la structure cristalline, et de la
chimie," Journal de Chimie Physique, Vol. 85, 1988, p. 759.
28
Maffre, P., and Peyrard, M., "Simulation of Impact-Induced Detonations in
Energetic Solids: Ideal and Heterogeneous Crystals," unpublished, 1990.
29
Mala, J., private communication, 1988.
30
Lambrakos, S. G., Boris, J. P., Oran, E. S., Chandrasekar, I., and Nagumo,
M., "A Modified SHAKE Algorithm for Maintaining Rigid Bonds in Molecular
Dynamics Simulations of Large Molecules," Journal on Computational Physics,
Vol. 85, 1989, p. 473.
31
Lambrakos, S. G., Peyrard, M., Oran, E. S., and Boris, J. P., "Molecular-
Dynamics Simulation of Shock-Induced Detonations in Solids," Physical Review
B: Solid State, Vol. 39,1989, p. 993; Lambrakos, S. G., and Peyrard, M., "Modeling
Complex Intramolecular Processes Using Constrained Molecular Dynamics," Jour-
nal of Chemical Physics, Vol. 93, 1990, p. 4329.
32
Boris, J., "A Vectorized Near Neighbors Algorithm of Order N Using a Mon-
otonic Logical Grid," Journal on Computational Physics, Vol. 66, 1986, p. 1.
33
Lambrakos, S. S., and Boris, J. P., "Geometric Properties of the Monotonic
Lagrangian Grid Algorithm for Near Neighbor Calculations," Journal on Com-
putational Physics, Vol. 73, 1987, p. 2546.
34
Hutchinson, J. S., Reinhardt, W. P., and Hynes, J. T., "Quantum Dynamics
Analysis of Energy Transfer in Model Hydrocarbons," Journal of Chemical Physics,
Vol. 79, 1983, p. 4247.
35
Walker, F. E., "Physical kinetics," Journal of Applied Physics, Vol. 63, 1988,
p. 5548.
36
Walker, F. E., "Calculation of Detonation Velocities from Hugoniot Data,"
unpublished, 1988; also, "Comparison of Detonation Velocities and Average Vi-
brational Motion of Atom Pairs in Organic Explosives," Shock Waves in Condensed
Matter 1987, edited by S. C. Schmidt and N.C. Holmes, Elsevier, 1988, p. 534.
37
Tsai, D. H., private communication, 1988.
38
Tsai, D. H., private communication, 1989.
Chapter 17
Overview of Spray Modeling
Clayton T. Crowe
T HE role of sprays in combustion systems is to break up a liquid fuel

so that there is more surface area for evaporation to provide gaseous
fuel for combustion. Some spray properties that are important for the
design of combustion systems are the evaporation rate of the fuel, mixing
of the fuel by entrainment and diffusion, and local cooling processes. To be
useful tools for understanding and predicting the behavior of sprays, nu-
merical models must be capable of providing this type of information, given
the basic properties of the incoming liquid jet and the background gas.
The physical development of a spray occurs in three stages corresponding
to three spatial regimes. First, given a liquid jet, the process of atomization
breaks up the jet and forms a dense cluster of droplets. In this second
dense droplet region, droplets interact with other droplets, leading to co-
alescence and further droplet breakup. The spray then diffuses and evap-
orates, forming a third dilute droplet region in which droplet-droplet contact
is minimal and the motion of droplets and changes in their properties are
governed by the local gas flowfield.
The physical mechanisms leading to atomization depend on the type of
atomizer used. Pressure atomizers convert the energy associated with pres-
sure to kinetic energy of thin liquid sheets. These sheets become unstable
due to frictional surface tension and fluid dynamic effects, and disintegrate
into droplets. The most common type of pressure atomizer is the centrifugal
pressure nozzle in which a swirling component of velocity is imparted to
the liquid forming a conical spray.
Two-fluid (or pneumatic) atomizers subject the liquid to a high-velocity
gas, which breaks up the liquid into droplets. In general, the breakup
occurs in two stages. First, the high-velocity gas stream fragments the liquid
into filaments and large droplets, and then the filaments and large droplets
are broken into smaller droplets. The process is very complex. The resulting
droplet size is influenced by liquid properties and the velocity and density
of the gas. The droplets emerge from the atomizer as a dilute spray.
reserved.
543
544 OVERVIEW OF SPRAY MODELING
The second, or dense, spray region is characterized by intensive inter-

actions among droplets. Droplets collide and may coalesce depending on
their relative velocity, size, surface tension, fluid properties, and type of
collision (grazing vs direct). During this period, the droplets may exchange
mass, momentum, and energy with the carrier gas, but the most important
consideration is the ultimate droplet size distribution, which continues into
the dilute spray region.
In the third, or dilute, spray region, droplet-droplet contact is minimal.
The change in droplet properties depends on interactions between droplets
and the background fluid. As the spray continues to disperse, the droplet
spacing increases to the point that a droplet can be modeled as an isolated
droplet. Of course, the droplets also affect the fluid properties. Evapo-
ration of droplets requires heat from the carrier gas, thus causing local
cooling. The momentum associated with the drag force between the gas
and the droplet causes gas to be entrained by the spray. The droplets also
modulate the fluid turbulence, which, in turn, affects the rate of mass,
momentum, and energy transfer. This coupling problem is the basis of the
numerical and physical difficulties encountered in modeling spray systems
accurately.
Numerical models for the dilute spray region can be classified as two-
fluid or trajectory models.1 The two-fluid model treats the droplet field as
a continuum with averaged properties. Thus, the gas-droplet mixture is
modeled as interpenetrating continua. Rather sophisticated models have
been developed for sprays, which require a number of empirical coeffi-
cients. The two-fluid model has a major problem due to the fact that the
dilute droplet field does not have the character of a continuum; the concept
of a droplet pressure is different in the two fluids in that there is no direct
analog in a dilute droplet field to molecular interaction in a gas.
The trajectory model is easy to implement and represents the essential
physics of the process. The model, however, is only as reliable as the data
provided and is limited in accuracy by the need to perform statistical av-
erages over the relatively small number of individual droplets that actually
can be considered. Information is needed on mass transfer, drag, and
energy transfer for droplets near other droplets. The primary disadvantage
of the trajectory model is the computational time required to achieve a
detailed description of the droplet field. This is particularly evident for
unsteady systems, where mass, velocity, temperature, and the position of
every computational droplet have to be stored at every time step.
A dominant mechanism controlling the spread rate of a spray is the
effect of fluid turbulence on droplet dispersion. In the two-fluid model,
droplet dispersion is modeled as a Fickian diffusion process, which requires
selecting a value for a dispersion coefficient for the droplets dispersing in
the turbulent fluid. However, the choice of a valid dispersion coefficient
is difficult because the available data are very inconsistent. Values varying
over two orders of magnitude are found in the literature. In the trajectory
model, a stochastic approach is used in which the turbulent velocities are
represented by random numbers and superimposed on the mean flowfield.
Considerable success has been made with this approach in relatively simple
C. T. CROWE 545
turbulent fields. 2 This stochastic approach requires more computational

time to reach stationary averages.
The presence of the droplets in the spray also modulates the turbulence,
which, in turn, affects the droplet dispersion and heat and mass transfer
rates. Analysis of available data shows that droplets can both increase and
decrease turbulent intensity, depending on the ratio of droplet size to
turbulence scale.3 Understanding and quantifying these effects is an im-
portant area of current research.
Sprays injected in large-scale turbulent fields, as generated by developing
shear layers, show interesting dispersion patterns. 4 In this situation, droplet
dispersion by a fluctuating turbulent velocity is insignificant compared with
the centrifuging effect of the large-scale vortex structures. Droplets are
centrifuged to the edge of the turbulent structures, leaving regions nearly
void of droplets at the vortex cores.
Droplet sprays for combustion systems represent a continuing research
challenge for the experimentalist and numerical analyst. Three areas lack-
ing fundamental information are
1) droplet-droplet interaction in the dense spray region and the resulting
coalescence or breakup of droplets;
2) droplet drag, mass, and heat transfer in droplet clouds; and
3) turbulent dispersion of droplets and modulation of turbulence by
droplets.
The numerical analyst is confronted with the challenge of finding more
efficient ways to incorporate coupling effects and droplet dispersion by
turbulence.
References
^rowe, C. T., "REVIEW—Numerical Models for Dilute Gas-Particle Flows,"
Journal of Fluids Engineering, Vol. 104, No. 3, 1982, pp. 297-303.
2
Faeth, G. M., "Mixing, Transport and Combustion in Sprays, 11 Progress in
Energy and Combustion Science, Vol. 13, 1987, pp. 293-345.
3
Gore, R. A., and Crowe, C. T., "Effect of Particle Size on Modulating Tur-
bulence Intensity," International Journal of Multiphase Flows, Vol. 15, No. 2, 1989,
pp. 279-286.
4
Crowe, C. T., Chung, J. N., and Troutt, T. R., "Particle Mixing in Free Shear
Layers," Progress in Energy and Combustion Sciences, Vol. 14, pp. 171-194.
Background
Most practical combustion occurs in extremely complex systems, as made
evident by the preceding chapters. The chemistry alone can be exceedingly
complex, even when many of the usual simplifying assumptions are legit-
imate, and the underlying flowfields are generally multidimensional and
usually turbulent. Furthermore, the systems are often multiphase and have
nonideal boundary conditions. This means that questionable or downright
unsatisfactory simplifications must be made throughout the models. Deal-
ing with this complexity in a meaningful way is both an intellectual challenge
and a form of art.
One of the most technically and economically important aspects of mul-
tiphase combustion is the physical behavior of fuel droplets and sprays
carried in a gas-phase reacting flow. Crowe has presented elsewhere a brief
overview introducing the subject. 1 It may be read in conjunction with the
three contributions here on spray combustion by Bellan, by Rangal and
Sirignano, and by Maclnnes and Bracco. Crowe's introductory comments
in this book help to place these contributions to the description of complex
multiphase spray phenomena into an overall perspective.
When the droplets in a spray are dilute, they behave as individual drop-
lets in a background flow; when they are densely packed, their behavior
changes due to limitations on heat and mass transfer that result from their
interactions with the background flow. In addition, drop collisions, coa-
lescence, and breakup can have a strong influence on the flow. In most
sprays, droplets start out in the dense-droplet regime and progress to a
dilute-droplet regime. Bellan focuses on the model of the interaction of a
cluster of droplets with the background fluid. This is a crucial physical
problem that can be isolated and studied and whose understanding con-
tributes to building better macroscopic spray models. In particular, Bellan
considers dense and dilute clusters of droplets and describes some of the
current models of droplet-cluster interactions with a background gas, and
proposes a better model for drop interactions. She shows how it may be
used as a subgrid model combined with a solution of the fluid-dynamic
equations.
Rangal and Sirignano consider parallel streams of droplets introduced
into a flow. This simplified parallel-stream configuration is particularly
useful as a calibrating benchmark. The authors present a series of solutions
of the problem in which the constraints and approximations are gradually
relaxed. Maclnnes and Bracco are also concerned with models of the drop-
lets used in sprays. They compare the results of using a stochastic model
of droplet interactions (in which the collision rates are estimated from
kinetic theory) to the results of computations using a deterministic model
in which droplet trajectories and collisions are computed. Maclnnes and
Bracco report a particularly interesting set of resolution convergence stud-
ies, relating them to the underlying changes in the droplet-model physics
that they uncover.
Not included in this book are the important calculations of Dwyer and
coworkers on the detailed behavior of individual droplets and their inter-
action with the surrounding flows.1 This is an extremely difficult problem
for which this group has worked extensively on numerical algorithms, as
well as on methods for inputting particular physical data. To date, they
have computed the behavior of evaporating and burning single droplets
and small groups of droplets and, to some extent, the interaction of these
droplets with the background flow. There has also been some parallel work
focusing on how droplets deform in the flow and the effects of surface
tension on their shape. In particular, Fyfe et al. have pointed out the
difficulties of treating the full nonlinear droplet-fluid problem with surface
tension and viscosity included. 2
The remaining chapters in this section deal with other complex phenom-
ena in which many of the important physical processes must be approxi-
mated by phenomenological submodels. Di Blasi describes the interesting
and complex problem of a combustible solid surface exposed to external
heating from thermal radiation. Depending on the rate of various diffusion
and evaporation processes, gases driven off the solid can ignite and a
combustion wave can propagate through the system along the surface of
the solid. To study such systems, she has constructed a two-dimensional
numerical model that includes more than the fluid dynamics, chemistry,
and heat and mass transport. For the solid phase, the model also describes
radiation absorption, surface consumption, and, for porous materials,
evaporation of water and crack formation. For the gas phase, the model
includes radiation absorption by the decomposition products and heat transfer
by radiation.
Combustion in furnaces is as complex as it is important. Because of the
concentrated energy release, furnace combustion is often a very dynamic
process. The chapter here by Barr and Dwyer describes numerical studies
of a particular type of dynamical furnace, the pulse combustor. This system,
used in some furnaces and boilers, is actually designed to take advantage
of combustion-driven oscillations. Pulse combustors enhance heat transfer
by large oscillations in the flowfield. That allows smaller furnaces to pro-
duce equivalent energy output. Barr and Dwyer explain the advantages of
this system and then present a series of computations that calibrate and
use a one-dimensional, time-dependent model to study a model pulse com-
bustor.
The last two chapters consider the very broad topic of fire modeling,
particularly in the study of macroscopic fires in enclosures such as rooms,
buildings, piping systems, and reactor vessels. Because of the vast range
of phenomena requiring treatment, enclosure fire models rely heavily on
phenomenological submodels and simple mathematical fits to empirical
relations. Mitler describes the range of models that have been used for
enclosure fires. The most computationally efficient of these are the zone
models which divide a domain into a small number of macroscopic zones.
These zone models are generally cast in terms of coupled ordinary differ-
ential equations and a number of associated constraints and algebraic con-
ditions.
When the spatial subdivisions (cells) are finer and the model is con-
structed by imposing sets of conservation conditions on the mass, energy,
and momentum transfer between these cells, the models are called field
models. These are in fact a natural extension of the finite-difference and
finite-volume models described in many of the preceding chapters. In field
as in zone models, submodels of complex chemical and physical processes
(such as multiphase chemical reactions, smoke formation, and soot for-
mation) and radiation effects must generally be included. Input data is
derived from a great number of experiments.
The last chapter, by Travis, Gregory, and Krause, considers the ex-
tremely serious problem of predicting and controlling fires in nuclear re-
actors and related configurations. Reactors contain many combustible
materials and possibilities for sparks and overheating to ignite a fire, de-
flagration, or detonation. Since reactor configurations differ greatly from
the corridor-connected room configurations treated by building-fire models,
such models cannot generally be used directly for nuclear-reactor studies.
Thus, this final chapter again emphasizes field-equation models to simulate
those aspects of full-scale reactor hazards and accidents and to consider
ways of controlling them.
!
Dwyer, H.A., and Sanders, B.R., "Calculations of Unsteady Reacting Droplet Flows,"
Proceedings of the 22nd Symposium (International) on Combustion, The Combustion Insti-
tute, Pittsburgh, PA 1989, pp 1923-1939.
2
Fyfe, D.E., Oran, E.S., and Fritts, M.J., "Surface Tension and Viscosity with Lagrangian
Hydrodynamics on a Triangular Mesh," Journal of Computational Physics, Vol. 76, 1988,
pp. 349-384.
Chapter 18
Liquid Drop Behavior in Dense and Dilute Clusters
Josette Bellan
Nomenclature
A = area, cm2
a = radius of sphere of influence, cm
Ac = cluster transverse area, cm2
CD = drag coefficient for one drop
=
CDC drag coefficient for a cluster of drops
Cp = heat capacity at constant pressure, cal/g-K
fB = mass of fuel burnt between ignition and drop disappearance/total
fuel mass
fFf = mass of fuel lost from cluster from t — 0 to drop disappearance/
total fuel mass
Lbn = effective heat of evaporation at normal boiling point, cal/g
Lp — penetration distance, cm [Eq. (4)]
m = evaporation rate for one drop, g/(cm2-s)
m = mass, g
n = drop number density, cm~ 3
N — number of drops
Nuc = Nusselt number for cluster
R = cluster characteristic length (radius, if cluster is spherical), cm
f = radial coordinate centered at drop center, cm
r = radial coordinate centered at cluster center, cm
R = radius of drop, cm
Re = 2Ru,Jv
Rf = radial location of flame measured on r coordinate, cm
R, = R/R°
R2 = a/R°
Shr = Sherwood number for cluster

547
548 LIQUID DROP BEHAVIOR IN CLUSTERS
T = temperature, K
t = time, s
u = velocity, cm/s
Vc = cluster volume, cm3
Yf = mass fraction of species i
z = r/R
6 = CpTILbn, nondimensional temperature
£ = similarity parameter
p = density, g/cm3
v = kinematic viscosity, jx/p, cm2/s
4> = air-to-fuel mass ratio
<|>5 = stoichiometric value of c|>
ij; = equivalence ratio 4>/<j)5
Subscripts
a — edge of sphere of influence; interstitial
d = drop
dl = dilute
evap = evaporation
F = fuel
g = gas
ndl = nondilute
r = relative
v = vapor
0.3 = instant when Rl = 0.3
Superscripts
0 = initial
oc = far field
I. Introduction
S PRAYS, created by atomizing liquid fuels, have been considered for a

long time as one of the most efficient ways of burning liquids. This is
because in a spray, the surface-to-volume ratio is increased greatly as
compared with a blob of liquid. The extra surface increases the efficiency
of heat transport to the liquid, thereby promoting evaporation, ignition,
and combustion of the fuel. Liquid sprays are also used in agriculture to
spray crops, in the food industry, in coating, and in printing.
To improve all of these processes, first, one needs to understand them.
Such understanding can be achieved in a variety of ways. For many years,
the only universally accepted way of improving a process was physical
experimentation, which involves building a prototype operating under a
variety of conditions. The results of these experiments are recorded and
compared. The prototype system can then be "tuned" to operate under
the optimal conditions achieved during the experiment. The simplistic "cut-
and-try" method described is making room for the growing use of analytical
methods and computational techniques based on the availability of large
computers. Experimental techniques, which themselves have become in-
J. BELLAN 549
creasingly sophisticated, are usually very expensive and are limited not
only by the sensitivity of the diagnostic but also by the range of conditions
that can be studied. In many cases, it is less costly to vary the parameters
in an existing computer code to simulate a different physical situation than
it is to build or alter an experimental setup to obtain new experimental
results. The flexibility of computer codes in terms of input conditions and
output results makes them a very powerful tool both for engineering cal-
culations and for understanding the fundamentals of physical phenomena.
This chapter describes how the computational approach has helped the
understanding of the fundamentals of spray evaporation, the interactions
between the drops in a spray, the interactions between the drops and the
gas surrounding them, and subsequent ignition and combustion. This
understanding, in turn, provides the basis for controlling these phenomena.
A typical liquid spray is composed of three main regions, although the
boundary between these regions is not always very distinct. Near the at-
omizer, there is a region of liquid filaments, which are the precursors of
drops. Adjacent to this region is a region of very closely spaced drops such
that the spacing between the drops is of the same order of magnitude as
their size; this is called the dense spray region. Further downstream, one
encounters a region where the drops are no longer closely spaced; this is
called the dilute spray region.
Recent observations of liquid-fuel sprays have revealed that although
the fuel flow to the injector is constant, drops appear to cluster in the spray
and remain clustered during combustion. 1 " 4 Thus, there is not one con-
tinuous flame surface in a burning spray, but instead there are many flame
surfaces, each one enclosing groups of drops.3-4
A characteristic feature of sprays is the wide range of space scales in-
volved in their description. For example, a few of the most obvious scales
are the scale of the enclosure, if any, where the liquid is sprayed; the many
turbulent scales associated with turbulence buildup and decay; the scale
of droplet interactions; the average distance between drops; and the scale
of drops themselves. These scales vary by orders of magnitude from the
largest to the smallest, and this implies that an accurate mathematical
description at all scales is impractical. To circumvent this difficulty, it has
been proposed5 that the macroscale, where phenomena occur that are of
interest to engineers, be described in detail, and that the microscale, which
encompasses the scales much smaller than the macroscale, be described in
an approximate and global manner. The coupling between the two scales
must be achieved through the boundary conditions at the microscale level.
The microscale formulation is also sometimes called a subscale, or subgrid
model, because the phenomena that are involved occur at a scale much
smaller than that of the grid size used to resolve the macroscale problem
computationally. Inherent to this two-level description is the understanding
that the subscale models are more approximate than the macroscale models
and lack the detail that the latter are required to have to be useful. This
concept is similar but not identical to the proposed particle-source-in-cell
model by Crowe et al.6
The observation of clusters of drops in sprays points to the natural choice
of the cluster size as one of the important microscales. Within this frame,
it is envisaged that each cluster should be followed on its trajectory in a

Lagrangian way and that the coupling between the cluster and its sur-
roundings should be achieved through proper boundary conditions at the
cluster surface. One example of the implementation of this concept can be
found in Ref. 7. Tambour7 partitions the spray into sections of drops of
known characteristics and follows each group of vaporizing drops along a
streamline in a Lagrangian manner while calculating one integral charac-
teristic quantity for each section. In his model, this integral quantity can
be either the number of drops, the surface area of the drops, or the volume
of the drops. A missing input of this model is a calculation of drop evap-
oration rate in each section; instead, Tambour cleverly estimates this rate
from the experimental data of Yule et al.8 using the dMaw approximation.
As expected, the characteristics of each group of drops change during
calculation. Thus, it is important to calculate these changes as they occur
rather than to fit them in an approximate way a priori using experimental
data. This example indicates that models of droplet clusters must be able
to handle the many areas where drops exist in a spray, going from the
dense region to the dilute region as discussed earlier.
Experimental observations show that dense and dilute collections of
drops behave very differently. In the dilute regime, the drops are so far
away from each other that they behave as if isolated for all practical pur-
poses. The behavior of individual, isolated drops under a variety of as-
sumptions is described by classical theory.9 In contrast, when drops are
closely spaced, their behavior changes due to limitations on heat and mass
transfer that result from drop interactions with the background gas. These
interactions are called indirect, or long-range, interactions. Direct, or short-
range, interactions, such as drop collisions, coalescence, or breakup, will
not be discussed here. Beer and Chigier10 show that when the quantity
obtained by dividing the difference between the initial droplet diameter
squared and the actual diameter squared by the time elapsed from the
initial condition is calculated for burning droplets, it is found that it varies
nonmonotonically with the average interdroplet separation distance. In
contrast, this quantity is a constant for an isolated drop and is called the
"burning constant." As the interdroplet separation distance is reduced from
large values for which drop interactions are negligible, the burning constant
increases to a maximum and then decreases. This behavior arises because
heat losses are reduced when the drops are brought slightly closer together,
thereby increasing the burning constant. However, when the separation is
greatly reduced, the amount of oxygen available for combustion diminishes
and, thus, decreases the burning constant.
This latter situation, which typically prevails in sprays, was investigated
experimentally by Koshland1142 and Koshland and Bowman. 13 Their results
show that, in agreement with Chigier and McCreath14 and Rao and Le-
febre,15 both evaporation and combustion of groups of interactive drops
depend on drop number density as well as oxygen concentration in the
ambient gas. Oxygen competition was identified11"13 as one of the processes
that can change the combustion mode of groups of drops from individual
flames around drops (in an oxygen-rich environment) to a group combus-
tion mode where the flame surrounds the droplet cloud (in an oxygen-lean
J. BELLAN 551
environment). Thus, the prediction of the global gas-phase characteristics

inside the drop cloud is of paramount importance for calculation of both
evaporation and burning rates.
Moreover, many of the pollutant-forming reactions also depend indi-
rectly on drop number density and oxygen concentration. Sangiovanni and
Liscinsky16 showed experimentally that soot production is a strong function
of the interspacing between drops. Their results are actually conservative
with respect to the situation in a spray, because in their droplet-stream
experiment there was no way to account for the influence of lateral spacing
between drops in the direction perpendicular to the stream. In a real spray,
the average drop is surrounded in all three dimensions by other drops. For
a variety of fuels and ambient oxygen concentrations, it was shown that
by increasing the droplet spacing, a substantial decrease can be obtained
in the soot emission.16 Reductions in spray flame temperature and NO V
levels were also associated with droplet interactions, as shown by Cher-
nansky and Sarv,17 who interpreted their results in terms of oxygen de-
pletion during burning. In addition, the question of whether the drops
disappear well before the entire amount of fuel vapor is burnt has direct
bearing on the formation of carbonaceous, porous particles, called cen-
ospheres. These particles form from each individual drop during combus-
tion of heavy-oil sprays once the more volatile hydrocarbons have evaporated
and the tarlike material is left behind. It appears that, contrary to the
classical d2-\aw assumption of equal vaporization and burning rates, even
for mildly dilute groups of drops (initial drop number densities of 18-80
drops/cm3), the vaporization rate is higher than the fuel vapor oxidation
rate.11 Thus, the drops disappear well before the vapor fuel is burnt, and
homogeneous burning prevails during the latter stage of drop cluster com-
bustion.
The preceding discussion indicates that models of droplet clusters must
be flexible enough to accommodate both dilute and dense configurations,
and that the cluster drop number density influences many of the crucial
aspects of combustion. The remainder of this chapter is organized as fol-
lows. Section II describes some current models of drop interactions and
shows that they have limitations important enough to preclude them from
being used as subgrid models. The third section presents an approach for
modeling drop interactions. The model has the potential for use as a subgrid
model. In addition, specific situations that were modeled and the results
obtained using this formulation are discussed. Finally, Sec. IV presents a
discussion of the needs for further research.
II. Survey of Drop Interaction Models

In the present context, models of droplet interactions describe the in-
teractions among a large number of drops. This precludes models of de-
tailed interactions between small numbers, such as two or three drops.
Although the few-drop models have helped to understand departures from
single, isolated-drop behavior, they themselves are not candidates for subgrid
models of dense sprays. Moreover, these few-drop models involve large
computational expense as well and, thus, are not suitable economically as

subgrid models.
One of the earliest contributions to the modeling of multiple-drop in-
teractions was made by Zung18 for the evaporation of drops in atmospheric
clouds. The cloud is approximated by a series of cubes having sides equal
to the distance between the centers of two adjacent drops. Thus, each cube
contains an identical drop in its center, and these cubes are approximated
further by spheres of equivalent volume. Zung assumes that no mass trans-
fer occurs between cells, and uses this assumption to predict the rate of
drop disappearance in terms of an average concentration inside the cell.
The results of this model show that the evaporation rate of a cloud is
strongly dependent on both the drop radius and the distance between drops.
In particular, saturation of the vapor in the ambient gas can occur before
complete evaporation in some cases.
The drop-in-a-cell idea is also used by Tishkoff, 19 who simulated the
interactions among evaporating drops by using an isolated drop in a "bub-
ble" of finite and constant radius. Because boundary conditions at the
bubble surface are specified as a function of time, the actual interactions
between drops during evaporation are given rather than calculated. How-
ever, by cleverly changing the size of the bubble, Tishkoff shows that the
initial spacing between drops strongly influences the outcome of evapo-
ration. When the bubble is large compared with the drop, the drop evap-
orates completely; whereas, when it is small, vapor-phase saturation may
occur before complete evaporation.
Chiu and his co-workers20"23 took a very different approach. In their
initial analysis, a parameter G, which is the ratio of total heat transfer
between the two phases to the heat needed to evaporate the drops, is
identified as a crucial number indicating the combustion mode for the drops
in the spray. If G exceeds a critical value, a single flame surrounds the
entire spray and the drops burn in what is called an "external group"
mode. In contrast, below this critical value of G, isolated droplet burning
occurs. This analysis was later22 refined to include other combustion modes
between these two extremes. The weakest aspect of this model is the
difficulty in obtaining information necessary to calculate G. Finding G
involves the use of empirical formulas characterizing the exchange of mass,
momentum, and energy between the phases. These formulas are based on
those of the classical single-drop evaporation in infinite surroundings. Re-
alizing this, an attempt was made recently24 to improve this model through
the use of a modified drop-in-a-bubble model. The "test" drop is sur-
rounded here by a "first coordination" shell, which corresponds to the
radial position of the largest number of drops close to the test drop. Thus,
statistically there can be other drops inside the cell, but the model will not
take this into account. Using the same idea, a series of second, third, etc.,
coordination shells is developed. The farthest shell from the drop is a
transition shell, the outer edge of which has a continuum of drops; at this
outer edge, the gas properties are prescribed for the determination of the
drop evaporation rate. The droplet arrangement inside these shells must
be provided either experimentally or theoretically through a pair-distri-
bution function that represents the joint probability of finding a drop lo-
J. BELLAN 553
cated at a given distance from the test drop. Since this information must
be provided at every instant of time, and it is very difficult to obtain, except
for the simplest of situations which can be treated using simple formalisms,
this latest model of Chiu et al. does not seem practical.
The idea of group combustion was also examined by Correa and Sichel.25
Their model assumed that the temperature of the gas in a spherical, uni-
form, monodisperse cloud of fuel drops immersed in a quiescent atmos-
phere was initially in saturated equilibrium and that it stayed so during
combustion. This assumption precludes the calculation of the influence of
drop proximity on the gas temperature inside the cloud and has a bearing
on the calculation of the drop evaporation rate. In addition, they also
assumed the ratio of the interdrop distance to the drop diameter to be of
order 10, so that the conclusions based on results obtained with this model
apply only to this regime and not to much denser collections of drops. The
largest
3
drop number density for which results were presented was 3 x 103
cm~ . The results predicted the existence of a thin, inwardly propagating
vaporization wave at the edge of the cloud, a decrease in cloud radius
following the classical d2 law with a modified "evaporation constant," and
a flame- to cloud-radius ratio and flame temperature independent of the
cloud parameters. The flame temperature was found to be independent of
the cloud parameters because the gas temperature inside the cloud was
assumed fixed at the saturated equilibrium condition during evaporation
and combustion.
Group combustion of drops was modeled in a totally different manner
by Labowsky and Rosner26 and Labowsky.27~29 Under the quasisteady gas-
phase assumption, two different approaches are used to model interacting
droplet burning. 26 In the first approach, the cloud is treated as a continuum
with the droplets acting as distributed sources of fuel and sinks of oxygen.
In the second approach, the flame location is calculated for cubical arrays
of up to 729 particles. The flame location is found by solving the Shvab-
Zeldovich equations. These equations are transformed further into Laplace
equations, which, in principle, can be solved by the method of images.
However, since the computational costs of the second approach become
prohibitive when the number of particles reaches about 20, the authors
solve the equations using a superposition method, which is an approxi-
mation of the method of images. In this approximation, higher-order terms
on the image series are neglected, with the result that particles are again
treated as point sources and their fields are simply superimposed. Thus,
again, interactions due to particle proximity are neglected, as these inter-
actions would have been contained in the higher-order terms. The contin-
uum method yields reliable solutions only when the number of particles is
large (in a dilute cloud), and the second method yields reliable solutions
only when the number of particles is small. The results show that in virtually
all practical situations of interest, clouds burn as a total group. This con-
dition is defined as that occurring when the cloud as an entity provides
sufficient vapor so that fuel and oxygen meet in stoichiometric proportion
precisely at the cloud boundary where they will burn.
Although drop interactions are not accounted for by Labowsky and
Rosner,26 one important contribution of this work is to have identified the
ratio of the fuel-cloud radius divided by the fuel-particle radius as the

important characteristic that determines the mode of combustion of the
drop in the cloud. If this ratio is less than a critical value, the particles
burn individually. Conversely, if this ratio is larger than another critical
value, the flame surrounds the entire cloud. Presumably, if the character-
istic ratio falls between these two critical values, there will be several
separate flames surrounding groups of drops inside the cloud. Although
the superposition approximation is not made by Labowsky in Ref. 27, the
actual calculations are limited to a three-droplet array, which limits the
generality of the results. In addition, the analysis uses the assumption that
the particle temperature is not affected by the drop interactions, an as-
sumption that is certainly not valid for closely spaced drops, as it will be
shown later. In fact, all of the comparisons between this theory using the
method of images and experimental observations28"30 are made with ex-
periments containing two, three, or a stream of drops. This is because the
method of images leads to financially prohibitive calculations for arrays of
more than about 20 drops.26 The largest number of drops considered in a
calculation performed using this method was 25, and the comparison was
made with a stream of interacting drops.30 For the two, three, or 25 linear-
drop array, the comparisons with experiments showed reasonable agree-
ment for the drop lifetime when compared with experiments involving two,
three, or a stream of drops, respectively.
Samson et al.31 developed a simple model to describe spray combustion
by combining the classical isolated drop theory devoid of hydrodynamic
effects and various statistical concepts. In this model, the burning rate of
the drops is identified with their evaporation rate, a fact that is contradic-
tory to experimental evidence.11 Some favorable comparisons with exper-
imental data are presented, but to obtain these results the unknown radius
of the spherical fuel cloud had to be assigned. The values chosen for this
radius were selected to obtain the best comparisons between experiments
and theory. However, the authors point out31 that the experiments used
are not really appropriate to compare with the theory because they involve
well-defined droplet arrays for which statistical models are not applicable.
Umemura32 identifies a function /, intrinsic to the geometrical configu-
ration of the cloud and independent of the combustion characteristics, as
the characteristic indicator of droplet interactions. In Umemura's 32 quasi-
steady quiescent-atmosphere theory, the drop evaporation rate is again
identified somewhat incorrectly with the burning rate, just as in the isolated
drop theory. Results are presented for a system of two droplets.
The effect of particle interactions—first on drop evaporations, and then
on drop burning—was also modeled by Ray and Davis33 and Marberry et
al.34 The isothermal droplet evaporation model of Ray and Davis33 further
assumes that, despite evaporating, the particles do not change size (rate
of change of size much smaller than the diffusive velocity of the transferred
species) and, thus, there is no convective transport. By assuming that the
particles can be treated as point sources or sinks, the authors preclude the
treatment of dense collections of particles where the concentration or tem-
perature field around each drop interacts with that of the neighboring
J. BELLAN 555
drops. Furthermore, the results show that the model becomes invalid if
the number of particles exceeds the ratio of the distance between particles
divided by the particle radius. This represents a serious restriction to the
model. The same assumptions of quiescent atmosphere and constant par-
ticle size are used by Marberry et al.34 The burning rate of each drop is
given by material and energy balance at the drop surface, and these are
the equations that now determine the source strength at each droplet lo-
cation. The results indicate significant deviations from the isolated droplet
burning rate if the distance between the drops is smaller than 20 times the
droplet radius. These results were obtained for systems of 2, 3, and 4
drops, tetrahedral and cubical arrays, so that none of the drops was sur-
rounded entirely by other drops. In that respect, the estimate of the pre-
ceding ratio of 20 should be understood as a lower limit.
Annamalai35 modeled the evaporation, ignition, and combustion of a
droplet cloud as well. However, the results cannot be expected to be
reliable because his formulation did not contain a genuine description of
droplet interactions.
The three-dimensional numerical study of Shuen36 was meant to inves-
tigate the effect of droplet interactions on transport phenomena for droplet
Reynolds numbers of 5-100, droplet spacings of 2-24 diameters, and
oxygen concentrations of 0.1-0.2. The physical situation studied is that of
a monosize, planar, semi-infinite droplet array oriented perpendicular to
the direction of the flow. Shuen also assumed a quasisteady gas phase with
respect to the liquid phase and no drop heating. The geometry studied
suggests immediately that droplet interactions will be underpredicted since
there is no drop that is surrounded on all sides by other drops in a three-
dimensional sense. Furthermore, since the array is semi-infinite, the amount
of heat available to the drops is not a limiting factor, as it is in real sprays.
Moreover, the neglect of droplet heating immediately suppresses one of
the most important modes of droplet interaction; that is, competition among
drops for the available heat. It is not surprising that Shuen finds that
interactions between drops become negligible for spacings greater than six
droplet diameters and Reynolds numbers greater than 10. He also finds
that drop interactions result in an insignificant change in drag per drop.
All of the previously-discussed models of drop interaction are deficient
in some important way, which precludes both the prediction of the behavior
of dense clusters of drops and their inclusion as subgrid models, as discussed
in the Introduction. In the following, models that have promise in both of
these respects are discussed. In contrast to the models discussed so far,
they include explicitly the effects of drop interactions. In addition, they
are formulated in such a way that boundary conditions with the surrounding
gas can be changed to describe a variety of situations.
III. Models of Drop Evaporation, Ignition, and Combustion Based on

Multiple-Drop Interactions
The models that follow pertain to a variety of situations, and a description
of their assumptions can be found in Table 1.
Table 1 Summary of subgrid models

No. Situation described
(Refs.) by the model Model assumptions
1. Constant-volume, a) Unsteady liquid phase
variable pressure b) Uniform drop temperature
cluster c) Gas phase quasisteady with respect to liquid phase
(27)
d) Equilibrium evaporation
-evaporation e) No convective effects
f) Cluster volume is adiabatically insulated from
surrounding gas.
a and c
g) Heat conduction in drops in the radial direction
h) Kinetic evaporation
(40) -evaporation Convective flow in the gas surrounding the cluster
Mass, species, momentum, and energy transfer across
the cluster boundary is negligible with respect to the
mass, species, and energy content of the cluster.
2. Constant-pressure, a, c, g, and h
variable volume k) Turbulent flow exists around the cluster. It can be of
cluster two types:
1) Turbulence model 1—turbulence is not present
initially, but develops with time.
(5) -evaporation 2) Turbulence model 2—turbulence is present
initially.
1) The drops inside the cluster move radially with respect
to each other. The cluster boundary is found by
assuming that all of the evaporated fuel exits the
cluster in the dense case. In the dilute case all of
the evaporated fuel is trapped inside the cluster. An
interpolating factor, called "trapping factor" is used
for intermediate situations.
m) Transport of mass, species, momentum, and energy
across the surface of the cluster is important.
a, c, g, h, k, and m
n) The drops inside the cluster move radially with respect
(46, 47)
-evaporation to each other in a self-similar manner. The
similarity parameter is the ratio of the radial
coordinate inside the cluster divided by the radius
of the cluster.
a, c, g, h, k, m, and n
(46) -evaporation and o) The drops are electrostatically charged with the same
ignition charge,
p) Ignition is dominated by diffusion.
a, c, g, h, k, m, n, and p
q) The burning rate is not necessarily equal to the
(51) evaporation rate.
-evaporation, r) Burning occurs according to the following scenario: a
ignition and very shortly-lived premixed flame depletes all the
combustion oxygen inside the cluster and is followed by a thin,
sheath-like, counterflow diffusive flame.
J. BELLAN 557
Constant- Volume (Variable-Pressure) Cluster Models

In the approach taken by Bellan and Cuffel, 37 a cluster of spherical,
monodisperse, uniformly distributed drops is considered, and the cluster
volume, Vc, can have an arbitrary shape and contain an arbitrary number
of drops, N. Figure 1 shows each drop of the cluster surrounded by a
fictitious sphere of influence whose radius is one-half the distance between
the centers of two adjacent drops. Thus, the volume of the cluster is, in
fact, the volume of these spheres of influence added to the volume of the
space between the spheres of influence. From simple geometrical consid-
erations usual in solid-state theory, 38 it is known that the radius of the
sphere of influence thus defined is
a = [0.74Vc/(/V 4TT/3)] (1)
where Vc is the cluster volume, N the total number of drops, and a the
radius of the sphere of influence. In Ref. 37, the volume Vc was adiabat-
ically insulated, but this restriction is removed in later studies, which are
discussed subsequently. On the length scale of many drops, the cluster is
considered spatially homogeneous in the thermodynamic quantities. This
means that each drop behaves similar to all of the other drops in the cluster.
Moreover, all dependent variables are assumed to be uniform in the in-
terstitial space between the spheres of influence. Thus, the conservation
equations for this cluster of drops consist of 1) the differential conservation
equations for each drop and gas phase inside a sphere of influence, 2) the
global conservation equations for all drops inside the cluster volume, and
3) the perfect gas law. In Ref. 37, the classical quasisteady assumption for
the gas phase is made, and, thus, the differential equations for the gas
phase inside the sphere of influence can be solved exactly. The quasisteady
assumption is well justified at atmospheric pressure since p^/p { — 10 ~ 3 ,
where p^ and p{ are the gas and liquid densities, respectively. For simplicity,
Fig. 1 Pictorial of drops in a cluster surrounded by fictitious spheres of influence.

it is assumed that thermodynamic equilibrium prevails in the gas phase at

the drop surface, and that the drop temperature is uniform but transient.
(These assumptions are also relaxed in the later studies discussed subse-
quently.) The solutions of these equations contain as parameters the un-
known values of species and temperature at the edge of the sphere of
influence. Since these quantities also appear in the global conservation
equations and the perfect gas law, the system of equations is closed. At
every time step, the solution consists of the evaporation rate, the radius
of the drop, the value of the fuel mass fraction and temperature at the
drop surface and at the edge of the spheres of influence, the density and
the pressure (both assumed uniform in Vc).
Comparing this model with some of the models previously described,
one notices that this model is different both from Tishkoff s bubble model19
and Zung's cell model,18 even though it also isolates each drop in a sphere
of influence. It is different from Tishkoff s model because the conditions
at the edge of the sphere of influence are not imposed, but instead are
calculated as a consequence of the droplet interactions. It is different from
Zung's model in quite a few aspects, but mainly because it does not impose
zero mass transfer between the spheres of influence and it relates the
evaporation rate to an exact mass fraction inside the sphere of influence
instead of an averaged value. When the volume of the cluster is constant
and the cluster is insulated from the surroundings, the quasisteady and
spacial homogeneity assumptions imply that there is no transport between
the spheres of influence because there is no transport out of the cluster
and because of symmetry. In contrast to Zung's model, however, we can
evaluate how well the quasisteady assumption holds by calculating the
gradient at the edge of the sphere of influence. The results show that,
because the quasisteady assumption deteriorates as the radial coordinate
r increases, the model improves as a in Eq. (1) is smaller, i.e., for dense
collections of drops.
All of the subsequent calculations are for n-decane. Symbols in the
figures and text are defined in the Nomenclature. Figure 2, reproduced
from Ref. 37, shows plots of both the cluster evaporation time, tendh ob-
tained with the model described earlier and the ratio of this time to that
of the evaporation time obtained with a model of dilute spray evapora-
tion,39 tendl/tedl. For dilute collections of drops, the predictions of the two
models agree. However, as soon as the nondimensional radius of the sphere
of influence, R2, is smaller than or about 10 (see Fig. 3), departures from
the dilute theory occur. In fact, when comparing with the dilute theory,
four universal regimes are identified as the equivalence ratio v|/ varies from
very large values (fuel lean) to very small values (fuel rich). In the first
very lean regime, the results of the two theories agree. In the second
regime, evaporation is completed before saturation, but the ratio of the
two evaporation times, tend^lted^ is a function of i|/. In the third regime,
nondilute spray theory predicts saturation before complete evaporation,
whereas the dilute spray theory predicts the opposite. Finally, in a fourth
regime, both theories predict saturation before complete evaporation but
at different times and values of the residual drop radius, R1. In the first
and second regimes, tendi is a decreasing function of vj;, whereas in the third
and fourth regimes it becomes an increasing function of i[>.
J. BELLAN 559
COMPLETE
R° = 2x10 cm
EVAPORATION
BEFORE Y =
SATURATION Fva °' °°
TO =350°K
gs
T° = 550°K
Fig. 2 Variation of ^/^ with i|/ for constant-volume, adiabatically insulated clus-
ter of drops.
170
160 3
R°=2x10 cm
150
140 Y°va=0.00
130
120 TgS - 350°K
110
T° = 550°K
100
™90
80
70
60
50
40 FUEL -FUEL
30 RICH LEAN'
20
10
0 -2
10 10 10 10 10 10
V
Fig. 3 Nondimensional radius of sphere of influence vs i|/.
Although these regimes are believed to be universal, the values of i[/

separating them change as a function of the fuel and the initial conditions.
Thus, nondilute effects can occur even at values larger than R2 = 10. For
example, nondilute effects can be found even for very dilute sprays when
the total mixture is rich. This is illustrated in Fig. 4. When \\i is constant
1 .UO
n -3
Fr=2xlo "cm
1.06 5.0
w
Y T° T°
gs ga
........... 1.0 500°K 750°K
1.04
\ — •— 0.5 500°K 750°K
\ — -— 0.1 500°K 750°K
1.02 _ V _
4.0
\
\
1.00 - ! \
—/ \
1 \
0.98 ± : \ 3.0
;/^ \
WN
0.96 - \\ i
•*v
0.94 \ 2.0
\
0.92 -uT "~ur""—"' x x "^

\ i
y
n QD \ i i ^- i 1.0
0.0 0.1 0.2 0.3 0.4
INITIAL FUEL MASS FRACTION AT r = a, Y°va
Fig. 4 Variation of ?ndilt?d(_ and tend{ with Y%va for constant-volume, adiabatically
insulated cluster of drops.
and the initial interstitial mass fraction of fuel vapor, YFva, increases, the
cluster becomes more dilute and the liquid-fuel mass decreases. Because
of this, the interstitial gas temperature, Tga, stays higher than at lower
values of Ypva since less heat is transferred to the evaporating drops. Thus,
as Y%va increases, two competing effects determine the rate of evaporation
and, therefore, the value of tendi\ Ypva is larger, which tends to decrease
the evaporation rate, but Tga is larger as well, which tends to increase the
evaporation rate. The interplay between these two effects is such that, as
Ypva increases from zero, the vapor-pressure effect dominates and, there-
fore, tendi increases with increasing Y°Fva. In contrast, the dilute spray theory
predicts that ted€ decreases monotonically with increasing Y%ua.
The results of Fig. 4 are not directly comparable with those of Kosh-
land,11 because in those experiments the oxidizer mass fraction and droplet
spacing were decreased independently. Here, as Y°Fva increases (and thus
the mass fraction of oxidizer decreases), the droplet spacing increases be-
cause the equivalence ratio is kept constant. However, since Fig. 4 shows
that tend€ is a very weak function of Y%va and since a decrease in droplet
spacing results in increasing tendi (see Figs. 2 and 3) for complete evaporation
before saturation, it is expected that tend( increases when the oxidizer mass
fraction decreases while the droplet spacing is constant. This yields the
same result as that of Koshland,11 who found the burning rate decreasing
J. BELLAN 561
with decreased ambient oxidizer mass fraction. The results of the theory
of Bellan and Cuffel37 also showed that the square of the normalized radius
does not decrease linearly with time, and thus the d2 law does not hold
for interacting droplets.
To improve the predictive abilities of this model, some simplifying as-
sumptions of the preceding formulation were relaxed in Ref. 40. In par-
ticular, the more realistic Langmuir-Knudsen kinetic evaporation law replaced
the Clausius-Clapeyron relationship, and the drop temperature was con-
sidered not only transient but a function of radial position.
The general idea of the model is to account for global effects. Thus,
although there is a net flow of gas and heat through the surface of the
cluster, these effects are not modeled in detail, but only globally. In this
particular case, mass and heat inflow or outflow through the boundary are
assumed negligible with respect to the mass and heat content of the cluster.
The results show that, in fact, this is a very good approximation for non-
dilute clusters, where penetration of the clusters by surrounding gas is
confined to a very thin shell. Conversely, for dilute clusters where pene-
tration is substantial, convective flow effects are appropriately taken into
account by correlations relating the evaporation rate in convective flows
to those in quiescent flow and to the Reynolds number. For intermediate
regimes of drop number density, the model is still expected to be a good
global approximation. Consistent with the assumption of small pressure
gradients, the exchange of momentum between drops and gas is treated
on a local basis and is considered to be due to 1) transfer due to evaporation,
2) transfer by fluid flow interaction in the form of a drag coefficient, and
3) transfer due to small pressure gradients. The ensuing momentum equa-
tions are
^ = -n{rh ur/pg + i [1 + Pg/(nmd)] AdCDu2r} (2)
2
Du + [nmd/(pg + nmd)] pgugutiCDcAc/N} (3)
where ur = (ud — ug) is the relative velocity between drops and gas inside
the cluster, m the evaporation rate, n the drop number density, p^ the gas
density, md the mass of one drop, N the total number of drops, Ac the
transverse area of the cluster, Ad the transverse area of a drop, and CD
and CDc are, respectively, the drag coefficient for one drop and the drag
coefficient for the cluster as an entity. These momentum equations are
coupled to the other equations. Here CDc is based on a length scale [Ac(ug/
u^/u]05 and is a function of the resultant Reynolds number only. The
Reynolds number is based on ud. In contrast, CD depends on both Re and
m. The dependence of CD on R2/Ri, and thus the "blockage" effect due
to drops shielding other drops from the flow, is neglected here. It is im-
portant to notice that these two equations yield the correct limits in the
cases of no evaporation (m —> 0), no slip (ur —> 0), and quiescent ambient
gas (ug-* 0).
Since the kinetic evaporation law and the equation of state form a non-
linear implicit set of equations for the pressure and evaporation rate, a
predictor-corrector scheme is used to solve the entire set of equations. The
drop temperature is solved as a four-term truncated series in powers of
the nondimensional radial coordinate z, and a GEAR integrator package
is used to solve the set of ordinary differential equations. For each inte-
grator step, the iteration starts with the prediction of the drop temperature.
Then the equation of state and kinetic evaporation law are solved using a
Newton-Raphson iteration for the pressure and evaporation rate. Next,
the convective correction is applied to the evaporation rate, and, finally,
the drop temperature is corrected. The iteration is repeated, starting with
the Newton-Raphson procedure, until convergence is obtained. Thus, each
time step requires a nested double-loop iteration to calculate the drop
temperature, evaporation rate, and pressure. The calculations can be per-
formed on a personal computer, or on a faster computer if more speed is
desired.
An interesting quantity to calculate is the penetration distance of the
surrounding flow into the cluster, Lp. This is done by following the cluster
on its trajectory and identifying this distance with the relaxation distance,
which is the distance traveled to the location where ur — 0. Thus, Eq. (2)
is rewritten in Lagrangian variables and integrated. The solution is
Lp = {2 41 + u°rpg(pg/md + n)AdCDl(2mn}]}l(pglmd + n)AdCD (4)
If the ratio LpIR, where R is the cluster radius, is larger than unity, pen-
etration is important and, thus, evaporation is controlled by convection.
In contrast, if LpIR is much smaller than unity, evaporation is controlled
by diffusion. Between these two regimes, there is an intermediate regime
where both convection and diffusion are important. In this intermediate
regime, the evaporation rate is very close to that in the convective evap-
oration regime since convective effects dominate diffusive effects during
evaporation. Calculations40 of LpIR with R = 10 cm show that the diffusive
regime corresponds to the very dense (n > 105 cm~ 3 ) and dense (105 cm~ 3
> n > 104 cm~ 3 ) clusters, whereas the convective regime corresponds to
dilute clusters (n < 103 cm~ 3 ; <j) > $s = 15.7 for n-decane, where cj> and
$s are the air/fuel and stoichiometric air/fuel mass ratios, respectively).
Figure 5 shows that the evaporation time reaches an asymptote fast as <j>
increases from the dense to the dilute regime. In addition, this figure shows
that the model is not very sensitive to the drag model used,41"43 provided
that the drag model accounts appropriately for the "blowing" effect due
to evaporation. This blowing effect tends to decrease the drag coefficient
by comparison with the nonevaporation case. The conclusion regarding
drag modeling is valid for the small-to-moderate range of relative velocities
used in the calculation. Furthermore, the plots of Fig. 5 show that u°r is a
relatively weak parameter for controlling evaporation in both the dense
and dilute regimes.
The variation of the relative velocity of drops evaporating in clusters of
various initial equivalence ratios is shown in Fig. 6. Figures 7 and 8 show,
respectively, the variation of the drop velocity and Reynolds number with
J. BELLAN 563
I I i
——— REF. 41 DRAG MODEL, u° - 200 cm/sec
- - - - - REF. 42 DRAG MODEL, (fr = 200 cm/sec
— •— REF. 43 DRAG MODEL, L/J! = 200 cm/sec
— — REF. 41 DRAG MODEL, u° = 1 000 cm/sec
14 - — •— REF.41 DRAG MODEL, u° = 40 cm/sec -
12 - n = 5x104cm-3 R ° = 2 x10~
3
cm
R 2 =7.6 0 _3
10
Vi
|\ TgS = 350°K —
8
%\ Tga = 1000°K
6 n = 4. 8 cm"3 ~
R 2 = 166.8
4 %^__
*-————iL-.Il- — —•——r — ^
2
RICH" • > LEAN
n I I I
10 10 10 10 10
Fig. 5 Evaporation time vs <f> for several initial relative velocities and several drag
models. R = 10 cm.
residual drop radius. Figure 6 shows that the relative velocity of a dense
cluster of drops decreases faster than does the relative velocity of a dilute
cluster. The opposite is true for the drop velocity, as shown in Fig. 7. This
is due to the fact that, when a dense cluster of drops moves through the
gas, it exposes a greater surface area to the flow, because at fixed R the
number of drops in the cluster is larger. For this reason, there is a stronger
interaction between drops and gas, and thus faster relaxation of ur to zero.
In contrast, ud depends on the inertial effect of the cloud. Since a dense
cloud has a larger mass, it slows down less than a more dilute cloud. Finally,
we have found that the Reynolds number decreases very fast during the
lifetime of the drop, which is in agreement with the isolated drop results
of Dwyer and Sanders.44 These results also show that, in convective flows,
the concept of cluster density is strongly related to the value of the relative
velocity between drops and gas. Dense clusters quickly become "nonpo-
rous," i.e., impenetrable to the outer flow. In this case, only the outer
shell of drops is exposed to the flow. In contrast, dilute clusters are "po-
rous" and the ambient flow substantially penetrates into the cluster.
The results of Bellan and Harstad40 show that the drop temperature
becomes uniform very fast (about 1% of the drop lifetime) in dense clusters
of drops. In addition, as depicted in Fig. 9, it is found that the drop
temperature also becomes quickly uniform even for dilute collections of
3
R°=2x10 cm
220
T° = 1000°K
200
180 Fva
u° = 200cm/sec
160
7.85
3.14
140
1.57
0.785
120 0.314
100
80
ARROW POINTS IN
60
THE DIRECTION OF
DECREASING <J>'S
40-
20-
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 6 Variation of relative velocity with residual drop size for several
<(>. R = 10 cm.
drops, although the heat-up time is now about 20% of the drop lifetime
and, thus, cannot be ignored. This is to be contrasted with computational
results for single drops, which showed that the temperature was remaining
nonuniform during most of the drops' lifetime.45 In agreement with the
results of Ref. 45, Fig. 9 shows that the drop temperature continues to
increase during the entire drop lifetime for dense and dilute clusters of
drops. It must be pointed out that the results of Ref. 40 were obtained
using a transient conduction equation to describe droplet heating. This is
justified because the ratio of the characteristic time for circulation to the
characteristic time for heat-up [M(p<C^)/(|Vp<)] is on the order of 3 x
10~2. Here X, Cp, and JJL are heat conductivity, heat capacity at constant
pressure, and viscosity, respectively, and the subscript € refers to liquid.
Thus, the heating time is indeed independent of the circulation time, jus-
tifying the use of the spherically symmetric, heat conduction equation.
Based on these results, it seems appropriate to ignore drop-temperature
J. BELLAN 565
220
200 -
0.0 0.2 0.4 0.6 0.8 1.0
Fig. 7 Variation of drop velocity with residual drop size for several
<(>. /? = 10 cm.
nonuniformities when one considers the evaporation of dense clusters of

drops for single-component fuels. Thus, it is appropriate to use a one-
dimensional spherically symmetric model to describe the evolution of the
drop temperature, instead of the more sophisticated model of Ref. 45.
Since subgrid models should be computationally economical, because they
are to be incorporated into large and expensive macroscale codes, the use
of a simpler model could yield substantial computational savings.
Constant-Pressure (Variable-Volume) Cluster Models
Evaporation
Although the models described earlier contribute to understanding the
difference between the evaporation of dense and dilute clusters of drops,
such models are not appropriate as subgrid models. The reason for this is
the lack of appropriate boundary conditions at the macroscale level to
describe appropriately and self-consistently the interactions with the sur-
n——i——i—— ~T
8.0 - R°=2x10~ 3 cm = 235

7.85
7.0 _ cua=3x10-3
Yr Fva 3.14
1.57
6.0 _ TgS = 350°K 0.785
0.314
5.0 T°a =1000°K
= 200cm/sec
4.0
ARROW POINTS IN
3.0
THE DIRECTION OF
DECREASING <J>' s,
2.0-
1.0-
0 I J_
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Fig. 8 Variation of Reynolds number with residual drop size for several
<)>. R = 10 cm.
roundings. These boundary conditions provide the coupling between the

microscale and the macroscale and, thus, are crucial to a subgrid model.
The models described next apply to clusters of drops that are as large
as, or larger than, the Taylor macroscale in a flow; that is, they are at least
as large as the largest dynamically significant eddies in the flow. This means
that, unless the scale of the computational grid is larger than the Taylor
macroscale, the following models cannot be used as subgrid models. Un-
fortunately, this means that, if such models are incorporated as submodels,
the grid may have to be rather coarse and, thus, the calculations will lack
detail. The development of submodels for a finer grid scale is an important
area of future research.
In Ref. 5, a model accounting for mass, species, and energy transfer
across the surface of the cluster was formulated. In this model, the drops
move with respect to each other and, thus, expansion and contraction of
the cluster may occur. Thus, in contrast to the models of Refs. 37 and 40,
the drop number density becomes a time-dependent variable, whereas
pressure is taken as constant. A pictorial representation of the situation
studied is shown in Fig. lOa. In this formulation, the gas density inside
each sphere of influence is no longer uniform, but becomes a function of
the radial position in the sphere. The model still keeps the previous basic
features with conservation equations for the drop and gas inside the sphere
of influence, global conservation equations for the entire cluster and the
perfect gas law, but there are two new elements: 1) a model for transport
from the surrounding gas to the cluster, and 2) a model for the drop motion
inside the cluster that pertains to transport from the cluster to the sur-
rounding gas.
Transfer from the surrounding gas to the cluster. Consistent with the
assumption that the cluster size is much larger than the size of the smallest
J. BELLAN 567
1.6 -I——I——I——T-
-3
cm
-2
T°s - 350°K
u° = 200cm / sec
1.5
_ 3
$ n, cm R
2
0.314 6.56x10 4 9.6
94.2 2.2 216.3
R O .04;
R1 = 0. 89; 6ga = 6. 52
1.4
R
1 = °'
R - 0. 95; 9ga = 6. 52
R1 = 0. 89; 0ga = 3. 53
R1 - 0. 94; 6ga - 4. 02
1.3 "'evap '"evap

<|>= 94.2 <|>= 0.314 R
1
0.0 0.0 1.000
0.0128 0.0006 0.999
0.16 0.950
0.01 0.940
0.26 0.16 0.890
1.2 1.0 1.0 0.040
R =0. 999; -6.52
= 0.999;8 ga =
R1=1;ega=6.52
1.1 I
0.0 0.2 0.4 0.6 0.8 1.0
z
Fig. 9 Nondimensional drop temperature vs nondimensional radial
coordinate in a drop for a dense and dilute cluster of drops at several
residual drop sizes. R = 10 cm.
dynamically important eddies, turbulent heat transfer may be described

using a Nusselt number approach. Furthermore, similarity between heat
and mass transfer is assumed and, thus, Shc = Nuc. Because it is very
important to understand how the history of turbulence affects cluster be-
havior, two turbulence models are considered and compared.
In the first turbulence transfer model, the drops do not act initially as
an entity, but rather as individuals, and turbulence builds up with time if
the cluster porosity, i.e., ur, diminishes significantly. In the second tur-
bulence model, the cluster surroundings are assumed to be initially tur-
bulent.
MASS OUT
a) THE "TRAPPING FACTOR" MODEL
MASS
TRANSFER
HEAT
TRANSFER
b) THE SIMILARITY MODEL. Er IS ASSOCIATED WITH THE

ELECTROSTATIC FORCE IF THE DROPS ARE CHARGED
Fig. 10 Pictorial of two different models to describe drop motion inside the cluster:
a) "trapping factor" model; b) similarity model. Er is associated with the electrostatic
force if the drops are charged.
Transfer from the cluster to the surrounding gas. Similar to the preced-
ing, two models are considered here as well. In the first model, transport
is represented using a "trapping factor" model for the gas velocity at the
edge of the sphere of influence45 to calculate the mass loss. The trapping
factor is an interpolation coefficient between a strictly steady gas-phase
limit, where there is strong evaporation and maximum new vapor passes
through the sphere of influence (representing the dense case, mg « md),
and the limit of null loss, where all new vapor is trapped inside the sphere
of influence (representing the dilute case md « mg).
The second model for the drop motion inside the cluster46 introduces a
similarity parameter, ^ = rlR. For a given drop, £ is a constant that varies
in the cluster between zero and unity. Whereas in the previous model drops
and gas were moving with velocities ud and UK, respectively, in the axial
J. BELLAN 569
direction, now there is an additional radial velocity component, respec-

tively, £ dR/dt and £ uge, where uge is the gas velocity at the edge of the
cluster. Thus, the radial slip velocity at the edge of the cluster is ure = dR/
dt — uge. Note that on a large length scale the evaporating drops form a
uniform mass source density, and this is consistent with the similarity as-
sumption. In this model, global momentum equations for radial and axial
velocities are formulated in a manner similar to Bellan and Harstad40;
unlike in Ref. 40, £ now appears in the momentum equations. In general,
the momentum equations are consistent with similarity with the exception
of nonlinear drag and convective derivative terms. To be consistent with
the other global equations where the radial dependence does not appear,
average equations are formed by integrating over £. This eliminates some
coupling between the axial and radial velocities due to convective derivative
terms, and also means that the tendency of nonlinear drag to destroy the
self-similar radial motion is neglected.
Figure 11, reproduced from Ref. 47, shows a comparison among the
predictions of the four models. These models are listed in Table 2. In the
dilute cluster regime, there is excellent agreement among the predictions
of all four models because evaporation of drops in this regime is not con-
trolled by transport processes. In contrast, in the dense cluster regime,
transport processes control evaporation, resulting in substantial quantita-
TURBULENCE CLUSTER
MODEL MODEL
1 "TRAPPING" MODEL •
2 ——— SIMILARITY MODEL •
OPEN SYMBOLS CORRESPOND

TO SATURATION BEFORE
EVAPORATION
— •— UNCERTAIN PART
OF THE CURVE
102
Fig. 11 Drop evaporation time vs c(>0 for four different models describing transport
from and to the cluster of drops.
en
3
O
Table 2 Models whose results are compared in Fig. 11 c
D
Symbol in Fig. 11 Model description D
ID
O
Model A -Turbulence Model 1—assumption (k.l.) in Table 1 TJ
-Drop motion described by the trapping factor model—assumption 1. in Table 1 03
m
Model B -Turbulence Model 1—assumption (k.l.) in Table 1
-Drop motion described by the similarity model—assumption n. in Table 1
Model C -Turbulence Model 2—assumption (k.2.) in Table 1
-Drop motion described by the trapping factor model—assumption 1. in Table 1
Model D -Turbulence Model 2—assumption (k.2.) in Table 1 O
r~
-Drop motion described by the similarity model—assumption n in Table 1
m
J)
c/)
J. BELLAN 571
tive disagreement among the results of the four models. Qualitatively, the
predictions should be similar: tevap reaches a maximum in the dense regime
and decreases as 4> decreases. The cluster is initially denser, and eventually
saturation is obtained before complete evaporation. Thus, validation of
dense cluster or dense spray models requires experimental results precisely
in this large drop number density regime.
The models of Refs. 5 and 47 also predict that, in the absence of internal
vortical motion inside the cluster, dense clusters will first contract due to
internal gas cooling and then expand due to turbulent transport of hot
external gas inside the cluster. The stronger the turbulent transport and
the smaller the cluster, the smaller the contraction and the larger the
recovery toward the original size. In general, it seems that clusters do not
recover their original size by the time the drops have evaporated. Total
fuel-mass loss from clusters turns out to be very small for large clusters of
drops, but increases substantially for small clusters of drops.47 Accurate
prediction of this quantity is crucial for dense clusters of drops because
ignition is expected to occur outside of the cluster48 and it is precisely this
ejected fuel vapor that ignites to establish a flame surrounding the cluster.
Ignition and Combustion

The prediction of droplet ignition in clusters is very challenging from
many points of view. In power generation systems, clusters of drops are
exposed to convective flows that sweep fuel vapor ejected from the cluster
into the wake regime behind the cluster. Thus, ignition will be a result of
the interplay between transport processes from the cluster to the surround-
ings that bring the fuel vapor from the interior to the exterior of the cluster,
convective effects outside of the clusters that sweep the fuel vapor into the
wake regime, and chemical kinetics. This interpretation is in agreement
with the experimental results of Sato et al.49 They infer from their exper-
iments that ignition occurs at the stagnation point of the tip of the fuel
spray. The ignition delay is due to the interplay between two effects de-
scribed by two characteristic times. The first effect is convection, and thus
the characteristic time is that spent for reducing the velocity gradient at
the spray tip below a critical velocity gradient at which ignition occurs.
The second effect is chemical kinetics, and thus the characteristic time is
that for ignition (i.e., chemical runaway) to occur at a given velocity gra-
dient. The conclusion of Sato and coworkers is that, in their particular
experiment, the latter time is much smaller than the former, and thus
chemistry dominates ignition delay times. In other experiments, convection
effects might dominate chemical effects. This picture is very different from
that of the ignition of a single, isolated drop in quiescent surroundings, or
even that of single-droplet ignition in a convective flow.50
It is important to notice that convective effects are important in cluster
ignition. If the cluster is dilute, the effect of convection around each drop
will be important. If the cluster is dense, the convective effects around the
drops located on a shell at the boundary of the cluster will be important
because outer flow penetration is confined to this shell. The denser the
cluster is, the thinner will be the shell.
The experimental observation of thin flames surrounding clusters of

drops2'3 suggests that it may be reasonable to describe ignition and com-
bustion using the classical flame-sheath concept.8 This means that instead
of using a chemical kinetic approach to ignition and combustion, an ignition
criterion combined with a flame-sheath approach might be used instead.
Such an ignition criterion was developed by Bellan and Harstad. 48 To
predict the ignition location, a two-dimensional map is built to compare
convective and diffusive effects. The convective effects are measured by
the extent of flow penetration into the cluster, quantified as the ratio of
the penetration distance to the radius of the cluster. If this ratio is much
smaller than unity, evaporation, and thus ignition, is diffusion controlled.
Therefore, the ignition location can be identified using the criteria devel-
oped by Labowsky and Rosner26 for quasisteady combustion to determine
whether particles burn individually or collectively. Moreover, in this dif-
fusion-controlled regime, the ignition timing is predicted using a Damkoh-
ler number criterion, which compares characteristic diffusion and reaction
times. This Damkohler number criterion is valid only for diffusion-con-
trolled situations, and thus the corresponding ignition criterion is valid only
for nondilute clusters of drops exposed to a moderate to weak convective
flow. When the ratio of the penetration distance to the cluster radius is on
the order of, or much larger than, unity, evaporation is convection-diffusion
and convection controlled, respectively. Since convection effects are always
stronger than diffusion effects, the role of convection must be important
during ignition. It is the author's opinion that a field computation, taking
fnto account both convection and chemical kinetics, is necessary to deter-
mine the ignition time in these situations. Ignition criteria similar to that
of the Damkohler number, but also involving the Reynolds number for a
single, isolated drop, would be irrelevant to cluster ignition, except for
very dilute clusters. Such criteria cannot take into account the very im-
portant transport processes between the cluster and the surroundings. Bel-
lan and Harstad48 show that ignition of nondilute clusters of drops invariably
occurs external to the cluster, and that the boundaries between the dif-
fusive, convective-diffusive, and convective regimes defined earlier are
functions of the fuel and the initial conditions.
Using the preceding model of cluster ignition, Bellan and Harstad51 have
formulated a model of nondilute cluster combustion. In this model, fuel
evaporated from the drops is ejected out of the cluster at a rate determined
by the balance of transport processes across the boundary of the cluster.
After ignition, there is a very short-lived, weak, premixed flame depleting
all oxygen inside the cluster, which is followed by a thin, counterflow
diffusion flame. One crucial aspect of this combustion model is that the
burning rate is not equal to the evaporation rate, in agreement with the
experimental results of Koshland.11 In addition, the flame temperature
depends on the cluster parameters, unlike in the work of Correa and
Sichel.25
Figure 12, reproduced from Ref. 51, shows the fraction of fuel burnt,
/#, at ignition and at the moment of drop disappearance for the weak and
strong turbulence models described earlier vs the initial air/fuel mass ratio.
As experimentally observed by Koshland,11 there is still a substantial amount
J. BELLAN 573
Fig. 12 Burned fuel fraction vs cf>°: O, at ignition; •, at drop disappearance,

turbulence model 1; A, at drop disappearance, turbulence model 2.
of fuel vapor left to burn when the drops disappear. Figure 13 shows the
ratio between the fraction of fuel burnt and the fraction of fuel lost from
the cluster, fBlfph when R = 0 vs 4>° for the same conditions as those of
Fig. 12. When the cluster is initially denser (smaller <j>°), this ratio is smaller
because more fuel is ejected from the cluster, making the flame stand
further from the cluster.
More vigorous fuel evaporation and fuel loss, and shorter evaporation
time, explains why the ratio between the flame radius and the cluster radius,
Rf/R, is larger for turbulence model 2 than for turbulence model 1. This
results in smaller values for fB/fFi when turbulence is strong (turbulence
model 2) than when turbulence is weak (turbulence model 1).
In all calculations carried out with this model, varying parametrically
the initial cluster size, the ambient gas temperature, and the initial drop
temperatures, the flame establishes itself very close to the cluster boundary
and Rf/R is at most 1.01. During burning, Rf/R varies but never exceeds
1.01. This is in complete agreement with the experimental observations of
Alien and Hanson2-3 and Nakabe et al. 4 for "pockets" of drops.
The restricted range of initial air/fuel mass ratios investigated with this
model corresponds to the range of moderately dense clusters. For larger
values of (j>°, the gas phase inside the cluster becomes lean at ignition. The
present model is limited to describing combustion after ignition when the
gas phase inside the cluster is rich. For smaller values of 4>°, the cluster
becomes very dense and ignition does not occur before the drops disappear.
This means that burning will occur only in the presence of the gas phase,
a situation that the present model does not treat properly. In this latter
o
II
DC
Fig. 13 Ratio of burned fraction to fraction of fuel that escaped the cluster, eval-
uated at drop disappearance vs cf>° for two models of turbulent transport to the
cluster. The similarity model was used to describe the drop motion inside the cluster.
situation, turbulence effects and flame wrinkling become dominant (these

phenomena are not modeled here).
Figure 14, reproduced from Ref. 51, shows that smaller clusters of drops
are more efficient at burning fuel, a fact that agrees with intuition.
Evaporation and Dispersion of Electrostatically Charged Drops

As mentioned in the Introduction, soot production is a strong function
of drop proximity, increasing as the drops get closer and trapping the fuel-
rich gas between them where soot formation reactions occur. Calculations
based on a simple model of electrostatic drop dispersion carried out by
Bellan52 indicated that electrostatic drop dispersion might be a viable con-
cept for rendering clusters more dilute and, thus, controlling soot produc-
tion. Recently, a more realistic model of evaporation of electrostatically
charged drops has been formulated by Harstad and Bellan,46 and some of
the results obtained with this model are presented subsequently to illustrate
another aspect where there is a difference of behavior between dense and
dilute clusters of drops.
J. BELLAN 575
0.5
0.4
0.3
0.2
0.1
0.0
0.1 1.0 10.0
ft 0 , cm
Fig. 14 Burned fuel fraction vs R°: O, at ignition; •, at drop disappearance,

turbulence model 1; A, at drop disappearance, turbulence model 2.
The model describing electrostatic dispersion and evaporation is the same

as that of Bellan and Harstad,47 with the addition of the charge effects.
The radial electric field and electrostatic force are taken proportional to
the similarity parameter £ defined earlier. Both momentum and energy
equations take into account these effects, which are averaged by integrating
over £. Figures 15-17, reproduced from Ref. 46, show, respectively, the
evaporation time and the volume ratio (final volume/initial volume), vs
the initial air/fuel mass ratio for various charge ratios, as well as the evap-
oration time vs the charge ratio for a dense cluster of drops. The charge
ratio is the charge divided by the maximum charge feasible for a given
drop size.53 As seen in Fig. 15, only dense clusters are affected by drop
charge. The charge acts to expand the cluster into the hot ambient air, as
seen in Fig. 16, thus promoting evaporation. For dilute clusters, the elec-
trostatic forces are too weak to produce any significant expansion.
Some of the results discussed earlier, together with the results illustrated
in Fig. 17, show that the predicted evaporation time does not depend on
turbulence models for nondense clusters of drops, independently of whether
the drops are charged or uncharged. In contrast, the turbulence models
can affect substantially the predictions for uncharged dense clusters of
drops.5'47 This is due to the restricted available thermal energy for evap-
oration in the relatively small interstitial region of dense clusters, along
with the comparative isolation from the hot ambient gas. For uncharged
drops, only turbulence can break this isolation, whereas for highly charged
I I
CHARGE RATIO
3
0 0 R°=2x10 cm H
1/4 n
1 o
Tgs =350°K
- (TURBULENCE MODEL 1)
C = 1000°K
= 500 cm/sec
12.0
10.0
| 8.0
CO
~d 6.0
4.0
2.0
1.E-01 1.E+00 1.E+01 1.E+02 1.E+03
AIR/FUEL MASS RATIO
Fig. 15 Time for drops to evaporate to 30% of their initial radius vs c|>0 for null
and various finite-charge ratios.
18.0.—
CHARGE RATIO R° = 2x10 3

cm
16.0-
0 0 Ypva-0
14.0- D 1/4
o 1 TgS = 350°K
12.0- (TURBULENCE MODEL 1) Tg°°=1000°K
10.0- u = 500 cm/sec
=10cm
8.0-
6.0-
4.0-
2.0-
0.0
1.E-01 1.E+00 1.E+01 1.E+02 1.E+03
AIR/FUEL MASS RATIO
Fig. 16 Volume ratio at t0 3 vs <J>° for null and various finite-charge ratios.
J. BELLAN 577
15.0
3
R°=2x10 cm
r
Fva = 0
TgS = 350°K
T°°= 1000°K
10.0 g
u° = 500 cm/sec
r
R°=10cm
<t>° =0.471
5.0 0 TURBULENCE MODEL 1

o TURBULENCE MODEL 2
0.00 _L
0.0 0.20 0.40 0.600.80 1.00
CHARGE RATIO
Fig. 17 Time for drops to evaporate to 30% of their initial radius vs the charge
ratio for two turbulence models.
drops, charge-induced expansion dominates. As seen in Fig. 17, for drop

charges greater than one-half of the value proposed by Kelly, 53 the evap-
oration time is independent of the turbulent model. In Figs. 15-17, the
evaporation time is that obtained at R^ = 0.3, which is the value close to
that at which Rayleigh instability occurs. The calculations are stopped at
this point, thus avoiding the modeling of the ensuing drop breakup.
IV. Conclusions
The models and results presented in Sees. II and III underscore the
difficulty of formulating appropriate subgrid models for use in computa-
tionally intensive codes that could describe a variety of combustors. The
difficulty is associated with the many phenomena involved, some of which
are important only in limited regimes. However, since subgrid models must
be reliable over a wide range of physical situations, it is important to
account for all of these phenomena.
The computational and experimental results obtained so far show that
dense and dilute collections of drops behave very differently. For dense
clusters of drops, the numerical results show that the initial turbulence
level of the surrounding flow is crucial in determining the evaporation time
and the dynamics of the cluster size. This is due to several interaction
factors. In the absence of vortical motion inside the cluster, any initial
relative velocity between drops and gas inside the cluster decays very fast
to zero and, thus, relative convection effects primarily enhance evaporation
initially. The internal temperature of the drops increases rapidly, becomes
uniform very fast compared with the drops' lifetime, and continues to
increase during the entire evaporation time. The energy thus transferred
from the gas phase to a relatively large quantity of liquid, particularly in
the absence of turbulent mixing with large volumes of hot gas, cools the
gas very rapidly during this initial period, and the cluster size decreases as
a result. The turbulent transfer of heat and mass from the ambient to the
cluster can influence evaporation significantly. If this transfer is vigorous,
evaporation continues, and the cluster size increases rather than decreases.
In the absence of turbulent mixing, saturation may occur before complete
evaporation.
In contrast, the evaporation of dilute clusters of drops is not controlled
by turbulence. In dilute clusters, the energy transferred from the gas to
the drops has a negligible impact on the total energy content of the gas.
Thus, in the absence of internal vortical motion, the cluster size stays nearly
equal to its initial value. Both the relaxation time of the relative velocity
between drops and gas inside the cluster, and the time taken for internal
drop temperature to become uniform are comparable to the drop lifetime.
The preceding results have direct bearing on the validation of such subgrid
models using comparison with experimental observations. They show that
results obtained only from dilute portions of a spray are inappropriate to
use for comparison because they cannot take into account the importance
of the transport phenomena, which are crucial in the dense regions of the
spray. This is especially important for the subgrid models since, as pointed
out here, appropriate boundary conditions for the microscale of the com-
putation are critical to couple to the macroscale. It should be noticed that
the models presented here are not truly appropriate as subgrid models,
except for calculations where the macroscale of the system is larger than
or comparable to the Taylor macroscale. Modified approaches must be
used to mathematically describe situations where the Taylor microscale
and the cluster size are of the same order of magnitude.
The preceding considerations of spray and cluster ignition have shown
that there is still a considerable uncertainty regarding the description of
this phenomenon, even qualitatively, at the microscale level. The difficulty
arises because in practical systems ignition is always controlled by convec-
tion. For dense clusters of drops, ignition occurs outside the cluster where
convective effects are important. For dilute clusters of drops, ignition might
occur inside the cluster. Since the relative velocity between drops and gas
has a characteristic relaxation time comparable to the lifetime of the drop
in this case, convection effects are important again. Thus, it appears that
it would be necessary to solve differential equations taking into account
both convection and chemical kinetics at the microscale level. Clearly, a
simpler method is desirable, making this an area where new strategies are
definitely needed to deal with the situation. It should be pointed out that
this discussion is relevant not only to the mathematical description of ig-
nition, but also to flame stability.
J. BELLAN 579
Cluster flames cannot be either quasisteady or thin just after ignition.

However, the flame-sheath approximation still seems to be a useful concept
that qualitatively describes flames around clusters, providing one accounts
properly for transport processes to and from the cluster and does not make
the assumption that the burning rate equals the evaporation rate. When
the flame-sheath concept is properly incorporated into a model, this model
yields two important results: 1) the flame sits just at the periphery of the
cluster, and 2) the drops disappear well before the entire fuel vapor is
burnt. Both of these results are in complete agreement with experimental
observations.
Finally, although numerical methods have not been discussed here in
detail, they are a very important component of these two-level (macroscale-
microscale) models and are expected to become even more important once
these subgrid models have to be incorporated into the macroscale calculations.
Most current large-scale codes are based on the assumption that the
volume element does not change in size and that the pressure is variable
within the volume.54 Lagrangian calculations, where the volume does change,
are also performed occasionally, but their implementation in multidimen-
sions requires much more effort. 54 Typically, all large-scale codes solve
nonlinear, coupled mass, species, momentum, and energy equations using
a variety of techniques. In unsteady codes, the solution is obtained by
iterating several times at each time step until convergence of the nonlinear
and coupled terms is obtained.
Since clusters of evaporating drops change their volume during the drops'
lifetime, even in the absence of internal vortical motion, it seems that the
Lagrangian codes are physically much more appropriate to describe the
macroscale in this two-level (macroscale-microscale) description. The La-
grangian description also offers the advantage of the simplest boundary
conditions at the surface of the cluster. These boundary conditions couple
the microscale to the macroscale, and they describe transfer of mass, spe-
cies, momentum, and energy across the surface of the cluster. The density,
composition, temperature, and velocity field of the gas surrounding the
cluster must be provided at each time-step iteration by the solution of the
macroscale equations. The motion of the cluster surface is given by the
solution of the global conservation equations for the cluster. In the Lan-
grangian description, the cluster has an identity because it contains the
same drops throughout the calculation and the volume of the cluster is
defined as the volume contained inside the envelope of these drops.
In contrast, in Eulerian calculation, a fixed volume element cannot rep-
resent the same drops throughout the calculation, and thus the cluster
representation is no longer meaningful. In the Eulerian representation,
drops move between volume elements and, thus, conservation equations
for the number of drops in each volume element are additionally needed.
Moreover, the boundary conditions at the fixed surface of the volume
element have to include new terms representing transport of drops, gaseous
mass, species, momentum, and energy require to keep the volume constant.
Similar to the microscale-macroscale coupling at each time-step iteration
in a Langrangian code, the macroscale Eulerian code must also provide
the density, composition, temperature, and velocity field of the gas sur-
rounding the cluster for coupling with the microscale calculations.
The present calculations5 of cluster evaporation, ignition, and combus-
tion are rather inexpensive. An average of less than 10 s of CPU time is
spent on a UNIVAC 1100 to calculate the entire history of the cluster for
. a given set of initial conditions and fixed values of the density, composition,
temperature, and velocity field in the gas surrounding the cluster. The
model and code are flexible enough to allow the use of more realistic
boundary conditions and to be coupled to a macroscale code. The final
cost of a code based on the two-level description will be determined by
the macroscale code and by the efficiency with which the microscale code
is coupled to the macroscale code.
Acknowledgments
The work reported herein was supported by the U.S. Department of
Energy Utilization Research, Energy Conversion and Utilization Tech-
nologies Program, Marvin Gunn, Jr., Program Manager, and by the Air
Force Office of Scientific Research, Directorate of Aerospace Sciences,
Airbreathing Combustion, Dr. Julian Tishkoff, Program Manager, both
through interagency agreements with NASA. The author would also like
to acknowledge numerous enlightening discussions with Dr. Kenneth G.
Harstad of the Jet Propulsion Laboratory.
References
^higier, N. A., Mao, C. P., and Oechsle, V., "Structure of Air-Assist Atomizer
Spray," CSS/WSS/Combustion Institute Spring Meeting, Paper 7-6A, April 1985;
also, private communication.
2
Allen, M. G. and Hanson, R. K., "Planar Laser-Induced-Fluorescence Moni-
toring of OH in a Spray Flame," Optical Engineering, Vol. 25, No. 12, 1986, pp.
1309-1311.
3
Allen, M. G. and Hanson, R. K., "Digital Imaging of Species Concentration
Fields in Spray Flames," Twenty-First Symposium (International) on Combustion,
4
Nakabe, K., Mizutani, Y., Tanimura, S., and Hirao, T., "Burning Character-
istics of Premixed Sprays in Gas-Liquid Coburning Mixtures," Proceedings of Joint
Conference WSS/JS/The Combustion Institute, The Combustion Institute, Pitts-
burgh, PA, 1987, pp. 227-229; also, Combustion and Flame, Vol. 74, 1988, pp.
39-51.
5
Bellan, J. and Harstad, K., "Turbulence Effects During Evaporation of Drops
in Clusters," International Journal of Heat and Mass Transfer, Vol. 31, No. 8,
1988, pp. 1655-1668.
6
Crowe, C. T., Sharma, M. P., and Stock, D. E., "The Particle-Source-In Cell
(PSI-CELL) Model for Gas-Droplet Flows," Journal of Fluids Engineering, Vol.
99, 1977, pp. 325-332.
7
Tambour, Y., "A Lagrangian Sectional Approach for Simulating Droplet Size
Distribution of Vaporizing Fuel Sprays in a Turbulent Jet," Combustion and Flame,
Vol. 60, 1985, pp. 15-28.
8
Yule, A. J., Ah Seng, C., Felton, P. G., Ungut, A., and Chigier, N. A., "A
Laser Tomographic Investigation of Liquid Fuel Sprays," Eighteenth Symposium
J. BELLAN 581

pp. 1501-1509.
9
Williams, F. A., Combustion Theory, Addison-Wesley, 1965, pp. 37-56.
10
Beer, J. M. and Chigier, N. A., Combustion Aerodynamics, Applied Science
Publishers Limited, London, England, 1972.
n
Koshland, C. P., "Combustion of Monodisperse Hydrocarbon Fuel Droplet
Clouds," High Temperature Gasdynamics Lab., Stanford Univ., Stanford, CA,
Topical Rept. T-248, 1985.
12
Koshland, C. P., "Combustion of Interacting Droplets in a Droplet Cloud
Flame," Western States Section/The Combustion Institute Fall Meeting, Paper 85-
10, 1985.
13
Koshland, C. P. and Bowman, C. T., "Combustion of Monodisperse Droplet
Clouds in a Reactive Environment," Western States Section/The Combustion In-
stitute Fall Meeting, Oct. 1983.
14
Chigier, N. A. and McCreath, C. G., "Combustion of Droplets in Sprays"
Acta Astronautica, Vol. 1, 1976, pp. 687-710.
15
Rao, K. V. and Lefebre, A. M., "Evaporation Characteristics of Kerosene
Sprays Injected into a Flowing Air Stream," Combustion and Flame, Vol. 26, 1976,
pp. 303-309.
16
Sangiovanni, J. J. and Liscinsky, D. S., "Soot Formation Characteristics of
Well-Defined Spray Flames," Proceedings of Twentieth Symposium (International)
17
Chernansky, N. P. and Sarv, H., "Pollutant Formation in Monodisperse Fuel
Spray Combustion," NASA CP 2268, 1982, pp. 39-47.
18
Zung, J. T., "Evaporation Rate and Lifetimes of Clouds and Sprays in Air—
The Cellular Motel "Journal of Chemical Physics, Vol. 46, No. 6, 1967, pp. 2064-
2070.
19
Tishkoff, J. M., "A Model for the Effect of Droplet Interactions on Vapori-
zation," International Journal of Heat and Mass Transfer, Vol. 22, 1979, pp. 1407-
1415.
20
Suzuki, T. and Chiu, H. H., "Multi-droplet Combustion of Liquid Propellants,"
Proceedings of 9th International Symposium on Space Technology and Science,
AGNE Publications, Inc., Tokyo, 1971, pp. 145-154.
21
Chiu, H. H. and Liu, T. M., "Group Combustion of Liquid Droplets," Com-
bustion Science and Technology, Vol. 17, 1977, pp. 127-142.
22
Zhou, X-Q. and Chiu, H. H., "Spray Group Combustion Processes in Air
Breathing Propulsion Combustors," AIAA Paper 83-1323, June 1983.
23
Kim, H. Y. and Chiu, H. H., "Group Combustion of Liquid Fuel Sprays,"
24
Jang, S. D. and Chiu, H. H., "Theory of Renormalized Droplet: II—Non-
steady Vaporization of Droplet in Non-Dilute Sprays," AIAA Paper 88-0639, Jan.
1988.
25
Correa, S. M. and Sichel, M., "The Group Combustion of a Spherical Cloud
of Monodisperse Fuel Droplets," Proceedings of Nineteenth Symposium (Inter-
981-991.
26
Labowsky, M. and Rosner, D. E., Evaporation-Combustion of Fuels, Advances
in Chemistry Series 166, American Chemistry Society, Washington, DC, 1978, pp.
63-79.
27
Labowsky, M., "A Formalism for Calculating the Evaporation Rates of Rapidly
Evaporating Interacting Particles," Combustion Science and Technology, Vol. 18,
1978, pp. 145-151.
28
Labowsky, M., "Burning Rates of Linear Fuel Droplet Arrays," American
Society of Mechanical Engineers, New York, ASME Paper 80-WA/HT-34, 1980.
29
Labowsky, M., "Calculation of the Burning Rates of Interacting Fuel Drop-
lets," Combustion Science and Technology, Vol. 22, 1980, pp. 217-226.
30
Sangiovanni, J. J. and Labowsky, M., "Burning Times of Linear Fuel Droplet
Arrays: A Comparison of Experiment and Theory," Combustion and Flame, Vol.
47, 1982, pp. 15-30.
31
Samson, R., Bedeaux, D., Saxton, M. J., and Deutch, J. M., "A Simple Model
of Fuel Spray Burning I: Random Sprays," Combustion and Flame, Vol. 31, 1978,
pp. 215-221.
32
Umemura, A., "A Unified Theory of Quasi-Steady Droplet Combustion,"
Proceedings of Eighteenth Symposium (International) on Combustion, The Com-
33
Ray, A. K. and Davis, E. J., "Heat and Mass Transfer with Multiple Particle
Interactions. Part I. Droplet Evaporation," Chemical Engineering Communica-
tions, Vol. 6, 1980, pp. 61-79.
34
Marberry, M., Ray, A. K., and Leung, K., "Effect of Multiple Particle Inter-
actions on Burning Droplets," Combustion and Flame, Vol. 57, 1984, pp. 237-
245.
35
Annamalai, D., "Evaporation, Ignition and Combustion of a Cloud of Drop-
lets," The Combustion Institute, Pittsburgh, PA, Paper 3-6B-85, 1985.
36
Shuen, J-S., "Effects of Droplet Interactions on Droplet Transport at Inter-
mediate Reynolds Numbers," AIAA Paper 87-0137, Jan. 1987.
37
Bellan, J. and Cuffel, R., "A Theory of Non-Dilute Spray Evaporation Based
Upon Multiple Drop Interaction," Combustion and Flame, Vol. 51, 1983, pp. 55-
67.
38
Kittel, C., Introduction to Solid State Physics, 3rd ed., Wiley, New York, Dec.
1966.
39
Law, C. K., "Adiabatic Spray Vaporization with Droplet Temperature Tran-
sient," Combustion Science and Technology, Vol. 15, 1977, pp. 65-76.
40
Bellan, J. and Harstad, K., "The Details of the Convective Evaporation of
Dense and Dilute Clusters of Drops," International Journal of Heat and Mass
Transfer, Vol. 30, No. 6, 1987, pp. 1083-1093.
41
Cliffe, K. A. and Lever, D. A., "Isothermal Flow Past a Blowing Sphere,"
International}ournal for Numerical Methods in Fluids, Vol. 5, 1985, pp. 709-725.
42
Yuen, M. C. and Chen, L. W., "On Drag of Evaporating Liquid Droplets,"
43
Eisenkham, P., Arunchalam, S. A., and Weston, J. A., "Evaporation Rates and
Resistance of Burning Drops," Proceedings of Eleventh Symposium (International)
44
Dwyer, H. A. and Sanders, B. R., "Detailed Computation of Unsteady Droplet
Dynamics," Proceedings of the Twentieth Symposium (International) on Combus-
45
Prakash, S. and Sirignano, W. A., "Theory of Convective Droplet Vaporization
with Unsteady Heat Transfer in the Circulating Liquid Phase," International Jour-
nal of Heat and Mass Transfer, Vol. 23, 1980, pp. 253-268.
46
Harstad, K. and Bellan, J., "Electrostatic Dispersion of Drops in Clusters,"
Combustion Science and Technology, Vol. 63, No. 4-6, pp. 169-172.
47
Bellan, J. and Harstad, K., "Transport-Related Phenomena for Clusters of
Drops," International Journal of Heat and Mass Transfer, Vol. 32, No. 10, 1989,
pp. 2000-2002.
48
Bellan, J. and Harstad, K., "Ignition of Nondilute Clusters of Drops in Con-
vective Flows," Combustion Science and Technology, Vol. 53, 1987, pp. 75-87.
49
Sato, J., Konishi, K., Okada, H., and Niioka, T., "Ignition Process of Fuel
Spray Injected in the High Pressure High Temperature Atmosphere," Proceedings
J. BELLAN 583
of Twenty-First Symposium (International) on Combustion, The Combustion In-

50
Dwyer, M. A. and Sanders, B. R., "A Detailed Study of Burning Fuel Droplets,"
Proceedings of Twenty-First Symposium (International) on Combustion, The Com-
51
Bellan, J. and Harstad, K., "Evaporation, Ignition and Combustion of Non-
dilute Clusters of Drops," Combustion and Flame, Vol. 79, 1990, pp. 272-286.
52
Bellan, J., "A New Approach to Soot Control in Diesel Engines by Fuel-Drop
Charging," Combustion and Flame, Vol. 51, 1983, pp. 117-119.
53
Kelly, A. J., "The Electrostatic Atomization of Hydrocarbons," Journal of the
Institute of Energy, 1984, pp. 312-320.
54
Oran, E. S., private communication, 1988.
Chapter 19
Spray Combustion in Idealized Configurations:

Parallel Droplet Streams
R. H. Rangel and W. A. Sirignano
Nomenclature
AT = pre-exponential factors for energy
AY — pre-exponential factors for species
B = transfer number
ct = liquid specific heat
cp = gas specific heat
CD = drag coefficient
CRe = convective correction of vaporization rate
3) = mass diffusivity
E = activation energy
G = Green's function
h = heat-transfer coefficient
hfg = latent heat of vaporization
H = kernel of integral equation
/ = generalized gas-phase variable
JT = total number of droplet streams or droplets
K = parameter defined by Eq. (22a)
KT = total number of fuel species in multicomponent droplet
L = channel half-width
m = droplet mass
m = droplet vaporization rate
M = molecular weight
N = number droplet density
p = pressure
P = parameter defined by Eq. (22b)
PI,... = dimensionless parameters
q = heat flux
Q = heat of combustion
reserved.
585
586 SPRAY COMBUSTION IN IDEALIZED CONFIGURATIONS
r = radial coordinate
R — droplet radius
/?° = gas constant
s = heat-transfer surface area
S = source term
T = temperature
u = velocity
v = velocity vector
x = transverse coordinate
Y = mass fraction
z = streamwise coordinate
a = thermal diffusivity
3 = dimensionless droplet temperature, Eq. (20d)
yn — parameter
8 = delta functions
e = arbitrarily small number
ek = fractional vaporization rate for fuel k
£ = dimensionless streamwise coordinate
6 = dimensionless gas temperature, Eq. (20c)
K = gas thermal conductivity
v = kinematic viscosity
£ = dimensionless streamwise coordinate, Eq. (20a)
p = gas density
X = dimensionless transverse coordinate, Eq. (20b)
Subscripts
chem = chemical reaction
eff = effective
F = fuel
g = gas
/ = inlet
/ = liquid
m = mean
O = oxygen
ref = reference
s = droplet sheet, droplet surface
th = axial injection location
v = vaporization
0 = initial
oc = ambient conditions, £ —> °c
I, II = first stream, second stream
I. Introduction
S PRAY systems are encountered in many engineering applications, in-

cluding liquid-fueled combustors, irrigation, drying, gas absorption, and
fire control. Extensive reviews of the spray combustion literature can be
found in Refs. 1-5. Analytical treatment of the equations governing typical
spray phenomena is usually impossible due to the characteristic nonlin-
R. H. RANGEL AND W. A. SIRIGNANO 587
earities associated, for example, with the convective motion, the coupling
between the phases, and the temperature dependence of the physical prop-
erties. In some cases, the formulation of certain simplifying assumptions
permits an analytical or semianalytical solution of the equations. Such
solutions are sometimes useful because they provide closed-form expres-
sions in which the important parameters and their effects can be easily
studied and understood. The combined insight gained by the use of both
theoretical and numerical results is always more beneficial than the use of
either one alone when explaining existing experimental behaviors.
The work reviewed in this chapter describes a model of heat and mass
transfer, vaporization, and combustion of a spray system consisting of a
number of parallel streams injected in an oxidizing gas flow. Our approach
has been one of gradually relaxing some of the simplifying assumptions in
going from a full analytical solution for a very idealized situation to a
numerical solution for a more complex one. A two-dimensional analysis
of steady channel flow with two streams of heating and vaporizing droplets
was considered first by Rangel and Sirignano. 6 Because the flow velocity
was assumed uniform and constant, an analytical procedure based on Green's
function approach was possible. Since the nonlinear vaporization terms
were retained in the gas-phase energy equations, the final solution, after
a suitable transformation to integro-differential equations, was obtained
by using numerical integration. The liquid-phase equations were also in-
tegrated numerically because of the nonlinearities introduced by the va-
porization process. The model has been extended more recently to include
the chemically reactive case by considering the one-step Arrhenius chemical
reaction of Westbrook and Dryer. 7 In the reactive model, the equations
for conservation of mass for the fuel and oxidizer have the same form as
the gas-phase energy equation, and since the nonlinear chemical reaction
is confined to the source terms, the Green's function approach is still valid. 8
In another work, 9 the high-Reynolds-number droplet heating model of
Tong and Sirignano10 was employed, and density variations due to energy
release by the chemical reaction were taken into account in the flow di-
rection. The resulting flow acceleration and its effect on the droplet velocity
through the droplet drag was thus considered. Transient effects in the gas
phase were later introduced so that the development of the reaction zone
could be investigated. In addition, a second transient effect, more properly
described as an inherent intermittency that results from the discrete char-
acter of the droplets, was described.11 Because of the increased complexity
of the physical model, the solution procedure was based on a finite-dif-
ference scheme. Currently, this model is being used to study the destruction
of hazardous liquid waste.12 A one-dimensional unsteady linear model of
heat transfer in a spray was developed by Sirignano13 by using the Laplace
transform to reduce the gas-phase partial differential equation to an or-
dinary differential equation in the space variable. Other work to be dis-
cussed here considers the two-dimensional heat transfer between a carrier
gas phase and a sheet of droplets without vaporization or combustion. In
this case, an analytical solution that yields closed-form expressions for the
droplet and gas temperatures is possible. The physical situation and the
solution procedure are described in the following sections.
The objective of this chapter is to present the development of the parallel

stream analyses and computations in a compact and cohesive form. Glob-
ally, the work provides insights to the behavior of sprays and to ignition
and flame propagation in fuel-air sprays.
II. Problem Statement
Consider a gaseous mixture flowing into a channel and a number of
droplet streams injected at various locations at the channel inlet. If an
ignition source is provided at some point downstream of the channel inlet,
a reaction zone may be initiated and can propagate through the channel
in a manner to be investigated. We shall focus on a planar, two-dimensional
geometry, and the boundary conditions at the channel walls may be a
constant temperature or a combination of a heat flux and insulated sections.
Once a reaction zone is established, large density variations can exist in
the flowfield, and large transverse velocities can cause the droplets to
deviate from their initial paths. This more elaborate flow requires a fully
numerical solution, such as the one described by Delplanque and Rangel 14
for the effects of thermal expansion on the droplets near a flat wall. By
making a sufficient number of simplifying assumptions, the governing equa-
tions can be made analytically tractable while retaining some of the im-
portant physics. The configuration chosen, illustrated in Fig. 1, is a gas
flow between parallel planes with prescribed boundary conditons. A num-
ber of droplet streams with prescribed initial temperature and size are
injected at the channel inlet. The droplets heat up and vaporize as they
travel downstream under the effect of the gas flow. The following sections
describe a comprehensive model, the analytical solutions for the steady
nonvaporizing, vaporizing, and reactive cases, and finally a numerical so-
lution for the unsteady reactive case. Future developments are suggested
in Sec. IX.
III. Comprehensive Model
Consider a multicomponent spray system such as the one described in
the previous section. The liquid phase consists of a number of droplet
O O O O O o o o o o o o
-© O O O O O O o o
O O O O O O O O O o o
Fig. 1 Schematic of the problem and coordinate system.

streams of prescribed initial composition, temperature, size, and velocity,

injected at the channel inlet. The model assumes spherical droplets, a
quasisteady gas film surrounding the droplets, transient internal heating
and species diffusion, as well as internal circulation in the droplets. The
monocomponent case is described by Abramzon and Sirignano.15 The
equations governing the conservation of energy, species mass fraction, and
mass of each droplet in a given stream are
dT _ l d i 2df
dt
In addition, there are equations governing the droplet trajectory and the
droplet monentum:
f = f f v- 2 ^-»> <4b)
In the preceding equations, the effective thermal and mass diffusivities
take into account the internal circulation of the droplet, so that
aeff - J 1.86 4- 0.86 tanh 2.245 log(^J (5)
and similarly for the mass diffusivity.

The gas-phase equations are those of conservation of mass, momentum,
energy, and species:
^ + V - (pi;) = 2 my*(r - r y ) (6)
x
-- + V • pvv = -Vp - V • T + 2 5my.8(r - r y ) (7)
. cpTsj - ^ emhfg,m 8(r - r y ) + STchem (8)

/=! \ m =l
f V - pvYl - V - pSVY,- = 2 S^-m^r - r,) + 5y/chem (9)
where two types of source terms are included. One type corresponds to
the mass source due to heating and vaporizing / droplets, and the other
corresponds to the chemical reactions. In the vaporization source term,
8/>v = 1 only if species i is present in droplet j. Otherwise, it is zero. In
Eq. (7), the source term SmJ contains the effect of droplet j on the gas
linear momentum. For simplicity, our model considers one-step chemical
reactions of each fuel species with oxygen, although the analytical approach
is independent of this assumption. For the oxygen species, / = O, the
chemical reaction source term is
SyOChem = - i AyofcYak*Y% exp[-£,/(/? () r)] (10)

k= 1
whereas for the kth fuel, the reaction source term is

Ek/(R°T)] (11)
The reaction source term for the energy equation is
5™ = 2 ATYpY% exp[-Ek/(R»T)} (12)

k= 1
The coupling between the gas and liquid phases requires the use of a
film model, which represents the gaseous layer surrounding the droplet.
The droplet and its gaseous layer are represented as a point source in the
gas-phase calculations. An evaluation of the point-source approximation,
and a quantitative measure of the errors involved in it, has been presented
by Rangel and Sirignano.16 Although a better model15 has been suggested
recently, we employ the spherically symmetric film model with a correction
for convective effects suggested by Clift et al. 17 :
m = 4TrRpQ)CRe /„(! + Beff) (13)
CRe = 0.5[1 + (1 + /te)1/3/te°-077], Re
- l (14)
- 0.5[1 + (1 + Re)1'3], Re < 1
M M \ I( K \
V
2j
vmsY
~
V vm~ )1/ / M
2j Y
I L
~
V vms] I
2j Y nsî
\^)
m =1 m—\ / I \
M -]
T
^ V . ,. I
T
Z^ e lfg,m
*-m' f h (16)
m=l J
where ew = rhjrh is the fractional vaporization rate corresponding to

the mth fuel in a multicomponent droplet containing M different fuels.
Other equations needed to complete the film model are the Clausius-
Clapeyron equilibrium relations between the droplet surface temperature
and the fuel mass fractions at the surface.
IV. Nonvaporizing Case: Analytical Solution

A one-dimensional unsteady analysis of heat transfer in a nonvaporizing
spray was conducted by Sirignano.13 In his analysis, Sirignano assumed a
uniform gas and liquid velocity and obtained a solution by means of the
Laplace transform and a series solution obtained by successive substitu-
tions. Here, we consider a two-dimensional steady situation and retain the
uniform-flow assumption. The domain extends from x = - L to L in the
transverse direction with the z axis running downstream and no dependence
on the direction normal to the x-z plane. Only one sheet of droplets, located
at x = 0, is considered. The gas-phase model assumes constant physical
properties and constant pressure, and neglects the momentum exchange
between the gas and liquid phases as well as other viscous effects. This
results in the constant and uniform gas velocity u^ and bypasses the con-
tinuity and momentum equations. By further neglecting radiation heat
transfer and axial conduction of heat, the conservation-of-energy equation
for the gas phase can be written as
pcpug - K = -Nhs*(x)(T - T,) (17)
with boundary conditions T = Tt at z = 0 and dT/dx = 0 at x = -L

and L.
The heat-transfer coefficient h and the heat-transfer surface s are as-
sumed constant and the same for all droplets. The dispersed phase consists
of the planar sheet of droplets of specified droplet density N, injected at
the inlet so that it coincides with the midplane of the channel (x = 0).
The velocity of this sheet of droplets is constant but different from the gas-
phase velocity. In a more realistic situation, the relative velocity between
the gas and the droplets would vary as a result of drag on the droplets and
gas-velocity variations due, for example, to chemical reaction. These effects
are included in a later section. The internal temperature of the droplets is
assumed uniform but time-varying. This limit is not typically achieved in
practice since internal droplet heating can take a large fraction of the
droplet lifetime. Vaporization is ignored in this analysis. The energy equa-
tion for a characteristic droplet in the sheet is
mclU, .= hs(Ts - T,) (18)
where Ts is the gas temperature at the location of the droplet sheet. The
inlet condition for the droplet energy equation is simply 7) = Tn at z = 0.
The source term in Eq. (17) can be replaced by an appropriate boundary
condition at x = 0 if Eq. (17) is integrated across the plane of the droplets
from x = 0~ to 0 + to yield
r^T"
2K — = Nhs(T - 7,) (19)

as the boundary condition to be applied at x = 0. The domain of the

problem is now reduced to half of the channel width, and the new gas-
phase equation is the sourceless version of Eq. (17).
The governing equations can be cast into an appropriate dimensionless
form with the introduction of the following dimensionless variables:
(20a)
X = x/L (20b
e = (T - r,)/(r/; - r,.) (20
P = (r, - r,)/(r, - r,) (20
so that Eqs. (17) and (18) become
--«
(2ib)
The dimensionless boundary conditions are 0 = 0 and (3 = 1 at £ = 0,
a0% = P(6 - 3)/2 at x = 0, and 66% = 0 at x = 1- In Eqs. (21), 9, =
=
0 (x 0) and the dimensionless parameters are
hs L2pcpuK
f - K/L
The parameter K acts as an effective dimensionless diffusivity and is also

the ratio between a characteristic droplet-heating length (used to nondi-
mensionalize z) and a characteristic length defined as the distance travelled
by the gas in a characteristic diffusion time. The parameter P is a dimen-
sionless droplet number density and has the effect of a modified Biot
number in the gas-phase boundary condition. A derived parameter formed
by the combination PK/2 will serve as the expansion parameter in the
solution. It represents the ratio of the convective heat capacity of the liquid
to the convective heat capacity of the gas, i.e. , PK/2 = (NmclUi)/(2LpcpuK).
The solution of Eqs. (21) is obtained in terms of an appropriate Green's
function G(£,x|£',x')? which satisfies the auxiliary problem
subject to the conditions dG/dx — 0 at x = 0 and x — 1. and G = 0 for

£ < £'. This Green's function can be interpreted as the effect at the point
(£,x) of a point source located at (£',x')« Green's function also satisfies the
symmetry condition G(£,xl€',x') = G(-g,x|-€,x')- 1 8
The dimensionless gas temperature 6 is obtained in terms of Green's

function and the boundary condition d0/dx at x = 0 as19
r (24)
The boundary condition at x = 0 is coupled to the liquid-phase solution,

which can be obtained as a function of the gas temperature by integration
of Eq. (21b):
(25)
The gas-phase temperature is obtained by introducing Eq. (25) in the

boundary condition d6/dx = P(0 - (3)/2 at x = 0 and using the result in
Eq. (24). After some manipulations, we obtain
^ //( X ,€lm<ir (26)
where
60 = 0 (27a
V^ -**-n
(27b)
7« - /
cos(mrx) (28)
« = 0 ^n - \
and
The gas-phase temperature at the location of the droplets is obtained by

putting x = 0 in the last equations. Thus,
PK PK [^
e, = 65>0 + — 6 J f l + — J () H(Q£\t')Qs d?' (29)
with Qs^0 = 0.
The complete gas-phase temperature may be obtained explicitly once
the gas temperature at the location of the droplets is known. Equation
(29) is an inhomogeneous Volterra integral equation of the second kind. 20
A solution of this equation that converges for values of the parameter
PK/2 less than or equal to 1 is obtained by a procedure consisting of
successive substitutions by letting
j
f PK\
U,, (30
The first two terms in this series are given by Eqs. (27), and the various
other terms can be successively found as
e,.y. = //(o.sioe,..-, dr, j = 2, 3, ...

The dimensionless droplet temperature is obtained if we let
so that, by combining Eqs. (25), (30), and (32), the various terms in Eq.
(32) are obtained from the recurrence relations
'>6,.od€' (33)
and
3,. = JV«-«'>9,,yd£'> / = 1 , 2 , ... (34)
The dimensionless mean gas temperature is defined by

ri pis I rs \
em(0 = Jo e& x ) d x = — î - <r< - Jo e-«-*'>e, dg'j (35)
In view of Eq. (30), this last expression may be written as
J
y=o
om, . (36)
where
6m,o = 0 (37a)
0 mJ = 1 - e-« (37b)
and
-«-e')e5J--i d ^'' / = 2 ' 3 > -• (38>

The zeroth-, first-, and second-order terms for the gas temperature at
the location of the droplet sheet 65, the mean gas temperature 6 m , and the
droplet temperature (3 are given in Ref. 19.
The form of the series solution obtained in the previous section is such
that the parameter PK/2 appears as the expansion parameter, whereas the
coefficients in the expansion are dependent on the parameter K through
the defined parameter yn. The zeroth-order solution corresponds to the
limit in which the effect of the droplet on the gas is negligible but the effect
of the gas on the droplets is significant. This occurs, for instance, if the
number droplet density N is very low, or, more precisely, when the pa-
rameter P is negligibly small. In such a case, the gas temperature remains
unchanged and equal to its inlet value while the droplet temperature grad-
ually approaches the gas temperature. Figure 2 illustrates the results for
P = 0.1. At this relatively low value of P, the effect of the droplets on
the gas is not very marked unless K is sufficiently large, which occurs if
the heat capacity of the gas is low. As seen in Fig. 2, the gas temperature
rises almost uniformly across the channel. If K is low, the zeroth-order
solution is sufficiently accurate because the temperature change in the gas
is very small. At higher K (Fig. 2), the final gas temperature is somewhat
lower (low gas heat capacity), and the second-order solution is recom-
mended. Figure 3 shows the results for P = 1. At this moderate value of
P, the effect of the droplets on the gas is more significant, and the zeroth-
order solution is generally inaccurate unless K is less than 0.002. For P =
1, the gas temperature profile does not evolve uniformly because the effect
of the droplets is locally strong. For the case of K = 0.1, illustrated in
Fig. 3, and with the gas initially at a higher temperature, the gas temper-
ature at the location of the droplet sheet first decreases and then increases
(the opposite behavior would occur if the liquid were initially hotter than
the gas) due to heating by transverse conduction from the high-temperature
gas away from the liquid sheet once the energy exchange with the droplets
becomes less significant (£ > 1). This effect is not observed when K is
increased to 0.4, because the larger thermal diffusivity implied by a larger
— , — , — , order 0
— — — — order 1
order 2
cx
I
.2
*oo
I
3
0.2 -
0.0
Fig. 2 Droplet temperature, gas temperature at the location of the droplets, and
mean gas temperature for P = 0.1 and K = 4 (zeroth, first, and second order).
-• order 0
— order 1
order 2
mean gas temperature for P = I and K = 0.1 (zeroth, first, and second order).
K distributes the energy more rapidly in the gas. In this last case, the
second-order solution is preferred over the first-order one.
Figure 4 corresponds to the case of P = 10. For this relatively high value
of P, the local energy exchange between the gas and the droplets has a
strong effect on the gas temperature, particularly during the intermediate
stages of the process. For a sufficiently low value of K (e.g., K = 0.002),
the first-order solution is sufficiently accurate. Despite the large effect on
the gas temperature in the exchange region (£ < 10), the final system
temperature is not significantly different from the inlet gas temperature.
At a value of K = 0.01, illustrated in Fig. 4, the effect is even more
pronounced and the total gas temperature change is somewhat higher. In
this case, the second-order solution becomes necessary. At higher values
of K, the second-order solution is inaccurate, and, therefore, those cases
are not considered here.
V. Vaporizing Spray
By still retaining the uniform-flow assumption but allowing for the va-
porization of the droplets, it is possible to obtain a semianalytical solution
for the gas-phase equations and a numerical solution of the liquid-phase
equations. Since the gas-phase differential operator is still linear and the
nonlinearities are confined to the source terms, the Green's function ap-
_,_. —. order0
— — — — order 1
order 2
0.0
log®
mean gas temperature for P = 10 and K = 0.01 (zeroth, first, and second order).
proach is still useful. This case was considered by Rangel and Sirignano, 6
and the main features and results are presented here.
The monocomponent droplets are presented as two sheets parallel to
the boundary planes which are kept at constant, but different, tempera-
tures. No momentum exchange exists between the gas and the liquid, and,
as a consequence, each phase moves with a constant, but different, velocity.
In terms of the following dimensionless variables,
= z/(tvapU,) (39a)
X = x/L (39b)
T = (39c)
the energy and species equations become
^ - Pl < = -s
II
y=i
+ - r;)5,.8'(x - x,) (40)
8YF n 82YF
(41)
where 8'(x - Xy) *s a dimensionless delta function and x/ the transverse

location of droplet stream j.
The liquid-phase equations assuming uniform internal temperature are
d£ C,\ Bef( P3
while the source term is Sv = R'CRe ///(I + Beti). The various parameters
appearing in the preceding equations are Pl = (R2)plU/)/(3L2pUg), P2 =
(NmoUMLpUJ, P3 = cp (Tc - r/0%, and P4 = 2R()(Ug - U,)/v.
Upon introducing the Green's function, expanding in a series solution,
and differentiating once with respect to £, the following set of ordinary
differential equations results6:
n
dT
-77 + T\nT'n = ^ST sin(«7TX,), n = 1, 2, 3, ... (44)
Q
b 7=1
u
AY
(45)
-Qb 7 =^
subject to the conditions T'n = YFn = 0 at ^ = 0 and with r\n = P2{n27i2.

The gas- and liquid-phase ordinary differential equations are solved si-
multaneously by using a Runge-Kutta scheme. Typically, a value of n
between 8 and 10 is required.
Figures 5 and 6 show the development of the fuel mass fraction and
temperature profiles in the gas phase. The two maxima in the mass fraction-
profiles correspond to the location of the droplet streams. These results
correspond to n-heptane droplets initially at 300 K and 150-jjim initial
diameter injected at 1 m/s in a gas traveling at 5 m/s. The initial gas
temperature profile is linear with a maximum of 2050 K at x = 0 and a
minimum of 1000 K at x = 1.
VI. Steady Reactive Case

The steady reactive case was investigated by Rangel and Sirignano8-9 by
still representing the droplets as contained within planar sheets and in-
cluding a one-step Arrhenius-type chemical reaction. Also relaxed here is
the constant droplet velocity assumption by including a drag force on the
droplet. The droplet equations are
^(W) = ° (46a)
t / ; = -S, (46b)
dT' c
'3U', - = (T - T',)(B-fl - B-i)S,, (47)
.2 .3 .4 .5 .6 .7 .8 .9 1.0
NORMflLIZED TRflNSVERSE COOROINflTE
Fig. 5 Gas-phase fuel mass fraction profiles for two vaporizing streams.
2.6
2.4 -
2.2 -
2.0 -
1.8 -
1.6 -
1.4 -
1.2 -
1.0
.1 .2 .3 .4 .5 .6 .7 .8 .9 1.0
N0RHRLIZED TRflNSVERSE CaBRDINflTE
Fig. 6 Gas-phase temperature profiles for two vaporizing streams.

599
(/; ^p - (Cd/l6)(U'g - U'^Re/R'2 (48)
where N} = W//V/0, /*' = /?//?0, and U/Ulo. The initial conditions are
Nl = 1, /?' = 1, 77 = 0, and £//' = 1. The dimensionless gas-phase
equations governing the conservation of species and energy can be written
in a generalized form as
—
3Y — P ! —
n 2 = — ^
V ^2^7°
P V fiYv v "l — ^PS ^Sc1 h e m
VX — Xj) (49">
\^y)
c/c, c/x j =i
where / may be either YF, Y0MF/(v0M0), or cp(Trcf - Tl(})T'/Q. This last

equation is equivalent to the one in the vaporizing, nonreactive case except
for the new combustion term, which is given by
Schem = YaFY^Gxp(-E/R°T) (50)
The method of solution is the same as in the vaporizing case, by means
of Green's function. However, because of the distributed reaction term,
this leads to a set of integro-differential equations in the streamwise variable
with an integration in the transverse equation
d/s + . j = f p ^
d£ " " j =i ~
for n = 0, 1,2, ... and subject to the inlet condition Jn = JQ /() cos
dx' at £ = 0.
Results from a calculation for normal-decane droplets with a 100-(xm
inlet diameter, inlet temperature of 300 K, and inlet velocity of 10 m/s are
taken from Ref. 8. The air flows at 5 m/s, the pressure is 10 atm, and the
inlet gas temperature is linear from 400 K at x = 0 to 1400 K at x = 1-
The liquid streams are located at x — 1/3 and 2/3, respectively. Figure 7
shows the gas temperature profiles at various axial locations, starting from
the channel inlet. The onset of a premixed reaction zone is observed to
the left of the first stream for £ = 0.25. This premixed flame propagates
to the right through the droplet streams, with the temperature peaking
first at the first droplet stream and farther downstream at the second droplet
stream. The contours of isoconcentration of fuel are shown in Fig. 8, where
the locations of the premixed and diffusion-controlled reaction zones are
indicated. Diffusion flames are formed once the droplets penetrate the
premixed reaction zone and the ozygen is depleted at their locations. The
effect of droplet size and fuel volatility is summarized in Table 1. The
smaller droplets vaporize faster, and the premixed flame is established
closer to the inlet. The ignition delay in Table 1 is defined as the down-
stream distance between the points where the premixed flame reaches the
first and second streams divided by the gas velocity.
The assumption of constant gas velocity was relaxed partially by Rangel
and Sirignano9 by accounting for the thermal expansion of the gas in the
axial direction only. The acceleration of the gas as the reaction increases
3.0
0.0
0.0 0.2 0.4 0.6 0.8
NORMALIZED TRANSVERSE COORDINATE X
Fig. 7 Gas-phase temperature profiles for two reacting streams.
Table 1 Droplet vaporization distances and lifetimes, ignition delays,

and flame propagation rates for different fuels and droplet sizes in
steady, parallel-stream configuration
Fuel n-Decane n-Hexane

Inlet diameter, |xm 100 50 100 50
Vaporization distance, cm
I 30 9 20 7
II 38 13 26 9
Droplet lifetime, ms
I 42 14 28 9
II 57 20 37 13
Ignition delay, ms 15 6 11 4
Flame propagation rate, cm/s 13 34 19 47
appears through a continuity equation of the form pUg = const and an

equation of state of the form pT = const. This effect is reflected in a
modifcation of the reaction source term, which becomes
J
chem = (T0/TY exp( - E/R"T) (52)
contour labels
scaled by 10000
I
NORMALIZED TRANSVERSE COORDINATE

Fig. 8 Gas-phase fuel mass fraction contours for two reacting streams.
The ignition process was investigated by using a volumetric integral of

energy obtained from integration of the energy equation across the channel
width:
1 = / dx = /„ +
p^'
JO ,= i
rr r^
The four terms on the right-hand side of this equation are the contri-
(53)
butions to the total gas thermal energy coming from the inlet flow, droplet
heating and vaporization, chemical reaction, and ignition source. The first
term is an additive constant and is irrelevant. Near the channel inlet, where
ignition is expected to occur, the heat transfer to the droplets is at least
an order of magnitude smaller than the other terms. Therefore, a reason-
able global criterion for the ignition location may be defined as the location
for which the contribution from the chemical reaction is equal to that from
the ignition source.
The effect of increasing the number of streams—while keeping the first
one fixed—on the overall ignition is shown in Fig. 9. We start with one
5.0
0 .02 .04 .06 .08 .10 .12 .14 .16 .18 .20 .22 .24 .26 .28 .30
DONNSTREflM DISTRNCE
Fig. 9 Volumetric energy contributions from the ignition source (broken line) and
from the reaction source (solid line) for one, two, four, and eight streams.
stream at one-ninth of the channel width and an overall equivalence ratio

of 0.25. For the next case, a second stream is added at x = 2/9 (overall
equivalence ratio is 0.50). Then, two more streams are added at x = 3/9
and 4/9 (overall equivalence ratio is 1), and, finally, four more streams are
added at x = 5/9, 6/9, 7/9, and 8/9 (overall equivalence ratio of 2). The
overall ignition distance decreases as the number of streams is increased
from one to two and from two to four, but no difference is observed when
increasing further from four to eight. Note, however, that the reaction
energy increases at the same rate up to £ = 0.05 in all cases. The case
with four streams can be viewed as an optimum, because it corresponds
to stoichiometric conditions. Additional increases in the number of streams
result in excess fuel that remains unburned after the oxygen is depleted.
The constant-density assumption in the steady, reactive case has been
relaxed by Delplanque and Rangel14 in a numerical investigation of the
behavior of the streams of droplets injected near a hot wall. The authors
consider a gaseous mixture flowing over a flat plate, and one stream of
droplets injected at a specified location. The ignition source provided at
the wall can be a heat flux or a constant-temperature hot wall. The bound-
ary-layer assumptions are used to simplify the Navier-Stokes equations. A

one-step chemical reaction is assumed. The droplets are represented as
point sources of mass and point sinks of energy. The temperature profile
inside the droplets is transient and nonuniform. Spherical symmetry is
assumed for the droplet solution, and the classical film theory is used to
couple the gas and liquid phases.
The final nondimensional form of the equations is obtained by use of
the continuity equation in the momentum, energy, and species equations:
+ S
dx dy '"'vap (54)
du du 1 d2U ,,„_
PW + P (55)
^ ^*^
3T l 3
^ -2 + 5^p + ST^m (56)
f 5y vap + 5y chcm (5?)

" '-
with
^ « ^^
Sc =
k
The effect of the droplets on the gas momentum is neglected. The non-
dimensional source terms are defined in Table 2, where the parameter
M = pÛJLh is a characteristic mass source.
Since only one droplet stream is considered and the flow is steady, the
parabolic nature of the problem is used in the numerical procedure: the
gas-phase equations are solved while marching downstream with a repre-
sentative droplet. The computational domain is discretized uniformly in
the y direction, but the mesh size in the x direction is determined by the
velocity of the droplet and the time step chosen. The gas-phase equations
are discretized using explicit finite differences. A semi-implicit split of the
reaction term is used for numerical stability. The droplet equations are
advanced explicitly in time, and the gas-phase quantities at the droplet
Table 2 Definition of source terms for gas-phase equations in droplet stream

flow past a hot wall
Vaporization Chemical reaction
Equation c(> ^ Sachem
Mass m M! 0
Energy T ~M~l(T(d) - T}* >)(! + 5-ffO M~ {T-lATk°Ji
Fuels Yf M~l(l - Yf) !-*AYf®.
O2 y0 —M~lYn f
~ l AY $1
locations are interpolated using an area-weighted average from their values

at the two nodes surrounding the droplet. 21 Since the mesh in the y direction
is uniform, the flow near the leading edge is not well resolved. Therefore,
the droplets are not injected at the leading edge of the plate, but after a
certain distance jt th where the boundary layer is well resolved. Between x
— 0 and jc th , the velocity of the gas is used to determine the marching step
A*. The energy equation of the droplet is solved with an implicit scheme
on a uniform mesh. To reduce computational times and for numerical
stability, the number of nodes is reduced as the droplet gets smaller. At
each x location, the equations are solved between y = 0 and y m a x , using
the values computed at the previous axial station *,-_!, and the following
algorithm: 1) the gas-phase species equation [Eq. (57)] is solved, 2) the
temperature field is computed using the energy equation [Eq. (56)] and
the wall temperature is determined if the boundary condition is a heat flux
condition, 3) the gas u velocity is obtained using the momentum equation
[Eq. (55)], 4) the equation of state gives the density, 5) the droplet equa-
tions are solved, and 6) the v component of the gas velocity is calculated
using the continuity equation. The base-case calculation employs 160 nodes
in the y direction and 8000 axial steps are taken in the streamwise direction.
The computational time is 15 min on a Convex C-240 supercomputer or
1.5 h on an Apollo DN-3000 workstation.
Droplets with diameters of 25, 50, 100, and 200 [Jim were tested (cases
6, 7, 1, and 8, respectively). As expected, the smaller the droplet the better
it follows the gas flow. Accordingly, the deviation from the wall is bigger
for smaller droplets. As shown on the velocity history plots of Fig. 10, the
i.oo
•o
c
0.00
0.0 9.90
Time (nd)
Fig. 10 Droplet velocity history: initial diameter effects.
v component is much more affected for smaller droplets. Interestingly

enough, this plot also shows that the v velocity could be monitored to
determine the time of ignition, since at that moment the droplet experiences
a strong acceleration due to the thermal expansion resulting from the
energy released by the chemical reaction. For very small droplets, 50 and
25 |xm, the u component of the velocity takes a positive slope after having
decreased because of the boundary-layer development. This is not the case
for larger droplets. Small droplets are more sensitive to a deflecting v-
velocity component of the gas. Therefore, after ignition, they are driven
more easily toward the boundary-layer edge. Figure 11 shows the stream-
lines, isotherms, and reaction rate contours for the case of 25- jjun droplets,
and Fig. 12 shows the corresponding fuel mass fraction and oxygen mass
fraction contours.
VII. Unsteady Reactive Case
By relaxing the steady-state assumption, we can consider two types of
transient phenomena of interest. The first one is the initial transient period
extending from the cold-flow conditions before the injection of the streams
to the steady-state pattern described in the previous section. The other is
more properly defined as an intermittency associated with the discrete
character of the droplets. Thus, the assumption of axially continuous drop-
let sheets is relaxed. The fluid mechanics are still of zeroth order since the
gas velocity is assumed uniform and transverse convection is neglected.
The gas-phase equations for the case of multicomponent droplets are
J_ ^T d2T N
pe
emhfgtmb(x - Xj)b(z - Zj) + Srchem (58)
,
dt dz Pe\dx2 dz2
J N
2 8;;A-77T~"
= j=i z 8 6 z z 5
UgLp > (* ~ */) ( - y) + Y,-chem (59)
and the liquid-phase equations were given in Sec. III.

In the monocomponent11 and multicomponent 12 cases, these equations
are solved by using a finite-difference technique. The gas-phase equations
are discretized using explicit formulas for the convection and diffusion
terms, and a semi-implicit split of the chemical reaction terms in the species
equations. A uniform grid is used, and the vaporization source terms are
calculated explicitly from the solution at the previous time step. A fine
mesh (of the order of the droplet size) must be used to resolve phenomena
occurring on this scale. The point-source approximation may become in-
valid unless corrections are introduced to correct the error in the vapori-
zation rate.16
Droplet combustion in Boundary Layer flow

1.71 -, -.- —
9.49
0.00
Reaction rate contours

Fig. 11 Streamlines, isotherms, and reaction rate contours for 25-fxm droplets.
The droplet energy and species equations are transformed into equivalent
fixed-boundary problems. In this way, the surface regression is masked by
the transformation while a convective term, easier to treat numerically, is
introduced. A time step typically smaller than the gas-phase time step is
required; thus, a time-split algorithm is used. In the multicomponent case,
due to the large liquid Schmidt numbers, species gradients near the surface
are much larger than the corresponding temperature gradients. In such
cases, a nonuniform grid is used to solve the species and energy equations
in the droplets. The Lagrangian droplet equations are advanced explicitly
Droplet combustion in Boundary Layer flow
0.00
Fuel mass fraction contours
0.00
o.oo ..-.,.--
Healflux x(nd)
Oxygen mass fraction contours

Fig. 12 Fuel and oxygen mass fraction contours for 25-|xm droplets.
in time, with linear interpolation from the neighboring gas points to eval-
uate the gas properties.
In the monocomponent case, a number of parametric studies were per-
formed by Rangel and Sirignano.11 In these calculations, the channel walls
are adiabatic except for a distributed heat source acting on a portion of
the boundary at x = 0. The following results are for 100-|jim-diam decane
droplets injected at 300 K in air at 600 K. Additional details are given in
Ref. 11. Figure 13 shows the ignition delay time (measured from each
droplet's injection time) for the first and second droplets in the first stream
(closest to the ignition source) as a function of the droplet-gas relative
velocity. Results are shown for three different locations of this stream (17.5,
12.5, and 7.5 diameters from the hot wall). A positive droplet-gas relative
velocity means a higher droplet velocity and vice versa. Starting from a
low droplet velocity (negative relative velocity), increasing the droplet
velocity reduces the time before the droplets reach the high-temperature
region. However, if they are injected too fast, they may flow past the hot
zone with too short a residence time for ignition to occur. An optimum
exists that is defined more sharply for the first droplet and for increasing
separation from the hot wall.
To investigate the transition from individual flames to group flames, the
lateral spacing between two streams is varied. Figure 14 shows the reaction
10
~ 6
:1 4
-20 0 20 40 60 80 100
Relative Velocity (cm/s)
Fig. 13 Ignition delay time for the first (1) and second (2) droplets in the first
stream when this stream is located 17.5 (— • —), 12.5 (———), and 7.5 (— —)
droplet diameters away from the hot wall.
rate contours and the droplet locations when the second stream is located
10 diameters away (Fig. 14a) and 5 diameters away (Fig. 14b) from the
first stream. Incipient group combustion is observed for a separation of 10
diameters and group combustion at a separation of 5 diameters. For the
case of 50-jjira-diam droplets, incipient group combustion starts at 15 di-
ameters, whereas droplets of less than 40 jjim in diameter fully vaporize
before or upon crossing the premixed reaction zone.
Recently, Delplanque et al.12 examined the destruction of hazardous
liquid wastes by means of incineration, using the parallel stream configu-
ration. Their analysis focuses on the utilization of conventional auxiliary
fuel as an aid in the burning of aromatic and chlorinated hydrocarbons.
Two strategies are being investigated, one in which the waste is mixed with
the auxiliary fuel prior to injection (multicomponent droplets), and another
in which the waste and auxiliary fuel are injected in separate parallel
streams. The results shown in Fig. 15 indicate that when an auxiliary hexane
stream is used, droplet lifetimes are shorter than when two streams of
benzene are used. The higher volatility of hexane results in the earlier
formation of a reaction zone in the gas phase, thus increasing the heat-
transfer rates to the benzene droplets.
a) contour interval: 8.57E-01 from: O.OOE+00 to: 4.29E+00
57E-01
b) contour interval: 8.48E-01 from: O.OOE+00 to: 4.24E+00
Fig. 14 Reaction rate contours and droplet locations at t — 21 ms when the second
stream is located at a) 10 and b) 5 droplet diameters away from the first stream.
VIII. Conclusions
The heat transfer, vaporization, and combustion of an idealized spray
consisting of a number of parallel liquid droplet streams injected in a gas
flow have been investigated in a number of recent papers. The approach
has been to develop the model in successive stages that gradually incor-
porate more complex effects. The parallel stream configuration is a very
useful one because it provides a benchmark arrangement against which
experimental and computational results can be compared and tested. Cor-
relations obtained in very detailed droplet studies can be tested in an
idealized configuration such as the parallel droplet streams. The model has
been developed initially by making as many simplifying assumptions as
required in order to obtain an analytical solution. Relaxing some assump-
tions to include vaporization and chemical reaction has yielded a combined
355
I
03
CD
Q.
03 0)
E
1
Q_
O
300
time (s) 3.00E-02
Fig. 15 Droplet volume and droplet surface temperature of a second benzene stream
when the auxiliary fuel in the first stream is benzene (— —) and when it is hexane
i______\
analytical-numerical solution, whereas the most current model requires a

fully numerical solution. Exact integration of the governing equations is
achieved by means of a Green's function approach that yields a Volterra
integral equation as the expression for the dimensionless gas temperature
at the location of the sheet of droplets. In the vaporizing and reactive cases,
the Green's function approach is still employed, although the series solution
is not achievable analytically because of the nonlinearlities. In this case,
the integral equations are differentiated once, and the resulting equations
are numerically integrated. The unsteady reactive case requires a finite-
difference solution. The results indicate the existence of complex flame
structures composed of premixed and diffusion flames. Typically, a pre-
mixed flame is established closer to the channel inlet and acts as the ignition
source of droplets crossing it.
IX. Future Challenges and Concluding Remarks

Future improvements of this model should include elimination of the
uniform-flow velocity assumption. This would require solution of the con-
tinuity and momentum equations. In a first approximation, the inviscid
Euler equations could be solved. This would allow consideration of the
thermal dependence of the gas density. The droplet streams would then
deviate from their initial paths. Calculations by Delplanque and Rangel14
for the boundary-layer flow of a droplet-containing gas past a wall show

that the droplets may indeed deviate from their initial paths if the thermal-
expansion effects are considered. These studies are important to determine
the behavior of droplets near a combustor wall. Extension of the parallel-
stream model to account for three dimensionality is straightforward for
the simplified case assuming uniform flow. The variable-density case re-
quires numerical solution of the gas-phase momentum equation. Also of
interest is an extension to an axisymmetric configuration because some
experimental work has been conducted in this geometry. At the high tem-
perature at which the droplets are exposed, thermal radiation can be im-
portant during the droplet heating period and should be addressed in a
future investigation. Thermal-radiation absorption can be nonuniform for
multicomponent droplets and may result in droplet disruption. In dense-
spray situations, direct droplet interaction becomes important and the pos-
sibility of droplet collision leading to droplet shattering or coalescence. In
some investigations, it is necessary to resolve on a scale which is comparable
to the droplet size and, at the same time, cover a region many times larger
than a characteristic droplet. Such situations are encountered when one
wishes to address phenomena that occur on the scale of a few droplet sizes,
e.g., isolated vs group-combustion behavior, 22 ' 24 ignition of a stream of
droplets,8'9 and interaction of a few droplets with a wall. 14 In the numerical
simulation of these problems, the droplets are usually represented by point
sources of mass and point sinks of energy. It has been shown16 that the
point-source approximation must be corrected when the characteristic mesh
size in the gas phase is comparable to the droplet size. The correction
factor required has been determind only for a few configurations, and
further refinements are necesssary.
Acknowledgments
The work reviewed here was supported in part by Grant 860016D from
the Air Force Office of Scientific Research and by Grant DAAG2984K0077
from the Army Research Office. The technical monitors were Dr. Julian
Tishkoff and Dr. David Mann, respectively.
References
!
Faeth, G. M., "Evaporation and Combustion of Sprays," Progress in Energy
and Combustion Science, Vol. 9, 1983, pp. 1-76.
2
Law, C. K., "Recent Advances in Droplet Vaporization and Combustion,"
Progress in Energy and Combustion Science, Vol. 8, 1983, pp. 169-199.
3
Sirignano, W. A., "Fuel Droplet Vaporization and Spray Combustion," Prog-
ress in Energy and Combustion Science, Vol. 9, 1983, pp. 291-322.
4
Sirignano, W. A., "Spray Combustion Simulation," Numerical Simulation of
Combustion Phenomena, edited by R. Glowinski, B. Larrouturou, and R. Temam,
Springer-Verlag, Heidelberg, West Germany, 1985.
5
Sirignano, W. A., "An Integrated Approach to Spray Combustion Model De-
velopment," Combustion Science and Technology, Vol. 58, 1988, pp. 231-252.
6
Rangel, R. H., and Sirignano, W. A., "Rapid Vaporization and Heating of
Two Parallel Fuel Droplet Streams," Twenty-First Symposium (International) on
7
Westbrook, C. K., and Dryer, F. L., "Chemical Kinetic Modeling of Hydro-
carbon Combustion," Progress in Energy and Combustion Science, Vol. 10, 1984,
pp. 1-57.
8
Rangel, R. H., and Sirignano, W. A., "Combustion of Parallel Droplet Streams,"
9
Rangel, R. H., and Sirignano, W. A., "Two-Dimensional Modelling of Flame
Propagation in Fuel Stream Arrangements," Dynamics of Reactive Systems Part
II: Heterogeneous Combustion and Applications, Vol. 113, edited by A. L. Kuhl,
J. R. Bowen, J.-C. Leyer, and A. Borisov, Progress in Aeronautics and Astro-
nautics, AIAA, Washington, DC, 1988, pp. 128-150.
10
Tong, A. Y., and Sirignano, W. A., "Analytical Solution for Diffusion and
Circulation in a Vaporizing Droplet," Nineteenth Symposium (International) on
n
Rangel, R. H., and Sirignano, W. A., "Unsteady Flame Propagation in a Spray
with Transient Droplet Vaporization," Twenty-Second Symposium (International)
12
Delplanque, J.-P., Rangel, R. H., and Sirignano, W. A., "Liquid Waste In-
cineration in a Parallel Stream Configuration: Effect of Auxiliary Fuel," 12th
International Colloquium on Dynamics of Explosions and Reactive Systems, 1989,
(to be published by AIAA).
13
Sirignano, W. A., "Linear Model of Convective Heat Transfer in a Spray,"
Recent Advances in Aerospace Sciences, edited by C. Casci, Plenum, New York,
1985.
14
Delplanque, J.-P., and Rangel, R. H., "Thermal-Expansion Effects on the
Behavior of Vaporizing Fuel Droplet Streams," Paper 89-63, Western States Sec-
tion of the Combustion Institute, The Combustion Institute, Pittsburgh, PA, 1989.
15
Abramzon, B., and Sirignano, W. A., "Droplet Vaporization Model for Spray
Combustion Calculations," International Journal of Heat and Mass Transfer, Vol.
32, 1989, pp. 1605-1618.
16
Rangel, R. H., and Sirignano, W. A., "An Evaluation of the Point Source
Approximation in Spray Calculations," Numerical Heat Transfer, Vol. 16, 1989,
pp. 37-57.
17
Clift, R., Grace, J. R., and Weber, M. E., Bubbles, Drops, and Particles,
18
Ozi§ik, M. N., Heat Conduction, Wiley, New York, 1980.
19
Rangel, R. H., "Heat Transfer Analysis of a Particle-Containing Channel Flow"
(to appear in Warme-und Stoffiibertragung).
20
Morse, P. M., and Feshbach, H., Methods of Theoretical Physics, Vol. 1,
McGraw-Hill, New York, 1953, Chap. 7.
21
Aggarwal, S. K., Fix, G. J., Lee, D. N., and Sirignano, W. A., "Numerical
Optimization Studies of Axisymmetric Unsteady Sprays," Journal of Computational
Physics, Vol. 50, 1983, pp. 101-115.
22
Annamalai, K., Ryan, W., and Caton, J., "Group Combustion: Applications
to Drop Array Studies," Combustion and Flame (in press).
23
Labowsky, M., "A Formalism for Calculating the Evaporation Rate of Rapidly
Evaporating Interacting Particles," Combustion Science and Technology, Vol. 18,
1978, pp. 145-151.
24
Twardus, E. M, and Brzustowski, T. A., Achiwum Process Spalania, Vol. 8,
1977, p. 347.
Chapter 20
Comparisons of Deterministic and Stochastic

Computations of Drop Collisions in Dense Sprays
J. M. Maclnnes and F. V. Bracco
I. Introduction
C OMBUSTION in direct-injection engines is influenced by the dynam-

ics of the sprays that are formed when the jets of liquid fuel encounter
the combustion-chamber gas. The evolution of such sprays, which typically
consist of millions of individual drops, is computed primarily by the sto-
chastic method of Dukowicz1 that was extended by O'Rourke and Bracco2
to account, again stochastically, for drop collisions and coalescences.
There are two main elements of the stochastic collision model: the es-
timate of the mean collision rate within a fixed volume and that of the
actual number of collisions a drop experiences in a given time interval.
The equations for the estimates are derived from the kinetic theory of gases
and are precise only when the drops are distributed uniformly within the
volume, undergo elastic collisions, experience no significant forces between
collisions, and have a volume that is small compared with the volume
containing them. In a spray, such conditions are not always met, and it is
difficult to evaluate accurately the corresponding errors because of the
complexity of the flow. In fact, the model has been evaluated primarily by
comparison of its predictions with measurements in the far field, i.e.,
beyond the region of the spray that is influenced directly and strongly by
drop collisions and coalescences.
In this chapter, the accuracy of the stochastic method of computing
collisions is examined by comparing its results with those obtained with
direct, deterministic calculations of drop trajectories and collisions. The
comparison is made for a practical spray of the sort used in diesel engines
with fine, but still practical, space and time resolutions. However, other
restrictions have been accepted that limit the full extension of our conclu-
sions to practical sprays. These include disregard of drop breakup (see,
for example, Refs. 3-5) and drop vaporization. For the present study, the
Copyright © 1991 by J. M. Maclnnes and F. V. Bracco. Published by the American Institute

of Aeronautics and Astronautics with permission.
615
616 DROP COLLISIONS IN DENSE SPRAYS
neglected effects are not essential but may be contributing. They will have
to be considered in future work aimed at evaluating the overall accuracy
of the spray model in practical applications.
It is shown that the stochastic method underestimates drop collision and
coalescence rates by about a factor of 2 with respect to those computed
deterministically for selected numerical cells in the stochastic calculation.
Most of the difference is due to the stochastic parcel treatment of the spray,
rather than to the stochastic algorithm used to compute drop collisions.
When the stochastic calculation of drop collisions and coalescence is carried
out using exactly the same sequence of drops as for the deterministic
calculation, rather than with the stochastic parcels of drops, the difference
between the two is much smaller. It is also found that near the injector
the stochastic drop parcel technique results in undesirable fluctuations of
the steady mean gas velocity. Far from the injector, insensitivity of the
mean gas and drop velocities (but not necessarily of the drop size) to the
physics and numerics of the region near the injector is confirmed. It is
suggested that future spray computations be made with an orthogonal mesh
that approaches spherical coordinates far from the injector.
In the following sections, the stochastic model is reviewed, and the
stochastic and deterministic methods of computing drop collisions are stated.
Then a specific steady spray and sample numerical cells within it are se-
lected, and the two methods of computing collisions are applied to those
cells. The results of some computations with greater numerical resolution
are then considered, and the main conclusions are summarized.
II. Calculation Approaches

A description of the two calculation methods is now given. First, the
stochastic method is reviewed, including details that pertain both to the
momemtum exchange between the drops and the gas and to drop collision.
The deterministic calculation approach is then described, followed by a
discussion of the method of comparing the stochastic and deterministic
results.
A. Stochastic Method
The stochastic approach to representing spray dynamics originates in the
work of Dukowicz,1 who used a Lagrangian calculation of sample parcels
of drops. The parcels are assigned properties at the inlet according to a
statistical distribution and are allowed to interact through an assumed drag
law with a continuum gas field, which is determined by solution of Eulerian
finite-difference equations. The effect of turbulent dispersion is included
by associating with each parcel a turbulent gas velocity, sampled from a
Gaussian distribution with a selected standard deviation, and allowing the
parcel to be affected by that velocity for a specified length of time. O'Rourke
and Bracco2 and O'Rourke 6 extended that work by placing the method on
a firm theoretical basis and by developing models for drop collisions pro-
ducing coalescence of smaller drops into larger ones.
Gosman and loannides7 used the two-equation k-e model to evaluate
scales of the gas turbulence and also improved the method of O'Rourke
J. M. MACINNES AND F. V. BRACCO 617
and Bracco2 for computing the residence time of a drop in a turbulent

eddy. Finally, the introduction of drops into the domain, corresponding
to the physical process of drops being stripped from the surface of the
injected liquid jet, is accomplished with a version of the line-source method
of Chatwani and Bracco.8 Parcels of drops with properties determined from
aerodynamic instability theory are injected along the jet axis from the
nozzle exit to the end of the intact liquid core. The implementation used
here differs from that reported in Ref. 8 in two ways. First, the size of the
drops in injected parcels is assigned by sampling the empirical size distri-
bution function of Hiroyasu and Kadota 9 instead of injecting each parcel
with the same drop size. Second, rather than introduce all parcels on the
axis, they are distributed uniformly in the radial direction from the axis
out to one nozzle radius.
Since the aim of the present study is the evaluation of the purely dynamic
aspects of sprays associated with collisions, mass and thermal energy ex-
changes between the drops and the gas are not considered and the cor-
responding terms are dropped from the equations. Those interested in the
complete model may consult the references given earlier. The equations
expressing the behavior of the gas field are
Mass continuity:
^ + V • (pgu) = 0 (1)
Momentum conservation:
dp u f f 4 3
—J- + V • (pguu) = - Vp + V • T, - I I p/ - iir !)/ dr dv (2)
Enthalpy balance:
-^ + V - (pghgu) = 6 ( — + u - Vp} + T,: VM
+ V - PgCpDtV(^-} - f f p^TrrX'C"
J
+ u'- v)fdrdv (3)
L \^pg/ J
Turbulence kinetic energy and dissipation rate:
-^- -h V • (pguk) = V-(^—' Vk] + G - p^,(E (4)

dt \ <jk /
+ ~k (CVG - C2p,e) + C3p,eV • u (5)
Here
G = T,: Vii - (2/3)p/V - u (6)
and
T, - pg£>, Vw + Vw T - (2/3)V • w/ (7)
The model relation for the diffusivity due to molecular effects and to
turbulent motions is Dt = D + C^k2/e. The model coefficients are Cl =
1.5, C2 - 1.9, C3 = -1.0, (jk = 1.0, ae = 1.3, and C^ = 0.09.
In these equations, the mean gas velocity is designated by u and the
drop velocity by v\ pg is the mass of gas per volume of space (which when
divided by the fraction of volume occupied by the gas — the "void" fraction,
6 — gives the gas density pg);f(r,v ,x,i) is the number distribution function
of the drops, i.e., the number of drops per unit volume in physical space
and drop property space; and the usual meanings are attached to the other
variables. In the gas momentum and energy equations, the integral terms
on the right-hand side represent the rates of momentum and energy ex-
change between drops and gas.
The equation governing /is the spray equation of Williams10 with terms
added by O'Rourke and Bracco to account for distribution changes due
to drop collisions:
f + V - (fv) + Vv - (fv) = \\ I vab[<rah - 8(r - ra)8(v - va)

Ot L J a Jb
- 8(r - rb)8(v - vb)] dra dva drb dvh (8)
The terms on the left-hand side describe changes in / resulting from un-
steadiness and convective effects in physical and velocity space. The in-
tegral on the right-hand side represents the rate of change of / resulting
from collisions. The integral accounts for binary collisions between all
possible drop pairs. The quantity vab is the collision rate per volume in
physical and phase space for drops with properties (vfl,ra) colliding with
drops with properties (vb,rb). The first term in the brackets represents the
source term due to collisions, being the probability that a collision between
a drop with properties a and one with properties b will produce a drop
with (v,r). The other two terms in the brackets describe the loss of drops
with properties (v,r) due to collisions.
Relations must be given for vab and orab to complete the spray equation.
These relations are central to the present investigation and will be presented
in the next section. Also needed are the equations governing the motion
of a drop having density p /9
v = va - Vp/Pl (9)
with
3 p U + U' - V\ ,
-
———— ———l '
and the equation of state

p = pgRhglcpK (10)
In the drop equation of motion, the empirical relation used for the drag
coefficient is
24/ D 2/3 A - 1 . 7 8 \
Q(e, Re) = ê-2-65 + ——-——J (ii)
where
Re = (2pg\u + u' - v\r)/iLg
As previously stated, the influence of turbulent gas motion is felt through
the drag term. The drop responds to the difference between the gas velocity
(u + u') and the drop velocity v. The turbulent velocity, u', that is assigned
to a given drop is determined by randomly sampling a Gaussian distribution
having zero mean and a standard deviation equal to the local value of the
turbulence velocity scale VIA:. The sampled turbulent velocity remains in
effect for a time corresponding to the lesser of the eddy lifetime, T,, and
the time taken for the drop to cross the turbulent eddy, T/. The characteristic
times are those proposed by Gosman and loannides 7 :
T, - min [T,, T/] (12)
where
rr ——
*t
and T, is determined from the implicit equation

rt + Ti
I l« + « ' - *
where
Collision Model
The collision characteristics that remain to be described are the collision
rate, vab, and the transition probability function, <jab, for a binary collision
between drops of class a and drops of class b. The collision frequency
relation used in the model is derived from the kinetic theory of gases and
is accurate for elastic particles occupying uniformly a negligible volume in
space and traveling in a vacuum. The collision frequency per unit volume
in physical space is
vab dva dra dvb drb (13)
where
V
ab = fafb^(ra + rb)2 Va ~ Vb\
The transition probability function vab is constructed with the view that
the outcome of collisions is either coalescence of the two drops or a collision
in which the two drops do not combine but do exert an influence on each
other. When the drops coalesce, the outcome of the collision is a single
drop having the velocity which preserves the total translational momentum
of the original two drops. The translational energy and angular momentum
with respect to the center of mass of the two drops is assumed to be
transformed into rotation and internal energy of the coalesced drop. When
the rotation energy of the colliding drops exceeds the drop breakup energy,
the model prescribes a "grazing" collision in which the two drops do not
combine, but in which the kinetic energy and angular momentum (in the
mass center coordinate system) are partially dissipated to internal forms
of motion. The degree of dissipation is required to conform to the limits
of complete dissipation when coalescence occurs and to zero dissipation
when the two drops just barely touch one another as they pass. Between
these two limits, a dissipation of angular momentum varying linearly with
the collision separation distance is prescribed. The fraction of dissipated
energy is selected so that the direction of motion of the two grazing drops
is not altered when viewed from the center-of-mass coordinate system. The
<jab function that follows from these considerations is rather complicated
and is not given here (see Ref. 6), since in the numerical scheme the
function is not used directly.
Numerical Implementation
The gas field equations are solved using a first-order pressure-implicit
scheme (the Implicit Continuous-Fluid Eulerian (ICE) method of Harlow
and Amsden11). These equations are coupled through the drag terms to
the drop equations and that coupling is included in the implicit pressure
solution using the scheme proposed by Dukowicz. 1 The spray equation is
solved indirectly using a Monte Carlo technique in which the paths of
stochastic parcels of drops are followed in physical, velocity, and radius
space. Each parcel contains a specific number of drops and is associated
with particular values of velocity, radius, and turbulent eddy parameters.
The parcels in the calculation domain at any moment represent the drop
number distribution. As the number of parcels is increased, the represen-
tation of the number distribution should become more accurate. In the
calculations, the numerical cell volume is used to compute collisions and
momentum exchange terms. In each time step, the drops in the parcels in
each cell are allowed to collide according to a scheme that stochastically
selects which parcel pairs collide and what the outcome of the collision is.
The properties and the number of drops in the colliding pairs of parcels
are adjusted as described in the previous section. The parcels replace the
drops of classes a and b. The mean collision rate between each pair of
parcels coexisting in a given numerical cell is given by
vÂÂivÂ*VVolcell - nav(ra + rb)2\va - vb\/(Volcell/nb) (14)
The ratio on the right-hand side is the rate at which the drops in parcel a
sweep the volume in which a collision with parcel b drops may occur,
divided by the volume in which the center of one drop from b will, on
average, be found. In the numerical scheme, the number of collisions
occurring between drops in parcels a and b in time A/ is assumed to have
a Poisson distribution about the mean number of collisions given by Eq.
(14). The number of collisions occurring is determined by the selection of
a random number uniformly distributed in the interval [0, 1] and allowing
the number of collisions to occur that corresponds to the cumulative Pois-
son-distribution probability equal to the random number.
Once it is determined that a pair of parcels collide, the outcome of the
collision is determined by selecting a second random number that, in effect,
specifies the impact parameter of the collision and, thus, determines whether
the collision results in coalescence of the drops or a grazing collision. The
decision is based on the coalescence parameter
(15)
where
Weab = p,(i>fl - vb)2mm[ra,rh]/v
and
F(y) = y3 - 2Ay2 + 2.1y (16)
2
When the second random number is greater than (b/bcrit) , the energy of
angular motion of the two drops exceeds the surface energy decrease in
forming a single drop, and a grazing collision occurs. Below this value, the
parcel drops coalesce. The function F expresses the effect that the relative
size of the colliding parcel drops has on the collision outcome. The param-
eter y is the ratio of the radii of the larger to the smaller size drops of the
parcel pair and is always greater than unity. From the form of F(y), drops
of greatly unequal size will be more apt to coalesce than drops of nearly
equal size. The collision outcome is also influenced by the collision Weber
number, Weab, which brings in a dependence on relative drop velocity
and absolute drop size. Larger Weah corresponds to reduced likelihood of
coalescence.
B. Deterministic Model
The results obtained with the stochastic method will be compared with
those obtained with a deterministic approach that uses the same algorithm
to determine the outcome of collisions, but where the decision of which
drops collide and when they collide is made deterministically. The com-
putational work is devoted almost entirely to determining the time to the
next collision. This requires calculating the future time of closest proximity
for each of the N(N - l)/2 possible pairs of TV drops in the computational
cell and determining whether the closest approach allows contact. Thus,
the computational time increases as A/ 2 . Reduction of this time is possible
when the trajectories of the drops between collisions are straight, but not
when they are curved. For the conditions of interest, the integration time
step was set equal to the time between successive collisions, and it was
typically less than the integration time used in the stochastic calculations.
The deterministic code has undergone thorough tests, including ones to
ensure that individual collisions are worked out correctly (so that momen-
tum and energy are conserved, and so that none that should occur is
overlooked). In addition, the code has been applied to calculating the
distribution of velocities in an elastic-sphere gas as well as the diffusion
coefficients of momentum and mass transfer. The equilibrium results showed
excellent agreement with the Maxwell-Boltzmann velocity distribution, and
the diffusion coefficients were in good agreement with those determined
from Enskog gas theory.
Correspondence of Deterministic and Stochastic Conditions
The strength of the comparison between the stochastic and deterministic
results depends greatly on how closely the conditions of the deterministic
calculation approach those of the corresponding stochastic calculation. What
is desired in the deterministic computation is a schedule with which drops
of various sizes, velocities, and associated gas turbulence properties are
introduced at the cell boundary. This schedule should reproduce the same
statistical distribution of the parcels entering the cell in the stochastic com-
putation. The method used in this work is based on treating the stochastic
results as statistically steady with the parcels entering a given mesh cell
treated as independent samples from the same, steady drop distribution
function. This view is consistent with the stochastic calculation method,
which solves the distribution function equation using a finite number of
sample parcels.
The task of devising a schedule for introducing drops at the cell boundary
in the deterministic calculation involves deducing the drop distribution
function from the parcels entering that cell in the stochastic calculation.
That is, a representation of
f(x9s)=f(x,vsJf,Tt) (17)
must be made, where s is used to abbreviate the list of properties, excluding
position, that characterize a drop entering the cell. Once/(.x,s) is available,
the number of drops that enter the cell through area AA and have properties
in (s,s -f As) is given by the convection formula
rc(jt,s,As,AO - f(x,s)ksv • AAAr (18)
In this relation, /As is simply the number of drops per unit volume at x
with properties in the interval [s,s + As], and v - AA is the rate at which
volume is passing across the area element at the velocity of the drops.
The stochastic computation results in some number of sample parcels np
that entered a cell over a collection period T. There are two approaches
to representing /(#,s) using the collected parcels. A smooth distribution
function could be fitted through the collected samples, perhaps requiring
the function to satisfy certain moments of the actual samples, or one could
make use of the sample parcels directly, which is equivalent to using a
discrete representation of the distribution function. The first method re-

quires constructing the distribution function and then sampling that func-
tion to arrive at the number of drops and their properties that should be
introduced in each time step in the deterministic calculation. However,
fitting a distribution introduces errors, and so the direct use of the collected
samples was selected.
The ndj drops in each of the np parcels collected over a total time T in
the stochastic calculation are introduced according to the following relation:
/x/*,AO = ndjAt/T9 j = l,np (19)
where ndj is the number of drops in the /th parcel, and HJ is the number of
drops with the properties of parcel j introduced at the cell surface in the
numerical time step Af . This formula may be derived by first constructing
the discrete distribution function having one parcel in each discrete volume
of physical space and phase space and substituting into the number flux
relation, Eq. (18). To demonstrate the relationship, one may substitute
the preceding formula into the number flux relation to arrive at the implied
discrete distribution function:
//*,*,, As,-) = n^sjOj • AA7), j = 1, np (20)
which is clearly the correct form of the discrete distribution function, being
the number of drops with certain properties s,- per property space volume
As,- and per physical space volume vj - A AT. The point at which the drops
are introduced is chosen randomly over the surface according to a uniform
distribution over the projected cell surface area in the direction of the
entering drop velocity. All drops introduced in a given time step are started
at the same time from the cell boundary. The time step is always much
smaller than the time for a drop to cross the numerical cell in order to
resolve the collision and turbulent dispersion processes.
In practice, when the number of parcels in the field is large enough to
provide reasonable statistical resolution, the number of drops from a given
parcel to be introduced in one time step may be much less than one. Thus,
it is necessary to accumulate fractions of drops over consecutive time steps
until a complete drop is accumulated and may be introduced. There is a
further difficulty that must be overcome. If many of the parcels in the
ensemble require a long time for a single drop to be accumulated, very
few drops will be introduced into the cell until sometime after the beginning
of the calculation. Then the introduction rate will rise suddenly. This prob-
lem is avoided in the calculation by randomly assigning to each sample
parcel an initial accumulated drop fraction in the range [0, 1].
The correct operation of the drop introduction scheme of Eq. (19) was
checked by comparing the average properties of the entering drops in the
deterministic calculation with those of the stochastic calculation. The av-
erage rates of entry for the number, mass, and momentum of drops agreed
to within 1% of the stochastic values. Similarly, the average drop velocity,
radius, and turbulence scales associated with incoming drops were the
same.
III. Stochastic Results

Since the investigation takes as its starting point the spray field and gas
field conditions obtained with the stochastic result, it is useful to examine
those results in some detail before considering the deterministic compu-
tations and comparing the two results. Thus, the results of previous sto-
chastic calculations of the spray examined in this chapter are presented,
compared with measurements, and used to select the numerical parameters
for our study. Then two sets of stochastic results are discussed that were
obtained with different treatments of the gas field.
A. Selected Spray and Earlier Comparisons with Experiments
The particular spray considered is that formed by a round jet of liquid
hexane at 300 K issuing at 19.4 cm/ms from a tubular nozzle with diameter
dn - 127 jjum into gaseous nitrogen at 293 K and 4.24 MPa absolute
pressure. This is case C studied by Wu et al., 12 who measured the steady-
state axial and radial drop velocity components at various locations for
x/dn > 300. The high gas density and low temperature make this case
attractive for the purposes of the present study since the dynamic effect
of drop momentum exchange is significant but the complicating process of
drop vaporization is practically eliminated. In addition, since drop densities
will be high in the developing near-field region of the spray, coalescence
of drops will be a strong effect. The intact core length is estimated to be
26 nozzle diameters from dc = 7rf /7 Vp/p^. 13
The stochastic computation method has been demonstrated to give an
adequate account of the drop velocity measurements for xldn > 400. Cal-
culations of this spray are reported in Refs. 8, 14, and 15. The results given
in Ref. 15 are used here to demonstrate the agreement achieved with the
measured drop velocities. This same calculation using an identical com-
putational grid and identical conditions has been repeated as a check in
this chapter.
Various statistical properties of the measured and calculated drop ve-
locities are plotted as functions of radial position in Fig. 1 at axial locations
ofx/dn = 400, which is the measurement location nearest the spray origin,
and at xldn = 500 and 600. The radial profiles of the mean, skewness, and
flatness of the axial component of drop velocity produced in the calculation
(Figs, la, Id, and le) show good agreement with the measurements. The
statistical uncertainty of properties at the edge of the layer is high because
few samples are collected there, both experimentally and computationally.
The calculated skewness and flatness of the radial component of drop
velocity show a similar level of agreement with experiment as for the axial
skewness and flatness shown. The root-mean-squared fluctuation of the
axial drop velocity (Fig. Ib) is somewhat underestimated and that of the
radial drop velocity (Fig. le) is somewhat overestimated. There are no
measurements of the drop size for which usefully small uncertainties have
been demonstrated. The quoted works had also shown that the computed
drop velocity field beyond xldn = 400 did not change significantly when
the numerical resolution was doubled.
Thus, at the beginning of the present study, indications were that the
model reproduces the far-field mean drop velocities of the sprays with
a) i.e b) "•* C) U
'^
Measured Computed Measured Computed Measured Computed
• z/4 = 400 O • x/4 = 400 O • x/4 = 400 °

0.3 ,-= 0.3 ' D D • x/4 = 500 D
Ud
1.2 ^P (^ A x/4 = 600 £
M- 3_
Jr ^W f
B x / 4 = 500D
A x /^ = 600 A
\M?_ ADA D D A x/4 = 600 A
(cm/ms)
d y
'° 0.2 ' n!i[5rf ^ A
*Vt* f 0.2
^ 2&«<no A
**%'***•t
0.6
* *' * £ 0.1 0.1
p^ • *.V^.6H^^ ô
* * •? • A. n .DA^.^J ,° «* « xi
1 2 3
0 0.4 08 1.2 ° '° '° '° 0 1.0 2.0 3.C
V ( cm) > y/y\ y/yi
3
d) -° Measured Computed
** o e) 10.0
Measured Computed
8.0 • x/4 = 400 0 °
• z/4 = 400 0 A
8 2.0 • x/4 = 500°
01 iff • z/4 = 500 D °• • ^ A A
A x/4 = 600 A •
ftF)' 3/2
v d
i.o
A z/4 = 600 A D * »
« /
•
AB * • rff 6
'° D • •.•<
4.0 A A^ ••
Â^BJ^l^j*** • 1 ÔSW^ /Q^rf*^ A
0
^^^ ^ D 2.0 D
-1.0
A . . ,
—————————————————————————————————— —— ——————
. -
< nu ————————————————————————————————————————————————————————
0 1.0 2.0 3.0 o i.o 2.0 3.0

^/2/i y/yi
Fig. 1 Comparison of stochastic calculation and measurements of Ref. 12 in a full-cone spray: a) average axial drop velocity, rms
fluctuation of b) axial and c) radial drop velocity fluctuation, and d) skewness and e) flatness factors of axial drop velocity fluctuation.
ym(x) is the position at which ud is one-half of its centerline value. (Figures from Ref. 15.)
about the same accuracy with which such velocities can, at present, be
measured but that there are uncertainties about the drop size.
The cylindrical grid used in all of the computations quoted earlier has
44 cells in the axial direction, 26 cells in the radial, a cell-to-cell axial
expansion factor of 4% and a radial factor of 7%, and minimum cell size
nearest to the injector nozzle of A_y = 0.5 mm and A* = 1.0 mm. This
grid is referred to here as the XI grid. To reduce the number of drops in
each numerical cell and improve numerical accuracy, the present study
employs the same number of cells and expansion factors but a minimum
cell size of Ay = 0.17 mm and Ax = 0.33 mm. This grid, shown in Fig.
2, is referred to as X3 since its resolution is three times greater than that
of the earlier studies. Note that the XI grid extends to xldn = 850, whereas
the X3 grid extends to x/dn = 130 with a corresponding reduction in the
radial direction. Because the drop velocities remain high in the shortened
X3 grid domain, the time required in the calculation to reach a steady state
and to collect an adequate number of samples thereafter is reduced con-
siderably. This partially offsets the decrease in computation time step size
and the greater computation time per step due to the finer grid and the
increased number of parcels needed to maintain resolution of the distri-
bution function in each grid cell. A total of about 1000 parcels was used
in the XI calculation, and around 10,000 in the X3 calculation.
Results of preliminary calculations were used to identify the cells for
comparisons of the stochastic and deterministic models. A region of the
grid was found that had cells with a high collision rate and an average
number of drops below a few thousand per radian of azimuthal angle. This
region extended from about 30 to 60 nozzle diameters downstream and
radially for about 5dn. Upstream of this region, drop numbers become too
high to allow use of the deterministic calculation. Downstream and further
away from the axis, the collision rates become small. From this region, six
specimen cells were selected. The cells are at three radial positions—
rldn = 0.7 (the cell that adjoins the domain centerline), 2.0, and 4.6—
and at the two axial locations—xld n = 38 and 49. The relation of the cells
y
i en cm
cm —-»
Ax = 0.033 cm » *
k-
0.5'1 cm
Ay = 0.017 cm
1 "=
t 0 50 100 x/dn
Fig. 2 X3 grid used for stochastic calculations herein.

to the spray and to the initial, continuous liquid core of the hexane jet is
illustrated in Fig. 3. The dots mark the locations of parcels at a particular
time under steady-state conditions in the stochastic calculation.
B. Stochastic Results with Calculated Gas Field

The average conditions found in the stochastic calculations are sum-
marized for each specimen cell in Table 1. In the first portion of the table,
the velocities, pressure gradients, and other cell gas properties needed to
make deterministic calculations are given. Because of the staggered grid
used in the numerical scheme, different values of velocity and static pres-
sure gradient apply to parcels in the same cell depending on which of the
staggered momentum cells (two for ug and two for vg) the parcel is in. The
staggered locations of the velocities and pressure gradients indicated in
Table 1 are shown for a typical cell in Fig. 4. Values given in the table are
cell averages corresponding to the time interval from 0.32 to 0.66 ms
following the start of the calculation when gas is at rest and no drops are
in the domain. The uncertainty of the average gas velocities is about ±1%
of their local values. The average pressures have an uncertainty of about
±1% of the local dynamic pressure (pgu - u). Other gas properties are
uncertain by about ±1% of their local values. All of these uncertainties
are based on the root-mean-squares of the values accumulated within the
time window from 0.32 to 0.66 ms. The second part of the table gives
(volume) average properties of the parcels that entered and exited each
specimen cell during the calculation, together with the rates at which par-
cels, drops, and drop mass entered the cell on average. The number of
• '*'" . • • . ' ' • ' • .'. • . . • > • « '

~..^..-~.k^.^îA£L^.''J»*v <'//> ''^(Af-
End of liquid core
0.483 cm
0.619 cm
Fig. 3 Positions of specimen cells in relation to the spray. Spray represented by

positions of all parcels at a particular instant of time at stochastic steady state.
Table 1 Average stochastic results with computed gas field
Specimen cell
1 2 3 4 5 6
Average gas properties in specimen cells
ug , cm/ms 12.3 9.01 2.76 10.6 8.75 3.82
Hg+, ... 11.9 9.00 3.07 10.1 8.61 4.04
v~ , ... 0.0 0.08 -0.09 0.0 0.09 0.02
v/, ... 0.08 0.04 -0.21 0.09 0.09 -0.07
w , ... -0.04 0.03 0.01 -0.03 -0.01 0.0
(Bp/Bx)-,+ g/(cm2-ms2) -0.4 0.5 -0.3 1.6 1.0 0.1
(d/?/dx) , ... 0.8 0.6 -0.3 -0.2 0.8 0.0
(Bp/Br)-, ... 0.0 -0.1 6.3 0.0 -1.9 3.6
(dp/dr) + , ... -0.1 4.8 4.4 -1.9 1.9 3.5
p , g/cm3 0.0449 0.0451 0.0464 0.0453 0.0454 0.0463
e* 0.938 0.957 0.999 0.962 0.960 0.996
V2F3, cm/ms 2.71 2.65 1.67 2.36 2.30 1.79
T,, |JLS 2.9 3.2 5.3 3.8 4.2 5.9
/,, ^m 77 84 88 90 97 106
Average properties of entering droplets
hp, parcels/ JJLS 5.89 4.46 0.96 2.93 3.48 1.33
hd, droplets/jjus 1200 831 162 290 463 172
md, g/s 0.0975 0.1741 0.0067 0.0543 0.1331 0.0223
tTd, cm/ms 12.52 10.80 3.60 11.37 9.90 4.58
^, ... 1.18 1.12 0.55 0.46 0.58 0.35
vv^, ... -0.07 -0.06 0.16 -0.03 0.15 -0.14
SMR, fjim 5.15 5.76 3.54 5.53 5.87 4.71
Td, |xm 1.35 2.69 1.22 2.62 3.46 2.19
Average properties of exiting droplets
«p, parcels/jxs 5.32 4.16 0.94 2.78 3.33 1.30
A^, droplets/ixs 892 670 158 262 432 157
md, g/s 0.1005 0.1727 0.0066 0.0540 0.1337 0.0223
M,/, cm/ms 12.02 10.34 3.55 10.97 9.66 4.83
Wd, ... 1.55 0.60 0.24 0.93 0.55 0.27
H^, ... -0.13 -0.01 -0.05 0.0 0.0 -0.14
SMR, ^m 5.30 5.84 3.68 5.68 5.98 4.78
7^, |xm 1.65 3.03 1.31 2.74 3.65 2.57
parcels collected is the sample population size; this is over 1000 for each
of the four cells nearest the axis and around 500 in the two outer cells.
The pattern observed in the results is that which is expected. The gas
and the drop velocities decrease with both radial and axial distances. The
gas volume fraction, 6, is lowest upstream and near the centerline, where
there is a greater drop density. The Sauter mean radius (SMR, the ratio
y
9 ' \dy
P9,
Fig. 4 Typical specimen cell showing storage locations of gas field variables.
of the average of the drop radius cubed divided by the average of the drop
radius squared) and the number average radius increase in the axial di-
rection because of coalescences. In general, they decrease radially because
small drops tend to be diffused by turbulence more effectively than large
ones.
Some results indicating the relative importance of various processes in
the specimen cells are listed in Table 2. The number of drops per unit
volume decreases both axially and radially. The proportion of collisions
resulting in coalescence—the coalescence fraction—is about 70-80% in
the upstream cells and about 40% in the downstream cells. The cause of
the change appears to be related to the tendency for colliding drops of
similar size to favor grazing over coalescence [which is represented by the
function F(y) in the transition probability model]. Indeed, the standard
deviation in size of drops entering cells 4 and 5 is noticeably lower than
for the other cells. The final quantity in the table is the rate at which drops
change to a new eddy in the calculation. The eddy exchange rate per unit
volume is nearly a constant proportion of the number of drops per unit
volume at each of the two streamwise positions. The exchange rate is,
however, a slightly greater proportion of the number density at the upstream
position compared with the downstream position. This reflects the fact that
the turbulent time and length scales are larger downstream (Table 1).
It is observed that, within the steady jet, the computed steady (mean)
velocity of the gas is, in fact, fluctuating in time and with an amplitude
that is comparable to the turbulence intensity. One can see from Table 3
that the root-mean-squared fluctuations of the "steady" axial and radial
gas velocity components are about 10% of the average axial velocity com-
Table 2 Cell conditions of the stochastic results with computed gas field
Specimen cell
1 2 3 4 5 6
np 13.2 11.2 5.5 7.3 9.4 6.9
«<//Vol cell , 493 136 27 146 81 25
millions/cm3
mj/Vol cell , g/cm3 0.0358 0.0279 0.0008 0.0293 0.0278 0.0027
Volcell, 10 ~ 5 cm3 0.463 1.389 3.241 0.4847 1.454 3.393
SMR, u,m 4.76 5.88 3.46 5.63 6.39 4.69
Collision rate, 108 20.1 0.2 9.7 9.1 1.3
millions/cm3- JJLS
Coalescence rate, 88 12.3 0.2 3.7 3.6 1.1
millions/cm3- (JLS
Grazing rate, 20 7.8 0.0 6.0 5.5 0.2
millions/cm3- JJLS
Eddy exchange rate, 205 57.1 6.4 47.4 31.8 5.1
millions/cm3- JJLS
ponent taken over a time window of 0.34 ms in the steady spray. The steady
void fraction and the turbulence kinetic energy also fluctuate by similar
magnitudes.
The cause of the gas field fluctuations is a variation in time of the number
of drop parcels that are within each numerical cell coupled with the sig-
nificant momentum exchange between drops and the gas. That is, the
roughness of the representation of the drop distribution function and the
strong coupling between gas and drops in the sample cells produces the
gas field fluctuations. The number of parcels in the sample cells is shown
at 0.005-ms intervals in Fig. 5. The large variations are a result of statistical
Table 3 Normalized rms fluctuations of the steady gas properties of the

stochastic results with computed gas field
Specimen cell
1 2 3 4 5 6
VftFT^, % 10.2 8.1 9.3 10.5 8.3 8.1
V^F/^! % 10.2 8.1 9.8 11.8 8.6 8.5
v^r/^", % 3.1 1.8 1.1 3.0 1.9 0.6

VF^, % ' 12.8 5.8 3.7 14.5 6.9 3.2
Vl^/e, % 7.9 4.3 0.5 6.7 4.5 0.7
VF^/ip ~u^\ % 109 100 316 140 111 207
\/77^/</> .^MRn. <% 2.1 0.8 1.6 5.5 1.8 1.5
Celll
10
0
CeU2
10
CD
§
cx Cell3
10
1-4
CD
Cell 4
10 -
0
10
0
Cell 6
10 h
0.4 0.5 0.6

t (ms)
Fig. 5 Time history of the number of drop parcels in the X3 specimen cells for the
computed gas field stochastic calculation, shown at intervals of 0.005 ms.
scatter associated with a relatively small average number of parcels in each

cell, hp, between 5 and 14 in the sample cells (Table 2).
That the fluctuation of the steady mean gas velocity is as strong away
from the axis as on the axis reflects the elliptic character of gas motions
in which a strong disturbance in one cell can propagate through adjustments
in the static pressure field to other locations. The time history of the axial
velocity at the upstream face of cell 2 is plotted in Fig. 6. Indicated on the
figure is the transit time of drops, T trans = kx/ud. The higher frequency
components of the fluctuations appear to scale with the transit time. The
shear-layer time scale 3>i/(wg)CL at the position of cell 2 is about equal to
T
trans- The low-frequency fluctuation is related to acoustic propagation of
disturbances reflected by the boundaries of the computational domain. It
is found16 that the amplitude of the high-frequency fluctuation decreases
to less than 4% of the average axial velocity for xldn > 200. Thus, these
fluctuations do not constitute a significant problem in the far field.
16
12
0.52 0.54 0.56 0.58 0.60 0.62
t (ms)
Fig. 6 Time history of u^ gas velocity in specimen cell 2 for the stochastic calculation
with computed gas field. The average over the entire duration is indicated with the
horizontal line at u^ = 9.0 cm/ms.
C. Stochastic Results with Fixed Gas Field

A large fluctuation of the steady mean gas velocity is highly undesirable
because, according to Eq. (9), it acts as an artificial turbulence. In partic-
ular, such fluctuations complicate the comparison between stochastic and
deterministic methods of evaluating drop collision frequencies because the
drop and gas properties needed in the deterministic computation are av-
erages obtained from the stochastic computation, as explained in Sec. II.
For this reason, we decided to perform a second stochastic computation
of the same drops injected into a fixed gas field. The fixed gas field was
obtained by averaging the gas properties over a time that is long compared
with the transit time from injection to the location of the furthest down-
stream specimen cell. It is realized that the flowfield obtained in this way
is somewhat inaccurate since the gas and drop fields are not from the same
solution of the governing equations. However, the point is irrelevant to
the comparison of drop collisions within numerical cells with given gas
properties. In fact, one can get a qualitative appreciation of the effect of
the fluctuations of the steady gas velocity on the computed near field of
the spray by comparing the results obtained with the fixed gas field with
those of the calculated gas field.
Tables 4 and 5 show the results for the fixed gas field, the entering and
exiting drop properties, and the processes of collision and eddy exchange
for the same specimen cells as before. Included in the table, as for the
calculated gas field results, are the rates of drop mass inflow and outflow
for each cell. It is clear from these mass fluxes that diffusion of drops in
Table 4 Average stochastic results with fixed gas field

Specimen cell
1 2 3 4 5 6
Average gas properties in specimen cells
u~ , cm/ms 12.3 9.04 2.84 10.5 8.753.84
11.9 9.01 3.15 10.3 8.56
4.02
v~, . . . 0.0 0.07 -0.07 0.0 0.080.03
Vg-, . . . 0.07 0.05 -0.18 0.08 0.09 -0.05
n^, . . . 0.04 0.0 0.0 -0.02 -0.06 -0.02
(dpldx) ~ , g/(cm2-ms2) 0.4 0.7 -0.4 -0.8 -0.3-0.1
0.2 0.0 -0.4 1.5 0.3 -0.1
(dp/dr) 0.0 -1.8 6.4 0.0 0.9 3.3
(dp/dr) -1.8 4.0 4.5 0.9 3.2 3.6
pg, g/cm3 0.0487 0.0440 0.0464 0.0506 0.0444 0.0462
e 0.858 0.979 1.0 0.850 0.974 1.0
V2&/3, cm/ms 2.69 2.65 1.69 2.39 2.35 1.80
T r , juts 3.0 3.2 5.3 3.9 4.2 5.9
80 86 89 92 98 107
Average properties of entering droplets
/ip, parcels/ JJLS 5.30 3.09 0.72 2.77 0.72
2.11
/i^, droplets/fjis 676 492 106 373 350 87
«„ g/s 0.1908 0.0830 0.0001 0.1808 0.0964 0.0028
w^, cm/ms 13.67 9.77 3.25 11.85 9.69
3.99
v^, . . . 0.03 0.02 0.52 0.05
-0.23 -0.07
H^, . . . 0.04 0.05 -0.13 -0.17 0.01 0.48
SMR, fjim 5.71 5.28 1.28 6.59 5.69 3.19
rd, (Jim 3.53 2.27 5.38 4.57 3.31
1.03
Average properties of exiting droplets
hp, parcels/ jxs 4.59 2.97 0.67 2.47 0.70
2.05
hd, droplets/ JJLS 578 446 98 310 314 79
md, g/s 0.1894 0.0830 0.0001 0.1783 0.0964 0.0028
ud, cm/ms 13.48 10.40 3.24 11.45 9.95
4.48
IT;, ... 0.25 0.35 0.12 0.10 0.17 0.79
H^, . . . 0.05 0.05 0.03 -0.29 0.07-0.08
SMR, jjim 5.83 5.33 1.23 6.90 5.79 3.24
3.88 2.61 4.65 4.86 3.65 1.19
Table 5 Cell conditions of the stochastic results with fixed gas field
Specimen cell
1 2 3 4 5
np 9.9 7.4 2.8 7.0 6.5 3.1
^/Volcell, 289 85.7 14.4 166 69.1 9.5
millions/cm3
m^/Volcell, g/cm3 0.0993 0.0149 0.0000 0.0914 0.0230 0.0000
SMR, fjim 5.98 5.66 1.49 6.84 6.09 1.75
Collision rate, 74.2 5.6 0.0 54.7 6.2 0.0
millions/cm3- JJLS
Coalescence rate, 34.3 3.5 0.0 12.1 2.8 0.0
millions/cm3- JJLS
Grazing rate, 39.9 2.1 0.0 42.6 3.4 0.0
millions/cm3- JJLS
Eddy exchange rate, 136 35.3 3.1 73.5 22.9 1.8
millions/cm3- (JLS
the fixed gas field case is reduced significantly from that in the calculated
gas field case. The axial cells show a strong increase in drop mass-flow rate
while the mass flow rate into the outer cells is reduced—and nearly zero
in the two cells furthest from the axis. Although drop mass-flow rates into
the cells on the axis increase, the number of drops in the cells decreases.
The increased average drop size reveals that the decreased drop numbers
in the cells on the axis must be largely the result of an increased coalescence
rate upstream of each cell, whereas away from the axis the decrease in
drop size indicates that reduced diffusion is the cause. Again, large fluc-
tuations in the number of parcels in a cell occur even though the gas field
is fixed. This confirms that the fluctuations are due primarily to the poor
resolution of the drop distribution function.
Thus, it has been established that gas fluctuations, which are set up by
the rough discretization of the distribution of drop properties—in partic-
ular, the drop number density—as represented by the parcels, cause con-
siderable augmentation of drop diffusion in the near field of the computed
spray. The drop number distribution is too crudely resolved with the num-
ber of parcels used in the calculations, even though the number of parcels
is greater than is practical in current spray applications.
IV. Deterministic Results

The deterministic calculations were performed using the spray and gas
properties obtained with both the calculated gas and the fixed gas fields.
In examining the deterministic results, the focus is on differences in what
occurs within the cells.
Most of the details of the procedure used to make the deterministic
calculations have been provided in Sec. II.B. Some additional aspects of
the calculation remain to be discussed. The calculation domain used in the
stochastic computation extends for 1 rad in the azimuthal direction. This
has been reduced to 0.2 rad since the larger domain increases computation
time considerably as a result of the proportionally larger number of drops
involved and is judged to be unnecessary. For each cell, the calculation is
made for a length of time equal to 40 T trans , which has been found to allow
for transient adjustment from the initial conditions and a suitably long
averaging time to form accurate statistical quantities. The transient time
was typically only a few transit times. Still, averages were performed only
over the final 20 T trans of the calculations. The number of drops entering
the cell and the number of collisions occurring during this time are, typi-
cally, of the order of thousands.
A. Comparison of Deterministic and Stochastic Results

Summaries of the main results of the deterministic and stochastic cal-
culations are provided in Tables 6 and 7 for the calculated gas field and
the fixed gas field, respectively. Starting with the calculated gas field results
in Table 6, the main difference between the stochastic and deterministic
calculations is in the collision rate. The deterministic calculation produces
a collision rate roughly 50% greater than the stochastic calculation.
A secondary effect is a reduction in the number density in the deter-
ministic calculation. The increase in the collision rate and, hence, in the
coalescence rate results in fewer drops of larger size. The larger drops
experience a smaller drag force in relation to their inertia and, therefore,
move more rapidly through the cell, which further reduces the number
density. In addition, as a result of the greater coalescence rate in the
deterministic calculation, the drop size is more uniform, which produces
a tendency toward lower coalescence fractions in the deterministic results.
The cause of the larger collision rate in the deterministic calculation was
thought perhaps to be the fluctuation of the steady gas velocity. In the
stochastic calculation, all parcels in a cell at any one moment are influenced
by a common mean gas field fluctuation, and, therefore, the fluctuation
does not contribute significantly to the relative velocity between drops,
which influences the collision rate. The drops that are collected from the
stochastic calculation to be injected into the cell in the deterministic cal-
culation, though, come from different parcels from different times and thus
exhibit a broader drop velocity distribution function as a result of the gas
velocity fluctuation. Larger drop relative velocity and, consequently, greater
drop collision rates would result from the broader distribution. In the end,
however, the deterministic collision rate is about 50% greater than the
stochastic one with both computed and fixed gas fields. Thus, the gas
fluctuation itself is not the dominant cause of the increased collision rates.
One reason for the underestimate of the collision rate in the stochastic
method follows. In the stochastic calculation, drops are assumed to be
distributed evenly throughout the numerical cell. In the deterministic cal-
culation, the drop number density is found to be significantly larger in the
upstream portion of a cell and then to decrease due to collisions and
coalescences. Since the collision frequency is proportional to the square
of the number density [Eq. (14)], the deterministic computations will give
greater collision frequencies when the number density changes significantly
Table 6 Comparison of deterministic and stochastic results with

calculated gas field
Specimen cell
1 2 3 4 5 6
rtrf/Vol ceH , millions/cm3
Deterministic 388 118 28 127 80 24
Stochastic 493 136 27 145 81 25
Collision rate, millions/cm3-|jis
Deterministic 156 34.3 0.4 28. 6 17.2 0.8
Stochastic 108 20.1 0.2 9. 7 9.1 1.3
Coalescence fraction,
coalescences/collision
Deterministic 0.72 0.51 0.75 0. 52 0.410.50
Stochastic 0.81 0.61 1.00 0.38 0.390.85
Eddy exchange rate, millions/(cm3-|ULs)
Deterministic 174 53.4 6.0 42. 8 27.0 5.0
Stochastic 205 57.1 6.4 47. 4 31.8 5.1
Table 7 Comparison of deterministic and stochastic results with fixed gas field
Specimen cell
1 2 3 4 5 6
3
7^/VolceU, millions/cm
Deterministic 227 81.0 17.7 158 63.3 12.7
Stochastic 289 85.7 14.4 166 69.1 9.5
Collision rate,
millions/(cm3-)ULs)
Deterministic 148 14.1 0.0 89.8 9.6 0.0
Stochastic 74.3 5.6 0.0 54.7 6.2 0.0
Deterministic 0.380.47 — 0.27 0.39 —
Stochastic 0.460.62 — 0.22 0.45 —
Eddy exchange rate,
millions/(cm3-|ULs)
Deterministic 100 30.6 3.2 62.1 23.0 2.6
Stochastic 136 35.3 3.1 73.5 22.9 1.8
within a numerical cell. As an alternative way of expressing the same

concept, it could be argued that the distance for significant change of the
drop distribution function is the drop mean free path relative to the grid,
since approximately every other collision yields a coalescence. Therefore,
the spatial discretization should be no greater than the drop mean free
path.
A further difference is associated with the resolution of the drop property

space by the stochastic parcels. In the limit in which only a single parcel
of drops is present in a cell, no collisions are computed even though in
reality the drops of the parcel would undergo some collisions. Typically,
spray calculations use a number of parcels that result in around 10 parcels
in cells where collisions are significant. The present stochastic calculations
result in an average of 3-13 parcels in the specimen cells (Tables 2 and
5). It is likely for this low number of parcels to result in underprediction
of the collision rate. Resolving this issue by increasing the number of parcels
in the calculation does not appear to be a viable approach. Already, the
number of parcels used in the fine grid calculations stretches available
computer resources—there are over 10,000 parcels in the domain with the
X3 grid. Notice that it is not possible to allow for collisions of the drops
within a parcel because such drops have the same velocity and, therefore,
their relative velocity and collision frequency are zero. It is possible to
alter the number of collisions (and to improve the resolution of the drop
distribution function) by splitting parcels, but the total number of parcels
within the domain is restricted by practical considerations.
B. Higher Resolution Computations

Since the reason for the difference between the stochastic and deter-
ministic results is mainly inadequate numerical resolution of the stochastic
computations, higher resolution calculations were made. As explained in
this section, they provided some additional insight, but also presented some
new difficulties.
Table 8 shows the resolution of two additional computations, the X6
and X12 cases. Again, the total number of cells was kept constant so that
the higher resolution cases cover smaller axial and radial domains. For the
XI resolution, the radial extent of the cell was 7.94 times larger than the
liquid core radius and use of a line-source technique was justified. For the
X12 computation, the ratio is 0.65; i.e., the first row of cells is within the
liquid core. Moreover, the spray model, constructed for dense sprays, still
requires the liquid volume fraction to be no more than 10% (6 > 0.90) so
that the average surface-to-surface interdrop distance could be at least
equal to the drop diameter; otherwise, the assumption of isolated spherical
drops is most likely incorrect. It is seen that conditions in the X6 and X12
computations violate the limit on 6. Thus, the X6 and X12 computations
push the model beyond its limits. Even when 6 > 0.9, breakup of drops
has been disregarded, and yet, according to the criterion We > 6, where
We = pg(u + u' — p) 2 r/<j, 5 a significant fraction of the drops in the domain
would undergo secondary breakup.
An additional problem is due to the cylindrical grid that seems to be
used routinely for spray computations and yet is not the natural grid for
the flowfield of a jet. In the far field, the angle of the jet becomes constant
and a spherical grid would maintain the same number of grid cells within
the jet boundary layer at all axial locations. A cylindrical grid gives the
optimal number of grid points at one axial location, but too few points
upstream of that location and too many downstream. Conversely, very
near the nozzle, the jet broadens slowly, and nearly cylindrical coordinates
Table 8 Features of the stochastic computations with calculated gas field

and for four grid sizes
Number of
A*min Ay min Drop mass drops
Grid dJ2 dJ2 emin (x/dn) with We > 6, % with We > 6, %
XI 15.9 7.94 0.96 (50) 97 12
X3 5.29 2.64 0.86 (20) 98 17
X6 2.64 1.32 0.72 (12) 96 3
X12 1.32 0.66 0.64 (7) 92 11
may be best. The optimal grid would maintain about the same number of
points within the boundary layer at all distances from the injector and
would tend to a spherical grid whose center is the virtual origin far from
the injector. Such a grid can be achieved with arbitrary orthogonal coor-
dinates. The present cylindrical grid brings about several interesting an-
omalies. As the resolution is increased, the computed near field changes
drastically, whereas the computed far-field velocity is scarcely altered. In
the near field, where the boundary layer is not resolved by the XI grid,
the increased resolution reduces various forms of numerical diffusion and
results in a narrower, faster, more penetrating spray. In the far field, where
the resolution of the velocity boundary layer was already adequate, the
higher resolution simply pushes downstream the virtual origin of the fully
developed jet, but the position of the virtual origin is irrelevant at a suf-
ficient distance from the nozzle. Thus, we can explain both the reported
insensitivity to numerical resolution of the computed far-field velocity that
was mentioned at the beginning of this chapter and the strong sensitivity
of the near-field results.
For all of the preceding reasons, the results of the X6 and X12 calcu-
lations are considered just computational experiments as far as the structure
of the near field is concerned. However, they do provide progressively
smaller numerical cells within which the gas and spray properties are known.
The local comparison between the stochastic and deterministic approaches
to computing drop collisions and coalescence within selected numerical
cells is still valid for cells with 0 > 0.9.
Thus, deterministic calculations for six specimen cells in the X12 grid,
taken here from the same region of the domain as the X3 specimen cells
(see Fig. 7), were made. For the cells of the X12 grid, AJC ~ 80 jxm,
Ay « 40 jxm, 2.1 < Td (jxm) < 3.8, and 0 > 0.93. Comparison of the
deterministic and stochastic results in these cells is shown in Table 9. Again,
the collision rate is lower in the stochastic computation than in the deter-
ministic one.
At this point, we must remind the reader that we have considered a very
severe test of the stochastic method of computing drop collisions and co-
alescences. Let us reconsider how the comparison between stochastic and
deterministic computations was set up. The stochastic computation is per-
formed. Stochastic cell drop collisions and coalescences are counted, and
contributions to / are accumulated over a time that is much longer than
Fig. 7 Relation of X12 grid specimen cells, a, b, c, d, e, and f to the X3 grid

specimen cells.
the drop cell transit time. By sampling the accumulated distribution func-
tion data, individual drops are introduced into the cell, and the determin-
istic computation of cell drop collisions and coalescences is performed.
Notice that the stochastic and deterministic computations of the collisions
and coalescences within the cell do not start from exactly the same drops
entering the cell; the stochastic collision computations use the parcels of
drops that enter the cell, whereas the deterministic collision computations
use individual drops sampled from the entering drop parcels. When the
same sequence of individual drops is introduced into the cell for both the
stochastic and the deterministic evaluation of the cell drop collision rate,
the difference in the two rates is greatly reduced, as can be seen from
Table 10. In the case of the X3 grid, the stochastic collision rates are no
longer systematically lower than deterministic ones and differ by 10% at
most. The stochastic collision rates do remain consistently lower for the
X12 grid but now by only about 20%. This second method of comparison
is fairer to the stochastic collision algorithm itself, but less representative
Table 9 Comparison of deterministic and stochastic results; X12 grid for

calculated gas field
X12 grid cell
«d/Vol cen , millions/cm3

Deterministic 74 117 90,.4 55,.7 78 .7 60 .4
Stochastic 170 152 95,.0 104 104 73,.0
Collision rate, millions/(cm3-|ms)
Deterministic 256 90.7 15,.7 63,.5 17..4 7,.8
Stochastic 52.0 35.3 11.8 18,.0 11 .9 6,.1
Deterministic 0.39 0.42 0.,55 0.,51 0,.61 0,.71
Stochastic 0.23 0.29 0.,61 0.,30 0..60 0.,67
Table 10 Comparison of deterministic and stochastic results when sequence of

entering drops is identical
X3 grid cell
1 2 3 4 5 6
3
Collision rate, millions/cm - [xs
Deterministic 157 36 .3 0.400 33.2 15 .8 0 .630
Stochastic 174 36 .2 0.396 36.0 15 .6 0,.612
Deterministic 0.713 0 .525 0.640 0.496 0 .431 0 .509
Stochastic 0.708 0 .503 0.639 0.480 0 .418 0,.503
X12 grid cell
Collision rate, millions/(cm3-|jLs)

Deterministic 98.6 71.3 14.3 24.8 13.8 7.70
Stochastic 73.1 50.0 11.4 17.1 10.8 6.3
Deterministic 0.328 0.403 0.495 0.467 0.549 0.653
Stochastic 0.235 0.342 0.499 0.381 0.572 0.685
of the accuracy of the stochastic computation of collisions within a sto-

chastic spray computation.
Other factors expected to influence the difference between the stochastic
and deterministic collision rates are drop breakup and drop vaporization,
but they have not been evaluated in this study.
V. Conclusions
Comparisons of stochastic and deterministic computations of drop col-
lision and coalescence within selected numerical cells of a self-consistent
multiphase spray-flow computation were presented. The numerical reso-
lution was greater than that currently used in practical spray computations.
The selected spray is narrow, fast, dense, and of the type found in diesel
and stratified-charge engines. However, the spray computed was steady,
and drop vaporization and breakup were not considered.
The stochastic method was found to underestimate drop collision and
coalescence rates by about a factor of 2 with respect to those computed
deterministically. A significant fraction of the difference is not due spe-
cifically to the stochastic technique of computing drop collision frequencies
but originates in the stochastic parcel treatment of the spray. In addition,
it is expected that the difference will decrease when drop breakup and
vaporization are considered, because both effects tend to narrow the drop
size and velocity distribution function, which, in turn, reduces both collision
and coalescence frequencies. Indeed, one origin of the difference is that
a large fraction of the drops that enter a numerical cell undergo collisions
and coalescences within the cell, thus causing significnt changes in the drop
number density. The stochastic technique assumes a uniform distribution
of drops within a cell, whereas the deterministic technique computes the
spatial variation of that distribution. The obvious remedy is to increase the
numerical resolution, but since the drop distribution function changes sig-
nificantly within distances of the order of the drop mean free path, im-
practically small grids may be necessary for all numerical techniques. Actually,
the reported work has progressed only to the point of identifying some
inaccuracies of the stochastic method of computing drop collisions and
coalescences. Further work is needed to determine the specific sources of
the inaccuracies and the conditions necessary to avoid them.
The present work has focused on the near field, where drop collisions
and coalescence are among the controlling processes. Several significant
difficulties were encountered. The cylindrical grid that was used, and that
is common for spray computations, is not the natural one; future work
should adopt a general orthogonal grid that tends to spherical coordinates
in the far field. Fluctuations, found in the steady gas velocity, are due to
the stochastic drop parcel technique. These fluctuations are comparable
in magnitude to the gas turbulence and act as artificial gas turbulence.
When progressively higher resolution was used, the physical accuracy of
the results became determined by physical processes that are not included
in the model.
For the far field, i.e., for xldn > 300 to 400, 12 - 14J7 certain more favorable
conclusions have been confirmed. The computed far-field mean drop and
gas velocities are largely independent of the near-field model and resolu-
tion. The model and resolution affect the computed location of the virtual
origin of the fully developed jet, but this location is immaterial at a suf-
ficient distance from the nozzle. The artificial fluctuations of the computed
steady mean gas velocity become of the order of 10% of the gas turbulence
and, therefore, are not of dominant importance. However, there remain
questions as to the sensitivity of the computed far-field drop size to the
near-field physical and numerical inaccuracies.
Even though the accuracy of far-field computations approaches accept-
ability, those interested in diesel and stratified-charge engines can derive
little comfort from it because in such applications it is the near field that
counts.
References
ûkowicz, J. K., "A Particle-Fluid Numerical Model for Liquid Sprays," Jour-
nal of Computational Physics, Vol. 35, 1980, pp. 229-253.
2
O'Rourke, P. J., and Bracco, F. V., "Modelling of Drop Interactions in Thick
Sprays and a Comparison with Experiments," Institute of Mechanical Engineers,
Pub. 1980-9, London, England, 1980, pp. 101-116.
3
Reitz, R. D., and Diwakar, R., "Effect of Drop Breakup on Fuel Sprays,"
Society of Automotive Engineers, Warrendale, PA, SAE Paper 860469, 1986.
4
O'Rourke, P. J., and Amsden, A. A., "The Tab Method for Numerical Cal-
culation of Spray Droplet Breakup," Society of Automotive Engineers, Warren-
dale, PA, SAE Paper 872089, 1987.
5
Reitz, R. D., "Modeling Atomization Processes in High-Pressure Vaporizing
Sprays," Atomisation and Spray Technology, Vol. 3, 1988, pp. 309-337.
6
O'Rourke, P. J., "Collective Drop Effects on Vaporizing Liquid Sprays," Ph.D
Thesis 1532, Princeton Univ., Princeton, NJ, 1981.
7
Gosman, A. D., and loannides, E., "Aspects of Computer Simulation of Liquid-
Fueled Combustors," AIAA Paper 81-0323, 1981.
8
Chatwani, A. U., and Bracco, F. V., "Computation of Dense Spray Jets,"
Proceedings of the International Conference on Liquid Atomization and Spray Sys-
tems (ICLASS 85), Paper IB/1/1, Institute of Energy, London, England, 1985.
9
Hiroyasu, H., and Kadota, T., "Fuel Droplet Size Distribution in Diesel Com-
bustion Chamber," Society of Automotive Engineers, Warrendale, PA, SAE Paper
740715, 1974.
10
1985.
H
Harlow, F. H., and Amsden, A. A., "A Numerical Fluid Dynamics Calculation
Method for All Flow Speeds," Journal of Computational Physics, Vol. 8, 1971,
pp. 197-213.
12
Wu, K.-J., Coghe, A., Santavicca, D. A., and Bracco, F. V., "LDV Meas-
urements of Drop Velocity in Diesel-Type Sprays," AIAA Journal, Vol. 22, Sept.
1984, pp. 1263-1270.
13
Chehroudi, B., Chen, S.-H., Onuma, Y., and Bracco, F. V., "On the Intact
Core of Full-Cone Sprays." Society of Automotive Engineers, Warrendale, PA,
SAE Paper 850126, 1985.
14
Martinelli, L., Reitz, R. D., and Bracco, F. V., "Comparisons of Computed
and Measured Dense Spray Jets," Dynamics of Flames and Reactive Systems, Vol.
95, edited by J. R. Bowen, N. Manson, A. K. Oppenheim, and R. I. Soloukhin,
Progress in Astronautics and Aeronautics. AIAA, New York, 1985, pp. 484-512.
15
Andrews, M. J., and Bracco, F. V., "On the Structure of Turbulent Dense
Spray Jets," Encyclopedia of Fluid Mechanics, Vol. 8, Gulf, Houston, TX, 1989.
16
Andrews, M. J., private communication, 1989.
17
Bracco, F. V., "Structure of High-Speed Full-Cone Sprays," Recent Advances
in the Aerospace Sciences, edited by C. Casci, Plenum, New York, 1985.
Chapter 21
Ignition and Flame Spread Across Solid Fuels
Colomba Di Blasi
I. Physical Description of Problem
E XPERIMENTAL and theoretical studies on ignition and flame spread

over solid fuels are motivated mainly by their importance to fire safety.
An improved understanding of controlling mechanisms of flame initiation
and spread can be very useful in prevention and control of fires. In addition,
studies related to the flammability of solid fuels are of current concern in
the ignition transients of solid propellants and hybrid rocket motors, in-
cineration of solid wastes, energy recovery from biomass pyrolysis, and
basic problems in combustion science.
When a solid surface is exposed to external heating from thermal radia-
tion and subjected to endothermic processes related to water evaporation
and phase change of the solid fuel, a temperature increase is observed. As
a consequence, the solid starts to degrade at a rate depending on the local
temperature, and fuel vapors, released into the atmosphere, mix with oxy-
gen. Solid-phase heating also causes the adjacent gas layer to be heated.
If the temperature becomes high enough to initiate and accelerate the
combustion reaction rate, flaming ignition occurs. In some cases, instead
of flaming ignition, a different phenomenon, called "glow,"1 is observed.
Indeed, some materials leave a porous char layer 2 when all volatile gases
are removed. As the solid continues to pyrolyze, the fuel vapors flow
through this layer, where they encounter higher temperatures. They may
undergo catalytic reactions, and also may react with the oxygen that had
diffused into the subsurface layer from the atmosphere. The temperatures
in this region are high because of the absence of endothermic processes
related to pyrolysis and because of large exothermicity of oxidation reac-
tions. Often, glow does not produce a high concentration of decomposition
products in the gas phase, and the gas-phase mixture is barely flammable.
A different behavior is shown by thermoplastic polymers, which, when
exposed to a radiant flux, start to liquefy a narrow subsurface layer3 where
reserved.
643
644 IGNITION AND FLAME SPREAD ACROSS SOLID FUELS
bubbles grow and transport volatile gases toward the surface. In this case,
oxidative reactions can be important since oxygen diffuses through the melt
layer. Furthermore, oxygen penetration inside the solid is favored by the
large holes left at the surface by burst bubbles.
In general, to get flaming ignition, three conditions must be met4: 1)
fuel and oxidizer must be available at a proper level of concentration to
give a mixture within the flammability limits; 2) the gas-phase temperature
must attain values sufficiently high to initiate and accelerate the combustion
reaction; and 3) the extent of the heated zone must be sufficiently large
to overcome the heat losses.
After the external heating source is removed, flame spread and possible
solid burning are established only under appropriate conditions. The ex-
othermicity of gas-phase reactions and reactions inside the solid fuel cause
a heat flux that, through solid- and gas-phase heat-transfer processes, heats
the unburned fuel. To have flame spread, the burning region must supply
a sufficient amount of heat to the unburned solid fuel to cause degradation.
At the same time, the conditions in the gas phase must allow the combustion
reaction to proceed. Many interacting physical processes are involved in
flame spread whose controlling mechanisms depend on factors such as the
flow configuration, oxygen level, orientation of the solid fuel, and fuel
properties.
II. Classification of Ignition and Flame-Spread Problems
Solid-fuel ignition phenomena can be either piloted ignition or autoig-
nition. However, a clear-cut division between the two categories does not
exist since, in the present case, the system of interest is open.
In general, piloted ignition is assumed to be caused by a device capable
of creating a region of very high temperature in the gas phase (pilot flames,
sparks, and hot wires). Autoignition is caused by thermal radiation, hot-
air streams, or hot surfaces. These different modes of ignition have been
investigated experimentally since each of them corresponds to a physical
situation of interest. For instance, cases of piloted ignition are related to
fires caused by accidental localized heat sources; thermal radiation is the
primary mode of heat transfer from large fires; a hot-air stream is important
in fire conditions too, and hot products of combustion of explosive materials
in an igniter can bring solid propellants to ignition; a hot surface is related
to room fires when solid materials can become combustible resulting from
an increase of the temperature caused by radiation from the hot combustion
products. The difference between piloted ignition and autoignition lies
essentially in the fact that, in the former case, only conditions 1 and 3 must
be met, since an effective pilot flame is provided beforehand.
Flame spread over solid fuels can be classified into two categories ac-
cording to flow conditions.5 One mode occurs when the flame spread is in
the same direction as the oxidizing flow. The second mode occurs when
the flame spreads against the oxidizing gas flow. These two modes of flame
spread are illustrated schematically in Fig. 1.
In the flow-assisted mode of flame spread, the concurrent flow pushes
the flame ahead of the vaporizing fuel surface. The heat transfer from the
hot mixture of reacting gases and the combustion products above the va-
C. Dl BLASI 645
a) concurrent flow
d irect i or fuel vapor
heated 1ayer
solid phase
b) opposed flow
gas phase
f lame flame l e a d i n g edge
p y r o l y s i s front
fuel vapor
f f f f f t d i r e c t ion
neated layer
solid phase'
Fig. I Schematic of flame-spread problems: a) flow-assisted flame spread; b) op-

posed-flow flame spread.
porized region to the unburned fuel surface favor flame propagation. The
resulting flame-spread process is very rapid and, consequently, of great
importance to fire safety science. For a very large range of flow velocities
and oxygen concentrations, this phenomenon appears to be controlled
simply by heat-transfer mechanisms. 5 However, the flow remains laminar
only in the initial stage when heat transfer from the flame to the fuel is
due mainly to convection. When the dimensions of the flame increase, the
flow becomes turbulent, and flame radiation becomes the dominant mode
of heat transfer.
In the case of opposed-flow flame spread, the heat transfer to the un-
burned fuel is more difficult since the flame and pyrolysis fronts are in the
same location. The opposed flow pushes the flame in the burning region,
and heat transfer to the unburned fuel, which occurs through solid- or gas-
phase conduction, is very slow. The flame front is well defined, and, gen-
erally, the size of the fire is controlled easily. The phenomenon shows an
interesting dependence on environmental conditions 5 ; it is dominated by
heat-transfer mechanisms at relatively low opposed-flow velocities and high
oxygen concentrations and by chemical kinetics at high flow velocities or
low oxygen concentrations.
Ignition and flame-spread processes depend strongly on the presence of
gravity and, in a normal gravity environment, on the orientation of the
material. 5 Upward flame spread—i.e., in the direction opposed to the
gravity vector—is faster than downward flame spread—i.e., in the same
direction as the gravity vector. In the case of upward flame spread, the
heat transfer from the hot combustible gases to the fuel is enhanced by
natural convection, while, in the opposite case, natural convection slows
the spread process, taking away the hot combustion products from the
unburned fuel. Upward and downward modes of flame spread show the
same characteristics of flow-assisted and opposed-flow flame spread. Dif-
ferent characteristics of spread are observed6 in a microgravity environment
when no natural convection is present.
Another important factor for flame-spread processes is fuel thickness.
In general, the definitions of "thermally thin" and "thermally thick" fuel
are used.5 Thermally thin indicates that the fuel thickness is small compared
with the characteristic thickness of the thermal diffusion layer in the solid,
along which no significant variation of solid properties is observed. De-
pending on the thickness of the solid fuel, the role played by the solid-fuel
thickness in the path for heat transfer to the unburned fuel changes from
being of no importance for thin fuels to becoming of primary importance
for thick fuels. If the combustible is thin, flame spread is characterized
further by the consumption of the solid fuel in the burning region. This
produces a propagating burnout front that affects the flame and pyrolysis
front propagation.
Material properties influence the behavior of the sample under fire con-
ditions. Essentially, it is possible to consider two types of materials: char-
ring materials—such as wood and, in general, cellulosic materials—and
noncharring materials—such as some thermoplastic polymers. Most ex-
periments on flame spread over thick solid fuels have been carried out with
polymethylmethacrilate (PMMA), which is a noncharring material, whereas
experiments related to thin fuels have used mainly paper, which is a char-
ring material.
III. Modeling Solid-Fuel Combustion Processes
In light of the description and classification given in the preceding par-
agraphs, a mathematical model of combustion phenomena related to solid
fuels should include the description of transport and chemical processes
occuring in the solid and gas phases and should account for their interaction.
The mathematical complexity of such a problem has prevented the devel-
opment of effective models, partly because of the changes needed in the
models as material and combustion conditions are changed. The models
formulated to date mainly address some particular aspect of the phenom-
enon by making assumptions that limit the range of their applicability. A
rather extensive literature, reviewed in Refs. 1 and 7-9, is available on
simplified models that give explicit expressions for global parameters such
as the ignition delay time and the spread rates as a function of material
properties and environmental conditions.
The simplest criterion of ignition is based on the solid reaching critical
value of temperature, and the mathematical models include one-dimensional
solid energy balances. Asymptotic techniques, valid in the limit of large
activation energies, have been applied to more advanced one-dimensional
models with finite-rate gas-phase kinetics.10 u
The most commonly used approximations in flame-spread models are
the constancy of the surface temperature of the solid fuel during vapori-
C. Dl BLASI 647
zation, the flame-sheet approximation, and the boundary-layer approxi-

mation. The assumption of constant vaporization temperature has been
criticized7 since, although solid fuels reach constant surface temperature
in the pyrolysis zone during the vaporization process, this value depends
on material properties, ambient pressure and temperature, solid-fuel thick-
ness, and heating conditions. The flame-sheet assumption implies the ex-
istence of an infinite gas-phase reaction rate, whereas boundary-layer theory
assumes that diffusive processes along the streamwise direction are much
smaller than those occurring in the cross direction. However, at the flame
leading edge, both for opposed-flow and flow-assisted flame spread, and
at the flame tip, for flow-assisted flame spread, the assumption of infinitely
fast kinetics is not verified. At the leading edge of the flame, the reaction
rate is slow since the flame is very close to the surface of the fuel which,
generally, pyrolyzes at rather low temperatures. The existence of a quench-
ing layer close to the surface allows the diffusion of fuel vapor upstream
and the formation of a premixed zone. At the concurrent-flow flame tip,
the flame leans toward the surface because of the lack of fuel and, for the
same reasons as at the leading edge, the reaction rate is slow. Consequently,
finite-rate kinetics effects are important in some regions involved in the
spread process and, of course, at near-ignition and extinction conditions.
The boundary-layer approximation is not verified at the flame leading edge
and tip. At the leading edge of the flame, strong streamwise temperature
and concentration gradients exist: the presence of a premixed zone makes
the conditions very similar to those of laminar premixed flames. Streamwise
diffusion at the leading edge of the flame is very important in opposed-
flow flame spread because, along with solid-phase heat conduction, it pro-
motes flame spread. It is also important in near-limit concurrent spread,
as verified by experiments12 that show that extinction begins just in this
region.
In general, the previously-described simplified models give correct so-
lutions only under conditions in which the employed assumptions are valid
and, except for explicit expressions of spread rates and ignition delay times,
no other information on the phenomena is obtained. More useful models
include balance equations for both gas and solid phases and remove some
of the approximations. These more complete models are not amenable to
analytical solutions, and it is necessary to use numerical approaches.
In the following sections, computer models formulated for ignition and
laminar flame spread will be examined. The main difference between ig-
nition and the flame-spread models is that most ignition models use a one-
dimensional description, while at least two-dimensional equations must be
considered to describe flame spread.
IV. Gas-Phase Models

A typical model that removes the assumptions previously mentioned is
that of Ref. 13. The gas-phase equations considered are for
Continuity:
dp ^ d(pu) + =0
dt dx dy
Momentum along x direction:

du du du\ dp d fdu\ d (du
+u +v = +
~ "
+ n r +r - <KP - P») sine (2)
a*Vd* ay
Momentum along v- direction:
+ u dv + v dv\ = dp+ d idv\ a (dv

-
u c/ i /
U I uvi
+ 3 p /
P()) cose _
x /O\
(3)
+ v a^\aA:
T(T- r
a^/ 7/ ~ # "
Chemical species:
ay, ay,
-f u —- + v —-
dx dy
>
= w/ + iL
dx[
pte )l + A[DP^'
\dxj] dyl
2 y, = 1, / = I, P, O, F (4)
Energy:
[ dT <*T\ = k—(—} k—(—} (5)

t + U
dx + V
dy) ~ q +
a^V^/ + ^\^/
where
iv, - -Agexp(-Eg/RT)YFY0MF/(viMl)p2 f = O, F
2
wp = AgQxp(-Eg/RT)YFY0MF/(vpMP)p w, = 0
q = -wFMl
Ideal gas law:
p = pRT (6)
In Eqs. (1-6), u and v are the streamwise and cross-velocity components,
p the pressure, jx the viscosity, Y the species mass fraction, p the density,
D the diffusion coefficient, cp the specific heat at constant pressure, Tthe
temperature, k the thermal conductivity, g the acceleration of gravity, 0
the angle of solid-fuel orientation with respect to gravity, Ag and Eg the
pre-exponential factor and the activation energy in the gas-phase reaction,
C. Dl BLASI 649
A// the heat of combustion, R the universal gas constant, v the stoichio-
metric coefficient, and M the mean molecular weight. The subscript 0 refers
to initial conditions and / to chemical species: F(fuel), O (oxygen), / (inert),
P (product).
In the formulation of balance equations, some assumptions are made:
constant thermal properties, negligible viscous dissipation and compressive
work, negligible soot and flame radiation, laminar flow, and validity of the
ideal gas law. Furthermore, once the pyrolysis process starts, the fuel vapor
(F) reacts with the oxygen (O) of the surrounding atmosphere, according
to a global second-order Arrhenius reaction: F -+- v0O = vPP. To describe
ignition and flame spread over solid fuels, Eqs. (1-6) must be coupled to
solid-phase balance equations through boundary conditions at the solid-
gas interface.
To date, flame-spread models numerically solved for the gas phase treat
the velocity field differently. The most complete models are those that
consider the solution of the full Navier-Stokes equations with density and
pressure variations described according to the ideal gas law. These models
are also the most expensive in terms of computer time. In addition to the
model of Ref. 13, the models in Refs. 14-16, used to predict the thermal
and convective structure of a diffusion flame in a counterflow environment,
can be included in this category. In the models of Refs. 14-16, only gas-
phase equations are solved, and solid-phase processes are described by
proper boundary conditions at the solid-gas interface. Using this approach,
the most important feature in the modeling of flame spread—i.e., coupling
between the degradation of the solid-fuel and gas-phase processes—is not
described. However, interesting information has still been obtained on the
extinction and blowoff of diffusion flames.
A second type of model,17 used to simulate flame spread, assumes the
Boussinesq approximation for the formulation of momentum equations.
Density is constant except for variations in the buoyancy-induced body
forces. In this approximation, the continuity equation is identically satisfied
and the pressure can be decoupled from the velocity field if the unsteady
Navier-Stokes equations are expressed in vorticity £ and stream function
X transport form for
Vorticity:
p — + u— + v —
dt dx dy
d /a .
\dx/ M* T- a
dy\dyj + P*S T- sme ~ T-
\dy dx
cos6
Stream function:
dx\dx] d
where
du dv
a x = u,
— —a x= -v, and A £ Y= — - —
dy dx dy dx
and B is the volumetric expansion coefficient. The pressure distribution
can be computed from a Poisson equation obtained by taking the diver-
gence of momentum equations, written in terms of primitive variables u
and v, and from the continuity equation:
dxdx dydy = 2 I - \dy\dx \dx\dy
This approach is less complex and less expensive in terms of computer

costs than the one using full Navier-Stokes equations.
The simplest models consider the solution of species and energy equa-
tions assuming that gas density and pressure are constant and the velocity
field is known. 18 " 25 The streamwise component of velocity is assigned
acceding to a chosen profile, while the effects of the mass outflow from
the vaporizing surface on the gas velocity field are neglected. Consequently,
these models account essentially for thermodiffusive aspects of flame-spread
phenomena. The computational costs are even lower, and so they have
been widely used.18"25 Predicted values of the opposed-flow spread rate
depend strongly on the velocity profile used in the computations. However,
the use of flowfields assigned according to the Oseen hypothesis (uniform
velocity profile) and to the Hagen-Poiseuille profile (parabolic velocity
profile) do predict the same functional dependence of the spread rate on
the maximum oxidizing flow velocity26 as observed in the experiments.
V. Solid-Phase Models
From the mathematical point of view, models of thermally thin fuels can
take advantage of the uniformity of variable distribution along the fuel
thickness, observed in the experiments, and use one-dimensional balance
equations. However, fuel consumption (burnout) cannot be neglected.
Two-dimensional heat transfer must be taken into account for thermally
thick fuels, but, in this case, fuel consumption is generally neglected. Mod-
eling fuels of intermediate thickness is very difficult, and both two-dimen-
sional and solid consumption effects are important. To the best of the
author's knowledge, no model of flame spread for this case has been solved.
All models of flame spread use very simple descriptions for the degra-
dation of the solid fuel. Thermally thick materials have been modeled by
neglecting in-depth mass transfer and regression rate, by describing the
chemical kinetics with a one-step Arrhenius reaction, and including heat
transfer.13'20"24'26 In modeling cellulosic materials, char formation is de-
scribed in very simple fashion and only for thin fuels. 18a9-26 More compli-
cated solid-fuel models have been solved where the effects of internal
C. Dl BLASI 651
bubble formation on the steady-state gasification of thermoplastic materials27

and the variable material properties and pressure effects in the pyrolysis
of thermally thick wood28 are considered. However, these models are one
dimensional, and combustion and transport phenomena in the gas phase
are not modeled at all.
A. Thermally Thick Fuels

The model presented here for thermally thick fuels is essentially that of
Refs. 13, 20-24, and 26. The heated polymeric fuel remains solid until it
finally gasifies at the surface according to a zero-order Arrhenius pyrolysis
reaction giving the corresponding monomer. In such a way, the use of a
vaporization temperature is not needed and the temperature at the inter-
face, which depends on fuel properties and environmental conditions, is
found as a result of the solution. The external thermal radiation, which is
used to cause ignition, is absorbed by the solid fuel according to Beer's
law. Heat capacity, thermal conductivity, and density are assumed con-
stant. Processes related to surface consumption of the solid fuel are ne-
glected. By these hypotheses, and bearing in mind that pyrolysis is a surface
process and must be taken into account through the boundary conditions,
the mathematical model for the solid phase can be expressed by a simple
two-dimensional heat conduction equation:
ar/l , d TarJ , a far,

— = — — ks — \—-
dy\_dy
T)] (10)
where cs is the solid heat capacity, p5 the density, ks the thermal diffusivity,
Ts the temperature, 70 the intensity of the incident radiant flux, (35 the
absorption constant, and T the solid thickness.
The pyrolysis mass flux m can be expressed as
m = Asexp(-Es/RTs)ps . (11)
where As is the pre-exponential factor and Es the activation energy of the
pyrolysis reaction.
B. Thermally Thin Fuels

The model presented here for thermally thin fuels has been formulated
for thin paper in Ref. 25. It can be assumed that variable distributions
across the solid thickness are uniform. For chemical processes, only py-
rolysis is described: the paper (S) goes to vapor fuel (G) and char (C) by
S —» vGG + vcC, whose kinetics is expressed according to a zero-order
Arrhenius law. As the paper pyrolyzes, the solid density, specific heat,
and thermal diffusivity change from their initial values to their final values
for the char. The fuel vapors, flowing out of the solid immediately after
vaporization, are in thermal equilibrium with the char. In addition, the
char surface is at the same location where the paper surface was, but the
burnout front propagates as burning occurs. The energy balance accounts
for the variation of solid-fuel enthalpy, energy convection due to the flow
of volatile gases through the porous solid, heat diffusion along the direction
of flame propagation, and energy release due to the pyrolysis reaction.
These processes, neglecting the accumulation of mass and energy of volatile
gases inside the solid, are expressed by
Gas continuity:
= _
ay at
Mass balance:
Energy balance:
PA - = - - { L + [- pV(ps0 - Pc)ca + pc/(ps0 -
OA (14)
°yI \°*/ v /
where
~ PcV(PsO ~ Pc)CaPsO + (P,0 ~
- Pc)ka + (p,o - p5)/(p50 -
where the subscripts are sO, the initial solid property values; c, the final
constant char value; a, the active part of the pyrolyzing material. L is the
heat of vaporization.
From Eq. (9), the pyrolysis mass flux at the surface can be obtained as
m = tAJ[l - (pc/pso)](ps ~ PC) e*p(-Es/RTs) (15)
The conductive term in the energy balance does not play an important role
in the spread process, and it can be dropped from the equation with no
change in the results.25
VI. Initial and Boundary Conditions

Solid- and gas-phase models are coupled through the boundary condi-
tions at the interface. The boundary conditions are obtained by writing
species mass flux and heat flux balances. For the chemical species, the
mass fluxes of species / convected out and diffused out must be equal to
the mass flux of species / generated on the surface. If K is the mass fraction
of the gasified solid, which is fuel vapor, species mass flux balances at the
interface give
C. Dl BLASI 653
£
= m(YF - K) (16)
p ^Xo = m(Yo _ l + K} (17)

°y
= my,, / = /,/> (18)
For thermally thick fuels, the heat flux balance is expressed as
At each instant at the solid-gas interface, the solid temperature must be

equal to the gas temperature,
Ts = T (20)
The cross component of velocity is given by a balance of mass fluxes,

m = pv . (21)
whereas for the streamwise component, a no-slip condition is usually as-

signed (u = 0).
For thermally thin fuels, Eq. (19) is directly included in the energy
balance. In addition, in this case, boundary conditions change as the burn-
out front progresses. Conditions (16-21) are applied until burnout. In the
burned fuel region, the pyrolysis mass flux is zero, zero-normal derivatives
on chemical species and temperature equations, and zero velocities are
imposed.
The treatment of artificial boundaries is required downstream of the
leading edge of the flame. In this region, boundary conditions are difficult
to assign. For opposed-flow flame spread and downstream of the leading
edge of the flame, the boundary-layer approximation is adequate and,
consequently, the first derivatives along the streamwise direction are zero.
The problem is more complex for flow-assisted flame spread. Downstream
of the flame leading edge, there is flame spread and the flame tip can
interfere with the boundary. In this case, zero streamwise derivatives are
again assumed, but computations are valid only as long as the flame tip is
far away from the boundary. The remaining boundary conditions are con-
ventional and are based on the physical problem to be modeled (open
boundaries or walls).
The models presented earlier have been used to simulate both opposed-
flow and flow-assisted flame spread. The initial conditions correspond to
ambient conditions, and ignition of a small portion of the solid fuel is
caused by thermal radiation for thick fuels 13 - 20 " 24 - 26 and by a pilot flame
for thin fuels.25
VII. Finite-Difference Approximations

The numerical solutions of the models of flame spread13-18 ~ 26 have been
found by using finite-difference methods to determine a set of algebraic
equations. In particular, in Ref. 13 and 20-26, discretization has been
done using the Euler implicit difference scheme for the unsteady terms,
the upwind schemes for the convection terms, and the central second-order
scheme for the diffusion terms. The staggered grid places velocity nodes
in between those used for pressure, temperature, chemical species, and
density.
The main difficulties that arise in the finite-difference discretization are
due to the existence of very steep spatial gradients of temperature, species
concentrations, and velocity, both in the streamwise and cross-stream di-
rections. In opposed-flow flame spread, there is a distinct flame front where
physical variables show a large change from the region of burning to the
region of unburned gases, upstream of the flame leading edge. This does
not occur for flow-assisted flame spread, which does not have a well-defined
flame front, and profiles of solid and gas variables are smoother along the
streamwise coordinate.
In flame-spread problems, it is very important to have an accurate eval-
uation of the cross gradients. This is particularly important for the heat
flux from the flame to the fuel surface, k(3T/dy), since spread or not and
predicted spread rates strongly depend on how the solid fuel is heated and,
therefore, on its degradation rate.
The use of very fine computational grid spacing is also required after
ignition, when the external heat source is removed, and the flame may
spread or die out. To obtain accurate predictions of flame spread or quench-
ing, it is important to evaluate the streamwise heat conduction accurately.
Problems associated with the presence of steep gradients in physical
variables have been treated only partially. Variable computational grids
have been used for opposed-flow22 and flow-assisted25 flame spread. The
grid along the cross direction22'25 must be very fine near the surface, where
the gradients are large, and can be coarse in the upper part of the com-
putational domain, where small variations are observed. In the streamwise
direction, the grid25 must be very fine near the ignition area, where the
gradients are large and diffusion is very important, and it can be coarser
downstream since, for flow-assisted flame spread, the gradients become
small. These grids are fixed in space and time, and, consequently, they do
not follow the time evolution of the flame spread. In particular, for the
cross direction, fine gridding close to the surface along the whole solid-
fuel length is used, which wastes computer time since far upstream gradients
are negligible. Furthermore, opposed-flow simulations have been carried
out only with a fine constant grid along the x direction, as dictated by the
requirements of the description of the leading edge of the flame, and,
consequently, computational times are high.
Many of these drawbacks could be avoided by using an adaptive grid,
which is also needed to model the pyrolysis of thick charring fuels. Such
problems are characterized by the propagation of a pyrolysis wave inside
the solid, and this reacting front is associated with very steep gradients in
C. Dl BLASI 655
temperature and density. The problem is made even more complex by the
variation of thermal properties, diffusivity, and porosity from the values
of unburned material to those of char and by the anisotropic character of
charring materials, which show different parallel and cross-grain values of
thermal properties and permeabilities.
VIII. Numerical Solution Procedures

Characteristic properties of the finite-difference equations that must be
solved are nonlinearity, some of which represents chemical production
terms, and coupling among the various physical processes. Proper linear-
ization techniques20-23 allow the first aspect to be handled efficiently. Block
implicit methods, which treat coupled systems of equations, have not yet
been used in flame-spread modeling. Conversely, semi-implicit methods
have been applied through operator splitting or iterations. 13 " 26
The solution methods are semi-implicit, so that the sets of finite-difference
equations corresponding to each differential equation are solved in turn.
In general terms, the solution procedure is as follows:
1) Solve the solid-phase model equations given the heat flux from the
flame to the fuel. In addition to the solid-phase variables, this step also
computes the pyrolysis mass flux and the temperature at the surface to be
used in the boundary conditions for the gas-phase equations.
2) Solve the gas-phase equations (momentum, energy, species) with the
boundary conditions that use the previous values of the pyrolysis mass flux
and surface temperature. This step also allows a new distribution of the
heat flux, k(8T/dy), to be computed and then used as a boundary condition
in the solution of solid-phase equations.
Given the coupling among the solid- and gas-phase variables, steps 1
and 2 must be repeated for each time step until the computed variables
converge. For each spatial node, it is required that
[£n + 1, m + 1 - & + l,m]/& + 1, m < e
where £ is a general dependent variable, n refers to the time level, and m
refers to the particular iteration. The number of iterations for convergence
depends on numerical parameters (essentially the time step) and on phys-
ical properties and data (environmental conditions).
The coupling between the solid- and gas-phase equations occurs both
through the conditions on chemical species and temperature [Eqs. (16-
20)] and the condition on velocity [Eq. (21)]. The coupling between the
gas-phase flowfield and the solid-fuel degradation is particularly important
at ignition conditions. The solid surface, exposed to external radiant flux,
can reach relatively high temperatures. Given the exponential dependence
of the reaction rate on temperature, the pyrolysis mass flux attains even
higher values, often leading to cross velocities larger than the forced stream-
wise velocities.
Needless to say, the solution of the solid-phase equations is straightfor-
ward. It is achieved by means of an alternating direction implicit (ADI)
method for the thick fuels, 13 ' 20 ~ 24 - 26 and by a splitting approach, which
employs a fourth-order Runge-Kutta method for chemical operators (two
ordinary differential equations) and a simple implicit method for the dif-
fusion operator, for thin fuels. 25 Numerical solution of gas-phase equations
is more complex.
A. Steady Gas-Phase Equations in Thermodiffusive Models

The simplest version of gas-phase models is the thermodiffusive one with
constant pressure and density and a known flowfield. Models of this type
essentially differ in that they use either steady 18 - 19 or unsteady 20 " 26 for-
mulations of gas-phase equations. Although steady gas-phase models can-
not be used to simulate ignition, they are useful for flame spread. Even
though spread rates do not depend on time (except for some cases of
concurrent spread), flame spread is an unsteady phenomenon. However,
a quasisteady treatment of the gas-phase equations can be used based on
the fact that the characteristic evolution times of the solid are much longer
than those of the gas phase, given that the flame-spread rate is much smaller
than the gas velocity. In this approach, gas-phase conditions can be con-
sidered as a series of steady states resulting from small changes in the
solid.18'19
The choice between the steady and unsteady approaches has been
examined20 to determine which one can save computer time in the simu-
lation of flame spread. The comparison has been made for the case of
opposed-flow flame spread starting from steady conditions of propagation.
Two methods have been tested in the solution of species and energy equa-
tions: the classical iterative Gauss-Seidel (GS) method and a line-by-line
(LbL) method, which is a convenient combination of the Thomas algorithm
for tridiagonal systems and of the Gauss-Seidel method. Computer times
obviously increase by decreasing the convergence parameter e for both
formulations. However, the increase is higher for the steady approxima-
tion, and even more so when the GS method is employed. The lower values
are obtained for the unsteady formulation and, in this case, they are almost
independent of the method used to get the solution. Therefore, the un-
steady formulation, in addition to being more general, also costs less. For
values of e currently employed (~ 10 " 4 ), the reduction in computer time
ranges from factors of 4 for the LbL-method to 7 for the GS method. For
this value of e and for unsteady gas-phase equations, both methods show
exactly the same performance. However, because implicit methods are
more stable, the LBL method has been used in the solution of energy and
species equations. 20 " 26
B. Solution of Momentum Equations

Two-dimensional models of ignition and flame spread17 that include a
solution of the momentum balance equations according to the Boussinesq
approximation have also been solved by the semi-implicit procedure out-
lined earlier. In the gas-phase step, the vorticity field has been computed,
and then the Poisson equation for the stream function and equations for
species and energy are solved. The new values of velocity and the boundary
values of vorticity are then computed.
The problem becomes more complex from a numerical point of view
when full Navier-Stokes equations with density and pressure variations are
C. Dl BLASI 657
considered.13 A fundamental point in this type of model is represented by

the coupling between pressure and velocity. The most popular methods
used to solve momentum equations account for this coupling by iterating
for each time step. Methods based on the SIMPLE approach28 have been
used in the numerical predictions of flame structure in an opposed-flow
environment in Refs. 14-16. However, it has been found that the iteration
procedure is much too time consuming. Moreover, convergence is not
always guaranteed unless ad hoc under-relaxing factors are used. A more
effective technique, which avoids iterations by splitting computational op-
erations into a few steps, was proposed by Issa for nonreacting flows.3031
This technique, named PISO (pressure implicit with splitting operators),
looks very promising for combustion problems where the costs of iterative
procedures are very high, given the larger number of balance equations to
be considered.
The PISO technique is an extension of classical splitting procedures used
to solve discretized equations, as ADI, to the treatment of the coupling
between pressure and velocity in the momentum equations by predictor-
corrector steps. This technique has been included in the semi-implicit so-
lution procedure outlined earlier for flame-spread models when the com-
plete momentum balance equations are considered.13 One predictor and
two corrector steps are considered. Starting from a known pressure field,
the velocity distribution is computed by an implicit solution of momentum
equations (predictor step). In general, this velocity field does not satisfy
the continuity equation. Hence, a corrector step is performed where a new
pressure distribution is computed by means of the latest available velocities.
Then the flow velocities are also updated, and temperature and species
mass fractions are also computed. At this point, the continuity equation
is satisifed, but to improve the accuracy of the solution, a new corrector
step for pressure and velocities is performed. It is possible to demonstrate
that solutions obtained with one predictor and two corrector steps are
second-order accurate in time.30-31 A higher formal order of accuracy can
be reached by increasing the number of corrector steps, but the second-
order approximation is sufficient for the problem under study.
IX. Numerical Simulations of Ignition and Flame Spread

In this section, some results of computations carried out using the com-
puter models described earlier are presented. Time and space evolution
of solid- and gas-phase variables from ignition to spread are discussed. The
predicted dependence of ignition delay times on the intensity of the external
radiative flux and of spread rates on velocity and oxygen concentration of
the oxidizing flow are also given.
A. Ignition and Opposed-Flow Flame Spread

Apart from very few models proposed for hot-air-stream ignition, 32 - 33
most computer models of ignition consider one-dimensional balance equa-
tions. Among these models, those of Refs. 34-36 are interesting in that
gas-phase absorption of radiation is modeled for the first time in solid-fuel
radiative ignition. It is assumed that fuel vapors absorb thermal radiation
according to Beer's law, which, for the gas phase, can be expressed as
d/
—•** = p/o
or
pF exp P - p/?F dy
ay Jo
where /?F is the partial pressure of fuel vapors and p a constant, global
absorption coefficient. Simulations of PMMA radiative ignition34" 36 pre-
dict that ignition occurs in the premixed gas phase, rapidly followed by
transition to a diffusion flame. As the heat flux is increased, higher surface
temperature and pyrolysis mass flux are reached, ignition is located closer
and closer to the fuel surface, and the ignition delay times decrease. Figure
2 shows the calculated ignition delay times,36 td, as a function of 70 and p,
and the experimental ignition delay times37 as a function of /0 and sample
orientation. Even though the shape of the curve giving td as a function of
/0 is unchanged as p is varied, there is a strong dependence of the predicted
values of td on low values of p (0.01-0.06 atm^-cm" 1 ) and a weaker
dependence for higher values, where td is essentially determined by solid-
fuel heating. It must be pointed out that when p = 0, ignition does not
occur, even if very high values of the radiative heat flux are used. This
confirms the controlling role played by gas-phase absorption of radiation
in the radiative ignition.
From experiments, it is shown that the ignition phenomena, and partic-
ularly the ignition delay times, depend on the sample orientation. Although
the description of both solid-fuel processes (pyrolysis, absorption of ra-
diation, heat conduction, and solid-surface regression) and gas-phase proc-
36. O
3O. O
6. O
O. O
4. O 8. O 12. O 16. O 20. O 2*. O
2
IQ [ W / c m ]
Fig. 2 Ignition delay times as a function of 70 for several values of P: ___, model36;
*, experiments, vertical orientation37; o, experiments, horizontal orientation.37
C. Dl BLASI 659
esses (combustion, variable density, convection and diffusion of heat and

mass, and absorption of radiation) has allowed ignition delay times to be
predicted surprisingly well, the model36 is one-dimensional, so that it cannot
in any way account for phenomena related to buoyancy or two-dimensional
convective effects.
The results of a detailed simulation of radiative ignition obtained using
the two-dimensional model based on the Boussinesq approximation is shown
in Figs. 3-5.17 Starting from ambient conditions in an opposed-flow en-
o .o
o.o x [cm]
Fig. 3 Constant contour levels of the gas-phase local heat-release rate from t =
0.150 to 0.1572 s with a time step of 0.8 ms. The contour values (cal/cm2-s) starting
from the left side to the right are 1, 10, 50, 100, 200, 400, 800, 1600, 2600.
vironment (u = 20 cm/s), ignition of an assigned part (0.0625 cm) of a

thermally thick cellulosic fluel slab, placed horizontally in a channel, is
caused by radiation (/0 = 20 cal/cm2-s, $s = 200 cm" 1 , r = 0). Even
though this model does not account for gas-phase absorption of radiation,
the main processes of radiative ignition are included. The flame rise is
shown by the constant contour levels of the local heat-release rate (Fig.
3), the fuel (solid lines) and oxidizer (dotted lines) (Fig. 4), the solid- and
gas-phase temperatures (Fig. 5), and the vector velocity field (Fig. 5) for
several times. After the heating stage, fuel vapors are produced and mix
with oxygen (Figs. 3a and 4a). For t = 0.152 s, ignition occurs, as confirmed
by high reaction rates in the left corner of the computational domain (Fig.
3b), by gas-phase temperatures considerably higher than those of the solid
phase (Fig. 5a), and by the consumption of fuel and oxygen (Fig. 4b).
Then the flame begins to heighten and to move forward over the fuel slab
(Figs. 3c-f, 5b) until the oxidizer present in the left zone is burned out
(Figs. 4c-d). The large amount of vapor fuel, generated during the heating
of the gas layer adjacent to the solid when solid thermal degradation has
already begun, is also burned out (Figs. 3g-j). For t = 0.1572 s, the flame
leading edge moves beyond the heated zone and the spreading process
starts.
From Fig. 6, which shows the physical variables at t = 0.162 s, the flame
structure in an opposed-flow environment can be analyzed. Opposed-flow
flame spread is dominated by the interaction of fluid dynamics, chemical
reactions, and heat transfer at the flame leading edge, where a combustible
mixture is established. This position is marked by a vertical line in Fig. 6a
and corresponds to the maximum heat flux from the flame to the solid
fuel. Given that the simulated conditions are far from extinction, the edge
of the heated layer in the gas is very close to that in the solid. Downstream
Fig. 4 Constant contour levels of fuel (solid lines) and oxygen (dotted lines) mass
fractions: a) t = 0.150 s; b) t = 0.152 s; c) t = 0.153 s; d) t = 0.157 s. The values
from the left side to the right are 0.2, 0.1, 0.15, 0.05 (oxygen); 0.001, 0.01, 0.1,
0.3, 0.5, 0.6, 0.7, 0.8 (fuel).
C. Dl BLASI 661
a)
Fig. 5 Constant contour levels of solid- and gas-phase temperature and vector
velocity field: a) t = 0.152 s; b) t = 0.154 s. The values (K) from the left side to
the right are 350, 400, 450, 500, 550, 600, and 650, and then increase with 250-K
steps. Velocity at the inlet is 20 cm/s.
of this position, the flame structure is typical of a diffusion flame in a

counterflow environment: isolines of reaction rate (Fig. 6c) correspond to
the flame position characterized by the maximum temperature and minima
of fuel and oxygen mass fractions. From the vector velocity field (Fig. 6a),
two main flows appear, one of vapor fuel from the pyrolyzing surface and
the other of air at ambient conditions from the inlet. It is interesting to
observe that the normal component of velocity at the interface, propor-
tional to the pyrolyzed mass flux, reaches the highest values behind the
flame leading edge. Consequently, the formation of the upstream com-
bustible mixture comes from fuel diffusion from the flame zone through a
quenching layer between the solid surface and the flame. The opposed
velocity goes to zero on the surface, allowing this fuel transport upstream.
In addition, the process is favored by the very steep gradients of fuel mass
fraction in this region, while it does not appear to depend on the surface
a)
O .O . O25 . O5 . O75 .1
x [cm]
Fig. 6 Constant contour levels of a) solid- and gas-phase temperature; b) fuel (solid
lines) and oxygen (dashed lines); c) local heat-release rate; and a) vector velocity
field for t = 0.162 s. Same contour values as Figs. 3-5.
temperature gradient. This mechanism of flame spread was also predicted

by previous numerical analyses.13" 22 26 An important effect, not described
in the present simulation, is the existence of a region of elevated pressure
upstream of the flame leading edge, caused by gas expansion. This effect,
described in Refs. 13-16, causes the outward deflection of the flow near
the flame front and a reduction of the gas velocity encountered by the
flame at its leading edge.
Numerical simulations of opposed-flow flame spread have also been used
to analyze the dependence of the spread process on gas-phase param-
eters18'21'22 and solid temperature. 19 Particularly interesting is the predic-
tion of the dependence of the flame-spread rate on the velocity and oxygen
concentration of the opposed gas flow. The spread rate for thick PMMA,
as measured38 and predicted by a thermodiffusive model,22 is reported in
Fig. 7 as a function of the opposed-flow velocity and for several values of
the oxygen mass fractions. The analysis, which qualitatively describes the
experimental results, predicts a spread rate that varies with the flow velocity
C. Dl BLASI 663
0.3
CO Exp . Mode 1 Y00
O • -"3
D • -329
0.24 - A A -230
o o
0.18
0.12
0.06
10 10 10 10 10
UMRX [CM/S]
Fig. 7 Dependence on the opposed-flow velocity of flame-spread rate for oxygen
mass fraction of 0.23, 0.329, and 0.533.22
for fixed oxygen concentration. This spread rate first increases and then
decreases as the flow velocity increases. For a given flow velocity, the
spread rate increases with the oxygen concentration, and the location of
the maximum is also displaced toward the higher velocity region as the
oxygen concentration increases. The major quantitative discrepancies be-
tween the theory and the experiments occur in the very low and very large
velocity regions, where the approximations of the thermodiffusive model
are not valid. Indeed, at low flow velocities, the phenomenon is controlled
by buoyancy, not accounted for in the model, whereas at high flow veloc-
ities, the flame structure is very sensitive to the gas velocity profile, not
well described in the model. The very large differences shown, at high
velocities and for oxygen mass fraction equal to 0.533, can be attributed
not only to the superimposed flowfield but to the gas-phase combustion
model, which is very simple and uses data determined for air. For inter-
mediate gas flow velocities, the flame-spread rate is controlled by the
transfer of heat from the flame to the solid, which is less sensitive to the
relatively small variations of the velocity profile. Thus, in this range of
velocities, the analysis of the spread of the flame is quantitatively better.
The results of numerical analyses have also been employed to determine
the controlling mechanisms of the flame-spread process and, thus, to verify
the phenomenological arguments used to explain the experimental results
in terms of thermally and chemically controlled mechanisms. Both the
maximum surface heat flux and maximum surface temperature increase
with the oxygen concentration and the flow velocity. The increase with the
oxygen concentration is due primarily to the increase in the flame tem-
perature; the increase with the flow velocity is due to the flame moving
closer to the surface. Both of these effects increase the heat transfer from
the flame to the fuel and the gasification rate of the solid combustible;
consequently, they would tend to increase the spread rate for all values of
flow velocity.
Detailed simulations of the flame-spread process have been made17 to
analyze extinction caused by high flow velocities or low oxygen concen-
trations. Although the mechanisms that lead to extinction are very different
for the two cases,17 extinction always can be related to a continuous de-
crease of the ratio of the flow time (a/u%) to the chemical time (p/wF), i.e.,
the Damkohler number. From both extinction cases simulated, indeed, the
flame leading edge recedes with respect to the heated solid fuel; thus the
solid heat conduction is not controlling at near-extinction conditions. The
interactions between fluid dynamics and chemical processes in the gas
phase, which are much faster than in the solid phase, are responsible for
extinction. In the former case, when the flow velocity is increased, the
slight increase in the chemical time, due to the increase in the fuel con-
centration, is much lower than that in the flow time. Thus, the Damkohler
number at the flame leading edge continuously decreases, leading to ex-
tinction. In the latter case, the flow time is unchanged, but the strong
increase in the chemical time, due to the decrease of the reaction rate,
again gives decreasing values of the Damkohler number and extinction
occurs.
B. Flow-Assisted Flame Spread

Only thermodiffusive models have been used to simulate flow-assisted
flame spread over thick23-24 and thin 25 fuels. A characteristic example of
the predicted flame structure and solid property distributions is shown in
Fig. 8 for thin paper. 25 The gas-phase isotherms (Fig. 8b) and contours of
equal species concentrations (Fig. 8c) show that the flame, except at its
upstream leading edge and tip, has a diffusional character with a relatively
large reaction rate. Four regions exist in the integration domain: the burn-
out region, the vaporization (or pyrolysis) region, the "combusting plume,"
and the "thermal plume." In the first region, from 0.08 to 0.068 m, neither
combustion nor degradation processes take place. The pyrolysis initiation
(x = 0.058 m) is characterized by a decrease in the solid density, which
continues until all of the fuel is pyrolyzed and only the constant-density
char is left. The termination of the pyrolysis process is also characterized
by a sharp increase in the solid temperature due to the accumulation of
C. Dl BLASI 665
a) gas f low
Dy no 1 y s is fr c
burn— out front
<
f lame3 tip
' .
t t t t *=F=T=?=-=*:~4
lid
f ue 1 10 ,,'*> '*!
1
400 . O
o. o o .o
O.O .01 .02 .03 .04 .05 .OB .07 .08
x Lml
c) . 02 1 .O
Qs
y Em] QsxlO [kg/
. 01 .5
———— f ~ - ~ ———"-^=-=^^=^^8-8.-,^,
\ ^
"
O.O .01 .02 .03 .04 .05 .06 .07 .08
x Em]
Fig. 8 a) Schematic of flow-assisted flame-spread problem. Predicted flame struc-

ture and solid property distributions for a mean flow velocity equal to 0.67 m/s, at
t = 2 s after ignition; b) gas isotherm field and solid temperature distribution; and
c) species equiconcentration fields and solid density distribution. For the gas phase,
the external isotherm is 600 K and the increment is 300 K. The isolines for the fuel
mass fraction correspond to 0.4, 0.3, 0.2, 0.1, 0.001; for the oxygen mass fraction,
the internal line is 0.05 and the increment is 0.05.25
heat by the char. Either constant char density or a sharp temperature
increase determines the location of the burnout front. The pyrolysis region,
which for thermally thin fuels is very narrow, is characterized by a nearly
constant surface temperature and by maxima of gas temperature and re-
action rate. Figure 8c shows that at the flame leading edge there is a
premixed zone similar to that shown in opposed-flow flame spread, where
the flame is thick. In the combusting plume region, the pyrolysis rate is
negligible. The fuel vapors, however, are pushed beyond the vaporization
region by the external flow rate, and burn giving high gas temperatures (x
= 0.025 m). The consequent high values of the heat flux at the surface
heat the solid fuel, which also reaches rather high temperatures. In the
final part of the combusting plume region, due to the proximity of the
flame to the surface and to the relatively low surface temperatures, the
reaction rate assumes low values, indicating that the flame-sheet approx-
imation is not valid here. The last zone at the left of the flame tip is the
so-called "thermal plume." Because of the absence of fuel vapors in this
zone, the reaction rate is zero. The species present are essentially the
oxidizer and the combustion products at a fairly high temperature. The
flame structure predicted for thick fuels is qualitatively similar to that
reported here except for the burnout region.
As time increases, the flame front progresses rapidly. Using a charac-
teristic gas-phase variable, such as the location where the local fuel-to-
oxygen equivalence ratio is 1, the flame length as a function of time can
be determined. As the flame front propagates, pyrolysis and burnout fronts
also propagate but at slower rates. The predicted25 and measured39 vari-
ation with time of the flame, pyrolysis, and burnout distances to the location
of the flame-spread initiation under several flow conditions are reported
in Fig. 9. The model agrees well with the experiments, which show a
spreading process that initially accelerates and then tends toward a steady
state as the flame spreads over the solid surface. An issue that can be
solved only with more complete analyses is the dependence of the spread
process on the flow velocity. The model predicts a spread rate that in-
creases, although weakly, with the flow velocity, whereas the experiments
show a spread rate that first increases at low flow velocities and then
becomes constant as the flow velocity is further increased.
The predicted spread rates23-24 increase with oxygen concentration and
flow velocity, and vary over a very large range of values. This is in agree-
ment with experiments when the combustible is thermally thick. Both the
increase of oxygen concentration and concurrent flow velocity increase the
heat flux from the flame to the fuel, and the spread process is enhanced.
This indicates that, unlike the opposed-flow case, flow-assisted flame spread
is controlled essentially by heat transfer. The effects due to finite-rate
kinetics are of increasing importance as extinction is approached, as a
consequence of low oxygen concentrations or very large flow velocities.
These effects appear mainly at the flame leading edge. Here the extinction
length, i.e., the distance of the flame leading edge from the edge of the
fuel slab, increases as the flow velocity is increased or the oxygen concen-
tration is lowered. Again, the extinction process can be explained in terms
of a decrease, below a critical value, of the Damkohler number at the
flame leading edge. The flame counteracts the decrease by moving in a
zone of high fuel concentration. The extinction process begins in this way,
but the downstream spread process goes on increasing flame and pyrolysis
lengths until the pyrolysis spread rate is greater than the upstream extinc-
tion rate. Complete extinction occurs when the extinction distance extends
to the position of the pyrolysis front. 24
X. Conclusions and Further Developments

Ignition and flame spread are the result of interactions among many
processes. For the solid phase, these include heat and mass transfer, in-
ternal absorption of radiation, chemistry of reactions interior to the solid
(thermal degradation and oxidation reactions), surface consumption, and,
for porous charring materials, evaporation of absorbed water, pressure
C. Dl BLASI 667
c b a
.3
X
.2
. 1
0.0
O .0 .O 8.O 12. O 18. O
t [s]
Experiments:
Gas velocity Flame Pyrolysis Burnout
0.625 m/s 0 4
1.00 m/s A A
2.00 m/s D
Model: —- —, flame; _, pyrolysis; — - - , burnout; mean gas velocity: a)
0.67 m/s, b) 1 m/s, c) 1.33 m/s, d) 2 m/s.25
Fig. 9 Predicted variation with time of the flame, pyrolysis, and burnout distances
for several flow conditions.
variations, and cracks caused by thermal stresses. For the gas phase, these
processes include convection and diffusion of mass, momentum, and en-
ergy, combustion reactions, ignition and flame propagation, radiation ab-
sorption by the decomposition products, and heat transfer by radiation.
Processes occurring at the solid-gas interface include convection and dif-
fusion of mass, momentum, and energy, reradiation from the surface, and
heterogeneous reactions between the gas phase and the solid phase. In
some situations, especially when the scale length of the phenomenon is
large, the problem is made more complex by gas-phase turbulence.
Analytical models have given only a global description of the problem.

Early theoretical studies of ignition and flame spread were devoted to the
formulation of explicit expressions for ignition delay times and spread rates.
These formulas are only partially able to describe qualitatively the de-
pendence on environmental conditions and property values of the solid
fuel. In general, the correct spread rates are given when phenomena are
controlled by heat-transfer mechanisms.
More detailed mathematical models, solved numerically, do not give
explicit expressions for global parameters. On the contrary, these param-
eters can be derived from the predicted time and space evolution of the
phenomenon, often by using the same approaches used in the experiments.
Indeed, the main difference between analytical and numerical solutions is
that, in the latter case, temperature, chemical species, and velocity distri-
butions can be obtained under different conditions, as it is done in the
experiments. The predicted ignition delay times and spread rate show
qualitative agreement with experiments under almost all conditions, and,
in some cases, quantitative agreement. The only partially quantitative
agreement can be attributed first to the simplified descriptions of the flow-
field. These simplifications have been removed in a more recent model
where the full Navier-Stokes equations have been solved,13 even though
the numerical computer code has not been applied extensively to simulate
flame-spread problems yet. Second, although the numerical models avail-
able to date are very advanced in relation to analytical solutions and en-
gineering approaches, they use very global descriptions of solid-phase
processes and of reaction chemistry for both phases.
Immediate improvements in the models should include the effects of
surface oxidative reactions and, for thermally thick fuels, of char formation.
To improve the numerical method, it is necessary to include algorithms
for adaptive grids. The unsteady formulation of gas-phase equations has
reduced the computer time by about a factor of 5, allowing the codes to
be run on minicomputers. However, the iterative procedure used to couple
the solid and gas phases is rather expensive, since, for each time step, the
two systems of equations for the two phases must be solved at least twice.
In the long term, there would be great progress in modeling of ignition
and flame-spread phenomena if a more detailed description of the different
processes, especially of chemical kinetics for combustion and solid deg-
radation, were available. This is not an easy task, and research in this
direction is at a very early stage. For example, very little has been done
to improve the models of solid-fuel degradation or to formulate detailed
kinetic mechanisms for polymer combustion. Once the submodels have
been validated, they can be included in a general model. The complexity
of the complete problem and the more and more detailed descriptions of
physical processes require the use of advanced numerical techniques and
high-speed computers.
The improvement in the predictions, however, do not depend only on
the cleverness of theoreticians and on the availability of high-performance
machines, but also on the advancements in the experimental analyses and
on the ability to transfer information from one field to another.
C. Dl BLASI 669
Acknowledgments
The author would like to express sincere thanks to Profs. S. Crescitelli
and G. Russo (University of Napoli), Prof. A. C. Fernandez-Pello (Uni-
versity of California), and Dr. T. Kashiwagi (NIST) for their encourage-
ment and support of her work in this field. Also, special thanks go to Dr.
E. Oran (Naval Research Laboratory) for many useful comments in writing
this chapter.
References
*Akita, K., "Ignition of Polymers and Flame Propagation on Polymer Surfaces,"
Aspects of Degradation and Stabilization of Polymers, edited by H. Jellinek, El-
sevier, New York, 1978, pp. 500-525.
2
Kanury, A. M., and Blackshear, P. L., Jr., ''Some Considerations Pertaining
to the Problem of Wood-burning," Combustion Science and Technology, Vol. 1,
1970, pp. 339-355.
3
Kashiwagi, T., and Ohlemiller, T. J., "A Study of Oxygen Effects on Nonflaming
Transient Gasification of PMMA and PE During Thermal Irradiation," Nineteenth
PA, 1982, pp. 815-823.
4
Kashiwagi, T., "Radiative Ignition Mechanisms of Solid Fuels," Fire Safety
Journal, Vol. 3, 1981, pp. 185-200.
5
Fernandez-Pello, A. C., and Hirano, T., "Controlling Mechanisms of Fire Spread,"
6
Olson, S. L., Ferkul, P. V., and T'ien, J. S., "Near-Limit Flame Spread Over
a Thin Solid Fuel in Microgravity," Twenty-Second Symposium (International) on
7
Sirignano, W. A., "A Critical Discussion of Theories of Flame Spread Across
Solid and Liquid Fuels," Combustion Science and Technology, Vol. 6, 1972, pp.
95-105.
8
Williams, F. A., "Mechanisms of Fire Spread," Sixteenth Symposium (Inter-
national) on Combustion, The Combustion Institute, Pittsburgh, PA, 1976, pp. 1281-
1294.
9
Fernandez-Pello, A. C., "Flame Spread Modeling," Combustion Science and
Technology, Vol. 39, 1984, pp. 119-134.
10
Kindelan, M., and Williams, F. A., "Theory for Endothermic Gasification of
a Solid by a Constant Energy Flux," Combustion Science and Technology, Vol.
10, 1975, pp. 1-19.
H
Kindelan, M., and Williams, F. A., "Gas Phase Ignition of a Solid with In-
depth Absorption of Radiation," Combustion Science and Technology, Vol. 16,
1977, pp. 47-58.
12
Loh, H. T., and Fernandez-Pello, A. C., "A Study of the Controlling Mech-
anisms of Flow Assisted Flame Spread," Twentieth Symposium (International) on
13
Di Blasi, C., Crescitelli, S., and Russo, G., "A Finite Difference Approach
of Navier-Stokes Equations for Heterogeneous Combustion," Proceedings of Eu-
ropean Conference on Computer Applications in the Chemical Industry, Dechema
Monographs, Germany, 1989, pp. 357-362.
14
Mao, C. P., Kodama, H., and Fernandez-Pello, A. C., "Convective Structure
of a Diffusion Flame over a Flat Combustible Surface," Combustion and Flame,
Vol. 57, 1984, pp. 209-236.
15
Chen, C. H., and T'ien, J. S., "Diffusion Flame Stabilization at the Leading
Edge of a Fuel Plate," Combustion Science and Technology, Vol. 50, 1986, pp.
283-306.
16
Kodama, H., Miyasaka, K., and Fernandez-Pello, A. C., "Extinction and
Stabilization of a Diffusion Flame on a Flat Combustible Surface with Emphasis
on Thermal Controlling Mechanisms," Combustion Science and Technology, Vol.
54, 1987, pp. 37-50.
17
Di Blasi, C., Crescitelli, S., and Russo, G., "Computer Simulation of Detailed
Processes Occurring at Near Extinction Conditions of Flame Spread Over Solid
Fuels," Lecture Notes in Physics: Numerical Combustion, Vol. 351, edited by A.
Dervieux and B. Larroutorou, Springer-Verlag, Heidelberg, Germany, 1989, pp.
233-244.
18
Frey, A. E., and T'ien, J. S., "A Theory of Flame Spread Over a Solid Fuel
Including Finite-Rate Chemical Kinetics," Combustion and Flame, Vol. 36, 1979,
pp. 263-289.
19
Borgeson, R. A., and T'ien, J. S., "Modeling of the Fuel Temperature Effect
on Flame Spread Limits in Opposed Flow," Combustion Science and Technology,
Vol. 32, 1983, pp. 125-136.
20
Di Blasi, C., and Continillo, G., "Numerical Solution of a Two-Dimensional
Unsteady Mathematical Model for Flame Spread over Solid Fuels," Proceedings
of Chemical Engineering Fundamentals XVIII Congress: The Use of Computers in
Chemical Engineering, Ainic, Rome, 1987, pp. 261-266.
21
Di Blasi, C., Continillo, C., Crescitelli, S., and Russo, G., "Numerical Sim-
ulation of Opposed Flow Flame Spread Over a Thermally Thick Solid Fuel,"
22
Di Blasi, C., Crescitelli, S., Russo, G., and Fernandez-Pello, A. C., "Predic-
tions of the Dependence on the Opposed Flow Characteristics of the Flame Spread
Rate Over Thick Solid Fuel," Proceedings of Second International Symposium on
Fire Safety Science, Hemisphere, New York, 1988, pp. 119-128.
23
Di Blasi, C., Crescitelli, S., and Russo, G., "Numerical Modelling of Flow
Assisted Flame Spread." Computer Methods in Applied Mechanics and Engineer-
ing, Vol. 75, 1989, pp. 481-492.
24
Di Blasi, C., Crescitelli, S., and Russo, G., "Near Limit Flame Spread Over
Thick Fuels in a Concurrent Forced Flow," Combustion and Flame, Vol. 72, 1988,
pp. 205-215.
25
Di Blasi, C., Crescitelli, S., Russo, G., and Fernandez-Pello, A. C., "Model
of the Flow Assisted Spread of Flames Over a Thin Charring Combustible," Twenty-
Second Symposium (International) on Combustion, The Combustion Institute,
Pittsburgh, PA, 1988, pp. 1205-1212.
26
Di Blasi, C., Crescitelli, S., Russo, G., and Fernandez-Pello, A. C., "On the
Influence of the Gas Velocity Profile on the Theoretically Predicted Opposed Flow
Flame Spread," Combustion Science and Technology, Vol. 64, 1989, pp. 289-294.
27
Wichman, I. S., "A Model Describing the Steady-State Gasification of Bubble-
Forming Thermoplastics in Response to an Incident Heat Flux," Combustion and
Flame, Vol. 63, 1986, pp. 217-229.
28
Kansa, E. J., Perlee, H. E., and Chaiken, R. F., "Mathematical Model of
Wood Pyrolysis Including Internal Forced Convection," Combustion and Flame,
Vol. 29, 1977, pp. 311-324.
29
Patankar, S. V., Numerical Heat Transfer and Fluid Flow, Hemisphere, Lon-
don, England, 1980, pp. 113-137.
30
Issa, R. I., "Solution of the Implicitly Discretised Fluid Flow Equations by
Operator-Splitting," Journal of Computational Physics, Vol. 62, 1985, pp. 40-65.
C. Dl BLASI 671
31
Issa, R. I., Gosman, A. D., and Watkins, A. P., "The Computation of Com-
pressible and Incompressible Recirculating Flows by a Noniterative Implicit Scheme,"
32
Kashiwagi, T., and Summerfield, M., "Ignition and Flame Spreading over a
Solid Fuel: Non-similar Theory of Hot Oxidizing Boundary Layer," Fourteenth
PA, 1973, pp. 1235-1247.
33
Kumar, R. K., and Hermance, C. E., "Gas Phase Ignition Theory of a Het-
erogeneous Solid Propellant Exposed to a Hot Oxidizing Gas," Combustion Science
and Technology, Vol. 4, 1972, pp. 191-196.
34
Amos, B., and Fernandez-Pello, A. C., "Model of the Ignition and Flame
Development on a Vaporizing Combustible Surface in a Stagnation Point Flow:
Ignition by Vapor Fuel Radiation Absorption," Combustion Science and Tech-
nology, Vol. 62, 1988, pp. 331-343.
35
Di Blasi, C., Crescitelli, S., Russo, G., and Cinque, G., "Modellistica al cal-
colatore dell'ignizione di combustibili solidi," Proceedings of Journees Franco-
Italiennes sur les Flammes, Centro Grafico Linate, S. Donato Milanese, 1989, pp.
231-246.
36
Di Blasi, C., Crescitelli, S., Russo, G., and Cinque, G., "Numerical Model of
Ignition Processes of Polymeric Materials Including Gas Phase Absorption of Ra-
diation," Combustion and Flame, 1990 (to be published).
37
Kashiwagi, T., "Effects of Sample Orientation on Radiative Ignition," Com-
bustion and Flame, Vol. 44, 1982, pp. 223-245.
38
Fernandez-Pello, A. C., Ray, S. R., and Glassman, I., "Flame Spread in an
Opposed Forced Flow: The Effect of Ambient Oxygen Concentration," Eighteenth
PA, 1981, pp. 579-589.
39
Loh, H. T., and Fernandez-Pello, A. C., "Flow Assisted Flame Spread Over
Thermally Thin Fuels," Proceedings of First International Symposium on Fire Safety
Science, Hemisphere, New York, 1986, pp. 65-74.
Chapter 22
Pulse Combustor Dynamics: A Numerical Study
Pamela K. Barr and Harry A. Dwyer
Nomenclature
a = sound speed
A = cross-sectional area of cell face
Aw — wall area (for friction and heat transfer)
cp = specific heat at constant pressure
cv = specific heat at constant volume
e = cvT + U2/2
f = friction factor
h = heat-transfer coefficient
ml = inertial fluid mass in valve
n = unit vector normal to cell face
p = pressure
Patm — atmospheric pressure
qw = local wall heat flux
Q = chemical energy release normalized by mean value
SL = chemical energy release per unit volume
R = ideal gas constant
s = entropy
t = time
T — temperature
Aflame = adiabatic flame temperature
Tw = wall temperature
U = velocity
t±Aw = cell wall area (for heat transfer and friction)
t±t = time-step size
A* = grid spacing
p = density
T = cycle time normalized by frequency
TU, = wall shear stress
673
674 PULSE COMBUSTOR DYNAMICS: A NUMERICAL STUDY
Superscripts and Subscripts

i = cell index
n — time-step number
( « + ! ) ' = predicted solution at the n + 1 time-step number
R = reservoir conditions, used in valve model
S = surface area of cell face
y = value at cell with smaller cross-sectional area
v = valve parameters
V = cell volume
I. Introduction
M OST types of combustion-driven heating devices have experienced

combustion instabilities.1 These instabilities can cause serious prob-
lems, such as mechanical vibrations, noise, enhanced corrosion, and un-
predictable performance. Because of the associated problems, most systems
are designed to prevent the combustion process from setting up large-scale
oscillations. However, some furnaces and boilers are designed to take
advantage of a periodic combustion process and the related phenomena
accompanying combustion oscillations. These systems are referred to as
pulse combustors. Putnam et al.2 point out that the choice of terminology,
calling the instabilities either ''combustion-driven oscillations" or "pulse
combustion," depends on the application.
The advantages of pulse combustors are numerous. The heat transfer is
enhanced by the large oscillations in the flowfield. Dec and Keller3 have
measured heat-transfer rates up to two and a half times greater than those
obtained for steady turbulent flows at the same mean Reynolds numbers.
The enhanced heat transfer means that a smaller furnace can be used to
provide the same energy output. In addition, pulse combustors do not
require additional equipment such as external blowers. They have few, if
any, moving parts. The pressure oscillations enable them to draw in their
own supply of fuel and air, and the resulting thrust is used to vent the
exhaust products. The maintenance expenses are low because there are
few movable parts. Another advantage of gas-fired pulse combustors is
the reduced NO^. level in the exhaust gas obtained without additional
pollution control equipment. The main disadvantages of pulse combustors
are caused by the large oscillations, which result in noise and vibrational
problems. However, the combustor's physical structure can be designed
to reduce these problems.
A variety of pulse combustors have been developed to take advantage
of unsteady combustion, but because of the highly coupled interactions of
the controlling mechanisms, pulse combustion systems have been difficult
to design and scale. Zinn4 and Putnam et al.2 present historical reviews of
the development and application of pulse combustors. Although the phe-
nomenon of combustion-driven oscillations first was reported by Higgens
in 1777 (see Ref. 5), a hundred years lapsed until Rayleigh6 explained the
mechanism. He concluded that, for the oscillations to be stable, the energy
from the combustion process must be added at a time in the cycle when
it adds energy to the acoustic wave. For the pressure oscillations to con-
P. K. BARR AND H. A. DWYER 675
tinue, the amount of energy added must compensate for the losses in the
system.
Several general designs of pulse combustors exist, all of which use the
periodic addition of energy to drive the resonant pressure wave. The class
under consideration in this chapter is the Helmholtz-type pulse combustor.
It consists of a closed cylinder, which acts as the combustion chamber, that
is attached to a long tail pipe through a small hole or a transition section
(see Fig. 1). Fresh charge is introduced into the combustion chamber through
either a flapper valve or an aerodynamic valve. Both types of valves allow
the gas to flow in one direction; the flow in the other direction is blocked
by a mechanical flapper in the flapper valve, and it is resisted strongly by
the geometry of the aerodynamic valve. The exhaust decoupler shown in
the figure is not needed to support the oscillations, rather it is used in
commercially available systems to terminate the acoustic wave, suppressing
sound emission.
Helmholtz-type pulse combustors act on a principle similar to Helmholtz
resonators. The inertia of the mass in the tail pipe is driven by the expansion
and compression of the mass in the reservoir. In the case of the pulse
combustor, the expansion of the gas in the combustion chamber is caused
by the combustion of the fresh reactant. This increases the pressure, send-
ing a compression wave down the tail pipe. This compression wave reflects
off the tail-pipe exit as an expansion wave. When this wave reaches the
closed end of the combustion chamber, it is reflected as another expansion
wave, causing the pressure in this chamber to drop below the reactant
supply pressure. When the expansion wave reaches the tail-pipe exit, it is
reflected off the open end as a compression wave. The low pressure in the
combustion chamber allows the flapper valve to open, and pulls fresh
reactant into the combustion chamber. The cold reactant mixes with hot
products from the previous cycles, and reignition occurs. For the combus-
tion to reinforce the pressure oscillation, the reignition and combustion
need to be in phase with the compression wave as it returns to the com-
bustion chamber, because both act to raise the pressure. The increase in
pressure closes the flapper valve and again sends a compression wave down
the tail pipe, and the cycle repeats itself.
Because the behavior of Helmholtz resonators depends on the relative
timing of the acoustic waves and the injection and combustion processes,
the design and scaling of these systems are difficult. Pulse combustion
systems for new applications currently are designed through trial and error,
or "cut and try." Design handbooks, such as can be found in Ref. 7, provide
a systematic approach to the design of a Helmholtz-type pulse combustor
for a particular application. However, they do not provide an understand-
ing of the fundamental mechanisms responsible for pulse combustor op-
eration.
Several numerical models have been developed to simulate the dynamics
of Helmholtz-type pulse combustors. A simple model based on the method
of characteristics was developed by Craigen8 to track the wave dynamics
in the tail pipe. Ponizy and Wojcicki9 combined the method of character-
istics with a well-stirred reactor model to simulate the combustion process.
Tsujimoto and Machii10 combined lumped-parameter models with the method
Injection Mixing + Chemistry Wave Dynamics Heat Transfer + Friction Sound Suppression + Additional Heat Transfer C/)
m
o
O
oo
c
o
JJ
o
o
C/)
combustion transition tail pipe exhaust
chamber exhaust pipe
section decoupler
-10x7.5- 10- 200x3 -20xd d e - 50x3 m

5
o
Fig. 1 Schematic of a Helmholtz-type pulse combustor, showing dominant processes that occur in each region. Geometry
consists of a combustion chamber, tail pipe, and decoupling chamber. (All dimensions are in centimeters.)
o
of characteristics to simulate the behavior in more complex pulse combustor

geometries, including both exhaust decoupler volumes and combustion
chambers. Lee11 combined lumped-parameter models for the large volumes
with finite-element models for the wave dynamics in both the tail pipe and
exhaust pipe. Winiarski 12 developed a one-dimensional control volume
model to simulate the dynamics in both the combustion chamber and the
tail pipe.
Pulse combustor dynamics have also been simulated with a one-dimen-
sional finite-difference method that includes losses due to wall friction and
heat transfer.13 However, this model required experimental measurements
of the cyclic energy-release and injection processes as input data. In the
current work, the finite-difference model is extended to include prediction
of the flow through the flapper valve. The sensitivity to the assumed energy-
release profile is also investigated, and the influence of an exhaust decou-
pler is included in the finite-difference model.
II. Problem Formulation, Basic Equations, and Numerical Method

A. Governing Equations
The approach taken in the present research study is to simulate numer-
ically the one-dimensional, unsteady equations of gasdynamics. Since most
applications in a pulse combustor involve heat transfer, wall friction, and
energy release from combustion processes, the standard form of the equa-
tions has been modified to include these influences. A typical control
volume used in the analysis is shown in Fig. 2, and the equations are
Continuity:
X momentum:
I fff pUdV + It pUU-ndA = -nH pdA - 1 1 TH, <L4B. (2)

Ol J J J V J JS J JS J J iv
Energy:
*V + J/u, 4* dA» (3)
where e = cvT + U2/2.

Equation of state:
P = ?RT (4)
In these equations, p, p, T, and U represent the thermodynamic pressure,
density, temperature, and gas velocity, respectively. The local duct cross-
sectional area is A, and the unit vector normal to it is n. The local duct
wall surface area is Aw, the wall shear stress is T W , and the local wall heat
flux is qw. 2, is the chemical energy release per unit volume; R and cv are
the ideal gas constant and the specific heat at constant volume, respectively.
Fig. 2 Control volume used in the numerical analysis.
The subscripts V, 5, and w indicate integration over the cell volume, the
cell face surface area, and the duct wall area, respectively.
B. Numerical Method of Solution

The mathematical nature of the basic system of one-dimensional equa-
tions given earlier is classified as hyperbolic, and the equations allow for
the propagation of pressure and shock waves through a duct. The method
of characteristics, which has been used to simulate the dynamics of pulse
combustors, is found by simplifying the preceding equations. If the influ-
ences of duct friction, wall heat transfer, chemical energy release, and
variable duct area are neglected, the characteristic equations that describe
the movement of disturbances in the gas become14
dp dt/ n
—- ± pa —— = 0 on (5a)
dt dr T-"
dt
dx
— = 0n on — = U ± a (5b)
dt
where s is the local gas entropy and a2 = (dp/dp)s =consl. The method of
characteristics has been used successfully for relatively simple flows, but
the isentropic assumption is restrictive. For applications where heat transfer
and friction change the entropy of the gas, as occurs in pulse combustors,
the method of characteristics is neither general nor flexible enough to
obtain detailed solutions, especially if contact discontinuities are present.
In the present research study, a numerical solution method has been de-
veloped and applied to solve the governing equations.
There are many possible numerical methods available to solve the pre-
ceding system of equations. The method developed by MacCormack15
remains one of the best for problems such as the wave dynamics of pulse
combustion systems. This method is an explicit technique, and the time
step chosen to integrate the equations is related closely to the local char-
acteristic wave speed. MacCormack's method is second-order accurate and
has a reasonable amount of dispersion. With the addition of artificial vis-
cosity or some form of monotone flux limiter, it is capable of good reso-
lution of shocks and contact surfaces. Its major advantages are its simplicity
and computational efficiency. Incorporation of complex physics, chemistry,
and multiphase models and phenomenologies are straightforward. In ad-

dition, complex geometries and realistic boundary condition models are
easier to implement in MacCormack's method than in many other methods.
The use of the control-volume formulation of the governing equations
allows for a clear physical interpretation of the numerical modeling of each
of the terms. This can be important because the original application for
which MacCormack's method was designed did not include the compli-
cating influences typically encountered in a pulse combustor. The basic
MacCormack's scheme is a two-step method, a predictor step followed by
corrector formulas. The predictor step for Eqs. (1-3) evaluated at a typical
computational cell i is
Continuity:
(6)
Momentum:
2
A)?+1 - (pU2A)?]
);Vi - (PA)?] + i[p?+1 + p?][Ai+1 - Af] - cAAv, (?)
Energy:
+ p}UA)1+l - ({pe + p}UA)>;\ = &M/A* + q»v£AWi

where the superscript n denotes values at the current time level and (n +
1)' indicates the predicted variables at the forward time level. The index
/ + 1 denotes values in the cell to the right of the cell under consideration,
as shown in Fig. 2. The equation of state is applied directly to determine
the pressure at (n + 1)'. The corrector step makes use of predicted values,
and the resulting equations are
Continuity:
(9)
Momentum:
Energy:
[(
+ l)
+ [({pe + p}VA^' ' ~ ({pe + p}UAW_\»'} (11)
where the superscript n + 1 marks new variable values at the next time
level, and the subscript / - 1 denotes the variables to the left of the cell
being studied. Again, the equation of state is used as the fourth equation
to obtain values of p, (7, 7\ and p at the new time level.
The preceding formulation indicates that wave and flow influences are
brought in alternately from the left and right sides of the cell, and a com-
bination of values from the previous time level and predictions of new
variables are utilized during the time step. A stability and accuracy analysis
of the method is beyond the scope of this work, but it has been shown
that the method is second-order accurate in time, 15 and the time-step size
is restricted by
A r < A*/(|t/| + a) (12)
For most calculations, a time step slightly smaller than the above value
should be utilized to ensure that the above inequality is satisfied during
both the predictor and corrector steps. Thus, very small cells should be
avoided because of the small time steps that they impose.
C. Friction, Heat-Transfer, and Energy-Release Models

The major additional numerical modeling issues introduced into the
gasdynamic equations by a Helmholtz-type pulse combustor are injection
of fuel and air into the combustion chamber, wall friction due to viscous
stresses, heat losses due to the cooling of the duct walls, and energy release
due to chemical reactions in the combustion chamber. The injection of fuel
and air into the combustion chamber is treated as a boundary condition
in the present analysis; therefore, it is not necessary to include this effect
explicitly in the flow equations. The other three terms have been evaluated
at cell centers, since they do not directly involve wave dynamics. The
influences of wall friction, wall heat transfer, and chemical energy release
involve multidimensional viscous and heat-transfer effects, as well as chem-
ical kinetics, and these can be introduced only in an approximate way in
a one-dimensional analysis. In the present model, both quasisteady ap-
proximations and experimental measurements from similar pulse combus-
tion systems have been used to supply parametric expressions for these
physical processes.
The wall friction and heat transfer have been represented in terms of
the standard quasisteady forms of the friction factor / and the film coef-
ficient h. Expressed in terms of these coefficients, the wall friction and
heat transfer become
qw = h(T - TJ (14)
where /and h are held constant over a pulse cycle. The friction coefficient
/ was estimated from the mean flow conditions16; however, there is con-
siderable uncertainty in the values employed. This lack of confidence comes
from the unsteadiness of the viscous flow near the wall, which can be out
of phase with the mean flow £/used in the one-dimensional analysis. 17 This
is because the fluid near the walls, which has less momentum than that
near the center, is more responsive to changes in the pressure field. How-
ever, these cycle-resolved results cannot be incorporated into a model that
does not examine the flow near the wall because they are incompatible
with the implicit assumptions of one-dimensional flow. The different phas-
ing between the mean flow and the momentum losses results in momentum
being added to the flow by the viscous losses when the losses are out of
phase with the one-dimensional velocity field.
The estimate of the overall heat-transfer coefficient h includes the effects
of forced convection on the inside of the duct, free convection on the
outside, and radiation losses from the duct walls. 13 - 18 The largest contri-
bution to h is the loss due to free convection from the outer duct wall.
Because the external atmosphere is quiescent, the outer free convection
flow is steady during a pulse cycle, and a quasisteady assumption is valid.
The external temperature Tw is assumed to remain constant at 300 K.
The release of energy due to chemical reactions is represented by the
term 12, in the energy equation. 2, is a function of the dynamic flow rate of
the premixed reactants, the mixing of cold and hot gases, and chemical
kinetics. To simulate the dynamic energy release occurring in a Helmholtz-
type pulse combustor, the model would have to predict the mixing of the
injected cold reactants with the hot products before the chemical reactions
can occur. In pulse combustors, these reactants are injected in transient
jets, whose behavior is influenced strongly by the local geometry of the
system. An accurate representation of these highly coupled processes is
well beyond the scope of one-dimensional modeling. In this work, the
cycle-resolved energy release is prespecified; the shape has been guided
by previous experimental results.19 The time dependence of this term, and
the sensitivity of the results to this profile, will be discussed in the Results
section of this chapter.
D. Boundary Conditions—Basic Inflow and Outflow

To close the system of equations at the end cells in both the combustion
chamber and tail pipe, boundary conditions must be imposed. At these
boundary locations, the fluxes or variables must be specified in a consistent
way and in a manner that reproduces the correct physical behavior. Because
the predictor-corrector algorithm uses both forward and backward differ-
encing to evaluate spatial derivatives, boundary conditions must be spec-
ified at both the combustion chamber inlet and the tail pipe exit.
The boundary conditions utilized at the entrance of the combustion
chamber are a function of time during a single pulse cycle. This is because
of the unsteady introduction of premixed fuel and air through a flapper
valve. As the pressure in the combustion chamber decreases and increases,
the flapper valve opens and closes, and an unsteady inflow is established.
When the valve is closed, the physical situation is simple and "reflection"
boundary conditions are used. The reflection conditions impose a zero
velocity at the end wall, and the gradients of all thermodynamic variables
are zero (see Table 1 for the closed valve conditions). As the valve opens,
other conditions must be employed because fresh fuel and air begin to flow
into the combustion chamber. (The calculation of the inlet flow magnitudes
is described in a following section.) Because the valves are flapper valves
in many of the Helmholtz-type pulse combustors, there is total blockage
of the dynamic pressure waves and the use of a pressure reflection condition
(dp/dx = 0) is appropriate. However, the inlet mass-flow rate and the other
thermodynamic variables must represent conditions upstream of the com-
bustion chamber. A consistent boundary condition is achieved by specifying
the inlet gas temperature and velocity, and determining the density from
the combustion-chamber pressure and the equation of state. The conditions
during injection are also given in Table 1.
The conditions employed for the tail pipe or outflow section of a pulse
combustor depend strongly on whether an exhaust decoupler, or sound
reduction device, is used on the system. When the decoupler volume is
large, or when the pulse combustor exhausts to a free atmosphere, the
boundary conditions are uncomplicated, and the classical outflow condi-
tions can be employed. The outflow conditions at the tail-pipe exit assume
that the time scale for equilibration between the external atmosphere and
the tail-pipe exit is short, and the tail-pipe exit pressure is assumed to equal
the surrounding ambient pressure. The outflow boundary conditions for
the other variables are found by setting the derivatives of velocity and
density to zero, and determining the temperature from the equation of
state (see Table 1).
If the low-pressure part of the pulse combustor wave generates an inflow
or reversed velocity at the tail-pipe exit, the surrounding atmosphere (or
the large decoupler) acts as a stagnation chamber from which gas is drawn
into the tail pipe. Under these conditions, the tail-pipe inflow is that from
a stagnation chamber with the Mach number determined from the local
conditions at the tail-pipe exit. When the inflow Mach number is small, as
is the case for pulse combustors under consideration here, the constant-
pressure boundary condition yields the same results as the stagnation con-
ditions.
Table 1 Boundary conditions
Combustion-chamber entrance
Tail-pipe exit
During injection: Valve closed: Entire cycle:
*= ° to = ° *= p -
d d
P n P n
T=TR78 ™ =0 ^ = 0
a* a^
m(r) 5w
w = u(t\ = —^- u =0 — = 0
pA dx
E. Valve Dynamics
In many pulse combustor systems, the fresh charges of fuel and air enter
into the combustion chamber through a valve that opens and closes in
response to the pressure oscillations in the system. The lumped-parameter
valve model used in the present study is based on a relatively simple
momentum balance across the flapper valve. It uses the time-dependent
pressure difference across the valve, the difference between the combus-
tion-chamber pressure and the upstream reservoir pressure (assumed to
be atmospheric), to accelerate an inertial mass representing the fluid con-
tained in the valve. Flow through the valve is resisted by frictional losses.
The dynamic equation for the flow velocity through the valve is
m, = (pR - Pl)At. - PlA/, (15)
where pR and pl are the upstream reservoir pressure and the pressure at
the first node of the combustion chamber, respectively; m/ represents the
inertial fluid mass; pt, is the fluid density in both the valve and the upstream
reservoir; and Uv is the flow velocity in the valve. The effective flow area
in the valve is given by Av, and/^, is the friction loss coefficient. The pressure
difference across the valve is set to zero whenever the combustion-chamber
pressure is greater than the reservoir pressure, indicating that under these
conditions the flapper blocks flow through the valve. Equation (15) is
solved using a simple explicit Euler method at each time step during the
computation. Because the response time of the fluid in the valve is usually
much longer than the acoustic time scale of the combustion gases, there
is little difficulty in obtaining an accurate solution.
The dynamics of the valve are controlled in part by the parameters ml
and/ v in the valve equation. These parameters have to be determined from
numerical experiments on the valve, and in the present studies this semi-
empirical procedure has been carried out. The performance of the previ-
ously-described valve has been good, and the valve opening and closing
characteristics compare favorably with experimentally measured flows. An
improvement to the valve model could include the dynamic behavior of
the gas in the lines connecting the combustion chamber, the flapper valve,
and the air and fuel reservoirs.
F. Models for Industrial Furnace Decouplers

An important component of most pulse combustion systems is an exhaust
decoupler added to the tail-pipe exit (see Fig. 1) to reduce sound emission
by terminating the acoustic wave before the exhaust is vented to the at-
mosphere. The decoupling chamber provides a large increase in the cross-
sectional area at the tail-pipe exit. This increase causes the wave motion
to be reduced drastically, and the flow out of the exhaust pipe is no longer
characterized by strong pulsations. Because a sudden change in cross-
sectional area produces large velocity gradients, large area discontinuities
must be handled in a special manner. The direct use of MacCormack's
algorithm at these locations produces serious errors and instabilities. Spe-
cial control volumes are utilized in the present work to obtain solutions
that have physical validity. The choice of these special control volumes for
a one-dimensional model is not unique. Because the flow is multidimen-
sional in regions of large area discontinuities, a one-dimensional model
will not be capable of simulating it accurately. Physical reasoning has been
used to deduce appropriate control volumes to approximate the flow as
one dimensional.
The geometries of these control volumes are based on identifying regions
around the area discontinuities that can affect the fluid in the neighboring
cell. That is, in the large cell adjacent to the area jump (both the expansion
and contraction sections are shown in Fig. 3), the pressure pushing against
the vertical wall cannot affect the gas in the smaller neighboring cell. Only
the pressure acting at the opening contributes to the gas motion in the
smaller cell. Likewise, the pressure in the smaller of these two cells does
not push directly against the gas located in the large cell along the far edges
of the adjoining face.
In the modified control-volume equations, the area terms have been
selected to represent these "zones of influence." The numerical differenc-
ing is consistent with MacCormack's predictor-corrector method as pre-
sented in Eqs. (6-11), so the modified differencing equations given later
are used at every time step. For the predictor step, which uses upstream
differencing for the spatial derivatives, evaluation of the equations at the
A:th cells in both drawings in Fig. 3 requires special handling. (Note that
the node labels in each of these drawings were selected to correspond with
the single set of equations presented subsequently.) To form a "good"
control volume, the smaller cross-sectional area is used as the area term
in all inertial and convective terms. That is, the variable Ay is set to the
smaller of Ak and Ak +l. However, because Eqs. (1-3) are in conservative
form, the actual area of cell k, Ak, is used in the time derivative term
regardless of whether it is the large or small cell.
Predictor Step
Continuity:
n+ir
[p( - &][(Akbc)ito] + [(puyk+l - (puyk]A, = o (16)
Momentum:
[(P C/ 2 )2 + 1 - (pU 2 )JflA a
- T'',A/U. (17)
Energy:
[(pe)l"+ 1>' - (pe)"k][(Akkx)/&] + [({pe + p}U)'k +1
- ({pe + p}U)?]Ay = £M*Ax + <,AA,. (18)

Note that the energy-release, friction, and heat-transfer terms are un-
changed.
For the corrector step, which uses downstream differencing to evaluate
the spatial derivatives, evaluation of the equations at the k + 1 cells in

Fig. 3 requires special handling. The solution is again to eliminate the large
area terms from the flux and pressure terms in the basic equations, and
the resulting corrector equations become
Corrector Step
Continuity:
(19)
Momentum:
(20)
Energy:
[({pe &Yr - «pe

(21)
The preceding formulation eliminates the large area in the flux and pressure
terms. In both the predictor and corrector steps of the momentum equation,
the term representing p(dA/dx) that appeared in Eqs. (7) and (10) drops
out because the area derivative is set to zero.
To ensure conservation, the application of these modified expressions
requires that the fluxes to cells adjoining the affected cells also be adjusted.
Expansion Section Contraction Section
k+1 k+2 k+3 k-2 k-1
Fig. 3 Area discontinuity, node indices, and cell face areas used in the modified
finite-difference equations at the entrance and exit of the decoupler.
The appropriate areas to use in the fluxes are shown in the bottom of Fig.
3. In this example, the mass flux at each cell face is presented. The ap-
propriate values of the effective cross-sectional area must be used consis-
tently throughout the evaluation of the continuity, momentum, and energy
equations.
A practical problem that can occur results from the reduced stability of
MacCormack's method when a large contact discontinuity passes or de-
velops in the flow. The regions close to the area discontinuities are espe-
cially sensitive to these steep gradients in flow properties. The contact
discontinuities usually result from the startup of the combustion process,
caused by the contact of the cold initial gas with the high-temperature
products. The problem is controlled by using a small amount of diffusion
or smoothing in the region around the decoupler entrance and exit. The
value of the diffusion coefficient, which is nonzero only around the de-
coupler ends, has been selected from numerical experiments such that the
oscillations are controlled but the wave motion is unaffected.
The preceding algorithm is capable of simulating flow through a discon-
tinuity in a one-dimensional model. The primary deficiency in the preceding
model is the use of one-dimensional analysis at a region of very rapid area
change, and the proper corrections are not known at the present time. It
will be shown that the model equations work well for large-amplitude
pressure pulses in the tail pipe or decoupler, but that excessive pressure
rises can occur in the tail-pipe exit for weak waves in the decoupler. The
primary reason seems to be an over-restriction of the flow between the
tail-pipe exit and the decoupler entrance. A logical solution is to use an
intermediate area between the tail pipe and decoupler. The finite-differ-
ence expressions presented in Eqs. (16-21) have been used at all of the
area discontinuities that appear in the present pulse combustor studies.
Another approach to modeling the behavior of a decoupler is to use a
lumped-parameter model. This approximation should be valid when the
volume of the decoupler is large compared with the pulse combustor vol-
ume. With a lumped-parameter model, the time-dependent pressure in the
decoupler serves as the quasisteady pressure boundary condition for the
tail pipe. It is expected that this type of model would lead to an over-
response of the tail pipe to the decoupler pressure. It is the opinion of the
present investigators that a combination of the two methods eventually
will lead to a more reliable model to simulate the interaction between the
tail pipe and decoupler.
III. Results
A. Simulation of a Pulse Combustor
The pulse combustor model described in this chapter was designed to
simulate two Helmholtz-type pulse combustor configurations. The first
configuration, consisting only of a combustion chamber and a tail pipe, is
often referred to as a "bare bones" pulse combustor. It is used primarily
as a research tool to study the fundamental operation of a Helmholtz-type
pulse combustor, containing the minimum features to operate. The other
pulse combustor configuration includes an exhaust decoupler system to
terminate the acoustic wave (see Fig. 1). Simulations involving the decou-
pling chamber will be discussed in a subsequent section. All other simu-
lations are for the bare bones configuration. In this numerical study, the
pulse combustor has a square cross section, simulating an existing design
used for experimental verification of the model. The dimensions of both
configurations are given in Fig.'1. One hundred grid cells are used in the
bare bones simulations and 145 are used in that containing the decoupler
system, resulting in the same grid spacing in both geometries.
In this work, the time-dependent energy-release rate is assumed to be
described by a sine curve, as shown in Fig. 4. Here, the normalized energy-
release profile is represented by (), where the mean value has been used
for normalization. This shape compares favorably with experimentally de-
termined profiles obtained in an operating Helmholtz-type pulse combustor
with the same geometry,19 thus it is a reasonable assumption, although
variations in operating conditions produce a more or less peaked energy-
release profile. A sinusoidal Q profile was used in the results presented
here to isolate the effects of other physical processes. As will be shown,
the operating frequency of the pulse combustor is not very sensitive to the
peakedness of the energy-release profile, although it does affect the mag-
nitude of the oscillation. Previous simulations used an experimentally ob-
tained energy-release profile to drive the wave dynamics. These results
produced good agreement with experimental measurements for pulsation
strength and operating frequency. 13
The model predicts the ''optimal design frequency," defined here as the
frequency that produces the strongest pulsations. The optimal design fre-
quency is determined primarily by acoustic resonance. The time scale of
the energy-release profile Q(f) is set by the prespecified forcing frequency
(the frequency at which the energy release is added to the gas in the
combustion chamber). In the model, the system is driven over a range of
prespecified forcing frequencies, rather than allowing the model to iterate
on the operating frequency, to provide information on the sensitivity of
the system to the operating conditions. The numerical algorithm could be
modified to use the pressure oscillations to determine the operating fre-
2-
0
Cycle Time
Fig. 4 Normalized sinusoidal energy-release profile Q, assumed in the numerical
simulation, as a function of the normalized cycle time.
quency, and the algorithm would converge on the same optimal design
frequency. However, this automated algorithm would not provide the ad-
ditional sensitivity information.
The use of a prespecified shape of the energy-release profile, regardless
of the operating frequency, assumes that in a physical system the com-
bustion process can be tuned to produce the desired frequency, either by
changing the mixing characteristics of the combustor or by modifying the
chemical kinetics of the reactants. For systems where both the fluid mixing
and the chemical kinetics are fixed, the energy-release profile cannot be
adjusted to a particular frequency. In this case, the operating frequency
of the pulse combustor is affected by the time scale between injection and
reaction, in addition to the time scale associated with acoustic resonance.
Keller et al. 20*21 discuss the importance of these characteristic times in
determining pulse combustor performance. Bramlette 22 has developed a
model based on quasisteady jet mixing that predicts the energy release
based on the rate at which the injected gas can mix with the hot products
present in the combustion chamber to reach a temperature at which it can
react. In this model, Bramlette calculates the characteristic time for mixing
as a function of the system geometry and operating characteristics (see
Ref. 22 for a discussion of the model). Barr et al. 23 and Keller et al.24 have
combined Bramlette's jet mixing algorithm with a wave dynamic algorithm
to produce a model that couples the mass injection with both the energy
release and the wave dynamics of a pulse combustor. To focus on the
importance of acoustic resonance in the performance of pulse combustors,
the results presented here use the operating frequency to determine the
characteristic time for reaction to occur.
In the simulations presented here, the fuel mixture is assumed to be
premixed methane-air at an equivalence ratio cj> of 0.8. The assumption of
injecting the reactants premixed, rather than injecting the fuel and air
separately, is necessary because a one-dimensional model is not capable
of simulating the mixing process. Current commercially available pulse
furnaces have utilized two flapper valve systems for safety reasons to avoid
storing a large reservoir of combustible gas. Recent experimental work by
Keller et al. 20-21 has shown that the pulse combustor operates much more
stably by premixing the fuel and air, producing repeatable pulsations as
indicated by their strength and frequency.
In the model, the combustion-generated energy release Q is distributed
uniformly within the combustion chamber, although the Q varies during
a cycle. Because energy is added continually to the gas in the combustion
chamber, the temperature continues to increase as the gas propagates
through this region. The pressure in the system evolves from this forcing
function. The resulting combustion-chamber pressure pulls the gas through
the simulated flapper valve, opening the valve when the pressure difference
across the valve is favorable, as described by Eq. (15).
The system parameters used in this simulation are listed in Table 2. The
forcing frequency is 53 Hz. The adiabatic flame temperature rflame is a
measure of the energy released during combustion of the methane-air
mixture. For simplicity, the equation of state assumes all gas has the same
specific heat cv and gas constant R, which are those for air. In addition,
Table 2 Pulse combustor parameters

Parameter Value
Equivalence ratio c|> = 0.8 -> rflame = 2000 K
Friction / = 0.01
Heat transfer hc = 13 W/(m 2 -K)
hf = 20 W/(m 2 -K)
Valve model At. = 25 cm2
/. = 2000
m, = 0.29 g
the friction factor / and the heat-transfer coefficients for the combustion
chamber and tail pipe, hc and hn remain constant throughout the pulsations,
as was discussed previously. In the transition section, the heat-transfer
coefficient varies linearly between these values.
The pulse combustor simulation is started by repeatedly imposing the
sinusoidal energy-release profile on the initially cool system, but the so-
lution is only of interest after a repeatable periodic wave pattern forms.
Because in the initially quiescent system no pressure waves are present,
an initial guess is used for the injection profile. After five cycles, the
assumed injection profile is replaced by the calculated injection from the
lumped-parameter valve model.
The appearance of a repeatable wave requires that the injection profiles
m(t) have also reached a repeatable shape. After 25 cycles using the dy-
namic injection model, the wave pattern is periodic, as shown in Fig. 5 by
the mass-injection rate, the combustion-chamber pressure, and the tail-
pipe exit velocity. The figure shows that it takes many cycles to establish
a periodic waveform. The amplitudes of the pressure and velocity waves
are determined by the interaction of all wave patterns that develop from
the imposed energy-release profile at the prespecified frequency. The wave
amplitudes and shapes result directly from the computation.
The periodic wave pattern cannot occur until the temperature distri-
bution is repeatable, requiring that enough gas be injected and burned to
replace all of the cool gas that initially filled the pulse combustor. De-
pending on the geometry, this can require many time steps: for the results
shown in Fig. 5, the computation required over 39,000 time steps to sim-
ulate these 25 cycles. It is essential that the solution has reached a re-
peatable wave pattern for the results to be meaningful. Although the startup
process is simulated, the energy-release profile is not expected to be ac-
curate: during this period.
The resulting repeatable injection profiles, pressures, and velocities are
shown in Fig. 6, obtained at the same locations as those in Fig. 5. The
pressure variations take on a sinusoidal shape. At the tail-pipe exit, the
simulation predicts a large reverse flow, a condition that has been measured
in experimental pulse combustors. The fact that these curves show a sinu-
soidal shape indicates that for these conditions the resonant wave in the
tail pipe is being driven strongly by the combustion process occurring at a
frequency of 53 Hz.
100
0
Cycle
Fig. 5 Profiles of mass injection, combustion-chamber pressure, and tail-pipe exit
velocity over 25 cycles from the initiation of the simulation with dynamic injection.
The configuration is that shown in Fig. 1 without the decoupling chamber. Frequency
is 53 Hz; other parameters are listed in Table 2.
CUT
2-
c
.o
CD 1-
125
100 H
Q.
75-
100-
0-
-100-
0 0.2 0.4 0.6 0.8 1

Cycle Time
Fig. 6 Profiles of mass injection, combustion-chamber pressure, and tail-pipe exit
velocity over one cycle after a repeatable wave pattern has been established. The
configuration is that shown in Fig. 1 without the decoupling chamber. Frequency
is 53 Hz; other parameters are listed in Table 2.
Figure 7 shows the spatial variation of pressure, velocity, and temper-

ature at a normalized cycle time T of 0.875. The temperature increases
within the combustion chamber because the energy release is distributed
within this section. Loss of energy due to heat transfer accounts for the
gradual decrease in temperature in the transition section and tail pipe.
Although the pressure varies smoothly throughout the system, the velocity
increases dramatically in the transition section as the area decreases.
This solution, which appears to model the wave motion of a pulse com-
bustor, is dependent on the frequency at which it is forced. In the simulation
illustrated here, the forcing frequency was chosen to be 53 Hz, which
happens to be the frequency that produces the strongest pulsations. The
following section shows how this optimal design frequency can be identified
using these numerical results.
B. Rayleigh's Criterion and Unsteady Energy Release

Although in the physical combustor the operating frequency occurs nat-
urally, in the one-dimensional numerical simulations the combustion proc-
ess is forced to repeat at a prespecified frequency. The most stable operating
-combustion
chamber
Fig. 7 Spatial distribution of pressure, velocity, and temperature at a normalized

cycle time T = 0.875 after a repeatable wave pattern has been established. The
horizontal dashed lines indicate (starting with the top figure) atmospheric pressure,
zero velocity, and the adiabatic flame temperature.
frequency for the numerical pulse combustor can be identified by corre-

lating the resulting oscillations in combustion-chamber pressure with the
periodic energy release. When the pressure wave is in phase with the
combustion oscillation, the combustor is stable, as was discussed by Ray-
leigh in 1878.6 That is, the energy released during the combustion process
causes the combustion-chamber pressure to increase. This increase in pres-
sure creates a pressure wave that propagates down the tail pipe and is
reflected off the open end, returning to the combustion chamber as an
expansion wave, causing the pressure to decrease enough to open the
flapper valve. This expansion wave then travels to the end of the tail pipe
and reflects as a compression wave. If both the expansion and compression
waves are in phase with the minimum and maximum in the energy-release
profile, respectively, the combustion process reinforces the wave pattern
and produces the strongest, most stable pulsations. This is commonly re-
ferred to as Rayleigh's criterion. (See Ref. 25 for additional discussion on
energy-driven oscillations.)
For the simulation described previously, the phase relation between the
energy-release profile and the combustion-chamber pressure wave is plotted
in Fig. 8. This curve was computed from the normalized oscillatory_com-
ponents of both the energy-release profile shown in Fig. 4, (Q - Q)IQ,
and the combustion-chamber pressure plotted in Fig. 6, (p - p)lp, where
the overbars represent the mean values (typically, p is greater than p atm ).
As can be seen from Fig. 8, when both the pressure and the energy release
are higher (or lower) than their respective mean values, the energy release
I-or
I ft,
-2
Cycle Time
Fig. 8 Phase relation between the energy-release profile (shown in Fig. 4) and the
combustion-chamber pressure (shown in Fig. 6) as measured by the cyclic variation
of Rp'Q'l(cppQ). The double peak shape indicates that both the expansion and
compression waves are in phase with the sinusoidal energy release.
is reinforcing the pressure wave and the value of the phase relation is
positive. The phase relation becomes zero when either the energy release
or the pressure passes through the mean value.
The cycle integral of the curve presented in Fig. 8 is used to quantify
the cyclic phase relationship. Based on an expression inspired by Kishimoto
(Ref. 16 in Keller and Westbrook19), Keller and Westbrook used the first
law of thermodynamics to determine the fraction of the energy release that
goes into maintaining the oscillations. They defined the Rayleigh efficiency
r\ as
TI = -?^ Jof * (p ~ P)(Q ~ (>)dT = -^ P p'G'dr (22)

cppQ cppQ Jo
where the overbar represents cyclic averages and the primes represent
deviations from these mean values. In addition, T is the cycle time nor-
malized by the frequency, cp the specific heat at constant pressure, and R
the gas constant. To determine the optimal design frequency, the simu-
lation is run over a range of forcing frequencies. The condition that pro-
duces the maximum value of r\ has the strongest pulsations, as will be
shown in the next section.
C. Use of Rayleigh's Criterion to Determine the Optimal Design

Conditions
To determine the optimal design frequency for the pulse combustor, that
which produces the strongest pulsations, the simulation is carried out at
different forcing frequencies, and the value of the Rayleigh efficiency is
computed after the pulsations have become repeatable. Large positive
values of r\ occur when the peaks in the pressure coincide with the peaks
in the energy release. The forcing frequency that produces the maximum
value of T| is defined as the optimal design frequency. The plot in Fig. 9a
shows the variation of r\ with frequency for the case described in the
previous section. Also shown in this figure are plots of the combustion-
chamber pressure vs the nondimensional cycle time (cycle time normalized
by the cycle period) that were obtained at different frequencies.
As shown in Fig. 9, the Rayleigh efficiency peaks at 53 Hz, which is the
forcing frequency used in the previous results. The pressure wave produced
at this frequency is essentially sinusoidal, as can be seen in the top figure,
and has the largest peak-to-peak amplitude of all the pressure waves. The
waves produced at 30 and 64 Hz are not strong enough to pull fresh
reactants through the flapper valve, as shown in Fig. 9b. Here, the mean
mass-flow rate is shown as a function of the forcing frequency. The injection
profiles are also shown in this figure, at the same frequencies as the pressure
plots in Fig. 9a. As expected, the strongest pressure wave produces a strong
injection profile. Good agreement has been obtained between the meas-
ured operating frequency in a physical system and that identified in the
numerical simulation from the maximum in the Rayleigh efficiency (see
Ref. 13). The fact that T| varies so smoothly and yet develops a distinct
maximum indicates that r\ can be used to determine the optimal design
frequency for a pulse combustion system.
a)
CD
30 40 50 60 70
Frequency (Hz)
b)
CD
CD
-^ 30 40 50 60 70
Frequency (Hz)
Fig. 9 a) Rayleigh efficiency r\ as a function of frequency, overlaid with the com-
bustion-chamber pressure waves produced at the indicated frequencies vs normalized
cycle time, b) Mean flow rate as a function of frequency, overlaid with the injection
profiles obtained at the indicated frequencies vs normalized cycle time. All subscales
in both figures are as shown in Fig. 5. The optimal design frequency is identified
from the maximum of if], which also produces the most sinusoidal pressure wave
and a large flow rate.
The preceding type of analysis is a good example of how numerical

simulations can be used to understand the design parameters of a practical
pulse combustion system. In these numerical simulations, the frequency
of the combustion process has been adjusted to determine the optimal
design frequency. In physical systems, the time scales associated with the
combustion process are fixed by the parameters governing the injection,
mixing, and chemical kinetics of the fresh charge. By carefully adjusting
these processes, the combustion process can be tuned to operate closer to
the optimum frequency.
D. Combustion-Chamber Processes
The driving force behind the operation of pulse combustors is the os-
cillatory nature of the combustion-generated energy release. In the pre-
vious simulation, a sinusoidal shape was assumed for <2, and it was assumed
to be distributed uniformly within the combustion chamber. In the follow-
ing simulations, the sensitivity to these assumptions is investigated.
To provide strong reinforcement of the oscillations, Helmholtz-type pulse
combustors are designed so that combustion occurs in the combustion
chamber and significant combustion is not occurring in the gas venting to
the tail pipe. Within the combustion chamber, the location where the
energy is released may vary. In the simulation discussed previously, the
combustion process was assumed to be distributed evenly in the combustion
chamber. Other spatial distributions of the energy release do not signifi-
cantly affect the wave dynamics of the system because the combustion-
chamber length is short relative to the length of the tail pipe. One energy
distribution tested varied linearly within the combustion chamber, with the
maximum energy release occurring at the chamber inlet and linearly de-
creasing to zero at the transition section. The resulting wave pattern is
virtually identical to the previously described spatially uniform result: the
optimal design frequency is 53 Hz and the cycle-resolved pressure and
velocity profiles are indistinguishable from those shown in Fig. 6.
Although the spatial distribution of the energy release in the combustion
chamber does not affect the wave motion, the temporal distribution of
Q has a significant affect. Figure 10 compares the wave dynamics obtained
from four different normalized Q profiles: a sine wave (described previ-
ously), a square wave, a triangular wave, and a pseudodelta function. The
energy-release profiles shown in the figure have been normalized by the
mean values, thus, the integral of each of these curves is 1. Figure 10 shows
the resulting periodic profiles of mass injection, pressure in the combustion
chamber, and velocity at the tail-pipe exit, each displayed at the optimal
design frequency for the specified energy-release profile (see Table 3). As
can be seen from the figure, the sinusoidal profile produces pressure and
velocity waves with the same harmonic content. The square-wave profile
produces waves of a similar amplitude, but they are not smooth like the
waves produced by the sinusoidal profile, indicating that higher frequencies
are present in the waveform. Although the triangular profile contains a
discontinuity, the waves show a smooth shape, with larger peak-to-peak
amplitudes as compared with those generated by either the sine wave or
the square wave. The oscillations generated by the pseudodelta function
sine wave square wave triangular wave delta function
33
ID
0
IE
m
DO
Cycte Time Cycle Time Cycle Time Cycle Time
Fig. 10 Comparison of the wave dynamics obtained from the four different energy-release profiles Q shown in the top row. Pressure
is in the combustion chamber; velocity is at the tail-pipe exit. Unless indicated otherwise, the scales are the same for each variable.
Optimal design frequencies for each of the four conditions shown (as identified by the maximum in the Rayleigh efficiency) are 53, 52,
55, and 54 Hz ± 1 Hz for the energy-release profiles shaped as a sine wave, a square wave, a triangular wave, and a pseudodelta
function, respectively.
Table 3 Sensitivity to energy-release profile

Q sine wave square wave triangular wave delta function
Frequency, Hz 53 52 55 52
m, g/s 3.01 3.29 3.40 3.33
Mmax 2.64 4.05 5.86 8.41
have the largest magnitudes of any of those under comparison. For nu-
merical stability, the pseudodelta function was given a finite thickness. The
resulting pressure wave clearly shows at which time in the cycle the energy
release is being added. As in the results produced by the square wave, the
velocity profile contains components with higher harmonics.
The results presented in Fig. 10 indicate that in order to produce strong
oscillations (as measured by the peak-to-peak amplitude), significant com-
bustion should not occur throughout the cycle, rather the energy should
be released all at once. Furthermore, the time lag between injection and
combustion must be compatible with the acoustic time scale: the flapper
valve only allows gas to be injected when the combustion-chamber pressure
is low, but to reinforce the wave motion, combustion of the injected gas
must not occur until the wave dynamics has increased the pressure. If large
magnitude oscillations are desired, the energy release should be adjusted
to resemble more closely a delta function by modifying the mixing char-
acteristics and changing the chemical kinetics.
E. Tail-Pipe Processes
The variation of the Rayleigh efficiency with the frequency represents
the sensitivity of the pulsation stability to the frequency at which combus-
tion occurs. The smooth variation of r\ with frequency shown in Fig. 9a
indicates that the Rayleigh efficiency is a very powerful tool for quantifying
the performance of the combustor under different conditions. By consid-
ering all of the system parameters, the Rayleigh efficiency r\ can be thought
of as a multidimensional surface whose shape characterizes the sensitivity
of the combustor operation to these parameters. The plot of r\ represents
a slice of that surface, with all parameters except frequency held constant.
By simultaneously varying two parameters, such as frequency and either
friction or heat transfer, a three-dimensional representation of this surface
can be obtained. Although changes in the operating conditions affect the
combustion process by altering the energy-release rate, it is assumed that
the mixing and chemistry can be controlled so that it still has the sinusoidal
shape shown in Fig. 4.
The optimal design frequency of the pulse combustor is highly sensitive
to the amount of heat loss, as indicated in the variation of the Rayleigh
efficiency r\ with the normalized heat-transfer coefficient shown in Fig. 11.
In this figure, the abscissa is the scaling factor for both the tail-pipe and
combustion-chamber heat-transfer coefficients, and so the value hn = I is
the condition listed in Table 1. The shift in the peak toward a higher
frequency for lower heat loss shows that when the mean sound speed of
the gas is changed, either by increasing or decreasing the heat loss, acoustic
resonance occurs at different frequencies. This is consistent with Rayleigh's

criterion because the sound speed controls the relative phase of the pressure
and energy release when the combustion process occurs at a fixed fre-
quency.
Figure 11 summarizes the results from more than 300 simulations. The
sharp drop in the Rayleigh efficiency as the frequency increases from the
optimal design frequency is caused by the sensitivity of the valve perform-
ance to the magnitude of the pulsations. The wiggles in the surface at high
Fig. 11 Variation of the Rayleigh efficiency TJ with heat transfer and frequency.
The surface represents results from 323 simulations. Heat-transfer coefficients are
normalized by the values listed in Table 2. The peak shifts toward higher frequencies
as the heat transfer decreases, indicating that the reduction in heat losses increases
the temperatures and, hence, the sound speed, thus, increasing the resonant fre-
quency of the system. The sharp drop in the surface is produced by the inability of
the pressure wave to open the valve sufficiently. The small bumps near the left
corner (at higher frequencies and higher heat transfer) result when the wave pattern
does not converge to a repeatable form. For reference, the curve in Fig. 9a is a slice
of this surface for the normalized heat-transfer coefficient of 1.
heat transfer and high frequency result from the inability of the pulse
combustor to maintain repeatable pulsations under these conditions.
Figure 12 compares the combustion-chamber pressure oscillations ob-
tained at the optimal design frequencies for conditions of high (hn — 2
and 42 Hz) and low (hn = 0.5 and 62 Hz) heat transfer. In both of these
cases, the pressure waves are sinusoidal in shape, as in Fig. 4. Although
the higher heat transfer removes more energy, the resulting pulsations are
stronger under these conditions than for the lower heat-transfer condition,
as indicated by both the magnitude of the oscillations shown in Fig. 12 and
the respective Rayleigh efficiencies of 3.13 for hn = 2 and 2.35 for hn —
0.5. The stronger pulsations also result in a larger mean mass-flow rate:
3.16 vs 2.92 g/s.
a)
125-
100 H
Q_
hn = 0.5
75-
0
Cycle Time
b)
125-
100 H
Q_
75-
Cycle Time
Fig. 12 Pressure in the combustion chamber obtained at the optimal design fre-
quency for different heat-transfer coefficients: a) hn = 0.5 and 62 Hz, and b) hn =
2 and 42 Hz. These frequencies represent the maximum in the Rayleigh efficiency
for their respective amounts of heat transfer. Note that the pressure waves are of
similar magnitude despite the fact that the available energy differs because the heat
transfer has been changed by a factor of 4.
Unlike the heat transfer, the amount of friction does not significantly
change the optimal design frequency, although it dramatically affects the
pulsation strength. The dramatic increase in the Rayleigh efficiency for
lower friction coefficients, shown in Fig. 13, is caused by reduced damping
of the waves. Careful inspection of the surface in Fig. 13 shows that at
lower friction, the peak shifts toward higher frequencies. This is demon-
strated in Fig. 14, which presents the cyclic pressure in the combustion
chamber at the optimal design frequencies for conditions of no friction (/
= 0 and 56 Hz) and high frictional losses (/ - 0.02 and 50 Hz). The shift
Fig. 13 Variation of the Rayleigh efficiency it] with wall friction and frequency.
The surface represents results from 95 simulations. At low values of friction, the
magnitude of the oscillations increases dramatically, as indicated by the value of the
Rayleigh efficiency. Close observation indicates that the peak shifts toward higher
frequencies as the friction decreases, caused by the increased mass-flow rates caused
by the large pressure difference across the valve. For reference, the curve in Fig.
9a is a slice of this surface for a value of / = 0.01.
a)
220-
160-
Q_
100-
f = 0
40
0
Cycle Time
b)
220-
160-
CL
100-
f = 0.02
40-
Cycle Time
Fig. 14 Pressure in the combustion chamber obtained at the optimal design fre-
quency for different amounts of friction: a)/ = 0 and 56 Hz, and b)/ = 0.02 and
50 Hz. These frequencies represent the maximum in the Rayleigh efficiency for their
respective amounts of friction. Note that the friction significantly affects both the
size and the shape of the oscillation.
in frequency between these two conditions can be explained by the different

mean mass-flow rates. The larger pressure oscillation obtained with no
friction causes more mass to be pulled through the flapper valve than in
the case with large frictional losses, producing mean flow rates of 4.60 and
2.74 g/s, respectively. The larger mass-flow rate, in turn, results in higher
mean velocities throughout the pulse combustor, providing less time for
energy removal and, hence, higher temperatures.
F. Influence of an Exhaust Decoupler

A simple configuration can be used in the design of a Helmholtz-type
pulse combustor that is required only to produce stable pulsations. This
bare bones configuration consists of just a combustion chamber and tail
pipe, and the previously presented results were for such a pulse combustor.
Commercially available pulse furnaces, either industrial boilers or resi-

dential space or water heaters, require the addition of an exhaust decou-
pling system to reduce the sound level by internally terminating the acoustic
waves. The decoupling chamber is simply a large reservoir into which the
pulsating flow from the tail pipe is exhausted (see Fig. 1). The gas is then
vented into the atmosphere through the exhaust pipe. The decoupling
chamber should be large enough so that the conditions within it are fairly
constant in time and are not affected significantly by the pulsating flow
from the tail pipe. This produces a constant flow through the exhaust pipe.
To minimize the noise radiated from the chamber walls, a sturdy construc-
tion of the chamber is also required.
In the numerical simulation, the decoupling system is included by adding
a large decoupling volume and exhaust pipe to the tail-pipe exit, as shown
in Fig. 1. The discontinuities in cross-sectional area require that the finite-
difference equations be modified at the nodes neighboring the expansion
and contraction sections [Eqs. (16-21)]. The volume of the decoupling
chamber affects the performance of the pulse combustor, as shown in Fig.
15 by the sensitivity of the optimal design frequency to the volume of the
decoupler. In this figure, the decoupler volume is normalized by the com-
bustion-chamber volume, although it is the entire system geometry that
affects the pulse combustor performance. The dimensions of the pulse
combustion system are those shown in Fig. 1; the decoupler diameter is
o
00
c:
<b
^j o _
Cr
O
CO
10 20 30 40
v decoupler / Vcombustion chamber

Fig. 15 Variation of frequency with decoupler volume. Pressure waves in the com-
bustion chamber and decoupler volume are out of phase in the upper curve and are
in phase in the lower curve.
varied to change the volume. The system parameters are listed in Table
2. The heat-transfer coefficient in the exhaust pipe is the same as in the
tail pipe and is assumed to be zero in the decoupler.
For a decoupler volume equal to 12.5 times the combustion-chamber
volume, the optimal design frequency is 54 Hz. Figure 16 shows the spatial
variation of pressure and velocity at one instant during the cycle. The
pressure varies smoothly through the system, except at the area discontin-
uities around the decoupler, where MacCormack's method has introduced
oscillations in the solution. The velocity profile contains steep jumps at
these locations, consistent with the conservation of mass at the area dis-
continuity. Throughout the cycle, the pressure in the decoupler is 180 deg
out of phase with the combustion-chamber pressure, as shown in Figs. 16
and 17. Thus, when the combustion-chamber pressure is high and pushing
gas down the tail pipe, the pressure in the decoupler is low, helping to pull
the gas toward it.
The smallest decoupler volume represented in Fig. 15 has the same
diameter as both the tail pipe and the exhaust pipe, and so this system acts
similar to a bare bones configuration with a long tail pipe. The optimal
design frequency predicted with the decoupler model, 39 Hz, is very close
Fig. 16 Spatial distribution of pressure and velocity for configuration with decou-
pler system at a normalized cycle time T = 0.875. Decoupler volume is 12.5 times
the combustion-chamber volume; operating frequency is 54 Hz. The horizontal
dashed lines indicate atmospheric pressure and zero velocity.
125
100-
CL
decoupler
75-
Cycle Time
Fig. 17 Pressures in the combustion chamber and decoupler obtained at the optimal
design frequency (54 Hz) for a decoupler volume 12.5 times the combustion-chamber
volume. Note that the two pressure waves are 180 deg out of phase.
to that predicted with the bare bones model for the same conditions, 40
Hz. The Rayleigh efficiencies are also very similar, within 3% for the two
methods. As the decoupler volume is increased from the minimum value,
both the optimal design frequency and the pulsation strength decrease.
When the decoupler volume is around five times the combustion-chamber
volume, two values for the optimal design frequency are obtained, as
identified by a pair of peaks in the Rayleigh efficiency. The lower frequency
corresponds with the behavior that is observed in a bare bones configu-
ration: a single node in the wave motion located at the exhaust-pipe exit.
The spatial variation of pressure and velocity are shown in Fig. 18 for the
lower frequency. Although the velocity changes discontinuously at the
decoupler entrance and exit, the combustion-chamber pressure is in phase
with the decoupler pressure, as shown in Fig. 19. For this geometry, the
higher-frequency operation corresponds to the pressure in the decoupler
being out of phase with that in the combustion chamber, similar to the
results presented in Figs. 16 and 17. As the decoupler volume is increased
further, the strength of the pulsations at the higher frequencies increases,
and the lower frequency solution is no longer identifiable from the Rayleigh
efficiency.
The switch in the phase relation between the combustion chamber and
the decoupler pressures, and the corresponding shift in frequency, have
been observed experimentally. 7 In the experiment, the exhaust-pipe length
was increased while the decoupler volume remained fixed. When the de-
coupler volume is very large relative to the volume of the bare bones
system, the pressure oscillation in the decoupler is essentially zero. In this
case, the pulse combustor should perform as if the decoupler was not
present. The model shows that as the decoupler size is increased, the
pressure oscillations in the decoupler decrease in comparison to the oscil-
Fig. 18 Spatial distribution of pressure and velocity for configuration with decou-
pler system at a normalized cycle time T = 0.875. Decoupler volume is 4.5 times
the combustion-chamber volume; operating frequency is 35 Hz. The horizontal
dashed lines indicate atmospheric pressure and zero velocity.
lations in the tail pipe. As the decoupler volume is increased, both the
pressure oscillation and the optimal design frequency do not asymptote to
the bare bones values. The results indicate that the flow from the tail pipe
into the decoupler is restricted significantly by the internal boundary con-
ditions at the area discontinuity. This is not a problem for small decouplers,
as verified by two results. First, good agreement was obtained between
the predicted performance of a bare bones pulse combustor (no decoupler
or exhaust pipe) with results for a pulse combustor with an effective tail
pipe that includes a decoupler and a vent pipe of the same diameter as the
tail pipe. In addition, the decoupler model is able to reproduce the ex-
perimentally observed dual operating frequencies for a small range of geo-
metries.

The behavior of a pulse combustor has been simulated by using a finite-
difference method to solve the one-dimensional equations of continuity,
momentum, and energy for a variable-area geometry, along with the per-
fect gas equation of state. MacCormack's method was selected to solve the
125-
Q_
75
Cycle Time
Fig. 19 Pressures in the combustion chamber and decoupler obtained at the optimal
design frequency (35 Hz) for a decoupler volume 4.5 times the combustion-chamber
volume. Note that the two pressure waves are in phase.
equations because it is capable of predicting wave propagation with a

reasonable amount of dispersion. Because it is an explicit technique, the
time step is linked directly to the grid resolution. Typical bare bones sim-
ulations presented here took 20 min on a VAX 8650 computer, an amount
of time that makes it reasonable to consider using the model to perform
design studies. If either higher resolution or additional time steps become
an issue, an implicit integration technique might be preferable to the ex-
plicit one used here.
In this model, the periodic energy-release profile, which represents the
combustion process, is used to drive the acoustic waves in the pulse com-
bustor. Fresh reactants are pulled through the flapper valve when the
pressure in the combustion chamber is lower than the pressure upstream
of the valve. The mass-flow rate is used to determine the net energy release.
The resulting pulsation strength and stability are quantified with the Ray-
leigh efficiency T| , which is an integrated measure of the cyclic phase relation
between the energy release and the combustion-chamber pressure. Strong,
stable pulsations occur for conditions that produce a positive value for TJ.
The optimal design frequency, that which produces the strongest pulsa-
tions, is identified by the frequency resulting in the largest value of r\. The
results show that because this frequency produces large pressure oscilla-
tions, it also produces strong mass flow through the system. The combustion
process can be tuned to this optimal design frequency by adjusting both
the mixing of the cold reactants with the hot products and by modifying
the chemical time scale.
The modified finite-difference equations used for the decoupling cham-
ber were shown to be valid for small to intermediate volumes. The de-
coupler model was able to predict the different phase relations between
the combustion-chamber pressure and the decoupler pressure that have

been observed experimentally. However, the decoupler model needs to be
improved for large volumes, where the decoupler pressure does not vary
enough over each cycle to help pull the gas in the tail pipe. For large
decoupler volumes, the current decoupler model damps the wave dynamics
of the system.
To improve the predictive capabilities of the model, several submodels
should be revised. Detailed heat-transfer measurements have been made
recently over a range of operating conditions in a Helmholtz-type pulse
combustor. 3 When available, correlations of these results with both pul-
sation amplitude and operating frequency should be included in the model.
Complementary experimental work showed that the cycle-resolved wall
shear stress TH, in an operating pulse combustor was out of phase with the
cycle-resolved mean flow. 17 This is because the fluid near the walls, which
has less momentum than that near the center, is more responsive to changes
in the pressure field. However, in the one-dimensional model, the frictional
losses are a function of the mean flowfield. The experimental work of Dec
et al.17 may lead to formulas specifying mean friction losses as a function
of operating conditions. However, the cycle-resolved wall shear-stress re-
sults are incompatible with a one-dimensional assumption.
To understand the complex mixing and combustion processes that occur
in the combustion chamber, a multidimensional model is needed that fo-
cuses on the flowfield in this region. Although experimental work can help
illuminate the fundamental mechanisms that occur here, parameters that
are not controlled easily in the experiment can be varied independently
with a numerical model. Ideally, this multidimensional model would be
developed in cooperation with an experimental study.
The ultimate goal of this pulse combustor model is twofold. In addition
to helping explain the complex interactions that occur in a Helmholtz-type
pulse combustor, the model should be able to assist in the design of pulse
combustors. Both goals have been demonstrated in the work presented
here. Strong coupling between the dominant processes affecting the op-
eration of a Helmholtz-type pulse combustor was demonstrated by the
model. The Rayleigh efficiency was used to quantify the sensitivity of the
pulse combustor performance to changes in the geometry and operating
conditions. The model can be used to identify designs that produce the
strongest, most stable pulsations, and to determine the optimal design
frequency for injection and combustion. Thus, the Rayleigh efficiency can
be used to obtain design information and insight.
Acknowledgments
The authors would like to thank their colleagues, Taz Bramlette, Jay
Keller, and John Dec, for continued support of the modeling work. This
work was supported by the U.S. Department of Energy, Office of Industrial
Processes, Advanced Industrial Concepts Division.
References
Tutnam, A. A., "General Considerations of Autonomous Combustion Oscil-
lations," Nonsteady Flame Propagation, edited by G. H. Markstein, Pergamon,
Oxford, England, 1964.
2
Putnam, A. A., Belles, F. E., and Kentfield, J. A. C, "Pulse Combustion/ 1
3
Dec, J. E., and Keller, J. O., "Pulse Combustor Tail-Pipe Heat-Transfer De-
pendence on Frequency, Amplitude, and Mean Flow Rate," Combustion and Flame,
Vol. 77, 1989, p. 359.
4
Zinn, B. T., "State of the Art and Research Needs of Pulsating Combustion,"
American Society of Mechanical Engineers, Paper 84-WA/NCA-19, 1984.
5
Tyndall, J., Sound, D. Appelton, New York, 1897.
6
Rayleigh, J. W. S., "The Explanation of Certain Acoustical Phenomena," Na-
ture, Vol. 18, 1878, p. 319; also, The Theory of Sound, Vol. 2, Dover, New York,
1945, p. 226.
7
Vishwanath, P. S., "Advancement of Developmental Technology for Pulse
Combustion Applications," Gas Research Institute, Chicago, IL, Rept. GRI-85/
0280, 1985.
8
Craigen, J. G., "Mathematical Model of a Pulsating Combustor," Ph.D. Thesis,
Univ. of Durham, England, 1976.
9
Ponizy, B., and Wojcicki, S., "On Modeling of Pulse Combustors," Twentieth
PA, 1984, p. 2019.
10
Tsujimoto, Y., and Machii, N., "Numerical Analysis of a Pulse Combustor,"
Twenty-First Symposium (International) on Combustion, The Combustion Institute,
Pittsburgh, PA, 1986, p. 539.
H
Lee, J. J.-H., "Computer Simulation of Pulsations in a Gas Fired Pulse Com-
bustion Device and Predictions of Their Exhaust Noise for Single and Dual Com-
bustion Chamber Designs," Ph.D. Thesis, Purdue Univ., Lafayette, IN, 1983.
12
Winiarski, L. D., "A Method of Calculating Non-Steady Compressible Flow
in a Propulsive Duct," Ph.D. Thesis, Oregon State Univ., Corvallis, OR, 1972.
13
Barr, P. K., Dwyer, H. A., and Bramlette, T. T., "A One-Dimensional Model
of a Pulse Combustor," Combustion Science and Technology, Vol. 58, 1988, p.
315.
14
Whitham, G. B., Linear and Nonlinear Waves, Wiley, New York, 1974.
15
MacCormack, R. N., "The Effect of Viscosity on Hypervelocity Impact Cra-
tering," AIAA Paper 69-354, 1969.
16
Blevins, R. D., Applied Fluid Dynamics Handbook, Van Nostrand Reinhold,
New York, 1985.
17
Dec, J. E., Keller, J. O., and Hongo, I., "Time-Resolved Velocities and Tur-
bulence in the Oscillating Flow of a Pulse Combustor Tail Pipe," Combustion and
Flame, 1991, (to be published).
18
Dec, J. E., private communication, Sandia National Labs., Livermore, CA,
1986.
19
Keller, J. O., and Westbrook, C. K., "Response of a Pulse Combustor to
Changes in Fuel Composition," Twenty-First Symposium (International) on Com-
bustion, The Combustion Institute, Pittsburgh, PA, 1986, p. 547.
20
Keller, J. O., Bramlette, T. T., Dec, J. E., and Westbrook, C. K., "Pulse
Combustion: The Importance of Characteristic Times," Combustion and Flame,
Vol. 75, 1989, p. 33.
21
Keller, J. O., Bramlette, T. T., Westbrook, C. K., and Dec, J. E., "Pulse
Combustion: The Quantification of Characteristic Times," Combustion and Flame,
Vol. 79, 1989, p. 151.
22
Bramlette, T. T., "The Role of Fluid Dynamic Mixing in Pulse Combustors,"
Sandia National Labs., Livermore, CA, Rept. SAND87-8622, 1987.
23
Barr, P. K., Keller, J. O., Bramlette, T. T., Westbrook, C. K., and Dec, J.
E., "Pulse Combustor Modeling: Demonstration of the Importance of Character-
istic Times," Combustion and Flame, Vol. 82, 1990, p. 252.
24
Keller, J. O., Barr, P. K., Bramlette, T. T., Evens, L., and Marchant, R. N.,
"Pulse Combustion: Demonstration of the Characteristic Mixing Time in a Com-
mercial Burner," Combustion Science and Technology, Vol. 66, 1989, p. 127.
25
Oran, E. S., and Gardner, J. H., "Chemical-Acoustical Interactions in Com-
bustion Systems," Progress in Energy and Combustion Science, Vol. 11, 1985, p.
253.
Chapter 23
Mathematical Modeling of Enclosure Fires
Henri E. Mitler
I. Preliminaries
A. Definitions
Pyrolysis is the heat decomposition of large molecules, generally in a
solid, into smaller molecules; these exit the solid in the form of gas or
vapor. Pyrolysis is often loosely referred to as burning. A fire is a set of
physical and chemical phenomena, which include combustion (the rapid,
localized exothermic chemical reactions involving an oxidizer, almost al-
ways atmospheric oxygen, which produce high temperatures and luminos-
ity), fluid flows, and (generally) pyrolysis or evaporation. If the combustion
is in the solid phase, as in char oxidation, it is a smoldering fire. When
the combustion occurs in the gas phase, the luminous part of the gas is
called the flame. Thus, fire is the more general phenomenon, which may
or may not involve a flame. Enclosure fire models attempt to describe all
of the important phenomena associated with a fire in an enclosure, such
as the production of various species, fluid flow throughout a building, and
structural damage due to heat. Flame models attempt to describe the struc-
ture, properties, and behavior of the gas-phase combustion zone only (see
Refs. 1-3).
When a fire in an enclosure grows large enough, a number of phenomena
frequently ensue: the temperature of the hot gases in the room rises with
increasing speed to 600°C or more; the accompanying heat fluxes to the
floor rise to 2 W/cm2 or more, and other objects ignite. By the time a
second object ignites, it is usually only a matter of seconds before most of
the flammable materials in the room burst into flame. The smokey layer
descends nearly to the floor, and often flames shoot out of the enclosure.
All or most of these events occur within a few seconds; this sequence of
events is called flashover. After flashover, the room is said to be "fully
involved."
711
712 MATHEMATICAL MODELING OF ENCLOSURE FIRES
This chapter discusses deterministic models of enclosure fires, which are

a subset of fire models. Thus, we have not included stochastic models of
room fires, 4 forest fires, 5 hybrid models (which contain features of two or
more distinct models), or smoldering fires, which range from cigarette fires
to fires in mines. 6- 7 The emphasis will be on numerical methods.
B. Reasons for Fire Modeling

It is desirable to have a good mathematical model for the development
of a fire in a structure in order to:
1) Avoid full-scale testing. For example, new materials and new or al-
tered suppression devices all need to be tested before they can be used in
new construction. Once the intrinsic properties of the materials are known,
it is much faster and cheaper to estimate the performance of such new
materials and designs on a computer than to carry out full-scale tests.
Moreover, the fire behavior of a new furniture item, for example, could
be tested in a number of configurations rather than just the one that would
be tried in a physical test. Similarly, the behavior of a fire when the venting
is altered, when the wall thickness is changed, or any combination of
structural changes are made, can be examined in a way that would otherwise
be impossible.
2) Help designers and architects. The most important parameters deter-
mining the course of any particular fire can be found by making a number
of runs, and so details of design can be optimized according to a desired
criterion.
3) Establish the flammability of materials. There is as yet no universally
accepted definition of "flammability." Indeed, the fire behavior of any
material is not an inherent property of it, but depends on the circumstances
and configuration in which the material is burning, as well. A reliable
model permits the assessment of material flammability in a particular con-
figuration.
4) Increase the flexibility and reliability of fire codes. At present, fire
codes generally prescribe rigid specifications. With models, performance
fire codes become possible.
5) Identify needed fire research. The construction and use of such models
very quickly reveal where there are significant gaps in our knowledge, and
indicate where research efforts should be focused.
6) Help in fire investigations and litigation. A good model will be very
useful in the investigation and reconstruction of fires, whether for litigation
or for other reasons. Indeed, some models have already been so used.
C. Brief History of Fire Science and Modeling

Mathematical modeling of enclosure fires dates back at least 60 years,
if we take Ingberg's research8 at the National Bureau of Standards as the
first attempt to understand scientifically the postflashover compartment
fire. Ingberg examined the fire behavior of a fully involved room, burning
in a quasisteady state. He related the temperature history of the room
gases to the available mass of fuel.
It took 30 years before that analysis was improved: Kawagoe9 found that
the burning rate was limited to a factor times /4\/H (the "ventilation
H. E. MITLER 713
factor"), where A is the area of the ventilating opening and H its height.
Kawagoe's theoretical results for a steady-state fire in a flashed-over room
where the gases are well mixed were confirmed largely by experiment. 10 " 12
Later, these results and analyses were extended by Odeen. 13 Thomas and
Hinkley14 and Thomas et al. 15 introduced the idea that the hot gases form
a separate upper layer, and that the thermal plume is the way in which
mass and energy pass into the upper layer.
Interest in fire research and in fire modeling grew rapidly after 1958,
and in 1959 the first symposium on fire research and the use of models
was held in Washington, sponsored by the National Academy of Sciences.16
Part of the reason for the enhanced interest was that research on the physics
and chemistry of fires had made significant strides. Equally important was
that the development of large, fast electronic computers permitted the
rapid solution of many simultaneous coupled equations and, therefore, the
possibility of more sophisticated and comprehensive mathematical models.
Research since then has been accelerating, as well as the development of
computer hardware and appropriate mathematical software.
D. General Remarks
Mathematical models can be put into two broad classes: stochastic (or
probabilistic) and deterministic. In stochastic models, the probabilities and
contingent probabilities of significant events are estimated, and then the
a priori probability of a particular outcome estimated. Little or no physics
needs to be introduced.4 -15 -17-46
In deterministic models, the problem and configuration are prescribed,
and then the laws of physics and chemistry, as reflected in the equations
chosen to simulate reality, are invoked. Well-determined correlations may
be used to describe some of the processes. These equations then determine
the evolution of the fire. No probabilities are involved.
There -are by now many exemplars of each type of model. We shall
concentrate on deterministic models. The review will not be comprehen-
sive; only a small subset of existing models will be described. The numerical
techniques used in these models, however, are fairly representative of those
in general use.
Finally, it should be made clear that the adequacy with which the physics
and chemistry of the processes taking place in fire is represented by the
model determines the adequacy of the results. That is, it is far more im-
portant than the precision of the numerical solutions; thus, integrating the
equations to within 1% (or even 5%) is perfectly adequate.
II. Physics of Fire

There are a number of discussions of the physics of fires (see, for ex-
ample, Refs. 18-22). Since the emphasis in this book is on numerical
methods, I shall sketch the rudiments of only the most important dynamical
processes.
A fire acts as a pump, since the hot gases produced by combustion
expand, rise because of buoyancy, form a thermal plume, and produce
large-scale convection. This plume contains products of combustion, in-
eluding CO and soot (in addition to CO, other common toxicants are HC1
and HCN), and is oxygen-vitiated. There are other associated flows, such
as entrainment of air into the thermal plume. The plume, upon striking
the ceiling, fans out into a ceiling jet', this rapidly flowing layer eventually
reaches the confining walls, and is turned downward, forming "wall jets."
Momentum carries the jets down the walls for some distance, but eventually
that is dissipated and buoyancy turns them back up. Then they form a hot
gas layer trapped under the ceiling; this layer heats up the ceiling and the
contiguous upper parts of the walls. It is sooty and oxygen-poor (i.e., lean
in oxygen), since the plume is. Initially, expansion drives cool air from the
room. When the layer becomes so thick that its lower boundary is driven
below the top sill of a free vent, the hot gases begin to escape from the
compartment, too. After a brief phase during which room gases—hot and
cold—continue to escape over the entire vent(s) due to the gas expansion,
the buoyant outflow reaches such a magnitude that external (generally
cooler) air is pulled into the room. In the region between the two gas
streams, a mixing region is formed, so that some of the hot gases are pulled
back into the room and mixed with the lower layer. The situation is shown
schematically in Fig. 1, taken from Ref. 23.
The flows are generally turbulent due to two mechanisms: first, as a gas
mass accelerates due to buoyancy, it will quickly achieve velocities such
that the flow becomes unstable; second, gradients in the flow velocity will
produce vortices, which eventually dominate the plume.
The flames produce radiation, which heats walls and other objects in
the enclosure. The heat feedback from other hot regions such as the hot
layer and the ceiling will also contribute, and exposed objects may be
heated to their ignition temperature. Moreover, the additional heat feed-
back generally will accelerate the spread and burning rates of the fire(s).
TEMPERATURE PROFILE, Tr
STABLY STRATIFIED
REGION
Fig. 1 Schematic of the burning room, showing the various mass fluxes, mi9 in-
duced by a fire near the floor. The mean plume inclination is 0, and Z, is the heights
of the interface, the soffit, the neutral plane, etc., above the floor.
H. E. MITLER 715
Thus, many processes occur in a fire, and must be described in a good

model. Among these are convective heat and mass transfer, radiative heat
transfer, ignition, pyrolysis, and the formation of soot and other species.
III. Field Models

A. Equations
Field models attempt to calculate the velocity and temperature fields in
an enclosure, given some (generally prescribed) heat source. Some field
models also calculate gas species concentrations as a function of position
and time. The equations that describe the flow of fluids are the conservation
of mass, momentum, and energy, plus the equation of state of the fluid.
These equations constitute a set of coupled nonlinear partial differential
equations (PDE's). The associated boundary conditions and initial con-
ditions are determined by the particular problem and its geometry (see,
for example, Ref. 24).
Conservation of mass is described by the continuity equation
~ + V - (pa) = S (1)
Here p is the local mass density of the fluid and u is its velocity. The 5 on
the right-hand side corresponds to sources or sinks of mass; for almost all
cases of interest, S is zero. The Navier-Stokes equation describes conser-
vation of momentum:
+ V • (pun) = -Vp - gpe3 + V • T (2)

01
where g'is the acceleration of gravity, p the absolute pressure, e3 the

upward-pointing unit vector, and T the stress tensor. The right-hand side
describes the forces acting on the differential element.
Because the kinetic energy and viscous dissipation terms are generally
small compared with the internal energy, they are usually neglected. The
pressure work is included in the PV term in the enthalpy. Hence, the
energy equation is approximated by
(pcpTu) = q'" - V-</ (3a)
where the right-hand side gives the following sources:

q'" = volumetric source q = heat flux = -kVT (3b)
The viscosity JUL implicit in Eq. (2) and the thermal conductivity k, which
is implicit in Eq. (3a) via Eq. (3b), are functions of p and T. Finally, there
must be an equation of state,
p — p(/?,7) (4)
for which the ideal gas law is generally used:

P = PR? (5)
Analytic solutions of these equations exist for special cases that have
simple boundary conditions. Thus, in the steady state, the momentum
equations are elliptic, and admit of analytic solutions for the two-dimen-
sional case. In the general two-dimensional case, a stream-function for-
mulation that simplifies the equations becomes possible, although they still
remain nonlinear. The assumption of no viscosity also simplifies the prob-
lem a great deal, as does the assumption that the fluid is incompressible,
in the sense that dp/dp = 0. In other special cases, simplifications become
possible that either permit analytic solutions to be found or make the task
of solving the numerical equations easier. However, the general set of
equations is not solvable analytically. Therefore, solving the equations
usually requires the use of numerical techniques.25 ~ 27J3()
To solve the PDE's numerically, the equations must be discretized; that
is, the region is divided, by a set of grids, into as many small-volume
elements as is practicable, and the PDE's are replaced by a set of difference
equations (see, for example, Ref. 28).
We will discuss the equations and the solution methods in more detail
subsequently. First, consider the approximations made to the basic equa-
tions. Since the equations are parabolic (elliptic in the steady state), the
boundary conditions must be specified over the entire boundary of the
computational region. The no-slip condition for the velocity components
on solid boundaries is used,29 and either the temperature, heat flux, or a
combination of the two is prescribed for the thermal boundary condition.
At an open boundary, such as a doorway or window, the free flow of gas
into and/or out of the computational region must be permitted, and the
flowfield in the outside region must be known or postulated.
The equations are sometimes simplified by assuming the fluid to be
incompressible. Moreover, the variation of other fluid properties, such as
viscosity, specific heat, and thermal conductivity, is small compared with
that of the corresponding eddy properties and is therefore often neglected.
Fusegi and Farouk30 allow full compressibility and variable properties, with
viscous flow. They use a k-e turbulence model (see Sec. III.B). (For ad-
ditional information, see Refs. 31-34.)
The source of convective and radiative energy is the chemical release
from combustion. This is generally specified as an input to the program.
On the other hand, Bagnaro et al.35 modeled the combustion rate by finding
a diffusion equation for the mixture fraction; this also provides species
concentrations.
B. Turbulence Modeling
The flows generated in room fires generally are turbulent, so that they
consist of eddies or vortices of many sizes. The energy that resides in large
vortices cascades down to smaller and smaller vortices, until it diffuses into
heat. The size of the turbulent eddies generated by the numerical approx-
imations to the solutions of the flow equations can be no smaller than at
H. E. MITLER 717
least several computational cells; the actual eddies exist down to the size
where the viscous forces dominate over inertial forces and energy is dis-
sipated into heat. This is commonly referred to as the Kolmogorov micro-
scale. For laboratory-sized turbulent fires, this scale is on the order of a
millimeter, which is much smaller than the usual computational mesh size.
Since the mesh spacing is usually much greater than the Kolmogorov mi-
croscale, special methods must be used to approximate the turbulence in
these numerical approaches. Thus, the turbulence is often modeled by
turbulent transport equations, which are then coupled to the full set of
equations.
When the turbulent fluctuations are very rapid, the instantaneous so-
lutions are difficult, if not impossible, to obtain. Indeed, turbulent fluc-
tuations generally exist even in what would macroscopically be described
as a steady state. Therefore, one often solves for the time-averaged, or
Reynolds-averaged, flows. A variable ()> is written as
cj> = 4> + <|>' (6)
where <J> is the time-averaged value and cj>' the fluctuating component.
When this is done, there are more variables than equations, and the Rey-
nolds stress terms (defined subsequently) must be given separately to achieve
closure. One method frequently used to handle those terms in turbulent
flow is the Spalding /c-e model of hydrodynamic turbulence. This turbulence
model is due to Saffman, 36 Rodi and Spalding, 37 Jones and Launder, 38 and
Launder and Spalding. 25 The generic name for this type of correction to
the basic equations is subgrid modeling.
The k-e model is obtained as follows: Each variable is expressed as a
time-averaged value plus a fluctuating component, as in Eq. (6). However,
it is generally assumed that density fluctuations p' are sufficiently small to
justify neglecting turbulent correlations involving them (the Boussinesq
approximation, which assumes that the flow has constant density except
in the buoyancy term). These variables are substituted into the flow equa-
tions (1) and (2), and then time averaged. The resulting equations are
div pit = 0 (7)

and
+ ^ . va = -v/5 + V • (M,V«) - (8)

dt
where the summation convention has been used in the last term, and the
tensor R is the Reynolds stress (per unit mass):
RU - <«;«;> w
where the brackets indicate time averaging. The k-e model arises from the
determination of the form of this correlation term. One approach that is
much simpler than using the full Reynolds stress model is to use an algebraic
expression. Thus, the Boussinesq turbulence, or eddy viscosity, JJL,, which
accounts for the momentum transfer due to turbulent fluctuations, can be

used by choosing
Prandtl suggested that, in analogy with jm in the kinetic theory of gases,

|ji, o c p M , L , (11)
where ut is a local turbulent velocity and Lt a characteristic turbulent length
scale. If one can find simple algebraic expressions for ut and Ln then the
problem formulation is complete. It is not difficult to show that
I*, = CipA:2/€ (12)
where
k - <«;«;>/2 . (13)
is the mean turbulent kinetic energy per unit mass, e the rate at which k
is dissipated, and Cl a constant to be determined. Since the two new
variables k and e have been introduced, it is necessary to solve two new
transport equations. These so-called two-equation models have been de-
scribed in Refs. 25, 36-38, 52, and 131 (also see Ref. 39, Chap. 9).
The k-e model ignores the effects of gravity, yet it is the gravitational
force that gives rise to buoyancy, which, in turn, accelerates flows and can
cause turbulence. When buoyancy is important, the anisotropy forces the
determination of the level of temperature fluctuations. This quantity is
symbolized by g:
2
g - <r ) (14
where T is the fluctuating component of T, and a third transport equation
is needed. This buoyancy-corrected k-e model should be used whenever
large-scale recirculating movements occur, such as in a fire.40 The k-e-g
model is described in Ref. 25.
The transport equations for turbulent quantities usually are treated by
the k-e model, but with no viscous terms, since the molecular viscosity is
so much smaller than the turbulent viscosity (see, for example, Ref. 30).
For turbulence, Ku et al.29 used the algebraic turbulence viscosity model
of Launder and Spalding25:
TXXW = 2^ —
I € •%
(15)
\ /
(17)
H. E. MITLER 719
where |xe and ke are effective values of viscosity and thermal conductivity.
The turbulent flow formulation is sometimes Favre-averaged, which means
that new variables are defined as integrals with p as the weight function.
This is used to describe flows with large density variations. 30
As Baum and Rehm41 point out, most turbulence models in use are of
the gradient diffusion type. Thus, they implicitly assume that a small-scale,
locally homogeneous turbulent field underlies the organized macroscopic
mean motion whose solution is sought. There is evidence that this is not
correct; for that reason, among others, Baum and Rehm do not use tur-
bulence models in their work. They aver that, on the contrary, the large
eddies control the overall fluid motion; therefore, theirs is a large-eddy
model. Since the energy cascades down from large to small eddies, the
kinetic energy would, with their model, accumulate (unphysically) at the
grid scale, rather than diffusing into heat. Therefore, they periodically use
Lanczos smoothing [which is equivalent to putting in an artificial viscosity
term (see Ref. 42)]. A detailed discussion of their model is given in Sec.
III.E.2.
C. Solving the Model Equations

The three principal means of discretizing the equations are the use of
finite differences, finite elements, and finite volumes. In using finite dif-
ferences, the differential operators are approximated by making a Taylor
series expansion of the operators in terms of differences on a grid, and
taking a finite subset (see Ref. 39, Chap. 2). The finite-element method
basically consists of reducing the PDE's to a set of ODE's or algebraic
equations; one typically assumes an expansion of the function in terms of
a basis. This could include polynomial fitting (see Ref. 39, Chap. 3). Finally,
the finite-volume method involves integrating the governing equations along
the faces of control volumes surrounding each node. This has the advantage
of satisfying the conservation laws in the difference equation form. This is
particularly easy to do for Eqs. (1-3), because they are written in diver-
gence, or conservation-law, form — so-called because, in integrating over
a volume, the divergence theorem can be used on the second terms on the
left-hand side to transform them into the fluxes of each quantity across the
enclosing surface. When the volume is that of a primitive cell, this guar-
antees that the flux leaving one cell flows into the adjacent one, conserving
the flux, both locally and globally. This leads to a set of algebraic equations
involving scalar terms evaluated at the center of the cell and flux terms at
the cell boundaries. Thus, the PDE's describing fluid flow are approxi-
mated by a (large) set of algebraic equations. That is true, of course,
whatever discretization method is used.
Birkhoff43 has written an excellent review article that places the nu-
merical solutions of these equations in perspective (see, especially, Sec.
13 of Ref. 43). A classic text on PDE's and their numerical solution is that
by Richtmyer and Morton44 ; a more up-to-date text is that by Peyret and
Taylor.39
The method most generally used for the solution of the nonlinear dif-
ference equations is some variant of the Newton-Raphson multivariate
procedure (see Sec. IV.E.I); this is equally useful for sets of nonlinear
equations. This technique generally requires the evaluation of the Jacobian
of the equations.
For the steady-state case, the Navier-Stokes equations are elliptic (in
the spatial variables). If the problem is one dimensional and there is no
radiation, the Jacobian is tridiagonal, and the solution method is easy. 44
The nonlinear Gauss-Seidel method (see Sections III.E.I and IV.E) can
also be used, but it is not as good as Newton's method for this case. For
the two-dimensional case, there are five diagonals and seven diagonals in
three dimensions, but the matrix is still sparse, and special methods can
be used to find the solution. The best way to solve them is by iterative
techniques, rather than "exactly," because truncation errors can build up
to prohibitive levels. When radiation is included in the problem, every cell
connects with every other (i.e., there are no zero matrix elements), and
no special solution methods are available.
The prototypical problem is a prescribed fire in a room with one or more
vents. Explicit methods have the theoretical advantage of yielding exact
results in one pass; however, they may require exceedingly short time steps.
Implicit methods are more stable and permit much larger time steps, usually
more than making up for the fact that iterations are necessary for conver-
gence to a solution. An early comparison of explicit vs implicit methods
is given in Ref. 45.
Because of computer memory and speed constraints, the equations have
generally been written assuming two-dimensional flow. For example, Ku
et al.29 wrote the program UNDSAFE to solve the Navier-Stokes equations
with local heating, in two dimensions. They solved the problem for rec-
tangular enclosures. They used the SIMPLE algorithm of Patankar and
Spalding.52 The governing equations and their boundary conditions are
approximated with finite-difference equations by a control-volume-based
method. The computing cells are chosen so that their boundaries coincide
with physical boundaries. A staggered grid system devised by Harlow et
al.60 was chosen. In this system, the pressure, temperature, and density
are evaluated at the centers of the cells while the velocity components are
evaluated at the cell boundaries. Thus, the momentum equations are writ-
ten for cells centered at the boundaries of the basic cells. The convection-
diffusion terms are discretized by a hybrid scheme (see Ref. 26, pp. 89-
90 and 126-131).
D. General Observations
Numerical approximations lead to approximate solutions; sometimes the
solutions are then so poor that the results are incorrect, but that might not
be apparent, and would be attributed to the physics of the problem. For
example, numerical diffusion or dispersion (or both) is (are) created by
discretization. Thus, fine structure is smeared out (or ripples created) via
the continuity equation, an artificial viscosity is created in the momentum
equation, and numerical conductivity in the energy equation. The following
H. E. MITLER 721
is one example: the limitations of computer speed and size forced Torrance
and Rockett 47 to use a relatively gross grid. The solution of the equations
showed that, at high Grashof numbers (Gr ~ 1011), large coherent struc-
tures were periodically formed and shed, and eventually dissipated. It was
surmised that this resulted from the large shear produced across the large
mesh (because of the coherence of the phenomenon, they avoided saying
that this was the onset of turbulence). They also found that their numerical
solution of the equations resulted in unphysical oscillations. Therefore,
they devised a smoothing algorithm that eliminated these oscillations; note
that this smoothing is equivalent to the addition of a zeroth-order viscous
term to the equations. In fact, later reconsideration of the same problem
using a smaller grid did not produce these eddies, showing that they were
an artifact of the (too large) grid size; the numerical viscosity was much
reduced when the grid size was made smaller.
This example and others lead to an important general guideline: it is
essential to make sure that the numerical procedure yields reasonable
solutions. This is perhaps best done by showing that the numerical pro-
cedure used yields the correct solution for a case where an analytic solution
is known. 43 For example, Torrance45 could not obtain an analytic solution
to the fluid flow equations for a problem in which the Grashof number
was greater than 105; therefore, a comparison was not possible. However,
the analytic solution for the problem with Gr — 0 was available, and the
validity of the solution in that limit, at least, was verified.
Again, Baum48 examined the k-z-g model; he found quasianalytic so-
lutions for two boundary-value problems of flow past a plate. He then
solved the same problems using the A-e-g model, and showed that the
parameter values normally used for the model lead to some internal sin-
gularities. To obtain numerical solutions, one integrates across these, and
the result depends on the relative location of the nodes and the singularities.
Thus, solutions obtained with the k-z-g model are not always well defined.
E. Solving the Difference Equations

Once the equations have been discretized, then they must be solved.
Many methods have been used to solve these sets of simultaneous algebraic
equations.
1. Linear Methods
Since the Navier-Stokes equations are nonlinear, the resulting difference
equations are also nonlinear. One solution method is to solve a sequence
of sets of linear equations, although this is seldom the most accurate way.
Therefore, first we consider solution methods for sets of linear equations;
that is, sets that can be written in the form
Ax = b (19)
where A is a given matrix, b a given vector, and x the vector we are trying
to find. An enormous literature exists on the solution of this problem, and
we will not attempt to describe all of the methods. A possible explicit
method is to find the inverse of the matrix A, and to set

x = A~lb (20)
This technique has many drawbacks and, therefore, is seldomly used; the
problem is usually attacked in the form of Eq. (19). Solving for x generally
involves Gaussian elimination or some variant of it. Even when the matrix
A is nonsingular, however, the procedure is inadequate when its condition
number c(A)132 is so large that the product of the relative error in the matrix
(due to rounding or truncation errors in the calculation) and c(A) is of
order of one or larger. This is made more precise by Eq. (9.4-20) in Ralston,
for example.
In addition to explicit methods, there are iterative procedures, such as
the Jacobi or Gauss-Seidel (G-S) successive substitution methods (de-
scribed subsequently). The G-S procedure is one of a class of relaxation
procedures, where one or more components of the residual vector
r, = b - Axt (21)
arising in the ith iteration are made to vanish by some modification ("re-
laxation") of the ith approximation xt. These methods are especially im-
portant because they also are useful when the equations are nonlinear.
There are a number of iterative techniques based on minimizing a quad-
ratic form. Two of these are the method of steepest descent and the con-
jugate gradient method. For ill-conditioned systems, the convergence rate
of steepest descent will be very slow; it is mentioned here because it appears
in the Levenburg-Marquard method, discussed in Sec. IV.G. The conjugate
gradient method132 will converge in a finite number of steps, but that
number may be large, and it is often not as efficient as the relaxation
methods. This is not always the case, however (see Ref. 49).
2. Current Methods
An implicit iterative method is generally used to solve the equations.
Ku et al., 29 who solved the flow problem in a compartment with a fire that
was modeled as a prescribed heat source, best describe the details of the
numerical procedure:
The flux terms at the cell boundaries are approximated through
upwind differencing* while the velocity and temperature gradients
on the cell boundaries are approximated by central differences.
The unsteady term is approximated by a simple forward difference
in time. The overall procedure is conservative in that it conserves
mass, momentum, and energy in any finite region of computation.
After the above procedures are carried out one obtains, at each
cell, five algebraic equations in five unknowns: pressure, density,
temperature, and the two velocity components. In low speed flows
the pressure is only weakly coupled to the equation of statet and
*For discussions of upwind differencing, see Refs. 50 and 51, and Ref. 39, Sec. 2.4.
tThat is, the pressure during well-vented fires remains approximately atmospheric.
H. E. MITLER 723
thus the primary coupling in the state equation is between the

density and temperature. Therefore the usual procedures employ-
ing the equation of state to calculate the pressure cannot be used.
Special procedures must be devised for treating the pressure.
The special procedure they use to find the pressure is an iterative scheme,
following Patankar and Spalding52:
... the temperature is evaluated from the energy equation. Then
the density is found from the equation of state. This reflects the
correct physical mechanism that density changes are caused pri-
marily by temperature changes. Next a pressure field is estimated
by taking it to be the pressure field at the previous time level.
Using this first estimate of the pressure field the velocity compo-
nents are calculated from the two momentum equations. These
velocity components, in general, do not satisfy the mass conser-
vation equation so there will be a residual mass source at each
cell. The pressure is then corrected in such a manner that the
residual mass sources are reduced in size. This corrected pressure
is then used as a second estimate to calculate the velocity com-
ponents. The process is repeated through a number of iterations
until the residual mass sources become as small as desired. Thus
the final pressure and velocity fields will satisfy the mass conser-
vation equation with a prescribed accuracy at each cell. The equa-
tion for the pressure correction is obtained by making the
approximation that the velocity corrections are linearly dependent
on the gradient of the pressure correction.
Ku et al. go on to describe the algorithm in further detail.
The two-dimensional steady-state problem was solved similarly in Ref.
53. The solution method was a finite-difference technique that combines
features of the SIMPLE and NEAT algorithms [Ref. 52 and Ref. 131, p.
565] together with a whole-field pressure-correction algorithm described
in Ref. 40.
A slight modification was used by Fusegi and Farouk30 ; they also use an
iterative scheme (a modified SIMPLE algorithm). However, their solution
method employs the Strongly Implicit Scheme54 as a multidimensional
simultaneous algebraic equation solver instead of the less efficient line-by-
line solver based on the tridiagonal matrix algorithm used in the standard
method.26 They solve the equations using a pressure-corrected iterative
technique.
In the late 1970s, Baum and Rehm launched a project to model the
combustion and convective flow processes that occur in enclosure fires,
from first principles. In Ref. 55, they established the equations to be solved,
taking the following conditions: the fluid is assumed to be inviscid, thermal
conductivity is ignored, the gas is ideal, the enclosure walls are perfectly
insulating (so that all of the processes are adiabatic), and the heat source,
which simulates a fire and drives the flow, is prescribed (both in time and
extent). The enclosure is taken to be two dimensional, with a rectangular
cross section. Radiation transport was ignored. Unlike many such studies,
however, the Boussinesq approximation (approximately constant density)
was not made. That is, the large density variations due to the equally large
temperature variations were taken into account, but compressibility effects

were suppressed. Such a fluid has been called thermally expandable.^ The
flow is fully time dependent, except that no turbulence model is included,
and acoustic oscillations are not in the equations. The latter result is achieved
by treating the pressure terms so that only the time derivative of the
spatially averaged pressure appears in the energy equation, but the spatial
gradients of the pressure are admitted in the momentum equation. There-
fore, buoyant or internal-wave motions can appear, but sound waves are
"filtered out."
These equations were solved in the second paper of the series.57 They
did not use an available "package"; rather, they constructed their own
scheme. First, the equations were nondimensionalized such that all de-
pendent quantities were made of order unity, which is one way to avoid
certain computational inaccuracies (see Sec. IV.E.3). Next, the equations
were discretized; they established a number of criteria that the differencing
scheme must satisfy. By casting the equations in conservation-law form,
the divergence and the curl (for the vorticity equation) are calculated
correctly when in discrete form. Second, the difference equations must be
at least second-order accurate in time as well as space. This was done by
using a three-time-level, leapfrog scheme to express time derivatives nu-
merically, rather than the usual two-level central difference formula. Third,
they should accurately reproduce internal-wave modes; in Ref. 41, Baum
and Rehm had already established which difference equations satisfy this
criterion. They also carefully avoided using differencing schemes that had
been shown to lead to computational instabilities. The scheme they chose
is called J3 by Arakawa and Lamb.58 They also were able to avoid an
instability, described in Ref. 59, by using a different order in the operations.
They use a typical staggered grid (see Ref. 60). The pressure (energy)
equation plus boundary conditions constitute a singular linear algebraic
system of equations; details of the algorithm used to solve them are given
in Ref. 49. They also carried out a linear stability analysis, and found a
condition on the maximum time-step size that can be taken at any step.
In the formulation of their field model, Baum et al.61 carefully and explicitly
checked the adequacy of their numerical solution against the exact analytic
solutions for some cases where the latter could be found; the importance
of this step was stressed in Sec. III.A.
More discussion of the numerical procedure is given in Ref. 62. In that
report, Baum and Rehm also give some experimental verification of the
results for the formation and buoyant rise of a thermal plume due to a
heat source in an enclosure. In Part II,63 they carry out the task that is
usually omitted in these numerical calculations: they verify the adequacy
of the algorithm. They find exact analytic solutions to the equations for
two special cases, and find that the numerically derived solutions agree
with those to about one part in 106 (the specified accuracy) in both cases.
In Ref. 57, the computation was extended to treat the coagulation of the
smoke aerosol produced by the fire. In Ref. 64, Baum and Rehm extended
the computation to the three-dimensional case. In their words, "the basic
assumption is that large-scale macroscopically resolvable motion is re-
sponsible for most transport of mass, momentum, and energy except at
H. E. MITLER 725
boundaries. Small-scale physical and chemical processes, which control

heat and mass transfer to boundaries, as well as most combustion and
smoke aerosol phenomena, are ultimately to be embedded in a systematic
way into the large-scale flow." As in the first calculations, heat transfer to
the boundaries is ignored, and no turbulence model is employed. The large-
scale transport is then calculated directly from the equations of motion.
F. Radiation
There has been much less uniformity of treatment for radiation than for
fluid flow. In fire models, radiation has either been omitted altogether, 29 - 40
or questionable assumptions have been made: radiation has been assumed
not to interact with the gas,65 the gas is assumed to be gray, 66 -67 or the
radiation field has been considered to be one-dimensional. 68
The interaction between turbulent convection and radiation in room fires
recently has been investigated by Fusegi and Farouk30 in a two-dimensional
calculation. Their calculation is typical: the number of nodes they use is
about 20,000. Since they use an implicit technique for the solution, they
must iterate: 50-100 iterations are needed per time step (but they are giant
time steps: Ar ~ 10-40 s!) to converge to an accuracy of |A///| ~ 10~ 4 at
each grid point; /is any of the variables. The results of their calculations
show that the interaction of radiation with soot strongly influences the
temperature field and, hence, can influence the circulation.
G. Calculations in Three Dimensions

Increases in machine capacities, together with advances in algorithms,
have resulted in the development of three-dimensional codes, as well.
Bagnaro et al.35 solved the fluid flow problem in an enclosure with a
prescribed fire in three dimensions for the steady-state case, including
radiation and combustion. They used a k-t model for the turbulence. Using
a method due to Lockwood and Shah, 69 radiation heat transfer was handled
by using a three-dimensional quadratic expression for the radiation inten-
sity / in the approximate radiation transfer equation (see Ref. 70):
(« • V)/ x = /Ut^/Tr - / x ] (22)
The coefficients giving the radiation are functions of position. After in-
tegrating Eq. (22) over various solid angles, there result three coupled
PDE's in the coefficients (which appear in the function that approximates
/). Bagnaro et al.35 solved this set of PDE's using a three-dimensional
version of the program TEACH.71
Other three-dimensional calculations of fire and smoke spread through
rooms have been made in the last few years (see Refs. 27, 72, and 73).
These calculations use the program JASMINE, which is based on a finite-
domain solution of the heat- and mass-transfer equations, including com-
pressibility. An extensive discussion of the use of UNDSAFE and UND-
SAFE II recently has been given in Ref. 74. A three-dimensional version
of UNDSAFE has been developed75-76 that uses an improved tridiagonal
matrix solution scheme.77 Also, as noted earlier, Baum and Rehm64 have
performed three-dimensional time-dependent computations using their

model.
H. General-Purpose Programs
Most of the programs mentioned so far have been developed with a
specific application in mind. A number of sophisticated, general-purpose
computer programs have been developed that permit the solution of various
kinds of equations (algebraic, ODE's, PDE's), or coupled sets of any of
the above. Shampine and Watts78 discuss software available for solving
ODE's, and Boisvert and Sweet79 describe PDE solvers, as does Walsh.80
Solvers for linear and nonlinear sets of algebraic equations are discussed
in Refs. 81 and 82. There are various compendia of available mathematical
software, e.g., IMSL or NAG; these are commercial packages available
at most research institutions. There is also software in the public domain,
such as LINPAC (linear equations solvers), DEPAC (ODE solvers), etc.
Kahaner et al.133 have written a text that discusses these and other software
packages.
There also are commercial codes currently available that have been
designed specifically to solve fluid-flow problems. Upon developing the
k-e model of turbulence, Spalding embedded it in a general-purpose fluid
dynamics code, which is commercially available as PHOENICS4 Com-
parable codes such as FLUENT, NEKTON, and FIDAP exist, and there
are doubtless others available. A brief description of these, including a
very brief description of the discretization and solution methods, as well
as turbulence models, is given in Refs. 83 and 84.
These codes have been applied to fire problems in an enclosure (see the
workof Cox, Markatos, and coworkers). In principle, these codes can be
applied to multiple rooms; as stated earlier, however, the limited number
of nodes available, and the correspondingly coarse resolution, especially
for three-dimensional configurations, makes such application, at best, ex-
tremely costly.
IV. Zone Models

A. Description of Models
In a zone model, the space is divided into a small set of large control
volumes, with conditions (generally) uniform throughout each. The heat
and mass transfer between zones can then be calculated, using the appro-
priate equations. In general, a different set of processes dominates in each
zone. For a single room, the zones that are usually chosen are an upper
(hot) region and a lower (cooler) region, and such models are referred to
as two-zone models. This is somewhat misleading, however, because other
zones are also often included: a combustion zone, the walls, the region
around vents, and one or more significant objects in the compartment.
The thermal plume that develops at and above a fire is another zone; its
^Mention of specific products throughout this chapter does not constitute an endorsement,
implicit or explicit, by the National Institute of Standards and Technology.
H. E. MITLER 727
volume, however, is (incorrectly!) never removed from the lower layer.

The plume carries mass, species (inert plus the products of combustion),
and enthalpy from the lower layer into the upper one. When the fire lies
within the upper layer, however, no mass is carried up from the lower
layer. Note that the upper and lower layers are zones whose volumes vary,
in distinction to field model cells.
Field models yield the velocity and temperature distributions (although
the resolution and accuracy of the results depend on the grid size); zone
models do this only crudely. The geometry of the room and its furnishings
can have substantial effects on the nature of the recirculation patterns;
hence, the higher spatial resolution of field models can sometimes be im-
portant. For example, detailed knowledge of the temperature and/or flow-
field near some critical item (such as a smoke detector or sprinkler) may
be necessary. Also, fluid-dynamic considerations are automatically built
into field models, rather than being forced into oversimplified approxi-
mations. Thus, field models follow the movement of the thermal plume,
rather than assuming that deposition of mass and energy from the plume/
combustion zone into the upper layer is instantaneous. Similarly, they
describe the spread of the ceiling jet to the entire upper layer, rather than
assuming instantaneous mixing.
Conversely, field models have certain drawbacks: they strongly empha-
size the fluid flow aspects of the fire problem, but generally give corre-
spondingly little weight to two other classes of processes central to
understanding fire spread and fire behavior—solid-phase (heating, pyrol-
ysis, ignition, charring) and combustion processes. There are other good
reasons, as well, to believe that zone models are often preferable to field
models, in the fire area. They will not be listed, however, to avoid the
appearance of making this a polemical tract. Nevertheless, it should be
pointed out that, due to the vast computer requirements, it is not now
practical—nor does it seem likely to become so in the near future—to
describe fires in multiple-room enclosures with a field model; zone models
are the only ones available for that purpose at present.
For any fire model to be truly useful, it will have to include phenomena
that are not so well understood at present. Progress needs to be made, for
example, in understanding the behavior of wall fires, of burning in vitiated
air, of extinguishment, the production of CO, HCN, and other toxicants,
etc.
Even when the basic equations for some process are either not known
or not solved, useful correlations may exist that can be used in a zone
model. Thus, many global phenomena can be included in zone models that
are not easily incorporated into field models. For these reasons, a great
deal of effort has gone into the development of zone models. In one sense,
zone models may be thought of as hybrids: field models solved via finite-
volume methods, where the cells are very large, and where, as a result, a
great deal of phenomenological subgrid modeling replaces expensive de-
tailed simulation performed with marginal resolution at best.
Zone models can be classified in several ways. We will begin by consid-
ering steady-state vs transient (time-dependent) models.
1. Steady -State Models

These models span the spectrum of sophistication and complexity from
a relatively simple compendium of individual equations yielding estimates
of upper-layer temperatures in a single room, time to flashover, etc.,85- 86
to quite sophisticated models that take many physical phenomena into
account and solve the equations of conductive, radiative, and convective
heat transfer from first principles. 87 In between are models such as
COMPBRN,88 89 which uses many empirical correlations, as well as basic
equations, to describe the results of various physical processes. The earliest
model9 treated the steady state.
Since time derivatives vanish, these models consist of sets of coupled
algebraic equations. The solution of sets of algebraic equations was dis-
cussed in Sec. III.E; further discussion is given in Sec. IV. D.
2. Transient Models
There are single- and multiple-room models. For comparisons among
some of these models, see Refs. 90 and 91. As is the case for field models,
these models have many features in common. They are mostly two-zone
models, with similar plumes and vent-flow descriptions. Many of them
include the radiative exchanges between gases and solids, and some cal-
culate the heating of target objects.
B. Formulation of the Problem

Only a small subset of the existing models will be examined. Primary
emphasis will be placed on the Harvard family of models, with which the
author has greatest familiarity. These, and the model FAST, use prototypes
of the numerical methods used in most zone models.
L FAST Model
Most of the models solve a relatively small set (about five per room) of
coupled, generally nonlinear, ODE's that describe the bidirectional flow
of gases through vents, the rate of change of mass and temperature in the
upper layer, etc. For example, FAST (which is based on BRI192) solves
the equations
P = s/($ - l)V (23)
tf = ^(TJPV&Et + V>/(P - 1)V1 (24)

where / = U,L (U = upper, L = lower) and
Vu = (cpmuTu + EV - V^s/VO/Pp (25)
Here
s = CprhyTy + cpmLTL + Eu + EL (26)
and
P - cp/R = y/(y - 1) (27)
H. E. MITLER 729
P, F, 7, R, cp, and y are, respectively, pressure, volume, temperature,

the gas constant, specific heat, and cplcv\ see Ref. 93 for the rest of the
nomenclature. The right-hand sides of these equations are the source terms.
The rooms are most strongly coupled through the pressure equations, which
determine the flows through vents connecting the rooms, and to the exterior.
2. Harvard Mark 5 Model

Mark 5 is a single-room model in which a small set of ODE's describes
the rate of change of what may be called the "primary" variables. These
are the mass and energy of the two layers, the size of the fire if it is growing,
and the pressure at the center of the floor. This set is supplemented by a
PDE that describes heat diffusion through solids, yielding the temperatures
of walls and heated objects. It is also supplemented by a number of al-
gebraic equations involving other variables, such as the radiation flux from
the hot layer to any object or wall, the convective heating/cooling of walls,
the pressure drop across any vent produced by the fire (i.e., the buoyancy-
induced flows), etc.
Mark 5 takes into account not only fluid flows but other dynamics of
the fire, e.g., the radiative feedback from the room to a pyrolyzing (burn-
ing) surface, the resulting spread of that fire on the surface, etc. Therefore,
many more variables must be taken into consideration than in FAST or
the Consolidated Compartment Fire Model (CCFM). 94 The additional var-
iables may be thought of as auxiliary variables, but are clearly of interest
for an accurate description of a compartment fire.
The plumes are described by equations that give the mass and enthalpy
injected into the upper layer due to a source in the lower layer. Several
choices are available in the program; the simplest is the point-source plume
described in Ref. 95. The mass flow into the upper layer from a thermal
plume of height x is given by
mp = Po[(4 • 357T2a4g^)/(54 cpp(>T())$ (28)
where p0 and T() are the ambient density and temperature, g the acceleration
of gravity, Q the convective heat-release rate, and a the entrainment con-
stant (for turbulent entrainment of air into the plume). The resulting energy
transport into the upper layer is
Ep = mpcpT(} (29)
Vent flows are given by a hydrodynamic approximation to the fluid flow
equations, and are described by Eqs. (37-39). To follow the transient
heating of walls or "target" objects, the heat diffusion equation
^ - a *I
2 y(30)
dt dx '
which is a PDE, must be solved (assuming that heat diffuses internally only
by conduction). Here it is assumed that
1) The thermal conductivity k, the specific heat c, and the density p are
constant (independent of x, r, and T). Hence, so is the thermal diffusivity
a = k/pc (31)
2) The flow of heat is one dimensional (slab geometry).
3) There are no sources or sinks of heat in the solid.
The numerical solution is given by Eqs. (65-70) in Sec. IV. E.I. A flam-
mable material is taken to ignite when its surface reaches an ignition tem-
perature. There are three kinds of fire algorithms built in: a growing fire,
a pool (constant size) fire, and a burner fire. The burning rate of a pool
fire depends on the flux incident on its surface, and, therefore, on the
temperature and emissivity of the hot layer, of the ceiling and walls, and
on the radiation from flames from other fires as well as from its own. The
burner fire, on the other hand, is unaffected by external fluxes; it is pre-
scribed by the user. Finally, the growing fire is similar to the pool fire, but
the radius is permitted to increase in size. The equation for R(t) is
R = AC(l + C/2 + C2/3) (32)
where A is the "growth rate parameter," C = <t>/°"^/> ()> is the net flux
falling on the surface, and 7} is the mean flame temperature.
Species production rates are evaluated simplistically, as a constant frac-
tion of the pyrolysis rate. Oxygen consumption is found in the same way;
then the rate of change of species / in the upper layer is readily found from
the equation
mf = m/i - mULYj (33)
where mf is the fuel flow rate, muv is the rate at which hot gases leave
the layer through the vent(s), Y, is the concentration of species / in the
layer, and/) is the "yield" of species / (in kilograms/kilogram). For oxygen,
f0 is negative, and another term is added to account for the injection of
oxygen into the upper layer through the plume.
Once the soot, CO2, and H2O concentrations are known, the absorption
coefficient of the upper layer is calculated. This is done in one of two ways.
The first is extremely simple: the soot concentration is multiplied by a
constant determined from an experiment. The second method includes the
absorption by the molecular bands of H2O and CO2; the algorithm is rather
complex (see Ref. 96). Once K is known, the absorption and emission of
radiation by the layer are readily calculated via a mean-beam-length ap-
proximation or better. The emission from flames is calculated assuming
that the flame shape is a cone; the emission coefficient of the flame is input
by the user. Heat exchange with the walls is by convection and radiation.
For example, the convective heating flux to the wall/ceiling from the hot
layer is
4>con - h(TL - Tw) (34)
where the temperature-dependent heat-transfer coefficient h is taken to
be
h = min[fl, b + (a - b)(TL - TW)/1QO] (35)
H. E. MITLER 731
with TL the layer temperature, Tw the wall temperature, and a and b the
maximum and minimum values that h can accommodate, respectively.
One poor assumption made in this model is that none of the oxygen in
the upper layer can participate in the combustion process. Hence, if the
interface height falls so low that the plume height is insufficient to entrain
enough O2 to provide combustion for the fuel, oxygen starvation is said to
ensue. The complete set of equations used in Mark 5 (i.e., the model) is
given in Ref. 97. A slightly earlier version was given in Ref. 98.
The equation for the pressure entails enough interesting points that a
detailed discussion is warranted: when the physical processes that affect
the pressure in the room occur on a time scale that is long compared with
the transit time of a sound wave in the enclosure, the compressibility of
the gases in the room can be ignored; that is, the gas acts as an incom-
pressible fluid. The appropriate equation for the mean pressure at the floor
of the room, /?, is then"
VP + yp 2^ (mi'Pi) — (y -~~ i)\

1) \Q S ((ccPpm
Q ++ ^ m
i* /)
y7}) (yw
r
T / r. . X ^ / • / \ / 1\ yO , X / • r \ /0^\
L jy J
where Q is the net heat release rate in the room, ra, the mass flow rate
out of the room through vent / of fluid of density p,-, and CprhjTj the inflow
enthalpy through vent j. This equation is very stiff, which slows down the
solution of the equation set considerably.
For fires that occur in enclosures that are not hermetically sealed, the
pressure stays nearly exactly constant. The pdV work done by the gases
while the hot layer is expanding (or shrinking) can be a large fraction of
the layer energy change; however, in Mark 5, enthalpy is used rather than
internal energy, and it turns out that Vdp is a very small fraction of the
layer energy change. Therefore, it is ignored; moreover, it is assumed that
the equation is so stiff that p is vanishingly small. Hence, the pressure is
found in a quasi-steady-state approximation. Thus, the pressure difference
at any vent is obtained by noting that, at any moment, the flows induced
by that pressure difference must yield mass and energy conservation. This
is exactly the same approximation as is made by Tanaka in the BRI2 model.
Thus, assuming the ideal gas law, neglecting the differences in molecular
weight among air, fuel, and product species, and (finally) neglecting the
change in specific heat with temperature, the total internal energy in the
room gases remains constant. Thus the rate of change of gas mass in the
room is
mR = m out — m in = EJcpTa — mL + mf (37)
where mL and EL are the mass and internal energy of the upper layer, and
mf the rate at which fuel gas mass is produced in the room, either by
pyrolysis or via a gas burner; this is derived in Ref. 18.
We will now indicate how the mass flow and pressure are related. The
pressures at various heights in the room are referred to that at the floor,
pf. The mass flow through a vent at a given height is calculated by using
the hydrostatic approximations to fluid flow. Thus, it is assumed that mo-
mentum is conserved at each horizontal plane. The vent is divided into a

number of horizontal strips, the number depending on how many changes
in density or flow direction are encountered between the bottom and the
top of the vent; there can be as many as six. The pressure difference from
the bottom to the top of each strip is a linear function of height. Consider
strip /; if the pressure difference at the bottom (top) of the strip at height
^/ (^/ + i) is A/?/ (A/?, + j), then the flow in that strip is related to the pressure
difference across the vent at that height according to
1 |)] (38)
where
G, = (sign A/?,)(2/3)Q B(hl+{ - /i,) V2^ (39)
where B is the width of the vent, Cd the orifice coefficient, usually taken
to be 0.68, and pa the density of the incoming (usually ambient) air. The
derivation is given by Mitler and Emmons 97 and DiNenno et al.22 (Sec. 1,
Chap. 8, "Vent Flows," by H. W. Emmons). This is a quasi-steady-state
expression, but is quite adequate, as long as there are no compressibility
effects to consider. Given ^pl (the pressure drop at floor height), all of
the other Ap/s are fixed, and, therefore, m, are found. Finally, the mass
flows must satisfy mass and energy conservation. The set of Eq. (38) cannot
be inverted (unless the entire vent opening is one strip) so as to give A/?j
in terms of the mass flows; if we write
•mR + min - m out = /(A Pl ) = 0 (40)
it is clear that A/?] is a zero of the function /; it is found by a Newton
technique.
C. Structure of the Equations

The equations of the system for Harvard Mark 5 and Mark 6 (the latter
is described in Sec. IV.G) are of the form
0 = /(w,i;,w) (41)
— = g(u,v,w) (42)
,w) (43)
y = e(u,v,w) (44)
where u, v, w, and y are arrays (''vectors") of variables. Each variable is
an (implicit) function of /, the time since the start of the calculation. There
is generally no explicit /-dependence in any of the functions/, g, h, and e,
which are vector functions of those variables. If the number of components
in array a is represented by [a], then
[/] = [«], [g] = [v], W = M, W = (y] (45)
must hold. Moreover,
H. E. MITLER 733
[/] + [g] + [*] + [e] = N (46)

where N is the total number of variables. Equations (41), (43), and (44)
are the algebraic equations, and the corresponding variables (the com-
ponents of M, H>, and j) the algebraic variables. The components of u,
corresponding to the differential equation (42), can be referred to as the
integrated variables. Equation (41) is a root-finding equation, and the com-
ponents of u are the root-finding variables. Equation (43) [cf. Eq. (51)] is
a fixed-point equation, and the components of w are the fixed-point vari-
ables. Finally, in the original formulation of the problem, there may be a
subset of variables that depends only on other variables, and not on each
other. They are represented by the vector j; thus, none of the variables y
are found in the arguments on the right-hand side of Eqs. (41-44). Those
equations represented by Eq. (44) are also called eliminated variable equa-
tions, and the components of y are the eliminated variables. They may
also be referred to as auxiliary, or dependent, variables.
Evidently, Eqs. (43) and (44) could be converted to the form of Eq.
(41), but the chosen formulation uses less computer time because a system
of implicit fixed-point equations, such as Eq. (43), can (usually) be solved
by a successive substitution method, which requires fewer operations than
the Newton method needed for Eq. (41). Eliminated variables are trivial:
one need only solve Eqs. (41), (42), and (43) simultaneously, and then
substitute the resulting values for u, v, and w back into Eq. (44) to find
y. Equation (41) is the nonlinear algebraic equation for the pressure.
D. Numerical Methods for Solving the Model Equations

Analytic solutions for these coupled sets of equations are generally not
possible. Hence, numerical methods must be used. As is the case for field
models, the differential equations are replaced by difference equations,
which are then solved numerically. In the BRI models, the equations are
solved explicitly, using explicit Runge-Kutta (R-K) methods (see, for ex-
ample, Refs. 100 and 132). The solution is obtained with R-K's of increasing
order, until the solution of that order agrees with that of the next lowest
order. If convergence does not occur by the time a sixth-order version is
used, the time step is cut to one-tenth its original value, and the process
is repeated. If the solution is still not obtained to the prescribed accuracy,
the calculation stops.
The time rate of change p is assumed to be extremely small, and,
therefore, the pressure equation is taken to be an algebraic equation. The
pressure equation is solved by using Grout's method to invert the matrix.
That gives new pressures that yield new flows; the whole procedure is
iterated until convergence is established. One difficulty with BRI1 was that
the initial values had to be guessed very shrewdly for convergence to occur.
Consider the /th member of the fixed-point equation [Eq. (43)]; when
Wj does not appear on the right-hand side, the equation is explicit, and the
solution for wf at any given iteration is exact. When wl does appear on the
right-hand side, the equation for wf is implicit, and the solution will gen-
erally be only approximate, assuming that an approximate value of H>,- is
available for use on the right-hand side. Thus, Eq. (41) [i.e., Eq. (38) for
A/?j] is, by definition, implicit in wt. We could make it explicit if we could
invert Eq. (38) and solve it for A/^; as noted earlier, however, this cannot
be done in general.
Finding kpl as the root of Eq. (40) presents two interesting problems:
first, any numerical "package" to be used in solving the equations must
include a root-finding part. Second, the following interesting conundrum
arises: the algebraic equations are solved by a successive substitution
(G-S) method; hence, it would appear to be a waste of effort to solve
exactly for Ap in any of the global iteration cycles, since 1) the solution
will differ in each subsequent iteration, and 2) the overall convergence of
the procedure should guarantee increasing accuracy. Conversely, it might
be that finding the "exact" solution in each of the global cycles would help
to accelerate the overall convergence process. In producing Harvard 5, the
former philosophy was adopted, and Ap was found with only one local
iteration, rather than iterated to local convergence. The second approach
has been used in the developmental version of FIRST; the result is as
expected: not only has convergence indeed been accelerated by finding Ap
"exactly" at each time iteration, but the program now runs to completion
in many cases where it would "crash" with the previous approach!
A problem with explicit methods is that the time-step size required is
prohibitively short when stiff equation sets are being solved. More fre-
quently, therefore, predictor-corrector methods are used to solve such sets
of equations. FAST was based on BRI1, but the numerical method was
much improved. Since the rates of the various dynamical processes vary
over orders of magnitude, these equations are stiff, and the time steps
required even by predictor-corrector methods may become prohibitively
small. However, FAST uses the Selected Asymptotic Integration Method
for this purpose (see Refs. 101-103), which is a modified predictor-cor-
rector method. In its simplest form, the method may be described as fol-
lows: given a set of ODE's of the form
*i - f i = Qi - LiXi (47)
where Qt is the formation rate of the quantity xh and Ltxt is the loss rate.
The equations are separated into two classes: "normal" and "stiff," ac-
cording to a simple criterion. (Canonically, one refers to the entire set of
coupled equations as stiff.) For the normal equations at time t 4- 8r, the
first iteration for xh the predictor, is
*,M = */.o + 8rF,,0 (48)
where xifQ is the initial value (at time f); for the corrector steps, the k +
1 iteration is given by
*/,*+ i = */,o + Sf[ô + FLk]/2 (49)
For the stiff equations, the predictor and corrector equations are, re-
spectively,
H. E. MITLER 735
and
*/,*+ i = */,o + [284G/,* ~ 4-,o*/,o + f/,o)]/{4 + 8f[L,.0 + L,,J} (51)
The method works well, and it would be most useful to go through a
rigorous stability/convergence/error analysis of the method.
In FAST, the source terms are calculated explicitly at each time step.
Thus, the initial iterates are based on the values the primary variables have
at the beginning of that step, making FAST a single-step model.
The more standard method for solving sets of stiff coupled ODE's nu-
merically is an implicit method using a backward-difference formula, such
as the Gear method.104 This is a multivariate Newton-Raphson technique
(see Sec. IV.E.I). CCFM94 uses a Gear-like package called DEBDF (Dif-
ferential Equations via Backward Difference Formula) to solve the ODE's.
DEBDF is part of a three-routine set called DEPAC, developed at Sandia
Labs.105 These packages are quite sophisticated; the programs internally
determine the required time-step sizes as they proceed. Some packages of
ODE solvers are so sophisticated as to include a heuristic that decides
whether the equations are stiff, or not, and choose internally which solution
algorithms to use! (See Ref. 106.)
In general, numerical "packages" are used to solve the ODE's; the PDE
must be solved separately. A solution method for Eq. (30) is described in
Sec. IV.E.I. However, it is possible to avoid having to solve a PDE, by
writing the surface temperature as a convolution integral. It can also be
done by transforming the PDE into a set of coupled ODE's. This can be
done in three ways: first, by using an integral technique to find the surface
temperature approximately, second, by using similarity solutions to extract
the spatial dependence analytically; and, third, by using a PDE solver such
as the method of lines.107 There is one ODE per mesh point in the solid
into which heat is diffusing. This subset of ODE's often produces great
stiffness in the overall set, however. Since a relatively fine mesh is used
to solve the PDE for all methods except the first, these constitute a field
model approach. Thus, the entire model should more properly be called
a hybrid (partly zone, partly field) model.
E. Harvard Mark 5
L Solution Methods Used

In Mark 5, every variable is calculated so that its value is consistent with
that of every other, according to all of the available relationships, within
the prescribed accuracy. That is, if there are N variables being followed,
there are N algebraic equations to be solved simultaneously. There are
four ODE's describing the rates of change of mass and energy in the layers,
three ODE's for each vent, and five for each burning object (describing
the mass and radius of the fuel packet, the energy release, and the mass
and enthalpy injected into the upper layer by the burning object's thermal
plume). Hence, for a room with only one burning object and one vent,
there are 12 ODE's for 12 primary variables, resulting in 12 (out of the
N) difference equations.
In FAST,93 a set of algebraic relationships similar to, but smaller than,

those in Mark 5 is assumed to be the source terms on the right-hand side
of the differential equations
*> = e,-w (52
where xt — xf(t) is the value of the /th primary variable, and the vector
symbol x stands for the set {*,-}. Equation (52) is essentially Eq. (42).
Since ODE's are transformed into algebraic equations, there is no dif-
ference in principle between the solution method used in FAST and other
ODE-based models and that used in Mark 5. In each case, a set of nonlinear
algebraic equations is solved numerically.
The numerical calculation procedures are described in considerable de-
tail in Ref. 97. A brief synopsis is presented here: rewrite the (linear or
nonlinear) "fixed-point" algebraic equation [Eq. (43)] in the form
xk = g*W (53)
The principal method of solution used is the successive substitution method,

in which the variables x are calculated iteratively. After k iterations, the
(k+ l)st value of variable xf is found by setting
x(* + » = g/(jt (A-)) (54)
where x(k) = {x(fk)} is the set of values found at the previous iteration. The
time dependence of each variable is not displayed in order to make the
equations more succinct.
There is one exception to the formulation of the equations all being of
the form of Eqs. (41-44): the PDE in Eq. (30) is not of any of the forms
of Eqs. (41-44), and must be solved independently, as discussed earlier.
The calculation for A/? is done inside a physics subroutine, since it cannot
be carried out in the numerics package. This is so because the technique
described by Eq. (54) is valid only for fixed-point equations, which Eq.
(41) [or Eq. (40)] is not. Thus, given all of the other jc/s at the nth (global)
iteration, we find a value for A/? — i.e., for x]n) — in the same way as Eq.
(53) yields value for all of the other jc/s. Therefore, we will include Eq.
(41) with the rest of Eq. (53). Note that in order to be able to find A/?,
we must have already evaluated some of the flow rates, so that this variable
is best calculated late in the cycle. This different treatment of A/? is an
exception to the modularity principle: a program should be structured so
that any physics algorithm can be replaced by an improved or corrected
version, or a new numerical procedure readily introduced, without re-
quiring extensive program changes. Generally, modularity is best achieved
by decoupling the various calculations into independent subroutines and
separating the numerical procedures from the physics.
Next, consider Eq. (42), which is solved by numerical integration:
Xi(t + Af) = Xi(t) + I OX*) dr (55)

H. E. MITLER 737
where Af is the time-step interval. In Harvard Mark 5, the trapezoidal rule

integration method is used:
xf(t + Af) - Xf(t) + Af[e,(f) + 6,(f + Af)]/2 (56)
where we have written 0(f) to represent 6[;t(f)] symbolically. Although this
is an explicit integration, the overall method, based on successive substi-
tutions, is implicit. Equation (56) is now also of the form of Eq. (53), and
can be included in that set, as well. Thus Eqs. (41), (42), and (53) [i.e.,
Eq. (43)] have all been cast into the form of Eq. (53), and the successive
substitution method can now be applied to them all. Use of the trapezoidal
rule seems crude, especially when compared with the methods used in FAST,
CCFM, and other models. However, the inherent error is -(Af) 3 6,(T)/12
per time step, where t < T < f + Af. Hence, the magnitude of the total error
in the fth variable (assuming uniform time steps only) is [(Ar) 3 /12]26,(T y ),
where / goes from 1 to n = f/Af. If we approximate the sum by an integral,
it can be shown that the cumulative numerical error in xf is
Error - - (Af) 2 {e,(f) - e,.(0)}/12 - (Af)3{3e,.(0
- 0,.(0)} + (.([Af] 4 ) (57)
At large times, 6, and 6, should both go to zero. Thus, knowing the function
6 7 (f) [and, therefore 0,-(0)], we can readily choose Af small enough to keep
the total error within satisfactory bounds. Indeed, note the comment at
the end of the Sec. IV.E.4.
In early versions of the program, the Jacobi method was used to solve
Eq. (53). In Mark 5, a variant of the G-S method is used, instead. An
important feature of the Jacobi method is that the results are essentially
independent of the order in which the equations are solved. This is not
the case for Gauss-Seidel, and a consequence is that some orderings will
require more iterations than others, for convergence. Indeed, for some
cases, one ordering may diverge, while another ordering produces con-
vergence. Define
P. = &-(xk + ^ xk + ^ xk + ^ xk xk xk} (^R}
When Mark 5 was first being constructed,

*j+1 = g_i.k (59)
was used. Still faster convergence was obtained with what might be called
successive underrelaxation (Ref. 108, Chap. 2): instead of Eq. (59), jtf + 1
is now given by
+1
xf = Xgj:A, + (1 — X)xf, X< 1 (60)
where X is the relaxation factor. Best results were found with X = 1/2.
The use of Eq. (60) rather than Eq. (59) has sometimes led to the con-
vergence of an otherwise divergent calculation. 97 When the method con-
verges, it does so linearly. That is, for all sufficiently large k (i.e., k greater
than some value &c),
||jc*+1 - jc|| < C||x* - *|h k> kc (61)
with m = I and \C\ < 1, where x is the exact solution at time t and the
double bars denote the norm of the vector.28 The Gauss-Seidel method
has yielded more superior results for the enclosure problem than the Jacobi
method: it converges over a wider domain, and usually with substantially
fewer iterations.
The starting values at a new time step are given, for each variable, by
a quadratic extrapolation of the values of the variable at the last three time
steps; this is analogous to the predictor step in predictor-corrector methods.
If, during the calculation, a variable is found to fall outside physically
reasonable (or permissible) bounds, its value is reset to the value it had
at the last time step. This procedure, carried out in sections called LIMITS,
is ad hoc and has drawbacks, but it made the calculation still more robust.
The second method used to solve the equations is the multivariate New-
ton-Raphson technique, normally applied to functions of the form of Eq.
(41). For fixed-point equations of the form of Eq. (54), one defines the
vector function / as
/(*) - x - g(x) (62)
The Newton technique for the case of N simultaneous equations takes the
form
** + ! = jt* - /-!/(**) (63)
where / is the N x N Jacobian matrix {dfj/dXj}. For a single variable, a
sufficient condition for convergence is that
|/(**)/V)/[/V)]2l < 1 (64)
for all k (see Ref. 109, paragraph 78). In Mark 5, the equations are solved
by Gaussian elimination with partial pivoting (see Ref. 108, and Ref. 109,
p. 270) using the code of Ref. 110. If an exact Jacobian were known at
each iteration, then the algorithm given by Eq. (63) would converge quad-
ratically, i.e., with m = 2 in Eq. (61). The finite-difference approximation
to the Jacobian used in Mark 5 is not exact, so that, in fact, we are using
a secant method. It can be shown that this leads to a rate of convergence
given by m = 1.839, in a single variable.111 When the condition number
is very large, the Newton method will not work properly; what can be done
about this is discussed in Sec. IV.G.
The full Newton method is quite time consuming, and some alternatives
were devised. IfJ(t) varies slowly enough, it requires recomputation only
once every few time steps. This simple approach is called Newton Super
Fast (NWSF). It works much of the time, and is indeed very much faster
than the multivariate Newton method (NWTN).
Using a numerics package that solves sets of simultaneous algebraic
equations (or ODE's and algebraic equations) will not suffice for the PDE
H. E. MITLER 739
[Eq. (30)], which describes heat diffusion. Hence, we must again deviate
from modularity and use a numerical procedure inside a physics subroutine.
The technique used in Mark 5 to solve for the surface temperature as a
function of time is a simple explicit finite-difference method: the wall (or
object) is divided into M slabs of thickness 8* by planes parallel to the
surface. The temperature profile within the wall is found at successive time
steps. The Courant criterion is satisfied when
8* > (8*)min = V2a(AO max (65)
where (A/) max is the largest time step taken in the problem and a is the
thermal diffusivity (assumed to be constant). Thus, M is given by
M = [e/(8x) min] (66)
where 6 is the wall thickness ([a] means "the integer part of a"). Then the
temperature at the front surface, at time t + A/, is given by
7\(f + AO - 7\(0(1 - 0AO + 0Af{72(0 + Hi (01 (67)
Here $l is the net heating flux to the front surface, A^ the current value
of the time step, Tl and T2 the temperatures at the first and second nodes
(front surface of wall and front surface of second slab, respectively), and
a and b are defined by
a = 2a/(8*)2 and b = §xlk (68)
where k is the thermal conductivity. For nodes 2 to M, the temperature is
given by
Tt(t + AO = TXOtt - *Af} + aAtfT^CO + Ti+1(t)}/2 (69)
The temperature at the back surface is
t) + HM + I«} (70)
A derivation of these equations is given in Ref. 18 (also see Ref. 22,

Chap. 3).
The overall computation proceeds as follows: the calculation runs along
using G-S until it fails to converge at t + Af() in a fixed, user-determined
number of iterations, with 35 as the default number. The program there-
upon cuts the time step in half and starts over (i.e., tries to converge to
t + iAr0), using newly extrapolated values from the previous time steps;
this guarantees a starting solution closer to the correct answer, which helps
to lead to convergence. If it then succeeds in converging, the calculation
continues with Ar = Ar0/2 for 10 steps (with some exceptions), then doubles
the time increment, since the difficulty may have been only temporary.
Failure to converge in 35 of these smaller steps, on the other hand, causes
another halving of the time step, to Ar()/4; and so forth. The time step will
be halved repeatedly until the solution converges at some point, or until
A/ = At0/16, where A£ is the newly cut time step and Af0 the originally
chosen time step, whichever comes first. If the procedure fails to converge
at this (smallest) time step, the program switches to NWTN. The program
tries to take one step in NWTN, using the time step Ar, and then (if it
converges to a solution) switches to NWSF. The solution mode will remain
NWSF until it ceases to converge (in which case, it goes back to NWTN,
and we start over), or until the size of the time step just taken is > Af () /4,
in which case it switches back to the G-S mode. If it takes a number of
steps successfully in the G-S mode, Ar is doubled when the total number
of time steps taken (from the beginning) is a multiple of 10. This doubling
procedure continues until Ar = Af () again, where it stays, or until trouble
occurs again.
2. Convergence Criterion
The convergence criterion used is that two successive iterations do not
change the value of any variable by more than some small, specified frac-
tion. Since we want the relative accuracy to be about the same for all
variables, we normally would demand that
for all /, where A*/ = Jtf + 1 — Jtf. However, when a variable xf takes on a
value */,mjn that is so small that it ceases to have physical significance, the
program stops testing it for convergence. When it takes on a value greater
than but comparable to jc /<min , we do not need the full accuracy given by
Eq. (71), and it is replaced by
, min /k|,e c .} (72)
The largest value of
is then called the NORM at this iteration. When NORM < et for some
iteration, the program has "converged" to a solution.
Suppose the yth variable is the small difference of two large numbers.
Then in order for x} to be computed to within e r , its large constituents
must be computed more accurately. However, this also implies that if we
take ec too small, truncation and roundoff errors become significant, and
it may be literally impossible to obtain "convergence." Hence, ec must be
judiciously chosen. The value er is user-chosen, with a default value of 3
x 10 ~ 4 . This value is small enough to give satisfactorily accurate answers
but large enough to permit rapid convergence in single-precision arithmetic
while avoiding the truncation errors just described.
3. Scaling of Variables
The physical variables in Mark 5 range in magnitude from energies on
the order of 107 J to masses on the order of 10 ~ 5 kg. For a variety of
reasons, it is difficult to work numerically with sets of variables that range,
as these do, through 12 or more orders of magnitude. One problem occurs
as a result of the finite accuracy of the computer's floating-point repre-
sentation of numbers; for example, in estimating a partial derivative ac-
H. E. MITLER 741
cording to the finite difference/^ + h) — /(*), the perturbation h cannot

be too much smaller than x in magnitude when x is already small, or the
machine will compute the resulting difference as zero, or at best inaccur-
ately. Again, to have maximum accuracy in solving the simultaneous (lin-
ear) equations in the Newton mode, we need to have all of the variables
of comparable size. This is accomplished by normalizing all variables by
their then-current values, to unity, whenever the transition is made to the
NWTN mode. Then the derivatives for the Jacobian are estimated nu-
merically.
4. Accuracy
The smaller we make ec, the more accurately we should converge to an
internally consistent solution of the algebraic equations. Likewise, the smaller
we make Af () , the basic time increment, the smaller will be the errors arising
from the use of numerical approximations of various types. There are some
limitations, however. First, the smaller we make ec., the more iterations
usually have to be taken to converge. Moreover, the more likely we are
to fail to converge at all. Indeed, as noted earlier, if we demanded too
much precision we would run into roundoff and truncation errors, and
probably could never converge. The same is true of A/0.
Consider the overall (rms, fractional) error per step. The mean conver-
gence error will be per, with (3 of order unity, and the errors will be assumed
to be random. If the numerical error (e,,) is also random and, therefore,
incoherent with the convergence error (3er, the typical error in one step
will be
E, = Ve,2, + P2e2 (73)
The total error after k steps, again assuming random errors, should be
about
Ek^VkE, (74)
For a given time step Af () , the (fractional, rms) numerical error is e,,,
which is proportional to some power of A^0:
€/I = a(Af () X (75)
If the basic time increment is Ar (), then for a run of duration /, the number
of time steps will be
k > r/Af() (76)
and, therefore,
Ek > Vt VaÂfo) 2 '- 1 + P 2 e 2 /Af ( ) (77)
The value ec — 3 x 10 ~ 4 was chosen after numerical experiments on a
VAX showed that the final answers (for a 500-s run of a standard problem)
did not differ by more than a few percent for this value, compared with
calculations with smaller ec. A series of runs with A/{) = 2, 1, i, i, and ?> s
was made, with ec kept at 3 x 10~4. It was found that all of the variables
appear to converge to asymptotic values as Ar 0 -^ 0. Indeed, the deviations
from those values were very nearly linear with A£O:
a,.Afo] all / (78)
Hence, the fractional rms error at time t must be
£(0 - <a(0>Af 0 (79)
where (a) is the rms value of a,. Only for Ar0 > 2 s does the relationship
Eq. (78) begin to fail. Since the values still appear to be converging toward
the exact answer with Ar0 = g, it appears that: first, the value of A/0 for
which we get minimum error is less than i s, and, second, £ec, which is
independent of Af 0 , must be much smaller than en even for Af0 — L There-
fore, we shall take E1 — en henceforth. Third, if the numerical errors are
random, then the linearity with A/0 in Eq. (79), plus Eq. (77), suggests
that 5 = 1.
A similar analysis was carried out assuming that the numerical errors
are systematic rather than random. In that case, Eq. (79) implies that s =
2. Finally, Eq. (78) appears to show that use of the trapezoidal rule for
integration is not the principal source of numerical errors, since that error
varies as (A/0)3 for one time step, and as (Af () ) 2 overall, according to Eq.
(57), rather than linearly, as in Eq. (78).
5. Numerical Problems
The principal numerical problems with Mark 5 were of two kinds. First,
the program would not run to completion some 10-15% of the time.
Second, when the program fails to converge in the G-S mode, it switches
to the Newton method; however, that overcame the difficulty only ap-
proximately 50% of the time.
Analysis has shown that when a numerical difficulty arises, it most likely
is due to some discontinuity inadvertently built into the program (i.e., the
physics was modeled poorly). When such a discontinuity was removed, the
numerical problem would disappear.
Moreover, discontinuities in first derivatives can be troublesome with the
Gauss-Seidel method, and deadly with the Newton method since, when
the left derivative is different from the right derivative, the Jacobian is not
well defined at the transition point. Finally, as suggested by Eq. (64), it is
desirable to have the functions at least twice differentiable. Again, the cure
is to avoid modeling a physical process with an expression that has dis-
continuities in/'. If such are found, one ad hoc way to remove them is to
round the function in that region.
For many problems, the solutions will approach asymptotic values after
a certain time. One would suppose that the solver would then zip through
the problem in record time. In fact, one sometimes finds that a program
will fail in this regime. The reason is that when the variables approach
convergence at the prescribed accuracy, it is often because of stiff com-
peting processes. The small differences between relatively large numbers
H. E. MITLER 743
are basically "noise," and, hence, the Jacobians that are calculated are
very poor. The result is an amplification of the noise. Some of the oscil-
lations will be damped out by the solver, but others continue to grow and
drive the solution outside of the region of convergence. A simple solution
to this problem lies in specifying that differences which fall below the noise
level (essentially, below the prescribed error tolerance) must be set to
exactly zero.
F. FIRST Model
1. Improvements over Mark 5

A number of improvements, mainly in the physics, have been made to
Mark 5; the resulting program is called FIRST. These include the effects
of layer mixing at vents due to the formation of a turbulent shear layer at
the interface, the possibility of having forced ventilation, and a number of
other items. The complete list of improvements is too long to be given
here; it is given in Ref. 112, App. C.
FIRST also incorporates several improvements in the numerics: the pro-
gram is now run in double precision, which minimizes problems arising
from working with small differences of large numbers. Indeed, this one
change alone to Mark 5 lowered the "bombout" rate to nearly 20% of the
previous rate, i.e., to 2-3%. It was also realized that one source of dif-
ficulties was the "machine e" problem, and this was removed. The inte-
grations shown in Eq. (55) are now carried out in a separate subroutine,
so that the integration technique can easily be upgraded to Adams-Moulton,
Simpson's, Weddle's, etc.
2. Further Contemplated Changes
The use of LIMITS [see remarks preceding Eq. (62)] to help obtain
convergence is an undesirable procedure: it can lead to discontinuities,
spurious results, and mask errors and inconsistencies in the program. FIRST
has had these removed from some of the subroutines. Newer versions of
FIRST currently under development have had the numerics improved in
a number of ways, including the removal of the LIMITS section from most
of the remaining physics subroutines.
Consider the values of the /th variable at the current (t + 8r), previous
(r), and previous-but-one (t - Ar) times. Then in the expression
x.(t + 8r) - Xi(f) -f u[Xi(i) - Xf(t - Af)]8f/Ar (80)
which can be used to make an initial estimate of the value of x{ at a new
time step, CD = 0 corresponds to no extrapolation and CD = 1 to linear
extrapolation. A scheme that merits investigation is the following: nu-
merical experiments can show what value of co gives the best convergence
for a series of runs. That value can then be used as the default value, but
the user would have the option to change the value of co if he or she so
desired.
Yet another technique that should be tried is to accelerate convergence
of the iterations by some method such as Aitken's S2 procedure (see Ref.
113, p. 70); that method is valid if the sequence of iterates is converging

with an error that is approximately in geometric progression.
G. Alternative Method
As was pointed out earlier, the Newton-Raphson method will fail when
the Jacobian matrix is ill conditioned. In Ref. 114 (Chaps. 15-17), Nash
discusses the unconstrained minimization problem and shows that variable
metric (quasi-Newton or matrix iteration) algorithms are better than the
Newton method. He shows that a variant of the Levenberg-Marquardt
method134-135 is very fast and extremely robust. When det J is "large," the
method approaches the Newton algorithm, with rapid convergence. When
det J is small, however, then rather than failing altogether, the method
approaches the steepest descent algorithm, which still converges, although
only linearly. This method has not yet been implemented in FIRST, how-
ever.
H. Harvard Mark 6
Harvard Mark 6 (referred to as Harvard 6, as Mark 6, or as'CFC VI)
was created as a multiroom version of Mark 5. 115 ~ 117 The physical content
is essentially the same. One design goal for Mark 6 was to construct the
system of equations so as to eliminate the many variables as possible (see
Sec. IV.C). Emmons118 showed that the order in which the equations are
solved greatly influences the apparent number of dependent variables.
Ramsdell119 devised an algorithm that finds the ordering yielding the min-
imal set of independent variables and wrote JUGGLE, a computer program
that implements this algorithm. 120 However, the program was written in
PASCAL and is too large to be used as a subroutine in Mark 6. It has
been used to find the best ordering for the default set of variables.
Equations (41-44) are solved using the Gear104 ODE/algebraic equation
solver. The solver had to be generalized somewhat to handle Eq. (41).
The ODE part uses an implicit method; the program decides what the
time-step size should be as it goes along, based on estimates of the local
truncation error, and what order of integration/extrapolation to use. It
handles stiff equations relatively well. It is a predictor-corrector method;
the algorithm reduces the integration problem to that of solving a system
of nonlinear algebraic equations. Solving this system of equations is called
the correction and is done via a muitivariate Newton's method, which
requires an estimate of the Jacobian matrix. This is not done at every time
step, but only when the corrector fails to converge or when 30 corrections
have elapsed since the last estimation of the Jacobian; this saves computing
time. Even when the Jacobian is estimated, it is not the full Jacobian that
is estimated. First, the structure of the equations implies that the Jacobian
columns for the eliminated variables are known a priori: they are zero
except for 1's on the main diagonal. Therefore, the U/L decomposition
only needs to be applied to the upper left-hand-corner matrix, whose di-
mension is given by the number of noneliminated variables in the system,
M + M + M- Thus, the algorithm first solves for the noneliminated vari-
ables using that portion of the Jacobian. It then uses back substitution with
the remainder of the Jacobian to find values for the eliminated variables.
H. E. MITLER 745
Second, the algorithm often does not even evaluate the Jacobian for all
of the noneliminated variables. When the off-diagonal entries of J are
small, one can often obtain convergence in the corrector by assuming that
those small entries are exactly zero. In this case, J is estimated by assuming
it is the identity matrix except in those columns corresponding to root-
finding variables: for those variables, it is seldom true that the off-diagonal
entries are small compared with the diagonal ones. Therefore, these col-
umns of J are estimated in the usual manner. In this case, only the upper
left-hand-corner [/] x [/] matrix need be subject to U/L decomposition.
All of the remaining variables can be found by back substitution. Conse-
quently, this technique is by far the least expensive correction technique.
Note that when this simplified Jacobian is used, the resulting correction
method has a successive-substitution form. The package does not handle
PDE's, and so, just as in Mark 5, the calculation of the heating of objects
and walls is handled independently of the general numerical package.
The calculation switches back and forth between the Newton-like tech-
nique and the successive-substitution-like technique. It uses the latter (faster)
method until the corrector fails to converge at twice the minimum time
step and then switches to the more robust method until the faster method
begins to work again. On the other hand, since the faster method is less
stable, the integrator tends to take more time steps when it is in use,
canceling much of the savings from the smaller Jacobian decomposition.
Moreover, it is not clear that the successive-substitution method will work
for systems of equations that include other root-finding equations, because
the upper left-hand submatrix may become singular.
The integrator cannot guarantee a solution to the system of equations
unless the system satisfies certain conditions (see Ref. 115). In part because
these are not always satisfied, the program fails to converge about 10% of
the time. However, one can usually avoid such a difficulty by making a
small change in the input, which has no real physical significance (e.g.,
making a room that is 3.6 m long into one that is 3.61 m long); just one
such change will work, 90% of the time, and the result is that the program
will work, 99% of the time. Occasionally, a calculation will reach a "sen-
sitive" point, where the time steps taken will be the smallest permitted
(100 |JLS, as the default) for some time, making the total computing time
quite long. The same, of course, is true of most such computer programs.
I. Validation
The results of using Mark 5 are discussed in a number of places.98-121"126
The model has been found to predict the outcome of a number of fires
reasonably well (also see Refs. 90 and 91). A number of studies of the use
of FIRST have also been carried O ut. 127 - 128 - 136 - 137 As might be expected,
FIRST is generally an improvement over Mark 5. The results of using
Mark 6 are discussed in Refs. 116 and 117.
V. Concluding Remarks
Zone models, at present, allow for a wider range of processes to be
incorporated than field models do, while requiring much less computer
time or memory—minutes vs hours. On the other hand, the capacities of

computers are increasing at such a pace that three-dimensional field models
with radiation and combustion are becoming possible. Moreover, new nu-
merical techniques and refinements are still appearing for the solution of
PDE's.
The numerical techniques currently used in zone models differ among
each other, although some variant of a multivariable Newton-Raphson
technique is used most frequently. It is not yet clear whether it is more
expedient to solve a set of differential and algebraic equations, or a set of
differential equations with very complicated sources. The nonlinear Gauss-
Seidel technique, with certain improvements, works very well, and other
possibilities (such as accelerating convergence, using the Levenberg-
Marquardt method, etc.) exist. A less well-known but powerful technique
for solving sets of ODE's, based on a Taylor series expansion of the func-
tions, is a program called ATOM FT129', perhaps its use in this context
should be examined.
Fire modeling is in the midst of a vigorous growth phase. Efforts are
being made in at least a dozen places throughout the world to develop
perhaps double that number of models. Although progress is being made,
these are isolated efforts, with a good deal of repetition and overlap. New
programs do not profit from others nearly as much as is desirable, so "the
wheel is (continually) being reinvented." Furthermore, each individual
effort is hampered by local limitations of time, manpower, and money.
Progress—especially in incorporating more physical and chemical phe-
nomena in the models—would be much faster if the same resources now
being applied to more than 20 efforts were applied to a much smaller
number—perhaps four or five. That would still allow for alternative av-
enues to be explored.
References
*Jeng, S.-M., Chen, L.-D., and Faeth, G. M., "An Investigation of Axisymmetric
Buoyant Turbulent Diffusion Flames," Nineteenth (International) Symposium on
2
Fernandez-Pello, A. C., "Flame Spread Modeling," Combustion Science and
Technology, Vol. 39, 1984, p. 119.
3
Sibulkin, M., "Free Convection Diffusion Flames from Burning Solid Fuels,"
4
Siu, N. O., "Probabilistic Models for the Behavior of Compartment Fires,"
Univ. of California, Los Angeles, Rept. NUREG/CR-2269, 1981.
5
Rothermel, R. C., "A Mathematical Model for Predicting Fire Spread in Wild-
land Fuels," U.S. Department of Agriculture Forest Service, Missoula, MT, Re-
search Paper INT-115, 1972.
6
Ohlemiller, T., "Modeling of Smoldering Combustion Propagation," Progress
in Energy Combustion Science, Vol. 11, 1985, p. 277.
7
Mitler, H. E., "Modeling Ignition," The Effect of Cigarette Characteristics on
the Ignition of Soft Furnishings, National Bureau of Standards, Gaithersburg, MD,
NBSTN 1241, Sec. 5, 1988.
8
Ingberg, S. H., "Fire Tests of Office Occupancies," National Fire Protection
Association Quarterly, Vol. 20, 1927, pp. 243-252; also, "Tests of the Severity of
H. E. MITLER 747
Building Fires," National Fire Protection Association Quarterly, Vol. 21, 1928, pp.
43-61.
9
Kawagoe, K., "Fire Behavior in Rooms," Building Research Institute, Tokyo,
BRI Rept. 27, 1958.
10
Thomas, P. H., "Studies of Fires in Buildings Using Models Part I—Experi-
ments in Ignition and Fires in Rooms," Research Applied in Industry, Vol. 13,
1960, p. 69.
n
Simms, D. L., Hird, D., and Wright, H. G. H., "The Temperature and Du-
ration of Fires, Part I: Some Experiments with Models with Restricted Ventilation,"
Fire Research Station, Boreham Wood, England, Fire Research Note 412, 1960.
12
Gross, D., and Robertson, A. F., "Experimental Fires in Enclosures," Tenth
PA, 1965, p. 931.
13
Odeen, K., "Theoretical Study of Fire Characteristics in Enclosed Spaces,"
Royal Institute of Technology, Stockholm, Sweden, Div. of Building Construction
Bull. 10, 1963.
14
Thomas, P. H., and Hinkley, P. L., "Design of Roof Venting Systems for
Single-Storey Buildings," Fire Research TP 10, Her Majesty's Stationery Office,
London, England, 1964.
15
Thomas, P. H., Heselden, A. J. M., and Law, M. B., "Fully-Developed Com-
partment Fires—Two Kinds of Behavior," Fire Research Tech. Paper 16, Her
Majesty's Stationery Office, London, England, 1967.
16
Berl, W. G. (ed.), International Symposium on the Use of Models in Fire
Research, National Academy of Sciences, Washington, DC, 1961.
17
Williamson, R. B., "Fire Performance Under Full Scale Test Conditions—A
State Transition Model," Sixteenth (International) Symposium on Combustion, The
18
Mitler, H. E., "The Physical Basis for the Harvard Computer Fire Code,"
Harvard Univ., Cambridge, MA, Home Fire Project TR-34, 1978.
19
Lyons, J. W., Becker, W. E., Jr., Clayton, J. W., Jr., Emmons, H. W.,
Fristrom, R. M., Glassman, L, Graham, D. L., McDonald, D. W., and Nadeau,
H. G., "Fire Research on Cellular Plastics: The Final Report of the Products
Research Committee," Products Research Committee, Washington, DC, 1980.
20
Drysdale, D., An Introduction to Fire Dynamics, Wiley, New York, 1985.
21
Lyons, J. W., Fire, Scientific American Library, 1985.
22
DiNenno, P., Beyler, C. L., Custer, R. L. P., Walton, W. D., and Watts,
J. M., Jr. (eds.), SFPE Handbook of Fire Protection Engineering, National Fire
Protection Association, Quincy, MA, and Society of Fire Protection Engineers,
Boston, MA, 1988.
23
Steckler, K. D., Quintiere, J. G., and Rinkinen, W. J., "Flow Induced by Fire
in a Compartment," Nineteenth Symposium (International) on Combustion, The
24
Batchelor, G. K., An Introduction to Fluid Dynamics, Cambridge University
Press, London, England, 1967.
25
Launder, B. E., and Spalding, D. B., Mathematical Models on Turbulence,
Academic, London, England, 1972.
26
Patankar, S. V., Numerical Heat Transfer and Fluid Flow, Hemisphere, Wash-
ington, DC, 1980.
27
Markatos, N. C., and Cox, G., "Hydrodynamics and Heat Transfer in Enclo-
sures," Physico-Chemical Hydrodynamics, Vol. 5, No. 1, 1984, p. 53.
28
Atkinson, K. E., An Introduction to Numerical Analysis, Wiley, New York,
1978.
29
Ku, A. C., Doria, M. L., and Lloyd, J. R., "Numerical Modeling of Unsteady
Buoyant Flows Generated by Fire in a Corridor,' 1 Thirteenth Symposium (Inter-
national) on Combustion, The Combustion Institute, Pittsburgh, PA, 1977, p. 1373.
30
Fusegi, T., and Farouk, B., ''Numerical Study on Interactions of Turbulent
Convection and Radiation in Compartment Fires," Fire Sciences and Technology,
an International Journal (to be published).
31
Galea, E. R., "On the Field Modeling Approach to the Simulation of Enclosure
Fires," Journal of Fire Protection Engineering, Vol. 1, 1989, p. 11.
32
Larson, D. W., and Viskanta, R., "Transient Combined Laminar Free Con-
vection and Radiation in a Rectangular Enclosure," Journal of Fluid Mechanics,
Vol. 78, 1986, p. 65.
33
Patankar, S. V., "A Computer Model for Three Dimensional Flow in Fur-
naces," Fourteenth Symposium (International) on Combustion, The Combustion
34
Yang, K. T., Lloyd, J. R., Kanury, A. M., and Satoh, K., "Modeling of
Turbulent Buoyant Flows in Aircraft Cabins," Combustion Science and Technology,
Vol. 39, 1984, p. 107.
35
Bagnaro, M., Laouisset, M., and Lockwood, F. C., "Moderation d'un In-
cendie Simule," First Specialists' Meeting (International) of the Combustion Institute,
36
Saffman, P. G., "A Model of Inhomogeneous Turbulent Flow," Proceedings
of the Royal Society of London, Series A: Mathematical and Physical Sciences, Vol.
317, 1970, p. 417.
37
Rodi, W., and Spalding, D. B., Warme- und Stoffiibertragung, Vol. 3, 1970,
p. 85.
38
Jones, W. P., and Launder, B. E., "The Prediction of Laminarization with a
Two Equation Model of Turbulence," International Journal on Heat and Mass
Transfer, Vol. 15, 1972, p. 301.
39
Peyret, R., andTaylor, T. D., Computational Methods for Fluid Flow, Springer-
Verlag, New York, 1983.
40
Markatos, N. C., Malin, M. R., and Cox, G., "Mathematical Modeling of
Buoyancy Induced Smoke Flow and Enclosures," International Journal on Heat
and Mass Transfer, Vol. 25, 1982, p. 63.
41
Baum, H. R., and Rehm, R. G., "Finite Difference Solutions for Internal
Waves in Enclosures," SI AM Journal of Scientific and Statistical Computations,
Vol. 5, 1984, p. 958.
42
Rehm, R. G., Baum, H. R., and Barnett, P. D., "Buoyant Convection Com-
puted in a Vorticity, Stream-Function Formulation," NBS Journal of Research,
Vol. 87, 1982, p. 165.
43
Birkhoff, G., "Numerical Fluid Dynamics," SI AM Review, Vol. 25, 1983,
p. 1
44
Richtmyer, R. D., and Morton, K. W., Difference Methods for Initial-Value
Problems, 2nd ed., Interscience, New York, 1967.
45
Torrance, K. E., "Comparison of Finite-Difference Computations of Natural
Convection," NBS Journal of Research, Vol. 72B, No. 4, 1968, p. 281.
46
Schmidt, L. A., "A Parametric Study of Probabilistic Fire Spread Effects,"
Institute for Defence Analysis, Paper IDA P-1372, Sept. 1979.
47
Torrance, K. E., and Rockett, J. A., "Numerical Study of Natural Convection
in an Enclosure with Localized Heating from Below," Journal of Fluid Mechanics,
Vol. 36, 1969, p. 33.
48
Baum, H. R., "An Analysis of Turbulent Mixing Layers Using a Two-Equation
Model of Turbulence," Proceedings from Second U.S./Japan Government Coop-
erative Program on Natural Resources (UJNR), 1976, unpublished.
H. E. MITLER 749
49
Lewis, J., and Rehm, R. G., "The Numerical Solution of a Nonseparable
Elliptic Partial Differential Equation by Preconditioned Conjugate Gradients,"
National Bureau of Standards Journal of Research, Vol. 85, 1980, p. 367.
50
Raithby, G. D., and Torrance, K. E., "Upstream-Weighted Differencing Schemes
and Their Application to Elliptic Problems Involving Fluid Flow, 11 Computers and
Fluids, Vol. 2, 1974, p. 191.
51
Drummond, J. P., "Modeling of Supersonic Reacting Flowfields," this volume.
52
Patankar, S. V., and Spalding, D. B., "A Calculation Procedure for Heat,
Mass, and Momentum Transfer in 3-D Parabolic Flow," International Journal of
Heat and Mass Transfer, Vol. 15, 1972, p. 1987.
53
Markatos, N. C., and Cox, G., "A Field Model for Smoke Flow in Enclosures,"
First Specialists' Meeting (International) of the Combustion Institute, The Combus-
tion Institute, Pittsburgh, PA, 1981, p. 653.
54
Stone, H. L., "Iterative Solution of Implicit Approximation of Multi-Dimen-
sional Partial Differential Equation," SI AM Journal of Numerical Analysis, Vol.
5, 1968, p. 530.
55
Rehm, R. G., and Baum, H. R., "The Equations of Motion for Thermally
Driven, Buoyant Flows," NBS Journal of Research, Vol. 83, 1978, p. 297.
56
Porsching, T. A., "A Finite-Difference Method for Thermally Expandable
Fluid Transients," Nuclear Science and Engineering, Vol. 64, 1977, p. 177.
S7
Baum, H. R., Rehm, R. G., and Mulholland, G. W., "Computation of Fire-
Induced Flow and Smoke Coagulation," Nineteenth Symposium (International) on
58
Arakawa, A., and Lamb, V. R., "Computational Design of the Basic Dynam-
ical Processes of the UCLA General Circulation Model," Methods in Computa-
tional Physics: General Circulation Models of the Atmosphere, Vol. 17, edited by
J. Chang, Academic, New York, 1977, p. 173.
59
Kreiss, H., and Oliger, J., "Methods for the Approximate Solution of Time-
Dependent Problems," GARP Pub. Series 10, Global Atmospheric Research Pro-
gramme, 1973.
60
Harlow, F. H., and Amsden, A. A., "Fluid Dynamics, A LASL Monograph,"
Los Alamos Scientific Lab., Los Alamos, NM, Rept. LA4700, 1971.
61
Baum, H. R. Rehm, R. G., Barnett, P. D., and Corley, D. M., "Finite Dif-
ference Calculations of Buoyant Convection in an Enclosure, I. The Basic Algo-
rithm," SIAM Journal of Scientific and Statistical Computations, Vol. 4, 1983, p.
117.
62
Baum, H. R., and Rehm, R. G., "Numerical Computation of Large-Scale Fire-
Induced Flows," Eighth International Conference on Numerical Methods in Fluid
Dynamics, Aachen, West Germany, 1982.
63
Rehm, R. G., Baum, H. R., Barnett, P. D., and Corley, D. M., "Finite
Difference Calculations of Buoyant Convection in an Enclosure, Part II: Verifi-
cation of the Nonlinear Algorithm," Applied Numerical Mathematics, Vol. 1, 1985,
p. 515.
64
Baum, H. R., and Rehm, R. G., "Calculations of Three Dimensional Buoyant
Plumes in Enclosures," Combustion Science and Technology, Vol. 40, 1984, p. 55.
65
Handa, T. Kawagoe, K., Yoshikawa, T., Mashige, J., and Joh, T., "Numerical
Simulation of Early Stage of a Compartment Fire," Fire Science and Technology,
Vol. 4, 1984, p. 91.
66
Bagnaro, M., Laouisset, M., and Lockwood, F. C., Fire Dynamics and Heat
Transfer, HTD-Vol. 25, edited by J. G. Quintiere et al., American Society of
Mechanical Engineers, New York, 1983, p. 107.
67
Markatos, N. C., and Pericleous, K. A., Fire Dynamics and Heat Transfer,
HTD-Vol. 25, edited by J. G. Quintiere et al., American Society of Mechanical
Engineers, New York, 1983, p. 115.
68
Lloyd, J. R., Yang, K. T., and Liu, V. K., Proceedings of First National
Conference on Numerical Methods in Heat Transfer, University of Maryland, Col-
lege Park, MD, 1979, p. 142.
69
Lockwood, F. C. and Shah, N. G., "An Improved Model for the Calculation
of Radiation Heat Transfer in Combustion Chambers," Proceedings ASME-AIChE,
National Heat Transfer Conference, ASME, New York, 1976.
70
Hottel, H. C., and Sarofim, A. F., Radiative Transfer, McGraw-Hill, New
York, 1967.
71
Gosman, A. D., Humphrey, J. A. C., and Vlachos, N. S., "TEACH 3E; A
General Computer Program for Three-Dimensional Recirculating Flow," Imperial
College, London, England, Mech. Eng. Dept. Report.
72
Kumar, S., and Cox, G., "Mathematical Modeling of Fires in Road Tunnels,"
Proceedings of Fifth International Conference on the Aerodynamics and Ventilation
of Vehicular Tunnels, Lille, France, 1985, p. 61.
73
Kumar, S., Hoffman, N., and Cox, G., "Some Validation of JASMINE for
Fires in Hospital Wards," Numerical Simulations of Fluid Flow and Heat/Mass
Transfer Processes, edited by N. C. Markatos et al., Springer-Verlag, Berlin, 1986,
159.
74
Galea, E. R., and Markatos, N. C., "Mathematical Modeling of Aircraft Cabin
Fires," Applied Mathematical Mo deling, Vol. 11, 1987, p. 162.
75
Satoh, K., Yang, K. T., and Lloyd, J. R., "Ventilation and Smoke Layer
Thickness Through a Doorway of a Cubic Enclosure with a Central Volumetric
Heat Source," Eastern Section of the Combustion Institute, The Combustion In-
stitute, Pittsburgh, PA, 1980.
76
Satoh, K., Lloyd, J. R., and Yang, K. T., "Flow and Temperature Oscillation
of the Fire in a Cubic Enclosure with Ceiling and Floor Vents; III. Non-dimensional
Relationship Between Oscillatory Frequency and Heat Source Strength," Fire Re-
search Institute of Japan, Tokyo, Japan, Rept. 57, 1984.
77
Yang, H. Q., "Laminar Buoyant Flow Transition in Three-Dimensional Rec-
tangular Enclosures," Ph.D. Thesis, Univ. of Notre Dame, Notre Dame, IN, 1987.
78
Shampine, L. F., and Watts, H. A., "Software for Ordinary Differential Equa-
tions," Sources and Development of Mathematical Software, edited by W. R. Cow-
ell, Prentice-Hall, Englewood Cliffs, NJ, 1984, Chap. 6.
79
Boisvert, R. F., and Sweet, R. A., "Mathematical Software for Elliptic Bound-
ary Value Problems," Sources and Development of Mathematical Software, edited
by W. R. Cowell, Prentice-Hall, Englewood Cliffs, NJ, 1984, Chap. 9.
80
Walsh, J., "Problems of Algorithm Design for Partial Differential Equations,"
Numerical Software—Needs and Availability, edited by D. Jacobs, Academic, New
York, 1978, Sec. V.4.
81
Reid, J. K., "Software for Sparse Matrices," Numerical Software—Needs and
Availability, edited by D. Jacobs, Academic, New York, 1978, Sec. II.3.
82
Duff, I. S., "A Survey of Sparse Matrix Software;" Sources and Development
of Mathematical Software, edited by W. R. Cowell, Prentice-Hall, Englewood
Cliffs, NJ, 1984, Chap. 8.
83
Hutchings, B., and lannuzzelli, R., "Taking the Measure of Fluid Dynamics
Software," Mechanical Engineering, Vol. 109, May 1987, p. 72.
84
Hutchings, B., and lannuzzelli, R., "Benchmark Problems for Fluid Dynamics
Codes," Mechanical Engineering, Vol. 109, June 1987, p. 54.
85
Lawson, J. R., and Quintiere, J. G., "Slide-Rule Estimates of Fire Growth,"
National Bureau of Standards, Gaithersburg, MD, Rept. NBSIR 85-3196, 1985.
86
Nelson, H. E., "FIREFORM—A Computerized Collection of Convenient Fire
Safety Calculations," National Bureau of Standards, Gaithersburg, MD, Rept.
NBSIR 86-3308, 1986.
H. E. MITLER 751
87
Quintiere, J. G., "The Growth of Fire in Building Compartments," American
Society for Testing Materials, Philadelphia, PA, ASTM Special Pub. 614, 1977.
88
Siu, N. O., "Physical Models for Compartment Fires," Reliability Engineering,
Vol. 3, 1982, p. 229.
89
Chung, G., Siu, N., and Apostolakis, G., "COMPBRN II: Code Description
and Simulation of Experiments," Univ. of California, Los Angeles, Rept. UCLA-
ENG-8404, 1984.
90
Parikh, J. S., Beyreis, J. R., and Christian, W. J., "Survey of the State of the
Art of Mathematical Fire Modeling," Underwriters Labs., Northbrook, IL, Rept.
File NC 554, 1983
91
Mitler, H. E., "Comparison of Several Compartment Fire Models: An Interim
Report," National Bureau of Standards, Gaithersburg, MD, Rept. NBSIR 85-
3233, 1985.
92
Tanaka, T., "A Model of Multiroom Fire Spread," National Bureau of Stand-
ards, Gaithersburg, MD, Rept. NBSIR 83-2718, 1983; also, Fire Science and Tech-
nology, Vol. 3, 1983, p. 105.
93
Jones, W. W., "A Multicompartment Model for the Spread of Fire, Smoke,
and Toxic Gases," Fire Safety Journal, Vol. 9, No. 1-2, 1985, p. 55.
94
Cooper, L. Y., and Forney, G. P., "Fire in a Room with a Hole: A Prototype
Application of the Consolidated Compartment Fire Model (CCFM) Computer
Code," Proceedings of Eastern Section of Combustion Institute, National Bureau
of Standards, Gaithersburg, MD, 1987, p. 104.
95
Morton, B. R., Taylor, G. I., and Turner, J. S., "Turbulent Gravitational
Convection from Maintained and Instantaneous Sources," Proceedings of the Royal
Society of London, Series A: Mathematical and Physical Sciences, Vol. 234, 1956,
P.I.Modak, A. T., "Radiation from Products of Combustion," Fire Research, Vol.
96
1, 1978/79, p. 339.
97
Mitler, H. E., and Emmons, H. W., "Documentation for CFC V, the Fifth
Harvard Computer Fire Code," Harvard Univ., Cambridge, MA Home Fire Proj-
ect T.R. 45, 1981; also, National Bureau of Standards, Gaithersburg, MD, Rept.
GCR 81-344, 1981.
98
Emmons, H. W., "The Prediction of Fires in Buildings," Seventeenth Sym-
1979, p. 1101.
"Quintiere, J. G., "An Approach to Modeling Wall Fire Spread in a Room,"
Fire Safety Journal, Vol. 3, 1981, p. 201.
100
Abramowitz, M., and Stegun, L A . (eds.), Applied Mathematics Series: Hand-
book of Mathematical Functions, No. 55, National Bureau of Standards, Gaith-
ersburg, MD, 1964.
101
Young, T. R., and Boris, J. P., "A Numerical Technique for Solving Stiff
Flow Problems," Journal of Physical Chemistry, Vol. 81, 1977, p. 2424.
102
Oran, E., Young, T., and Boris, L, "Application of Time-Dependent Nu-
merical Methods to the Description of Reactive Shocks," Seventeenth Symposium
p. 43.
103
Young, T. R., "CHEMEQ—A Subroutine for Solving Stiff Ordinary Differ-
ential Equations," Naval Research Lab., Washington, DC, NRL Memo. Rept.
4091, 1980.
104
Gear, C. W., "Simultaneous Numerical Solution of Differential-Algebraic
Equations," IEEE Circuit Theory Transactions, Vol. 18, Jan. 1971, pp. 89-95.
105
Watts, H. A., and Shampine, L. F., "DEPAC—Design of a User-Oriented
Package of ODE Solvers," Sandia Labs., Albuquerque, NM, Rept. SAND-79-
2374, 1979.
106
Radhakrishnan, K., "Combustion Kinetics and Sensitivity Analysis Compu-
tations," this volume.
107
Aiken, R. (ed.), Stiff Computation, Oxford University Press, England, 1985.
108
Smith, G. D., Numerical Solution of Partial Differential Equations, Oxford
University Press, New York, 1965.
109
Scarborough, J. B., Numerical Mathematical Analysis, 5th ed., Johns Hopkins
Press, Laurel, MD, 1962.
110
Forsythe, G., and Moler, C. B., Computer Solutions of Linear Algebraic
Systems, Prentice-Hall, Englewood Cliffs, NJ, 1967.
H1
Jarrat, P., Numerical Methods for Nonlinear Algebraic Equations, edited by
P. Rabinowitz, Gordon and Breach, NY, 1970.
112
Mitler, H. E.-, and Rockett, J. A., "Users1 Guide to FIRST, a Comprehensive
Single-Room Fire Model," National Bureau of Standards, Gaithersburg, MD,
Rept. NBSIR 87-3595, 1987.
113
Henrici, P., Elements of Numerical Analysis, Wiley, New York, 1964.
114
Nash, J. C., Compact Numerical Methods for Computers, Wiley, New York,
1979.
115
Gahm, J. B., "Computer Fire Code VI—Volume 1," Harvard Univ., Cam-
bridge, MA, Rept. TR-58; also, National Bureau of Standards, Gaithersburg, MD,
Rept. NBS-GCR 83-451, 1983.
116
Rockett, J. A., and Morita, M., "The NBS/Harvard VI Multi-Room Fire
Simulation," Fire Science and Technology, Vol. 5, No. 2, 1985, p. 159.
117
Rockett, J. A., Morita, M., and Cooper, L. Y., "Comparison of NBS/Harvard
VI Simulations and Full-Scale Multiroom Test Data," Fire Safety Journal, Vol. 15,
1989, p. 115.
118
Emmons, H. W., "A Note on Minimizing the Unknowns for Computation of
a Large System of Equations," Division of Applied Sciences, Harvard Univ.,
Cambridge, MA, Home Fire Project TR, 26, 1978.
119
Ramsdell, J. D., "Finding Minimal Feedback Vertex Sets," Harvard Univ.,
Cambridge, MA, Home Fire Project TR, 47, 1981.
120
Ramsdell, J. D., "Juggle User's Guide," Applied Sciences Div., Harvard
Univ., Cambridge, MA, 1981.
121
Mitler, H. E., and Rockett, J. A., "How Accurate Is Mathematical Fire
Modeling?," US-USSR Joint Seminar on Mathematical Modeling of Fire, Tbilisi,
USSR, July 1981; also, National Bureau of Standards, Gaithersburg, MD, Rept.
NBSIR 86-3459, 1986.
122
Rockett, J. A., "Modeling of NBS Mattress Tests with the Harvard Mark V
Fire Simulator," National Bureau of Standards, Gaithersburg, MD, Rept. NBSIR
81-2440, 1982.
123
Rockett, J. A., "Park Service Room Fire Test Simulations Using the Harvard
Level 5.2 Computer Fire Model," National Bureau of Standards, Gaithersburg,
MD, Rept. NBSIR 83-2805, 1984.
124
Rockett, J. A., Morita, M., and Handa, T., "Using the Harvard Fire Simu-
lation," Fire Science and Technology, Vol. 3, No. 1, 1983, p. 57.
125
Blomqvist, J., and Andersson, B., "Modelling of Furniture Experiments with
Zone Models," Fire and Materials, 1984.
126
Mitler, H. E., "The Harvard Fire Model," Fire Safety Journal, Vol. 9, No.
1-2, 1985, p. 7.
127
McCaffrey, B. J., and Dagher, S. N., "The Use of FIRST for Forced Coun-
terflow-Ventilated Fires," 1989.
H. E. MITLER 753
128
Duong, D. Q., "The Accuracy of Computer Fire Models: Some Comparisons
with Experimental Data from Australia/ 1 Fire Safety Journal (submitted for pub-
lication).
129
Corliss, G., and Chang, Y. F., "Solving Ordinary Differential Equations Using
Taylor Series," ACM Transactions on Mathematical Software, Vol. 8, No. 2, 1982,
p. 114.
13()
Roache, P. J., Computational Fluid Dynamics, Hermosa Publishers, NM, 1976.
131
Spalding, D. B., "The Calculation of Free Convection Phenomena in Gas-
Liquid Mixtures,' 1 Heat Transfer and Buoyant Convection, edited by Spalding and
Afgan, Hemisphere, Washington, DC, 1977.
132
Ralston, A., A First Course in Numerical Analysis, McGraw-Hill, NY, 1965.
133
Kahaner, D. K., Moler, C., and Nash, S., Numerical Methods and Software,
Prentice-Hall, Englewood Cliffs, NJ, 1989.
134
Levenberg, K., "A Method for the Solution of Certain Nonlinear Problems
in Least Squares, 1 ' Quarterly of Applied Mathematics, Vol. 2, 1944, p. 164.
l35
Marquardt, D. W., "An Algorithm for Least-Squares Estimation of Nonlinear
Parameters," Journal of the Society for Industrial and Applied Mathematics, Vol.
11, 1963, p. 431.
136
Gross, D., and Davis, W. D., "Burning Characteristics of Combat Ship Com-
partments and Vertical Fire Spread," National Institute of Standards and Tech-
nology, Gaithersburg, MD, NISTIR 88-3897, Dec. 1988.
137
Todd, N. W., and Ryan, J. D., "Better Codes Via Fire Technology," Pro-
ceedings of the Pacific Rim Conference of Building Officials, International Con-
ference of Building Officials, April 1989, p. 293.
Chapter 24
Nuclear Systems
J. R. Travis, W. S. Gregory, and F. R. Krause
I. Introduction
T HE occurrence of a number of severe fire accidents in nuclear facilities

has motivated research activities to model combustion phenomena in
these systems. The most publicized accidents are the electrical wire insu-
lation (cable-tray) fire at the Brown's Ferry nuclear power plant 1 and the
hydrogen combustion event during the Three-Mile Island accident.2 Severe
fires involving small radiation confinement compartments (glove boxes)
and high-efficiency particulate air (HEPA) filters have occurred in the
Rocky Flats facility.3 Significant, but less severe, flammable-liquid fires
have occurred in diesel generator rooms and turbines. 4 Table 1 presents a
summary of the combustible materials in nuclear facilities.
It is important to analyze entire facilities to determine the transient
pressure and heating loads due to combustion events on the containment
or ventilation structures and safety-related equipment. These analyses are
also needed to plan emergency evacuation, to validate existing fire pro-
tection programs or suggest changes, and to integrate fire safety with other
safety analysis requirements. Fire-safety requirements are different for
nuclear reactor containments than for nonreactor sytems such as nuclear
fuel cycle facilities.
Nuclear reactor systems are designed to produce power or nuclear ma-
terials. A disruptive reactor transient, in which the nuclear core loses much
of its cooling capacity, may lead to a severe accident.5 For example, roughly
750 kg of hydrogen was generated during the Three-Mile Island accident
by the oxidation of approximately one-half of the zirconium fuel cladding
in the core. Because of thermal and nuclear properties, zirconium is often
used as a cladding or thin protective metal layer covering the nuclear fuel.
In the unlikely event of a reactor pressure vessel failure, additional hy-
drogen and carbon monoxide could be generated by the attack of degraded
or molten core materials on the concrete basement and support structures.
755
756 NUCLEAR SYSTEMS
Table 1 Nuclear facility fire hazards
Type of Type of
combustion Fire location fire growth Hazards
Cable-tray fires Containment Local accumulation Loss of
(electrical building; cable of flammable nuclear- or
insulation) spreading and toxic vapors fire-safety
systems;
reignition of
deep-seated
fires
Filter fires Confluence of fire Ignition of Failure of
products and additional filter supply or
fresh air fires and duct exhaust fans
fires and HEPA
filters
Flammable- Diesel generator Flame elongation Failure of fire
liquid fires fuel; coolant by external barriers, fire
pump lubricants feedback of suppression
oxygen-depleted systems, and
air; wall safety-
temperature rise related
in sealed equipment
containment
buildings
Hydrogen fires Tritium Accumulation and Deflagrations/
production combustion of detonations;
facilities; flammable gases; rupture of
nuclear reactor hydrogen release radioactive
containment into ventilation containment
buildings; system; reversal barriers;
reactor off-gas of ventilation survival of
systems system flows safety-
related
systems
Solid Glove boxes and Gas temperature Release of
hydrocarbon HEPA filters in below flashover smoke,
fires forced threshold radioactivity,
ventilation or toxic
chemicals
into
ventilation
passages
Thus, fires in nuclear reactor systems resulting from severe accidents may
initiate and subsequently propagate during a transient release of heat,
carbon monoxide, and hydrogen.
Nonreactor systems involve activities other than power or nuclear ma-
terial production. Typical activities include glove box operations, fuel fab-
rication, and spent fuel reprocessing or storage. Industrial hygiene usually
requires that such operations be conducted in closed rooms in which a
plant ventilation system prevents the release of radioactivity or toxic chem-
J. R. TRAVIS ET AL. 757
icals by maintaining a negative pressure differential toward contaminated

areas during normal operations. These systems have complex ventilation
systems that exhaust air from multiple rooms or laboratories.
Section II reviews traditional fire-analysis methods 6 and identifies those
capabilities that can be used for modeling nuclear facilities. Additional
capabilities are outlined that will be needed to accommodate the unique
architecture and design-basis accidents of nuclear facilities. In subsequent
sections, the Los Alamos modeling approach is presented for both non-
reactor and reactor systems.
II. Overview of Numerical Models

Traditional models simulate the release and motion of combustion prod-
ucts in residential and commercial buildings such as hotels, shopping malls,
and warehouses. This traditional approach describes combustion by scaling
empirical heat- and mass-release data with the surface area of the irradiated
fuel. 7 The movement of combustion products in a building is calculated by
tracking the evolution of hot-gas layers in the burn room and in adjacent
rooms.8
Building-fire models by themselves cannot be used for nuclear facilities
because the unique architecture of nuclear facilities introduces special modes
of fire growth. In addition, fire hazards exist in nuclear facilities that are
absent in commercial buildings. Most of the observed fire-growth phenom-
ena and design accidents listed in Table 1 fall into this special category.
Traditional fire models must be extended to represent the special char-
acteristics and architecture of nuclear facilities. The associated modeling
requirements are derived subsequently by reviewing traditional models for
heat release, combustion-product formation, and the motion of combustion
products. The desired additional capabilities are derived by comparing
traditional models with existing lumped-parameter codes and transient,
three-dimensional conservation equations (Navier-Stokes equations or field
models). This chapter identifies the lumped-parameter codes and field
models that can pedict combustion-product flows in a room where fuel is
burning (burn room) and throughout a nuclear facility.
A. Traditional Fire Models

Much research and testing have gone into the development of computer
models that reliably predict the release and motion of combustion products
in buildings.6 The following review identifies which of these models are
applicable to both nuclear facilities and buildings.
1. Release of Combustion Products and Heat

The rate of combustion-product formation per irradiated fuel area is
calculated in a two-step process, which is described mathematically in Sec.
III. The first step is to estimate the heat flux to the fuel by tracking the
exchange of heat radiation among the fuel and surrounding fires, hot layers,
and hot walls. The net heat across the radiation-exposed fuel surface then
is converted to the mass-burning rate of hydrocarbon fuels by multiplying
it with an empirical heat of gasification. This heat has been measured in
small-scale calorimeter fire tests. 7
758 NUCLEAR SYSTEMS
The second step is to convert the volatilization mass-burning rate to a

rate of heat release and smoke release using two additional empirical pa-
rameters. The first parameter is an apparent heating value that has been
measured in compartment-size9 calorimeter tests. Multiplying the mass-
burning rate by this parameter gives the rate of heat release. The second
parameter denotes the "theoretical" heating value that has been measured
in an oxygen bomb calorimeter. The ratio of these two heating values is
the combustion efficiency; deviations of the total conversion efficiency from
1.0 are used to estimate smoke release. Thus, in this approximation, smoke
denotes volatiles that will not burn in a room fire, but will burn in an
oxygen bomb calorimeter. This smoke is made up mostly of soot and may
include unburned flammable gases such as CO or fuel droplets.
Calorimeter tests measure oxygen consumption, which then can be used
to estimate the heating value by assuming that 1 g of oxygen consumption
releases 13 kJ of heat regardless of fuel configuration, combustion envi-
ronment, or chemistry.9 The apparent heat of combustion is equivalent to
an entrainment coefficient. Thus, the combustion efficiency should de-
crease and smoke release should increase with oxygen depletion in the
combustion environment. This decrease has been observed in underven-
tilated calorimeter fire tests,10 where the presence of oxygen-depleted air
is indicated by increased gas temperatures.
Forced-ventilation compartment-fire tests11 show that oxygen intake by
complete combustion increases by a factor of 2 when little or no oxygen
is available through entraining the burn-room atmosphere.12 Therefore,
numerical models of combustion-product release in nuclear facilities should
track all constituents that could be recirculated to the fuel inside and outside
the burn room. These constituents are listed in Table 2. Inefficient com-
bustion and the associated release of smoke and unburned flammables in
a ventilation-controlled fire can occur when the fire becomes fuel rich, i.e.,
when the flow of oxygen into the burn room is low.
The extension of the building-fire release rates to nuclear facilities may
be unsatisfactory because empirical factors, which do not change during
the oxygen-rich stage of a fire, change with the environment of ventilation-
controlled fires; i.e., they are sensitive to temperature, smoke, flammable
gas, and amount of oxygen. Accurate release rates for ventilation-con-
trolled fires may require an extraordinary number of calorimeter fire tests
unless other approaches to entrainment coefficients can be found that are
insensitive to the accumulation of combustion products near the fire. One
thermodynamic approach is to predict entrainment from wall heat losses
and experimental flame temperatures.13 Wall heat losses are available from
lumped-parameter codes or three-dimensional field models, as discussed
in Sees. Ill and IV, or might be estimated from compartment-fire tests.
They are known to be insensitive to the accumulation of combustion prod-
ucts in all ventilation fire tests. Empirical flame temperatures for venti-
lation-controlled fires are available from investigations of flames burning
near extinction,14 These temperatures are insensitive to fuel chemistry and
the combustion environment. An initial application of the thermodynamic
approach generated acceptable pretest values for crib and pool-fire burning
rates.13
Table 2 Simulation capabilities of lumped-parameter codes
Phenomenon
modeled Desired capabilities of models COMPARE model CONTAIN model FIRAC model
Fluid components CO2, CO, H2O, H2, O 2, N 2: All except smoke and All except smoke and All except H2
smoke; water droplets; fuel fuel vapor fuel vapor
vapor
Continuous diffusion Fuel vapor combustion, H2 H2 combustion at H2 and CO None
reaction rate combustion, CO fixed rate combustion at
combustion, temperature- fixed rate
dependent rates
Wall heat transfer Wall variation, time All except free All except free Forced convection
variation, forced convection and convection and and radiation with o
convection, free wall variation wall variation time variation
convection, radiation,
evaporation, condensation
Radioactive particle Source terms, mass fluxes in None Source terms and Souce terms,
concentration ventilation passengers,
filter fires, three-
mass fluxes ventilation
passages, fluxes =1>
dimensional circulation in
selected rooms
Ventilation controls Fans, filters, control Fans and doors Fans and doors All
dampers, doors
Bidirectional flow Entrainment by fire plumes, None None None
entrainment by vent jets,
radiation heat exchange
with fuel
Particle motion Gravitational settling, Drop condensation Gravitational settling, Gravitational settling
resuspension, drop and evaporation drop condensation, and resuspension
condensation, and drop and evaporation
evaporation
en
CD
760 NUCLEAR SYSTEMS
2. Motion of Combustion Products

Building-fire models describe the motion of combustion products through
the evolution of hot-gas layers in the burn room and adjacent rooms.8 A
set of coupled ordinary differential equations is used to predict the time-
dependent composition and temperature of individual layers from the con-
servation laws for mass, momentum, and energy. The locations of gas layers
are selected by the modeler and represent local heat and mass exchange
phenomena, which are represented by source terms. Corridors are treated
as other rooms. The models are restricted to multiroom architectures and
open-door ventilation passages.
Certain local phenomenological parameters are needed for a reliable
prediction of gas layers outside the burn room. For unidirectional flow,
we need to know the time variation of heat-transfer coefficients at the wall
and the mass flux into and out of individual layers. For bidirectional flows,
we need to know the amount of entrainment of fire plumes, the amount
of entrainment of jets through doors, and the radiation heat exchange
among fire plumes, hot layers, and walls.
The idealization of three-dimensional flows as hot layers with bidirec-
tional exchange is the key for operating building-fire models on very modest
computers, e.g., personal computers. This simplification may not be de-
sirable in nuclear facilities because the locations where combustion prod-
ucts and heat accumulate often cannot be supplied as input and may not
take the form of horizontally extended hot layers. The heat- and mass-
exchange modules used for building-fire simulations are independent of
the location of local accumulations and are applicable to nuclear facilities
as well as to buildings.
B. Desired Capabilities for Nuclear Facility Models

The building-fire models cannot be used directly to model nuclear fa-
cilities because the architecture of nuclear facilities differs from standard
corridor-connected rooms. The following desired modeling capabilities are
based on such architectural differences.
1. Nonreactor Facilities
Nonreactor facilities have ventilation sytems that are designed to prevent
the accumulation of toxic or radioactive products in rooms where people
are active. Thus, the development of horizontally extended hot-gas layers
is unlikely and may violate industrial hygiene requirements. However, the
local accumulation of radioactive constituents, combustion products, and
unburned flammables must be accounted for in the model.
The ability to predict local accumulations of heat and combustion prod-
ucts is needed to avoid corrosion and heat loads on safety systems, such
as fire detectors, safety sensors, computer chips, and the degradation of
ventilation control systems, such as control dampers, fans, and filters. The
capability to predict the temperature and concentration of locally accu-
mulated flammable gases or fuel vapors is needed to predict fire spread
by fireballs and duct fires.
Models for calculating sources of radioactive constituents are needed to
simulate nuclear-facility-specific design accidents. These models might be
the resuspension of radioactive particles from burn-room floors by locally

high velocities and from smoke produced from combustion in glove boxes
or contaminated storage devices, and clothing.
The release of radioactive trace constituents may violate industrial hy-
giene regulations. The concentration of radioactive trace constituents must
be predicted throughout the entire facility to identify such potential vio-
lations. Thus, we need to model the three-dimensional flowfield in rooms
with glove boxes or in nuclear-material testing laboratories, and we will
need to model particle motion with gravitational deposition and resuspen-
sion throughout the facility.
Filter fires occur because flammable liquids accumulate in or on the
filter, and the filter is exposed to high temperatures. Vital filter banks are
usually protected by automatic water sprinkler sytems; therefore, the ability
to simulate droplet evaporation and wall heat transfer is needed to reliably
predict the temperature of filters, local combustion products, and concen-
trations of flammable gases upstream of filters.
2. Reactor Facilities
Reactor facilities are tall, sealed containment buildings. The combustible
materials are flammable liquids and cable trays. The combustion of hy-
drogen and carbon monoxide, which are generated during severe accidents,
requires special additional modeling capabilities.
Flammable-liquid fires produce large amounts of heat and may raise
temperatures throughout the containment building by more than 100°C.
Strong updrafts move hot air and moderate amounts of smoke throughout
the building and control the oxygen supplied to the fire location. Fires
become ventilation-controlled very quickly, even without a ventilation sys-
tem. The ability to model the feedback of heat and smoke to the fuel is
needed because this feedback affects the rate of release of heat and com-
bustion products.
New methods for estimating flame height also are needed because elon-
gated flames may raise local temperatures to flashover thresholds (600°C)
or spread out of the burn room. Current research has shown that flames
burning near the extinction limit elongate because they proceed by the
reignition of locally fuel-rich pockets of flammable vapors. 15 Raising the
inlet gas temperature should extend flames because reignition requires less
heat release by local combustion. Elongated flames become possible when
oxygen-depleted air is fed back to the burn room. Such feedback may be
generated by three-dimensional flows in the vicinity of the burn room ("a
bathtub vortex") that cause bidirectional flows in vertical flow passages.
Flames may then extend out of the burn room in search of oxygen, because
oxygen is depleted from the gas flow into the burn room.
The ability to predict flame elongation by heating of inlet air or recir-
culation of oxygen-depleted air is very important for reliable prediction of
fire spread. Simple models of a temperature-dependent and continuous
diffusion reaction rate have been successful in predicting the response of
oxygen-depleted air. This simple model is discussed in Sec. IV.
Cable fires burn slowly but produce very large amounts of smoke, toxic
combustion products (hydrochloric acid), and flammable vapors. The same
762 NUCLEAR SYSTEMS
cable tray can release combustion products at very different rates, de-
pending on both the temperature and the smoke in the combustion envi-
ronment.16 Thus, hazards due to cable fires are sensitive to ventilation
conditions because the slow heat-release rate does not break up local and
extended accumulations of smoke and corrosive or flammable gases. There-
fore, the concentrations of combustion products must be modeled accu-
rately throughout the facility. The same capability is needed in reactor
facilities because of health hazards and the degradation of safety or com-
puter systems by acid and hot cable-fire smoke.
Accumulations of hydrogen and carbon monoxide may burn locally as
diffusion flames, burn as deflagrations in premixed volumes, or detonate
under certain combinations of local temperature and flammable-gas con-
centration. Some modeling advances for simulating such fires have been
made and are discussed next.
C. Available Advanced Models
Advanced lumped-parameter computer codes and field equation models
have been developed for preparing nuclear safety analysis reports. Each
of these models can provide many of the desired capabilities. Table 2
compares the available and desired capabilities to illustrate mature models
that are developed, or nearly developed, for predicting the spread of flam-
mable gases and hazardous combustion products throughout a nuclear
facility.
1. Lumped-Parameter Codes
The local accumulation and combustion of hydrogen and carbon mon-
oxide in reactor facilities have been simulated with lumped-parameter codes
such as COMPARE17 and CONTAIN.18 The combustion of hydrocarbon
fuels in nonreactor nuclear facilities has been simulated by using the FIRAC
code.19 All three codes have been compared with experiments and together
are considered to represent the current state of the art in modeling com-
bustion in nuclear facilities. Table 2 gives a detailed review of these three
lumped-parameter codes.
There are other lumped-parameter codes that we know less about. The
COBRA-NC20 code was derived from a three-dimensional field equation
model by eliminating the momentum flux, turbulent shear stress, and vis-
cous shear stress terms from the momentum equation and the turbulent
diffusion term from the mass and energy equations. The RALOC21 code
was developed for safety analysis of reactor facilities in West Germany.
Both codes have been verified partially in full-scale fire tests.
Table 2 compares the current modeling capabilities available in lumped-
parameter codes with the set of desired modeling capabilities that were
identified in the preceding section. All capabilities are grouped according
to seven important phenomena that control the spread of fire and the
hazardous exposure of people and safety equipment. The comparison shows
that six out of seven phenomena may be simulated with lumped-parameter
codes. None of the codes provides for all capabilities that are necessary to
simulate the relevant phenomena. However, a complete coverage of de-
sired capabilities could be available by combining the FIRAC code with
one of the other codes. In Sec. Ill, we discuss FIRAC to illustrate com-
bustion analysis with lumped-parameter codes.
2. Field Equation Models

Three-dimensional field equation models are needed to calculate local
phenomena such as flammable or toxic gas component concentrations and
heat and mass exchange in bidirectional flows. Lumped-parameter codes
cannot simulate such phenomena. Hydrogen accumulation and combustion
in large containment buildings have been predicted by the three-dimen-
sional COBRA-NC20 and HMS22 codes, and hydrocarbon combustion in
nuclear facilities was simulated by using the PHOENIX23 and HMS24 codes.
These three field equation models have been verified partially in full-scale
fire tests.
When field equation model capabilities are compared with desired bi-
directional transport modeling capabilities, we find that the HMS code is
the only code that has been applied to both hydrocarbon and flammable-
gas combustion. Therefore, we selected it to illustrate the application of
field equation models to nuclear facilities.
When field equation modeling capabilities are compared with those de-
sired for unidirectional flows, lumped-parameter codes often are found to
be more applicable then field equation models. This is most obvious for
small rooms and narrow ventilation passageways in nonreactor facilities.
The best model for combustion in nuclear facilities most likely will use
lumped-parameter codes to link the inflow and outflow to multidimensional
field equation models. This hybrid model could represent burn rooms with
complex architecture and buoyant plumes in large rooms, whereas lumped-
parameter models could connect one-dimensional pathways between the
larger volumes.
III. Nonreactor Systems

A. Introduction
By nonreactor systems, we refer to processes or operations in facilities
dealing with, for example, nuclear fuel fabrication, fuel reprocessing, and
spent-fuel storage. Accidents involving combustion in these facilities are
considered by the Department of Energy and the U.S. Nuclear Regulatory
Commission (USNRC) in their safety analyses. Los Alamos National Lab-
oratory has developed a handbook devoted to the analysis of accidents in
these types of facilities for the USNRC. 25
The unique features of these facilities that affect combustion accidents
in nonreactor nuclear facilities are the forced ventilation in a confined
volume, complex network flow paths, and the presence of radioactive
materials. Nuclear facilities do not have open doors and windows. Air is
brought into, through, and out of the facility by forced-ventilation systems.
The air passes from less contaminated zones to more contaminated zones.
These forced-ventilation systems can consist of hundreds of ducts, many
rooms and glove boxes, multiple fans, and many filtration devices all in
an interconnected network. The possibility of fire in these systems, the
involvement of radioactive materials, and the fact that the ventilation sys-
764 NUCLEAR SYSTEMS
tern provides the primary pathway to the environment have led to the
development of numerical models to analyze these systems for combustion
accidents.
B. Numerical Method and Models

The numerical method we developed for modeling nonreactor systems
that involve forced-ventilation systems is a lumped-parameter method. In
this method, the interconnected ductwork is described by linkages that are
called elements or branches and by junction points called nodes. The models
used to describe fires in nonreactor systems must include submodels for
combustion, gasdynamics, heat transfer, and material transport. These
models are discussed in detail next.
1. Combustion Model
The combustion model describes the behavior of a fire in a compartment
containing radioactive material. The combustion model gives the energy
release and quantity of smoke and radioactive material that become air-
borne during the fire. This source of material and the calculated thermal
energy generated provide the driving potential for the gasdynamics dis-
cussed in the next section. The gasdynamics, along with the heat-transfer
and material-transport models, distribute the effects of the fire throughout
the rest of the facility and the release of radioactive contaminants into the
environment.
The combustion model was developed by Battelle Pacific Northwest
Laboratories (PNL) and is composed of three basic modules that interact
with each other: 1) fire source term, 2) heat and mass transfer, and 3)
radioactive source term. 26 The fire source-term module is based on Te-
warson's theory of steady-state heat balance on the surface of an element
of burning material. Special experiments were carried out at Factory Mu-
tual Corporation in Norwood, Massachusetts, to determine the energy,
particulates, and gases released from combustible materials commonly found
in nonreactor system facilities.10 The model for the fire source-term equa-
tions is developed subsequently.27
Mass loss rate (M). The mass loss rate depends on the heat flux on the
material in a fire and the heat required to generate a unit mass of com-
bustible vapors,
M = qJL (1)
where M is the mass loss rate of the material, qn the heat flux on fuel per
unit fuel surface area, and L the heat required to generate a unit mass of
fuel vapor. The mass loss rate because of pyrolysis is
Mp = (qe - qrr)IL (2)
where (qe - qrr) is the net heat flux, qe the external heat flux, and qrr the
surface reradiation heat loss. The mass loss rate from combustion is
Mb = (qe + qfs ~ qrr)/L (3)
where (qe + qfs — qrr) is the net heat flux and qfs the net heat transfer
due to combustion.
Heat-release rate (Qa). The heat-release rate is calculated to determine
the rate of fire growth and the size of the fire. This is expressed as
Qa = HaMb (4)
where Ha is the heat of combustion. The ratio of actual to complete heat
of combustion is HJHt = Xa, so that
Qa = XaHtMh (5)
or
Qa = Xa(HtIL}(qe + qfs - qn.} (6)
Heat and mass balance. The heat- and mass-transfer module uses a
two-layer (hot and cold) model. This type of model frequently is referred
to as a zone model. Each layer is assumed to be homogeneous, or well
mixed. Radiative, convective, and conductive heat transfer to compartment
boundaries (walls, ceiling, floor) and equipment in the room are included
in the model, as is a mass balance of oxygen, nitrogen, and five different
combustion products. Fluid flows, such as ventilation through the com-
partment and entrainment of air from the fire plume, also are modeled.
The major types of heat-transfer phenomena considered in the zone model
are direct radiative transfer, convective heat transfer, and conductive heat
transfer. The individual models for these processes are found in Ref. 27.
The expression for the heat balance within the burning zone is developed
subsequently.
The rate of change of internal energy in the hot layer is
C^V^THL), - p2V2(THL)2]/& (7)
where Cp is the specific heat capacity of the hot layer at constant pressure;
pj and p2 the densities of the hot layer at new and old times, respectively;
Vl and V2 the volumes of the fire-zone compartment occupied by the hot
layer at present and previous times, respectively; (THL)^ and (THL)2 the
hot-layer temperatures at the present and previous times, respectively; and
Af the time-step increment.
The flux of enthalpy into (out of) the hot layer after differentiation equals
the heat gained (lost) by convection and radiation from (to) the ceiling,
floor, and walls of the fire compartment plus the heat gained (lost) by
ventilation to (from) the hot layer. This is expressed as
(Qrlc + Qrlw + Qrlf + Qclc + Gc/vv +
+ mCp[(THL)2 - rj/Af (8)

where Qrlc, Qriw, and Qrlfare radiative heat transfer between the hot layer
and the ceiling, the wall, and the floor, respectively; m is the mass of hot
gases in the smoke layer obtained from the mass balance; and Tf is the
initial temperature of the compartment before the fire.
766 NUCLEAR SYSTEMS
The convective energy from burning added to the hot layer is

(9)
The heat lost to equipment at risk inside the fire compartment after dif-
ferentiation equals the heat lost by hot-layer convection and radiation to
equipment, or
(Qrle + Gc/e)/Ar (10)
where Qrle is the hot-layer radiation to equipment and Qcle the hot-layer
convection to equipment. The enthalpy in the hot layer after differentiation
is
p2V2Cp[(THL)2 - n (11)
Combining the preceding equations and solving for (THL\, we get
(THL\ = (THL)2 + — - {Qrhl + Qrlc + Qrlw + Qrl
+ Qci* + Qdf + mCp[(THL}2 - TJ + Qc - Qcle (12)

)2 - Tt]}
This equation is used to predict the advanced time hot-layer temperature.
The expression for the mass balance in the fire compartment incorporates
the assumption that no mixing occurs between layers. The ideal gas law is
used to obtain total mass and molar information for completing the gas-
dynamics and heat-transfer calculations. Oxygen concentrations can be
calculated from the mass conservation equations, and this information is
then used to determine the burning rates.
Smoke-generation rate. The smoke-release rate can be calculated by
using
Gs = xsMb (13)
where xs is a fractional yield of smoke. Tabulated values of Xa and xs for
various materials are shown in Table 3.
The fire source term, heat and mass release, and radioactive and com-
bustion material generation can be used as input to the rest of the network.
The effect of the combustion process can be determined by using gasdyn-
amics, heat transfer, and material transport throughout the rest of the
system.
Radioactive source generation rate. The radioactive source term gives
the release rate of airborne radioactive particles from burning contaminated
combustible solids, burning contaminated liquids, heating contaminated
surfaces, heating radioactive liquids, overpressurization of closed con-
tainers containing radioactive powders and liquids, and burning radioactive
pyrolytic metals. The models used are empirical and based on a series of
experiments performed at PNL.28 Table 4 shows the equations used to
model the release rate from burning of contaminated combustibles.
Table 3 Smoke and energy burn characteristics
Material x.a x*
Polymethylmethacrylate 0.94 0.21
Polyvinylchloride 0.35 0.086
Polystyrene 0.68 0.15
Cellulose 1.00 0.001
Polychloroprene 0.41 0.15-0. 38a
Kerosene 0.91 0. 002-0. 087a
Wood 0.70 0.015
a
High values are recommended because they are based on
experimental data. However, low values have been found in the
literature. Therefore, a range is given.
Table 4 Radioactive source term equations for burning

contaminated combustibles
Combustible Contaminant form Equation
Cellulose Air-dried UNH M = lAOE-lxWxBxQ
UNH liquid M = 6.08E-7xWxBxQ
DUO powder, M = l.l$E-9xWxBxVxUxQ
flaming
Polychloroprene UNH liquid M = 0.385 x W x 5

DUO powder or M = 0.058xWxS
air-dried UNH
Polystyrene UNH liquid M, = 0 . 0 2 X W
PMMA Air-dried UNH M, = 0.007xW

UNH liquid Mt = 0.02x W
DUO powder M, = 0.05xW
30% TBP/kerosene Uranium M = l.3SxSxW

M = mass-release rate of radioactive particles, g/s
Mf = total mass release of radioactive particles, g
B = burn rate of combustible, g/s
Q = external heat flux to the combustible, kW
S = smoke-release rate, tls, divided by total grams of fuel
U = uranium concentration, g uranium/g combustible
V = air velocity, cm/s
W = mass of radioactive material, g
UNH = uranyl nitrate hexahydrate
DUO = depleted uranium dioxide
768 NUCLEAR SYSTEMS
2. Gasdynamics
The conservation of mass, energy, and momentum are used to describe
the gasdynamics within a network system. The equations are developed as
follows:
2 NkpkQk = 2 Nkmk = 0 (14)
k k
where mk and Qk are the mass and volumetric flow rates, respectively; p^
is the density; and Nk is used to adjust for the proper flow direction for
branch k. For example, Nk = +1 for the downstream node of a branch,
and Nk = -1 for the upstream node. Equation (14) can be generalized
to allow for mass accumulation at the node by the relationship
^ = - [2 Nkmk + Ms] (15)

d/ V \jk \
where Ms is an arbitrary mass source per unit time for the volume V of
the node.
The second relationship required in this analysis is the energy equation,
with its corresponding effect on both nodal temperatures and pressures.
The energy equation used in the analysis is
* - f [Z Wr. + £) + H.T, + f 1 (16)

at V |_ k V 2CP/ CpJ
where vk is the branch velocity, R the gas constant, y the ratio of specific
heats, Cp the specific heat at constant pressure, Tk the branch temperature,
Ts the source temperature, and Es the energy source. This equation is
developed in detail in Ref. 29.
In this analysis, we require a relationship between the pressure and mass
flow rates in addition to Eq. (14). Using Fig. 1, the momentum equation
for branch flow m is
Sj
^ = (P. - P,) - j^f (17)
where the branch inertia is
-f • f ¥•
+ +
"^ A 2^, <' 8 >
A •:~) •
A, |_ Aj
H—————H^————H-«—————H
i j
Fig. 1 Branch with sudden area change.
and
with
s = +1 if m >0 (20a
- -1 if m< 0 (20b
The branch-flow equation [Eq. (17)] is derived in more detail in Ref. 29.
The true upstream node density and the pressure differential are (Pl —
P 2 )> not (Pi ~ Pj)- Therefore, when the flow is positive from / to y",
P2 = P j (21)
However, P1 and P, are connected through the isentropic relation
PIIPI = rp = [Tyrjv^v (22
with
r,. + (v?/2C,) = r, + (v?/2Cp) (23)
For the density relationship, we have
PI/P, = rp = [TyrjV'v (24
Therefore, the continuity between / and 1,
m = pfVfAf = PiVlAl (25)
can be solved to obtain rp for a given m and node condition.
We can cast the flow equation into the following form by adding both
the perturbation pressure and density:
m = A - Cbp - £8p (26)
where A, C, and E are constants, and 8/? and 8p are temporary constants
associated with the density and pressure perturbation terms, respectively, as
Pi = P/ + 8p/ (27)
and
Pi - Pi + 8A- (28)
The values for A, C, and E can be found in Ref. 30. After much manip-
ulation, the perturbation pressure and density can be formulated as
anbpf + fl128p, - fcj (29)
and
a21bpf -f a22§pi = b2 (30)
770 NUCLEAR SYSTEMS
Matrix manipulation gives

8p/ = 0*22^1 - a12b2)/(ana22 - a]2a2l) (31)
and
8p/ = (anb2 - a2lbl)/(ana22 - a12a21) (32)
The values of the coefficients of the perturbation values for the pressure
and density can be found in Ref. 29. At this point, the procedure is to
iterate on Eqs. (31) and (32) for 8/?, and 8p, until these values approach
zero. When this occurs, the network is balanced for a distribution of pres-
sure, density, and temperature. The solution is unsteady and can proceed
until a steady state is reached.
3. Heat Transfer
Heat transfer from the duct walls must be calculated during the fire
transient, and the numerical model19 to accomplish this is discussed sub-
sequently.
Because the heat-transfer processes are nonlinear in temperature, solving
the equations requires that a set of nonlinear coupled differential and
algebraic equations be solved. The temperature (TR) discussed earlier in
the energy equation must be determined at each time step.
The energy balance across a duct section gives
Tout = Tm - (G,/mCp) (33)
where Qt is the net amount of energy transferred from the gas to the duct
wall. Heat transfer from the combustion gas and duct walls may be broken
into two components:
Qi = Qa + Qn (34
where Qci is the net amount of energy transferred from the gas to the duct
inside surface by forced convection heat transfer — i.e. ,hA(Tg — r,), where
A is the wall heat-transfer area, h the heat-transfer coefficient, Tg the bulk
gas temperature, and Tf the inside duct wall temperature — and Qri is the
net amount of energy transferred from the gas to the duct wall because of
radiation heat transfer — i.e., AEf(Ig — /J, where e, is the emissivity factor,
which is a function of gas composition and temperature, Ig is the intensity
factor, which is a function of the gas composition and temperature, and Is
is the intensity factor, which is a function of the gas composition and wall
temperature. A fit of empirical data for Ig and Is gives
= 190[(T + 460)/760]5 (35)
We now have
Qn = fW, Tf) (36)
and
Qt = f(Tg, Th rg, Tf) (37)
Because the total energy transfer is a function of the wall temperature as

well as the gas temperature, a model for duct wall temperature is needed.
Heat conduction through the wall is accomplished by solving a set of
coupled differential equations of the form
-^(pwCpwTN) = aNTN_l + bNTN + Q0 (38)
where p^ is the density of the wall, Cpw the constant-pressure specific heat,
and aN and bN the constants.
To solve the set of equations, the derivative terms are put into finite-
difference form, and a matrix solution is obtained for TN. To complete the
energy balance equations, we evaluate the amount of heat transfer from
the outside walls, Q0, with
Qo - Qco + Qro (39)
where Qco is the net amount of energy transferred from the duct wall
because of natural convection, i. e., HA (TN — T0), and Qro is the net amount
of energy transferred from the outside duct wall to the atmosphere from
radiation, i.e., &A( e T4N — aT4o), where T0is the outside air temperature.
The final set of equations is
G, = f ( T g , Th T*, 7?) (40)
QO = /en) (4
T, - f(Qn Qo, T,} (42)
TN = f(Qn Qo, Tf) (43)
The equations are nonlinear, and a direct solution is not possible. There-
fore, the equations are solved using an iterative method. Tilde (temporary)
quantities are defined as the best (latest) estimates of the exact solution
of the coupled equations. For the first iteration in a time step, these quan-
tities are estimated to be the solution for the equations at the previous
time step. The tilde quantities are calculated in the following order:
Q0 = f ( T g , TN) (44)
Q, = f(T8, T,) (45)
T, = f(Qa, Q,) (46)
TN = f(Q0, Qi) (47)
where the functional form is defined by the models discussed earlier. The
duct outlet temperature is then evaluated:
Toui = Tm - (QJmCp) (48)
772 NUCLEAR SYSTEMS
4. Material Transport
The purpose of the material transport algorithm is to provide an estimate
of the smoke, gas, and radioactive movement throughout the ventilation
network in the nuclear facility. Ultimately, the objective is to predict the
quantity released into the environment. The transport of material is by
convection with two forms of depletion taken into account: deposition by
gravitational settling and deposition on fibrous filters in the ventilation
duct work. Re-entrainment of aerosols is also taken into account.
The gasdynamics are uncoupled completely from the material flow. A
more complete discussion of the theory used here can be found in Ref .
31. The continuity equation for the mixture is
where up is the particulate velocity, pp the density of the particulate based

on the mixture volume, V the volume, s the boundary, and Mp the material
source. The mass conservation equation can be transformed into
w**' + M"~ x»v f ] (50)
where xp is the mass fraction of the particulate, p the density of the mixture,
AJ the flow area, and upi the flow velocity normal to the area.
Because ra, = p/Wj/4/, Eq. (50) becomes
-i[2
P L *
Mp - xpV (51)
Equation (51) is a differential equation for the unknown xp. When the
gasdynamics quantities (p, ra,) are known, the equation can be integrated
to obtain xp at a new time.
C. Applications to Heat-Transfer and Material-Transport Experiments

Verification of the FIRAC computer code has proceeded by testing
individual models. Here we describe verification tests for the duct heat-
transfer module and the material-transport module. 3233 Separate effects
tests to obtain empirical data for the blower, damper, and filters have been
performed and modeled. The heat-transfer and material-transport exper-
iments were carried out in a specially built apparatus located and operated
by New Mexico State University for the Los Alamos National Laboratory.
1. Experimental Apparatus
The multicompartment ventilation system apparatus is installed in a con-
crete building that provides environmental control. The experimental ap-
paratus is designed to accommodate thermal, pressure, and aerosol inputs.
The thermal inputs are generated by a commercially available duct heater
rated at 92,000 kcal/h fired by natural gas and limited to a maximum of
300°C. The duct heater is on casters, which allows thermal input from
different locations in the system. Pressure pulses are limited to 140-kPa

overpressure and must originate in the rectangular volume shown in Fig.
2. The pipe and square ducts are designed to have a mechanical safety
factor of 3. The particulate mass input is provided by a commercial dust
feeder having variable supply rates between 1 and 40 g/min (particulate
material density of 1 g/cm3). The maximum particulate mass concentration
is approximately 1.4 g/m3 for unit-density particulate material.
The experimental apparatus arrangement of ducts and volumes is shown
in Fig. 2. For economy, 30.5-cm-diam Schedule 20 pipe is used for the
bypass loop around the two volumes and for the connections between the
two volumes. The 0.6-m2 ducts were made from 0.64-cm steel plate and
were used for the remainder of the ventilation system. The system's straight
length is approximately 24.4 m. The inlet and outlet round-duct connections
are within the lower third of the rectangular tank and differ in height by
0.69 m vertically. Figure 2 shows one circular duct connecting at the top
of the cylindrical volume and the other connecting at the side of the volume.
The centerline distance of both the round and square ducts above the floor
is 1.1 m.
The two volumes are steel tanks modified as shown in Fig. 2. The rec-
tangular tank (17.0 m3 in volume) has 5-cm-thick walls and can withstand
the largest pressure pulses. The remainder of the system limits the maxi-
mum pressure pulse to about 140-kPa overpressure. The cylindrical tank
(22.6 m3 in volume and 4.6 m high) is upright to provide the maximum
possible stratification and serves as a location for thermal and/or particulate
mass input. The two fans in the facility provide positive or negative pres-
sure, relative to the ambient atmosphere, in the system.
The round-duct dampers are reversible 1.27-cm steel plates secured by
the pipe flanges. These dampers are either fully open or closed. In addition,
three dampers are located in the square duct: between the HEP A filter
gravimetric balance and fan, upstream from the square-duct tee, and down-
stream from the square-duct tee. Numerous model system configurations
can be obtained by opening and closing the dampers. The HEP A filter
gravimetric balance is designed especially to measure the collected mass
on an HEP A filter installed in the system. The balance uses a null technique
and an electronic force transducer to achieve a resolution of 2 g.
This experimental apparatus is three dimensional because of the vertical
height associated with the cylindrical volume. Thus, thermal loads or "test
fires" (possibly in conjunction with a particulate mass) can be input to the
base of the cylindrical volume, and the transport of both thermal energy
and particulate material can be observed in the system. In addition, the
gravimetric balance can determine the collected particulate mass on the
HEPA filter. Careful collection of particulate material on the internal
surfaces yields information on total deposition.
2. Duct Heat-Transfer Experiments

FIRAC's heat-transfer model predicts how the gas cools as it flows through
the ventilation system. The following heat-transfer processes are modeled
by solving a set of four coupled nonlinear algebraic equations: 1) forced
convection between the gas and the inside duct walls, 2) radiation between
r Gravimetric Balance
and
HEPA Filter Location
8.0 m
Cylindrical Volume - 22.7 m Rectangular Volume -17.0 m
0.3 m
Fig. 2 Full-scale model ventilation system used for mass-transport study.

J. R. TRAVIS ET AL 775
the gas and the side duct walls, 3) heat conduction through the duct wall,
and 4) natural convection and radiation from the outside duct wall to the
surrounding air.
Experiment description. The thermal tests reported here used only a
portion of the experimental apparatus shown in Fig. 2. The subsystem used
comprises the section of the 30.5-cm pipe between dampers 5 and 7 (in-
dicated in Fig. 2), the pipe tee joint adjacent to dampers 7 and 8, and the
entire section of the 0.6-m2 duct. The gravimetric balance and exhaust fan
are not included. The commercial duct heater, which includes the supply
blower, injects thermal energy into the system at damper 5. Dampers 4,
6, and 8 are closed for these tests.
The internal and external wall temperatures are measured at the lon-
gitudinal center of each segment of pipe and duct. The thermocouple
generally is placed on the vertical center of one wall. In one section of
duct, thermocouples are placed on each vertical wall, in the top, and in
the bottom of the duct to measure the thermal gradients in the duct cross
section. Pipe and duct wall temperature measurements are made at 15
locations. In addition to the wall temperature measurements, the gas flow
temperature is measured at the center of each segment of the 0.6-m2 duct.
Twenty-gage, type-J thermocouple wire was used for these measurements,
and the external and internal wall thermocouples were mounted by peen-
ing. The internal gas flow thermocouple assemblies were made of a 0.64-
cm-diam stainless-steel sheath with an aluminum heat and a type-J ther-
mocouple. All internal thermocouple wire insulation was replaced by a
fiberglass sleeve. The gas flow was obtained by a pitot tube in a five-section
traverse of equal concentric areas at the system outlet duct with the pressure
readings made with an inclined manometer.
The pressure is measured with a 0.0- to 62.3-Pa manometer at the center
of each segment of the 0.6-m2 duct and other selected locations, such as
on each side of the 90-deg elbow and on each side of the commercial
dampers. These locations were chosen to provide data spatially compatible
with data outputted by the FIR AC code.
The experimental sequence began with the duct heater blower running
until steady-state flow was reached. With the blower continuing to run, a
30-min heat pulse then was injected into the system at a predetermined
thermal rate. The blower then remained on, and the system returned to
ambient conditions. Gas flow and wall temperatures typically were re-
corded by the data acquisition system at 40-s intervals throughout the
experiment, and pressures and flow rates were measured manually during
each phase of the experiment. The steady-state volumetric flow rate meas-
ured before the heat pulse was 0.085 m3/s, and the pressure drop across
the system was about 38 Pa. At the end of the 30-min heat pulse, the flow
rate was 0.179 m3/s, with a 100-Pa pressure drop across the system. As the
system cooled, the measured flow rate decreased to 0.080 m3/s.
Results. The numerical simulation of the experiment using FIRAC be-
gan with the experimentally determined steady-state flow, temperature,
and pressure as initial conditions. A temperature-time function represent-
ing the measured temperature during and after the heat pulse was input
into the network model node 8, which corresponds to the temperature at
776 NUCLEAR SYSTEMS
the center of the first section of the 0.6-m2 duct. This is the section of duct
connected to the 30.5-cm-diam pipe.
A comparison of the experimental and FIRAC-generated temperatures
as a function of time and space showed generally good agreement. In Fig.
3, the gas flow temperature as a function of time is shown for nodes 11,
14, and 22. In Fig. 2, node 11 is in the center of the duct section downstream
of the 90-deg bend, node 14 is in the center of the duct section upstream
of commercial damper 9, and node 22 is in the center of the duct section
just upstream of the filter and gravimetric balance. The code predictions
for the temperatures at nodes 11 and 22 are within 5% of the experimental
data during the heat pulse and consequent rise in temperature. The node
11 error increases to nearly 12% as the temperature decreases. Because
the nodes with the highest and lowest temperatures agree relatively well
with the experimental values during the heat pulse, we were surprised that
node 14 showed a larger error—about 12% at the peak temperature. We
checked the nodes adjacent to node 14, and this indeed is the error pattern
(see Fig. 4). Here the temperature is plotted as a function of duct length
at three selected times: at the end of the heat pulse, 150 s after the heat
pulse, and 830 s after the heat pulse. The largest errors (up to 12%) are
between 9 and 18 m, which correspond to the locations of nodes 11-16.
FIRAC Prediction
• — • Node 11
Node 14
......................... Node22
Experimental
Nodes 11, 14,22
40
20
0 1 1 1 1
1000 2000 3000 4000 5000 6000
Time (s)
Fig. 3 Measured and calculated gas temperatures at nodes 11, 14, and 22.
240
FIRAC Prediction
220
— • — • t = 1840 sec.
200
t = 2670 sec.
o
Experimental
0)160
All Cases
Q
^ 140 ~
CD
V-
2 ,20
k—
CD
Q- 100
CD
H- 80
60 -
40 -
20 I
10 15 20 25 30 35
Length (m)
Fig. 4 Measured and calculated gas temperatures along duct length at 1840, 1990,
and 2670 s.
The FIRAC code, duct heat-transfer model simulates the amount of

thermal energy the gas loses as it passes through a duct section. This energy
loss is a function of radiation and convective processes on the inside and
outside of the duct wall and heat conduction through the wall. The com-
puted temperature loss is directly proportional to the energy loss and in-
versely proportional to the mass flow rate. Heat loss is not modeled at
branches other than ducts. The model for heat conduction through the
duct wall is based on standard models, such as those given in Ref. 34.
Temperatures at several nodal points through the wall can be calculated
with this model. Because the wall is only 0.64 cm thick, the measured
temperature differential across the wall is always less than 1°C. Therefore,
we modeled the wall with only one nodal point; thus, the calculated tem-
perature is always the average duct wall temperature.
The four wall temperatures at node 14 were averaged to compare them
with the computed average. The experimental and computed temperatures
as a function of time are shown in Fig. 5. The slope of the computed curve
is less than that of the experimental curve both during and after the heat
pulse. A maximum error of about 10% is seen at late times.
3. Material-Transport Experiments
The material-transport model in FIRAC is very elementary. The only
depletion mechanism acting on the aerosol as it moves through the duct-
work is gravitational settling and collection by HEPA filters.
778 NUCLEAR SYSTEMS
FIRAC Generated
Experimental
20
1000 2000 3000 4000 5000 6000
Time (s)
Fig. 5 Average measured and calculated wall temperature for node 14.
Experiment description. The particulate transport tests described here

used only a portion of the experimental apparatus shown in Fig. 2. The
subsystem used consists of the 30.5-cm pipe between dampers 5 and 7 and
the entire 0.6 x 0.6-m duct from damper 7 through the HEPA filter
gravimetric balance section. The outlet blower was not used, and all flow
and energy input originated at the gas-fired furnace located at damper 5.
As shown, dampers 6, 8, and 4 were closed, and, in addition, a temporary
wall was placed in the large tee to eliminate a recirculating cavity at this
location. Temperature measurements were made at the centerline of the
vertical wall of the duct at the center of each 2.44-m section of duct using
type-J thermocouples peened into the wall. The temperature of the flowing
gas was measured at the duct centerline at the center of each 2.44-m section
of duct using type-J shielded thermocouples. Temperatures were recorded
every 5-min during each experiment using a Hewlett-Packard data acqui-
sition system. Flow rate into the Hauck gas-fired burner, which supplied
both the heat energy and the airflow through the system, was measured
using a standard orifice in a 6.1-m-long, 0.203-m-diam inlet pipe to the
flower of the Hauck burner. The signal from the pressure transducer meas-
uring the pressure difference across the orifice was recorded by the data
acquisition system to provide a continuous flow rate measurement. Energy
input into the system was calculated from the lower heating value of natural
gas supplied by the gas company, the measured flow rate of the natural
gas entering the burner, and an assumed 95% burner efficiency.
The experiments began by adjusting the Hauck burner to deliver the

desired flow rate and heat input into the duct system. After 1 or 2 h of
operation (depending on the flow rate), the system reached steady state,
and particulate was injected into the hot-gas flow in the middle of the first
duct section downstream of the corner in the square duct. The total length
of 0.6 x 0.6-m ducting between the point of particle injection to the HEPA
filter at the duct outlet was 21.93 m. The particulate consisted of small
glass beads, nominally spherical, with a mean diameter of the volume
distribution of 11.86 jam and a calculated surface area of 0.9762 m2 /cm3.
The measured mean density of the glass beads was 2325.9 kg/m3.
An NBS-II dust generator was used to inject glass beads into the system.
A 1.6-cm-diam copper tube consisting of four 0.64-cm copper tubes con-
nects the dust generator to a manifold in the duct. These four outlet tubes
are aligned with the flow direction of the hot gas within the duct. For each
experiment, approximately 200 g of beads were placed in the hopper of
the dust generator. After the system had achieved steady-state operation
at the desired flow rate and heat input, the dust generator was started,
and the glass beads were injected into the gas flow. The time needed to
empty the hopper of beads was measured with a stopwatch. A small quan-
tity of beads always fell to the floor of the dust generator during its op-
eration. These beads were weighed after each experiment, and their weight
was subtracted from the original amount placed in the hopper. An average
injection rate then was calculated.
The distribution of the particulate deposition along the duct length was
determined by very careful sweeping of all four walls of each 2.44-m section
of the 0.6 x 0.6-m duct after each experiment. The weight of the glass
beads in each section of ducting was measured using an Ohaus triple beam
balance. Hence, the total amount of particulate was checked against the
known amount injected. The average concentration of the particulate in
grams per cubic meter was calculated by dividing the total weight in each
duct section by the surface area of each duct-section floor. This assumption
was made because photomicrographs of particulate deposited at the center
of the vertical walls indicated that this material amounted to less than 10%
of the material deposited on the duct floor. The concentration of particulate
on the floor calculated in this manner was assumed to be at the longitudinal
center of the duct section.
Results. Table 5 presents the parameters of the six experiments run to
verify the FIR AC treatment of material transport. Three different levels
of energy input were used: 1) no heat input, 2) low heat input sufficient
to bring the gas temperature of node KK to 150°C, and 3) high heat input
sufficient to bring the gas temperature of node KK to 300°C. However,
because the temperature of the gas flow at the particle injector was used
in the computer calculations, these values are given in Table 5, as well as
the temperatures at node KK. Burner energy input also is given.
Figure 6 shows the locations along the 0.6 x 0.6-m duct, beginning at
the particle injector at which the measurements in grams per square meter
of material deposited were made experimentally and the locations at which
FIR AC predicted the amount of material deposited (indicated by "NC"
for numerical calculation). Figures 7-12 compare the deposition concen-
tration of the glass particulate in grams per square meter predicted by
r Gravimetric Balance
and
HEPA Filter Location
NC NC EM NC
CO
o
8.0 m
Cylindrical Volume • 22.7 m Rectangular Volume -17.0 m
0.3 m
t— - — - — a—E
Fig. 6 Location of experimental measurements (EM) and numerical calculations (NC) of mass deposition.
Table 5 Material-transport experiments
Injection Injection Unacct.

Test T-KK, T-Injec., Heat input, Flow rate, rate, time, mass,
no. °C °C kJ/h ft-Vmin g/min mm %
1 21 21.0 _ 251 9.77 20.32 12.4
2 21 21.0 — 501 4.20 42.92 4.4
3 150 76.7 61,706 246 10.41 18.74 17.5
4 170 107.7 151,891 506 13.98 13.66 9.5
5 300 134.9 174,042 383 5.49 36.64 11.2
6 325 218.3 320,659 499 14.00 13.64 4.9
FIRAC with the deposition concentration actually measured in the exper-

iments. Notice that, for all six runs, the experimental data indicated a
higher initial deposition rate than did the code. That is, in the first 5 m
downstream of the injection point, the experimental deposition rate was
significantly higher than the predicted value (probably because of a higher-
than-expected agglomeration of the particles leaving the injector, which
resulted in a rapid gravitational settling). Photomicrographs of the depos-
ited particles at various locations along the length of the duct verify that
significant agglomeration does exist in the first 5 m downstream of the
injection point, whereas little agglomeration is found beyond this point to
the end of the duct. Several things could be responsible for the higher
percentage of agglomerated particles leaving the injector; for example, 1)
the particles in the hopper of the dust generator may have been highly
agglomerated, and the velocity of the flow through the injector manifold
may not have been large enough to produce the shear forces necessary to
break the clumps of particulate apart before entering the 0.6 x 0.6-m duct,
and 2) the velocity in the injector manifold may have been so slow that
the number density of the particles became a controlling factor (coagulation
is proportional to the number squared), and agglomeration actually oc-
curred within the injector. We assumed that the particulate injected was
monodispersed in the numerical calculations. Therefore, we cannot expect
the numerical results to match closely the experimental results in the first
5 m of the simulation.
Beyond 5 m from the injector to the end of the duct, the experimental
and numerical values of deposition concentration match closely and follow
the same trends. Less particulate is deposited on the duct floor and more
is trapped in the HEP A filter as flow rate is doubled, no matter what the
rate of heat input. This suggests that the settling velocity of the particles
changes little and the particulate in the higher flow rate stream is simply
swept further down the duct before striking the floor. At the same flow
rate, increasing the rate of heat input also appears to decrease the amount
of particulate deposition and increase the amount of particulate trapped
on the HEPA filter. This is consistent with Stokes settling in which the
settling velocity is inversely proportional to the viscosity. An increase in
temperature will increase the viscosity of the gas and, thus, decrease the
Floor Concentration (g/m ) Floor Concentration (g/m )
—• —• ro
o o O
I I I
3
TO" 3
CTQ
00 *
w
5 W
«§ oo" o X
3 5> g
3 3 c/)
2 CD °°
o
o CD
3 o
O
3
CD
CD
O x
m ~n H:
^
§: 3 -g 55 5 •CD >5
CD > ^ =- O
S e ^ O g"
s CO 00 ^
O5 CD
O O5 ^
CQ (Q ^-
^ 5-
> =»
H Tl
o
FIRAC Prediction PH250.2
Experimental (MAT08) Results
Filter Mass Gain

" FIRAC 92.2 g
Exper. 49.0 g
05 Flow Rate: 246 CFM

k_
•*-* Heat Input: 61706 KJ/hr
c:
CD
O
C
o
O
\_
o
o
0 2 4 6
Distance from Injector (m)

Fig. 9 Experiment 3, particle deposition.
12
FIRAC Prediction PH500.1
11
Experimental (MAT09) Results
10
C\J Filter Mass Gain
9
FIRAC 120.3 g
8 Exper. 101.4 g
c
o
7
OJ Flow Rate: 506 CFM
6 Heat Input: 151891 KJ/hr
0)
o
c 5
o
o 4
1_
o
o 3
0
1^ I I I
10 12 14 16 18 20 22
Distance from Injector (m)

Fig. 10 Experiment 4, particle deposition.
783
Floor Concentration (g/rrf Floor Concentration (g/m )
ro O 00 o ro o 00
i i i i i i i i i i I I I I I I I
3
TO* 3
CTQ*
W
t§ O
? 0)"
<
5* ^'
II -H 3 $u
3
O 1
a °
Cg o\ CD
y- ®
*» 3
^ 3
rt> 13 ro
S"^
m -n & §
X =n
"° 2? O
? o I TJ
§? § i
CQ (Q 5"
H
_L
TJ
-£
3? °
$ KD
c_
C/)
settling velocity, which, in turn, causes the particulate to travel further

down the duct before striking the floor.
The amount of particulate trapped by the HEPA filter at the outlet of
the 0.6 x 0.6-m duct was smaller for the experiment than it was for the
numerical prediction for all six cases studied. This also appears to be a
result of the unexpected agglomeration of the particle in the experiments.
Two nominal flow rates were used, 250 and 500 ftVmin, but the actual
flow rates attained are somewhat different.
D. Summary
The models and methods developed previously describe an approach
that can be used to analyze the effects of a fire in a nonreactor system.
The energy from the combustion process and the material generated, in-
cluding both nonradioactive and radioactive material, are calculated by a
zone-type combustion model. This model serves as a source term for dis-
tribution of energy and material throughout the nonreactor system net-
work, which is usually the facility's ventilation system. Solutions of the
gasdynamic equations give the pressures, densities, mass flow rates, and
temperatures. Models for the heat transfer at network duct walls are in-
cluded. The distribution of combustion products, radioactive materials,
and temperatures throughout the system is obtained to complete the so-
lution procedure.
IV. Nuclear Reactor Containment Systems

A. Introduction
Combustion in light-water reactor containments can cause high pressures
or temperatures that can, in turn, cause damage to the containment or
affect important safety-related equipment. After the Three-Mile Island
accident (a severe, or degraded-core, accident), it was found that significant
quantities of hydrogen had been generated from the chemical reaction
between the zirconium cladding (the thin protective covering of the nuclear
fuel) and the water vapor. When released into the containment, this hy-
drogen burned by one or more combustion modes and thereby threatened
the containment integrity, internal structures, and safety-related equip-
ment. Other combustibles, such as electrical cable insulation and hydro-
carbon lubricating oils for cooling electric pump bearings, also have come
under examination as possible fuels. In fact, the 1975 Brown's Ferry nuclear
plant fire, in which the electric-cable insulation ignited from a candle used
to detect leaks, caused direct losses of $10 million and indirect losses of
$30 million related to business interruption.
Modeling the geometries of containment buildings is a challenge. Figure
13 shows a cross section of the Heiss Dampf Reactor (HDR) containment
near Frankfurt, West Germany. This building is 60 m high and 20 m in
diameter. It contains two stairwells, an elevator shaft, several vertical open
hatchways, and about 70 rooms. This particular containment has roughly
11,300 m3 of free volume, or approximately one-sixth the free volume of
a typical U.S. pressurized water reactor containment. Experiments simu-
786 NUCLEAR SYSTEMS
Fig. 13 HDR containment building cross section.
lating cable-tray, hydrocarbon lubricating oils, and severe accident com-

bustion phenomena currently are being conducted.
The USNRC has supported research at Los Alamos and Sandia National
Laboratories to develop combustion models to evaluate threats to the
reactor containment and safety-related equipment. Current levels of re-
search will coordinate model validation with ongoing experiments at the
HDR facility. We describe the Los Alamos field model approach in the
following sections.
B. Mathematical and Physical Models

The partial differential equations that govern the fluid dynamics, species
transport, and turbulence model, and the equations modeling the com-
bustion, heat transfer, and condensation processes, are presented in this
section.
L Mixture Equations
The mixture mass conservation equation is
d
f( + V-(pu) = 5 (52)
where p is the mixture density or the sum of the macroscopic densities for
each individual species, p a, u the mass-averaged velocity vector, and S the
mass loss (condensation) per unit volume and time.
The mixture-momentum conservation equations are given by
+ V - (puu) = -Vp + V - T + p s - D (53)

dt
where p is the pressure, T the viscous stress tensor, g the gravitational

vector, and D the internal structure drag vector.
The viscous force V • f is the usual Newtonian one, where, for example,
the radial component is given by
i A ( r r ) + ! £ * _ Zee + ?!« (54)
rr) {
r dr <• r 30 r dz '
where
(55)
....
(56)
and
[dv dul
l +
[dr dz\
The coefficient of viscosity, JJL, which appears in the viscous stress tensor,
is interpreted as an "apparent" or "turbulent" viscosity. Here we have
used the second viscosity coefficient X = — 2fx/3, which is equivalent to
assuming the bulk viscosity to be zero. The calculation of the apparent
viscosity through the turbulence model will be discussed subsequently.
788 NUCLEAR SYSTEMS
The structural drag vector represents the resistance of internal structures

such as pipes, I-beams, catwalks, and such configurations that are impos-
sible to resolve. These internal structures play an important role as heat
sinks and to a lesser degree as momentum sinks. However, we do attempt
to model their momentum effects by
D = CDp(A/Vc)u\u\ (59)
where (A/VC) represents frontal structural area divided by the fluid volume
for every computational cell containing internal structures. Note that A is
dependent on the orientation of the structures. For example, there is little
resistance to flows parallel to the gratings of the catwalks but quite a
different resistance to flows normal to the gratings. The drag coefficient
CD is usually defined to have values between zero and unity. The mixture
internal energy density equation is
V-q+Q (60)
where / is the mixture-specific internal energy and Q the energy source or

sink per unit volume and time due to combustion, condensation, and energy
exchange with internal structures, floors, ceilings, and walls. Because con-
densation effects can dominate the work term (the first term on the right-
hand side), we must account for the remaining gas in a computational cell
expanding into the volume change associated with the condensate. We
account for this effect by making use of the ideal gas equation of state to
arrive at
m. _ RH*>T s
v at vcP ''•-
• ~
where /?H->o is the gas constant for steam, T the gas mixture temperature,
and S/ ra, the sum of all mass per unit time of steam condensing on all
surfaces internal to or bounding the computational cell Vc. Note that S is
the same mass loss per unit volume and time as Eq. (52). The energy flux
vector q is given by
q = $VT + PT S Wi ~ } (62)
a \ P/
where 4> is the apparent or turbulent conductivity, 7 the apparent or tur-

bulent mass diffusivity, and ha the enthalpy for species a.
2. Species-Transport Equations
The transport or mass equation of the individual species is given by
+ V • (PttM) - V • | P7V( ^ ) I = 5a (63)

where pa is the mass per unit volume (macroscopic density) of species a.

The source or sink term, Sa, represents the species mass created or de-
stroyed by chemical reactions and the loss of steam as a result of conden-
sation. When Eq. (63) is summed over all species, the result is the mixture
mass [Eq. (52)].
3. Constitutive Relationships
The specific internal energy of species a is related to the temperature
by
/« = (4). + £ (C,)a dr (64)
For the species in which we are mostly interested (O2, N 2 , H 2 , and H 2 O),
we can approximate the specific heats by
(CJa = Aa + (BJVT) 4 (CJT) 4 (DJT2) (65)
After integrating Eq. (64), the specific internal energy as a function of
temperature is approximated with the quadratic function
/„ = Ea 4 FaT 4 G^T2 (66)
over the temperature range 200-2500 K. The total specific internal energy
then is given by
/ = 2 xja (67)
where xa is the mass fraction for species a.

For consistency with Eq. (66), we approximate each species' specific
heat by differentiating Eq. (66) with respect to temperature to yield
(Cv)a = Fa 4 2GaT (68)
instead of applying Eq. (65) directly.
The equation of state for the fluid pressure p is given by the ideal gas
mixture equation:
P = T 2 * ap a (69)
a
where Ra is the gas constant for species a. The constant-pressure specific

heats are calculated by
(Cp)a - Ra + (Cr)a (70)
4. Heat-Transfer and Condensation Relationships

The heat exchange between the gas mixture and the internal structure
Qs is given by
Qs = qJVc
790 NUCLEAR SYSTEMS
where
qs = hsAs(Ts - T) (72)
In these expressions, Vc is the computational cell volume, 7^ the structure
surface temperature, hs the heat-transfer coefficient between the gas mix-
ture and the internal structures, and As the area of exposed structure in a
computation cell.
The heat flux between a fluid computational cell adjacent to a solid
boundary, such as a wall, ceiling, or floor, is given by
Qw = qJVc (73
where
qw = hn,Aw(Tw - T) (74)
The thermal boundary layer is taken into account by using a modified
Reynolds analogy formulation, 35 which is simplified to obtain .the heat-
transfer coefficient:
hw = (Tjuc)Cp (75)
The rates of heat transfer and condensation increase when the mass fraction
of steam becomes a relatively large fraction of the mass of the gas mixture.
As the mass-transfer rate increases, the thermal and concentration bound-
ary layers become thinner because of the suction effect of the condensation
process. This reduction in the boundary layer further increases the tem-
perature and concentration gradients near the boundary and, consequently,
increases the heat- and mass-transfer coefficients. Bird et al.36 developed
correction factors based on film theory that can be used to determine the
increase in the heat- and mass-transfer coefficients. The corrected heat-
transfer coefficient then becomes
ht = &Thw (76)
where
&T = $Tl(<*r - i) (77)
and the rate factor <f> r is given by
d > r = -[mw(Cp)LJhwAn, (78)
where mw is the wall condensation rate and (Cp)L, the specific heat of the
water vapor at constant pressure. Note that, in the presence of condensing
water vapor, 4>r is negative, which increases the correction factor &T and
the heat-transfer coefficient h*. The internal structure heat-transfer coef-
ficient is computed in an analogous fashion.
The heat-transfer coefficient expression [Eq. (75)] contains the com-
putational cell-centered average velocity uc and the wall shear stress TU ,,
which is related to the fluid density and the wall shear speed ul by
TH, = pw2 (79)
We are unable to resolve turbulent boundary layers near solid walls with
any practical computing mesh, so we match our solution near solid bound-
aries or internal structures with the turbulent law of the wall37:
uju, = A /;,(yu*/v) + B (80)
This expression requires an iterative solution for u*. We find that it is more
convenient and almost as accurate to use an approximation obtained by
replacing u* in the argument of the logarithm in Eq. (80) by the one-
seventh-power law.38 The one-seventh-power law may be rearranged to
give
yu.lv = O.l5(yuc/v)7/* (81)
which, when substituted into Eq. (80), yields
uc/u* = 2.19 /,,(yuc/v) + 0.76 (82)
when A = 2 . 5 and B = 5.5. It is now straightforward to find the shear
speed u*, where y is the distance from the wall to the cell-centered average
speed uc9 and v is the gas mixture molecular kinematic viscosity.
The local Reynolds number yujv may be small, indicating that the cell
center lies in the laminar sublayer and the law-of-the-wall formulation is
not valid. In this case, Eq. (82) is replaced by the corresponding laminar
formula:
uju* = (yuc/v)m (83)
The transition between Eqs. (82) and (83) is made at the value where they
predict the same w*, which is yujv = 130.7. Therefore, u* is calculated by
Eq. (82) when yujv > 130.7 and by Eq. (83) when yujv < 130.7. In the
laminar case, the wall heat-transfer coefficient [Eq. (75)] reduces to hw =
pv/y, which results in a simple difference approximation to the laminar
heat flux for a molecular Prandtl number of unity when substituted into
Eq. (74).
Condensation can occur on any structural surface (walls, ceiling, floors,
and internal structural) provided the surface temperature is less than the
saturation temperature of the water vapor next to the surface. The amount
of energy resulting from condensation delivered to the wall surface area
Aw is
qc = "U(/H2o)* - (/H2o)J (84)
where (/H2o)*> *s tne specific internal energy of the water vapor in the
computational cell adjacent to the wall with volume Vc, and (/H2o)s *s tne
specific internal energy of the liquid water film that is on the surface.
(Note: We have taken this to be equal to the surface temperature of the
wall.)
Actually, the term in brackets in Eq. (84) states that the energy per unit
mass deposited on the surface caused by condensation is
(Iôh ~ (Igs)s + hfg + (Ifg)s - (In2o)s (85)
792 NUCLEAR SYSTEMS
where (Igs)s is the specific internal energy at the saturation temperature of

the water vapor, hfg the latent heat of condensation, and (Ifg)s the specific
internal energy at the saturation temperature for the liquid film.
The condensation rate is described as
mw = h%As[(pH20)b - (pH2o)s] (86)
where hd is the corrected mass-transfer coefficient, (pH o)& the water vapor
density in the bulk, and (pH2oX the water vapor density on the condensing
surface.
The mass-transfer coefficient hd then can be expressed in terms of the
heat-transfer coefficient h39 as
hd = h/(PCp) (87)
Following similar ideas as with the heat-transfer coefficient for relatively
large steam mass fractions, we correct the mass-transfer coefficient by
h*d = ®mhd - (88)
where
©„, - [/„(/? + I)]//? (89)
In this expression, the flux ratio R is given by
R = (nw - nh)ll - nw (90)
where nw is the steam mole fraction at the wall and nb the steam mole
fraction in the bulk.
We model the radiation heat transfer in a relatively simple fashion. We
assume that 15% of the total chemical energy of combustion is radiated
from a point source at the computational cell center.40 This energy is
radiated spherically away from each computational cell where combustion
occurs to solid surfaces such as floors, ceilings, and walls with the appro-
priate geometric view factors. In some cases, such as the hydrogen diffusion
flames in the Mark III containments, 41 the wet-well walls become quite
hot, and energy is reradiated from each wall segment by
qwr = AW<T(*WT*W - T^ (91)
where <r is the Stefan-Boltzmann constant, tw the wall emissivity, and rref
the reference temperature.
For every computational cell side interfacing with a wall, ceiling, or floor,
the one-dimensional transient heat-conduction equation
^
dt
= Pdx0
2 (92)
with the wall boundary condition
qw + qc + 2 qr - qwr = - kAw — (93)

oX
is solved from the temperature distribution T and the wall surface tem-
perature Tw. On the left-hand side of Eq. (93), the four terms represent
energy delivered to a wall section by convection, condensation, radiation
from all computational cells with combustion occurring in line-of-sight
contact, and reradiation of energy from hot surfaces, respectively. To sim-
plify the reradiated energy analysis, we have assumed this energy to be
deposited in the surrounding gas.
5. Turbulence Modeling
We used three turbulence models during these investigations: algebraic,
subgrid scale, and K-E.
Algebraic model. For the algebraic turbulence model, we adapted the
approach of Launder and Spalding,42 and write the turbulent viscosity as
^ = £>K1/2^ (94)
where C^ is a constant (typically 0.05), p is the fluid density, K is the

turbulent kinetic energy, and € is interpreted as a length scale of the energy-
carrying eddies. It is often estimated that 10% or less of the mean flow
energy is contained in the turbulent kinetic energy, so
K1/2 = [O.l(l/2)u2]m (95)
For containment studies, the length scale usually is set equal to 0.25-
0.5 m.
Subgrid scale model. This model was formulated in terms of the tur-
bulent kinetic energy per unit mass, K, which represents turbulent length
scales too small to resolve in the computational mesh, and the local fluid
density, p. The transport equation for the product PK is given by
- (PK) + V • (PKM)
dt
= - p K V - u + T : Vu + V - (^VK) - pK3/2/€ (96)
where the second term on the left-hand side represents the convection of
turbulence by the established velocity field, and the terms on the right-
hand side represent the effects of turbulence generation by compression,
production of turbulence by viscous dissipation, diffusion of turbulence,
and decay of turbulent energy into thermal energy, respectively. The latter
term appears with opposite sign as a source term in the thermal internal
energy density equation [Eq. (60)]. After dividing up the density, the
turbulent kinetic energy K is substituted into Eq. (94) to compute the
turbulent viscosity. It is interesting to note that, in the quasisteady solution
where the production term equals the decay terms because of viscous
dissipation, this generalized model reduces to the original algebraic subgrid
scale model of the type used by Smagorinsky43 and Deardorff. 44 45
The K-e model. This turbulence model is the two-equation model de-
veloped by Launder and Spalding35 with an extension to treat buoyancy
794 NUCLEAR SYSTEMS
effects. The transport equation for the product PK is given by
+ V
dt
= VI — VK I + T : Vu + jjuxg • VT - pe -h K.S (97)
and the transport equation for the product of the density and the dissipation
of the turbulent kinetic energy e is
dt
+ C> fag • Vrj - C2p - + eS (98)

K. K.
The turbulent viscosity is calculated using the Kolmogorov hypothesis:

M, = (pC^2)/e (99)
Buoyancy effects are treated by the third term on the right-hand sides of
Eqs. (97) and (98), where a is the coefficient of thermal expansion. For
this analysis, we assume that the turbulent Prandtl and Schmidt numbers
equal unity. Therefore, the turbulent conductivity 4> is
4> = ^Cp (100)
and the turbulent mass diffusivity 7 is
Both of these transport coefficients are expressed in terms of the turbulent

viscosity and thermodynamic properties of the gases.
6. Chemical Kinetics
For diffusion flames involving hydrogen 22 or hydrocarbon fuels, 24 we
have used a simple global chemical kinetics model that grossly oversim-
plifies the actual chemical processes. In this model, the only reaction mod-
eled is
fl(fuel) + ft(0 2 ) -^ c(H20) + d(CO2) + e(C) (102)

here, for example, if the fuel is H2, 0 = 2, fe = l , c = 2, d = 0, and e
— 0. For hydrocarbon fuels, we have balanced Eq. (102) with a = 1.0, b
= 0.7328, c = 0.4828, d = 0.4914, and e = 0.258. In modeling nuclear
reactor containment buildings, typical computational cell volumes are
1-2 m3, and even larger in some cases. We try to keep cell volumes in
regions where diffusion flames are expected to about 1 m3. For this spatial
resolution, there is no attempt to describe flame structure; we simply rep-
resent combustion energy release in a complex geometric containment.

Furthermore, chemical reaction time scales generally are short compared
with fluid motions in these combustion modes, so the many elementary
reaction steps and intermediate chemical species can be neglected in this
first approximation.
The reaction rate o> in Eq. (102) is modeled by a modified Arrhenius
law. The general expression for the reaction rate is
rate (mole/m3-s) - C^r/rj^fuelôxidizer]0 cxp(-E/RT) (103)
where E is an activation energy; t, /, and o are exponents; "fuel" and
"oxidizer" indicate local concentrations in molecules per cubic meter; and
T0 is the adiabatic flame temperature. We usually set the exponent t = 0,
but, in some hydrocarbon simulations, we set t = 1/2 in an attempt to
model molecular motion.
Now, the finite-rate chemical kinetics can be written for the fuel,
(104)
and the oxidizer,
— = -bM0u (105
dt
where Mf and M0 are the molecular weights of fuel and oxygen, respec-
tively. The chemical energy of combustion is computed as a source for the
energy equation [Eq. (60)] by
Qc = Cccb (106)
5 5
where Cc = 4.778 x 10 J/mole for hydrogen fuel and 3.05 x 10 J/mole
for hydrocarbon fuel. The typical molecular weight for hydrocarbon fuel
is 7.17.
In practice, when solving the finite-rate chemical equations [Eqs. (104)
and (105)], we integrate the fuel [Eq. (104)] when the fuel-oxidizer mixture
is fuel lean and the oxidizer [Eq. (105)] when the fuel-oxidizer mixture is
fuel rich. By using Eq. (102), all components of the combustion process
are determined. We have compared results of this model with the exper-
imental data for the one-fourth-scale test facility 46 and the HDR contain-
ment building,47 and have found surprisingly good agreement for the general
circulation patterns in complex geometries, concentrations of combustion
products, and temperature distributions throughout the containment
buildings.
C. Computational Model
An abbreviated description of the computational model is given here;
refer to Ref. 48 for details. The solution algorithm follows the Los Alamos
ICEd-ALE49~53 methodology for solving multidimensional, time-dependent
fluid flow equations. For example, a transient fluid dynamics time step is
broken into three distinct phases, which are discussed next.
796 NUCLEAR SYSTEMS
1. Explicit Lagrangian Phase

In this phase, the densities, velocities, and specific internal energy fields
are updated by the effects of all chemical and physical processes. This
includes combustion, condensation, heat transfer, body forces, and tur-
bulence effects. The change in each computational cell volume Vc is cal-
culated by
(Vc ~ V"C)IM = V?V - u" (107)
The mass change for each species because of combustion, condensation,
and turbulent diffusion is calculated by
VZS£ (108)
When comparing Eqs. (63) and (108), we see that
5" = V - [p<YV(p a /p)] + Stt (109)
Note that this algorithm differs from most ICEd-ALE methods in that
velocities are centered on the computational volume, Vc, faces rather than
positioned at vertices. Because we are interested only in the Eulerian
solution of the flow equations, a full continuous rezone always will be
applied (see Sec. IV. C. 3), and this Lagrangian phase is only an interme-
diate step toward the full solution. A more efficient solution procedure
can be realized by using computational-volume-faced velocities. We define
a staggered volume Vs, which is equal to half of each Vc volume having a
common face. The components of the velocity field then can be found
from
-V?Vp" + V?M" (110)
where, by comparing with Eq. (53),
M" - V • T + pg - D (111)
We now can find the change in the specific internal energy:
n n
oIV
p c - po I"Vc
"
where, upon comparing with Eq. (60),
Qn = V • q + Q (113)
remembering Q contains energy of combustion, condensation, and heat-
transfer effects. Finally, after inverting Eq. (66) for the temperature T,
the updated pressure is determined by
P = rS*«P«
a
(114
2. Implicit Pressure Iteration Phase
In this phase, an implicit evaluation of the time-advanced densities,
velocities, pressure, and specific internal energy fields is achieved. The

purpose of this phase is to allow calculations of low-speed (low-Mach-
number) flows without any time-step restrictions from the fluid sound
speeds. In this phase, the computational cell volume is written as
K?V • (u* - u") (115)
the species mass equation is written as
(pf K* - p,Vg/Ar = 0 (116)
the change in velocities is written as
u* - H = -[M(p*V?)]V2V (p* - p») (117)
and the specific internal energy is given by
(p*/*V* - p/tg/Af = -V?p"V • (u* - u") (118)
Using the equation of state [Eq. (69)] and
(119)
we rewrite the left-hand side of Eq. (118) as

c
_ ^) „„ Vcp] (120)
A? *âR. a
A/ £„a Xa
where xa is the mass fraction of species a. Combining Eqs. (115), (118),

and (120) and rewriting, we obtain
(p* -p) = - LL 2,« *«*«// CV(T) J\P" + pJHV: - v^/v: (121)

This equation can be linearized by rearranging (V* - VC)/V* and then
applying a binomial series to obtain
(V; - VC)/V* - (I/* - VC)/VC (122)
provided (V* — VC)IV* « 1. Note that this quantity is monitored during

a transient simulation to ensure an accurate solution. As the time step
decreases, this quantity approaches zero.
Combining Eqs. (115), (117), (121), and (122) yields the pressure equa-
tion in the form
L
?V • (u" -~u) + (p - p") (123)
798 NUCLEAR SYSTEMS
where
{ ) = f I E ^Ka/cxr)IP" + P \ (124
IL « / J
and
8p = p* - pn (125)
Equation (123) is second order and linear in §p. To solve this Poisson-
type pressure equation, an efficient matrix solution algorithm called the
conjugate gradient method54 is used.
This implicit solution of the pressure equation allows for greater effi-
ciency than a purely explicit calculation with reduced time steps. The nu-
merical stability achieved permits pressure waves to traverse more than
one computational cell in a time step. In practice, after solving Eq. (123)
for 8/?, we evaluate u* from Eq. (117), V* from Eq. (115), p* from Eq.
(116), p* from Eq. (121), T* from Eq. (69), and /* from Eqs'. (66) and
(67).
3. Rezone Phase
Phase 3 explicitly performs all of the advective flux calculations. This
phase completes a Eulerian calculation and, therefore, completes a time
step. The equations for mass, momentum, and energy advection, respec-
tively, are as follows:
V?V • (pfZT) = 0 (126)
V?V • ( P *w*«*) = 0 (127)
and
fon + ljn + lyn _ p */*^*)/Ar] + K?V ' (p*/*M*) - 0 (128)
By dividing the solution algorithm into these three phases, one can use
whatever advection or rezoning algorithm they wish. We have made use
of simple donor cell, interpolated donor cell, van Leer,55 and flux-corrected
transport.56"58 The time cycle is completed by computing Tl + l from Eq.
(66) and/?" + 1 from Eq. (69).
There is an additional calculation that must be made, i.e., the explicit
evaluation of the turbulence model. This will elevate the turbulence trans-
port coefficients to the (n + 1) level in preparation for phase 1 of the next
time step. The new time step is controlled by checking the entire com-
putational mesh for a material velocity Courant condition and a diffusion
stability limit condition.
D. Application to Full-Scale Fire Experiments
We have used this time-dependent, three-dimensional model to simulate
the large oil-pool fire tests in the HDR containment 24 shown in Fig. 13.
For this simulation, important insight into verifying the general Navier-
Stokes solution algorithm and the submodels for condensation heat trans-
fer, turbulence, and chemical kinetics can be gained.
1. Experimental Facility and Description

This containment building is 60 m high and 20 m in diameter with a free
volume of 11,300 m3 in about 70 interconnected passageways and com-
partments. Total interior steel surface area is in excess of 30,000 m2. The
location of the oil-pool fires is shown in Fig. 14 as compartment 1.906
between the 25-m and 30.85-m levels on the left-hand side of the contain-
ment. This compartment is also shown in Fig. 13 in the upper left-hand
side just below the dome volume, and it is displayed schematically in Fig.
15. Fires of 2000, 3000, and 4000 kW were burned over 30- to 35-min
periods. We have calculated test T52.14, which is the 3000-kW fire ex-
periment. The fire compartment and adjoining passageways and rooms are
detailed in Figs. 16 and 17, where all dimensions are given in millimeters.
By continuously weighing the oil pan, the rate of combustion of the hy-
drocarbon fuel can be determined, and this burning rate is presented in
Fig. 18. Integration of this burning rate curve indicates that roughly 120
kg of fuel were consumed during the 35-min experiment.
Experimental measurements are extensive; continuous monitoring of
CO2, O2, and H2O concentrations, velocities, and temperatures inside,
entering, and exiting the burn compartment as well as throughout the
containment itself are recorded during the entire experimental sequence.
A horizontal raster or grid consisting of a 5 x 5 lattice of measuring devices
is located at 25.5 m (the fresh-air intake to the mounting hatch, designated
raster III, see Fig. 15), 31 m (mounting hatch passageway entering the
containment dome, designated raster II), and 38 m (the dome region,
mounted on the overhead crane, designated raster I). In addition, several
video cameras were focused on the combustion-room doorway, mounting
hatch compartment, and entrance to the dome region. More details of the
experiment and the data acquisition systems can be found in Ref. 47.
In general, the experimental results show that combustion occurs in the
fire compartment, forming a hot layer consisting of combustion products,
nitrogen, and excess oxygen (the burning occurs in an oxygen-rich envi-
ronment). There is an exception to this behavior early in the experiment
(3-12 min), when as a result of the feedback of combustion products into
the burn room, the combustion process is fuel rich and flashover occurs
with the flame spreading into the adjoining mounting hatch compartment.
This phenomenon is shown clearly in Fig. 19, where time histories of the
principal combustion product CO2 and oxygen are plotted for the upper
middle doorway position between the combustion compartment and the
mounting hatch (28.25 m). Notice the decrease in the concentration of
CO2 during this time while the flame has extended outside the combustion
room to burn mainly in the upper regions of the mounting hatch com-
partment. This very interesting effect was observed by video cameras lo-
cated in the mounting hatch compartment. The major flow paths or circulation
patterns are very complex, but, in effect, two large vortices are set up
during the burning event. Hot combustion products and gases exit the
combustion room, forming a hot plume that rises through the mounting
hatch passageway into the dome region. The return flow is vertically down
the staircase and equipment hatch region shown on the right-hand side of
Figs. 13 and 15 to about the midplane of the containment vessel before
800 NUCLEAR SYSTEMS
270° j 90°
Fig. 14 HDR containment with room numbers and elevation levels.
moving horizontally (compartment 1611 in Fig. 15) toward the spiral stair-
case to complete the top, or clockwise, convective loop. The other major
convective loop is counterclockwise; it is located in the lower half of the
building. This lower convective loop has a much lower velocity than the
upper hot-plume-driven loop. Refer to Ref. 59 for a more detailed dis-
cussion concerning this experiment.
Concrete Shell
Steel Containm.
Fire Fighting Hoses

Cable Trays
Measure Grids
Fig. 15 Schematic view of HDR containment showing relative location of fire room
and measurement rasters or grids.
2. Experimental and Calculated Results

Since the mass-burning rate is known from Fig. 18, it has been used as
a mass and energy source for the fuel species and energy conservation
equations [Eqs. (63) and (60)], respectively. This mass and energy fuel
source is distributed uniformly in the computational cells overlaying the
physical location of the oil pan. In Figs. 20 and 21, we present the gas
temperature time histories from locations at the top of the combustion
room and near the top of the doorway between the combustion room and
the mounting hatch, respectively. There is excellent agreement between
experiment and calculated values, which indicates that the overall energy
released in the combustion process has been modeled adequately. Com-
parisons between gas concentrations in Fig. 22 at the upper middle position
in the burn-room door indicate good agreement. The feedback and flash-
Fire
Compartment
Fig. 16 Horizontal layout of the geometry for the burn compartment, oil pan, and
connecting rooms. (All measurements are in millimeters.)
Smoke Exit
to Reactor
Dome 18
g^
Mounting Hatch Fire

Compartment
- ~1780 -
x
Oil-pan and
Mass Balance
Fresh Air Inlet Section A-A

Fig. 17 Vertical layout (Sec. A-A from Fig. 16) of the geometry for the burn
compartment, oil pan, and connecting rooms. (All measurements are in millimeters.)
802
(/)
75)
03
oj
DC
0)
13
CD
10.00 20.00 30.00 40.0
Time ( m i n u t e s )
Fig. 18 Measured fuel consumption for HDR test T52.14.
C
CD
O
CD
CL
CD
E
5 -~ —
10.00 20.00 30.00 40.00
Time (minutes)
Fig. 19 Measured CO2 and O2 concentrations in the upper burn
compartment door.
804 NUCLEAR SYSTEMS
1600
1200 --
o
CD
800 --
CD
Q.
E
CD
400 —— -
I
.00 10.00 20.00 30.00 40.00
Time (minutes)
Fig. 20 Measured and calculated temperatures in the upper region of the burn
compartment.
1600
1200 - - -
O
CD
i—
ID
800 -- -
CD
Q.
E
CD
400
I
.00 10.00 20.00 30.00 40.00
Time (minutes)
Fig. 21 Measured and calculated temperatures in the upper burn
compartment door.
c
CD
O
CD
CL
CD
E
_E 10---
o
>
5 ~~~ —
10.00 20.00 30.00 40.00
Time (minutes)
Fig. 22 Measured and calculated CO2 and O2 concentrations in the upper burn
compartment door.
over trends are captured with some degree of accuracy; however, the exact
production of CO2 is off by as much as 50% during the flashover event
(4-12 min) and roughly 10-15% during the more or less steady combustion
process (12-35 min). Concentration measurements are a very sensitive
standard*for assessing the combustion model, and, in this case, the simple
one-step chemical kinetics described by Eq. (102), or at least the coefficient
d in this equation, could be modified.
Temperatures from the three raster or grid locations are displayed in
Figs. 23-25. At the fresh-air inlet to the mounting hatch compartment,
raster III at 25.5 m, Fig. 23 gives an indication of the strong temperature
distribution in this region, especially during the early part (first 10 min) of
the experiment. Because of the lack of spatial resolution, the calculation
is not able to resolve this distribution, but predicts more of an average
during the early time. The peak temperature is delayed about 2.5 min, and
after the peak, the temperature history is higher than the average of the
five measured values. In general, it is greater than the maximum of the
five observed values. Since the gas inflow temperature is a very strong
function of the loop velocity and the heat exchange with all structural
components on the return loop from the dome volume, it is likely that the
heat transfer from the gas to the structural components is somewhat in
error. However, considering the spatial resolution and the complexity of
the containment flow paths, we consider this comparison to be acceptable
and in fair to good agreement with the experimental data. For raster II at
806 NUCLEAR SYSTEMS
120 -- — -+ — — — —
CD
i_
13 80 -- —
2
CD
Q.
E
CD
40 -- —
10.00 20.00 30.00 40.00
Time (minutes)
Fig. 23 Measured and calculated temperatures at raster III (25.5-m level).
31 m, where the fire plume enters the dome, the measured temperatures
have been averaged and plotted in Fig. 24 along with the calculated values.
We see in both observed and computed histories that the maximum tem-
perature of about 250°C occurs with flashover as the flame extends high
into the mounting hatch compartment. At 12 min, the combustion event
occurs largely within the burn room and the observed temperature drops
at raster II to 160°C with a slow increase to 225°C at the time the experiment
ends. Here, the calculated temperature increases to about the right peak
value at the right time, but it does not dip to the correct value at 12 min.
From then until the end of the experiment, it remains nearly constant at
215-220°C, with a slight increase to 225°C just before the end of the
experiment. Measured and computed temperatures for raster I (located at
the 38-m level) are plotted in Fig. 25. As seen by the observed results,
there is a large temperature gradient across the hot-gas plume. The two
calculated temperature histories are in excellent agreement up to 9 min,
reaching the maximum values at the correct time. A local minimum occurs
at 11-12 min and then tempratures recover near their maximum values by
the end of the experiment. The computed temperatures, however, continue
declining after the peak values. This is attributed to the fact that the
algebraic turbulence model used in this calculation promotes far too much
mixing and cold gases are entrained into the plume during this period, thus
spreading the temperature profile beyond the observed plume width. In
addition, the donor-cell advection scheme implemented for this simulation
is known to be diffusive, and, therefore, numerical diffusion contributes
to this effect.
350
300 - - -
0)
Q.
E
50 -
0
-4.00 .00 4.00 8.00 12.00 16.00 20.00 24.00 28.00 32.00 36.00 40.00
Time (minutes)
Fig. 24 Measured and calculated temperatures at raster II (31-m level).
245-
I
210 1
I I I
Calculation
175-h - -\- - -Jf- - -h •
0 9 18 27 36
Time (minutes)
Fig. 25 Measured and calculated temperatures at raster I (38-m level).
808 NUCLEAR SYSTEMS
Overall, even with the simplified models and rather coarse resolution,
the calculated results are in fairly good agreement with the observed data.
When fluid dynamics effects dominate the analysis with hot plumes and
complex convective circulation patterns, one can often be successful cal-
culating diffusion-type flames with simple combustion models that have
been calibrated to release the correct amount of heat due to the combustion
process. We have had similar success calculating standing hydrogen dif-
fusion flames in MARK III boiling water reactor containments. 41
E. Summary
Using the preceding field equation model coupled with finite-rate global
chemical kinetics, we successfully analyzed hydrogen and hydrocarbon dif-
fusion flames occurring in a nuclear reactor containment under accident
conditions. These combustion modes are the easiest to model and analyze
when compared with other modes of combustion, such as propagating
flames in premixed fuel/oxidizer volumes. Deflagrations, flame accelera-
tion and transition from deflagration to detonation, and detonations are
all important combustion modes that have not been modeled successfully
in complex reactor containment structures. In the next section, we will
address some of these issues and recommend approaches for solutions.
V. Research Directions
A. Nonreactor Combustion Modeling
The basic framework necessary for modeling combustion phenomena
associated with nonreactor systems has been established. However, re-
search in several areas is needed, including radioactive and nonradioactive
combustion characteristics, material transport model improvement, lumped-
parameter and field model coupling, development of system safeguard
models, and experimental validation and verification of system (network)
modeling codes.
Very little is known about how radioactive material burns, particularly
how it may combine with nonradioactive burning material. Currently, highly
empirical data are being used to determine the airborne fraction of radio-
active release from contaminated material. Research needs to be performed
to develop better models using a larger experimental database.
The transport of smoke with its radioactive component is perhaps the
single most important aspect of nuclear materials fires. Models that ac-
curately simulate the behavior of particulate and gaseous material as it
moves through the nuclear facility are needed. In particular, interparticle
dynamics along with particle depletion models need to be developed and
refined.
For large volumes, such as large rooms, in nonreactor systems, the lumped-
parameter approach is not suitable. In this case, a multidimensional model
is needed and can be coupled with the lumped-parameter ventilation net-
work. This capability currently does not exist, and research needs to be
devoted to establishing this capability.
Research directions toward developing system safeguard models also
need to be pursued. These models would simulate the effect of sprinklers,
halon discharge, and demisters. With these models, the adequacy of the
fire protection system can be evaluated.
Finally, the lumped-parameter codes and models need to be validated
through experiments. Large systems of interconnected ductwork, rooms,
filters, dampers, and fans need to be included in the experimental system.
This experimental system, with proper instrumentation for measuring tem-
peratures, pressures, flows, and aerosol concentrations, would be invalu-
able for obtaining experimental data.
B. Reactor Combustion Modeling

The field equation approach has been applied successfully to diffusion
flame combustion modes in complex reactor containment geometries. To
evaluate other major concerns involving combustion phenomena during a
severe accident in a reactor containment, we must develop modeling ca-
pabilities to investigate flame propagation in premixed, complex, multi-
dimensional volumes. This includes laminar and turbulent deflagrations,
flame acceleration and transition from deflagration to detonation, and
detonations. We also wish to model radioactive and nonradioactive aerosol
dynamics and the interaction of the nonoxidized and oxidized aerosols with
the propagating flame. Accident mitigation concepts such as water sprays
and their influence on a steam-air-hydrogen environment must be assessed.
All of these phenomena must be evaluated in terms of thermal and me-
chanical loads on the containment and on safety-related equipment. Some
of these research areas include adaptive gridding, detailed chemical ki-
netics, turbulence, radiation heat transfer, aerosol dynamics, and model
validation with experimental data.
For propagating flames through complex geometries, spatial resolution,
and accuracy of the flame and near the flame is important. This will require
dynamic implementation of adaptive gridding algorithms60 to resolve suf-
ficiently the flame as it propagates through the containment.
Global finite-rate chemical kinetics, which were used for diffusion flame
analysis, will not be adequate for propagating flames. For the wide range
of conditions (pressure, temperature, and gas composition) likely to be
found in reactor containments, it will be necessary to couple the fluid
dynamics with a detailed chemical kinetics reaction set, as suggested by
Oran et al.61 The disadvantage of directly coupling chemical kinetics with
48 reactions and 9 species is that, for large, time-dependent, three-
dimensional problems, the solution algorithm is currently computationally
prohibitive. An alternative approach62-63 could be to solve the detailed
chemical kinetics for the expected range of pressures, temperatures, and
gas composition. From these conditions and solutions, we can determine
an induction time, a reaction time, and the amount of chemical energy
released. These parameters could be approximated with analytic functions
or tabulated in a parameter space table. A modified combustion parameter,
P, transport equation can be written as
+ u • VP = V(0VP) + * (129
810 NUCLEAR SYSTEMS
where T/ and © represent the induction time and turbulent transport coef-
ficient, respectively. Initially, P = 0, so that when P > 1, the induction
time has elapsed and the chemical energy of combustion is released over
a period of time equal to the reaction time, T^. In this way, an approximate
method for coupling the fluid dynamics and detailed chemical kinetics can
be achieved. The disadvantage of this procedure is that fluid dynamics
effects such as acoustic waves and turbulence intensities actually would not
influence the individual chemical reaction rates and, therefore, the reaction
time. It may be possible to add an additional variable to the parameter
space table to account for turbulence and the influence of turbulence on
the reaction rates, induction time, and reaction time. If the reaction rates
could be correlated to the intensity or kinetic energy of the local turbulent
conditions, then the first term on the right-hand side of Eq. (129), the
turbulent diffusion of P, could be eliminated from the equation.
For diffusion flame modeling, we have found that turbulent buoyant
plumes generally are represented adequately with the two-equation K-e
model. However, Zyvoloski64 has shown that the most accurate plume
dynamic predictions are provided with the three-equation K-e-T' 2 model.
For this model, the additional transport equation for the average temper-
ature fluctuations squared is solved. Reynolds stresses and turbulent energy
fluxes then are calculated and used directly in the Reynolds-averaged Na-
vier-Stokes equations. At this point, it is not clear whether this three-
equation model is needed to describe diffusion flame plume dynamics. If
the survival and proper functioning of safety-related equipment located
near diffusion flames remain unresolved safety issues, then it may be nec-
essary to use a three-equation model. In addition, if the diffusion flame
plume impinges on the steel shell liner, a higher temperature can be pre-
dicted than what actually is observed. Although this result is conservative,
there is actually more entrainment of cooler material into the plume than
predicted. It may be necessary for "best-estimate" calculations of thermal
loads on the containment shell to use the three-equation model.
Radiation heat transfer from diffusion and propagating flames can be of
major importance. For example, in optically thick regions near hydrocar-
bon diffusion flames, the coupling between carbon or soot, the fluid field,
and the radiation field will be strong, whereas in optically thin regions far
from the flame, this coupling will be weak. To include these effects, we
consider adopting an extension of a nonequilibrium radiation model orig-
inally developed by Alme et al.65 and then extended by Daly.66 In this
model, the local energy densities of the radiation field, the fluid, and the
particles can be different. Because the speed of light is inherent in these
equations, any practical solutions must be obtained from an implicit finite-
difference form of the equations.
Water sprays and aerosol dynamics are similar problems. For example,
in a steam-air-hydrogen environment, the spray water droplets and aerosol
particles provide condensation nuclei to reduce the steam concentration
in the mixture. Two undesired physical effects occur: 1) hydrogen concen-
trations increase as the steam condenses, and 2) turbulence levels increase
with condensation and droplet and particle momentum exchange with the
continuous gas phase. Nonoxidized aerosols in the presence of hydrogen
J. R. TRAVIS ETAL 811
combustion could increase the energy release as aerosols contribute to the

combustion process. Conversely, water sprays such as mists can be huge
energy sinks and provide mitigation methods for suppressing or controlling
potentially damaging combustion modes. These tradeoffs must be evalu-
ated and assessed to provide input for severe accident management policies.
These modeling questions and the coupled phenomena are difficult, and
the physical geometry to which they must be applied is so complex that
model and code verification is essential. This can be accomplished only
with modelers and experimenters working very closely in a collaborative
effort to resolve and understand the relevant phenomena. Separate-effect
experiments must be performed to validate individual physical and chemical
processes wherever possible. Integrated tests at several scales will provide
data and the confidence for computer models to be used to extrapolate to
full-scale geometries. Only in this way can numerical modeling and sim-
ulations address the issues and answer the questions concerning the prob-
lems involved in nuclear reactor safety.
Acknowledgment
The authors thank the U.S. Department of Energy and the U.S. Nuclear
Regulatory Commission for their continued support during this work.
References
J
Pryor, A. J., "The Browns' Ferry Nuclear Plant Fire," Society of Fire Protection
Engineers, Quincy, MA, TR-77-2, 1977.
2
Henrie, J. O., and Postma, A. K., "Analysis of the Three Mile Island (TT1I-2)
Hydrogen Burn," Proceedings of Second International Topical Meeting on Nuclear
Reactor Thermal Hydraulics, ANS, Hinsdale, IL, Jan. 1983, p. 1157.
3
Patterson, D. E., "The Rocky Flats Fire," Fire Journal, Vol. 64, 1970, p. 5.
4
Dungan, K. N., and Lorenz, M. S., "Nuclear Power Plant Fire Loss Data,"
Electric Power Research Institute, Palo Alto, CA, CR EPRI NP-3179, July 1983.
5
Church, J. P., "Safety Analysis of Savannah River Production Reactor Oper-
ation," Savannah River Lab., Aiken, SC, Rept. DPSTSA-100-1, Sept. 1983.
6
SFPE Handbook of Fire Protection Association, Society of Fire Protection En-
gineers, Quincy, MA, Sept. 1988.
7
Tewarson, A., "Generation of Heat and Chemical Compounds in Fire," SFPE
Handbook of Fire Protection, Society of Fire Protection Engineers, Quincy, MA,
Sept. 1988, p. 179.
8
Jones, W. W., "A Multicomponent Model for the Spread of Fire, Smoke, and
Toxic Gases," Fire Safety Journal, Vol. 9, 1985, p. 55.
9
Babrauskas, V., "Burning Rates," SFPE Handbook of Fire Protection, Society
of Fire Protection Engineers, Quincy, MA, Sept. 1988, p. 2-1.
10
Steciak, J., Tewarson, A., and Newman, J. S., "Fire Properties of Combustible
Materials Commonly Found in Nuclear Fuel Cycle Facilities," Factory Material
Research Corp., Norwood, MA, Rept. FMRC J.I.OG3R8.RC, Feb. 1983.
n
Alvares, N. J., and Krause, F. R., "Experimental Simulation of Forced Ven-
tilation Fires," Los Alamos National Lab., Rept. LA-UR-84-1691, 1984.
12
Alvares, N., Foote, K., and Pagni, P. O., "Forced Ventilation Enclosure Fires,"
812 NUCLEAR SYSTEMS
13
Krause, F. R., and Gregory, W. S., "A Thermodynamic Fire Zone Model for
Tall Containment Buildings," 1987 Status Report on the HDR Safety Program,
Kernforschungzentrum, Karlsruhe, West Germany, Dec. 1987.
14
Hamins, A., and Seshardi, K., "The Structure of Diffusion Flames Burning
Pure, Binary, and Ternary Solutions of Methanol, Heptane and Toluene," Com-
bustion and Flame, Vol. 68, 1987, p. 295.
15
Masri, A. R., Bilger, R. W., and Dibble, R. W., "Conditional Probability
Density Functions Measured in Turbulent Non-Premixed Flames of Methane Near
Extinction," Combustion and Flame, Vol. 74, 1988, p. 267.
16
Krause, F. R., and Schmidt, W. H., "Bum Mode Analysis of Horizontal Cable
Tray Fires," Sandia National Labs., Livermore, CA, Rept. NUREG/CR-2431,
Feb. 1982.
17
Gido, R. G., "COMPARE-H21 Implicit Code for Containment Hydrogen-
Burn Analysis," Los Alamos National Lab., Los Alamos, NM, FIN A7288-Test 1
letter report, Feb. 1987.
18
Bergeron, K. D., Clavser, M. D., Harrison, B. D., Murata, K. K., Rexroth,
P. E., Schelling, F. D., Sciacca, F. W., Senglaub, M. E., Shire, P. R., Trebilcock,
W., and Williams, D. C., "Users Manual for CONTAIN 1.0, A Computer Code
for Severe Nuclear Reactor Accident Containment Analysis," Sandia National
Labs., Livermore, CA, Rept. NUREG/CR-4085, May 1985.
19
Nichols, B. D., and Gregory, W. S., "FIRAC Users Manual: A Computer
Code to Simulate Fire Accidents in Nuclear Facilities," Los Alamos National Lab.,
Los Alamos, NM, Rept. NUREG/CR-4561, April 1986.
20
Wheeler, C. L., Thurgood, M. D., Guidotti, T. E., and Bellis, D. E., "COBRA-
NC: A Thermal-Hydraulic Code for Transient Analysis of Nuclear Reactor Com-
ponents," Pacific Northwest Lab., Richland, WA, Rept. NUREG/CR-3262, Aug.
1986.
21
Jahn, H. J., "Status Report on the Hydrogen Distribution Following a Coolant
Losi Accident," Gesellschaft fur Reactor Sicherheit, Munich, Germany, Rept.
GSR-A-333, Aug. 1979.
22
Travis, J. R., "A Heat, Mass, and Momentum Transport Model for Hydrogen
Diffusion Flames in Nuclear Reactor Containments," Nuclear Engineering and
Design, Vol. 101, 1987, p. 149.
23
Cox, G., "A Field Model of Fire and Its Application to Nuclear Containment,"
Proceedings of CSNI Specialist Meeting on Interaction or Fire and Explosion with
Ventilation Systems in Nuclear Facilities, Los Alamos National Lab., Los Alamos,
NM, Rept. LA-9911-C, Vol. I, April 1983, pp. 199-209.
24
Travis, J. R., and Krause, F. R., "An Analysis of HDR Test T52.14: Diffusion
Flames in a Full Scale Nuclear Reactor Containment," Los Alamos National Lab.,
Los Alamos, NM, report in preparation.
25
Ayer, J. E., Clark, A. T., Loysen, P., Ballinger, M. Y., Mishima, J., Ow-
czarski, P. C., Gregory, W. S., and Nichols, B. D., "Nuclear Fuel Cycle Facility
Accident Analysis Handbook," U.S. Nuclear Regulatory Commission, Rockville,
MD, Rept. NUREG/CR-1320, 1988.
26
Ballinger, M. Y., and Owczarski, P. C., "Fire Source Term Modeling Using
the FIRIN Code," Proceedings of the 20th DOEINRC Nuclear Air Cleaning Con-
ference, 1988; also Nuclear Regulatory Commission, Washington, DC, NUREG
CP-0098.
27
Chan, M. K., Ballinger, M. Y., and Owczarski, P. C., "User's Manual for
FIRIN: A Computer Code to Estimate Accidental Fire and Radioactive Source
Term Releases in Nuclear Fuel Cycle Facilities," Pacific Northwest Lab., Richland
WA, Rept. PNL-4532, NUREG/CR-3037, 1988.
28
Halverson, M. A., Ballinger, M. V., and Dennis, G. W., "Combustion Aero-
sols Formed During Burning of Radioactively Contaminated Materials—Experi-
J. R. TRAVIS ETAL 813
ment Results," U.S. Nuclear Regulatory Commission, Rockville, MD, Rept.

NUREG/CR-4736, 1987.
29
Tang, P. K., Andrae, R. W., Bolstad, J. W., Duerre, K. H., and Gregory,
W. S., "Analysis of Ventilation Systems Subjected to Explosive Transients," Los
Alamos National Lab., Los Alamos, NM, Rept. LA-9094-MS, 1981.
30
Gregory, W. S., Nichols, B. D., and Idzorek, R., "Node-Based Network Anal-
ysis and Its Application to Mine Ventilation," Proceedings of Third Mine Ventilation
Symposium, Pennsylvania State Univ., University Park, PA, 1987.
31
Tang, P. K., "Material Convection Model," Los Alamos National Lab., Los
Alamos, NM, Rept. LA-9393-MS, 1982.
32
Gregory, W. S., Nichols, B. D., Fenton, D. L., and Smith, P. R., "Fire-
Accident Analysis Code (FIRAC) Verification," Proceedings of Nineteenth DOE/
NRC Nuclear Air Cleaning Conference, U.S. Dept. of Energy, Washington, DC,
Rept. CONF.-860820, NUREG/CP-0086, Vol. 1, May 1987.
33
Gregory, W. S., Nichols, B. D., White, B. W., Smith, P. R., Leslie, I. H.,
and Corkran, J. R., "Fire Aerosol Experiments and Comparisons with Computer
Code Predictions," Proceedings of Twentieth DOE/NRC Nuclear Air Cleaning
Conference, U.S. Dept. of Energy, Washington, DC, Rept. CONF.-880822, NUREG/
CP-0098, Vol. 2, May 1989.
34
Patankar, S. V., Numerical Heat Transfer and Fluid Flow, Hemisphere, New
York, 1980.
35
Launder, B. E., and Spalding, D. B., "The Numerical Computation of Tur-
bulent Flows," Computational Methods in Applied Mechanical Engineering, Vol.
3, 1974, p. 269.
36
Bird, R. B., Stewart, W. E., and Lightfoot, E. N., Transport Phenomena,
Wiley, New York, 1960.
37
Schlichting, H., Boundary-Layer Theory, McGraw-Hill, New York, 1968.
38
Brodkey, R. S., The Phenomena of Fluid Motions, Addison-Wesley, Reading
MA, 1969.
39
Rohsenow, W. M., and Choi, H., Heat, Mass, and Momentum Transfer, Pren-
tice-Hall, Englewood Cliffs, NJ, 1961.
40
Zalosh, B., "Energetics of Hydrogen Combustion," personal communication,
1984.
41
Travis, J. R., "Hydrogen Diffusion Flames in a MARK III Containment,"
Joint ANS/ASME Conference on Design, Construction, and Operation of Nuclear
Power Plants, ASME BOOK G00253, ASME, New York, 1984.
42
Launder, B. E., and Spalding, D. B., Mathematical Models of Turbulence,
43
Smagorinsky, J., "General Circulation Experiments with the Primitive Equa-
tions," Monthly Weather Review, Vol. 91, 1963, p. 99.
44
Deardorff, J. W., "A Numerical Study of Three-Dimensional Turbulent Chan-
nel Flow at Large Reynolds Numbers," Journal of Fluid Mechanics, Vol. 41, 1970,
p. 453.
45
Deardorff, J. W., "On the Magnitude of the Subgrid Scale Eddy Coefficient,"
46
Tamanini, F., Ural, E. A., and Chaffee, J. L., "Hydrogen Combustion Ex-
periments in a 1/4-scale Model of a MARK III Nuclear Reactor Containment,"
Factory Mutual Corp., Norwood, MA, Interim Rept. Y101-1, May 1987.
47
Valencia, L., "Brandschutzversuche im Containment HDR-Versuchsgruppe
BRA Versuche T52.1/t52.2," Kernforschungszentrum Karlsruhe PHDR Karlsruhe,
Germany, Rept. 5.075/86, Feb. 1987.
48
Travis, J. R., and Dukowicz, J. K., "A Linearized ALE Method for the Solution
of the Navier-Stokes Equations," Los Alamos National Lab., Los Alamos, NM,
report in preparation.
814 NUCLEAR SYSTEMS
49
Amsden, A. A., and Hirt, C. W., "YAQUI: An Arbitrary Lagrangian-Eulerian
Computer Program for Fluid Flow at All Speeds," Los Alamos Scientific Lab.,
Los Alamos, NM, Rept. LA-5100, March 1973.
50
Ramshaw, J. D., and Dukowicz, J. K., "APACHE: A Generalized-Mesh Eu-
lerian Computer Code for Multicomponent Chemically Reactive Fluid Flow," Los
Alamos Scientific Lab., Los Alamos, NM, Rept. LA-7427, Jan. 1979.
51
Amsden, A. A., Ruppel, H. M., and Hirt, C. W., "SALE: A Simplified ALE
Computer Program for Fluid Flow at All Speeds," Los Alamos Scientific Lab.,
Rept. LA-8095, June 1980.
52
Cloutman, L. D., Dukowicz, J. K., Ramshaw, J. D., and Amsden, A. A.,
"CONCHAS-SPRAY: A Computer Code for Reactive Flows with Fuel Sprays,"
Los Alamos National Lab., Los Alamos, NM, Rept. LA-9294-MS, May 1982.
53
Amsden, A. A., Ramshaw, J. D., O'Rourke, P. J., and Dukowicz, J. K.,
"KIVA: A Computer Program for Two- and Three-Dimensional Fluid Flows with
Chemical Reactions and Fuel Sprays," Los Alamos National Lab., Los Alamos,
NM, Rept. LA-10245-MS, Feb. 1985.
54
Chandra, R., "Conjugate Gradient Methods for Partial Differential Equa-
tions," Ph.D. Thesis, Yale Univ., New Haven, CT, 1978.
55
Van Leer, B., "Towards the Ultimate Conservation Difference Scheme V. A
Second-Order Sequel to Godunov's Method," Journal of Computational Physics,
Vol. 32, 1979, p. 101.
56
Boris, J. P., and Book, D. L., "Solution of the Continuity Equation by the
Method of Flux-Corrected Transport," Methods in Computational Physics, Vol.
16, 1976, p. 85.
57
Zalesak, S. T., "Fully Multidimensional Flux-Corrected Transport Algorithms
for Fluids," Journal of Computational Physics, Vol. 31, 1979, p. 335.
58
Baer, M. R., and Gross, R. J., "A Two-Dimensional Flux Corrected Transport
Solver for Convectively Dominated Flows," Sandia National Labs., Albuquerque,
NM, Rept. SAND 85-0163, 1986.
59
Muller, K., "Stand der Branduntersuchungen im HDR Containment Versuchs-
gruppe T52.1," Kernforschungszentrum Karlsruhe PHDR-Arbeitsbericht 5.139/87,
Feb. 1988.
60
New York, 1987.
61
Oran, E. S., Young, T. R., Boris, J. P., and Cohen, A., "Weak and Strong
Ignition. I. Numerical Simulations of Shock Tube Experiments," Combustion and
Flame, Vol. 48, 1982, p. 135.
62
Oran, E. S., and Boris, J. P., "Weak and Strong Ignition. II. Sensitivity of the
Hydrogen-Oxygen System," Combustion and Flame, Vol. 48, 1982, p. 149.
63
Oran, E. S., Boris, J. P., Young, T. R., Flanigan, M., Picone, M., and Burks,
T., "Simulations of Gas Phase Detonations: Introduction of an Induction Parameter
Model," Naval Research Lab., Memo. Rept. 4255, June 1980.
^Zyvoloski, G., "Simulation of Intense Fires: A Comparison of Turbulence
Models," Los Alamos National Lab., Los Alamos, NM, report in preparation.
65
Alme, M. L., Westmoreland, C., and Fry, M. A., "Non-Equilibrium Radiation
for the HULL Code," Air Force Weapons Lab., Albuquerque, NM, Rept. AFWL
TR-76-244, 1976.
66
Daly, B. J., "Modifications of the CONCHAS-SPRAY Code for Entrained
Flow Gasification Studies, Final Report 5/1/84-12/31/85," Los Alamos National
Lab., Los Alamos, NM, Rept., LA-10754-MS, June 1986.
Author Index
Baer, M. R. ...................... 481 Maclnnes, J. M. ................ 615

Barr, P. K. ....................... 673 McMurtry, P. A. ............... 257
Bellan, J. ......................... 547 Mitler, H. E. .................... 711
Boris, J. P. .......................421 Nunziato, J. W. ................. 481
Bracco, F. V. .................... 615 Odiot, S. ..........................513
Brown, N. J. .......................37 Oran, E. S. ...................... 421
Crowe, C. T. .................... 543 Page, M. ............................. 3
DiBlasi, C. ...................... 643 Peters, N. ......................... 155
Drummond, J. P. ............... 365 Peyrard, M. ...................... 513
Dwyer, H. A. ................... 673 Pitz, W. J. ..........................57
Embid, P. F. .....................481 Pope, S. B. .......................349
Frenklach, M. ................... 129 Radhakrishnan, K. ...............83
Ghoniem, A. F. ................ 305 Rangel, R. H. ................... 585
Givi, P. ............................257 Sichel, M. ........................ 447
Gregory, W. S. ................. 755 Sirignano, W. A. ............... 585
Kailasanath, K. ............225,421 Smooke, M. D. ................. 183
Krause, F. R. .................... 755 Travis, J. R. ..................... 755
Lengsfield, B. H., Ill ............ 3 Westbrook, C. K. ................57
815
This portion of the volume does not have page
numbers. The page number range listed in the
file name was used so the file will display at the
correct location in a directory listing.

Numerical Approaches To Combustion Modelling

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Numerical Approaches To Combustion Modelling

Загружено:

Авторское право:

Numerical

A. Richard Seebass, Editor-in-Chief

Published by the American Institute of Aeronautics and Astronautics, Inc.

American Institute of Aeronautics and Astronautics, Inc.

John R. Casani John E. Keigler

Alien E. Fuhs Daniel P. Raymer

Part 1. Chemistry in Combustion Modeling

Chapter 1. Ab Initio Quantum Chemistry for Combustion . . . . . . . . . . . 3

Chapter 2. Rate Coefficient Calculations for Combustion

Chapter 3. Numerical Modeling of Combustion of Complex

Chapter 4. Combustion Kinetics and Sensitivity Analysis

Chapter 5. Reduction of Chemical Reaction Models . . . . . . . . . . . . . . . 1 2 9

Part 2. Flames and Flame Structure

Chapter 6. Length Scales in Laminar and Turbulent Flames . . . . . . . 1 5 5

Chapter 7. Numerical Modeling of Laminar Diffusion Flames . . . . . . 1 8 3

Chapter 8. Laminar Flames in Premixed G a s e s . . . . . . . . . . . . . . . . . . . 2 2 5

Chapter 9. Spectral Simulations of Turbulent Reacting F l o w s . . . . . . . 2 5 7

Chapter 11. Combustion Modeling Using PDF M e t h o d s . . . . . . . . . . . . 3 4 9

Chapter 12. Supersonic Reacting Internal Flow Fields . . . . . . . . . . . . . 3 6 5

Chapter 13. Studies of Detonation Initiation, Propagation, and

Chapter 14. Numerical Modeling of Heterogeneous Detonations.... .447

Chapter 15. Deflagration-to-Detonation Transition in Reactive

Chapter 16. Toward a Microscopic Theory of Detonations

Chapter 17. An Overview of Spray M o d e l i n g . . . . . . . . . . . . . . . . . . . . . 5 4 3

Chapter 19. Spray Combustion in Idealized Configurations:

Chapter 20. Comparisons of Deterministic and Stochastic

Chapter 21. Ignition and Flame Spread Across Solid Fuels . . . . . . . . . 6 4 3

Chapter 22. Pulse Combustor Dynamics: A Numerical Study . . . . . . . 6 7 3

Chapter 23. Mathematical Modeling of Enclosure Fires . . . . . . . . . . . . 7 1 1

Chapter 24. Nuclear S y s t e m s . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 5 5

Author Index for Volume 135 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 815

Ab Initio Quantum Chemistry for Combustion

Michael Page and Byron H. Lengsfield III

A DVANCES in theoretical and computational methods, coupled with

QC impacts combustion modeling in three nearly distinct ways. To ap-

Fig. 1 Schematic representation of link between first-principles quantum chemistry

approaches will be described in Sec. IV. These approaches incorporate the

II. Electronic Structure Methods

Given a set of fixed basis functions x, the /th MO is thus specified by a

A. Bond Breaking and the MCSCF Method

corresponding to the two-electron bond that is being broken is replaced

If the coefficients Al and A2 are determined variationally, the result is

can be constructed by distributing the active electrons among the active

CASSCF/MRCI wave function was found to be in excellent agreement

l) - X,(2)X8(2) dTl dT2 (5)

basis set expansion, providing a strong motivation for the development of

D. Perturbation Methods and Bond-Additive-Correction Schemes

(MP2). Perturbation theory methods are not variational: the computed

This is equivalent to a direct atomization energy calculation incorporating

III. Potential Energy Surfaces

atom molecule requires 106 calculations; a five-atom system, 109 calcula-

A. Born-Oppenheimer Potential Energy Derivatives

In analytic derivative evaluation, the energy expression for a particular

B. Locating Critical Points

mination of molecular structure. Equilibrium molecular structures are those

If the initial-guess geometry is close to the minimum, so that the surface

the matrix manipulations, e.g., diagonalizing or inverting the force constant

IV. Reaction Path Approaches

Another limitation of conventional TST occurs when motion across the

v(s) = (s) = - = (s) (9)

where the normalization factor c is defined as

If one imagines a trajectory starting at the saddle point and proceeding

A. Reaction Path Hamiltonian

Here Q is the generalized normal coordinate and co the associated harmonic

^1 = [FV - (v'Fv)v]/Vc (12)

B, Following Reaction Pathways