Вы находитесь на странице: 1из 329

Simulation and Modeling

of Turbulent Flows
ICASE/LaRC Series in Computational Science and Engineering

Series Editor: M. Yousuff Hussaini

Wavelets: Theory and Applications


Edited by
Gordon Erlebacher, M. Yousuff Hussaini, and Leland M. Jameson

Simulation and Modeling of Turbulent Flows


Edited by
Thomas B. Gatski, M. Yousuff Hussaini, and John L. Lumley
Simulation and Modeling
of Turbulent Flows

Edited by
Thomas B. Gatski
NASA Langley Research Center

M. Yousuff Hussaini
ICASE

John L. Lumley
Cornell University

New York Oxford


Oxford University Press
1996
Oxford University Press
Oxford New York
Athens Auckland Bangkok Bogota Bombay
Buenos Aires Calcutta Cape Town Dar es Salaam
Delhi Florence Hong Kong Istanbul Karachi
Kuala Lumpur Madras Madrid Melbourne
Mexico City Nairobi Paris Singapore
Taipei Tokyo Toronto
and associated companies in
Berlin Ibadan

Copyright © 1996 by Oxford University Press, Inc.


Published by Oxford University Press, Inc.
198 Madison Avenue. New York, New York 10016
Oxford is a registered trademark of Oxford University Press
All rights reserved. No part of this publication may be reproduced,
stored in a retrieval system, or transmitted, in any form or means,
electronic, mechanical, photocopying, recording, or any means otherwise,
without the prior permission of Oxford University Press.

Library of Congress Cataloging-in-Publication Data


Simulation and modeling of turbulent flows / edited by Thomas B. Gatski,
M. Yousuff Hussaini, John L. Lumley.
p. cm. — (ICASE/LaRC series in computational science and engineering)
Includes bibliographical references and index.
ISBN 0-19-510643-1
1. Turbulence—Mathematical models. I. Gatski, T. B.
II. Hussaini, M. Yousuff. III. Lumley, John L. IV. Series.
TA357.5.T87S56 1996
532'.0527'015118—dc20 96-11573

1 3 5 7 9 8 6 4 2

Printed in the United States of America


on acid-free paper
PREFACE

This book is based on the lecture notes of ICASE/LaRC Short


Course on Turbulent Modeling and Prediction, held on March 14-18,
1994. The purpose of the course was to provide the scientists and
engineers with a knowledge of the state-of-the-art turbulence model
development including the latest advances in numerical simulations
and prediction of turbulent flows. The lectures focussed on topics
ranging from incompressible, low-speed flows to compressible, high-
speed flows. A key element of this short course was the systematic,
rational development of turbulent closure models and related aspects
of modern turbulent theory and prediction.
The first chapter is based on the lecture by John Lumley written
in collaboration with Gal Berkooz, Juan Elezgaray, Philip Holmes,
Andrew Poje and Cyril Volte. It covers the basic physics pertaining
to turbulent scales and spectral cascades for both equilibrium and
nonequilibrium flows. It also includes a discussion of proper orthogo-
nal decomposition and wavelet representation of coherent structures
in turbulent flows.
The next two chapters are on the numerical simulation of tur-
bulent flows - the direct numerical simulation (DNS) by Anthony
Leonard, and the large-eddy simulation (LES) by Joel Ferziger. Both
have been among the pioneers in this field. The chapter on DNS
explains the critical issues of numerical simulation, and discusses
various solution techniques for the Navier-Stokes equations, in par-
ticular divergence-free expansion techniques and vortex methods of
which the author has been a leading proponent. The chapter on
LES examines the modeling issues, surveys the various subgrid-scale
models, and describes some accomplishments and future prospects.
In the fourth chapter written in collaboration with I. Staroselsky,
W. S. Flannery and Y. Zhang, Steven Orszag provides an introduc-
tion and overview of modeling of turbulence based on renormaliza-
tion group method, which he and his group has pioneered for over a
decade. Although the application of RNG method to Navier-Stokes
equations is by no means rigorous, this chapter illustrates its use-
fulness by the quality of results from its application to a variety of
turbulent flow problems. A key feature is the emphasis on the grey
areas which require further analysis.
vi Preface

The fifth chapter is based on the lecture by Charles Speziale on


modeling of turbulent transport equations, whereas the sixth chap-
ter, by Brian Launder, on the prediction of turbulent flows using
turbulent closure models. Both the authors are well-known leaders
in the area of turbulence modeling which is receiving ever increasing
attention in national laboratories and industries.
We want to thank Ms. Emily Todd for her usual skillful attention
to detail and organization which resulted in a very smooth week of
lectures, and to Ms. Barbara Stewart and Ms. Shannon Keeter who
typed or reformatted some of the manuscripts. Thanks also goes to
Ms. Leanna Bullock for her expert revamping of many of the figures
so that they could be electronically assimilated into the text.

Thomas B. Gatski
M. Yousuff Hussaini
John L. Lumley
CONTENTS

PREFACE v

INTRODUCTION 1

1 FUNDAMENTAL ASPECTS OF INCOMPRESSIBLE


AND COMPRESSIBLE TURBULENT FLOWS 5
John L. Lumley

1 INTRODUCTION 5
1.1 The Energy Cascade in the Spectrum in Equilibrium
Flows 6
1.2 Kolmogorov Scales 9
1.3 Equilibrium Estimates for Dissipation 10
1.4 The Dynamics of Turbulence 11
2 EQUILIBRIUM AND NON-EQUILIBRIUM FLOWS 13
2.1 The Spectral Cascade in Non-Equilibrium Flows ... 13
2.2 Delay in Crossing the Spectrum 14
2.3 Negative Production 19
2.4 Mixing of Fluid with Different Histories 20
2.5 Deformation Work in Equilibrium and
Non-Equilibrium Situations 23
2.6 Alignment of Eigenvectors 25
2.7 Dilatational Dissipation and Irrotational Dissipation 26
2.8 Eddy Shocklets 28
3 PROPER ORTHOGONAL DECOMPOSITION AND
WAVELET REPRESENTATIONS 29
3.1 Coherent Structures 29
3.2 The Role of Coherent Structures in Turbulence
Dynamics 32
3.3 The POD as a Representation of Coherent
Structures . 33
3.4 Low-Dimensional Models Constructed Using
the POD 37
3.5 Comparison with the Wall Region 42
3.6 Generation of Eigenfunctions from Stability
Arguments 52
3.7 Wavelet Representations 67
viii Contents

3.8 Dynamics with the Wavelet Representation in a


Simple Equation 68
4 REFERENCES 72

2 DIRECT NUMERICAL SIMULATION OF


TURBULENT FLOWS 79
Anthony Leonard

1 INTRODUCTION 79
2 PROBLEM OF NUMERICAL SIMULATION 80
3 SIMULATION OF HOMOGENEOUS
INCOMPRESSIBLE TURBULENCE 85
4 WALL-BOUNDED AND INHOMOGENEOUS FLOWS 86
5 FAST, VISCOUS VORTEX METHODS 91
6 SIMULATION OF COMPRESSIBLE TURBULENCE 100
7 REFERENCES 104

3 LARGE EDDY SIMULATION 109


Joel H. Ferziger

1 INTRODUCTION 109
2 TURBULENCE AND ITS PREDICTION 111
2.1 The Nature of Turbulence 111
2.2 RANS Models 112
2.3 Direct Numerical Simulation (DNS) 115
3 FILTERING 116
4 SUBGRID SCALE MODELING 118
4.1 Physics of the Subgrid Scale Terms 118
4.2 Smagorinsky Model 119
4.3 A Priori Testing 123
4.4 Scale Similarity Model 125
4.5 Dynamic Procedure 127
4.6 Spectral Models 132
4.7 Effects of Other Strains 135
4.8 Other Models 137
5 WALL MODELS 138
6 NUMERICAL METHODS 141
7 ACCOMPLISHMENTS AND PROSPECTS 143
8 COHERENT STRUCTURE CAPTURING 146
8.1 The Concept 146
Contents ix

8.2 Modeling Issues 148


9 CONCLUSIONS AND RECOMMENDATIONS . . . . 149
10 REFERENCES 150

4 INTRODUCTION TO RENORMALIZATION GROUP


MODELING OF TURBULENCE 155
Steven A. Orszag

1 INTRODUCTION 155
2 PERTURBATION THEORY FOR THE
NAVIER-STOKES EQUATIONS 159
3 RENORMALIZATION GROUP METHOD FOR
RESUMMATION OF DIVERGENT SERIES 162
4 TRANSPORT MODELING 169
5 REFERENCES 182

5 MODELING OF TURBULENT TRANSPORT


EQUATIONS 185
Charles G. Speziale

1 INTRODUCTION 185
2 INCOMPRESSIBLE TURBULENT FLOWS 187
2.1 Reynolds Averages 187
2.2 Reynolds-Averaged Equations 189
2.3 The Closure Problem 189
2.4 Older Zero- and One-Equation Models 190
2.5 Transport Equations of Turbulence 192
2.6 Two-Equation Models 193
2.7 Full Second-Order Closures 210
3 COMPRESSIBLE TURBULENCE 220
3.1 Compressible Reynolds Averages 221
3.2 Compressible Reynolds-Averaged Equations 221
3.3 Compressible Reynolds Stress Transport Equation . . 223
3.4 Compressible Two-Equation Models 226
3.5 Illustrative Examples 227
4 CONCLUDING REMARKS 234
5 REFERENCES 236
x Contents

6 AN INTRODUCTION TO SINGLE-POINT
CLOSURE METHODOLOGY 243
Brian E. Launder

1 INTRODUCTION 243
1.1 The Reynolds Equations 243
1.2 Mean Scalar Transport 245
1.3 The Modeling Framework 246
1.4 Second-Moment Equations 247
1.5 The WET Model of Turbulence 253
2 CLOSURE AND SIMPLIFICATION OF THE
SECOND-MOMENT EQUATIONS 255
2.1 Some Basic Guidelines 255
2.2 The Dissipative Correlations 257
2.3 Non-Dispersive Pressure Interactions 258
2.4 Diffusive Transport dij, dio 273
2.5 Determining the Energy Dissipation Rate 275
2.6 Simplifications to Second-Moment Closures 278
2.7 Non-Linear Eddy Viscosity Models 281
3 LOW REYNOLDS NUMBER TURBULENCE NEAR
WALLS 284
3.1 Introduction 284
3.2 Limiting Forms of Turbulence Correlations
in the Viscous Sublayer 286
3.3 Low Reynolds Number Modelling 288
3.4 Applications 299
4 REFERENCES 302

INDEX 311
LECTURERS

Joel H. Ferziger John L. Lumley


Thermosciences Division Sibley School of Mechanical
Mechanical Engineering and Aerospace Engineering
Department Upson and Grumman Halls
Stanford University Cornell University
Stanford, CA 94305-3030 Ithaca, NY 14853-7501
(415) 723-3615 (607) 255-4050

Brian E. Launder Steven A. Orszag


Department of Mechanical Program in Applied
Engineering and Computational Mathematics
UMIST 218 Fluid Dynamics Research Center
P.O. Box 88 Forrestal Campus
Manchester M60 1QD Princeton University
ENGLAND Princeton, NJ 08544-0710
(44) 161 200 3701 (609) 258-6206

Anthony Leonard Charles G. Speziale


Graduate Aeronautical Department of Aerospace
Laboratories and Mechanical Engineering
California Institute of Boston University
Technology 110 Cummington Street
Pasadena, CA 91125 Boston, MA 02215
(818) 356-4465 (617) 353-3568
This page intentionally left blank
Simulation and Modeling
of Turbulent Flows
This page intentionally left blank
Introduction

The aim of this book is to provide the engineer and scientist with the
necessary understanding of the underlying physics of turbulent flows,
and to provide the user of turbulence models with the necessary back-
ground on the subject of turbulence to allow them to knowledgeably
assess the basis for many of the state-of-the-art turbulence models.
While a comprehensive review of the entire field could only be
thoroughly done in several volumes of this size, it is necessary to focus
on the key relevant issues which now face the engineer and scientist
in their utilization of the turbulent closure model technology. The
organization of this book is intended to guide the reader through the
subject starting from key observations of spectral energy transfer and
the physics of turbulence through to the development and application
of turbulence models.
Chapter 1 focuses on the fundamental aspects of turbulence phys-
ics. An insightful analysis of spectral energy transfer and scaling
parameters is presented which underlies the development of phe-
nomenological models. Distinctions between equilibrium and non-
equilibrium turbulent flows are discussed in the context of modifica-
tions to the spectral energy transfer. The non-equilibrium effects of
compressibility are presented with particular focus on the alteration
to the turbulent energy dissipation rate. The important topical issue
of coherent structures and their representation is presented in the
latter half of the chapter. Both Proper Orthogonal Decomposition
and wavelet representations are discussed.
With an understanding of the broad dynamic range covered by
both the turbulent temporal and spatial scales, as well as their modal
interactions, it is apparent that direct numerical simulation (DNS) of
turbulent flows would be highly desirable and necessary in order to
capture all the relevant dynamics of the flow. Such DNS methods,
in which all the important length scales in the energy-containing

1
2 Introduction

range and in the dissipation range are accounted for explicitly is


presented in Chapter 2. Emphasis is on spectral methods for incom-
pressible flows, including the divergence-free expansion technique.
Vortex methods for incompressible bluff body flows are described
and some techniques for compressible turbulent flow simulations are
also discussed briefly.
Unfortunately, while the utopic desire is to perform simulations
of tubrulent flows without recourse to models of any type, the reality
of the broad spectral range of such flows, discussed and exemplified
in the first two chapters, precludes such calculations in most flows
of engineering interest. Thus, in computational studies using the
large eddy simulation (LES) method, the largest scales of motion
are represented explicitly, and the small scales are approximated or
modeled. A chapter on LES of reasonable length can not possi-
bly be comprehensive. However, Chapter 3 updates earlier reviews
and provides a relatively comprehensive, yet succint discussion on
the subject. It begins with some cryptic remarks on the nature of
turbulence and the prediction methods. Then the stage is set for
the discussion of LES with a brief overview of methodologies for the
Reynolds-aver aged Navie-Stokes equations and DNS. This overlaps
nicely with the preceding and following chapters which deal in some
detail with the DNS and phenomenological modeling of turbulence
respectively, and explains their relation to LES. It contains a long
section on subgrid-scale modeling which is a distinguishing feature of
LES, and further includes a section on numerical methods in practice
for solving the relevant equations. A key feature is the discussion of
accomplishments and exploration of the feasible boundaries for LES
applications. It concludes with a discussion of modeling issues and
author's retrospect and prospect.
As the reader can assess from both the DNS and LES formula-
tions, the attempt is to directly compute either all (or most) of the
turbulent scales, or just the large turbulent scales, explicitly. In the
LES approach, the concept of turbulence modeling makes its first ap-
pearance. The remainder of the book indeed focuses on the present
state-of-the-art approaches to the turbulence modeling problem. Un-
like the DNS and LES approaches, however, no attempt is made to
explicitly represent any of the turbulent scales through direct com-
putation. Rather, the modeling approach is to represent the effect
of the turbulence on the mean flow in toto. This can be done in a
variety of ways, but the emphasis in the remaining chapters will be
Introduction 3

on the development of transport equations for turbulent single point


second moments, such as the turbulent Reynolds stresses or turbu-
lent kinetic energy, solved in conjunction with a suitable turbulent
scale equation.
The concept of renormalization group (RNG) applied to the de-
velopment of turbulent closure models has been shown to be rather
successful. Chapter 4 discusses the basis for the RNG method and
its application to a variety of flow problems. Once again, an under-
standing of the spectral structure of turbulent flows, specifically, the
fundamental assumption of the universality of the small scales, plays
a key role in the application of the RNG technique. RNG theory
then provides a description, or model, of the small scales which can
be used to isolate the large scales. This leads to equations of mo-
tion for the large scales, and turbulence models for the prediction of
large-scale flow properties.
A more common formulation in turbulent flow prediction meth-
ods is the utilization of modeled transport equations for the turbu-
lent Reynolds stresses and/or turbulent kinetic energy. In Chapter
5, the theoretical foundations of Reynolds stress models in turbu-
lence are assessed from a basic mathematical standpoint. It is shown
how second-order closure models and two-equation models with an
anisotropic eddy viscosity can be systematically derived from the
Navier-Stokes equations for incompressible turbulent flows that are
near equilibrium and only weakly inhomogeneous. Properly cali-
brated versions of these models perform extremely well in the pre-
diction of two-dimensional mean turbulent flows that are not too
far from equilibrium. The development of reliable Reynolds stress
models for more complex turbulent flows, particularly those involv-
ing large departures from equilibrium or high-speed compressible ef-
fects, presents greater difficulties. In regard to the latter flows, re-
cent progress in the modeling of compressible dilatational terms is
discussed in detail. The central points of the chapter are illustrated
by a variety of examples drawn from compressible as well as incom-
pressible turbulent flows.
Higher-order models, such as the Reynolds stress or second-mo-
ment closures, are being used more frequently in solving complex tur-
bulent flows. Chapter 6 presents both the methodology and specific
closure proposals for modeling the turbulent Reynolds stresses. Par-
ticular attention is given to second-moment closure in which evolu-
tion equations are solved for the Reynolds stresses themselves. While
4 Introduction

the capabilities of 'the basic model', used in current CFD software,


are briefly reviewed, most attention is given to a new non-linear
closure based rigorously on realizability constraints. This scheme
permits many types of near wall flows to be handled without 'wall-
reflection' corrections. Also presented is a new non-linear eddy vis-
cosity model that links the turbulent stresses explicitly to strain and
vorticity tensors up to third order. This leads to a scheme with a bet-
ter sensitivity to secondary strains that is possible at the quadratic
level.
As the reader can readily see, these chapters provide both a thor-
ough introduction and state-of-the-art assessment of predicting tur-
bulent flows through simulations or transport equation modeling.
With this overall view of the field, the reader can begin to get a
clearer understanding of the focus of turbulent modeling research at
this time, and become sensitized to the important link between sound
mathematical and physical analysis underlying the development of
well-posed turbulence models.
Chapter 1

FUNDAMENTAL ASPECTS
OF INCOMPRESSIBLE AND
COMPRESSIBLE
TURBULENT FLOWS

John L. Lumley
Gal Berkooz, Juan Elezgaray, Philip Holmes
Andrew Poje, Cyril Volte

1 INTRODUCTION

Turbulence generally can be characterized by a number of length


scales: at least one for the energy containing range, and one from
the dissipative range; there may be others, but they can be expressed
in terms of these. Whether a turbulence is simple or not depends
on how many scales are necessary to describe the energy containing
range. Certainly, if a turbulence involves more than one production
mechanism (such as shear and buoyancy, for example, or shear and
density differences in a centripetal field) there will be more than one
length scale. Even if there is only one physical mechanism, say shear,
a turbulence which was produced under one set of circumstances
may be subjected to another set of circumstances. For example,
a turbulence may be produced in a boundary layer, which is then
subjected to a strain rate. For a while, such turbulence will have
two length scales, one corresponding to the initial boundary layer

5
6 J. L. Lumley et al.

turbulence, and the other associated with the strain rate to which the
flow is subjected. Or, a turbulence may have different length scales
in different directions. Ordinary turbulence modeling is restricted
to situations that can be approximated as having a single scale of
length and velocity. Turbulence with multiple scales is much more
complicated to predict. Some progress can be made by applying
rapid distortion theory, or one or another kind of stability theory,
to the initial turbulence, and predicting the kinds of structures that
are induced by the applied distortion. We will talk more about this
later. For now, we will restrict ourselves to a turbulence that has a
single scale of length in the energy containing range. We may take
this to be the integral length scale:

where p(r) is the autocorrelation coefficient in some direction, say

where x — {xi, x^, £3}, a is an arbitrary direction, and we adopt the


convention of no sum on Greek indices. In just the same way, any
turbulence will have at least two velocity scales, at least one for the
energy containing eddies, and one for the dissipative eddies. Exactly
the same remarks apply here, and many flows of technological interest
may be expected to have more than one scale of velocity. Here,
however, we will restrict our attention to flows that have a single scale
of velocity in the energy containing range. We will take as our scale
of velocity the r.m.s. turbulent fluctuating velocity, u =< u,-u,- > 1//2 ,
where we use the Einstein summation convention; that is, if an index
is repeated, we understand a sum over i = I — 3. Note that, in the
above, we are supposing that < . > is a long time average, or an
average over the full .space, or an ensemble average.

1.1 The Energy Cascade in the Spectrum in Equilibrium


Flows
Fourier modes are too narrow in wavenumber space to represent
physical entities. In physical space, the corresponding statement is
that Fourier modes extend without attenuation to infinity, while we
know that the largest entity of any physical significance in a turbulent
Fundamental Aspects 7

flow has a size not much larger than /. Hence, we must seek another
physical entity. Traditionally, we have talked about "eddies," but
these have never been very well denned. We want to introduce here
the wavelet. We will talk more later about wavelets, and how to use
them for a complete representation of the velocity field, and how to
construct physical models using wavelets. For now, we want to use
the simplest properties of wavelets. If we consider a clump of Fourier
modes, as the band in Fourier space becomes wider, the extent in
physical space becomes narrower. A reasonable size appears to be a
band lying between about 1.62/c and K/1.62; the numerical value can
be obtained by requiring that KO, — K/a = K, from which a = 1.62.
That is, the bandwidth is equal to the center wavenumber. This
results in a wavelet in physical space that is confined to a distance
of about a wavelength.
We can now discuss how these wavelets interact in Fourier space.
Reasoning in the crudest way, a wavelet exists in the strain rate
field of all larger wavelets. This strain rate field induces anisotropy
in the wavelet, which permits it to extract energy from the larger
wavelets. This energy extraction process is associated with vortex
stretching: when a vortex is stretched by a strain rate field, the
strain rate field does work on the vortex, increasing its energy, and
losing energy in the process. This process, of extracting energy from
the larger wavelets, and feeding it to the smaller wavelets, is known as
the energy cascade. In Tennekes and Lumley (1972), it is shown that
the cascade is not particulatly tight — a given wavelet receives half its
energy supply from the immediately adjacent larger wavelet, and the
other half from all its neighbors. Similarly, of all the energy crossing a
given wavenumber, three quarters goes to the next adjacent smaller
wavelet, and the remainder is distributed to all the even smaller
wavelets. Nevertheless, we can to a crude approximation consider
that the energy enters the spectrum at the energy containing scales,
and then is passed from wavelet to wavelet across the spectrum until
it arrives at the dissipative range, where it is converted to heat.
There is some discussion in the literature of what is called "back-
scatter." This is an unfortunate choice of words, since it suggests that
energy that started in one direction, is turned around and ends up
going in the other direction. This is not at all what is meant. Rather,
this refers to transfer of energy in the spectrum in the direction from
small to large wavelets. Now, there is no question that, taking av-
erages over long times or large regions of space, the energy transfer
8 J. L. Lnmley et al

in the spectrum of three-dimensional turbulence is from large scales


to small. In two-dimensional turbulence, however, the energy cas-
cade goes in the opposite direction, since the mechanism is totally
different. There is no vortex stretching in two dimensional turbu-
lence; instead, vortices coalesce to form larger vortices, and this is
the mechanism for energy transfer. Now, it is perfectly possible in
a three dimensional turbulence, if one considers short time averages,
or averages over only a small region of physical space, to have co-
alescing vortices, and hence locally, temporarily, energy transfer in
the "wrong" direction. Some initial instabilities are two-dimensional,
and for a while the energy transfer will surely be in the wrong di-
rection, until the flow is thoroughly three-dimensionalized, and the
energy transfer can proceed in the usual direction. Probably many
flows of technological importance, which are young — that is, not
fully developed, have highly anisotropic remnants of initial instabil-
ities, and turbulent structures that are highly anisotropic, and may
well have for limited times or over limited regions, energy transfer
in the "wrong" direction. It should not be difficult to build such a
process into a model of the energy transfer, using the idea that it is
probably associated with two-dimensionality.
Now, let us consider the transfer of energy from one wavelet to
another. If V(CLK) is the velocity typical of a wavelet with center wave-
number CLK, which has a size of roughly 2?r/aK, then the energy (per
unit mass) in this wavelet is v2(aK), and the rate typical of the energy
transfer will be set primarily by the strain rate of the wavelet at K/a,
which is f (K/a)/(27ro//v). Hence, the rate of energy transfer should
be approximately v2(aK,)v(K/a)/2Ka/K). If the Reynolds number is
large, and if the turbulence is in equilibrium, then we may expect this
quantity to be approximately constant across a considerable range of
wavenumbers in the middle of the spectrum, and it must be equal to
the dissipation s. Now, since a is not a large number, we may prob-
ably approximate this by •y 3 (K)/(2?r/K) = e. We should probably
include a constant to be on the safe side, but we would expect it to
be of order one, since physically the two sides should be of the same
magnitude. In fact, it turns out experimentally that the left hand
side is equal to about 0.3£. This, in fact, gives the classical form of
the high Reynolds number, inertial subrange equilibrium spectrum,
E ~ ae 2 / 3 /^" 5 / 3 , where we interpret EK — v2(K,), since the width of
the wavelet in Fourier space is K. We are approximating the integral
of E from K/a to OK by K£.
Fundamental Aspects 9

1.2 Kolmogorov Scales

In 1941, Kolmogorov suggested that, as the energy was passed


from wavelet to wavelet, it would lose detailed information about
the mechanism of energy production. If the number of steps in the
cascade was sufficiently great, we could presume that all information
would be lost. The small scales would know only how much energy
they were receiving. They might be expected to be isotropic (having
lost all information about the anisotropy of the energy-containing
scales). Note that this state of isotropy would exist only at infinite
Reynolds number (infinitely many steps in the cascade). At any
finite Reynolds number, the small scales would be expected to be
less anisotropic than the energy containing scales, but still somewhat
anisotropic.
Note that these ideas cannot be applied directly to the spectrum
of a passive scalar. It can be shown fairly easily that, in the presence
of a mean gradient of the scalar, velocity eddies can produce sharp
fronts or interfaces between quite different scalar values; these sharp
jumps correspond to very high wavenumber processes, and the ori-
entation of the jumps appears to be determined by the orientation of
the mean gradient. Hence, serious anisotopy is introduced into the
smallest scales.
We may also mention at this point that there is permanent an-
isotropy even in the smallest scales of the velocity spectrum of a shear
flow. This is discussed in Lumley (1992). The anisotropy exists be-
cause the energy from the mean flow is fed into one component, and
must be redistributed to the other two. However, while the amount
of anisotropy remains fixed as the Reynolds number increases, it
steadily decreases when considered as a proportion of the total mean
square velocity gradient. Hence, it is still correct to say (from at least
one point of view) that the velocity spectrum becomes increasingly
isotropic in the small scales as the Reynolds number increases.
With these reservations in mind, at very high Reynolds number
the smallest scales in the velocity spectrum will be aware only of
the amount of energy they receive, e (in an equilibrium situation).
Hence, we can make scales dependent only on v and £, and these are
known as the Kolmogorov scales:

If we adopt the Kolmogorov 1.962 position, and consider £ r , averaged


10 J. L. Lumley et al.

over a sphere of radius r, then of course we can define scales rf and


vr based on the local value of er. If we consider that within each
material domain there is a cascade, depending on the local value of
£ r , producing a Kolmogorov spectrum locally, then averaging this
spectrum (with certain assumptions on the distribution of £ r ) will
produce slight changes in the power of K.

1.3 Equilibrium Estimates for Dissipation

Now, it is a slight stretch to apply the ideas of section 1.2 in


the energy containing range, since it is not fair to assume that the
rate of energy transfer is determined entirely by the strain rate of
the next largest wavelet, since we are in a range of wavenumbers
where the wavelets are under the direct influence also of the mean
flow. However, we can certainly ask whether u2u/l is constant, and
possibly equal to £. In fact, it turns out that u3/I = e to within about
10%, which is at first surprising. More mature consideration suggests
that it is probably only true in flows that are in equilibrium, which is
to say, those in which the rate associated with the energy containing
eddies u/l is equal to the rate associated with the mean flow Uij. We
expect any turbulent flow to try to equilibrate all these rates, and
ultimately they will all be equal (or at least evolve proportionately).
We can use £ = u3/l to generate convenient forms for the various
scales. For example,

We can use these ideas to obtain an upper bound on the variation


of £ as a function of Reynolds number. Suppose that the dissipation
is so unevenly distributed that it is all in one tiny region. Call this
maximum value em. This region must have a size of order rjm, based
on £ m . Hence, the total dissipation must be given by

where we are assuming that the average dissipation < £ > is deter-
mined by averaging over regions of the size of the integral scale. This
is true in most flows - i.e.- the integral scale is of the order of the size
of the flow. Using the definition of r)m, we can easily obtain

where RI — ul/v.
Fundamental Aspects 11

1.4 The Dynamics of Turbulence


With this prelude about turbulent scales and the turbulent spec-
trum, we can now turn to a discussion of the energetic dynamics of
the turbulence. In what follows, we are paraphrasing Tennekes and
Lumley (1972), where all the missing details can be found.
We first split all quantities into a mean and a fluctuating part,
e.g. Ui = Ui + Ui, where < Ui >= £/,-, and < U{ >= 0, and we
designate the instantaneous total velocity by -u,-, with q2 given by
U{U{. The equation for the turbulent fluctuating kinetic energy can
be written as:

where < . > denotes some kind oi averaging (possibly time; one,
two or three-dimensional space; phase; or ensemble), but if we do
not indicate otherwise, we will take the average to be an ensemble
average. For the moment we will consider the flow to be incompress-
ible, Uiti = 0, and later we will consider what modifications we will
have to make for the case of compressibility. We are also designating
d ( . ) / d x j = (.),j. Then, the mean strain rate 5,-j is = ([/,-j + Ujii)/2,
and e is the dissipation of turbulent fluctuating kinetic energy. Prop-
erly speaking e = 1v < SijSjj >, where Sij is the strain rate of the
fluctuating motion; however, at high Reynolds number this can be
written as v < UijUij >. We have neglected a number of other terms
in equation (1.4.1) which can be shown to be small at high Reynolds
numbers. These are all of the form of transport terms - that is, they
can be written as divergences of something, and hence, if integrated
over a closed region, contribute nothing to the net turbulent kinetic
energy budget, but simply move kinetic energy from place to place.
These neglected terms are of importance in the neighborhood of the
wall, for example, where the local Reynolds number is low.
Recall that we are supposing that < . > is a long time average,
or an average over the full space, or an ensemble average. In an
inhomogeneous flow (where, of course, we must use either a time
average or an ensemble average), such a quantity may be a function
of position, or in an instationary flow (where we must use either a
space average or an ensemble average) it may be a function of time,
but it will be a smooth function, and will not be a random variable.
This was the kind of average envisioned originally in the early work
of Kolmogorov (1941). There are, of course, other possibilities. For
12 J. L. Lumley et al.

example, one could average over a sphere of radius r, centered at x,


and designate such an average as v < u^jU^j > r = er(x). This is
now a random variable that varies erratically from time to time and
from point to point in physical space; how much it varies depends
on the value of r. This was the point of view taken by Kolmogorov
in 1962. The 1962 point of view results in slight (but sometimes
significant) changes in the conclusions from the 1941 point of view.
We will generally take the 1941 position, unless otherwise noted.
The first term on the right of equation (1.4.1) is called the tur-
bulent kinetic energy production, or production for short. It can
also be identified as the deformation work, that is, the work done
to deform an arbitrary volume against the stresses induced by the
turbulence. Such a term appears as a drain in the equation for the
mean flow kinetic energy, and, of course, the work done against these
stresses goes into the mechanism responsible for the stresses, the
turbulence. There is a similar term describing deformation work
against the stresses induced by the molecular motions, but it is easy
to show that this work is small compared to the deformation work
done against the turbulent stresses, at high Reynolds numbers. This
means that, at high Reynolds numbers, the energy flow is from the
mean flow to the turbulence, and then to the molecular motion.
The second term on the right is the transport of fluctuating en-
thalpy p / p + q2/I. Equation (1.4.1) can be regarded equally as the
equation for the mean fluctuating enthalpy, < p / p + q2 /2 >, since
< p >= 0. The transport of any quantity <j> by the turbulence can be
written as — < (f>Ui >,;. That is, — < <j>U{ > is the flux of <j> per unit
time into a surface with a positive normal in the z-direction. (p is the
concentration of <^>-stuff per unit volume, while —Uj is the volume
per unit time per unit area entering the surface. It was hoped at
one time that the term < pui > /p would be small, largely because
it was difficult to measure and hard to predict. However, it is now
realized that it is probably of about the same magnitude as the other
term, and may be of the opposite sign. At least one model for this
term suggests that < pui > /p = —C < (g 2 /2)-u; >. DNS results
suggest that this may not be a bad model in nearly homogeneous
circumstances.
Homogeneous turbulence is observed to be approximately Gaus-
sian in the energy containing scales (turbulence is never Gaussian
in the small scales, due to the spectral transport, but more about
that in the next section). A Gaussian distribution has all zero third
Fundamental Aspects 13

moments, and hence all fluxes of the form < U{U3u^ > will be zero,
and hence < (g 2 /2)u; > will vanish. In a homogeneous flow, all
transport vanishes, since the transport terms are of the form of a
divergence, and spacial derivatives are all zero in a homogeneous sit-
uation. More than this, however, by crude physical reasoning, we
expect that probably all fluxes will vanish in a homogeneous flow - if
everything is statistically the same everywhere, there is no reason for
anything to flow from one place to another. Hence, non-zero fluxes,
and thus transport, are associated with a departure from a Gaus-
sian probability density, specifically with the appearance of skewness
in the density, associated with inhomogeneity. A formal expansion
has been developed for the case of weak inhomogeneity, relating the
third moments with the gradients (Lumley, 1978), which does not
work badly in practice (Panchapakesan and Lumley, 1993).
Finally, the last term is the dissipation of turbulent kinetic energy
per unit mass discussed previously. This is, technically, the rate at
which turbulent fluctuating kinetic energy per unit mass is converted
irreversibly to heat (to entropy). We will see later, however, that in
an equilibrium situation, this is also the rate at which kinetic energy
is removed from the energy containing scales, and the rate at which
kinetic energy is passed from scale to scale across the spectrum. In
a non-equilibrium situation, of course, these three quantities are not
necessarily equal to each other.

2 EQUILIBRIUM AND NON-EQUILIBRIUM FLOWS

2.1 The Spectral Cascade in Non-Equilibrium Flows

In this section we describe briefly a model which was first pre-


sented in Lumley (1992). First, in a steady turbulent flow, we believe
that the level of dissipation is determined by the rate at which turbu-
lent kinetic energy is passed from the energy-containing eddies to the
next size eddies; that is, by the rate at which the turbulent kinetic
energy enters the spectral pipeline, eventually to be consumed by
viscosity when it reaches the dissipative scales. Hence, this quantity
is only secondarily dissipation, and maybe should be called some-
thing like spectral consumption. This picture has the consequence
that the level of dissipation should be independent of Reynolds num-
ber at infinite Reynolds number; a change in the turbulent Reynolds
14 J. L. Lumley et al.

number should change only the wavenumber where the dissipation


takes place. These are all ideas of Kolmogoiov (1941).
There appears to be not much question that the level of dissipa-
tion in a steady state is determined by scales of the energy containing
range. Much more to the point, however, is the behavior of the dis-
sipation under changing conditions, both spacially and temporally.
We need a dynamical equation for the dissipation. Of course, we can
write down an exact equation for the dissipation, but as has been
detailed elsewhere (Tennekes and Lumley, 1972) this is expressed en-
tirely in terms of the small scales. To first order this is a balance
between two large terms, representing the stretching of fluctuating
vorticity by fluctuating strain rate, and the destruction of fluctuating
vorticity by viscosity. At the next order, these terms are slightly out
of balance, and the imbalance, of course, is governed by the energy
containing scales, but in ways that we do not fully understand. At
the moment, the entire equation must be modeled phenomenologi-
cally (see Lumley, 1978) - by analogy with energy, we suspect that
there must be production of dissipation and destruction of dissipa-
tion, and we suppose that production of dissipation should keep pace
with the production of energy, while the destruction of dissipation
should keep pace with the destruction of energy, and these concepts
give us a model that works reasonably well. We believe everyone who
uses this model, however, is uneasy.

2.2 Delay in Crossing the Spectrum

We can place this model on a much sounder physical footing.


The quantity that appears in the equation for the turbulent kinetic
energy is the true dissipation, as opposed to the rate at which en-
ergy is lost from the energy containing wavenumbers to the next
smaller wavenumbers. In a steady state, of course, these are equal,
but in an unsteady situation they may not be (see figure 2.2.1).
It seems likely that there will be a lag in the development of the
true dissipation, corresponding to the time it will take the energy
lost to the energy containing wavenumbers to be reduced in size
to the dissipative wavenumbers. We can compute this time lag,
using the model suggested in Tennekes and Lumley (1972). If
we divide the spectrum logarithmically into eddies centered at OK,
a 3 K, ... where a = (1 + 5 1/2 )/2 (fig. 2.2.2), so that the eddies
have the same width in wavenumber space as their center wavenum-
Fundamental Aspects 15

Figure 2.2.1. Impressions of the distribution of spectral flux against


wavenumber in steady state (a) and in unsteady state (b).

ber, then the time required to cross the spectrum is something like
I(O,K) / U(O,K,) + l(a3K,)/u(a3K) + l(a5K)/u(a5K) + ... + l(aN K,}/u(aN K),
where K is at the peak of the energy containing range, say nl = 1.3,
and aNKr/ = 0.55, to place this at the peak of the dissipation range.
Note that these eddies, which we introduced in 1972 are, in fact,
wavelets, as pointed out by Sreenivasan (Zubair et al. 1992). Adding
this up (supposing that all terms are within the inertial subrange)
we find that the total time T = 2(//u)(l - 1.29JRJT1/2), where / and u
are scales characteristic of the energy containing range. Bear in mind
that we should not pay too much attention to the numerical values
of the coefficients in the expression for T; probably the only thing
that is significant is the general form, and the value of the exponent.
Note also that, for low Reynolds number, the time shrinks to a very
small value, since the energy is dissipated at essentially the same
wavenumber where it is produced, while for high Reynolds number
it goes to 2l/u. Now, the idea of a simple lag suggests that the cas-
cade is tight, that all the energy must pass through each wavenumber
16 J. L. Lumley et al.

in order to arrive at the dissipative scales, as we have suggested in


figure 2.2.2. This would produce a hyperbolic behavior, a sort of
telegraph equation, like diffusion with a finite velocity (only in time
instead of space - see Monin and Yaglom 1971). We know, however,
from the discussion in Tennekes and Lumley (1972) that at each step,
although most of the energy is passed to the next wavenumber, a di-
minishing fraction is passed to all higher wavenumbers (see figure
2.2.3). Thus, the dissipative wavenumbers receive a small amount of
energy almost immediately, and increasing amounts as time goes on,
finally receiving it all in a time of the order of T.

Figure 2.2.2. Simplified view of spectral flux being passed from eddy
to eddy.

Figure 2.2.4 is reproduced from Meneveau et al. (1992) (their fig-


ure 11). The results are obtained from direct numerical simulation
of forced isotropic turbulence. The figure represents the time evolu-
tion of the energy in logarithmic wavenumber bands after a pulse of
energy was added to the first band. Band n represents the energy
in wavenumbers 2™~ 1 < k < 2 n , normalized to its value at t = 0.
Band 1 would be represented by a horizontal line at a value of 2.
The figure completely supports our speculation regarding the energy
transfer (at the end of the last paragraph).
The fractions and times involved are known (Tennekes and Lum-
ley, 1972) and it is, in principle, possible to work out as a function of
time, the energy received at the dissipative wavenumbers resulting
from a step in. input at the energy containing wavemtmbers. It is also
possible to determine this from exact simulations, using the codes of
Domaradzki et al. (1990). This approach is rather complicated, how-
Fundamental Aspects 17

Figure 2.2.3. More realistic view of spectral liux. Now the flux
crossing the wavenumber K goes mostly to eddy &K, but a decreasing
fraction goes to a 3 K, a 5 K, etc. In its turn (at the second step), that
which had gone to a,K is redistributed to a 3 K, a 5 K, etc., at the same
time that which had gone to a3K is redistributed to &5K, etc., and
that which had gone to a 5 K is redistributed to a 7 K and a 9 K, and . ..
This is only the second step. On the third step, each packet must
again be redistributed.

ever, and it seems likely that this can be modeled satisfactorily by an


exponential. We would thus expect that dissipation would be given
by something like

We are ignoring, for simplicity, the possibility that T is a function


of time, which it will be in general. Hence, e(i) would be governed
by the following differential equation:

T here can now be a function of time. All of these considerations


relate to a given mass of fluid of energy-containing scale, carried
along by the mean flow and the energy- containing velocities. Hence,
we must add to this equation an expression for the advection by the
mean flow and the turbulence. From the point of view of turbulence
modeling, this equation is of no use, since we need a value for £ in
order to determine a value for /. Hence, we do not have a value for /.
(Note that it would be satisfactory to use e/g 2 for u/l in T, whereas
it would not on the right hand side of the £ equation, since its use
18 J. L. Lnmley et al.

Figure 2.2.4. From Meneveau et al. (1992). The time evolution of


the energy in logarithmic wavenumber bands after a pulse of energy
was added to the first band. Band n represents the energy in wave-
numbers 2 n ~ 1 < k < 2n, normalized to its value at t = 0. Band 1
would be represented by a horizontal line at a value of 2.

there would vitiate the equation). The most likely candidate as a


model for u3/l is the turbulent energy production, —Uij < UiUj >,
although this is not quite right for several reasons, among them that
some of the turbulent energy extracted from the mean flow goes to
increase the kinetic energy, and some is transported, and in fact
the difference between the production and the transport and rate of
increase in kinetic energy is the dissipation, and not the rate at which
energy enters the spectral pipeline. However, it has been the best we
could think of, and has become standard, giving as an equation

This is the equation that has been used to obtain £ in turbulence


modeling essentially since the beginning, with the slight modification
that our time scale T is now a weak function of Reynolds number.
Fundamental Aspects 19

We have introduced a constant c, presumably of order unity, since


the turbulent energy production is only an approximation for the
rate at which energy enters the spectral pipeline.
We see that the second term (on the right) does not represent de-
struction of dissipation, but rather reflects the presumed exponential
rise of the energy arriving at the dissipative wavenumbers in response
to a step input to the spectral pipeline. It is an approximation to
the extent that the rise has been modeled as exponential. We can
see two ways to improve the equation: first, we can examine more
closely (on the basis, for example, of the simple model in Tennekes
and Lumley, 1972) the rate of arrival of energy at the dissipative
wavenumbers; it will surely be possible to develop successive approx-
imations, improving on the simple exponential behavior. We do not
think this is likely to make much difference.

2.3 Negative Production

Second, we can try to find a better approximation for the rate


at which energy enters the spectral pipeline. This is a lot harder
to do, and it is not at first clear where to search. Let us restrict
our attention to isothermal shear flows; if we can successfully handle
these, we can later consider other flows, buoyantly driven, for exam-
ple. The turbulent kinetic energy production is not a.bad guess in
many circumstances; it is usually estimated as u 3 //, but, of course,
this estimate is an equilibrium estimate, assuming the production is
approximately equal to dissipation, and dissipation is approximately
equal to the rate at which energy enters the spectral pipeline. The
current model suggests that in equilibrium situations production will
approximately equal dissipation, which is true at the edges of wakes,
jets and shear layers, for example. We are interested here, however,
in precisely those situations that are not in equilibrium. For exam-
ple, there are regions in many turbulent flows where the turbulent
kinetic energy production is zero (or even negative in small regions),
partly due to the vanishing of the mean strain rate, but also due to
the vanishing of the Reynolds stress resulting from advection by the
fluctuating velocity field of material with different strain histories.
The centerlines of wakes and jets are examples where the produc-
tion goes to zero. However, energy will still be entering the spectral
pipeline in these regions. Hence, we may expect our model using
the turbulent kinetic energy production to fail in these regions. We
20 J. L. Lumley et al.

might hope to build a slightly better model if we take the magni-


tude of the mean strain rate as an estimate of u/l = [ S i j S i j ] 1 / 2 , and
write u3/I = [SijSij]l/2q2/3. This has a similar problem, in that the
strain rate vanishes at extrema of the mean velocity profile, where
energy is certainly still entering the spectral pipeline. It also suggests
that at equilibrium the (inverse) time scale based on e/q2 is equal
to that based on [ S i j S i j ] 1 ' 2 , which is probably only true in regions
far from extrema where most of the fluid has been subjected to the
same strain history during living memory.

2.4 Mixing of Fluid with Different Histories

To get a model that avoids these difficulties, we must consider


what it is that determines the time scale of a material region. If
a material region remains subject to the same sign and magnitude
of strain rate during its entire lifetime (that is to say, for times of
the order of l/u), then we expect that the characteristic time scale
of the region, l/u, will become equal to the magnitude of the (in-
verse) strain rate [ S ^ j S i j ] ' 1 / 2 . In an inhomogeneous flow, however,
at a given point, as time passes, fluid will arrive from many different
regions, where the sign and magnitude of the strain rate are quite dif-
ferent (figure 2.4.1). This is why the Reynolds stress on the centerline
of a wake is zero - each packet of fluid brings its own value of Rey-
nolds stress, depending on its history, and since the Reynolds stress
is both positive and negative, and on the centerline equal quantities
of fluid are seen, that have spent their lifetimes in regions of positive
and negative strain rate, the net value is zero. Hence, we should
consider averaging over the magnitude of the strain rates in the ar-
eas from which fluid is advected to the point in question; ideally,
backward along the mean trajectory though the point in question,
with a growing Gaussian (in first approximation) averaging volume,
with a fading memory (figure 2.4.2). We need not actually write
this as an integral in our equation; we can rather write an auxil-
iary equation for the quantity. We might also consider averaging
the production itself (instead of the strain rate magnitude), which
would give a different weighting. However, the rate at which energy
enters the spectral pipeline should be determined by the local value
of the energy, and the local value of the time scale, the latter being
determined by history.
Fundamental Aspects 21

Figure 2.4.1. Zone of influence in a turbulent flow. The point shown


is influenced by the histories of material points arriving (at different
times) along the trajectories indicated.

If we consider an auxiliary quantity, an inverse time scale, or rate,


which we may call S, we need an equation of the type

where VT — c'q4/e, and T = ciq2/£, where c' and c\ are constants


of order unity. This will give just such a spreading Gaussian average
with fading exponential memory of scale 7~ back along the mean
streamline. The exact values of c' and c\ probably are not too critical,
since they only determine the exact size of the region over which the
averaging is done. Then we can write
22 J. L. Lumley et al.

Figure 2.4.2. The real process of figure 2.4.1 replaced by a model,


in which the point shown is influenced by an integral back along
the mean streamline through the point, with a spreading zone of
influence and a fading memory.

where c" is another constant of order unity. Note that this is now
a non-local theory. Stan Corrsin (1975) pointed out some years ago
that the k — e model was also non-local for similar reasons.
If we apply this model to grid turbulence we may obtain some
relations among the constants. We should identify S with (propor-
tional to) the value of u/l determined by whatever mechanism. In
a grid turbulence, the initial value of u/l is determined by the grid;
thereafter, there is no further input to determine the value of u/l
(since Sjj is identically zero) and it simply relaxes, or decays. In
an equilibrium homogeneous shear flow (which may not exist), S
will take on the value [SijSij]1/2 asymptotically, which will have the
value f/'/A/2 = u/l^/2. If we identify the initial value S0 (at time
Fundamental Aspects 23

t 0 ] of S in the decaying grid turbulence as the value of u/l\/2 at


t0, we have S0t0/3 — n/2\/2, where q2 oc t~n. In order for similar-
ity to hold, we require cj = n/2, and c" = \/2 — C22\/2(ft + l)/ft,
where clearly we require c2 < n/2(n + 1). A typical value of n is
about 1.25, although there is a weak variation with Reynolds num-
ber that should be investigated. From our reasoning we expect that
1 /o
c2 oc (1 — c^Rl ). It seems reasonable to require both c\ and c2
to have this behavior with Reynolds number, and hence to require
that n should have this behavior also. We should examine the data
of Comte-Bellot and Corrsin (1966) for confirmation.

2.5 Deformation Work in Equilibrium and


Non-Equilibrium Situations
We can construct a simple picture of the turbulence production
—Sij < UiUj >. Imagine a velocity field that consists of randomly
oriented vorticity with equal amounts in all directions. If we write
the production in principal axes of the mean strain rate, it becomes

We presume that the 1-direction is the direction of maximum pos-


itive strain rate, while the 3-direction is the direction of maximum
negative strain rate. The strain rate in the 2- direction is intermedi-
ate in value; it may be positive or negative, but it will be smaller in
magnitude that the other two. Now, the vorticity in the 1-direction
will be stretched and intensified, while the vorticity in the 3-direction
will be shrunk and attenuated. Associated with the 1-direction vor-
ticity is 2- and 3-direction velocity, which will be intensified, while
associated with the vorticity in the 3-driection is velocity in the 1-
and 2-direction, which will be attenuated. We can make a chart:

where a horizontal arrow indicates no change, while an up arrow


indicates an increase, and a down arrow a decrease. As a result of
24 J. L. Lumley et a].

£22, W2 may go up or down somewhat, and there will be consequent


small changes in < u\ > and < u§ >, but these will be smaller
than the changes due to wj and u^. We have designated these as "no
change" with a horizontal arrow, for simplicity. If the initial vorticity
is more-or-less uniformly distributed, it is evident that there will be
a net decrease in < u\ > and a net increase in < u| >, while < u\ >
will remain essentially unchanged. As a result, < u\ > — < u? >< 0,
while < w§ > - < u\ » 0. Since 5n > 0, and 633 < 0, both terms
of the expression for the production will be positive.
What we have described is an equilibrium situation. That is, we
have described a situation in which the anisotropy of the turbulence
has been generated by the strain rate field. The anisotropy of the
turbulence is consequently in equilibrium with the strain rate field.
In the real world, of course, this sometimes happens; however, it is
also quite likely that the turbulence will have been generated by one
mechanism, will have lived its entire lifetime under this mechanism,
and will be in equilibrium with that mechanism, in the sense that
the time scales will have equilibrated, and the principal axes will
have taken on an equilibrium orientation, and that this turbulence
will then be subjected to a distortion of a wholly different nature.
For example, the boundary layer formed on the pressure surface of
a leading edge slat on an aircraft wing (in take-off configuration) is
suddenly subjected to the strain rate due to passage through the
gap between the slat and the wing. Under these circumstances, the
production can take on both positive and negative values. A good
example of this is the wall jet, which has a maximum which is not
symmetric. In this vicinity, material is sometimes swept past which
has come to equilibrium with the strain rate field on the wall side of
the maximum, and sometimes from the other side of the maximum,
where the strain rate has the opposite sign. Because of the lack of
symmetry, the amounts of the two types of material are not equal,
and hence the net value of the Reynolds stress does not vanish at the
point where the strain rate vanishes. There is, thus, a narrow region
in which the production is negative. A more impressive situation
can be generated in a special wind tunnel. Let the flow be in the x\
direction, and the tunnel be arranged to produce a positive strain
rate along the x? axis, and a negative strain rate along the #3 axis.
If the turbulence is subjected, to this distortion for long enough, the
turbulence structure will have come to equilibrium with the strain
rate. At this point, reverse the direction of the strain rate - that is,
Fundamental Aspects 25

let the positive strain rate be in the x3 direction, and the negative
strain rate be in the x-i direction. This is easy to arrange in practice -
the direction that had been shrinking now begins to expand, and vice
versa. Experimentally, it is found that the production immediately
becomes negative throughout the tunnel and remains so until the
anisotropy of the turbulence can adjust itself to the new value of the
strain rate field, which takes some time. Any model for turbulence
which hopes to deal with non-equilibrium situations must take this
into account - this means, in practice, that a separate equation must
be carried for the Reynolds stress; only in equilibrium situations can
the value of the Reynolds stress be related directly to the strain rate
field.

2.6 Alignment of Eigenvectors

The question of whether the turbulent field is in equilibrium


with the mean strain rate field arises also in connection with the
eigenvectors. This is really just another way of looking at the same
question, perhaps a more enlightening way. In a shear flow, say
Ui : Ui(x2),0,0, the principal axis of positive strain rate is at 7T/4,
while the principal axis of negative strain rate is at 37T/4. In a pure
strain (i.e. with no rotation), the principal axes of the Reynolds stress
would be aligned with those of the strain rate field. In our shear flow,
however, we have rotation as well as strain rate. A material region
is being continually rotated clockwise in this flow. Hence, although
the strain rate tries to align the principal axes of the Reynolds stress
with its own principal axes, the material with its principal axes is
rotated clockwise. Hence, we expect to find the principal axes of
the Reynolds stress rotated clockwise from those of the strain rate.
The relaxation time of the turbulence is of the order of l/u, and
the mean angular velocity is (\/1)dU\/dx^. The net angle through
which the axes will rotate is perhaps half of this, to give something
like ( l / 4 ) ( l / u ) d U i / d x 2 = 1/4, or some 15°. By picking the value of
1/2, we are trying to account for the fact that the relaxation is going
on continually - we are replacing the real process with an artificial
one which does not relax at all until l/2u, and then relaxes com-
pletely. In fact, the principal axes of the Reynolds stress are found
at 30° and 120°, rotated clockwise exactly 15° from the axes of the
mean strain rate. The agreement with our crude calculation is too
good to be true, but at least the direction and order of magnitude
26 J. L. Lumley et al.

are both correct and believable.


In turbulence modeling of the k — £ type, the Reynolds stress is
parameterized as being proportional to the mean strain rate, which
means that the two tensors are forced to have the same principal axes.
The relationship is calibrated to give the correct value of the off-
diagonal stress, which means that the diagonal stresses, the turbulent
intensities, must be wrong. In calculating simple shear flows, this
does not matter much, since the normal stresses are not used. In
more complex situations; however (such as separated flows), where
any of the components may be the important one, it matters, slit is
also not clear in such situations how to calibrate the relationship.
In non-equilibrium situations, it is essential that the equation for
the Reynolds stress be used, so that the rotation and relaxation of
the principal axes of the Reynolds stress can gradually accommo-
date to the mean strain rate history, being at any instant probably
misaligned with those of the mean strain rate.

2.7 Dilatational Dissipation and Irrotational Dissipation

Up to this point, we have ignored compressibility. In fact, in the


boundary layer at low Mach number U/u* is about 30 (between 25
and 35) between a length Reynolds number of 107 and 5 X 10s. The
speed q is about 2.65u*. Hence, the mean velocity'is about 11 times
the turbulent speed. As the Mach number rises, the wall temper-
ature rises, and the density and, hence, the skin friction drops, so
that this ratio increases somewhat. This means that the fluctuating
Mach number is of the order of 1/11 or smaller of the mean flow
Mach number. As a result, unless the mean flow is hypersonic, with
a Mach number in the neighborhood of 12-15, the turbulent Mach
number will not be anywhere close to one. Hence, at moderate mean
flow Mach numbers (say, below 5) we may expect the effects of com-
pressibility on the turbulence to be relatively small. An exception, of
course, is interaction with a shock wave; boundary layer turbulence
may be essentially incompressible, but if it passes through a shock
compressibility effects will be felt. Note that the relative fluctuations
in the isentropic speed of sound are of order (7 — l)m 2 /4, where m
is the fluctuating Mach number. Hence, even at a fluctuating Mach
number of unity, the fluctuations in the isentropic speed of sound are
of order 10%, and can be ignored in the definition of the fluctuating
Mach number.
Fundamental Aspects 27

Let us consider the dissipation. Let us begin by considering the


stress in a compressible flow. This can be written as

where 9 = Ui^,s'- is the deviatoric strain rate, and p is the thermody-


namic pressure, that is, p = (7 — l)/oe, where e is the internal energy.
[iv is the bulk viscosity. This is zero in a monatomic gas, and is of
the order of Q.66/J, in Nitrogen (Sherman, 1955). The (negative of)
the average normal stress —r,-;/3 is not equal to the thermodynamic
pressure because there is a lag between the rotational temperature
and the translational temperature. Under compression (9 negative),
all temperatures are rising, but the rotational temperature is lagging
behind the translational temperature, so that the translational tem-
perature is a little higher than the thermodynamic relation predicts.
Thus the actual (negative) average normal stress is higher than the
thermodynamic pressure (see, for example, Light hill, M. J., 1956).
From (2.7.1) we can see that the viscous stress depends on the di-
latation only through the bulk viscosity. In a monatomic gas there
would be no dependence.
If we now form the equation for the turbulent fluctuating energy
in a homogeneous flow, we can write

The first term, of course, is the pressure-dilatation correlation, or the


recoverable work. The remainder is the entropy production.
Now, we may write (if 5,-j = (uij + Uj^/2, r\,j = (uitj - -Uj i ,-)/2),

The second term may be written as < u^jUj^ > — < O2 > (in a homo-
geneous flow), where 9 = w ti ,. Hence, we can write (in a homogeneous
situation)

In addition, we may write < TIJTIJ >=< Wj-u;; > /2, and < SijSij > =
< s'-s'- > + < O2 > /3, so that we may write
28 J. L. Lumley et ad.

Finally, the total entropy production may be written as

Let us consider the Helmholtz decomposition. This purely kine-


matic decomposition states that any vector field Uj can be written
uniquely as the sum of two components, say Vi + W{ , where «,- is
solenoidal (but rotational), while w^ is irrotational (but compress-
ible). Viewed from this perspective, Ui is associated only with the
solenoidal component v,-, while 0 is associated only with the irrota-
tional component w^. It thus seems fairly safe to identify // < u;;uz- >
as the conventional (solenoidal) dissipation, and (fiv + 4///3) < # 2 >
as the dilatational dissipation. Note that, in an inhomogeneous situ-
ation, there are other terms that vanish only as the Reynolds number
becomes infinite.
Both Zeman (1990) and Sarkar (1992) have developed models
for the compressible component of the dissipation. In the compress-
ible mixing-layer flow (Sarkar and Lakshmanan, 1991), this com-
pressibility correction is essential for predicting the spreading rate
correctly. In an extension of the dilatational dissipation model-
ing, Sarkar (1992) has also examined contributions to the pressure-
dilatation using an analogous decomposition of the pressure field into
incompressible and compressible parts. The proposed model has
been compared to results from DNS of compressible homogeneous
shear flow (Sarkar et a/., 1991).

2.8 Eddy Shocklets


We tell our classes, and we believe, that in incompressible turbu-
lence the level of the dissipation is controlled by the rate at which
energy is fed into the spectral pipeline at the large scale end. Now,
even when the fluctuating Mach number is above unity, we expect
to find relatively incompressible turbulence separating a distribu-
tion of randomly oriented shocklets, which are relatively thin, and
hence correspond to high wavenumbers. Thus, we expect the energy
containing range (characterized by motions which, though energetic,
change only slowly in space) to be relatively incompressible and the
rate at which energy is transferred into the spectral pipeline in the
usual way should be unchanged from the incompressible case. Hence,
the ordinary dissipation should be unchanged (Zeman, 1990). Pre-
sumably in a compressible turbulence, the compressibility can be
Fundamental Aspects 29

characterized by the spectrum of < $2 >, which might be expected


to rise with wavenumber, so that the compressibility would occur
primarily at the larger wavenumbers. This supports our physical
argument, and is consistent with the shocks being thin.
Consider the dilatational dissipation from a physical point of
view. Energetic eddies, which locally exceed m = 1, form shock-
lets, and the passage through these shocklets removes energy from
the eddy. The dilatation is almost completely confined to the shock-
let, hence to high wavenumbers. Somehow, the energy is getting
from the low wavenumbers to the high wavenumbers, where it is dis-
sipated. This certainly does not seem to be by the usual cascade
process of vortex stretching. It seems rather, that in the passage
through the shock, the eddy is compressed in the direction normal
to the shock, and this reduces its scale, and increases its wavenum-
ber. In addition, the existence of the shock alters the velocity field
approaching it; the flow tries to avoid the shock, turning aside to go
around it if possible, and this may also result in a reduction in scale.
Certainly, it results in a transfer of energy from one component to
another. This is presumably < p9 > at work. There are indications
from the work of Zeman (1991) that < p9 > stores energy during
passage through the shock, returning it to the vortical mode down-
stream of the shock, causing an increase in the turbulent intensity
on the downstream side. On the other hand, certain simple models
suggest that under some circumstances the term < p9 > will have
the form of an additional dissipation. It probably does all of these.
Incidentally, it is clear from our discussion above that < p9 >
and [iv < d2 > are just two parts of the same thing: < (—r,-,-/3p)# >,
the work that is done by the normal stress during the compression,
not all of which is recoverable.

3 PROPER ORTHOGONAL DECOMPOSITION AND


WAVELET REPRESENTATIONS

3.1 Coherent Structures

If we examine pictures of various turbulent flows, we will dis-


cover that the proportion of organized and disorganized turbulence
in each flow is different. For example, if we look at mixing layers
30 J. L. Lumley et al.

from undisturbed initial conditions (with only thin laminar bound-


ary layers on the splitter plate), we find that there is an energet-
ically large organized component, which only relatively slowly be-
comes three-dimensional and disorganized, although the nearly two-
dimensional organized structures have from the beginning a stochas-
tic component, so that their occurrence is not precisely periodic,
nor are their strengths equal. On the other hand, if we examine a
mixing layer from quite disturbed initial conditions, with a thick,
turbulent boundary layer on the splitter plate, we find that the pro-
portion of organized component is considerably less - although the
organized component is still visible, it is no longer dominant. Thus,
in the same type of flow, we find that the initial conditions change
the relative strength of the organized and disorganized components.
We examine a different flow, for example a jet, we find the same
difference - that is, if the flow from the orifice is initially undis-
turbed - thin laminar boundary layers on the inner surface of the
nozzle - then there is initially a laminar instability which gradually
becomes three-dimensional and undergoes transition, leaving in the
downstream development of the turbulence the remnants of the in-
stability structure. On the other hand, if the boundary layers on the
inside of the nozzle are initially thick and turbulent, there is no ini-
tial laminar instability, and there are no visible organized remnants
in the turbulent motion. More than this, however, there is evidently
a substantial difference between this flow and the mixing layer. Even
in the undisturbed state, the organized component of this flow is a
very great deal weaker than that in the mixing layer under the same
circumstances. It is bearly discernible; when the initial conditions are
disturbed, the organized structure becomes essentially undetectable.
Thus, we can conclude that different flows, even under similar condi-
tions, have different relative strengths of organized and disorganized
components.
So far it seems that the organized structures, to the extent that
they are present, are the remnants of initial laminar instability. Let
us look at the wake of a flat plate normal to the stream. If the
wake is visualized close to the plate, we see one kind of organized
structure, which evidently is the remnant of the initial laminar in-
stability. However, if the wake is visualized far from the body, we
see a somewhat, different, though similar organized structure. Ev-
idently the organized structures initially present decayed, and new
distinct, though similar, structures arose. If we examine the wake of
Fundamental Aspects 31

a plate (still normal to the stream) which is sufficiently porous so as


not to give rise to an initial instability, we find at first a turbulent
wake without organized structures. However, after a time the or-
ganized structures present in the late part of the initially disturbed
wake spontaneously appear in this initially undisturbed wake (we say
undisturbed, but of course the wake is initially turbulent, and hence
disturbed; however, organized structures are not initially present.)
It seems reasonable to conclude that the organized structures that
appear in the late wake are a type of instability of the developed tur-
bulent flow, drawing on the mean velocity profile to obtain energy,
and giving up energy to the turbulent transport. The precise form
then, will be a function of the mean velocity profile, as well as of the
distribution of the turbulent stresses. The energy budget for such ex-
isting organized structures is complex, because they have reached a
non-linear energetic equilibrium, and their transport is modifying the
mean velocity profile, as well as the turbulent stresses. Their initial
growth is complicated also, since they must be imagined to grow from
a mean velocity profile (and a profile of turbulent stresses) existing
without the organized structures. What these profiles are is moot.
We will return to these ideas later when we discuss the prediction of
these organized structures.
Hence, in any given situation we may expect to find organized
structures the relative strength of which are a function of the initial
conditions of the flow, the type of flow, where we are in the flow (how
far downstream), and which may be remnants of initial instabilities
or may be a new instability of the turbulent flow.
If we restrict our attention to narrow two-dimensional shear flows
(jets, wakes, mixing layers) we will find that the organized structures
occur with more-or-less the same orientation and distance from the
flow centerline each time; that, is, their orientation and position
in the cross-stream, inhomogeneous direction is largely fixed by the
boundaries of the flow; in the streamwise, or homogeneous direction,
the location is more random; the existence of one structure seems to
suppress the presence of another, but as soon as we are sufficiently
distant from a structure, another one appears.
Organized structures are much more difficult to find in homo-
geneous flows, mostly because it is not clear where to look. There
is nothing in the flow to pin them down to a particular location.
Consider, for example, the homogeneous shear. In direct numeri-
cal simulations, Moin and his coworkers have found hairpin vortices
32 J. L. Lumley et al.

throughout the flow. The orientation of these hairpin vortices is


determined by the direction of the mean velocity shear, but their
location in three dimensions is random. They are thought to arise
from a type of non-linear instability connected with the same in-
stability that produces Langmuir cells in the ocean surface mixed
layer, or that produces streamwise rolls in a turbulent boundary
layer. Both of these flows are inhomogeneous in the direction normal
to the surface, and hence the location of the organized structures is
fixed by this inhomogeneity. The instability mechanism depends on
transverse vorticity associated with the mean shear being deflected
vertically (in the direction of the mean gradient), and then being
transported in the streamwise direction at different rates at different
heights, due to a gradient in the Stokes drift, resulting in a stretching
and intensification of the streamwise component of vorticity.
We may probably conclude that any turbulent flow will have a
more-or-less organized component, the strength of which will be a
complex function of the type of flow, the age of the flow, and the
initial conditions, and which may, depending on the situation, be
random in orientation and location in up to three dimensions.

3.2 The Role of Coherent Structures in Turbulence


Dynamics

Whether it is necessary to take into account the presence of orga-


nized structures in a turbulent flow, when considering the dynamical
behavior, will have a different answer in different situations. When
we study a turbulent flow for practical purposes, we are seldom in-
terested in more than the Reynolds stress. This is not a very so-
phisticated property of a turbulent flow; it is uninfluenced by subtle
changes in the structure of the flow, and tells us little about the flow.
Often, if the coherent structures in the flow scale in the same way as
the disorganized motions, they can all be lumped together and the
evident differences ignored. For example, in the turbulent mixing
layer, if both the turbulence and the organized structures have the
same origin, and have been growing together since the origin, then
they will scale in the same way, and need not be considered sepa-
rately. See Shih et al. (1987), where a compressible turbulent mix-
ing layer was successfully predicted using a second order turbulence
model that completely ignored the presence of coherent structures.
Many of the features of the mixing layer that are thought to be as-
Fundamental Aspects 33

sociated with the coherent structures (e.g. the asymmetry of the


entrainment at the two sides of the mean velocity profile) are in fact
mandated by the dynamics of the situation and must be produced by
whatever physical mechanism is doing the transporting of momen-
tum, whether organized or disorganized. However, if a turbulence
produced under one set of circumstances, and consequently having
a given set of scales, is subjected to a different set of conditions (a
new strain rate field, for example) it is quite likely that the new con-
ditions will give rise to a new instability of the turbulent profiles,
giving rise in its turn to a new organized structure. Until the scales
have had a chance to equilibrate, we will have a situation consist-
ing of background disorganized turbulence with one set of scales, on
which are growing organized structures with a different set of scales.
The transport produced by this combination will be quite difficult
to predict unless explicit account is taken of the organized structure.
We must also consider that the initial turbulence may already have
an organized structure with which it is in equilibrium.

3.3 The POD as a Representation of Coherent


Structures

The extraction of deterministic features from a random, fine


grained turbulent flow has been a challenging problem. Zilberman
et al. (1977) write: "there are no consistent methods for identifica-
tion which are independent of the techniques and the observer" and
"we cannot unambiguously define the signature of an eddy without a
priori knowledge of its shape and its location relative to the observa-
tion station and cannot map such an eddy because we do not have a
proper criterion for pattern recognition." In contrast, Lumley (1967)
proposed an unbiased technique for identifying such structures. The
method consists of extracting the candidate which is the best corre-
lated, in a statistical sense, with the background velocity field. The
different structures are identified with the orthogonal eigenfunctions
of the proper orthogonal or Karhunen-Loeve decomposition theorem
of probability theory (Loeve, 1955). This is thus a systematic way
to find organized motions in a given set of realizations of a random
field. The method applied here is optimal in the sense that the series
of eigenmodes converges more rapidly (in quadratic mean) than any
other representation.
The use of these modes for a low dimension dynamical system
34 J. L. Lumley et al.

study requires a very fast convergence of the series. The method


we propose here is limited in application to certain types of flows
in which large coherent structures contain a major fraction of the
energy. It has been demonstrated that axisymmetric turbulent jet
mixing layers (Glauser et a/., 1985) and wall regions of turbulent
boundary layers (Moin, 1984; Herzog, 1986) belong to this group.
Specifically, we will develop a model for the wall region of the
boundary layer (from x 2 + = 0 to x%+ = 40 in wall units (Tennekes
and Lumley, 1972)), using the proper orthogonal decomposition of
Lumley (1967, 1970, 1981) in the direction normal to the wall, in
which the flow is strongly non-homogeneous. In the stream wise and
spanwise directions the flow is essentially homogeneous, and Fourier
modes will suffice. Used in conjunction with Galerkin projection,
the proper orthogonal decomposition yields an optimal set of basis
functions in the sense that the resulting truncated system of ODEs
captures the maximum amount of kinetic energy among all possible
truncations of the same order. The method has obvious advantages
over a priori decompositions, based on linear normal modes, but it
does not appear to have been used before due to the difficulty of
computing the proper orthogonal modes. For this one requires three
dimensional autocorrelation tensors averaged over many realizations
of the flow in question, data only obtainable from lengthy experi-
ments and analyses or from detailed numerical simulations. In our
case complete data is only available from experimental work in a glyc-
erine tunnel (Herzog, 1986), although Moin (1984), has derived two-
dimensional orthogonal modes from large eddy simulations. How-
ever, as we shall see, knowledge of the autocorrelation tensor, and
use of the Navier-Stokes equations, does allow one to uniquely deter-
mine the unsteady flow, in contrast to Cant well's (1981) expectation.
Lumley (1967) proposed a method of identification of coherent struc-
tures in a random turbulent flow. An advantage of the method is its
objectivity and lack of bias. Given a realization of an inhomoge-
neous, energy integrable velocity field, it consists of projecting the
random field on a candidate structure, and selecting the structure
which maximizes the projection in quadratic mean. In other words,
we are interested in the structure which is the best correlated with
the random, energy-integrable field. More precisely, given an ensem-
ble of realizations of the field, the purpose is to find the structure
which is the best correlated with all the elements of the ensemble.
Thus, we want to maximize a statistical measure of the magnitude of
Fundamental Aspects 35

the projection, which can be given by the mean square of its absolute
value. The calculus of variations reduces this problem of maximiza-
tion to a Fredholm integral equation of the first kind whose sym-
metric kernel is the autocorrelation matrix. The properties of this
integral equation are given by the Hilbert Schmidt theory. There is
a denumerable set of eigenfunctions (structures). The eigenfunctions
form a complete orthogonal set, which means that the random field
can be reconstructed. The coefficients are uncorrelated and their
mean square values are the eigenvalues themselves. The kernel can
be expanded in a uniformly and absolutely convergent series of the
eigenfunctions and the turbulent kinetic energy is the sum of the
eigenvalues. Thus, every structure makes an independent contribu-
tion to the kinetic energy and Reynolds stress. The most significant
point of the decomposition is perhaps the fact that the convergence
of the representation is optimally fast since the coefficients of the
expansion have been maximized in a mean square sense. The mean
square of the first coefficient is as large as possible, the second is
the largest in the remainder of the series once the first term has
been subtracted, etc. We have described here the simplest case, that
of a completely inhomogeneous, square-integrable, field. If the ran-
dom field is homogeneous in one or more directions, the spectrum of
the eigenvalues becomes continuous, and the eigenfunctions become
Fourier modes, so that the proper orthogonal decomposition reduces
to the harmonic orthogonal decomposition in those directions. See
Lumley (1967, 1970, 1981) for more details.
The flow of interest here is three dimensional, approximately ho-
mogeneous in the stream wise direction ( x \ ) and span wise direction
(23), approximately stationary in time (/), inhomogeneous and of
integrable energy in the normal direction (x^).
We want a three dimensional decomposition which can be sub-
stituted in the Navier-Stokes equations in order to recover the phase
information carried by the coefficients. We have to decide which
variable we want to keep. Time is a good candidate since we are
particularly interested in the temporal dynamics of the structures.
Such a decomposition is possible and we do not need a separation of
variables in the eigenfunctions of the type <f>(x,u) = A(ui)tl>(x} (as
suggested by Glauser et a/., 1985) if we do not use any decomposi-
tion in time and choose the appropriate autocorrelation tensor. The
idea is to measure the two velocities at the same time and determine
< Ui(xi, £2, £3, t~)uj(x'i, x'2, £3,2) >= Rij. Since the flow is quasista-
36 J. L. Lumley et al.

tionary, Rij does not depend on time and nor do the eigenvalues and
eigenfunctions. The information in time is carried by the coefficients
a(") which are still "stochastic," but now evolve under the constraint
of the equations of motion. Thus the decomposition becomes

and we have to solve equation (3.3.2) for each pair of wave numbers
(&i, £3). 4>ij now denotes the Fourier transform of the autocorrelation
tensor in the #1,3:3 directions.
Our second change to the decomposition is a transformation of
the Fourier integral into a Fourier series, assuming that the flow is
periodic in the x^ and x3 directions. The periods L\,Lz are de-
termined by the first non zero wave numbers chosen. Finally, each
component of the velocity field can be expanded as the triple sum

In this case, a "structure" is denned by:

and the entire velocity field is recovered by the sum of all the struc-
tures (over n).
The candidate flow we are investigating is the wall region (which
reaches x 2 + = 40) of a pipe flow with almost pure glycerine (98%) as
the working fluid (Herzog, 1986). The Reynolds number based on the
centerline mean velocity and the diameter of the pipe is 8750. The
corresponding Reynolds number based on the shear velocity UT is 531.
From this data the autocorrelation tensor at zero time lag (t — t' = 0)
between the two velocities, RIJ(X\ — x ' l , X 2 , x ' 2 , X 3 — x'3)t_ti-0, was ob-
tained and the spatial eigenfunctions were extracted by numerical
solution of the eigenvalue problem. The results show that approxi-
mately 60% of the total kinetic energy and Reynolds stress is con-
tained in the first eigenmode and that the first three eigenmodes
capture essentially the entire flow field as far as these statistics are
Fundamental Aspects 37

concerned. This very fast convergence of the decomposition in the


near wall region is in good agreement with Moin's results (1984).
From a large eddy simulation data base, Moin uses the proper or-
thogonal decomposition successively in one and two dimensions in
the wall region (up to :r2+ = 65). His first structure contains 60%
of the total kinetic energy and 120% of the Reynolds stress (this
apparent paradox occurs because the contribution of higher order
structures to the Reynolds stress is negative). Ninety percent of the
kinetic energy is captured by the first three terms.

3.4 Low-Dimensional Models Constructed Using


the POD

We decompose the velocity - or the pressure - into the mean


(defined using a spatial average) and fluctuation in the usual way.
We substitute this decomposition into the Navier-Stokes equations.
Taking the spatial average of these equations we obtain, in the quasi
stationary case, an approximate relation between the divergence of
the Reynolds stress and the mean pressure and velocity.

This is substituted in the Navier-Stokes equations, giving an equation


for the fluctuating velocity. Equation (3.4.1) may be solved to give
the mean veloctity U in terms of the Reynolds stress < u\u^ >
in a channel flow in a manner which gives some feedback to the
system of equations as the fluctuation varies. We will see that this
feedback is necessarily stabilizing for the first structure (according to
the experimental results) and increases as the Reynolds stress gets
stronger. In other words, this term controls the intensity of the rolls,
by reducing the mean velocity gradient as the rolls intensify, thus
weakening the source of energy.
The expansion of the Fourier transform U{ of the fluctuating ve-
locity Ui, defined by

is achieved by use of the complete set of eigenfunctions <fi(n> 's in an


infinite sum:
38 J. L. Lumley et al.

Since we want to truncate this sum, we use a Galerkin projection


which minimizes the error due to the truncation and yields a set
of ordinary differential equations for the coefficients. After taking
the Fourier transform of the Navier-Stokes equations Ni = 0 and
introducing the truncated expansion, we apply Galerkin projection
by taking the inner product

Finally we obtain a set of ordinary differential equations of the form:

where A and B are matrices. Here A is the identity matrix (since


the complete set of eigenfunctions is orthogonal) and N.L. are non
linear terms. The non linear terms are of two sorts: quadratic and
cubic. The quadratic terms come from the non linear fluctuation-
fluctuation interactions and represent energy transfer between the
different eigenmodes and Fourier modes. Their signs vary. The role
of the Reynolds stresses < U{Uj > on these terms should be men-
tioned. They vanish for all wave number pairs except for (ki,ks) =
(0,0) for which they exactly cancel the quadratic term. Therefore
they prevent this mode from having any kind of quadratic interac-
tions with other Fourier modes. Since the cubic terms are zero too,
the (0,0) mode just decays by action of viscosity and does not par-
ticipate in the dynamics of the system.
The cubic terms come from the mean velocity-fluctuation inter-
action corresponding to the Reynolds stress < u\u?. > in the mean
velocity equation (the other part of this equation leads to a linear
term). Since the streamwise and normal components of the first
eigenfunction have opposite signs, they make a positive contribution
to the turbulence production and hence provide negative cubic terms
which are thus stabilizing. We remark that this is not necessarily the
case for higher-order eigenfunctions.
By use of the continuity equation and the boundary conditions

and
Fundamental Aspects 39

it can be seen by integration by parts that the pressure term would


disappear if the domain of integration covered the entire flow volume.
Since this is not the case (rather the domain is limited to X£ — 40),
there remains the value of the pressure term at the upper edge Xi
of the integration domain which represents an external perturbation
coming from the outer flow.
The exact form of the equations obtained from the decomposi-
tion, truncated at some cut-off point (&i c , k^c, n c ), does not account
for the energy transfer between the resolved (included) modes and
the unresolved smaller scales. The influence of the missing scales will
be parameterized by a simple generalization of the Heisenberg spec-
tral model in homogeneous turbulence. Such a model is fairly crude,
but we feel that its details will have little influence on the behavior
of the energy-containing scales, just as the details of a sub-grid scale
model have relatively little influence on the behavior of the resolved
scales in a large eddy simulation. This is a sort of St. Venant's
principle, admittedly unproved here, but amply demonstrated ex-
perimentally by the universal nature of the energy containing scales
in turbulence in diverse media having different fine structures and
dissipation mechanisms (see Tennekes and Lumley, 1972 for a fuller
discussion). The only important parameter is the amount of energy
absorbed.
We begin by defining a moving spatial filter which removes from
the total field the unresolved modes. The details of the definition
are not important - it is sufficient to conceive of the possibility of
such a filter. This filter is also an averaging operator. The velocity
field may now be divided into the resolved and unresolved field by
using this filter. The Reynolds stress of the unresolved field may now
be defined as the average using our filter operator, of the product of
the unresolved velocities; this acts on the resolved field. We suppose
that the deviator of this Reynolds stress is proportional to the strain
rate of the resolved field. We neglect the Leonard stresses, which
essentially supposes that there is more of a spectral gap than really
exists. This is what is done in the Heisenberg model, without ill
effect. The way in which we are treating the effect of the unresolved
modes on the resolved ones is very much like what is done in large
eddy simulation, and is called sub-grid scale modeling; our model
would probably be called a Smagorinsky model (there are minor dif-
ferences in the definition of the equivalent transport coefficient). Let
us agree to designate the resolved field as ut-< and the unresolved
40 J. L. Lumley et aL

field as tij>, while an average of <j>, say, over the unresolved modes
(the filtering process) can be designated as < d> >^. Thus:

with

and

Here < denotes the sum over all the modes (ki,k3,n) such that
k\ < kic,ks < &3C, n < nc and > denotes the sum over all the
modes (ki,kz,n) such that k\ > k\c or ^3 > k%c or n > n c , where
(k\c, ksc, n c ) is the cut-off mode. The characteristic scales of the
parameter Vf are those of the higher modes. We have introduced
an explicit dimensionless parameter ai, and will exclude adjustable
constants from i>x- By observation that the energy decreases rapidly
with increasing n and fc, we assume that these relevant scales are
given by characteristic scales of the first neglected modes. This is
probably a good approximation as far as the eigenmodes are con-
cerned since they are separated by large gaps in the spectrum and it
is a reasonable assumption for the Fourier modes since the steps of
our Fourier series are also large.
Finally, the parameter VT is taken equal to

(where u> and /> are characteristic scales of the neglected modes).
This can be expressed in terms of the eigenvalues and eigenfunctions
of the first neglected modes in the three directions (see figure 3.4.1).
We will refer to «i as a Heisenberg parameter. We will adjust a\
upward and downward to simulate greater and smaller energy loss
to the unresolved modes, corresponding to the presence of a greater
or smaller intensity of smaller scale turbulence in the neighborhood
of the wall. This might correspond, for example, to the environment
just before or just after a bursting event, which produces a large
Fundamental Aspects 41

Figure 3.4.1. Inhomogeneous Heisenburg model applied to the spe-


cific truncation discussed in the text. Legend: • resolved modes, x
first neglected modes which are considered for the computation of
the characteristic scales of the Heisenberg model.

burst of small scale turbulence, which is then diffused to the outer


part of the layer.
A term — l/3£,-j(< Uk>Uk> >> — « Ufc>tifc> > > > ) appears in
the equation for the resolved field. This term could be combined
with the pressure term and would not have any dynamical effect if
the integration domain covered the entire flow volume. In our case,
it needs to be computed since, like the pressure term, it leads to a
term evaluated at X^- We assume that the deviation (on the resolved
scale) in the kinetic energy of the unresolved scales is proportional
42 J. L. Lumley et al.

to the rate of loss of energy by the resolved scales to the unresolved


scales. This pseudo-pressure term gives some quadratic feed-back.
The rate of loss of energy from the resolved scales to the unresolved
scales is 2aiVTSi]<Si:j<, so that a free parameter appears also in this
term. Because this approximation involves a further assumption,
and to give ourselves greater flexibility, we call this parameter a 2 ,
although in all work presented in this paper, we have set a.\ = a 2 -
Thus the Heisenberg model introduces two parameters in the sys-
tem of equations, one, ai, in the linear term, the other one, 0*2, in
the quadratic term. The equations therefore have the following form:

where L and L' represent the linear terms, Q the direct quadratic
terms, Q' the quadratic pseudo-pressure term and C the cubic terms
arising from the Reynolds stress.

3.5 Comparison with the Wall Region


This study has two conflicting requirements. On one hand we
wish to keep as few modes as possible in order to obtain a low-
dimensional system, permitting us to apply the techniques of dy-
namical system analysis. On the other hand, we would like to retain
at least a qualitatively correct dynamical representation of the turbu-
lence production phenomenon. A necessary condition for the second
requirement consists of including as much of the energy and Reynolds
stress as possible in our system. This is already satisfied in the inho-
mogeneous direction by the proper orthogonal decomposition itself,
since the first structure is the most energetic. Given the energetic
gap between the two first eigenfunctions (Figure 3.5.1), keeping only
one structure seems quite reasonable. The choice of wave numbers is
now of great importance, especially as far as the Reynolds stress is
concerned (an important part of the Reynolds stress is contained in
the higher modes). The best selection is probably the one for which
the cut-off modes correspond to the experimental measurement cut-
off. Since we retain only a few modes, the steps of our Fourier series
are large. However, provided that the periodic length is larger than
a few integral scales, the Fourier transform of the autocorrelation
tensor is unchanged and the eigenfunctions are still the same.
We now seek a minimum truncation. The experimental results
show that the ratio between the streamwise and spanwise character-
Fundamental Aspects 43

Figure 3.5.1. Convergence of the proper orthogonal decomposition


in the near-wall region (x^~ = 40) of a pipe flow according to experi-
mental data. Turbulent kinetic energy in the first three eigenmodes,
A( n )(n =1,2,3) function of: the spanwise wave number (from Herzog,
1986).

istic length scales is of order ten (see figure 3.5.1.). Our first approx-
imation therefore neglects streamwise variations. We need at least
three terms in the spanwise direction (see below). Thus the minimum
truncation consists of one eigenmode, one streamwise wave number
(&! = 0) and three (i.e. two active) spanwise wave numbers (0, k, 2k).
In this paper, up to six spanwise wavenumbers (0,..., 5&) will be con-
sidered. In this case k has the value 3 X 10~3 and the lengths of the
periodic box are L\ — Ly, — 333. Even the model having six spanwise
Fourier modes is still very crude, although the truncation seems to
be the one which contains an optimally large amount of the total
44 J. L. Lumley et al.

energy among those of the same dimension. The zero cut-off mode
in the streamwise direction in particular is a very rough approxima-
tion. Such a truncation causes a drop of the spanwise and normal
root mean square values of the velocity which is particularly signifi-
cant in the upper half of the layer. For this reason, we do not expect
our velocity field reconstructed without rescaling to have more than
qualitative significance.
As a preliminary approach, we study the set of equations for a
truncation limited to the first eigenfunction (n = p = q = r = 1),
the zero streamwise wavenumber fci = 0 and up to six spanwise
wavenumbers k$ = 0, k, ...5k, for a suitably chosen k.
When only the zero streamwise wavenumber is considered, the
equations become much simpler. Indeed, because of the symme-
tries of the eigenfunctions (Herzog, 1986) in the (&i, £3) wavenumber
plane, the first and second components are purely real and the third
component is imaginary on the k<3 - axis (i.e. for k\ — 0). Using these
symmetries, the equations for the complex modal coefficients e41=0 ks
can be readily derived. Letting ^3 take the values jk,j = 0,1, ...,5
and writing a^. '_0 k __-k = Xj+i j/j, a typical equation takes the form:

In these equations the zero Fourier mode a(l)oo = XQ + iyo has


been removed since its imaginary part is identically zero and its real
part decays to zero under the influence of the viscosity (dxo/dt =
aox0;ao < 0). The coefficients of the quadratic and cubic terms are
each the sums of two quantities: c^^-k'i Ck-k',k and d k t k ' , d k , - k ' - The
influence of the Heisenberg model appears in the linear and quadratic
terms as follows:

where a\ and 0*2 are the proportionality (Heisenberg) coefficients


already mentioned. The computation of the nondimensionalized
transport coefficient i>x for this particular truncation gives the value
I/Rex = 6.28.
Fundamental Aspects 45

Numerical integrations of 3, 4, 5 and 6 mode models have been


carried out, but we shall only report in detail on the 6 mode (5 ac-
tive mode) simulations here. We summarize here the behavior of the
system. For a > 1.61 a unique circle of globally attracting stable
fixed points exists in the 2/4 subspace (with a window of instabil-
ity between 2.0 and 2.3): the /l, 7*3,7-5 components, along with the
trivial TQ component, are zero. As a increases, the magnitude of the
r-2 and r 4 components decreases until, at a ~ 2.409, this non-trivial
circle of fixed points coalesces with the origin, which for a > 2.409
is the unique and globally attracting fixed point for the problem.
For 1.37 < a < 1.61, an 51-symmetric family of globally attracting
double homoclinic cycles F exists, connecting pairs of saddle points
which are TT out of phase with respect to their second (x^^y^) com-
ponents. The points r + , r ~ discussed above are typical members of
this family. The existence of the cycles F implies that, after a rela-
tively brief and possibly chaotic transient, almost all solutions enter
a tubular neighborhood of F and thereafter follow it more and more
closely. As they approach F, the duration of the "laminar" phase of
behavior (in which r 2 and r 4 remain non-zero and almost constant
and ri^r3,r5 grow exponentially in an oscillatory fashion) increases
while the bursts (in which TI, 7-3 and r$ collapse) remain short. In an
ideal, unperturbed system, the laminar duration would grow with-
out bound, but small numerical perturbations, such as truncation
errors, presumably prevent this occurring in our numerical simula-
tions. More significantly, the pressure perturbation will limit the
growth of the laminar periods. Thus, there is an effective maximum
duration of events, which is reduced as a is decreased from the crit-
ical value aj ~ 1.61.
For the present study the intermittent behavior exhibited by the
six mode model for a between 1.35 and 1.61 appears to be of great-
est interest, since it corresponds in a fairly clear way to the physical
instability, sweep and ejection event observed in boundary layer ex-
periments. We now describe this intermittent behavior, starting with
some general remarks.
In the theory of dynamical systems, three types of intermittency
have been distinguished (Pomeau and Manneville, 1980). They are
associated with solutions repeatedly passing close to a weakly unsta-
ble fixed point or periodic orbit. The solution spends a long "lami-
nar" phase near the point or orbit until it reaches a critical amplitude
and a brief turbulent "burst" ensues, in which it travels far and fast
46 J. L. Lumley et al.

in phase space before returning. (These terms were appropriated by


the dynamical systems community and have been used in a predomi-
nantly metaphorical fashion thus far). The laminar phase is governed
by the linearized dynamics near the fixed point or periodic orbit, but
the burst and return are associated with a "global reinjection mecha-
nism," usually a homoclinic orbit or heteroclinic cycle (Tresser et al.
(1980), Silnikov (1965, 1968, 1970), Tresser (1984), Sparrow (1982),
Guckenheimer and Holmes (1983, §6.5).)
While evidence of intermittency has been detected before in fluid
systems and models (Pomeau and Manneville (1980), Berge et al.
(1980), Maurer and Libchaber (1980), Dubois et al, 1983) there has
been little evidence of type II, associated with a subcritical Hopf bi-
furcation, which is what we observe in the present model at a = 1.61.
We will, therefore, explore the connection between this dynamical
phenomenon and its analogue in the turbulent boundary layer in
some detail. Our model displays a "regular" intermittency, in con-
trast to the chaotic intermittency of Pomeau and Manneville (1980),
in which event durations are distributed randomly.
We describe here the reconstructions of the velocity field from
the expansion, using the computed values of xt- and y,-. Probably the
most interesting sets of solutions are those exhibiting intermittency,
obtained for 1.3 < a < 1.61. In the flow field, the rapid event which
follows the slowly growing oscillation and the repetition of the process
reminds one of the bursting events experimentally observed (Kline et
al. (1967), Corino and Brodkey (1969), etc.). For that reason we call
it a "burst." We will analyze its effect on the streamwise vortices.
In Figure 3.5.3 we show an enlargement of the time histories of
the modal coefficients during a burst for a = 1.4 (cf. Figure 3.5.2).
A description of the motion of the eddies during the burst is given in
Figure 3.5.4 for a •= 1.4 by plotting u? and 113 at the different times
indicated on Figure 3.5.3. Before and after the event, two pairs of
streamwise vortices are present in the periodic box, a structure very
similar to that previously obtained for 1.61 < a < 2.41. Nonetheless,
pictures 1 and 14 are shifted in the span wise direction by 333+/4,
corresponding to the phase shift A#2 = T, A#4 = 2?r. The bursting
event leads to variations of positions and amplitudes of the basic
streamwise rolls and formation of other vortices. The oscillation,
death and rebirth of vortices make the streak spacing vary, in agree-
ment with experimental results (Kline et a/., 1967). However, since
the intermittent solution is always very close to the real subspace or
Fundamental Aspects 47

a rotation of it, the vortices remain symmetric and paired. Moreover


it is possible to adjust the value of the viscosity parameter (a ~ 1.5)
so that the bursting period is 100 wall units as experimentally ob-
served. It is found that, in this case, the "burst" lasts 10 wall units
which is also the right order of magnitude.

Figure 3.5.2. Time histories of the real (x;) and imaginary (y,-) parts
of the coefficients for a value of the Hewisenberg parameter a = 1.4

The behavior of the eddies corresponding to a chaotic solution


(1.0 < a < 1.3) is shown in Figure 3.5.5, as before, by plotting
the u2, u3 velocity components at specific times. The behavior is
less regular and isolated vortices sometimes emerge. This is consis-
tent with the flow visualization experiments of Smith and Schwarz
(1983) who observed a significant number of solitary vortices among
the predominant vortex pairs. These patterns are also very rich in
48 J. L. Lumley et al.

Figure 3.5.3. Intermittent solution corresponding to an Heiseberg


parameter a = 1.4. Time history of the real and imaginary parts of
the model coefficients Xj, y; ( i =1 1,2,.. .5).

dynamics. We intend to study the regime further in future work.


Due to the truncation of our model to a single eigenmode and
one streamwise Fourier component, the absolute and relative energy
of the various velocity components is affected. First, the choice of
streamwise length scale affects the absolute values of all the compo-
nent energies, but not their relative values. We chose a streamwise
length scale of 333. Keeping only the Fourier mode at zero stream-
wise wavenumber, this is equivalent to supposing that the distribu-
tion of the first eigenvalue with wavenumber is flat to 333, and drops
to zero. Comparison with the true distribution of the first eigenvalue
with streamwise wavenumber indicates that this should overestimate
the energy in this eigenvalue. A more reasonable length scale would
be between 400 and 500, a value determined by the requirement that
the area between it and the origin is the same as the true area under
Fundamental Aspects 49

Figure 3.5.4. (left) Intermittent solution corresponding to an Heisen-


berg a = 1.4. Vector representation of u.2, 113 in an X2-X3 plane at
times indicated on figure 3.5.3.
Figure 3.5.5. (right) Chaotic solution for Heisenberg parameter a =
1.2. Vector representation of U2, us in an X2-xs plane at uniformly
spaced times (close enough to resolve the motion).

the distribution of the first eigenvalue with streamwise wavenumber.


However, examination of the equations indicates that changing the
value of LI leaves the equations invariant. Hence all phenomena re-
ported here would be unchanged, with simply a decrease in energy
level of all the velocity components. This would decrease the over-
estimated streamwise energy contained in our system but would not
necessarily improve our model since it would also decrease the energy,
50 J. L. Lumley et al.

already deficient, in the normal and cross-stream components.


The truncation also affects the relative energy in the various ve-
locity components. We have examined the distribution with x 2 + of
the energy in the three velocity components for truncations of 3, 6,
and 17 cross-stream wavenumbers, 1, 2 and 7 stream wise wavenum-
bers, and 1 and 3 eigenmodes. It is clear that the energy in the u^
and 11,3 components is relatively low (compared to the u\ component)
when only one streamwise wavenumber is included. This is partic-
ularly true in the upper part of the layer. While inclusion of more
than 6 cross-stream wavenumbers has no effect, addition of stream-
wise wave numbers does make a difference. In particular, addition of
another streamwise wavenumber helps considerably. We must con-
clude that there is energy in cross-stream and normal motions of
higher streamwise wavenumbers, that makes little relative contri-
bution to the streamwise velocity component. As a consequence,
our predicted rolls are somewhat weak compared to reality, and we
rescaled them. This rescaling will not be necessary in future work,
when we will add streamwise wavenumbers.
One naturally asks if the addition of streamwise modes will dras-
tically change the behavior of our model. In this respect, we observe
that the same symmetries, inherited from invariance of the Navier-
Stokes equations to spanwise translation and reflection, and stream-
wise translation, operate at any order of truncation. In particular,
this implies that the subspace spanned by basis functions with zero
streamwise wavenumber is invariant and that the coefficients of the
equations restricted to this subspace remain unchanged as stream-
wise components are added. Thus, the model studied in this paper,
together with its dynamics, exists unchanged, as a subsystem in the
larger set of differential equations, much as the 2-4 subsystem exists
in the present model. This raises the interesting possibility that the
2-4 branch may be destabilized by streamwise modes as well as by
the 1, 3, 5 "intersticial" spanwise Fourier modes.
We do not have present in our modal population a mechanism to
represent the production of higher wavenumber energy when an in-
tense updraft is formed, presumably as a result of a secondary insta-
bility. Thus, although our eddies are capable of exhibiting the basic
bursting and ejection process, the labor is in vain - there is no sequel,
no production of intense higher wavenumber turbulence. A contri-
bution is made only to the low wavenumber part of the streamwise
fluctuating velocity and the Reynolds stress. We could easily sim-
Fundamental Aspects 51

ulate the production of this high wavermmber turbulence, although


we do not expect its inclusion to change the dynamics of our system
qualitatively. However, we have held the value of our Heisenberg pa-
rameter constant, whereas its value should rise and fall with the level
of this intense higher wavenumber turbulence in the vicinity of the
wall. The transport effectiveness of the intense higher wavenumber
turbulence would be expected to damp the system, suppressing the
interesting dynamics until the higher levels are either blown down-
stream or lifted and diffused to the outer edge of the boundary layer.
The production of higher values of the Heisenberg parameter could
easily be parameterized by a single time constant first order equa-
tion tied to a measure of the amplitude of the coefficients. The
major effect would simply be to cut off each burst more rapidly than
at present. We plan to study the behavior of such a model in the
future.
The occurrence of the bursts for values of the Heisenberg pa-
rameter between 1.4 and 1.6 appears to be pseudo-random. This is
essentially due to round-off error in the calculations. The transitions
from one solution to the other and back are extremely sensitive to the
precise solution trajectory, and a minute change can make a consider-
able change in the time at which the next transition occurs. Initially
we did not exercise the pressure term. Recall that the pressure term
appeared due to the finite domain of integration. It represents the
interaction of the part of the eddy that we have resolved with the
part above the domain of integration, which is unresolved. The or-
der of magnitude that we estimated for this term was small, and
for that reason we at first neglected it. It has, however, an impor-
tant effect, while not changing the qualitative nature of the solution.
The term has the form of a random function of time, with a small
amplitude. This slightly perturbs the solution trajectory constantly;
away from the points r+ and r- this has little effect, but when the
solution trajectory is very close to these points, the perturbation has
the effect of throwing the solution away from the fixed point, so that
it need not wait long to spiral outward. This results in a thorough
randomization of the transition time from one solution to the other,
while having little effect on the structure of the solution during a
burst. While in the absence of the pressure term (and round-off er-
ror), the interburst time tends to lengthen as the solution trajectory
is attracted closer and closer to the heteroclinic cycle (this effect is
visible in figure 3.5.2), with the pressure term, the mean time stabi-
52 J. L. Lumley et al.

lizes (see figure 3.5.6). The mean interburst times for various values
of the pressure signal are shown in figure 3.5.7. Probably the most
significant finding of this work is the identification of the etiology of
the bursting phenomenon. That is, the bursts appear to be produced
autonomously by the wall region, but to be triggered by pressure sig-
nals from the outer layer. Whether the bursting period scales with
inner or outer variables has been a controversy in the turbulence lit-
erature for a number of years. The matter has been obscured by the
fact that the experimental evidence has been measured in boundary
layers with fairly low Reynolds numbers lying in a narrow range, so
that it is not really possible to distinguish between the two types of
scaling. The turbulent polymer drag reduction literature is particu-
larly instructive, however, since the sizes of the large eddies, and the
bursting period, all change scale with the introduction of the polymer
(see, for example, Kubo and Lumley, 1980, Lumley and Kubo, 1984).
The present work indicates clearly that the wall region is capable of
producing bursts autonomously, but the timing is determined by trig-
ger signals from the outer layer. This suggests that events during a
burst should scale unambiguously with wall variables. Time between
bursts will have a more complex scaling, since it is dependent on the
first occurrence of a large enough pressure signal long enough after
a previous burst; "long enough" is determined by wall variables, but
the pressure signal should scale with outer variables.

3.6 Generation of Eigenfunctions from Stability


Arguments
Ideally, one would like to apply the POD approach to a wide range
of flows where coherent structures are known to play an important
role in the dynamics. The POD procedure, however, requires the
two-point velocity autocorrelation tensor as input, thus necessitating
complete documentation of the flow before the analysis can proceed.
For flows with very high Reynolds numbers or complicated geome-
tries this can be prohibitively expensive given current computational
and experimental capablities. In this section we will describe an
analytic procedure for extracting basis functions (structures) which
approximate those given by the POD but which requires much less
a priori statistical information about the flow.
The method presented is based on energy stability considerations
put forth by Lumley (1971). First, the instantaneous flow field is de-
Fundamental Aspects 53

Figure 3.5.6. Pressure perturbations: solutions for a = 1.6 with


pressure term.

composed into three components in order to isolate the large scale


structures. Evolution equations can then be written for the coherent
velocity field and the coherent kinetic energy. A procedure can then
be formalized to search for the structures which maximize the in-
stantaneous growth rate of coherent energy, the rationale being that
the structures which on average have the largest growth rates will
compare well with the structures which contribute the most to the
average turbulent kinetic energy (POD eigenfunctions).
As an example we consider turbulent channel flow assumed sta-
tistically homogeneous in both the downstream (zj) and crosstream
(x 3 ) directions. In order to extract spatial structures from the total
velocity field, we avoid traditional Reynolds averaging and instead
decompose the instantaneous velocity field into three components:
the spatial mean (U), the coherent field (i;) and the incoherent back-
54 J. L. Lumley et al

Figure 3.5.7. Mean interburst duration T as a function of eigenvalue


A + and pressure signal amplitude £. Solid curves indicate prediction
of Aubry et al. (1988), open squares indicate results of numerical
simulations.

ground turbulence (it').

The spatial mean is an average over the x\ — 23 plane; we will indicate


it by [...]. We introduce a second averaging procedure, denoted by
< ... >, which eliminates the small-scale turbulence while leaving the
Fundamental Aspects 55

coherent field intact.

Practically this can be accomplished in several ways (Reynolds and


Hussain, 1972; Gatski and Liu, 1980; Liu, 1988; Brereton and Kodal,
1992; Berkooz, 1991). We will refer to this average as a phase aver-
age. For our purposes here it is sufficient that the phase average and
space average commute, and that the cross correlations be negligible.

Given these averaging procedures, we can manipulate the Navier-


Stokes equations to arrive at evolution equations for the coherent
velocity field.

where D/Dt denotes the mean convective derivative, Sij the mean
rate of strain, and v the kinematic viscosity. TJJ represents the rec-
tified effects of the small scale fluctuations on the coherent field and
is defined by

This can be thought of as a perturbed Reynolds stress, which is un-


known and will ultimately require modeling. In the limit of a com-
pletely random turbulence containing no structure (i.e. < ... >= 0)
this quantity is equal to the usual Reynolds stress. In the case when
the turbulence is completely structured so that < ... >= [...], T{J is
identically zero.
We now follow classical energy method stability analysis for the
coherent field. First, the growth rate of the volume averaged coherent
energy E is defined as a functional of the coherent velocity field.

Integration by parts and continuity are used to eliminate the nonlin-


ear convective and pressure terms. We seek the solenoidal velocity
field which maximizes A. Application of the calculus of variations
then gives the Euler equations for the maximizing v field in the form
of an eigenvalue relation.
56 J. L. Lumley et al.

We consider coherent fields which are periodic in the homogeneous


directions. This allows a decomposition into poloidal and toroidal
components which satisfy continuity exactly (Joseph, 1976)

The two scalar functions are then expanded in normal modes in the
streamwise and spanwise directions.

Substituting the above into equation (3.6.7) and eliminating the pres-
sure TT results in two coupled equations, forming a differential eigen-
value problem.
In order to precede we need to specify a mean velocity field and
a model for the unknown stress terms (see figure 3.6.1.).

Figure 3.6.1. Model inputs; (a) mean velocity and mean gradient;
(b) eddy viscosity and Reynolds stress.

We have investigated two different models for the unknown stress


terms appearing in the eigenvalue relation. It should be noted that,
modulo the modeled terms, equation (3.6.7) is linear in the coher-
ent velocities, providing an inexpensive means of determining basis
functions. This linearity is an essential advantage of the method and
for this reason we will constrain any stress model to be both linear
and homogeneous in the v field insuring that the governing equation
Fundamental Aspects 57

remains a regular eigenvalue problem. Tensorially this requires

The nature of the averaging procedure implies that the scales of the
coherent field and the background turbulence are different. Assum-
ing that the background turbulence evolves on much shorter time
and length scales then the structures, it seems plausible that a New-
tonian stress-strain relationship like that for the molecular stresses
will provide the basis for a model. We set

Due to the inhomogenity of the turbulence in the wall normal direc-


tion, we specify vt as a function of x?. corresponding to experimentally
determined values of the traditional eddy viscosity. We will refer to
this basic model as the isotropic eddy viscosity model.
Using the basic stress model and an analytic expression for the
fully turbulent mean profile (Reynolds and Tiederman, 1967) we have
solved the resulting equations numerically. Figure 3.6.2. shows com-
parisons between the calculated eigenvectors and the POD results
of Moin and Moser (1989) obtained from a numerical data base.
Although there are qualitative similiarities in the shape of the struc-
tures, the modes predicted by the stability method fall off much
more rapidly away from the wall than do the POD functions. The
eigenvalue spectrum clearly shows that the stability analysis favors
modes which have a much higher wavelength than the maximum
energy modes of the POD.
Although there may be a number of reasons for this discrepancy,
we choose to first examine more closely the closure model. Since both
the POD analysis and the stability method favor modes which are
infinitely long in the streamwise direction, we examine our equations
setting k\ = 0. We find that the isotropic eddy viscosity model
creates no coupling between the different components of the coherent
velocity. When there is no streamwise variation of the coherent field
the only coupling terms in the equations are those multiplying the
mean gradient. For realistic mean profiles, regions of high shear are
confined to thin regions near the wall, and the structures predicted
may be expected to fall off as quickly as the shear.
We now seek to develop a stress model that allows for some an-
isotropy in the eddy viscosity, and thus couples the component equa-
tions through the stress terms. We begin with the evolution equation
58 J. L. Lumley et al.

Figure 3.6.2. Isotropic eddy viscosity model: •, POD; , isotropic


model; (a) k3 = 6.00, (b) k3 = 9.00, (c) fc3 = 12.00, (d) k3 = 15.00.

for the Reynolds stresses, where we are obliged to model a number


of terms to obtain a closed system. We use standard second-order
turbulence models. We model the pressure-strain correlation by
a return-to-isotropy term and an isotropization-of-production term
(Naot et al., 1970); we use an isotropic dissipation. We assume the
stresses are in local equilibrium: D[uiUj]/Dt — 0. This reduces the
evolution equation for the Reynolds stress to an algebraic expression.
Now we set up a perturbation expansion in terms of mean field
quantities, taking the coherent field as an order e perturbation to
the spatial mean. On physical grounds, we argue that the perturbed
stress field is due entirely to the presence of the structures and conse-
quently we restrict the model to include only production due directly
to coherent velocity gradients. This is in agreement with a cascade
Fundamental Aspects 59

analogy for the complete flow: the coherent structures are fed energy
directly by the mean gradients while the small scale turbulence is in
turn fed by gradients of the coherent field. If we identify the Oth
order stresses with an eddy viscosity tensor, then the closure model
can be written as

where the tensor viscosity has the following structure in this specific
case: z/13 = 1/33 = 0, 1/33 = 1/22-
Despite the absence of mean production terms, this model is still
a major improvement over the isotropic eddy viscosity formulation.
In the simple model the effects of the mean field have been neglected
entirely. Here we have allowed for modulation of the perturbation
stresses by the mean field through the 0th order stresses appearing
in the production terms. Also we have unconstrained the model
in an important way since the tensorial form of the eddy viscosity
allows the pricipal axes of the stress tensor to be unaligned with
the axes of the rate of strain. This is more realistic considering the
three-dimensionality of the coherent field. This model leads to the
expected cross coupling of the equations through the stress terms.
Figure 3.6.3. shows eigensolutions for several values of k3. The
results compare well with the POD eigenvalues, especially for wave-
numbers at or below the peak in the POD spectrum. The improve-
ment with decreasing wavenumber is expected given the modeling
considerations. The separation of scales between the background
turbulence and the coherent structures increases as the wave number
decreases adding to the expected accuracy of the stress model. The
comparison of the two models indicates significant improvements in
the results given by the anisotropic eddy viscosity form. The energy
method procedure with the more refined closure model appears ca-
pable of extracting structures which closely approximate those given
by the POD at least at the energy containing scales of motion.
Despite the general improvement, it is still clear that more needs
to be done. From Figures 3.6.2. and 3.6.3., it is evident that the eigen
spectrum produced by solution of Equation (3.6.1),while improved
by the use of the anisotropic closure model, still predicts structures
with maximum growth rate that are a factor of 2 smaller than those
containing the most energy (as given by the POD). We next con-
sidered the effect on the spatial mean velocity field of the growing
coherent perturbation (see figure 3.6.4.).
60 J. L. Lumley et al.

Figure 3.6.3. Anisotropic eddy viscosity model: •, POD; iso-


tropic model; - - - anisotropic model: (a) k3 = 6.00, (b) £3 = 9.00,
(c) fc3 = 12.00, (d) k3 = 15.00.

At this point we consider the role of the mean velocity in the


two methods. The POD structures are derived from solutions to
the non-linear Navier-Stokes equations which allow for complicated
interaction between the different scales of motion. The structures
evolve in a mean velocity field that is changing due to the presence
of the structures themselves. Conditionally averaged mean profiles
clearly show the evolution of the local shear in the presence of co-
herent structures (see Figures 3.6.3, 3.6.4). We see that structures,
in the relatively long period before bursting, act to erode the shear
that they see. The POD eigenfunctions are given by averages of
contributions from different mean profiles.
The stability method on the other hand docs not allow for any
Fundamental Aspects 61

Figure 3.6.4. Comparison of the POD spectrum with the eigenspec-


trum of the stability problems.

interaction between the mean and the coherent field. The mean flow
is imposed and the resulting structures are calculated. The mean
profiles we have used are time averages which mask any contribution
from the coherent field. As such the stability analysis predicts that
the highest growth modes are those which can best extract energy
from the time averaged mean shear which is concentrated in the small
near-wall region. Since the structures have an aspect ratio of about
1, the narrow region of high imposed shear leads to a peak in the
eigen spectrum at a large wave number.
To allow the mean field to evolve under the influence of the coher-
ent field, we follow Liu (1988) and write time evolution equations for
the energy density of the coherent field. We allow the mean profile to
depend on the coherent velocity as it does in reality. We expect that
equilibrium solutions for the energy density as a function of cross
stream wave number will approximate the average energy content as
given by the POD spectrum.
We assume that the coherent field is given by the eigenvalues of
62 J. L. Lumley et al.

the stability problem, but we now allow them to vary in time

where V>i = vexp{ikx3}, i/>2 = ikil}exp{ikx3}, -03 = —Difjex.p{ikx3}.


By examining the evolution equation for r,-j =< M;UJ > — [t^Wj], the
forcing terms are of the form < UiUj > v^j. Consequently, we assume
the perturbation stresses also to be a product:

Since we have used an eddy viscosity in obtaining the coherent forms


WP fnrt.Vipr assume tlia.t'

where z/n = 1/22 = ^33 = ^r,^i2 = ^21 = Ar^is = "23 = 0. All that
remains is to model the mean profile. For this we adopt the quasi-
steady model used in Aubry et al. (1988). This allows the mean to
respond to growing structures providing the necessary feedback to
the evolving modes. Using the friction velocity, ur and the channel
half height, a, the scaled equation for the mean gradient is:

The rate of dissipation of turbulent energy is given by a simple model


adopted from second order closures schemes.

Substituting these various models into the energy equations results


in a set of three, coupled ODEs for the temporal evolution of the
energies and dissipation.
In order to evaluate the integrals appearing in these equations,
we need to assume the spatial form of the averaged turbulence quan-
tities Bij and the dissipation D(x-i). For this simple model we have
assumed that while the intensity of the turbulence varies its spa-
tial dependence remains unchanged. We use experimental data for
fully developed turbulent channel flow to determine both B and D.
The coherent structures are found by energy stability analysis, as
described above.
Fundamental Aspects 63

Figure 3.6.5. Temporal evolution of coherent energy density for pa-


rameter values corresponding to different wavenumbers.

We show in figure 3.6.5. the temporal evolution of the coher-


ent energy density for parameter values corresponding to different
wavenumbers.
In order to quantify comparisons of the single mode evolution
model and the POD eigenspectrum, we define an average of the life-
time of an individual structure. Taking A2(ti) = A 2 ( t f ) , where /,• is
the initial time and tt is the final time, we designate

where the integral is over t{ to fy, and T = tj — £,-.


Comparison between {Ai}(k^} and the spectrum of the POD is
shown in figure 3.6.6. The single mode model, while not capturing
the shape of the POD spectrum exactly, gives a good indication of
which wavenumbers are the most energetic. The discrepancy at high
wavenumbers is not surprising when we consider the simplifications
involved in the model. Our equations describe only the interactions
of a single mode with the local mean and the background turbulence.
The effects of interactions between modes may be negligible for the
64 J. L. Lumley et a].

large scale structures, but must become important for the smaller
scales. We will consider such interactions below.

Figure 3.6.6. Comparison of ensemble averaged energy content with


POD spectrum.

In order to estimate the effects of larger scale motions on the


evolution of the smaller scales, we examine the interaction between
a fundamental wave disturbance (small scale) and its subharmonic
(large scale). The coherent field is then made up of two components,
vi = v'j + vi'i, where the fundamental u" is periodic in x$ with a
wavelength A/2 and the subharmonic v[ has a wavelength A.
Since the periods of the disturbances have been artificially pre-
scribed, two phase averages can be denned to decompose the velocity
field analytically:

In this way the coherent contribution from the instantaneous velocity


is given by
Fundamental Aspects 65

and the fundamental can be separated from the subharmonic using


the « . » average:

The volume averaged kinetic energy equations for the two compo-
nents of the coherent field are:

As before, the coherent velocities are taken as the eigenfunctions of


the stability problem:

and the mean velocity model now contains contributions from both
the fundamental and the subharmonic:

New coupling terms appear in the evolution equations for the coher-
ent energy densities due to the interaction between the different size
modes:

/|2 and /I1 represent the effect on the mean shear production of
fundamental (subharmonic) coherent energy due to the presence of
the subharmonic (fundamental). /4 is a measure of the direct energy
transfer between the two modes due to the working of the sunhar-
monic stresses against the fundamental rate of strain. This quantity
appears with opposite sign in each equation.
As an example, we have calculated the interaction integrals using
the mode £3 = 20 as the fundamental. The results of integrating
equations (3.6.28) for these coefficient values are shown in figure
3.6.7. The mode-mode dynamics are dominated by terms due to the
mean velocity feedback model. The direct interaction term J4 is much
66 J. L. Lumley et al.

smaller. The effect of the large scale subharmonic on the behavior


of the small scale fundamental is dramatic. Although J^1 is twice
as large as I^, the slowly growing subharmonic mode is practically
unaffected by the presence of the fundamental. The fundamental,
however, is quickly damped by the larger scale motion. This agrees
with our intuitive picture of the physics. The small scale motions see
not only the mean shear, but the strain rate due to all larger scales.

Figure 3.6.7. Temporal evolution of coherent energy densities show-


ing the effect of mode-mode interaction.

While the situation analyzed here is admittedly artificial, since


a real turbulent flow will contain structures of all sizes at any given
instant, there is no reason not to expect a qualitatively similar effect
on the smaller scale motions in the real flow. Since the larger eddies
are relatively unaffected by the smaller scales, the arguments pre-
sented earlier and the small wavenumber results shown in the figures
will carry over to the case of manyn interacting eddies. The spurious
slow fall-off in the spectrum of {^42}, however, would be eliminated.
Fundamental Aspects 67

3.7 Wavelet Representations

All the flows of interest to us have one or more homogeneous di-


rections. We are accustomed to use in these directions the Fourier
transform, which is the homogeneous equivalent of the POD. How-
ever, the Fourier transform is not nearly so appropriate in the ho-
mogeneous case as is the POD in the inhomogeneous case. This is
because the Fourier modes are not confined to a neighborhood, but
extend to infinity without attenuation. All disturbances in a fluid,
and coherent structures in particular, are localized. There is there-
fore considerable motivation to find another representation that is
more appropriate.
In Tennekes and Lumley (1972) it was suggested that a more ap-
propriate quantity would be the energy surrounding a wavenumber
K, say from K/O, to UK, where a = 1.62. In physical space, this packet
with appropriate phase relations is confined to a region, essentially
dropping to zero in about lit/'K from the origin. Tennekes and Lum-
ley called these "eddies," but they are an example of what are now
called wavelets.
While wavelets appear to make more physical sense, we might
worry because we would be discarding the optimality of the Fourier
representation; would convergence be much slower, so that we would
need many more terms, or would we lose considerable energy if we
used the same number of terms? A main result of a recent paper
(Berkooz et al, 1992) is that very little energy is lost when using
a wavelet basis instead of a Fourier basis. Although wavelets are
physically appealing, it would also be nice to have reassurance from
calculations that physical behavior would be preserved in a wavelet
representation. To set our minds at rest on this point, Berkooz et
al. (1992) also display a relatively low-dimensional wavelet model of
the Kuramoto-Sivashinsky equation that shows dynamical behavior
similar to the full equation.
Without getting involved with mathematical details, an orthonor-
mal wavelet basis is constructed by starting with a function, say t^(x],
similar to the eddy suggested by Tennekes and Lumley (1972). From
this, construct a set •0j,o(;r) = V'( a; 2j), j = 1 , 2 , . . . . Each of these is
shrunk affinely, but is geometrically similar to the original function.
Now consider the translates of V^o : ^j,k(x} — ^j,o(x — k2j~),k =
1,2,...
Berkooz et al. (1992) consider a periodic, homogeneous stochastic
68 J. L. Lumley et si.

process. It is then obvious that the POD decomposition becomes


identical to the Fourier decomposition. Now if, for a given £, we
need N ( e ) POD modes in order to satisfy

then, if the {&;} are the coefficients in a wavelet basis, we get for
some constant C, depending only on the process (and not on e):

for some -V(e) > N(e) slightly bigger than -/V(e) (the precise state-
ment is given in Berkooz et al. 1992).

3.8 Dynamics with the Wavelet Representation in a


Simple Equation

Berkooz et al. (1992) wished to apply these ideas to a simple


situation. The three-dimensional, three component Navier-Stokes
equations are too complicated for a first effort.
The one-dimensional, scalar Kuramoto-Sivashinsky (K-S) equa-
tion appears in a variety of contexts, such as quasi-planar fronts,
chemical turbulence, etc . It shares some properties with Burgers'
equation and the Navier-Stokes equations, but is much easier to deal
with. UT> to some rescaling, this equation can be written as:

u periodic on [0, L], where L, the length of the spatial domain, is the
only free parameter in the problem.
Although the dynamical behavior for small values of L is fairly
well understood (see Hyman et a/., 1986 for an overview), many
open questions remain concerning the limit L —» oo (Zaleski, 1989;
Pomeau et al, 1984). As can be seen from the numerical simulations,
for L —> 30, a chaotic regime involving both space and time disorder
occurs (see Figure 3.8.1, where we plot a space-time representation
of a typical solution, L = 400, 0 < t < 100).
In order to check the estimate (3.7.2), we compare the energy
resolved by a given number of modes using either a Fourier (POD)
Fundamental Aspects 69

Figure 3.8.1. Space-time representation of a solution to the K-S


equation for L = 400, 0 < t < 100.

or wavelet basis. Note that, to compare the Fourier and wavelet


bases, all the translates in the wavelet basis must be considered.
Here are some results (see Fig. 3. 8. 2):

No. of modes wavelets (m — 6) wavelets (m = 8) Fourier


64(j = 6) 70.84% 71.5% 72.2%
96(j = 6,5) 79.1% 79.43% 83.3%
127(0 < j < 6) 84.1% 84.9% 89.7%
255(0 < j < 7) 99.9% 99.9% 99.9%
The scale j = 6 which captures most of the energy on the average,
70 J. L. Lumley et al.

Figure 3.8.2. Energy resolved by a given number of modes using a


Fourier or wavelet basis.

corresponds to a characteristic length 2~6i = Q"1, which is also the


length scale associated with the most unstable wavelength qm. In
agreement with the general shape of the energy spectrum, the scales
0 < j < 5 are shown to capture more energy than the scales in
the dissipative range (j > 7). The above figures show that (for
sufficiently smooth splines) the wavelet projection captures almost
the same amount of energy as the Fourier (= POD) decomposition
(within 5%).
These results prompted us to conclude that from an average en-
ergy point of view wavelets are a reasonable candidate for a modal de-
composition of the K-S equation. The localized nature of the wavelets
may give us a unique view of the spatial attributes of the coherent
Fundamental Aspects 71

structures. We outline our approach. We conjecture the existence


of a dynamically relevant length scale LC such that interactions be-
tween physical regions of distance greater than LC are dynamically
insignificant (a dynamical St. Venant principle). Determining the va-
lidity of this conjecture is part of our study. We use this conjecture
to remove terms in a wavelet-Galerkin projection that correspond to
interactions between regions of distance greater than LC- To study
whether the dynamics of coherent structures are indeed locally de-
termined we construct truncations corresponding to a small box size
LB in the larger box of size L.
We need to address the role of unresolved physical space (i.e.
modes located outside the box of size LB}- It is obvious that the
Dirichlet type of boundary condition imposed will create a bound-
ary layer which will affect the dynamics, especially in small boxes
which are of interest to us. There are two plausible approaches to re-
move this effect. One approach uses a stochastic boundary condition
(which is hard to implement numerically and treat analytically).
The other approach appeals to the conjecture on the existence of
LC- One takes a box of size LB greater that 2Lc so that one can
periodize the small model using resolved relatively distant modes
instead of unresolved ones. We opted for the second approach.
We present some preliminary results of the integration of one
such model. We resolved a box of size LB = 50 (this is 1/8 of the
original box). Figure 3.8.3 shows the spatio-temporal evolution of
the full system, with a Fourier basis. Figure 3.8.4 shows the spatio-
temporal evolution of a (rescaled) model with LC — 50 X 3/8, which
is in excellent qualitative agreement with the dynamics of the full
system. If LC is too small, after a long initial transient, the system
eventually settles down to a periodic oscillatory state, in which no
interaction between the localized structures is observed. This can
be avoided by an increase in LC, which adds significant non-linear
interactions between relatively distant physical space locations. It
might also be avoided by a stochastic boundary condition.

ACKNOWLEDGMENTS

Supported in part by Contract No. F49620-92-J-0287 jointly


funded by the U. S. Air Force Office of Scientific Research (Control
and Aerospace Programs), and the U. S. Office of Naval Research, in
72 J. L. Lumley et al.

Figure 3.8.3. (left) Spatio-temporal evolution of the full system using


a Fourier basis.
Figure 3.8.4. (right) Spatio-temporal evolution of a (rescaled) model
with Lc = 50 x 3/8.

part by Grant No. F49620-92-J-0038, funded by the U. S. Air Force


Office of Scientific Research (Aerospace Program), and in part by
the Physical Oceanography Programs of the U . S . National Science
Foundation (Contract No. OCE-901 7882) and the U. S. Office of
Naval Research (Grant No. N00014-92-J-1547).
Parts of these Notes previously appeared in: Aubrey, N., Holmes,
P.J., Lumley, J.L. and Stone, E. f990. The behavior of coherent
structures in the wall region by dynamical systems theory. In Near-
Wall Turbulence eds. S.J. Kline and N. H. Afgan, pp. 672-691.
Washington, DC: Hemisphere; Berkooz, G., Elezgaray, J., Holmes,
P., Lumley, J. and Poje, A. 1993 The proper orthogonal decompo-
sition, wavelets and modal approaches to the dynamics of coherent
structures. In Eddy Structure Identification in Free Turbulent Shear
Flows, ed. J. P. Bonnet and M. N. Glauser, pp. 295-310. Dordrecht
etc.: Kluwer; and Lumley, J. L. 1992. Some comments on turbulence.
The Physics of Fluids, A 4(2): 203-211.

4 REFERENCES

Aubry, N., Holmes, P., Lumley, J.L., and Stone, E., 1988. "The
dynamics of coherent structures in the wall region of a turbulent
boundary layer," J. Fluid Mech. 192: pp. 1-30.
Fundamental Aspects 73

Berge, P., Dubois, M. Manneville, P. and Pomeau, Y., 1980. "In-


termittency in Rayleigh-Benard convection," J. Phys. Lett. 41
L341.

Berkooz, G., 1991. "Turbulence, coherent structures and low di-


mensional models," Ph.D. Thesis, Cornell University.

Berkooz, G., Elezgaray, J. and Holmes, P., 1992. "Coherent struc-


tures in random media and wavelets," Physica D. 61 (1-4):
pp. 47-58.
Brereton, G.J. and Kodal, A., 1992. "A frequency domain filtering
technique for triple decomposition of unsteady turbulent flow,"
J. Fluids Engineering 114 (1): pp. 45-51.

Cantwell, B.J., 1981. "Organized motion in turbulent flow," Ann.


Rev. Fluid Mech. 13: pp. 457-515.

Corino, E.R., and Brodkey, R.S., 1969. "A visual investigation


of the wall region in turbulent flow," J. Fluid Mech. 37 (1):
pp. 1-30.

Comte-Bellot, G. and Corrsin, S., 1966 "The use of a contraction


to improve the isotropy of grid-generated turbulence," J. Fluid
Mech. 25: pp. 657-682.

Corrsin, S., 1975. Private communication.

Domaradzki, J.A., Rogallo, R.S. and Wray, A.A., 1990. "Interscale


energy transfer in numerically simulated homogeneous turbu-
lence," Proceedings CTR Summer Program, Palo Alto: Stan-
ford.

Dubois, M., Rubio, M.A. and Berge, 1983. "Experimental evidence


of intermittencies associated with a subharmonic bifurcation,"
Phys. Rev. Lett. 51, pp. 1446-1449.
Gatski, T.B. and Liu J.T.C., 1980. "On the interactions between
large-scale structure and fine-grained turbulence in a free-shear
flow III. A numerical solution," Phil. Trans. Roy. Soc. 293
(1403): pp. 473-509.
74 J. L. Lumley et aJ.

Glauser, M.N., Leib, S.J. and George W.K., 1985. "Coherent struc-
ture in the axisymmetric jet mixing layer," Proc. of the 5th
Symp. of the Turbulent Shear Flow Con/., Cornell University.
(Springer selected papers from TSF).

Guckenheimer, J. and Holmes, P.J., 1983. "Nonlinear oscillations,


dynamical systems and bifurcations of vector fields," Springer
-Verlag, New York. Corrected second printing, 1986.

Herzog, S., 1986. "The large scale structure in the near-wall region
of turbulent pipe flow," Ph.D. thesis, Cornell University.

Hyman, J.M., Nicolaenko, B. and Zaleski, S., 1986. "Order and


complexity in the Kuramoto Sivashinsky model of weakly tur-
bulent interfaces." Physica D, 18: p. 113.

Joseph, D.D., 1976. "Stability of Fluid Motion," Springer Tracts in


Natural Philosophy, Berlin.

Kline, S.J., Reynolds, W.C., Schraub, F.A. and Rundstadler, P.W.,


1967. "The structure of turbulent boundary layers," J. Fluid
Mech. 30 (4): pp. 741-773.

Kubo, I. and Lumley, J.L., 1980. "A study to assess the potential
for using long chain polymers dissolved in water to study tur-
bulence," Annual Report, NASA-Ames Grant No. NSG-2382,
Ithaca, NY: Cornell.

Kolmogorov, A.N., 1941. "Local structure of turbulence in an in-


compressible fluid at very high Reynolds numbers," Doklady
AN SSSR 30, pp. 299-303.

Kolmogorov, A.N., 1962. "A refinement of previous hypotheses


concerning the local structure of turbulence in viscous incom-
pressible fluid at high Reynolds number," J. Fluid Mech. 13
(l),pp. 82-85.

Lighthill, M. J., 1956. "Viscosity effects in sound waves of finite


amplitude," in Surveys in Mechanics, eds. G. K. Batchelor
and R. M. Davies, Cambridge, UK: The University Press, pp.
250-351.
Fundamental Aspects 75

Liu, J.T.C., 1988. "Contributions to the understanding of large


scale coherent structures in developing free turbulent shear
flows," Advances in Applied Mechanics, 26: pp. 183-309.

Loeve, M., 1955. "Probability Theory," Van Nostrand, New York.

Lumley, J.L., 1967. "The structure of inhomogeneous turbulent


flows," Atmospheric Turbulence and Radio Wave Propagation,
A. M. Yaglom and V. I. Tatarski, eds. p. 166. Moscow: Nauka.

Lumley, J.L., 1970. "Stochastic tools in turbulence," Academic


Press, New York.

Lumley, J.L., 1971. "Some comments on the energy method," De-


velopments in Mechanics 6, L.H.N. Lee and A.H. Szewczyk,
eds. Notre Dame, IN: N. D. Press.

Lumley, J.L., 1978. "Computational modeling of turbulent flows,"


Advances in Applied Mechanics 18, edited by C.S. Yih, p. 123,
New York: Academic Press.

Lumley, J.L., 1981. "Coherent structures in turbulence," Transition


and turbulence, edited by R.E. Meyer, Academic Press, New
York: pp. 215-242.

Lumley, J.L., 1992. "Some comments on turbulence," Phys. Fluids


A 4 (2): pp. 203-211.

Lumley, J.L. and Kubo, I., 1984. "Turbulent drag reduction by


polymer additives: a survey," The Influence of Polymer Addi-
tives on Velocity and Temperature Fields, IUTAM Symposium
Essen 1984, Ed. B. Gampert, pp. 3-21, Springer Berlin/Heidel-
berg.

Maurer, J. and Libchaber, A., 1980. "Effects of the Prandtl number


on the onset of turbulence in liquid helium," J. Phys. Paris
Lett. 41, L-515.

Meneveau, C., Lund, T.S. and Chasnov, J., 1992. "On the local
nature of the energy cascade," Proceedings of the Summer Pro-
gram, Center for Turbulence Research. Stanford/NASA Ames:
CTR.
76 J. L. Lumley et al.

Mom, P., 1984. "Probing turbulence via large eddy simulation,"


AIAA 22nd Aerospace Sciences Meeting.
Moin, P. and Moser, R.D., 1989. "Characteristic-eddy decomposi-
tion of turbulence in a channel," Journal of Fluid Mechanics
200: pp. 471-509.
Monin, A.S. and Yaglom, A.M., 1971. "Statistical fluid mechan-
ics: mechanics of turbulence," edited by J.L. Lumley, 1, Cam-
bridge, MA: MIT Press.
Naot, D., Shavit, A., and Wolfshtein, M., 1970. "Interaction be-
tween components of the turbulent velocity correlation tensor,"
Israel J. Tech., 8, 259.
Panchapakesan, N.R. and Lumley, J.L., 1993. "Turbulence mea-
surements in axisymmetric jets of air and helium, Part I: Air
Jet and Part II: Helium Jet." ,7. Fluid Mech. 246: pp. 197-223,
pp. 225-247.
Pomeau, Y. and Manneville, P., 1980. "Intermittent transition to
turbulence," Comm. Math. Phys. 74, pp. 189-197.
Pomeau, Y., Pumir, A. and Pelce, P., 1984. "Intrinsic stochasticity
with many degrees of freedom," /. Stat. Phys 37: pp. 39-49.
Reynolds, W.C. and Hussain, A.K.M.F., 1972. "The mechanics of
an organized wave in turbulent shear flow. Part 3," Theoretical
models and comparisons with experiment, J. Fluid Mech. 54:
pp. 263-287.
Reynolds, W.C. and Tiederman, W.G., 1967. "Stability of turbu-
lent channel flow with application to Malkus' theory," J. Fluid
Mech. 27: pp. 253-272.
Sarkar, S. & Lakshmanan, B., 1991. "Application of a Reynolds-
stress turbulence model to the compressible shear layer," AIAA
J. 29 (5): pp. 743-749.
Sarkar, S., 1992. "The pressure-dilatation correlation in compress-
ible flows," Phys. Fluids .A 4 (12): pp. 2674-2682.
Sarkar, S., Erlebacher, G., & Hussaini, M.Y., 1991. "Direct sim-
ulation of compressible turbulence in a shear flow," Theoret.
Comput. Fluid Dynamics 2: pp. 319-328.
Fundamental Aspects 77

Sarkar, S., Erlebacher, G., Hussaini, M.Y. & Kreiss, H.O., 1991.
"The analysis and modeling of dilatational terms in compress-
ible turbulence," J. Fluid Mech. 227: pp. 473-493.

Sherman, F.S., 1955. "A low-density wind tunnel study of shock-


wave structure and relaxation phenomena in gases," NACA TN
3298.

Shih, T.-H., Lumley, J.L. and Janicka, J., 1987. "Second order
modeling of a variable density mixing layer," J. Fluid Mech.
180: pp. 93-116.

Silnikov, L.P., 1965. "A case of the existence of a denumerable set


of periodic motions," Soviet Math. Dokl 6: pp. 163-166.

Silnikov, L.P., 1968. "On the generation of a periodic motion from


trajectories doubly asymptotic to an equilibrium State of Sad-
dle type," Math. U.S.S.R. SbornikG, pp. 427-438.

Silnikov, L.P., 1970. "A contribution to the problem of the structure


of an extended neighborhood of a rough equilibrium state of
saddle-focus type," Math U.S.S.R Sbornik 10 (1), pp. 91-102.

Smith, C.R. and Schwarz, S.P., 1983. "Observation of streamwise


rotation in the near-wall region of a turbulent boundary layer,"
Phys. Fluids 26 (3): pp. 641-652.

Sparrow, C.T., 1982. "The Lorenz equations: bifurcations, chaos


and strange attractors," Springer-Verlag, New York, Heidel-
berg, Berlin.

Tennekes, H. and Lumley, J.L., 1972. "A first course in turbulence,"


Cambridge, MA: MIT Press.

Tresser, C., 1984. "About some theorems by L.P. Silnikov," Ann


de L'Inst. H. Poincare, 40: pp. 441-461.
Tresser, C. Coullet, P. and Arneodo, A., 1980. "On the existence
of hysteresis in a transition to chaos after a single bifurcation,"
J. Physics (Paris) Lett. 41, L2 43-246.
Zaleski, S., 1989. "Stochastic model for the large scale dynamics of
some fluctuating interfaces," Physica D 34: p. 427
78 J. L. Lumley et al.

Zeman, O., 1990. "Dilatation dissipation: the concept and appli-


cation in modelling compressible mixing layers," Phys. Fluids
A. 2: pp. 178-188.
Zeman, 0., 1991. "On the decay of compressible isotropic turbu-
lence," Phys. Fluids A 3: pp. 951-955.
Zilberman, M., Wygnanski, I. , and Kaplan, R.E., 1977. "Transi-
tional boundary layer spot in a fully turbulent environment,"
Phys. Fluids Supp. 20 (10): S258-S271.

Zubair, L., Sreenivasan, K. R. and Wickerhauser, V. 1992. "Char-


acterization and compression of turbulent signals and images
using wavelet-packets," In The Lumley Symposium: Recent De-
velopments in Turbulence, edited by T. Gatski, S. Sarkar, C.
G. Speziale, Berlin: Springer, pp. 489-513.
Chapter 2

DIRECT NUMERICAL
SIMULATION OF
TURBULENT FLOWS

Anthony Leonard

1 INTRODUCTION
The numerical simulation of turbulent flows has a short history.
About 45 years ago von Neumann (1949) and Emmons (1949) pro-
posed an attack on the turbulence problem by numerical simulation.
But one could point to a beginning 20 years later when Deardorff
(1970) reported on a large-eddy simulation of turbulent channel flow
on a 24x20x14 mesh and a direct simulation of homogeneous, iso-
tropic turbulence was accomplished on a 323 mesh by Orszag and
Patterson (1972). Perhaps the arrival of the CDC 6600 triggered
these initial efforts. Since that time, a number of developments have
occurred along several fronts. Of course, faster computers with more
memory continue to become available and now, in 1994, 2563 sim-
ulations of homogeneous turbulence are relatively common with oc-
casional 5123 simulations being achieved on parallel supercomputers
(Chen et al., 1993) (Jimenez et a/., 1993). In addition, new algo-
rithms have been developed which extend or improve capabilities in
turbulence simulation. For example, spectral methods for the simu-
lation of arbitrary homogeneous flows and the efficient simulation of
wall-bounded flows have been available for some time for incompress-
ible flows and have recently been extended to compressible flows. In

79
80 A. Leonard

addition fast, viscous vortex methods and spectral element methods


are now becoming available, suitable for incompressible flow with
complex geometries. As a result of all these developments, the num-
ber of turbulence simulations has been increasing rapidly in the past
few years and will continue to do so. While limitations exist (Rey-
nolds, 1990; Hussaini et a/., 1990), the potential of the method will
lead to the simulation of a wide variety of turbulent flows.
In this chapter, we present examples of these new developments
and discuss prospects for future developments.

2 PROBLEM OF NUMERICAL SIMULATION

We consider an incompressible flow whose time evolution is given


by the Navier-Stokes equations for the velocity, u (x, t), and the
pressure, p (x, t) as

along with appropriate initial and boundary conditions. It is assumed


that the density = 1.
The character of the solution depends on the Reynolds number of
the flow, Re = UL/V, where U and L are a characteristic velocity and
a characteristic length of the large scales and v is the kinematic vis-
cosity. For small Reynolds numbers, one obtains a laminar flow that
is smoothly varying in space and time; for large Reynolds numbers,
one obtains a turbulent flow. Turbulent flows have been described as
random, chaotic, vortical, three-dimensional, and unsteady, and they
are known to contain a wide range of scales. It is the combination
of all these attributes that makes the numerical simulation of such
flows extremely challenging.
In turbulent pipe flow, for example, we estimate, according to
universal equilibrium theory (see, e.g. Batchelor, 1967), the smallest
important scale of turbulence to be proportional to the dissipation,
or Kolmogorov, length, rj = (i/ 3 /e) 1//4 , where e is the energy dissipa-
tion rate per unit mass, and the largest important scale to be some
multiple of the pipe diameter. Using the volume-aver aged s given by
Direct Numerical Simulation 81

where U is the mean velocity and D is the pipe diameter, we find


that,

where Re = UD/z/ and f is the friction factor,

and UT is the wall shear velocity given by

The friction factor, given implicitly by the formula (Hinze, 1975),

is only weakly dependent on Reynolds number so that the required


number of mesh points on a three-dimensional grid would be propor-
tional to (D/??)3 oc Re 3 ' 4 . Figure 1 shows the energy spectrum mea-
surements of Laufer (1954) for high-Reynolds-number (Re = 500,000)
pipe flow. The pipe diameter is 25.4 cm. The wave number corre-
sponding to the Kolmogorov length, k^ = 27r/?y, is seen to be well
beyond the measured data. To simulate reliably the dissipation of
turbulence energy, the grid spacing must be somewhat smaller than
the length scale corresponding to the peak in the dissipation spec-
trum. If isotropy of the small scales is assumed, the dissipation
spectrum is proportional to kfEi(kj). In Laufer's experiment this
peak, away from the wall, corresponds to a length of 150 77 or 0.03D
or k?7 = 0.04.
Figure 2 shows energy spectra in a high speed boundary layer
measured by Saddoughi and Veeravalli (1994). Note again that the
peak in the energy spectrum occurs at k?y well below 2?r, this time
near kij K 0.06. Thus we expect that a resolution of the fine scales
such that kmax7j = 1 should be sufficient and this, indeed, seems to
be the case (Huang and Leonard, 1994).
Therefore, as an estimate of the mean spacing between grid points
A, required in the direct, simulation of turbulent pipe flow, we take
A = 3r/. Table I gives corresponding estimates of the number of
mesh points required for several Reynolds numbers, assuming that
82 A. Leonard

Figure 1. Longitudinal energy spectra measured in pipe flow at Re


= 500,000, r' is the distance from the pipe wall. The pipe diameter,
D, is 25.4 cm (Laufer, 1954).

the computational domain extends 10 diameters in the streamwise di-


rection. (This estimate could be off by a factor of 3 either way. Some
measurements and their interpretation suggest correlation lengths of
20D, others correlation lengths of 2D; see Coles(1981).) It appears
that the two lowest Reynolds number cases would be accessible to
present day supercomputers (1010 floating point operations per sec-
ond, 109 words storage). In fact, Kim et al. (1987) previously per-
formed a direct simulation of plane channel flow at Re = 3300, based
on channel half width, using 4 X 106 grid points, roughly correspond-
ing to the Re = 5000 case for the pipe as given in Table I.
In addition it should be verified that the spacing A = 377 is
Direct Numerical Simulation 83

Figure 2. Longitudinal and transverse energy spectra measured in a


turbulent boundary layer at a momentum thickness Reynolds num-
ber, Re<? of 370,000 or a Taylor scale Reynolds number, RA of 1400.
Measurements taken at y + w 16,000. Solid lines are fits to the data
(Saddoughi, 1994).
84 A. Leonard

Table I. - Mesh-Point Requirements

Number of
Kolmogorov length mesh points
(A = 37?)
Reynolds n+
3
Number 7?/D wall^units N = 10^(ff
5 X 103 0.0045 1.6 3.1 X 106
1 X 104 0.0028 1.8 1.3 X 107
5 X 104 0.00093 2.4 3.6 x 108
1 X 105 0.00058 2.8 1.5 x 109
5 x 105 0.00019 3.8 4.2 x 1010

sufficiently small to allow resolution of all important turbulence phe-


nomena near the wall. The grid spacing measured in wall units is
given by

For Re from 5,000 to 500,000, A + ranges from 4.8 to 11.4 (see Table
I). This spacing would be marginally sufficient resolution to repro-
duce all important wall-layer structures (such as streamwise streaks),
which have characteristic lengths of 50-100 wall units with some
structures down to 20 units in size.
The number of time-steps, N s , required to follow one realization
for a time T and obtain reasonable statistics also depends on Rey-
nolds number. The time-step At is roughly limited to

Using the above estimate for A and (2/f) 1 ' 4 = 3 we find that
Direct Numerical Simulation 85

And if T = 100 D/U, then

or 6,500 steps for Re = 5,000.

3 SIMULATION OF HOMOGENEOUS
INCOMPRESSIBLE TURBULENCE

A variety of homogeneous turbulent flows can be treated by writ-


ing the velocity field u as the sum of a mean component and a tur-
bulent romnonfiTit.

and assuming that the components of U have the form

where repeated indices are summed (see Rogallo, 1981). By trans-


forming to coordinates, x, moving with the mean flow we obtain
momentum and continuity equations for u' that contain no explicit
dependence on x. Then assuming that the turbulent component of
the homogeneous flow is periodic in x-space, with period Lm in di-
rection m, no further boundary conditions are required, and spatial
derivatives can be computed accurately by Fourier interpolation.
Thus. u'(x, t) is represented by

Here the mth component of k is

and I ranges over -N/2 + 1 < I < N/2 - 1. The Navier-Stokes equa-
tions become a 3(N — I) 3 system of ordinary differential equations
(ODEs),

/here k 2 = k n k n , and
86 A. Leonard

(In the above, zero mean flow, Umn = 0, has been assumed in order
to simplify the presentation.) To avoid explicit evaluation of the
convolution sums u^un requiring 0(N 6 ) operations per step, fast
Fourier transforms (FFTs) are used to return to physical space where
the required products are formed and then transformed (by FFTs)
back to Fourier space. Consequently, only 0(N 3 logN) operations
per step are required.
Suppose the tensor U^m is decomposed into a symmetric (R) and
an antisymmetric (ft)tensor; then the only constraint on R is that it
be traceless Rnn = 0, but 17, related to the vorticity, must satisfy the
evolution equation

Besides zero mean flow, four examples are: plain strain, TJ22 =
—Uss = const; axisymmetric strain, 1)22 — Uss = —(l/2)Un =
const; shear, Uia = const; and rotation, TJis = -Uai = const. See
Rogallo (1981) and Rogallo and Moin (1984) for more discussion and
application.

4 WALL-BOUNDED AND INHOMOGENEOUS FLOWS

The direct simulation of wall-bounded and other inhomogeneous


flows presents a new set of difficulties. For wall-bounded flows, part
of the problem is due to the small-scale physical processes taking
place near the wall. The thin shear layer next to the wall is con-
tinually breaking up via three-dimensional (3-D) inertial instabilities
resulting in a violent 3-D wrinkling of the vortex layer. In the span-
wise direction, the scale of the breakup is of the order of the thickness
of the layer or tens of wall units and, perhaps, somewhat more in the
streamwise direction.
Another part of the problem is algorithmic and is due either to
the no-slip boundary condition at the wall or the presence of semi-
infinite domains of fluctuating potential flow in the case of free shear
flows or turbulent boundary layers. One can no longer use Fourier se-
ries for the spectral expansion in inhomogeneous directions. Rather,
to obtain rapid convergence independent of boundary constraints,
one should employ global polynomials related to the eigenfunctions
of a singular Sturm-Liouville problem (Orszag, 1980). Chebyshev
Direct Numerical Simulation 87

and Legendre polynomials are popular choices for channel flow but,
for example, in the case of pipe flow, other choices may be preferable
because of special conditions of the problem at hand. Whatever the
choice may be, this change in basis functions complicates the impo-
sition of, for example, the no-slip condition and the satisfaction of
the continuity constraint. These two conditions become global con-
straints on the expansion which are generally difficult or costly to im-
pose simultaneously. By contrast, in the simulation of homogeneous
flows using Fourier expansions in all three directions, the bound-
ary conditions are built into the expansion and the divergence-free
constraint is satisfied by a simple projection which is local in wave-
number space.
In the following, we will describe a technique for overcoming the
algorithmic difficulties described above, at least for flows in simple
geometries. The technique consists of expanding the velocity field
in terms of a set of divergence-free vector functions satisfying the
appropriate boundary conditions. First we write the Navier-Stokes
equations in rotational form:

Here o> — V x u '1S the vorticity. The boundary condition at a


wall is u|waii = 0. Other boundary conditions, such as periodicity or
freestream conditions at infinity, are imposed as appropriate.
The role of the pressure in incompressible flows is to enforce the
continuity condition. This may be expressed formally be recalling
that an arbitrary vector field f may be uniquely decomposed into a
sum of a divergence-free field satisfying tangency at the boundary
and the gradient of a potential.

Let <p be the projection operator that accomplishes this decomposi-


tion, that is.

applying the projection operator p to equation (18a) we obtain


88 A. Leonard

eliminating the dynamic pressure (Chorin and Marsden, 1979). The


above equation is the starting point for the numerical scheme de-
scribed below.
Vector expansion method:
We write u as the exoansion

where each if>n satisfies

and the homogeneous boundary condition on u. We need to derive


a system of ODEs for the coefficients an(t) (n = 1, 2, . . . , N). We
do so by substituting the expansion (22) into (18a) and taking the
inner product of the result with a set of weight vectors £m (m = 1,
2, . . . , N) satisfying

and

If the £m form a complete set and N -> oo, this operation is equiva-
lent to applying the projection operator because

for an arbitrary scalar field (f>. The result is the following system of
ODEs:

where

and gm is quadratic in the an s,

Thus, each evaluation of the a n 's requires the solution of a linear


system and the computation of the nonlinear term. The choice of the
Direct Numerical Simulation 89

vector functions -0n and £m is crucial to the success of the method.


Mathematically, of course, the sets {ifin} and {Cm} must be com-
plete in appropriate spaces of functions satisfying (23) plus bound-
ary conditions and (24) and (25), respectively. From a computation
standpoint we would like (i) rapid convergence of our expansions of
u; (ii) minimum (and ordered) coupling of the modes through the
time-dependent and viscous operators, (for example, banded struc-
ture for A and B); (iii) efficient construction of the matrices A and
B; and (iv) efficient computation of the nonlinear term g.
In practice, the index n ranges from one to three representing
each spatial direction. The matrices A and B are diagonal in the
homogeneous directions where Fourier series may be used but non-
diagonal in the direction normal to the wall. For example, in the
application to pipe flow (Leonard and Wray, 1983), A and B are
banded with the same number of bands. Thus, implicit treatment of
the viscous term is possible at no extra cost.
Equation (27) is a complete statement of the dynamics. No extra
equations are required to enforce boundary conditions or continuity,
and no fractional time-steps are needed. In addition, the number
of equivalent grid points in the computation is N/2. Thus, only
two unknowns per mesh point are required because of the constraint
(23) on expansion vectors, allowing considerable savings in memory
requirement.
As a relatively simple example, consider two-dimensional channel
flow with the velocity field assumed to be periodic in the streamwise
direction x with period L. The velocity u(x, y,t) is expanded as fol-
lows:

where k = 0, ±2?r/L, ±4?r/L,... and y is the coordinate perpendicular


to the channel walls located at y = ±1. The condition that each
expansion function be divergence free requires that

Boundary conditions and completeness of the expansion with rapid


convergence are satisfied if we choose for k ^ 0,

where I n (y) 1S the Lhebychev polynominal.


90 A. Leonard

The divergence condition (30) demands that

If k = 0 we choose

Table II shows the exponential convergence of the method when ap-


plied to the eigenvalue problem for the Orr-Sommerfeld equation.
The Table gives the real part of the time eigenvalue A using the
expansion functions for even v(y) : (1 - y 2 ) 2 T 2 n -2(y)?n = 1,2, ...,N.

Table II. Channel Flow Eigenvalues

R e = 10 4 ,k x = l,k z = 0

N Real (A)
15 .00372 . . . .
20 .0037398 . . . .
25 .003739669 . . .
30 .003739670616.
35 .0037396706227
40 .0037396706223
45 .0037396706216
50 .0037396706222

In treating vorticity-containing layers of finite thickness, bounded


by a semi-infinite domain of unsteady potential flow, it is desirable
to include, in the expansion, terms that represent the potential flow
explicitly. This is so because potential flow contributions, in general,
decay much more slowly, in the direction normal to the layer, than
the vortical contributions (Spalart et ai, 1991). Thus, for boundary
layer flows, for each (k x , kz) one adds an extra vector to the expansion
Direct Numerical Simulation 91

given by

such that

as y —> oo. The vortical expansion functions behave as

for large y, where y0 is on the order of the boundary layer thickness.


Two applications that incorporate the above technique for finite-
thickness vorticity layers are Spalart's (1989) simulations of turbu-
lent boundary layers and Moser and Rogers' (1991) study of plane
mixing layers. Other efforts that have used the general strategy
of divergence-free vector expansions include studies of transitional
pipe flow (Leonard and Wray, 1982) (Leonard and Reynolds, 1988),
curved channel flow (Moser et al, 1983), vortex rings (Stanaway et
al, 1988) and spherical couette flow (Dumas and Leonard, 1994).

5 FAST, VISCOUS VORTEX METHODS

Traditionally, vortex methods have been used to model unsteady,


high Reynolds number incompressible flow by representing the fluc-
tuating vorticity field with a few tens to a few thousand Langrangian
elements of vorticity. Now, with the advent of fast vortex algorithms,
bringing the operating count per timestep down to 0(N) from 0(N 2 )
for N computational elements, and recent developments for the accu-
rate treatment of viscous effects, one can use vortex methods for high-
resolution simulations of the Navier-Stokes equations. Their classical
advantages still hold — (1) computational elements are needed only
where the vorticity is nonzero, (2) the flow domain is grid-free, (3)
rigorous treatment of the boundary conditions at infinity is a nat-
ural byproduct, and (4) physical insights gained by dealing directly
with the vorticity field - so that vortex methods have become an
interesting alternative to finite difference and spectral methods for
incompressible turbulence simulation.
To describe vortex methods, we start with the three-dimensional
vorticity equation for the vorticity field w in a constant density, in-
compressible flow,
92 A. Leonard

Combining the incompressibility condition for the velocity field u, V


u = 0, and the definition of vorticity, y x u — u>, we find that

We will consider vortex methods in the context of bluff-body flows


and so we are interested in solutions to (37) and (38) corresponding
to no-slip at the surface of a rigid body moving with velocity Ub,

and free-stream conditions at infinity

To simplify the description of the method, we restrict the discussion


to non-rotating bodies. The solution to (38) satisfying (39) and (40)
is given in terms of the infinite-medium Green's function

In our numerical approach described below, it is important to recog-


nize that the no-slip boundary condition is maintained by a continual
flux of vorticity from the body surface into the fluid. Mathematically,
this flux is such that co is nonzero only in the fluid (i.e., external to
the body) and that (41) gives the result that

where t is any tangent vector at the body surface. That (41) and
(42) imply (39) may be shown as follows. Let -0, the stream function
for the imaginary fluid within the body, be given by

Thus, within the body B,

and
Direct Numerical Simulation 93

because all vorticity is external to B. Now we use Green's identity

with u = v = ib[. Here n is in the outward normal direction. From


(42) and (44),

at every point on the surface <9B for any tangent vector t. Thus,
~y~ — 0, so that (46) reduces to

Hence, -0' = const and

In the present method, the vorticity field is represented by a dense


collection of N moving, vector-valued computational elements or par-
ticles.

where the particle vector amplitudes, 0:^ , have units of circulation


times length and each particle has a radially symmetric spatial dis-
tribution defined by

where a is the effective core radius of the particle, and £ has unit
volume integral. See Leonard (1985) and Winckelmans and Leonard
(1993) for further discussion including choices for the function £. For
example, the distribution for the gaussian particle is given by

To satisfy the inviscid components of motion of the fluid, each par-


ticle moves with the local velocity
94 A. Leonard

and its vector a^ is stretched and rotated according to

The velocity field u is obtained by substituting (50) into (41) to


obtain

The function q is given by

For a variety of simple choices of the distribution function £ (e.g.,


gaussian) the truncation error is O(cr 2 ) if the typical interparticle
distance, h, satisfies h < <r. Higher order schemes are possible. See
Leonard (1980) and Beale and Majda (1982) for more details.
Although equation (54) for the inviscid evolution of the x^ seems
most natural, other possibilities, based on the fact that

also yield convergent, accurate schemes as discussed by Winckelmans


and Leonard (1993). For viscous diffusion we need to update the
vorticity field following the equation

with the boundary condition implied by (39). The random walk


method, a discussed by Chorin (1973), has been applied widely for
a number of years but has the disadvantage of being slowly conver-
gent. More recently, the technique of particle strength exchange (see
Degond and Mas-Gallic (1989)) has been proposed. This method has
good convergence properties but requires an occasional remeshing of
the particles as discussed below. In this algorithm we approximate
y 2 as an integral operator and discretize the integration over the
particles. Thus, we use
Direct Numerical Simulation 95

where, for example, with gaussian particle distributions,

Consider first the application to an infinite domain. The approxima-


tion (59) is good to second order in cr, as we can see by taking the
Fourier transform of both sides,

Using (59) and the particle representation, we find that (58) can be
written

Note now that by approximating the integral on the right hand side
of (62) over the particles and using (60) we obtain the evolution
equation for the problem (58) in an infinite domain,

where v is the particle volume.


The above algorithm conserves total vorticity exactly, that is,
E-^dctj/dt = 0. Furthermore, it is found that if (63) is integrated in
time with Euler timestepping and

the results are superaccurate (Pepin 1990, Leonard and Koumout-


sakos 1993). This phenomenon may be explained by noting that the
exact solution to (58) in an infinite domain at time t + At is given
by

For wall-bounded flows we enforce the no-slip condition in the form


(42). In addition, however, for a substep of the diffusion step we use
(63), that is we do not enforce du/dn — 0 at the wall during this
substep (Koumoutsakos et a/., 1994). This leads to an additional
96 A. Leonard

spatially varying flux of vorticity from the surface. This effect along
with the pressure gradient produces a wall slip (Usijp) at time t + At.
We compute this slip as an average over computational panels on
the wall directly from the Green's function integral for the velocity
potential, with a form analogous to (41), using the fast algorithm.
Thus the total flux to be emitted into the flow for the other substep
of the diffusion process is given by

This flux is then distributed to the particles by discretizing the


Greens' integral for the inhomogeneous Neumann problem for the
diffusion equation. See Koumoutsakos et al. (1994) and Winckel-
mans et al. (1995).
Rerneshing:
In order for the numerical simulation to be accurate, the particles
must, to a certain degree, be uniformly distributed. This is required
for accuracy in the convection step as well as the diffusion step.
On the other hand the local strain rate following a particle may
generate a substantial contraction or crowding of particles in one or
two directions accompanied by an expansion in the other directions,
similar to the situation at a hyperbolic stagnation point in steady
flow. When remeshing is deemed necessary, we overlay a regular
grid (the new particles) over the old particles, keeping the average
particle density constant, and interpolate the old vorticity onto the
new particle locations. We use a 27-point scheme to distribute the
old vorticity to the new mesh. Specifically, away from a wall, the
ith old vortex with circulation I\ and location (xi,yi,Zj) contributes
AFj ( i circulation to new mesh point (xj,yj,Zj), according to

where the interpolation kernel A is given by

If the old particle is less then a distance h from a wall the interpo-
lation is modified to maintain the same conservation properties. See
Direct Numerical Simulation 97

Koumoutsakos (1993) and Hockney and Eastwood (1981) for further


discussion. This scheme conserves the circulation, linear and angular
momentum.
Fast Vortex Methods:
The straightforward method of computing the right-hand sides of
(53) and (54), using (55) for every particle, requires O(N 2 ) operations
for N vortex elements. This precludes high resolution studies of bluff
body flows with more than say 50,000 elements.
Recently, fast methods (see e.g., Barnes and Hut (1986), Green-
gard and Rohkin (1987), and Salmon et al. (1994)) have been devel-
oped that have operation counts of 0(N log N) or 0(N) depending
on the details of the algorithm. The basic idea of these methods is to
decompose the element population spatially into clusters of particles
and build a hierarchy of clusters (tree) - smaller neighboring clus-
ters combine to form a cluster of the next size up in the hierarchy
and so on. Figure 3 shows an example of particle clustering in two
dimensions.
The contribution of a cluster of particles to the velocity of a
given vortex can then be computed to desired accuracy if the particle
is sufficiently far from the cluster in proportion to the size of the
cluster and a sufficiently large number of terms in the expansion is
taken. This is the essence of the particle/box (PB) method, requiring
0(N log N) operations. One then tries to minimize the work required
by maximizing the size of the cluster used, while keeping the number
of terms in the expansion within a reasonable limit, and maintaining
a certain degree of accuracy.
The box-box (BB) scheme goes one step further as it accounts for
box-box interactions as well. These interactions are in the form of
shifting the expansions of a certain cluster with the desired accuracy
to the center of another cluster. Then those expansions are used
to determine the velocities of the particles in the second cluster.
This has as an effect the minimization of the tree traversals for the
individual particles, requiring only 0(N) operations.
Applications:
To illustrate the use of the viscous vortex method in two dimen-
sion, we show in Figure 4 the vorticity distribution in accelerating
flow past a flat plate at 90° angle-of-attack. (P. Koumoutsakos, pri-
vate communication 1994). For similar applications to cylinder flow,
98 A. Leonard

Figure 3. Example of particle clustering for an elliptical spiral dis-


tribution of 1000 particles

see Koumoutsakos and Leonard (1995). A three-dimensional exam-


ple is shown in Figure 5. Shown is the deformation of an initially
spherical vortex sheet corresponding to potential flow past a sphere.

Spectral Element Method:


A powerful alternative approach for complex geometries is the
spectral element method (Patera, 1984) (Karniadakis, 1989) in which
local spectral (polynomial) expansions are used over quadrilateral
subdomains in two dimensions or hexahedral subdomains in three
dimensions. See Chu and Karniadakis (1992) for the application to
turbulent flow over streamwise grooves or riblets. Recently, Hender-
son and Karniadakis (1994) extended the method to a nonconforming
Direct Numerical Simulation 99

Figure 4. Vorticity contours for accelerating freestrearn flow past a


flat plate at 90° incidence. Viscous vortex method. Reynolds number
based on acceleration, a, = VaL 3 /z/ = 1,296. (a) Nondimensional
time = t<i/a/L — 2.5, number of vortex particles ~ 210,000; (b)
time = 3.0, number of vortex particles « 290,000 (P. Koumoutsakos,
private communication 1994).
100 A. Leonard

Figure 5. Evolution of a spherical sheet of vorticity corresponding


to potential flow past a sphere. Three-dimensional, inviscid vortex
particle method with 81920 particles (Salmon et a/., 1994).

grid structure so that local mesh refinement is possible. In addition,


Sherwin and Karniadakis (1994) have extended the technique to tri-
angular (2D) and tetrahedral (3D) subdomains.

6 SIMULATION OF COMPRESSIBLE TURBULENCE


The previous sections have considered the direct simulation of
incompressible turbulence for constant density flows. To set the
mesh spacing in this case for a fixed-grid computation, one need
Direct Numerical Simulation 101

only take into account the generation of small-scale vorticity struc-


tures brought about by the straining of the vorticity field. Mesh
requirements are dictated by the smallest of these scales that sur-
vive viscous diffusion, and the timestep scales with the mesh width
divided by a convection velocity. In addition, the pressure is just
a Lagrange multiplier whose purpose is to keep the velocity field
divergence-free (recall Eq.(21) above for a succinct, pressure-free de-
scription of incompressible fluid mechanics). In compressible turbu-
lence, the pressure is a thermodynamic variable and has dynamical
significance. It is, in principle, computed from the equation of state,
knowing the density and the internal energy. These variables, in
turn, are determined from the mass conservation equation,

and the total energy equation, respectively. We can convert the


energy equation to an evolution equation for the pressure. For a
perfect gas, the resulting pressure equation is

These two equations now join the momentum equation

and the equation of state

to give a complete system of equations for the dynamics. In the


above, the density, pressure p, and temperature T have been made
nondimensional by their respective reference values p a ,p a ,and Ta,
while distance, time and velocity are nondimensionalized using a
reference velocity, U, and length, L. The absolute viscosity fj, and the
thermal conductivity K are also normalized by reference values fia
and Ka respectively. In addition,

is the viscous stress tensor (with the bulk viscosity assumed zero),
L is the identity tensor,
102 A. Leonard

is the viscous dissipation function, 7 is the ratio of specific heats,


M = U/V^RTa is the reference Mach number (R is the universal
gas constant), Re = paUL/Ma is the Reynolds number, and Pr =
MaCp/K a is the Prandtl number (Cp is the specific heat at constant
pressure).
With the additional thermodynamic degrees of freedom, there
are other modes besides vorticity modes that can become active. In
particular, Kovasznay (1953) identified acoustic modes and entropy
modes that evolve independent of the vorticity modes and indepen-
dent of each other in linearized flow. Acoustic modes can, of course,
be in the form of sound waves and are present, for example, at high
amplitudes in the case of an eddy shocklet discussed below. Entropy
modes can arise even in the incompressible limit, in which the ther-
modynamic pressure is constant and no acoustic modes are present.
In this limit, density, and hence entropy, fluctuations travel with the
local fluid velocity and interact with the vorticity field through the
nonlinear baroclinic source term for vorticity,

which arises when one takes the curl of the momentum equation to
obtain the vorticity transport equation.
The initial conditions for, say, isotropic, homogeneous turbulence
are, of course, more varied for compressible flows. The initial ve-
locity field may be decomposed (Helmholtz decomposition) into a
solenoidal or "incompressible" component, u1 and a dilatational or
"compressible" component, u c , i.e.,

where

Similarly the pressure may be expressed as

where p1 is the incompressible pressure given by


Direct Numerical Simulation 103

Thus, for initial conditions and in addition to considerations of initial


spectra, one can have a varying initial fraction of compressible kinetic
energy

and an arbitrary initial compressible potential energy, | < p c2 >


/7 2 , where <> denotes volume average.
Also, initial density fluctuations may be imposed independently.
However, linearized analysis of low-Mach number flows (Erlebacher
et al. 1990, Sarkar et al. 1991) indicates that the compressible
potential energy and the compressible kinetic energy come into equi-
librium rapidly, i.e. on an acoustic timescale that is 0(M) times the
turbulence timescale.
Direct simulations of isotropic turbulence (Sarkar et al. 1991, Lee
et al. 1991) have substantiated this equilibrium process as well as
showing the importance of the turbulent Mach number, defined by

(where c = speed of sound = \/7pa//>a) in determining, for exam-


ple, the level of compressible dissipation. When the turbulence Mach
number is sufficiently large and when the Reynolds number is rea-
sonably large, Lee et al. (1991) have observed the formation of eddy
shocklets, local regions in the flow that have all the proper jump
conditions of a shock wave. Simulations of homogeneous shear tur-
bulence (Sarkar 1994) have shown the importance of gradient Mach
number Mg = S l/c, (where S is the mean dimensional shear rate
and £ is a representative integral length scale of the turbulence) as
well as turbulent Mach number. Studies of inhomogeneous turbu-
lence include the efforts of Sandham and Reynolds(1991) to study
variations in the structure of three-dimensional large-scale eddies in
compressible plane mixing layers as the convective Mach number is
changed. In such flows, the convective Mach number is given by

where Ui and U2 are the freestream speeds on either side of the layer,
and GI and C2 are the corresponding speeds of sound. It is known from
experiment that compressible layers have growth rates significantly
104 A. Leonard

lower than their incompressible counterparts, apparently depending


on the convective Mach number.

7 REFERENCES
Barnes, J. E., and Hut, P., 1986. "A hierarchical 0(N log N) force
calculations algorithm," Nature, 324, pp. 446-449.

Batchelor, G. K., 1967."The theory of homogeneous turbulence,"


Cambridge University Press, Cambridge.

Beale, J. T., and Majda, A., 1982. "Vortex methods II: High order
accuracy in two and three dimensions," Math. Comp. 39, pp.
29-52.

Chorin, A. J., and Marsden, J. E., 1979. "A mathematical intro-


duction to fluid mechanics," Springer-Verlag, New York.

Chorin, A., 1973. "Numerical study of slightly viscous flow," J.


Fluid Mech., 57, pp. 380-392.

Chu, D. C., and Karniadakis, G. E., 1992. "A direction numeri-


cal simulation of laminar and turbulent flow over streamwise
aligned riblets," J. Fluid Mech., 25, pp. 1-42.

Coles, D., 1981. "Prospects for useful research on coherent structure


in turbulent shear flow," Proc. Indian Acad. Sci. (Engg. Sci.),
4, pp. 111-127.

Deardorff, J. W., 1970. "A numerical study of three-dimension-


al turbulent channel flow at large Reynolds numbers," /. Fluid
Mech., 41, pp. 453-480.

Degond, P., and Mas-Gallic, S., 1989. "The weighted particle method
for convection-diffusion equations, Part I: the case of isotropic
viscosity, Part II: the anisotropic case," Math. Comp. 53, pp.
485-526.

Dumas, G., and Leonard, A., 1994. "A divergence-free spectral


expansions methods for three-dimensional flows in spherical-
gap geometries," J. Comput. Phys., Ill, pp. 205-219.
Direct Numerical Simulation 105

Emmons, H. W., 1949. "The numerical solution of the turbulence


problem," in Proceedings of Symposia in Applied Mathematics,
1, McGraw-Hill, New York, pp. 67-71.
Erlebacher, G., Hussaini, M. Y., Kreiss, H. 0., and Sarkar, S., 1990.
"The analysis and simulation of compressible turbulence," The-
oret. Comput. Fluid Dyn., 2, pp. 73-95.
Greengard, L., and Rohklin, V., 1987. "A fast algorithm for particle
simulations," J. Comput. Phys., 73, pp. 325-348.
Henderson, R. D., and Karniadakis, G. E., 1994. "Unstructured
spectral element methods for turbulent flows," J. of Comput.
Phys., submitted.
Hinze, J. 0., 1975. "Turbulence", 2nd ed., McGraw-Hill Inc., New
York, p. 722.
Hockney, R. W., arid Eastwood, J. W., 1981. "Computer simula-
tions using particles,", McGraw- Hill New York.
Huang, M.-J., and Leonard, A., 1994. "Power-law decay of homo-
geneous turbulence at low Reynolds numbers," Phys. Fluids,
6, pp. 3765-3775.
Hussaini, M. Y., Speziale, C. G., and Zang, T. A., 1990. "The
potential and limitations of direct and large eddy simulations,"
Whither Turbulence? Turbulence at the Crossroads, Springer-
Verlag, Berlin, pp. 354-368.
Jimenez, J., Wray, A. A., Saffman, P. G., and Roggalo, R. S., 1993.
"The structure of intense vorticity in isotropic turbulence", J.
Fluid Mech., 255, pp. 65-90.
Karniadakis, G. E., 1989. "Spectral element simulations of laminar
and turbulent flows in complex geometries," App. Num. Math.,
6, p. 85.
Kim, J., Moin, P., and Moser R., 1987. "Turbulence statistics in
fully developed channel flow at low Reynolds number," J. Fluid
Mech., 177, pp. 133-166.
Koumoutsakos, P., and Leonard, A., 1995. "High resolution simu-
lations of the flow around an impulsively started cylinder using
vortex methods,", J. Fluid Mech. (to appear).
106 A. Leonard

Koumoutsakos, P., Leonard, A., and Pepin, F., 1994. "Bound-


ary conditions for viscous vortex methods," J. Comput. Phys.,
113, pp. 52-61.

Koumoutsakos, P., 1993. "Large scale direct numerical simulations


using vortex methods." Ph.D. thesis, Caltech.

Kovasznay, L. S. G., 1953. "Turbulence in supersonic flow,", Jour-


nal of Aeronautical Sciences, 20, pp. 657-682.

Laufer, J., 1954. "The structure of turbulence in fully developed


pipe flow," NACA 1174.

Lee, S., Lele, S. K., and Moin, P., 1991. "Eddy shocklets in decaying
compressible turbulence," Phys. Fluids A, 3, pp. 657-664.

Leonard, A., and Chua, K., 1989. "Three-dimensional interactions


of vortex tubes," Physica D, 37, pp. 490-496.

Leonard, A., 1985. "Computing three-dimensional incompressible


flows with vortex elements," Ann, Rev. Fluid Mech., 17, pp.
523-529.

Leonard, A., and Koumoutsakos, P., 1993. "High resolution vortex


simulation of bluff body flows," J. Wind Eng. and Indust.
Aero., 4G&47, pp. 315-325.

Leonard, A., and Reynolds, W. C., 1988. "Turbulence research


by numerical simulation," Perspectives of Fluids Mechanics,
Lecture Notes in Physics, Vol. 320, Springer-Verlag.

Leonard, A., and Wray, A., 1982. " A new numerical method for the
simulation of three- dimensional flow in a pipe," Proceedings
of the 8th International Conference on Numerical Methods in
Fluids Dynamics, June 28-July 2, 1982, Aachen, W. Germany,
Springer-Verlag, New York, pp. 335-342.

Leonard, A., 1980. "Vortex methods for flow simulation," J. Com-


put. Phys., 37, pp. 289-335.

Moser, R. D., and Rogers, M. M., 1990. "Mixed transition and the
cascade to small scales in a plane mixing layer," Phys. Fluids
A, 5, pp. 1128-1134.
Direct Numerical Simulation 107

Moser, R. D., Moin, P., and Leonard, A., 1983. "A spectral numer-
ical method for the Navier-Stokes equations with applications
to Taylor-Couette flow," J. Comput. Phys., 52, p. 524.

Orszag, S. A., and Patterson, G. S., Jr., 1972. "Numerical simula-


tion of three-dimensional homogeneous isotropic turbulence,"
Phys. Rev. Lett., 28, pp. 76-79.

Orszag, S. A., 1980. "Spectral methods for problems in complex


geometries," J. Comput. Phys., 37, p. 70.

Patera, A. T., 1984. "A spectral element method for fluid dynamics;
laminar flow in a channel expansion," J. Comput. Phys., 54,
pp. 468-488.

Pepin, F., 1990. "Simulation of the flow past an impulsively started


cylinder using a discrete vortex method," Ph.D. thesis, Caltech.

Reynolds, W. C., 1990. "The potential and limitations of direct


and large eddy simulations," Whither Turbulence? Turbulence
at the Crossroads, Springer-Verlag, Berlin, pp. 313-343.

Rogallo, R. S., 1981. "Numerical experiments in homogeneous tur-


bulence," NASA TM-81315.

Rogallo, R. S., and Moin, P., 1984. "Numerical simulation of tur-


bulent flows," Ann. Rev. Fluid Mech., 16, pp. 99-137.

Saddoughi, S. G., and Veeravalli, S. V., 1994. "Local isotropy in


turbulent boundary-layers at high Reynolds-number," J. Fluid
Mech., 268, pp. 333-372.

Salmon, J. K., Warren, M. S., and Winckelmans, G. S., 1994. "Fast


parallel tree codes for gravitational and fluid dynamical N-body
problems," Int. J. Supercomputer Applications, 8, pp. 129-142.

Sandham, N. D., and Reynolds, W. C., 1991. "Three-dimensional


simulations of large eddies in the compressible mixing layer,"
J. Fluid Mech. 224, pp. 133-158.

Sarkar, S., 1994. "The stabilizing effect of compressibility in tur-


bulent shear flow," NASA Contractor Report 194932, ICASE
Report No. 94-46, pp. 1-36.
108 A. Leonard

Sarkar, S., Erlebacher, G., Hussaini, M. Y., and Kreiss, H. 0., 1989.
"The analysis and modelling of diltational terms in compress-
ible turbulence," J. Fluid Mech., 227, pp. 473-493.
She, Z.-S., Chen, S.-Y., Doolen, G. D., Kraichnan, R. H., and
Orszag, S. A., 1993. "Reynolds-number dependence of iso-
tropic Navier-Stokes turbulence," Phys. Rev. Letters, 70, pp.
3251-3254.
Sherwin, S. J., and Karniadakis, G. E., 1994. "A triangular spec-
tral element method: algorithms and flow simulations," Pro-
ceedings of the 14th International Conference on Numerical
Methods in Fluids Dynamics, Bangalore, 11-15 July.
Spalart, P. R., 1987. "Direct simulation of a turbulent boundary
layer up to R<? = 1410," J. Fluid Mech., 187, pp. 61-98.
Stanaway, S. K., CantweU, B. J., and Spalart, P. R., 1988. AIAA
Paper 88-0318.
von Neumann, J., 1963. "Recent theories of turbulence," 1949 re-
port to the Office of Naval Research, reprinted in John von
Neumann, Collected Works, 6, A. H. Taub, ed., Macmillan
Co., New York, pp. 437-472.
Winckelmans, G. S., Salmon, J. K., Warren, M. S., and Leonard,
A., 1995. "The fast solution of three-dimensional fluid dynam-
ical N-body problems using parallel tree codes: vortex element
method and boundary element method," Seventh SIAM Conf.
on Parallel Processing for Scientific Comp., Feb. 1995, San
Francisco.
Winckelmans, G. S., and Leonard, A., 1993. "Contributions to vor-
tex particle methods for the computation of three-dimensional
incompressible unsteady flows," J. Comput. Phys., 109, pp.
247-273.
Chapter 3

LARGE EDDY SIMULATION

Joel H. Ferziger

1 INTRODUCTION

Over a decade ago, the author (Ferziger, 1983) wrote a review of


the then state-of-the-art in direct numerical simulation (DNS) and
large eddy simulation (LES). Shortly thereafter, a second review was
written by Rogallo and Moin (1984). In those relatively early days
of turbulent flow simulation, it was possible to write comprehensive
reviews of what had been accomplished. Since then, the widespread
availability of supercomputers has led to an explosion in this field so,
although the subject is undoubtedly overdue for another review, it is
not clear that the task can be accomplished in anything less than a
monograph. The author therefore apologizes in advance for omissions
(there must be many) and for any bias toward the accomplishments
of people on the west coast of North America.
In the earlier review, the author listed six approaches to the
prediction of turbulent flow behavior. The list included: corre-
lations, integral methods, single-point Reynolds-averaged closures,
two-point closures, large eddy simulation and direct numerical sim-
ulation. Even then the distinction between these methods was not
always clear; if anything, it is less clear today.
It was possible in the earlier review to give a relatively complete
overview of what had been accomplished with simulation methods.
Since then, simulation techniques have been applied to an ever ex-
panding range of flows so a thorough review of simulation results is

109
110 J. H. Ferziger

no longer possible in the space available here. Simulation techniques


have become well established as a means of studying turbulent flows
and the results of simulations are best presented in combination with
experimental data for the same flow. There is also a danger that the
success of simulation methods will lead to attempts to apply them
too soon to flows which the models and techniques are not ready to
handle. To some extent, this is already happening.
Direct numerical simulation (DNS) is a method in which all of
the scales of motion of a turbulent flow are computed. A DNS must
include everything from the large energy-containing or integral scales
to the dissipative scales; the latter is usually taken to be the viscous
or Kolmogoroff scales. For any reasonable Reynolds number, this
requires a large number of grid points or modes and is very costly.
Despite the cost, the ability of the DNS to yield all of the flow vari-
ables at a large number of spatial locations for many instants of time
has made it a valuable tool for investigating the physics of turbu-
lence. Indeed, for a number of simple flows, it is now the method
of choice. This subject is covered in greater detail in Chapter 2 by
Leonard and will be dealt with only briefly here.
By large eddy simulation (LES), we mean an approach in which
the largest scales of motion are represented explicitly while the small
scales are treated by some approximate parameterization or model.
Large eddy simulations are three dimensional and time dependent,
and thus, expensive although, but as we shall see, they may be much
less costly than a DNS of the same flow.
As a result, LES has come to occupy a kind of middle ground.
Whenever DNS is feasible for a given flow, it should be the method
of choice, especially when a study of the detailed physics of the flow
is the goal. However, cost often makes DNS infeasible. Use of LES
is a good choice for investigating flows that are too complex to be
computed economically by DNS; in practice, with today's comput-
ers, this means any flow which is inhomogeneous in more than one
direction. LES is now becoming powerful enough to be worthy of
consideration as a method to be employed selectively by the work-
ing (as opposed to the research) engineer; we shall comment on this
below.
In this article, we describe the methods employed in LES; in this
area, there is a great deal of overlap with DNS. We shall concentrate
on the differences, principally the models that need to be employed
to represent the small scales and numerical methods that can be
Large Eddy Simulation 111

used in more complex geometries. First, brief overviews of Reynolds


averaged methods and direct numerical simulation will be given to
set the stage for the discussion of LES. We shall then explore the
boundaries of what is feasible with LES today. We begin with some
remarks on turbulence and prediction methods for practical flows.

2 TURBULENCE AND ITS PREDICTION

2.1 The Nature of Turbulence

It is difficult even for expert researchers to agree on a definition


of turbulence. Required elements that are generally agreed upon
include three dimensionality, unsteadiness, strong vorticity, and a
broad-banded spectrum. Beyond that, the issue becomes difficult.
Most would agree that turbulent flows are highly random and/or
noisy; the term chaotic could be used but it has been given a stricter
meaning in recent years.
However, there is more to turbulence than randomness. It is gen-
erally agreed that so-called coherent structures exist in nearly all tur-
bulent flows; we say 'so-called' because, despite considerable discus-
sion, an agreed definition of this term does not yet exist. Moreover,
although the coherent structures account for only a small fraction of
the turbulent energy, they are apparently responsible for more than
their fair share of the transport of properties such as species, mass,
momentum and energy. The coherent structures of a particular flow
are not identical and do not appear regularly in either time or space.
The remainder of the turbulent energy is apparently due to truly ran-
dom motion (which may be the remains of old coherent structures)
and probably causes the irregularity of the coherent structures.
This picture helps explain why turbulence is such a difficult prob-
lem. If it were random, statistical methods would probably have
solved the problem by now. If it were purely deterministic, com-
puter simulation might have solved the problem. In fact, turbulence
is sufficiently incoherent that the signal-to-noise ratio of the coherent
structures is very low; at the same time, the lack of a clear definition
of a structure makes eduction of a pattern from noisy data nearly
impossible.
When we add to this picture the probability that the coherent
structures are different in each flow, we see that the likelihood of
112 J. H. Ferziger

finding a simple method for predicting all flows is exceedingly small.


The search for a single universal method capable of predicting all
turbulent flows has gone on for along time and, while it has produced
many useful results, the goal remains a long way off.

2.2 RANS Models

In Reynolds averaged approaches to turbulence, all of the un-


steadiness is averaged out; this means that all unsteadiness is re-
garded as part of the turbulence. The non-linear terms in the Navier-
Stokes equations give rise to the Reynolds stress term in the Reynolds-
averaged Navier-Stokes (RANS) equations. This term must be mod-
eled if the equations are to be closed. The complexity of turbulence
makes it unlikely that any single model will be able to represent all
turbulent flows. Thus, turbulence models should be regarded as en-
gineering correlations or approximations rather than scientific laws;
this interpretation allows one to 'tune' models, hopefully in sensible
ways, for particular features that may arise in individual turbulent
flows. Experience with RANS-based turbulence models has yielded
both successes and failures. It is the lack of consistency that has led
to interest in new approaches, such as large eddy simulation.
Although the subject of this work is large eddy simulation it is
helpful to note a few things about RANS models. The mean velocity
field may be defined by ensemble, time, or spatial averaging; in any
case, the RANS equations are:

where [/,- =< ul >, the brackets denoting one of the averages listed
above, and T;J is the Reynolds stress tensor,

(actually, a stress divided by the density) and needs to be modeled.


In this paper, we shall assume that the flow is incompressible except
where noted otherwise (see Section 4F). We also use the convention
that any repeated index is summed over.
In laminar flows, energy dissipation and transport of mass, mo-
mentum, and energy normal to the streamlines are all mediated by
the viscosity, so it is natural to assume that the effect of turbulence
Large Eddy Simulation 113

can be represented as an increased viscosity. This leads to the eddy


viscosity model:

where VT is the eddy viscosity.


In the simplest description, turbulence can be characterized by
its kinetic energy, k, or equivalently, a velocity q = \/2&, and a length
scale, L. The eddy viscosity, which carries dimensions Iength 2 time~ 1 ,
must be:

In mixing length models, k is determined from the mean velocity


field using q = LdU/dy and L, which should be the integral scale
of the turbulence, is prescribed in terms of a physical length scale
or a shear layer thickness. Accurate prescription of L is possible
for simple flows but not for separated or highly three dimensional
boundary layers. The simplicity of mixing length models allows them
to be easily modified to account for pressure gradients, curvature,
transpiration, etc.
Two-equation models retain the Boussinesq eddy viscosity con-
cept but use a partial differential equation for the turbulent kinetic
energy k to determine the velocity scale. To obtain the dissipation
and the length scale L, we note that in so-called equilibrium turbu-
lent flows i.e. ones in which the rates of production and destruction
of turbulence are in near-balance, the following relation among" the
dissipation, e, and k and L:

may be used. Eq. (2.5) allows one to use an equation for the dissi-
pation as a means of obtaining both e and L. No constant is used in
Eq (2.5) because this constant combines with others in the complete
model.
An exact equation for the dissipation can be derived from the
Navier-Stokes equations and has a form similar to the energy equa-
tion. The modeling applied to the dissipation equation is so severe
that it is probably best to regard the entire equation as a model in
itself. Difficulties associated with the dissipation equation (or any
114 J. If. Ferziger

other equation used to determine the length scale) are the most dif-
ficult ones in two-equation modeling.
Some of the significant deficiencies of models based on Eq. (2.3)
are direct consequences of the eddy viscosity relationship itself. In
three dimensional flows, the eddy viscosity may no longer be a scalar;
measurements and numerical simulations show that it becomes highly
anisotropic, i.e. it is actually a tensor quantity.
The effect of an eddy viscosity can be interpreted in another
way. Enough viscosity is added to the equations to assure that the
computed flow is stable i.e., the solution of the RANS equations is
an effective laminar flow with a velocity field that is the mean (in the
Reynolds sense) of the turbulent velocity field. In two dimensions, it
is always possible to define a spatially dependent eddy viscosity that
produces the correct mean flow. In general, it is not possible to find
this eddy viscosity without knowledge of the solution but it is useful
to know that it exists in principle. In three dimensional flows, the
eddy viscosity may be a tensor of either second or fourth rank.
Anisotropic or tensor models have been proposed. If a model
is to be applicable to a wide range of flows, it should possess in-
variance properties, i.e. it must give the same results independent
of the coordinate system is used in the calculation. Many early
anisotropic models were not properly invariant. Recently, invariant
tensor and/or non-linear models have been proposed, for example,
see Speziale (1987) and Horiuti (1990). These models take the form:

where Sij is

Although these models contain more constants than the scalar eddy
viscosity model, some of them are fixed by requiring invariance. A
detailed discussion of these types of closures can be found in Chapters
5 and 6.
The most complex models in common use today are Reynolds
stress models which are based on dynamic equations for the Rey-
nolds stress tensor itself. As these are complicated, and because
their application as subgrid scale models in large eddy simulation
Large Eddy Simulation 115

has been limited, we shall not describe or discuss them here. A de-
tailed discussion of these types of closures can be found in Chapter
5 and 6.
With few exceptions, RANS models cannot be applied to flow
near a surface without modification. Special near-wall versions of the
models, especially the k — c model, have been developed and work
quite well for attached boundary layers. One of the most recent of
these models was proposed by Rodi and Mansour (1992), who include
references to other models.

2.3 Direct Numerical Simulation (DNS)


The most exact approach to turbulence simulation is to solve the
Navier-Stokes equations without averaging or approximation. The
result is a single realization of a flow and is equivalent to a short-
duration laboratory experiment; this approach is called direct nu-
merical simulation (DNS).
It is important to recognize that the considered domain must be
at least as large as the largest turbulent eddy; from a practical point
of view this means that the linear dimension of the domain must be
at least a few times the integral scale L. On the other hand, for a
simulation to capture all of the dissipation, which occurs on the small
scales on which viscosity is active, the grid must be no larger than
the viscously determined Kolmogoroff scale, 77. For homogeneous
turbulence, the simplest type of turbulence, there is no reason to
use anything other than a uniform grid. In this case, the number of
grid points in each direction must be at least L/r); it is easily shown
(Tennekes and Lumley, 1976) that this ratio is proportional to .Re3/4.
Since this number of points must be employed in each of the three
coordinate directions, and the time step is related to the grid size,
the cost of a simulation scales as Re3.
This means that direct numerical simulation can be carried out
only at relatively low Reynolds numbers. For homogeneous turbulent
flows, the Reynolds number of interest must be based on the turbu-
lent velocity and length scales. As these scales are typically an order
of magnitude or more smaller than the corresponding macroscopic
scale, the ability to compute flows with turbulent Reynolds numbers
of 100 actually allows DNS to reach the low end of the range of Rey-
nolds numbers of engineering interest. For further details of DNS see
Chapter 2.
Having described the methods that bracket it, we now turn to
116 J. H. Ferziger

the principal subject of this paper, large eddy simulation.

3 FILTERING

We now begin the description of large eddy simulation (LES).


The idea is to simulate the larger scales of motion of the turbulence
while approximating the smaller ones. One can think of it as ap-
plying DNS to the large scales and RANS to the small scales; it is
a compromise between the two approaches, a concept that we shall
explore in more detail later. The justification for such a treatment
is that the larger eddies contain most of the energy, do most of the
transporting of conserved properties, and vary most from flow to
flow; the smaller eddies are believed to be more universal and less
important and should be easier to model. It is hoped that univer-
sality is more readily achieved at this level than in RANS modeling
but this assertion remains to be proven.
As in the RANS case, it is essential to define the quantities to be
computed precisely. To do this it is essential to define a velocity field
that contains only the large scale components of the total field. This
is best done by filtering (Leonard, 1974); the large or resolved scale
field is essentially a local average of the complete field. We shall use
one dimensional notation for convenience; the generalization to three
dimensions is straight-forward. The filtered velocity is defined by:

where G(x, x'), the filter kernel, is a localized function or a function


with compact support i.e., one which is large only when x and x' are
not far apart. Filter kernels which have been applied in LES include:
• Gaussian. The Gaussian has the advantage of being smooth
and infinitely differentiable in both physical and Fourier space. In
fact, its Fourier transform is Gaussian in wavenumber space.
• Box. This is simply an average over a rectangular region. It
is a natural choice when finite difference or finite volume methods
are used. Two versions of this filter have been used and ought to be
distinguished:
In the moving box filter, the average is taken over a region of
space surrounding any chosen point. According to this definition,
Large Eddy Simulation 117

Ui(x) is a continuous function of x. This filter is similar in many


ways to the Gaussian.
A filter which is an average over a grid volume of a finite difference
or finite volume mesh is tied more closely to the numerical method.
According to this definition, u is a piecewise constant function of x
(Schumann, 1973).
• Cutoff. This filter is defined in Fourier space and eliminates all
of the Fourier coefficients belonging to wavenumbers above a partic-
ular cutoff. It is natural to use this filter in conjunction with spectral
methods as it leaves more energy in the large scale field than the fil-
ters defined above. However, it is difficult to apply to inhomogeneous
flows.
When the Navier-Stokes equations are filtered one obtains a set
of equations very similar in form to the RANS equations.

Of course, the definitions of the velocities appearing in Eqs (2.1)


and (3.2) differ but the closure issues are very similar. Since

a modeling approximation for the difference between the two sides


of this inequality,

must be introduced. In the context of large eddy simulation, r,-j is


called the subgrid scale (SGS) Reynolds stress. It plays a role in LES
similar to the role played by the Reynolds stress in RANS models
but the physics that it models is different. The SGS energy is a
much smaller part of the total flow than the RANS turbulent energy
so model accuracy may be less crucial in an LES than in RANS
computations.
Subgrid scale modeling is the most distinctive feature of large
eddy simulation and is the subject of the next section, the longest
one in this work.
118 J. If. Ferziger

4 SUBGRID SCALE MODELING

4.1 Physics of the Subgrid Scale Terms

The models used to approximate the SGS Reynolds stress (3.4)


are called subgrid scale (SGS) models. This nomenclature is derived
from the kind of LES in which one applies a finite volume approx-
imation directly to the Navier-Stokes equations; the filter is then
closely connected to the grid used to discretize the equations. This
technique was used in the earliest large eddy simulations and the
nomenclature has stuck. More generally, there need not be a con-
nection between the filter and the grid used in the solution method
so this nomenclature is more restrictive than necessary but it is too
late to change it.
One important difference between filtering and Reynolds aver-
aging is that, in general, filtering a field a second time does not
reproduce the original filtered field:

The exception is the cutoff filter for which equality does hold. The
difference between the two sides of this inequality will be exploited
for modeling purposes later.
For now, we note that the difference represented by Eq (4.1)
means that the both the physics and modeling of the subgrid scale
Reynolds stresses (SGSRS) may be more complicated than for the
RANS Reynolds stresses. By using the kind of decomposition of
the velocity field used in RANS modeling i.e. writing the complete
velocity field as a combination of the filtered field and a subgrid scale
field, we can decompose the SGSRS into three sets of terms:

which can be ascribed physical significance. In particular, these


terms represent the following physics:
• The first term, which can be computed explicitly from the fil-
tered velocity field, u, represents the interaction of two resolved scale
eddies to produce small scale turbulence. It has been called the
Leonard term and, sometimes, the outscatter term.
• The second term represents the interaction between resolved
scale eddies and small scale eddies. This term, also called the cross
Large Eddy Simulation 119

term, can transfer energy in either direction but, on the average,


transfers energy from the large scales to the small ones.
• The third term represents the interaction between two small
scale eddies to produce a large scale eddy and is called the true
subgrid scale term; as it produces energy transfer from the small to
the large scales, it is also called the backscatter term; as noted above,
the cross term may produce backscatter as well.
In the past, it was thought that, as each term represents a differ-
ent physical phenomenon, it ought to be modeled separately. How-
ever, modeling (either SGS or RANS) is far from exact and the un-
certainty in the modeling defeats any attempt at precision. Con-
sequently, in the recent past, it has become common to model the
entire subgrid scale Reynolds stress as a single unit.
It should also be noted that the subgrid scale Reynolds stress
is a local average of the small scale field. This means that models
for it should be based on the local velocity field or, perhaps, on the
past history of the local fluid. The latter can be accomplished by
using a model that solves partial differential equations to obtain the
parameters needed to determine the SGSRS.
We next look at subgrid scale models in some detail.

4.2 Smagorinsky Model

By far the most commonly used subgrid scale model is the one
proposed by Smagorinsky (1963). It is an eddy viscosity model that
can be thought of as an adaptation of the Boussinesq concept of Eq
(2.3) to the subgrid scale. It is:

This model can also be derived in a number of other ways. These


include heuristic methods, equating production and dissipation of
subgrid scale turbulent kinetic energy, and via turbulence theories
such as the direct interaction approximation (DIA) (Leslie, 1973),
the eddy damped quasi-normal Markovian (EDQNM) approximation
(Lesieur, 1992), and renormalization group theory (RNG) (Yakhot
and Orszag, 1986).
The form of the subgrid scale eddy viscosity can be derived by
dimensional arguments. We shall present a heuristic argument but
120 J. H. Ferziger

it can also be derived from theories including Kolmogoroff-like argu-


ments (Lilly, 1965) and the theories mentioned above. The following
argument contains some elements of these approaches.
At high Reynolds numbers, the dissipation in a turbulent flow
takes place at very small scales while energy is introduced at the
largest scales. Between these is a regime in which there is neither
significant production nor dissipation of turbulent energy, the iner-
tial subrange. In this range, only inviscid mechanisms are active
and energy is transferred from large to small scales. Since it is the
non-linear (advective) term in the Navier-Stokes equations that is
responsible for the energy transfer, the rate of transfer to the small
scales may be estimated as the magnitude of the contribution of this
term to the kinetic energy equation, which is (!/2}d(uiUiUj}/dxj.
As the large energetic scales supply the largest contribution to this
term, the magnitude scales as:

where Q is a velocity scale for the energetic eddies and L is the


integral scale of the turbulence.
Let us further assume that the largest subgrid scales are far re-
moved from the viscous scales. A repeat to the above argument then
shows that:

where q is a velocity typical of the subgrid scale field (most of which


resides in the largest subgrid scales) and A is the size of the largest
subgrid eddies, the length scale associated with the filter.
In a large eddy simulation, the large or resolved scales lose energy
by transferring it to the subgrid scales. From the point of view of the
large scale eddies, this appears to be dissipation i.e., it is energy lost
never to be recovered. A model of the eddy viscosity type represents
this energy transfer as effective viscous dissipation. Since the model
most affects the smallest resolved scales (which are of size A), the
magnitude of the effective dissipation may be estimated as:

Equating (4.5) and (4.6) shows that the eddy viscosity must take
the form:
Large Eddy Simulation 121

which could have been derived via dimensional arguments. We now


find q by using Eqs. (4.4) and (4.5) and substitute it into Eq. (4.7)
to obtain:

Finally, estimating Q as:

and inserting a model parameter to produce equality, we have:

As noted above, this result can be derived in a number of other ways.


The theories provide estimates of the constant as well as the form of
the model.
The presence of the integral scale L in the formulation for the
eddy viscosity makes the model difficult to use. Computing the in-
tegral scale, especially in inhomogeneous flows, could require a great
deal of effort. For this reason, the substitution:

is often used, leading to the usual form of the Smagorinsky model:

Other derivations lead directly to this form of the model.


As noted, the theories also predict the value of Cs, which is more
appropriately called a parameter than a constant. Most of these
derivations are truly valid only for isotropic turbulence but they all
agree that Cs ~ 0.2. LES of isotropic turbulence also shows that
this value of the parameter is optimum; varying it by about ten
or fifteen percent produces acceptable results. It should be noted,
however, that the range of Reynolds numbers studied is relatively
narrow so this may not be a very severe test of the model. Indeed
the substitution (4.11) used to produce the standard version of the
Smagorinsky model may mean that the parameter, Cs, is not a true
constant, but rather a function of A/Z< which is, in turn, a function
122 J. H. Ferziger

of Reynolds number. We should not be surprised if we find that the


parameter Cs needs to be a function of Reynolds number or other
non-dimensional parameters or is different in different flows.
The Smagorinsky model, although relatively successful, is not
without problems. For example, it has been found that, to simulate
channel flow, several modifications are required. The first is that
the value of the parameter Cs in the bulk of the flow has to be
reduced from 0.2 to approximately 0.065 which amounts to reducing
the eddy viscosity by almost an order of magnitude. Secondly, in
the region close to the surface, the value has to be reduced even
further. A recipe that has been found to be successful is the van
Driest damping that has long been used to reduce the near-wall eddy
viscosity in RANS models:

where y+ is the distance from the wall in viscous wall units (y+ =
yur/v, where ur is the shear velocity (r//?) 1 ' 2 and T is the shear stress
at the wall) and A+ is a constant usually taken to be approximately
25. Although this modification produces the desired results, it is
difficult to justify in the context of LES. The SGS model should
depend solely on the local properties of the flow and it is difficult
to see how the distance from the wall qualifies as an appropriate
parameter in this regard.
The purpose of the van Driest damping is to reduce the subgrid
scale eddy viscosity near the wall; it is generally believed that VT ~ y3
in this region and models ought to respect this property. It follows
that an alternative to van Driest damping is a subgrid scale model
which reduces the magnitude of the viscosity in the proper manner
when a subgrid scale Reynolds number (the obvious one is |5|A2/!/)
becomes small. Models of this kind were suggested by McMillan and
Ferziger (1980) and by Yakhot and Orszag (1986): the latter used
renormalization group theory to derive their model. It is necessary to
point out that these issues focus on the fully turbulent flow. Applica-
tion to transitional flows present further problems and modifications
which have been addressed initially by Piomelli et al.
A further problem is that, near a wall, the structure of the flow
is very anisotropic. Regions of low speed fluid (streaks) are created.
They have dimensions of approximately 1000 viscous units in the
strearriwise direction and perhaps 100 viscous units in the spanwise
Large Eddy Simulation 123

and normal directions. Resolving these streaks requires a highly


anisotropic grid and the question arises: what is the appropriate
length scale to use in the SGS model in this region? The usual choice
is (A1A2A3) 1 / 3 but another possibility is (Af + A?, + A^) 1 / 2 and still
others are easily constructed. It is possible that, with a proper choice
of length scale, the damping (4.13) would become unnecessary. An
alternative model for the near-wall region was proposed by Schumann
(1973); this model employs the horizontally averaged velocity and
thus does not satisfy the condition that a model should be based
only on the local velocity field. These issues were discussed in some
detail by Moin and Kim (1982) and Piomelli et al. (1989).
It has been found that, in a stably stratified fluid, it is again
necessary to reduce the value of the parameter in the Smagorinsky
model. Stratification commonly occurs in geophysical flows, where
the practice is to make the parameter a function of the Richardson
number, a non-dimensional parameter that represents the relative
importance of stratification and shear. Similar effects occur in flows
in which rotation and/or curvature play significant roles.
Thus, there are many difficulties with the Smagorinsky model. It
may be that the principal reason why this model has been relatively
successful is that most of the flows for which accurate results have
been obtained are relatively simple low Reynolds number cases; an
exception is buoyancy-driven flows for which good results have been
obtained with the Smagorinsky model even at relatively high nondi-
mensional parameter values. In such flows, the energy in the subgrid
scales and the rate of energy transfer to these scales are both rela-
tively small. Then the model may not need to reproduce the actual
subgrid scale Reynolds stress very accurately to produce acceptable
results; it may suffice to simply dissipate energy at the proper overall
rate. However, if we wish to simulate more complex and/or higher
Reynolds number flows, it may be important that the model be more
accurate in detail.

4.3 A Priori Testing

The traditional test of a model consists of applying it to the so-


lution of a problem and comparing the prediction with experimental
results for the same flow. This is an obvious and practical way of
testing models which we shall call the a posteriori approach.
The availability of direct numerical simulations (and, possibly, in
124 J. H. Ferziger

the future, detailed experimental data) makes another kind of testing


possible. Let us accept that the results of a DNS represent an exact
realization of a turbulent flow. Having DNS results, one can ask how
LES would fare for the same realization. In particular, it becomes
possible to evaluate exactly those terms which must be modeled in
the LES and, at the same time, the model estimates of them. Let
us see how this might work; for simplicity, we assume that the flow
under consideration is homogeneous.
Given the exact velocity field at an instant, u,-, it is straight-
forward to filter it to obtain the large eddy component of that field,
Ui. It is then not difficult to compute the subgrid scale Reynolds
stress tensor, u^Uj — U{Uj. Finally, one can compute the model esti-
mate of the Reynolds stress. We thus have data on the exact Rey-
nolds stress (R) and its model representation (M) at essentially every
point in the flow. To test the accuracy of the model, one need merely
compare the two. Two popular methods of doing so are by computing
a correlation coefficient:

and by producing a scatter plot i.e., a plot of the exact values of


a quantity vs. the corresponding model values. In this way, an
unambiguous test of a model can be produced. Examples of such
scatter plots are given in Fig. 1; it presents the Reynolds stress and
the Smagorinsky estimate of it at approximately 4000 points. If the
model was exact, all of the points would fall on a single straight
line. It is clear that the model is far from perfect. The correlation
coefficient of the data shown in this figure is approximately 0.35;
as the square of the correlation coefficient represents the fraction of
the data predicted by the model, this means that the model only
represents about ten percent of the data!
In a similar way, by comparing the magnitudes of the model and
exact values, one can obtain a value of the model parameter. The val-
ues obtained in this way are in good agreement with those obtained
by other means (cf. Clark et a/., 1979).
It is thus clear that, in the precise sense that a priori testing
provides, the Smagorinsky model is not very accurate. Clark et al.
(1979) used the a priori test to show that the problem does not arise
from the value of the parameter but rather from the fact that the
SGS Reynolds stress tensor and the strain rate of the resolved field,
Large Eddy Simulation 125

Figure 1. Scatter plot of the Smagorinsky model prediction of the


subgrid scale Reynolds stress and the exact value obtained from a
direct numerical simulation. From Bardina et al., 1980.

which the Smagorinsky model assumes to be proportional, actually


have little relation to each other. In particular, the principal axes
of the two tensors (which need to be identical for proportionality to
hold) are not well correlated.
A problem with the a priori method is that it is in some ways
akin to an in vitro biological test; great differences may be found
when in vivo testing is performed. Similarly, using a model in an
actual LES may give results that differ from what the a priori test
finds. The velocity field computed will differ from the large scale part
of the DNS field used in the test. Indeed, despite the poor rating
the Smagorinsky model receives in a priori tests, it seems to perform
reasonably well in LES.

4.4 Scale Similarity Model

The concept that the small scales of a simulation can be used to


study modeling has a number of interesting extensions. One leads
to an alternative model for the subgrid scales, the scale similarity
model (Bardina et al., 1980). A more important extension will be
presented in the following section.
The idea behind the scale similarity model is that the important
interactions between the resolved and unresolved scales involve the
smallest eddies of the former and the largest eddies of the latter
i.e., eddies that are a little larger or a little smaller than the length
126 J. H. Ferziger

scale, A, associated with the filter. These can be extracted from the
velocity field in the following manner.
From the complete velocity field, v,j, we can compute the resolved
or large scale field ¥,• by filtering and the small or subgrid scale
field u\ — U{ — U{ by subtraction. From these we can construct a
further subdivision. The very largest resolved scales may be defined
by filtering a second time to obtain ¥; so the smallest resolved scales
are defined by Uj — u,-. The largest unresolved scales are defined by
u\, A simple calculation shows that these are identical. This leads
to the following possibility as a subgrid scale model:

No constant is used because it can be shown (Reynolds, pri-


vate communication, and Speziale, 1983) that Galilean invariance
demands that the constant be unity. From its construction, it is not
surprising that this model correlates very well with the actual SGS
Reynolds stress in a priori tests, see Fig. 2. The argument of the
preceding suggests that, in essence, it is an identity. When applied in
a large eddy simulation, it is found that this model hardly dissipates
any energy and thus cannot serve as a 'stand alone' SGS model. It
does transfer energy from the smallest resolved scales to the larger
resolved scales, which is useful.

Figure 2. Scatter plot of the mixed scale similarity - Smagorinsky


model prediction of the subgrid scale Reynolds stress and the exact
value obtained from a direct numerical simulation. From Bardina at
aL, 1980.
Large Eddy Simulation 127

To correct for the lack of dissipation, it is necessary to combine


the Smagorinsky and scale similarity models to produce the 'mixed'
model. This model does indeed improve the quality of the simulations
as one can see from Fig. 3.

Figure 3. Comparison of the spectra obtained from a large eddy


simulation of decaying homogeneous isotropic turbulence with and
without the mixed scale similarity - Smagorinsky model. From Bar-
dina et a/., 1980.

When the large scales are defined by the cutoff filter, ¥; = ¥,-,
and the scale similarity model produces nothing i.e., Eq. (4.15)
evaluates to zero. This difficulty can be removed by noting that, for
the Gaussian filter, filtering twice is equivalent to a single filtering
with A replaced by \/2A. This is easily mimicked for the cutoff filter
by defining the double overline filter as a cutoff filter corresponding
to this width; however, the correlation is not as good as it is when
the Gaussian filter is used, cf. Bardina et al. (1980).

4.5 Dynamic Procedure

The concepts of the preceding sections can be taken one step


further, leading to the concept of a dynamic model, an idea originally
128 J. H. Ferziger

proposed by Germane et al. (1990). It might better be called a


procedure than a model as it takes one of the models described above
as its basis. Perhaps the simplest way to explain the concept is
the following. Suppose we are doing a large eddy simulation on a
relatively fine grid. We could regard it as a DNS and use its velocity
field as the basis for a priori estimation of the subgrid scale model
parameter. This can be done at every spatial point and time step.
The scales used in such a test are, of course, the smallest resolved
scales of the LES. If we assume, as we did in constructing the scale
similarity model, that the behavior of these scales is very similar to
that of the subgrid scales, the parameter so obtained can be applied
in the subgrid scale model of the large eddy simulation itself. In this
way, a kind of self-consistent subgrid scale model is produced.
The actual procedure of Germano et al. is a bit more formal than
what we have just suggested. We now present this formal procedure.
The subgrid scale Reynolds stress that must be modeled in the actual
LES is:

The second or test filter (the one used to determine the parame-
ter) is similar to the second filter used in the scale similarity model
but is denoted by a tilde (~) to make explicit the idea that the orig-
inal and test filters need not be identical. The subgrid scale stress
that must be modeled in the test-filter level LES is:

Now let us define the large scale component of the SGS Reynolds
stress at the test filter level. This is the portion that is directly
computable from the LES field by filtering:

and is essentially the Leonard stress associated with the test filter.
Now it follows directly from these definitions that:

This is a mathematical identity that is a consequence of the defini-


tions given above; it has come to be known as Germano's identity
and provides the basis of the dynamic model.
Large Eddy Simulation 129

The basic assumption that leads to the dynamic model is that


particular model applies on both filter levels with the same value of
the parameter(s). We shall use the Smagorinsky model as an example
but there is no reason why other models cannot be used; indeed, the
mixed model has been used as a base model (Zang et al., 1993). On
the original LES level, the Smagorinsky model is:

On the test filter level the Smagorinsky model is:

Now we substitute the last two equations into Eq (4.18) to obtain:

Everything on both sides of this equation is computable from


the velocity field computed in the LES, ¥;. This means that it can
be used to compute the constant, C. However, as (4.22) represents
five independent equations, C is overdetermined. Germane et al.
suggested that the scalar product of Eq (4.22) with Sij be taken. An
improvement was made by Lilly (1992), who suggested computing
the optimum in the least squares sense. If we call the right hand side
of Eq (4.22) CM;j, one can show that this is equivalent to taking the
scalar product of Eq (4.22) with M.^ and yields:

Thus, the model parameter can be computed, at every spatial


grid point and at every time step, directly from results produced by
the LES itself. In other words, we have a kind of self-consistent or,
as it is more commonly known, dynamic model.
Although this concept is very appealing, there are significant
problems with the resulting model. First of all, it was assumed in de-
riving Eq (4.22) that the model parameter is a constant; this allowed
it to be removed from the test filter and evaluated. The resulting
expression for C, (4.23), is a function of the spatial coordinates and
time, violating an assumption made in the derivation. Furthermore,
in actual simulations, C is found to be a very rapidly varying function
which takes on large values of both signs, leading to eddy viscosities
130 J. H. Ferziger

of both signs. Although negative eddy viscosity is not prohibited (it


may be considered a way of representing backscatter), if the eddy
viscosity remains negative over too large a spatial region or for too
long a time, numerical instability may result and the simulation must
be stopped; this occurs in actual simulations. Clearly, a cure for this
problem needs to be found.
The negative eddy viscosities occur because the numerator or
denominator of Eq (4.23) may become negative. Similarly, large
values are a generally a consequence of the denominator being small.
In turn, a small value of the denominator means there is relatively
little energy in the highest wavenumbers resolved in the LES. This
further implies that there is not much energy in the subgrid scales
and therefore that the eddy viscosity should be small. One cure for
the problem is thus to simply set any eddy viscosity wj < —z/, the
molecular viscosity, equal to —v. This has been used successfully but
is not very satisfying so other methods have been developed; some
of these are discussed below.
One useful alternative is to employ averaging. For a homogeneous
flow, we may average the scalar product of Eq (4.22) with M,-j over
the entire domain prior to computing C. A more satisfactory deriva-
tion of this result is to apply the least squares method to L±j over a
finite spatial region. This leads to the replacement of Eq (4.23) with

where the brackets (<>) represent an average over the spatial re-
gion to which the least square method was applied. This technique
produces excellent results; it has been used to compute a variety of
homogeneous turbulent flows, fully developed channel flow, and tran-
sitional channel flow, all with excellent results. This version of the
dynamic model removes many of the difficulties described earlier:
• It was noticed that, in shear flows, the required value of the
Smagorinsky model parameter is much smaller than in isotropic tur-
bulence. The dynamic model produces this modification automati-
cally.
• Near walls, the value of the model parameter has to be reduced
even further, for example by using van Driest damping (4.13). In
channel flow, the averaging in Eq (4.24) is usually averaging over
planes parallel to the wall. When this is done, the model automati-
cally decreases the parameter near the wall.
Large Eddy Simulation 131

• The definition of the length scale is unclear when the filter is


not isotropic. This issue becomes moot with the dynamic model
because, if the length scale is incorrect, the model compensates by
changing the value of the parameter. Essentially, the model actually
computes the eddy viscosity, not the model parameter.
To overcome the problem created by the large negative eddy vis-
cosities generated by the simpler forms of the dynamic model several
cures have been suggested. Two were already presented above—
averaging over homogeneous directions and limiting the magnitude
of negative eddy viscosities. These are successful but the former is
available only in flows with some degree of homogeneity and the lat-
ter is not satisfying from an esthetic point of view. This has led to a
search for other methods of dealing with the problem.
One such approach is to use a combination of local spatial and
temporal averaging which are available in any flow (cf. Piomelli,
1992). These have proven successful so long as one can find a spatial
region large enough to smooth out the parameter variation but small
enough that it does not contain significant inhomogeneity.
Another approach is based on the recognition that part of the
problem arises from the removal of the model parameter from the
filter. In order to do so, it was assumed that the parameter is con-
stant but the resulting values are far from constant, invalidating the
assumption. Instead, one can regard Eq. (4.22) (with the param-
eter inside the filtering operation) as an integral equation for the
parameter. This integral equation is then solved for the parameter.
It turns out that this removes some, but not all, of the variation of
the parameter and thus does not completely cure the problem and
increases the computational effort somewhat.
A further improvement is obtained by subjecting the integral
equation referred to in the last paragraph to a constraint that the to-
tal viscosity (eddy plus molecular) be everywhere non-negative. The
resulting problem can then be solved only in a least squares sense,
leading to a constrained optimization problem. This can produce
excellent results at a cost of some increase in computer time and has
been called the dynamic localization model (Moin et a/., 1994).
Finally, we mention that the arguments on which the dynamic
model is based are not restricted to using the Smagorinsky model
as the base model. One could, instead, use the mixed Smagorinsky-
scale similarity model or one of the models presented below. The
mixed model was used in this regard by Zang et al. (1993) with
132 J. H. Ferziger

considerable success. However, the flow to which this method was


applied is a transitional flow and it is not known whether the findings
extend to fully developed flows. Ghosal et al. (1994) applied the
dynamic procedure to a one equation (or turbulent kinetic energy)
subgrid scale model, obtaining good results.

4.6 Spectral Models

The Smagorinsky and scale similarity models are not the only
ones that have been used to represent subgrid scale turbulence. For
guidance as to how improved models might be constructed, one can
turn to turbulence theories. In order to deal with the distribution
of turbulent energy over a range of length scales, in most turbulence
theories the principal variables are the Fourier transforms of the ve-
locity components:

or, more frequently, its squared amplitude, the energy spectrum:

where the integral is over all wavevectors A; on a sphere of radius


k and, as usual, a sum over the index i is implied. Use of a spec-
tral representation of the velocity field implies that these theories
are applicable only to homogeneous turbulence and, usually, only
to isotropic turbulence. Despite these limitations, they can provide
considerable insight into the issues and extensions to more complex
flows are possible.
A number of turbulence theories exist; most produce similar re-
sults. Let us begin with an observation. From the Fourier transform
of the Navier-Stokes equations, one can derive a dynamic equation for
the energy spectrum. In this equation, the advective terms, the only
non-linear terms in the equations, transfer energy from one wave-
number to another but neither produce nor destroy total energy.
The pressure terms disappear entirely; their function is to transfer
energy from one component of the turbulence (say wf(fc)) to another
(say u\(k)} and so play no role in the energy equation. The viscous
term is responsible for the dissipation:
Large Eddy Simulation 133

which acts as an energy drain on the turbulence. For further details


of these theories and the roles of the various terms, it is recommended
that the reader consult the book by Lesieur (1992).
Results derived from turbulence theories make it possible to de-
fine an effective or spectral eddy viscosity. As just noted, the non-
linear term transfers energy from one wavenumber to another. One
can imagine doing a large eddy simulation (because spectral theories
are formulated in terms of spectra, only the cutoff filter is normally
used) and ask how much energy is transferred from a given wave-
number k to wavenumbers above the cutoff. If we call this energy
transfer rate T>(k) and think of it as a dissipation at wavenumber
k, we can define an effective spectral eddy viscosity by:

An example of such an eddy viscosity, taken from Lesieur (1992), is


given in Fig. 4. The decrease in the effective viscosity at low wave-
numbers is of little consequence because little energy is transferred
from these wavenumbers to the small scales. The rise at the high
wavenurnbers is due to the local nature of the interactions in tur-
bulence; simulations have shown that incorporating the rise in the
viscosity into the subgrid scale model is capable of producing simu-
lations in which the spectrum maintains an inertial subrange shape
up to the cutoff wavenumber.
Although the rise in the eddy viscosity at high wavenumber does
not follow any simple law, if spectral computational methods are
used, a fit to the results can be constucted and used. Alternatively,
the curve can be approximated by a constant plus a term propor-
tional to some power of k. This suggests that using the Smagorinsky
model (which roughly approximates the constant component of the
eddy viscosity) together with a hyperviscosity (a dissipative term-
containing even-order velocity derivatives of order higher than sec-
ond) might be a good choice for a subgrid scale model. The simplest
choice to implement is a so-called fourth order viscosity which intro-
duces a term proportional to the fourth derivative into the Navier-
Stokes equations. Two possibilities for the added terms, which we
shall call r^4 , are:
134 J. H. Ferziger

Figure 4. Spectral eddy viscosity computed from eddy damped quasi-


normal Markovian theory (EDQNM). From Lesieur, 1992.

and

These become identical if the eddy viscosity is constant. If we wish


to model these terms in the spirit of the Smagorinsky model, dimen-
sional analysis suggests that an appropriate expression for the fourth
order viscosity might be:

The introduction of such a term into the model increases the


order of the partial differential equation and raises the possibility that
additional boundary conditions might be needed. This unpleasant
Large Eddy Simulation 135

possibility can be avoided if the new viscosity 2/4 vanishes rapidly


enough in the vicinity of the wall.
Another means of using turbulence theories is to simulate the
large scales in the usual LES manner and use a theory to describe
the subgrid scale motions statistically. A method that used EDQNM
for the subgrid scales was developed by Aupoix (1987). He obtained
good simulations of high Reynolds number flows, but only for isotro-
pic turbulence, and at an order of magnitude increase in cost relative
to the Smagorinsky model. Clearly, if this method is to be practical,
it will need to be simplified to make the cost more reasonable.
Other suggestions for models have been made. We shall not cover
most of these here because they were constructed for use in Fourier
space and cannot be easily converted to physical space models. The
absence of such a conversion possibility renders a model almost im-
possible to use with finite difference or finite volume discretizations,
restricting their usefulness. One attempt in this direction was made
by Metais and Lesieur (1992) who devised what they called a struc-
ture function model; in practice, this model is very similar to the
Smagorinsky model.
We note that it is possible to use spectral eddy viscosity models
in the dynamic context. Although, to the author's knowledge, this
has not been attempted, it seems an interesting possibility.
We end with a brief note on another approach to turbulence sim-
ulation. A number of simulations have been made which claimed to
be direct numerical simulations of complex flows; for one of many
examples, see Kawamura and Kuwahara (1985). A brief analysis
will quickly convince one that these cannot possibly be DNSs in the
sense defined earlier. These simulations use third order upwind ap-
proximations to the spatial derivatives which produce fourth order
error terms similar to the fourth order viscosities presented above.
So these simulations can be re-interpreted as large eddy simulations
with a fourth order subgrid scale model. The danger in this approach
is that the eddy viscosity is determined by the grid and the solution
may therefore depend on the grid used.

4.7 Effects of Other Strains

The models described above have been designed for flows without
'extra strains' ( e . g . , rotation, compressibility and curvature); despite
this, we have seen that the dynamic model can handle some of these
136 J. H. Ferziger

without problems.
Meteorologists and oceanographers who predict global circulation
deal with flows that are nearly two-dimensional; an eddy viscosity
is used to represent the unresolved motions. At the smallest scales,
three-dimensional equations may be used; simulations are routinely
done on several levels. A single model (with a single parameter)
that can account for phenomena at all the various scales probably
does not exist. A systematic approach is needed to build a firm
foundation for modeling in these areas. The task is difficult and
progress may come slowly. A possible approach is the following. At
the lowest level, one can simulate the small-scales e.g., the planetary
boundary layer or the ocean mixed layer and use the data produced
to construct parameterizations that to be used represent motions
that belong to the subgrid scale on the next larger scale, perhaps
the regional scale. To assure that all possibilities are included, a
range of cases containing all physically possible phenomena must be
simulated to ensure that the full range of parameters are included
in the database. By bootstrapping in this way, and allowing two-
way interaction between simulations at different scales, it may be
possible to develop methods that allow phenomena on all scales to
be predicted. It should be obvious that there are difficulties in this
scenario for which solutions are yet not available.
Extra strains can be roughly divided into two classes. Some, such
as rotation, curvature, and stratification, affect the large scales more
strongly than the small scales. In these cases, SGS models designed
for incompressible flows without extra strains can probably be used
without major modification. For example, large eddy simulations of
a stable planetary boundary layer performed with the Smagorinsky
model (Mason and Derbyshire, 1990) agree very well with both direct
simulations (Colernan et a/,, 1989) and field data.
On the other hand, for 'strains' whose action is principally in the
small scales, the situation is less clear. Compressible turbulence at
low Mach numbers can be treated with incompressible models. At
higher Mach numbers, small shock waves ('eddy shocklets') develop,
and the flow behavior can be quite different (Blaisdell et a/., 1991,
Lee et a/., 1993). If one is to do large eddy simulations of such flows,
there are two possibilities. The first approach is similar to standard
SGS modeling. In the absence of shocks, SGS models applicable to
the Favre-filtered equations can be developed in a manner analogous
to the incompressible case (Erlebacher et a/., 1992; Speziale et a/.,
Large Eddy Simulation 137

1988). We note that, as a real shock is too thin to be resolved by


the grid, the viscosity and thermal conductivity need to be increased
so that the simulated shock becomes thick enough to be resolved;
this approach has been used in many 'shock capturing' methods in
aerodynamics. The second approach is to replace the actual curved
shock by a straight one, using a subgrid scale model to account for
the larger dissipation of the curved shock; this is akin to the 'shock
fitting' approach to aerodynamics.
In combusting flows, flames are normally thin with respect to
even the smallest scales of the turbulence and LES is again very dif-
ficult. Again, one can imagine two types of LES that are similar
to the ones described for shocks above. The first is applicable only
when the chemistry is simple enough to be characterized by one or
two constants. In this case, one can modify the reaction rate and dif-
fusivity so as to increase the thickness of the flame while maintaining
the flame speed constant. In the second approach, the flame is ide-
alized as an infinitely thin sheet. The function of the SGS model is
then to account for the 'wrinkles' that occur on scales that are not
resolved by increasing the local reaction rate. Such a suggestion has
been made by Ashurst et al. (1988). A different approach based on
the use of probability density functions (which are commonly used in
computing reacting flows) was suggested by Gao and O'Brien (1993).

4.8 Other Models


It is possible to use more complex models for the subgrid scale.
Any model used in RANS calculations can be modified and adapted
as an SGS model. In particular, models based on solving partial
differential equations may be used but this has been done only a few
times.
In RANS, the next step beyond a mixing length model (which cor-
responds to the Smagorinsky model in LES) is a two-equation model
which requires equations that determine the turbulence velocity and
length scales. The model length scale is not normally an issue in SGS
modeling because a natural length scale, the filter width, is available,
so two equation models see little service in LES. One equation mod-
els can fill this role. In such a model, a partial differential equation
for the subgrid scale kinetic energy is constructed and solved. There
is only a little experience with such models. Schumann (1978) found
that it provided no significant improvement over the Smagorinsky
138 J. H. Ferziger

model for fully developed channel flow. This is no surprise because,


as we noted earlier, the major deficiency of the Srnagorinsky model is
that the principal axes of SGS Reynolds stress and rate of strain ten-
sors are not aligned and this is not addressed by the model. On the
other hand, some benefit was obtained in transitional flow; however,
the dynamic model performs as well in transitional flows so the need
for the more complex model has not been demonstrated. It is worth
mentioning that Ghosal et al. (1994) have constructed a dynamic
model that includes a differential equation for the turbulent kinetic
energy equation that appears promising.
The most complex RANS models in use today are Reynolds stress
models in which a set of equations is derived for the Reynolds stress
and the various terms are modeled. It is, of course, possible to derive
equations for the SGS Reynolds stress components as well; they are a
bit more complicated that the corresponding RANS equations due to
the properties of the filtering operator. LES with an SGS Reynolds
stress model has been tried only once, and that in a relatively early
simulation of the atmospheric boundary layer by Deardorff (1974).
He found a huge increase in the cost but almost no improvement in
the results.
In both the turbulent kinetic energy and Reynolds stress subgrid
scale models, the constants were taken from RANS models. This is
probably inappropriate as the physics of subgrid scale turbulence is
different from that for all the turbulence; the relative importance of
the various terms is probably different in the two cases. Unfortu-
nately, at present, there is little guidance for improving the models.
DNS data could probably be used to guide the development of models
but this has not yet been attempted.

5 WALL MODELS

Another issue of great importance is modeling of the flow in the


vicinity of a wall. This question receives less attention than SGS
modeling because it is not as amenable to theoretical treatment, but
it is at least as important. Before discussing the wall models, we
shall review some results of experiments and direct and large eddy
simulations with no-slip wall boundary conditions.
Shear flows near solid boundaries contain alternating thin streaks
of high- and low-speed fluid. If they are not adequately resolved, the
Large Eddy Simulation 139

turbulence energy production near the wall (which is a large fraction


of the total energy production) is underpredicted (Kim and Moin,
1986), resulting in reduction of the Reynolds stress and the skin
friction.
Simulations suggest that wall-region turbulence and the region far
from the wall are relatively loosely coupled. Chapman and Kuhn's
(1986) simulation with an artificial boundary condition imposed at
the top of the buffer layer (y+ = 100) displayed most of the charac-
teristics of the wall layer. Thus, accurate prediction of the flow near
the wall does not require accurate simulation of the outer flow.
On the other hand, Schumann (1973) and Piomelli et al. (1987)
among others have shown that, relatively crude lower boundary con-
ditions can represent the effect of the wall region in a simulation of
the central part of a channel flow. Thus, details of the flow in the
wall region need not be known in order to simulate the outer region,
i.e. either region can be well-simulated if given a reasonable approx-
imation of the conditions at the interface between it and the other
zone.
These results suggest that useful simulations can be done without
resolving the entire flow. This is important because a fine grid is
required to resolve the wall region. If it can be represented via a
model, huge savings are possible. For rough walls, one has little
choice but to use a model to represent the wall region.
DeardorfPs (1970) original model contained weaknesses that were
remedied by Schumann (1973). The latter's model, with modifica-
tions, is still widely used. It assumes that the instantaneous velocity
at the grid point nearest a wall is exactly correlated with the shear
stress at the wall point directly below it:

where y\ is the height of the first grid point, < TW > is the mean wall
shear stress, and Ui(yi) is the mean velocity at y\.
Mason and Callen (1986) assumed that the logarithmic profile
for the mean velocity in the buffer region, holds locally and instanta-
neously. This assumption is incorrect but their boundary condition
is often used by meteorologists. Piomelli et al. (1987) found it to be
inadequate for engineering applications. This is an example of how
the differing needs of two disciplines can lead to opposite conclusions
about the effectiveness of a model.
140 J. R. Ferziger

Piomelli et al. (1987) used direct simulation results to test wall


layer models including Schumann's model and two new models. The
first of these is based on the idea that Reynolds-stress producing
events do not move vertically away from the wall but, rather, at a
small angle to it. This leads to the so-called shifted model,

where A s — 3/1/cos8° is a spatial shift, 8° being the observed mean


angle of event trajectories.
The second model notes that significant Reynolds-stress contain-
ing events involve vertical motion, so the vertical component of the
velocity rather than the horizontal should be correlated with the wall
shear stress:

Both of these models give improved agreement with experiments and


direct simulations for channel flow, including cases with transpiration
and high Reynolds number flows.
In fully developed channel flow at Reynolds number 15,000, use
of these conditions reduced the time of a simulation from 100 hours
to 10 hours (Piomelli et a/., 1989) so their value is unquestionable.
Finally, we mention and interesting proposal by Bagwell et al.
(1993). They used a linear estimation method developed by Adrian
to determine the best estimate of the skin friction given the velocity
distribution at some distance from the wall. They found that the skin
friction estimate could be improved (relative to those given above)
by using a weighted average of the velocity on the computational
plane closest to the wall. A disadvantage of this method is that
the two point correlation, which becomes a complicated function of
the coordinates in complex flows, and especially near separation and
reattachment, is required.
All of the models of this section have been applied only to flows on
flat walls with mild pressure gradients. They are almost certainly in-
adequate for separated flows (with or without reattachment), or flows
over complex-shaped walls. They may work in three-dimensional
boundary layers because the direction of the horizontal component
of the velocity changes slowly with distance from the wall in the
lower part of the flow. Because experimental data are scarce and
Large Eddy Simulation 141

lack detail, the development of trustworthy methods for simulating


these flows will probably require simulations with no-slip conditions.

6 NUMERICAL METHODS

A wide variety of numerical methods have been employed in large


eddy simulation. Almost any method of computational fluid dynam-
ics can be used. Because these methods are adequately described
elsewhere (see, for example, Ferziger and Peric, 1993), we shall not
describe particular methods here. Instead, a few generalities about
issues peculiar to LES will be discussed.
The most important requirements on numerical methods for LES
arise from the goal of producing an accurate realization of a flow
that contains a wide range of length scales. The need to produce
a time history means that techniques used for steady flows must be
rejected. Time accuracy requires a small time step and it is important
to know whether the time-advance method is stable for the time step
demanded by accuracy. This is generally the case so most simulations
use explicit time advance methods. An exception occurs close to
walls where fine grids must be used in the normal direction and
instability may arise from the viscous terms; in this case, only the
viscous terms involving derivatives normal to the wall are treated
implicitly. The numerical methods most commonly used in LES are
of second to fourth order; Runge-Kutta methods have been used most
commonly but others, such as Adams-Bashforth and leapfrog have
found application.
The need to handle a wide range of length scales means renders
some concepts of computational fluid dynamics relatively unimpor-
tant. The most common means of describing the accuracy of a spatial
discretization method is its order, a number that describes the rate
at which the discretization error decreases as the grid size goes to
zero. To see why this is not applicable in LES, it is useful to think in
terms of the Fourier decomposition of the velocity field. A discrete
version of Eq (4.25) is

The highest wavenurnber k that can be resolved on a grid of size


Ax is 7T/A:r, so it is sufficient to restrict the range of k to {0, TT/Ao:}.
142 J. H. Ferziger

The derivative of (6.1) may be taken term by term and, if we ignore


the effect of boundary conditions, it is sufficient to consider the ef-
fect of the discretization on a single Fourier mode, elkx'. The exact
derivative is, of course, ikelkx. All discrete approximations replace
this by ikejjelkx where kejf is called the effective wavenumber. For
example, the central difference approximation:

when applied to e , gives:

so

for this method. For small &, the Taylor series approximation:

shows the second order nature of the approximation. A plot of keff


is given in Fig. 5 which shows that the Taylor series approximation
is useful only for k < 7r/2Ax, the first half of the wavenumber range
of interest. Other discretizations give different expressions for the
effective wavenumber. Upwind approximations give effective wave-
numbers that are complex, reflecting the dissipative nature of the
discretization error for these schemes. A similar treatment of the er-
rors in approximations for the second derivative is easily constructed
but will not be described here.
The problem in LES is that the spectrum of the solution (its
distribution over wavenumber) covers a significant part of the wave-
number range {0,?r/Aa;}. The order of the method is no longer
sufficient to define the accuracy of a scheme. A better measure of
the error is:

Again, similar expressions can be given for the second derivative.


Using the measure (6.6), Cain e.t al. (1981) found, for a spectrum
Large Eddy Simulation 143

Figure 5. Effective wavenumber of the central difference approxima-


tion to the first derivative.

typical of isotropic turbulence, that a fourth order method had half


the error of a second order method, much more than most people
would have anticipated.
The final point is that the methods and step sizes in time and
space need to be chosen together. The errors made in the spatial and
temporal discretizations ought to be as nearly equal as possible i.e.
they should be balanced. This is not possible in detail but, if this
is not done, one is using too fine a step in one of the independent
variables and the simulation could be made with little loss of accuracy
at lower cost.

7 ACCOMPLISHMENTS AND PROSPECTS

Large eddy simulation has been applied to a range of flows too


large to be covered in a single paper; there no need to do so. The pur-
pose of many simulations was to study the physics and modeling of
flows and the results produced by DNS and LES are often treated as
experimental data. For that reason, it makes more sense to consider
the results together with experimental data on a flow-by-flow basis.
We shall, therefore, give only a short overview of the kinds of flows
144 J. H. Ferziger

that have been treated with LES, a snapshot of the state-of-the-art,


and a discussion of what may be possible in the next few years.
In the early days (in the 1970's), LES was used, at least in en-
gineering, for investigating simple flows in order to understand the
physics of turbulence and the accuracy of RANS models. Flows
that were treated in this way included all the homogeneous flows,
plane channel flow, and free shear flows that are inhomogeneous in
one direction (2D mixing layers, wakes, and jets). Computers have
now become sufficiently large and fast that these flows can be dealt
with by direct numerical simulation. Since DNS does not have any
uncertainty arising from subgrid scale modeling, it is the preferred
technique for this kind of investigation and should be used when-
ever possible. In the early and mid-1980's LES almost fell into dis-
use; however, interest was rekindled (Hussaini et al., 1990; Reynolds,
1990), and at the outset of the 1990's application of the method was
on the rise.
LES is now being applied to flows that remain beyond the reach
of DNS. A very important engineering issue is that of flow separation
and reattachment, phenomena that occur in many technological flows
and, with few exceptions, are not well predicted by RANS methods.
They are also, at present, outside the reach of DNS. The simplest
separating flow is the backward facing step in which a plane channel
flow encounters a sudden expansion on one wall of the channel. DNS
of this flow was performed by Le and Moin (1993) and although
good results were obtained, 1100 hours were required on a single
processor Cray-YMP. An LES of this flow by Akselvoll and Moin
(1993) using the same computer required only about 30 hours. The
latter figure, while much more than a working engineer would, care
to pay for a simulation, brings the cost to a point at which it may
be sensible to do a simulation occasionally to check the validity of
RANS results and/or the models used in RANS calculations. Other
flows of this kind which have been simulated recently include the two
dimensional obstacle (Yang and Ferziger, 1993) and flow over a cube
(Mauch, 1991; Shah and Ferziger, 1994). The former introduces a
flow-determined separation not found in the backward facing step
flow. The latter introduces three dimensional separation.
For the near future, a sensible role for LES to play is as a check on
the validity of RANS turbulence models and predictions for complex
flows. It will be possible to perform large eddy simulations of some
flows of engineering interest but the method will remain too costly
Large Eddy Simulation 145

for routine engineering use for a long time to come. Further, LES
can and will be used as a complement and partial substitute for
experimental testing. An occasional LES can be compared to RANS
predictions to test the adequacy of a design and/or the methods used
to develop it. It will also continue to be used to directly test the
accuracy of RANS models, a role that it has played with distinction
throughout its history. By using LES to tune RANS models, it should
be possible to obtain most of the benefits of LES at a small fraction
of the cost.
Flows that may be good candidates for LES in the near future
include the turbine blade passage and the internal combustion engine
cylinder. These are both relatively low Reynolds number flows and
of obvious technological importance. Both of these flows also contain
many extra strains that renders the development of RANS models
for them exceedingly difficult, making the possibility of using LES
directly in the design process and interesting one.
A word of caution is necessary. LES and its subgrid scale models
have been validated only for relatively simple flows at fairly low Rey-
nolds numbers. In these flows, most of the energy is in the resolved
scales and, even if the subgrid scale model is not very accurate, its ef-
fect on the results may not be too important. If one uses the success
of these simulations as a justification for applying LES to much more
complex flows, although reasonable looking results may be obtained,
placing one's trust in them may be risky. The leap is simply too great
to allow expectation of success in this kind of endeavor. Simulations
of this kind have been made but, in the author's opinion, they have
been premature and their value is questionable.
It is also important that the goal of a simulation be defined in
advance. Doing a simulation merely to show that it can be done is
of limited value. It is known in advance that it can be done and, if
enough resources are deployed, good results will be obtained. The
value is in learning about the physical nature of the flow, how it may
be modeled, or, perhaps, in making a contribution to the improve-
ment of a design.
It would be very valuable to have models that eliminate the need
to specify no-slip conditions at a wall. Boundary conditions of this
kind exist for attached flows and were discussed above. What is
not known is whether conditions of this kind can be constructed for
separated flows. Doing so for RANS models has proved exceedingly
difficult and there is no reason not to expect the task to be at least
146 J. H. Ferziger

as difficult for LES.


In flows in complex geometries, it is impossible to construct opti-
mum grids prior to the calculation, even for steady flows; one simply
cannot know in advance where the maximum resolution is required.
To compensate for the absence of this information, methods which
modify the grid as the solution procedure converges have been devel-
oped. The author's favorite such method is one developed by Berger
and Oliger (1984). This method has been adapted for elliptic flows by
Caruso et al. (1985). It also combines well with the multigrid solu-
tion procedure, one of the best methods for solving elliptic problems
(Thompson and Ferziger, 1989).
It is difficult to specify the grid requirements for the flows men-
tioned above at this time. The numbers will depend on whether
techniques of the kinds described in the last two paragraphs can be
developed. All that is certain is that they will need to be larger
than those now in use. The large parallel machines now coming
onto the scene will allow simulations to be done on grids containing
512 X 512 X 512 points (or other grids containing roughly the same
number of points). It is conceivable that simulations of turbine blade
passages and engine cylinders can be done with these grids.

8 COHERENT STRUCTURE CAPTURING

8.1 The Concept

Up to the present, researchers have attempted to build LES from


the ground up. The idea is to start with simple flows (preferably ones
that can be treated with DNS), use them to learn about SGS mod-
eling, and then go on to increasingly more complex flows. To date,
most researchers have taken care to simulate only flows in which a
large fraction of the energy of the turbulence can be resolved. The
importance of the model is thereby minimized and good results have
been achieved. This success does not assure equal success for LES of
complex flows; this might be the case if sufficient computer resources
were available but, for most flows of technological interest, there is
little possibility this will happen in the foreseeable future. Further-
more, the objective is usually to obtain just a few selected properties
of the flow at minimum cost. That being the case, LES and DNS are
best not used as everyday tools.
Large Eddy Simulation 147

A better choice for the near term is to perform LES and/or DNS
on 'building block' flows, i.e. flows that are structurally similar to
the ones of actual interest. From the results of such simulations,
RANS models that can be applied to the more complex flows can
be validated and improved. RANS computations can then be used
as the everyday tool. LES need be performed only when there are
significant changes in the design or as an occasional check on the
validity of the RANS results.
As noted earlier, there have been attempts at large-eddy and
direct numerical simulation of complex flows. Unfortunately, in most
of these, the subgrid scale model was uncontrolled and the results are
of uncertain value. This appears to be a case of reaching too far too
fast; we shall not present examples here.
Since answers to questions involving technologically significant
flows are required, the following questions arise. Is there a method
that will enable more complex flows to be simulated on available
machines? Are there flows of importance that are good candidates
for simulation via LES in the relatively near future?
The answer to these questions appear to be a qualified yes. Other
than the flows mentioned earlier, particularly good candidates are
flows in which there are a small number of important, energetic, and
easily identified coherent structures. In all the cases that have been
suggested, the large structures are vortices. Let us consider two such
cases.
Flows over bluff bodies usually produce strong vortices in their
wakes. The vortices produce strong fluctuating forces on the body in
both the streainwise and spanwise directions whose prediction is very
important in many applications. The latter include buildings (wind
engineering) and ocean platforms, among others. If the vortices are
sufficiently larger than the bulk of the motions that constitute the
'turbulence' it should be possible to construct a filter that allows the
vortices to be retained in the resolved field while removing all of the
smaller scale motions. We have called a method that accomplishes
this 'coherent structure capturing' or CHC (Ferziger, 1993). The au-
thor and others earlier called it very large eddy simulation or VLES,
a term we now find less descriptive. Methods of this kind were sug-
gested a long time ago (Ferziger, 1983) but deliberate simulations
of this type do not appear to have been attempted other than a few
cases which apparently gave unsteady results when a steady flow was
expected. An exception to this might be the recent work of Orszag
148 J. H. Ferziger

discussed in Chapter 4.
The cylinders of internal combustion engines provide another ex-
ample. This flow is inherently unsteady, so there is no possibility of
modeling it as a steady flow. Several interesting issues arise which
lead to the following questions. What does RANS mean in such a
flow and how should RANS results be compared with experimental
results? Since the flow is unsteady, LES can only produce a single re-
alization; can such a simulation provide sufficient information about
the flow? A partial answer to the first question is that the RANS
mean velocity should probably be defined as an average over many
cycles and the turbulence as the deviation from the multi-cycle aver-
age. LES should simulate a single cycle. To see what the differences
are consider the following. After the intake stroke, the flow contains
a strong vortex whose location, size, and strength varies from cycle
to cycle; this vortex is important to engine behavior. In a RANS
calculation, the result should contain an average vortex, one that is
relatively large and of average circulation. In LES, the vortex should
be smaller, of similar circulation, but its location should vary from
realization to realization, it should be possible to construct a filter
that can separate the vortex from the rest of the turbulence field.

8.2 Modeling Issues


The models to be used in CHC should be different from those
used in both RANS and LES. Indeed one needs to be very careful
and considerable experience is probably required before this kind of
simulation can be trusted.
According to the Smagorinsky model, the length scale to be used
in LES is the filter width, A. But, in CHC, the filter width may
become quite large; it may indeed become larger than the length
scale used in RANS models, which is an approximation to the integral
scale, L. If this is the case, the LES viscosity could exceed the RANS
viscosity, in violation of the concept that the RANS viscosity should
be large enough to remove all of the unsteadiness (at least all of it
that is considered turbulence) from the flow. This situation should
not be allowed to arise. To prevent it from occurring, it may become
necessary to introduce an equation for the length scale to be used
in the subgrid scale model and LES may then inherit many of the
difficulties that RANS models have with length scale modeling. The
only thing that is clear is that considerable effort will be needed to
Large Eddy Simulation 149

make this approach work.

9 CONCLUSIONS AND RECOMMENDATIONS

After years of being regarded as a method of second choice rela-


tive to direct simulation, LES is receiving increased attention. The
principal reasons are dissatisfaction with the performance of RANS
turbulence models on the one hand and the inherent limitations and
cost of direct simulation on the other.
Improved models for both the small-scale turbulence and the wall
layer are also needed if LES is to become a useful engineering tool.
The dynamic model offers promise of removing many of the difficul-
ties that have plagued LES and to give it an important advantage
with respect to RANS modeling. Improved models for the wall re-
gion, especially for separating and reattaching flows, are needed just
as badly and are an important subject for future research.
However, if LES is to prove useful in truly complex high Reynolds
number flows, a great deal more work may be needed. For the near
future, it is probably best to use LES to understand the physical
nature of the flow and to tune RANS turbulence models in a way
that will allow them to produce more accurate predictions.
Some 'extra strains,' namely those that mainly affect the large
scales, appear to be relatively easy to incorporate into large eddy
simulations; little, if any, modification of the SGS models is required.
Others, which act on scales smaller than the Kolmogoroff scale, for
example, compressibility, may require significant changes in the SGS
model.

ACKNOWLEDGMENTS

The author has been active in this field for over twenty years and
the list of people who have helped him is too long to be covered in a
short acknowledgment. I will, therefore, limit myself to mentioning a
few people who have influenced my thinking in this area in the past
few years. These include my colleagues: Profs. Peter Bradshaw, Jef-
frey Koseff, Parviz Moin, Stephen Monismith and William Reynolds,
and my students Matthew Bohnert and Kishan Shah. The support
received from a number of agencies over the year has also been very
150 J. H. Ferziger

important; these include NASA, the Office of Naval Research and


the Air Force Office of Scientific Research.

10 REFERENCES
Akselvoll, K. and Moin, P., 1993."Large eddy simulation of a back-
ward facing step flow," ASME Fluids Engr. Conf., Washington,
DC, June.

Ashurst, W.T., Sivashinsky, G.I., and Yakhot, V., 1988. "Flame


front propagation in nonsteady hydrodynamic fields," Comb.
Sci. Tech., 62, p. 273.

Aupoix, B., 1987. "Application de modeles dans 1'espace spectral a


d'autres niveaux de fermature en turbulence homogene," dis-
sertation, Universite Claude Bernard-Lyon I.

Bagwell, T.G., Adrian, R.J., Moser, R.D. and Kim, J., 1993. "
Improved approximation of wall shear stress boundary condi-
tion for large eddy simulation," in Near Wall Turbulent Flows
(R.M.C. So, C.G. Speziale and B.E. Launder eds.), Elsevier.

Bardina, J., Ferziger, J.H., and Reynolds, W.C., 1980. "Improved


subgrid models for large eddy simulation," AIAA paper 80-
1357.

Berger, M.J. and Oliger, J., 1984. "Adaptive mesh refinement for
hyperbolic partial differential equations," J. Comp. Phys,, 53,
p. 484.

Blaisdell, G.A., Mansour, N.N., and Reynolds, W.C., 1991. "Nu-


merical simulations of compressible homogeneous turbulence,"
Report TF-50, Stanford University, Dept. of Mechanical Engi-
neering.

Bohnert, M.J. and Ferziger, J.H., 1993. "The dynamic subgrid scale
model in large eddy simulation of the turbulent Ekman layer,"
in Engineering Turbulence Modeling and Experiments 2, W.
Rodi and F. Martelli eds., Elsevier.
Large Eddy Simulation 151

Cain, A.B., Reynolds, W.C., and Ferziger, J.H., 1981. "A three
dimensional simulation of transition and early turbulence in a
time developing mixing layer," Report TF-14, Dept. of Mech.
Engr., Stanford Univ.
Caruso, S.C., Ferziger, J.H., and Oliger, J., 1985. "An adaptive grid
method for incompressible flows," Report TF-23, Dept. Mech.
Engr., Stanford Univ.
Chapman, D.R. and Kuhn, G.D., 1986. "The limiting behavior of
turbulence near a wall," J. Fluid Mech., 70, pp. 265-92.
Clark, R.A., Ferziger, J.H., and Reynolds, W.C., 1979. "Evalua-
tion of subgrid scale turbulence models using a fully simulated
turbulent flow," J. Fluid Mech., 91, p. 92.
Coleman, G.N., Ferziger, J.H. and Spalart, P.R., 1990. "A nu-
merical study of the stratified turbulent Ekman layer," Rept.
TF-48, Thermosciences Div., Dept. of Mech. Engr., Stanford
Univ.
Deardorff, J.W., 1970. "A numerical study of three-dimensional
turbulent channel flow at large Reynolds number," J. Fluid
Mech., 41, p. 452.
Deardorff, J.W., 1974. "Three dimensional numerical modeling of
the planetary boundary layer," Boundary Layer Meteorology,
1, p. 191.
Erlebacher, G., Hussaini, M.Y., Speziale, C.G., and Zang, T.A.,
1992. "Toward the large-eddy simulation of compressible tur-
bulent flows," J. Fluid Mech., 238, p. 155.
Ferziger, J.H., 1983. "Higher level simulations of turbulent flow," in
Computational Methods for Turbulent, Transonic, and Viscous
Flows, J.-A. Essers, ed., Hemisphere.
Ferziger, J.H., 1993. "Simulation of complex turbulent flows: recent
advances and prospects in wind engineering," in Computational
Wind Engineering 1, S. Murakami ed., Elsevier.
Ferziger, J.H. and Peric, M., 1994. "Computational methods for
incompressible flow," in Computational Fluid Dynamics, (M.
Lesieur and J. Zinn-Justin eds.), Elsevier.
152 J. E. Ferziger

Gao, F. and O'Brien, 1993. "A large eddy simulation scheme for
turbulent reacting flows," Phys. Fluids A 5, p. 1282.
Germane, M., Piomelli, U., Moin, P., and Cabot, W.H., 1990. "A
dynamic subgrid scale eddy viscosity model," Proc. Summer
Workshop, Center for Turbulence Research, Stanford CA.
Ghosal, S., Lund, T.S., Moin, P., and Akselvoll, K., 1994. "A dy-
namic localization model for large eddy simulation of turbulent
flows," submitted to /. Fluid Mech..
Horiuti, K., 1990. "Higher order terms in anisotropic representation
of the Reynolds stress," Phys. Fluids A 2, p. 1708.
Hussaini, M. Y., Speziale, C. G., and Zang, T. A., 1990. "The
potential and limitations of direct and large eddy simulations,"
Whither Turbulence? Turbulence at the Crossroads, Springer-
Verlag, Berlin, pp. 354-368.
Kawamura, T., and Kuwahara, K., 1985. "Direct simulation of
a turbulent inner flow by a finite difference method," AIAA
paper 85-0376.
Le, H., 1993. "Direct numerical solution of turbulent flow over a
backward facing step," Dissertation, Dept. of Mech. Engr.,
Stanford Univ.
Lee, S., Lele, S.K., and Moin, P., 1993. "Simulation of spatially
evolving turbulence and the applicability of Taylor's hypothesis
to compressible flow," Phys. Fluids A 5, pp. 1521-1530.
Leonard, A., 1974. "Energy cascade in large eddy simulations of
turbulent fluid flows," Adv. in Geophys., ISA, p. 237.
Lesieur, M., 1991. "Turbulence in fluids," Second ed., Kluwer, Dor-
drecht.
Leslie, D.C., 1973. "Theories of turbulence," Oxford U. Press.
Lilly, D.K., 1965. "On the computational stability of numerical so-
lutions of time-dependent, nonlinear, geophysical fluid dynamic
problems," Mon. Wea. Rev., 93, p. 11.
Lilly, D.K., 1992. "A proposed modification of the Germano subgrid
scale closure method," Phys. Fluids A 4, p. 633.
Large Eddy Simulation 153

Mason P.J., 1989. "Large eddy simulation of the convective atmo-


spheric boundary layer," /. Atmos. Sci., 46, p. 1492.
Mason, P.J., and Calien, N.S., 1986. "On the magnitude of the sub-
grid scale eddy-coefficient in large eddy simulation of turbulent
channel flow," J. Fluid Mech., 162, p. 439.
Mason, P.J., and Derbyshire, S.H., 1990. "Large eddy simulation
of the stably stratified atmospheric boundary layer," Bound.-
Layer MeteoroL, 53, p. 117.
Mauch, H., 1991. "Berechnung der 3-D umstroemung eines quadr-
erfoermigen koerpers in kanal," Dissertation, Univ. Karlsruhe.
McMillan, O.J. and Ferziger, J.H., 1980. "Tests of new subgrid
scale models in strained turbulence," AIAA paper 80-1339.
Metais, O. and Lesieur, M., 1992. "Spectral large eddy simulations
of isotropic and stably stratified turbulence," J. Fluid Mech.,
239, p. 157.
Moin, P. and Kim, J., 1982. "Large eddy simulation of turbulent
channel flow," J. Fluid Mech., 118, p. 341.
Moin, P., Carati, D., Lund, T., Ghosal S., and Akselvoll, K., 1994.
"Developments and applications of dynamic models for large
eddy simulation of complex flows," 74th AGARD Fluid Dy-
namics Panel, Chania, Greece, April.
Piomelli, U., Ferziger, J.H. Moin, P., and Kim, J., 1989. "New
approximate boundary conditions for large eddy simulations of
wall-bounded flows," Phys. Fluids A, 1, p. 1061.
Piomelli, U., Zang, T.A., Speziale, C.G. and Hussaini, M.Y., 1990.
"On the large-eddy simulation of transitional wall-bounded
flows," Phys. Fluids A, 2(2), p. 257.
Piomelli, U., 1991. "Local space-time averaging in the dynamic
subgrid scale model," Bull. Amer. Phys. Soc., Vol. 35, No.
10.
Reynolds, W. C., 1990. "The potential and limitations of direct
and large eddy simulations," Whither Turbulence? Turbulence
at the Crossroads, Springer-Verlag, Berlin, pp. 313-343.
154 J. H. Ferziger

Rodi, W. and Mansour, N.N., 1992. "Modeling the dissipation rate


with the aid of direct simulation data," Studies in Turbulence,
(T. Gatski et al. eds.), Springer.
Rogallo, R.S. and Moin, P., 1984. "Numerical simulation of turbu-
lent flows," Ann. Revs. Fluid Mech., Annual Reviews.

Schumann, U., 1973. "Em untersuchung ueber der berechnung


der turbulent stroemungen im platten- und ringspalt-kanalen,"
Dissertation, University Karlsruhe.
Shah, K. and Ferziger, J.H., 1994."Simulation of flow over a wall-
mounted cube," in preparation.
Smagorinsky, J., 1963. "General circulation experiments with the
primitive equations, part I: The basic experiment," Mon. Wea.
Rev., 91, p. 99.
Speziale, C.G., 1987. "On nonlinear k — I and k — e models of
turbulence," /. Fluid Mech., 178, p. 459.
Speziale, C.G., Erlebacher, G., Zang, T.A. and Hussaini, M.Y.,
1988. "The subgrid-scale modeling of compressible turbulence,"
Phys. Fluids, 31, p. 940.
Tennekes, H., and Lumley, J.L., 1976. "A first course in turbu-
lence," MIT Press.
Thompson, M.C. and Ferziger, J.H., 1989. "A multigrid adaptive
method for incompressible flows," J. Comp. Phys., 82, p. 94.
Yakhot, V. and Orszag, S.O., 1986. "Renormalization group meth-
ods in turbulence," J. Sci. Comp., 1, p. 3.
Yang, K.S. and Ferziger, J.H., 1992. "Large eddy simulation of
turbulent flow with a surface-mounted obstacle," Fifth Asian
Cong. Fluid Mech., Dae Jon, Korea.
Zang, Y., Street, R.L. and Koseff, J.R., 1993. "A dynamic mixed
subgrid scale model and its application to turbulent recirculat-
ing flows," Phys. Fluids, 5, p. 3186.
Chapter 4

INTRODUCTION TO
RENORMALIZATION
GROUP MODELING OF
TURBULENCE

Steven A. Orszag
I. Staroselsky, W. S. Flannery, Y. Zhang

1 INTRODUCTION

The renormalization group (RNG) and related e-expansion meth-


ods are a powerful technique that allow the systematic derivation of
coarse-grained equations of motion for turbulent flows and, in par-
ticular, the derivation of sophisticated turbulence models based on
the fundamental underlying physics. The RNG method provides a
convenient calculus for the analysis of complex physical effects in
complex flows. The details of the RNG method applied to fluid
mechanics differ in some crucial respects from how renormalization
group techniques are applied to field theories in other branches of
physics. At the present time, the RNG methods for fluid dynamics
are by no means rigorously justified, so their utility must be based
on the quality and quantity of results to which they lead. In this
paper we discuss the basis for the RNG method and then illustrate
its application to a variety of turbulent flow problems, emphasizing
those points where further analysis is needed.

155
156 S. A. Orszag et al.

The application of a field-theoretic method like the RNG tech-


nique to turbulence is based on the fundamental assumption of uni-
versality of small scales in turbulent flows. Such universal behavior
was first suggested over 50 years ago in the seminal work of A. N. Kol-
mogorov who argued that the small-scale spectrum of incompressible
turbulence is universal and characterized by two numbers, the rate
of energy dissipation £ per unit mass and the kinematic viscosity
v. In fact, Kolmogorov predicted that the energy spectral density of
turbulence has the universal form

where k is the wavenumber and

is called the dissipation wavenumber and defines a scale below which


turbulent eddies are directly affected by viscosity. Kolmogorov ar-
gued that this form of the energy spectrum would apply at scales
small compared to those characterizing the inhomogeneities of the
average flow. It is implicit in the Kolmogorov theory that the rate
of energy dissipation £ is essentially independent of the viscosity v]
this first fundamental law of turbulence implies that £ is indepen-
dent of Reynolds number (R = VL/v, where V is the rms velocity
of the turbulence and L is an appropriate large scale) and therefore
6 = O(l) as R —>• oo. If this first fundamental law of turbulence is
true, then (2) shows that kd = 0(#3/4).
The Kolmogorov universal spectrum has two interesting special
cases. First, if k < kd, E(k] = CK£2/3k~5/3, where CK = .F(O),
so the energy spectrum is a power-law with the universal exponent
—5/3. In this so-called inertial range the energy spectrum is inde-
pendent of viscosity. Second, if k >• kd it can be argued that F(x)
must be of the form Cixae~C2X, so the energy spectrum decays ex-
ponentially fast with increasing k in this so-called far dissipation
range (Chen et al., 1993). The wavenumber k^ separates inertial-
range scales for which viscosity is not important and dissipation range
scales.
Experimental data demonstrating that £ — O(l) is still very lim-
ited. The situation is somewhat better for the Kolmogorov spectrum
(1) where a variety of experimental measurements have shown that
Renormalization Group 157

(1) is at least approximately satisfied. Measurements of the energy


spectrum in the inertial range demonstrate that deviations from Kol-
mogorov's 5/3 power law are small, although these deviations have
dominated turbulence research for over two decades now through the
search for so-called intermittency corrections to Kolmogorov's the-
ory. The RNG technique discussed in this paper is based on the
universality of the Kolmogorov spectrum. In contrast to the fluctua-
tion theory of phase transitions, the RNG theory of turbulence is not
used to calculate scaling exponents (like -5/3), but rather to calcu-
late amplitudes (like C/<). If corrections to perfect Kolmogorov scal-
ing are established at some future time, it will be possible to redo the
RNG theory account for these effects in a systematic way. The RNG
theory of turbulence does not prove the validity of the Kolmogorov
theory; rather, the RNG theory provides quantitative predictions of
the behavior of turbulent flows assuming that Kolmogorov's theory
holds.
The goal of a statistical theory of incompressible turbulence is to
give a probabilistic description of a solenoidal velocity field v(x, t)
governed by the incompressible Navier-Stokes equations (NSE):

supplemented by the boundary conditions that the velocity v at a


boundary or interface matches the velocity of the boundary or inter-
face. Ideally, such a statistical description of turbulence would yield
the multi-point probability distribution function, which is equiva-
lent to knowledge of all correlation functions of the system. A more
limited goal would be to obtain only certain velocity correlation func-
tions, say, two-point correlation functions, which include information
on the energy spectrum as well as enough information for nearly all
turbulence modeling requirements. Unfortunately, whereas the NSE
are an explicit, well-defined system of partial differential equations,
it is difficult to develop an efficient workable formalism to perform
statistical averages on them that allows the development of predic-
tive equations for the statistical properties of turbulence. In addition
to the formal problem of developing efficient and effective averaging
techniques, it is necessary to build a formalism that reflects the uni-
versality of the small scales of the flow; it is necessary to find some
158 S. A. Orszag et al.

way to de-emphasize the non-universal aspects of turbulence at large


scales imposed by boundary and initial conditions. This is not a
trivial matter because it is intuitively clear that boundary and initial
conditions are directly responsible for turbulent energy production.
The goal of the RNG theory is, as mentioned above, the develop-
ment of a quantitative description of small scales in turbulent flows.
With such a quantitative theory of small scales available, it is pos-
sible to "remove" the small scales from the turbulence dynamics,
thereby deriving effective equations of motion for large scales and,
in particular, turbulence models for the prediction of large-scale flow
properties. At finite Reynolds numbers, the 5/3 law holds in the
limited wavenumber range 1/L <C k <C k^. This range is limited be-
cause the boundary conditions are imposed at a finite scale of order
L and the Reynolds number is finite.
To pursue the RNG analysis we first attempt to remove the con-
straint due to boundary and initial conditions at the finite scale L
by, in effect, taking the limit L —>• oo. We do this by mimicking the
effect of initial and boundary conditions by imposing an artificial
random force f in the NSE:

Once an appropriate force that represents random forcing at in-


finitely large scales is imposed, the RNG theory proceeds to describe
the resulting infinitely long inertial range. In particular, the RNG
method is based on the assumption that the inertial range dynamics
is invariant under rescaling transformations, an assumption justified
by the infinite extent of this range and its independence of viscous
dissipation effects. As part of the RNG analysis, it is assumed that
the turbulent dynamics in the inertial range can be described by a fi-
nite number of effective transport coefficients, like eddy viscosity for
momentum and eddy diffusivity for heat and mass transport. While
the rigorous basis of these steps is individually still open to question,
the results to which they lead appear to have a remarkably close
agreement with reality.
Renorm&lization Group 159

2 PERTURBATION THEORY FOR THE


NAVIER-STOKES EQUATIONS

In this Section we introduce statistical perturbation methods to


organize the solution of the Navier-Stokes equations as an expansion
in powers of effective Reynolds number. We do this in the context of
an unbounded flow in d dimensions but, for analytical convenience,
we can equivalently view this flow as being confined to a cube hav-
ing side £ on which periodic boundary conditions are imposed and
then the implicit limit £ —> oo being taken at the conclusion of all
calculations. The Fourier transform of the Navier-Stokes equations
(3') is

where k = (k,w), u(fc) [f(A;)j is the space-time Fourier transform of


v(x,*)[f(x,i)],

and dq denotes the (d+ 1)-dimensional integral over the wavevector


and frequency components of q. In (5), AQ = 1 is a parameter that
allows us to conveniently determine the order of perturbation theory
to which we are working. The projection operator P;j(k) ensures the
incompressibility of the flow; the term kikj/k2 in P,-j represents the
pressure gradient Vp.
It is convenient to adopt a symbolic notation that absorbs both
the wavenumber and vectorial components into a single numerical
symbol, e.g., •u(l) = Uj1(ki) and where summation over repeated in-
dices and integration over wavevectors is implied. Then (5) becomes

where the coupling operator g is defined as

The formal solution of the nonlinear stochastic partial differential


equation (7) is given by the Neumann series obtained by iterating
(7):
160 S. A. Orszag et al.

Equation (9) expresses the fluctuating velocity u in terms of a func-


tional power series in the random force /.
The goal of statistical perturbation theory is to obtain correla-
tion functions of the random field u by averaging the series (9) over
fluctuations of the random noise /. Of particular interest is the two-
point correlation function (7(12) =< u(l)w(2) > which is directly
related to the energy spectrum and the nonlinear Green's (impulse
response) function C?(12) =< <!m(l)/<5/(2) > which characterizes the
full nonlinear system response to small perturbations. The symbol
<> indicates an ensemble average over all realizations of the random
field /. The properties of the random noise / are chosen to facilitate
the evaluation of the terms in the expansions of U and G. Indeed, it
is convenient to choose the force / to be zero-mean, divergence-free,
white-in-time, and gaussian of the form /(I) = P,- 1)t - 2 p(2) where p is a
random noise with the multi-point probability distribution function

The two-point correlation function of the random noise / is then

where

and the ^-function accounts for translational invariance in time and


space:

With this choice of gaussian random noise, multi-point averages


reduce to sums over products of pair correlations (10). In this case,
the solution to the linearized Navier-Stokes equations leads to the
two-point velocity correlation function

By applying these rules for averaging over the gaussian random noise
/ to the velocity field and the nonlinear Green's function G, we ob-
tain the full perturbation series for U and. G. The terms in these
Renormaliz&tion Group 161

series are usually classified based on the number of vertex opera-


tors g, which is also twice the number of integrations over space-
time wavenumbers. These integrations may or may not be indepen-
dent, depending upon the detailed structure of the term. Certain
terms may be identified as vanishing or non-existent, such as those
containing 0(123)^(23) oc Pili2ia(ki)6(l + 2 + 3)A'(23)<S(2 + 3) oc
P t - l i 2 i 3 (fci)tf(l) = 0, or 0(123)G(23), or any "closed cycle" of Green's
functions of the type G(12)G(23)...G(nl). Most generally, the full
series can be conveniently represented using diagram methods [Wyld,
1961; Kraichnan, 1961].
The lowest order nonvanishing terms appearing in the series for
U and G are those of second order

It should be emphasized that these elementary terms already provide


basic information about large-scale properties of the system. Analyz-
ing the divergence of the associated integrals in wavenumber space, it
is immediatedly found, for example, that the lowest order nonlinear
correction to the Green's function is (omitting tensorial indices)

In order to describe fully developed turbulence, the forcing would


have been chosen at the largest spatial scales, i.e. D(q) a £ d (q).
This would immediately yield a formally infinite result ~ / 6(q)dq/q^.
However, an infinite series of formally infinite terms may well give
finite and meaningful results, if summed properly. This motivates us
not to deal with the delta-functions directly but rather to introduce
an auxiliary power-law stirring force D(k] — Dok~y (see Orszag et
al., 1993a for more details). Then (14) becomes

Equation (15) hints at the importance of the parameter e = 4 + y — d


which controls the infrared divergence of the perturbation series.
162 S. A. Orsza,g et al.

3 RENORMALIZATION GROUP METHOD FOR


RESUMMATION OF DIVERGENT SERIES

Whereas there are methods of direct resummation of perturba-


tion series of logarithmically diverging terms, there is no general
method to handle series with power-law divergences. However, there
is a different route, viz. RNG, that is based on ideas of scaling and
universality. Suppose that at very large scales the effective coupling
becomes weak so that we can solve the problem perturbatively. On
an intuitive level, the main physical assumption is that there ex-
ists an "eddy viscosity" which yields an effective (scale dependent)
Reynolds number Re/f which is O(l) even as the "bare" Reynolds
number R, based on i/0, approaches infinity as the scale L —*• oo.
In comparison with techniques of direct summation of perturbation
series, RNG methods are more robust and more appealing to our
physical intuition about the significance or insignificance of eddy dy-
namics at large scales.
The technical approach involved in RNG is that, rather than
trying to sum up the entire perturbation series, we treat the problem
explicitly using perturbative methods. In other words, we consider
all expansions as performed in a true small coupling parameter g. It
will be shown that, as R —> oo, the resummed perturbation series
self-similarly approaches a limiting form, called a fixed point. The
main assumption of the RNG theory is that, at this fixed point, the
nonlinear effective coupling (or the effective Reynolds number Reff)
is small enough that useful results can be obtained by a perturbation
expansion in powers of Re/f. The logic of the RNG weak coupling
approach is checked by assuming it to hold, computing Reff and
other quantities based on this assumption, and then checking the self-
consistency of the results. It will be shown below that Reff oc A/£
where e = 4 + y — d was introduced above. The case of (. — 0 will be
shown to lead to zero coupling at the largest scales, so that Re/f —> 0
as e —>• 0.
The basic idea of our approach is to iteratively remove narrow
bands of wavenumbers, <5A, from the dynamic equations and thereby
obtain new equations for the remaining variables. This is done as fol-
lows. Suppose that at some stage of this renormalization process the
dynamic variables involve wavenumbers from the band 0 < k| < A.
When we begin this renormalization scheme, we choose A = kj but
at later stages the moving cutoff A satisfies A < kj. We remove the
Renormalization Group 163

dynamic variables w > (^) in the band A — 6A < |k| < A by using the
Neumann series (9) to obtain new expressions for the remaining vari-
ables u < (A;), 0 < |k| < A — £A, expressing the dynamic variables for
A — <*>A < |k| < A in terms of the random force in this narrow band,
and then averaging over the corresponding subensemble of random
forces. At each stage of this process, the use of the Neumann series
(9) is justified because <5A is small. This perturbation technique in <5A
generates so-called recursion relations for the terms in the reduced
dynamical system satisfied by n < (A;). The next step of the RNG pro-
cedure is to solve these relations to obtain new dynamic equations
in the case where many narrow bands of modes are eliminated. Of
course, the errors in the reduced dynamical system that result from
solving these relations need no longer be small because we are accu-
mulating many small contributions. However, under the assumption
that our Reff is small we proceed to use the dynamical system that
results from the recursion relations. The justification for this last
step has not been given for turbulence, so it is only possible to judge
the validity of the RNG equations by the accuracy of the results to
which they lead. The advantage of the present explanation of the
RNG technique is that the e—expansion (Re/f <C 1) is required only
to eliminate high-order nonlinearities in the dynamical equations for
tt < (A;) and not to evaluate constants appearing in the lowest order
dynamical equations for u < (A;).
When the velocity field is decomposed as

(7) generates equations for the "<"-band and the '>"-band:

No approximations have yet been invoked. Equations (16) and (17)


exactly represent the fluctuating Navier-Stokes dynamics in the cor-
responding wavenumber bands. Computation of the Green's func-
tion G < (12) =< <5it < (l)/(5/ < (2) > involves averaging these stochas-
tic PDEs over the subensemble of realizations / > (A;) based upon the
rule (12) applied to the "<"-band equation resulting from the Neu-
mann series (9):
164 S. A. Orszag et al.

Upon taking the functional derivative < <5/<5/ < (6) > in (18) and
gaussian averaging over the subensemble in which the Fourier modes
of / are held fixed for |k| < A — <5A, one obtains the correction to the
Green's function which is second-order in the coupling constant:

When explicit notations are introduced, say, (1) = (a,u — 0,k)


and (6) = (/3,o; = 0,k'), (19) translates into the following analytic
expression for the correction to the viscosity:

where the symbol / indicates integration over the band A — <5A <
q < A, and we have included the factor Pa;g(k) on the left side of
(20) due to the incompressibility of u(k). Notice that this procedure
of unconditional shell-averaging is asymmetric with respect to the
arguments of the Green's function 6*0 and the correlation function
UQ] the wavenumber of UQ is restricted to the shell A — 6A < k < A
while there is no similar constraint upon the wavenumber of GQ.
To compute the integral in (20) we proceed as follows. First,
using (6) and performing the integral over the internal frequency %,
noting that

we obtain
Renormalization Group 165

where

In the limit k/q —> 0, this latter integral has the asymtotic form

Using the following properties of the projection operators P a/ g(k) and


Pa^(k),

we find

This integral can be expressed in terms of the elementary angular


integrals

where Sd is the area of a d—dimensional unit sphere. Thus, we


evaluate II a /3(k, q, n) as
166 5. A. Orszag et al.

Substituting this result into (20/), we obtain the viscosity correction

where r = log (fcj/A). Using analogous methods one finds that there
is no "bare force renormalization", i.e. no contribution to the forcing
D0k~y. However, a "thermal" noise proportional to k2 is generated,
namely,

We will now rewrite this system in terms of scale-dependent non-


linearities and thereby make contact with classical turbulence phe-
nomenological concepts like local Reynolds numbers, effective viscos-
ity, etc. We introduce two nondimensional "charges" of the theory:
namely, the effective nonlinear coupling constant <?(A) = D 0 /z/ 3 (A)A c
and the 'thermal' coupling constant <7:r(A) = _DrA c '~ 2 /i' 3 (A) which
is the dimensionless intensity of the generated 'thermal noise' DyA;2.
The charge g(A) is proportional to [Reff(A.)}2. Indeed,

may be evaluated in terms of the random force correlation function


< // > noting that w(A) oc G(A)/(A) where G(A) is of order of the
characteristic response time o^A)"1 oc [z/(A) A2]"1:

where < // >oc D0A yui(A)Ad. Therefore, we obtain

The system dynamics can then be described in terms of the be-


havior of 0,tfr). From (21), (22) it follows that
Renormalization Group 167

Once the autonomous system (23) is solved, the scale-dependent


viscosity Z'(A) is determined from the differential equation:

that follows from (21).


Observe that when d — 2 > 0, there is a fixed point of this system
in which QT decays rapidly to 0 for r increasing beyond (d — 2)"1.
Therefore, for large r, or, equivalently, at very large spatial scales,
r oc e"1, the equation for g? becomes redundant since g? is small.
At these large r, the equation for g becomes

When r —>• oo, the RNG transformation given by (25) has a simple
fixed point in which the coupling constant is indeed small:

as conjectured above. In the vicinity of this fixed point, the viscosity


behaves as V(T) oc e~ 1 ' 3 exp(—c/3r) and Reff ~ ^/^•
When d < 2, the constant g? quickly approaches a finite limit
of order 1. There is no solution in which g and g? are small as
€ —» 0. This means that, unlike the previous case, the applicability of
second order approximation is violated and there is no self-consistent
asymptotic expansion in e — 4+y-d. The case d — 2 requires special
consideration which is beyond the scope of this paper.
It is a key part of the RNG procedure to obtain simple reduced
dynamical equations for «<(&) in which various terms in the formal
perturbation expansion can be argued to disappear in true dynamics
as the integral scale £ -^ oo. An example of such an "irrelevant
term" is the self-generated thermal noise. It is easy to see that when
e < 1 and d > 2,

which is small compared with the driving force of the system. We


caution that as of yet there is no rigorous proof that all such terms
are negligible.
168 S. A. Oisz&g et al.

These general RNG results can be used to analyze three-dimen-


sional fully turbulent flow (d = 3, e = 4). In this case, there is no
force renormalization and the effective viscosity at the cutoff A is
given by the solution of (25) with the initial condition f(A 0 ) = i>0:

It should be emphasized at this point that since the effective Rey-


nolds number is proportional to e1'2 to lowest order in €, all pertur-
bation expansions used in the theory are valid at lowest order in e
for e —> 0. It should also be noticed that the constant A^ is indepen-
dent of e, in contrast to the result obtained by Yakhot and Orszag
(1986). With the present approach, it is not necessary to e—expand
AJ. to obtain the RNG result f(A). The difference from the earlier
derivation lies in symmetry of the integral (20) with respect to the
arguments of the Green's function and the correlation function. In
the limit e —> 0, the symmetrized and unsymmetrized versions of (20)
become the same.
The unknown amplitude D0 is determined by Dannevik et al.
(1987) using the property of energy conservation by the nonlinear
terms. Expanding to second order in e they recover the EDQNM
equations introduced by Orszag (1977) which express the mean dis-
sipation rate £ in terms of the lowest-order energy spectrum E(k)
and the effective viscosity v(k}. Based on this representation, as well
as upon Equation (27), they obtain the relation

Using (28), one finds

The coefficients of (29) and (30) agree well with the experimentally
observed parameters.
Renormalization Group 169

4 TRANSPORT MODELING

The methods described above have also been used to derive trans-
port models in which all the scales up to the integral scale of tur-
bulent flow are averaged over. The RNG transport model describes
the dynamics of the mean flow /7;, turbulence kinetic energy li', and
energy dissipation rate £ and defines the eddy transport coefficients
in terms of the latter. For example, the eddy viscosity in the RNG
K — £ model is given by

with C/j, = 0.0845, in good agreement with the "standard" value.


RNG theory gives this result by eliminating the length scale I be-
tween the expression (27) for turbulent viscosity and the following
expression for the total kinetic energy in isotropic inertial range ed-
dies at scales smaller than I:

Here the constant 0.71 is obtained from the Kolmogorov constant


CK by integrating the inertial range energy spectra to the cutoff
scale / = 2ir/k. Another advantage of the RNG theory is that it also
allows us to interpolate (31) into the low Reynolds number regions
and obtain more general expressions which are valid across the full
range of flow conditions from low to high Reynolds numbers.
The high Reynolds number form of the RNG K — £ model is
given by

where i/j° = veddy + vmo\ and the rate-of-strain term R given by


170 5. A. Orszag et al.

This latter term is expressed in the RNG K — £ model equations as

where 77 = SK/£ and S2 = ISijSij is the magnitude of the rate-of-


strain. The RNG theory gives values of the constants Csl = 1.42
and C'£2 = 1.68, and a = 1.39, in comparison with the "standard"
values (7^ fy 1.4 and Cs2 « 1.9, and a = 1.
For turbulent heat convection problems, the temperature field is
governed by:

The inverse turbulent Prandtl numbers a in (34)-(35) and (38) are


obtained from the RNG scalar heat transport relation derived anal-
ogously to (27):

where for the heat transfer problem a0 refers to the molecular inverse
Prandtl number.
There are two interesting properties of the RNG transport model
described above which lead to reduced eddy viscosity. First, the re-
duced value of (7f2 compared with the standard S -equation coefficient
(^standard ^ ^<^ j^g ^g interesting consequence of decreasing both
the rate of production of K and the rate of dissipation of S, leading
to smaller eddy viscosities. In regions of small 77, the .R-term tends
to increase eddy viscosity somewhat, but it is still typically smaller
than its value in the standard model. In regions of very large 77,
where strong anisotropy exists, R can become negative and reduces
the eddy viscosity even more. This feature of the RNG model is
perhaps responsible of the marked improvement in anisotropic large-
scale eddies. The /^-dependency of C'£2 makes it possible to have a
spatially as well as a temporally varying balance between production
and dissipation terms in the ^-equation. This feature of dynamic
balance, based on the RNG theory, is unique. Secondly, in regions
of low turbulence Reynolds numbers the RNG eddy viscosity has a
cut-off below which the eddy viscosity is zero.
The RNG K — £ model has been successfully applied to a number
of difficult turbulent flow problems. Below, we present some results
Renormalization Group 171

to illustrate the ability of this model to capture complex flow phe-


nomena such as transitional behavior in turbomachinery heat trans-
fer, and time-dependent behavior of large-scale structures.
Turbomachinery Heat Transfer
Two cases of predicting turbomachinery heat transfer will be dis-
cussed, the so-called Langston cascade (Graziani et al., 1979) which is
large-scale and low-speed, and a transonic nozzle guide vane. For the
former case, the main difficulty is to accurately predict the pattern
of the heat transfer coefficient distribution over the suction surface
corresponding to flow relaminarization and subsequent transition to
turbulence. In the latter case, compressibility and an extremely high
turning angle further complicate the subtle heat transfer behavior
over the surface.
In Figs. 1-2 we plot heat transfer data compared with the results
of Graziani et al. as well as with results obtained from alternative
turbulence/transition models. The data is expressed in terms of the
Stanton number, Si = h/(pCpUo), where h is the heat transfer co-
efficient defined as h = qgen/AT, qgen is the heat flux at the blade
surface, and p, Cp, and UQ are the values at the inlet of the density,
specific heat, and velocity. In Fig. 1, we plot a comparison of the
RNG results and the measured distribution of the Stanton number.
Good qualitative agreement with the experimental data is observed.
The quantitative agreement is excellent in the leading edge region
and good in the relaminarization region. In Fig. 2, we plot a com-
parison of the Stanton number predictions for the RNG and standard
K — £ models and for laminar heat transfer - in the laminar case, the
flow field is the turbulent flow field determined by the RNG model
but the turbulent heat transfer coefficients predicted by the RNG
theory are replaced by laminar heat transfer coefficients. Evidently,
the standard K — £ model overpredicts the heat transfer for this case
by nearly a factor 2 and does not capture the transitional flow ef-
fects observed near the leading edge on the suction side. On the other
hand, the laminar heat transfer model does do a good job near the
leading edge, but does not account for the increase in heat transfer
near the transition point. Overall, the RNG model has demonstrated
a significant improvement over existing prediction methods, without
the use of an empirically fitted transition point.
Extensive parametric studies of heat transfer in this cascade were
also performed to investigate the effects of Reynolds number and in-
172 S. A. Orszag et al.

Figure 1. Distributions of Stanton number on the blade surface for


the Langston cascade: RNG-based computations versus experiment.

let turbulence intensity level on heat transfer patterns. Boundary


conditions for the heat transfer problem were the same for each run
and corresponded to constant heat flux qgen — I.31kW/m? at the
airfoil surface and ambient flow temperature at the inlet. The origi-
nal experiment was done at a Reynolds number (based on axial chord
and inlet velocity) of ReQ = 5.5 X 105 and inlet turbulence intensity
level of 1%. In Fig. 3, we show the distributions of Stanton number
along the airfoil surface corresponding to different Reynolds num-
bers. The curves are labeled by the Reynolds number relative to
RCQ. These data indicate that boundary layer transition occurs at
Re between 3.3 X 10s and 3.6 X 105: the Stanton number distributions
at 3.3 X 105 and lower match the laminar pattern. Heat transfer pat-
terns corresponding to Re0 = 5.5 x 10s and different inlet turbulence
intensity levels are shown in Fig. 4. Virtually no variation in the
Stanton number distributions was detected at turbulence intensities
lower than 0.5%.
The second case is a simulation of experiments performed in the
Renormalization Group 173

Figure 2. Comparison of computed distributions of Stanton number


on the blade surface for the Langston cascade.

von Karman Institute short duration Isentropic Light Piston Com-


pression Tube facility CT-2 of a high pressure turbine nozzle guide
vane (von Karman Institute for Fluid Mechanics Technical Note 174,
1990). In Fig. 5 we plot the distribution of heat transfer coefficient
in W/m^K as a function of the distance along the blade surface,
with the suction side shown on the left side of the plot and pressure
side on the right side. The accurate quantitative prediction of blade
heat transfer over almost the entire airfoil surface is readily observ-
able. The two spikes in computed heat transfer coefficient represent
purely numerical artifacts due to the junctions of the computational
domains. By comparison the standard K — £ model overpredicts heat
transfer by a factor of 2-5 due to significant overproduction of turbu-
lence throughout the passage. At the same time, the distribution of
turbulent viscosity provided by the RNG model leads to an quanti-
tative prediction of blade heat transfer over most of the entire airfoil
surface. Again, it should be emphasized that both the RNG and
standard model - based calculations were done using general pur-
174 S. A. Orszag et al.

Figure 3. Distributions of Stanton number over the blade surface as


a function of Reynolds number. The curves are marked according to
the fraction of the nominal Reynolds number of 5 X 105 ( based on
inlet velocity and axial chord ).

pose codes which contain no adjustable parameters characterizing


the turbine geometry.

Time-dependent Turbomachinery Computations

Some recent results from the computation of the flow over a com-
pressor trailing edge indicate that steady-state computations prove
inadequate in flows with large-scale unsteady structure. It will be
seen that when these calculations have a steady-state solution it does
not correspond to the time-averaged flow field. On the other hand,
the RNG transport model, which has a higher effective Reynolds
Renormalization Group 175

Figure 4. Distributions of Stanton number over the blade surface


as a function of inlet turbulence intensity. The curves are marked
according the intensity at the inlet in percent (100 X u'/Uo).

number than the standard K—£ model, may be used to perform time-
dependent, so-called Very Large Eddy Simulations (VLES). VLES
refers to the capability of the model to correctly predict the behav-
ior of anisotropic eddies with a size of the order of the characteristic
size of the problem, in this case the thickness of the trailing edge.
The Reynolds number for the flow under consideration is 56,400
based on the diameter of the half-cylinder trailing edge. The inlet,
taken at 10.6 diameters upstream, is a symmetrical zero pressure
gradient turbulent boundary layer set to match experimental con-
ditions. A number of runs were computed including steady-state
and time-dependent cases, and a steady-state case with a splitter
176 S. A. Orszag et al.

Figure 5. Comparison of computed and experimental distributions of


heat transfer coefficient on the blade surface for a high-load turbine
nozzle guide vane.

plate. First, steady-state computations of the trailing edge failed to


predict the strength of the vortex in the wake yielding significantly
over-predicted base pressures in the wake. As seen in Fig. 6, the
"symmetrizing" effect of the steady-state computation on the distri-
bution of the pressure coefficient Cp is close to that obtained by the
addition of a splitter plate in the wake. The streamlines shown in
Fig. 7 also demonstrate that the size of the vortex is the same in
both cases.
Next, a time-dependent flow with the RNG K - £ model was
computed. The computed Strouhal frequency is approximately 0.2 in
reasonable agreement with experiment. In Fig. 8 the time-averaged
distribution of Cp for the RNG K - £ run is plotted. The time-
averaged VLES result is also in reasonable agreement with exper-
iment and gives rise to the strong pressure minimum in the wake.
The time-dependent run gives a strong vortex as indicated in Fig. 9,
where both the magnitude and location of the velocity minimum a,re
in reasonable agreement with the experimental values of-0.2 and 0.4
Renormalization Group lj",

Figure 6. Pressure coefficient distributions over the surface and in


the wake of a compressor trailing edge for steady-state computations
with and without a splitter plate, and experimental splitter plate
results (Re = 56,400 based on freestream velocity and trailing edge
(TE) thickness).

respectively.
Finally, a time-aver aged calculation using the standard K — £
model was performed for comparison. A comparison of the time-
averaged results for Cp with experiment is shown in Fig. 8. As with
the steady-state RNG computations, the base pressure in the wake is
significantly overpredicted by the standard model. Excessive turbu-
lent production leads to large eddy viscosities and hence mean wake
vortices which are too weak. It appears that the VLES RNG model,
by producing less turbulence as a result of lower eddy viscosities, can
predict the time-dependent behavior of the (very) largest eddies.
Flapping-Hydrofoil Simulations
The flow over a two dimensional hydrofoil subject to vertical gusts
at high reduced frequency has been simulated, and a comparison
with the experiments performed in the MIT Variable Pressure Water
Tunnel was done. A modified NAG A16 hydrofoil with chord length
178 S. A. Orszag et al.

Figure 7. Computed steady-state streamlines in the wake of com-


pressor trailing edge with and without a splitter plate.

of 18 inches and maximum thickness of 8.84% is mounted in the


test section on the centerline of the tunnel. Upstream there are
two flappers operating at a reduced frequency K — 3.6 and with
an amplitude of 6 degrees. The Reynolds number based on test
foil chord length and freestream velocity is 3.78 X 106. Turbulence
intensity was measured in the empty test section and was found to
be 1%. In the case of steady flow, measurement of velocities and
pressure was taken on three nested boxes bounding the test foil.
Surface pressure data were taken on eight locations and boundary
layer velocity data were taken in nine location on the suction side
and seven locations on the pressure side. There were also two sets of
wake data taken.
The computations were performed using the FLUENT code with
the RNG K — £ turbulence model. For the case reported below, the
simulation was for steady flow at an angle of attack of 1.18°. The
computational domain spans the tunnel height with the upstream in-
let coinciding with the location where experimental measurement of
the flow data is available. The geometry and computational domain
are depicted in Fig. 10. A C-type grid was built around the test foil
and there are 301x166 total grid points with approximately 35 grid
points spanning the boundary layer. The upstream boundary con-
ditions were taken from the experimental data, in the two small re-
Renorma,liza,tion Group 179

Figure 8. Pressure coefficient distributions over the surface and in


the wake of a compressor trailing edge for time-aver aged computa-
tions of RNG VLES models and Standard K — e with experiment
(Re — 56,400 based on freestream velocity and trailing edge (TE)
thickness).

gions between upper/bottom tunnel wall and the flapper wake where
experimental data is not available, uniform velocity profiles were as-
signed.
Two sets of runs were performed with the same grid and bound-
ary conditions, but using different turbulence models. One used the
standard K — C model and the other used RNG K — £ model. The
skin friction coefficients are plotted in Fig 11. On the suction side
with x/C > 0.4 where experiment data were available, the computa-
tion with the RNG K — £ model produced excellent agreement while
the computation results using the standard K — E model severely
overpredicted the skin friction. The results from the RNG K — £
model predicted the separation point (shown by the location where
skin friction first turns negative) at x/C = 0.968, while the exper-
iment data showed the separation point between x/C — 0.972 and
x/C — 0.990. The standard K — £ model, on the other hand, in-
180 5. A. Orszag et al.

Figure 9. Comparison of the computed time-averaged centerline ve-


locity in the near wake of the compressor trailing edge.

Figure 10. The geometry and computational domain for flapping-


hydrofoil simulations.

dicates no separation. The RNG K — £ model computation also


predicted transition regions at x/C ~ 0.09 on the suction side and
at x/C ^ 0.13 on the pressure side. Since the experiment used a
trip device to trigger transitions on the test foil, comparison is not
available in these regions.
There now have been well over fifty substantially different test
flows on which the RNG K — 8 equations have been successfully
tested (e.g. Orszag et al, 1993b). This broad range of successes
Renormalization Group 181

Figure 11. Comparison of computed and experimental skin friction


coefficients for the suction side of the hydrofoil.

provides, we believe, an important indirect verification of the ideas


entering the RNG theory of turbulence. Indeed, turbulence models
are very sensitive to their details and we believe it is quite unlikely
that a "wrong" theory would be likely to produce the breadth of
significant engineering results available so far. In effect, this means
that we have a reasonable working theory of turbulence at least for
engineering calculations. Much further work remains to be done to
extend the range of applications, determine the limits of validity of
the theory, and, hopefully, provide a rational basis for it as a physical
theory of turbulence.

ACKNOWLEDGMENTS

This work has been supported by the ONR, NASA, and United
Technologies Research Center. Discussions with Victor Yakhot have
been a constant source of stimulation. We would also like to ac-
knowledge discussions with Torn Barber and Dochul Choi of UTRC.
Finally, we should like to acknowledge our collaborations with Fluent,
Inc., especially Ferit Boysan, Nelson Carter, Dipankar Choudbury,
182 S. A. Orszag et al.

Joe Maruzewski, and Bart Patel.

5 REFERENCES
Chen, S., Doolen, G., Herring, J. R., Kraichnan, R. H., Orszag,
S. A. and She, Z.-S., 1993. "Far dissipation range of turbu-
lence," Phys. Rev. Letts. 70, p. 3051.
Dannevik, W., Yakhot, V., and Orszag, S. A., 1987. "Analytical
theories of turbulence and the e—expansion," Phys. Fluids A
30, p. 2021.
Graziani, "R. A., Blair, M. F., Taylor, J. R., and Mayle, R. E.,
1979. "An Experimental Study of Endwall and Airfoil Surface
Heat Transfer in a Large Scale Blade Cascade," ASME Paper
79-GT-99.
Lurie, E. H., 1993. "Unsteady Response of a Two-Dimensional Hy-
drofoil Subject to a High Reduced Frequency Gust Loading,"
M.S. Thesis, Massachusetts Institute of Technology.
Kraichnan, R.H., 1961. "Dynamics of nonlinear stochastic sys-
tems," J. Math. Phys. 2, pp. 124-148. Erratum: 3, p. 205,
1962.
Orszag, S.A., 1977. "Statistical theory of turbulence," In: Fluid
Dynamics 1973, Les Houches Summer School in Physics. Eds.
R. Balian and J.-L. Peabe, Gordon and Breach, pp. 237—374.
Orszag, S.A., Staroselsky, I., and Yakhot, V., 1993a. "Some basic
challenges for large eddy simulation research," In: Large Eddy
Simulation of Complex Engineering and Geophysical Flows. (B.
Galperin and S.A. Orszag, eds). Cambridge University Press.
Orszag, S. A., Yakhot, V., Flannery, W. S., Boysan, F., Choudbury,
D., Maruzewski, J., and Patel, B., 1993b. "Renormalization
group modeling and turbulence simulations," In: Near-Wall
Turbulent Flows. Eds. So, R. M. C., Speziale, C. G. and
Launder, B. E., Elsevier Science Publishers, pp. 1031-1046.
Paterson, R.W., 1984. "Experimental Investigation of a Simulated
Compressor Airfoil Trailing Edge Flowneld," AIAA Paper 84-
0101.
Renormalization Group 183

Yakhot, V., and Orszag, S.A., 1986. "Renormalization group anal-


ysis of turbulence. I. Basic theory," J. Sci. Comp., 1, pp. 3-51.

Wyld, H.W., 1961. Formulation of the theory of turbulence in an


incompressible fluid," Ann. Phys. 14, pp. 143-165.
This page intentionally left blank
Chapter 5

MODELING OF
TURBULENT TRANSPORT
EQUATIONS

Charles G. Speziale

1 INTRODUCTION

The hlgh-Reynolds-number turbulent flows of technological im-


portance contain such a wide range of excited length and time scales
that the application of direct or large-eddy simulations is all but im-
possible for the foreseeable future. Reynolds stress models remain
the only viable means for the solution of these complex turbulent
flows. It is widely believed that Reynolds stress models are com-
pletely ad hoc, having no formal connection with solutions of the
full Navier-Stokes equations for turbulent flows. While this belief is
largely warranted for the older eddy viscosity models of turbulence,
it constitutes a far too pessimistic assessment of the current gen-
eration of Reynolds stress closures. It will be shown how second-
order closure models and two-equation models with an anisotropic
eddy viscosity can be systematically derived from the Navier-Stokes
equations when one overriding assumption is made: the turbulence
is locally homogeneous and in equilibrium.
A brief review of zero equation models and one equation mod-
els based on the Boussinesq eddy viscosity hypothesis will first be
provided in order to gain a perspective on the earlier approaches to

185
186 C. G. Spezide

Reynolds stress modeling. It will, however, be argued that since tur-


bulent flows contain length and time scales that change dramatically
from one flow configuration to the next, two-equation models con-
stitute the minimum level of closure that is physically acceptable.
Typically, modeled transport equations are solved for the turbulent
kinetic energy and dissipation rate from which the turbulent length
and time scales are built up; this obviates the need to specify these
scales in an ad hoc fashion. While two-equation models represent the
minimum acceptable closure, second-order closure models constitute
the most complex level of closure that is currently feasible from a
computational standpoint. It will be shown how the former models
follow from the latter in the equilibrium limit of homogeneous turbu-
lence. However, the two-equation models that are formally consistent
with second-order closures have an anisotropic eddy viscosity with
strain-dependent coefficients - a feature that most of the commonly
used models do not possess.
For turbulent flows that are only weakly inhomogeneous, full
Reynolds stress closures can then be constructed by the addition
of turbulent diffusion terms that are formally derived via a gradient
transport hypothesis. Properly calibrated versions of these models
are shown to yield a surprisingly good description of a wide range of
two-dimensional mean turbulent flows that are near equilibrium. In
particular, the stabilizing or destabilizing effect of a system rotation
on turbulent shear flows is predicted in a manner that is consistent
with hydrodynamic stability theory. However, existing second-order
closures are not capable of properly describing turbulent flows that
are far from equilibrium and have major problems with wall-bounded
turbulent flows.
High-speed compressible turbulent flows present a whole new
range of problems to Reynolds stress modeling. It has now become
clear that the traditional approach of using variable density exten-
sions of incompressible Reynolds stress models is not sufficient for
the reliable prediction of the high-speed compressible flows of aero-
dynamic importance. Reynolds stress closures are needed where tur-
bulent dilatational effects are accounted for in a physically consistent
fashion. Recent efforts in the modeling of the dilatational dissipation
and pressure-dilatation correlations have led to significant advances
in the description of free turbulent shear flows at high Mach num-
bers. However, a variety of problems still remain. All of the central
points of the paper will be illustrated by examples and an assessment
Turbulent Transport Equations 187

will be made concerning future directions of research.

2 INCOMPRESSIBLE TURBULENT FLOWS

The governing equations of motion for incompressible turbulence


are:
Navier-Stokes Eq.

Continuity Eq.

where

Here, P is a solution of the Poisson equation

2.1 Reynolds Averages

The velocity and pressure are decomposed into mean and fluctu-
ating parts, respectively, as follows (cf. Hinze 1975):

where an overbar represents a Reynolds average. This Reynolds av-


erage can take a variety of forms for any flow variable 0:
Homogeneous Turbulence
188 C. G. Speziale

Statistically Steady Turbulence

General Turbulence

(f) = <j)(x,t) (Ensemble Average over N repeated experiments)

It is assumed that the Ergodic Hypothesis applies. In a homogeneous


turbulence,

whereas in a statistically steady turbulence,

Reynolds Averaging Rules


Given that

then
Turbulent Transport Equations 189

2.2 Reynolds-Averaged Equations

The Reynolds average of the equations of motion (1) - (2) are


given by:
Navier-Stokes Eq.

Continuity Eq

where

is the Reynolds stress tensor (note that dui/dxi = 0).

2.3 The Closure Problem

The Reynolds-averaged Navier-Stokes equation (10) is not closed


unless a model is provided that ties the Reynolds stress tensor T{J to
the global history of the mean velocity Ui in a physically consistent
fashion. In mathematical terms, r^ is assumed to be a functional of
the global history of the mean velocity field, i.e.,

The oldest Reynolds stress closures are based on the Boussinesq eddy
viscosity hypothesis:

where

For incompressible turbulent flows, the isotropic part of the Reynolds


stress tensor, given by |/f (here, T;J — \K&ij + DTij where K = |r,-,-
is the turbulent kinetic energy), can simply be absorbed into the
mean pressure.
190 C. G. Spezi&le

2.4 Older Zero- and One-Equation Models

The eddy viscosity is given by

where to = turbulent length scale and to = turbulent time scale.


Zero-Equation Models
Here, both IQ and to are specified algebraically by empirical means.
The first successful zero equation model based on the Boussinesq
eddy viscosity hypothesis was Prandtl's mixing length theory formu-
lated in the 1920's. In the mixing length theory of Prandtl (1925),

where IQ = Ky is the mixing length (K is the Von Karman constant


and y is the normal distance from a solid boundary). This represen-
tation is only formally valid for thin turbulent shear flows - near a
solid boundary - where the mean velocity is of the simple unidirec-
tional form v = w(j/)i.
Four decades later, this type of model was generalized to three-
dimensional turbulent flows:
Smagorinsky (1963) Model

where

is the mean rate of strain tensor;


Baldwin and Lomax (1978) Model

where u> = Vxv is the mean vorticity vector. The former model has
been primarily used as a subgrid scale model for large-eddy simula-
tions whereas the latter model has been used for Reynolds-averaged
Navier-Stokes computations in aerodynamics (see Wilcox 1993).
Turbulent Transport Equations 191

One-Equation Models
The eddy viscosity in these models is typically given by

where K = \U{U{ is the turbulent kinetic energy which is obtained


from a separate modeled transport equation. Models of this type
were first proposed independently by Prandtl (1945) and by Kol-
mogorov (1942) during the Second World War.
More recently, Baldwin and Earth (1990) and Spalart and All-
maras (1992) proposed one-equation models wherein a modeled trans-
port equation is solved for the eddy viscosity z/x- One-equation mod-
els are superior, to zero-equation models in that the time scale to of
the eddy viscosity is built up from turbulence statistics rather than
from the mean velocity gradients. However, both of these types of
models have limited predictive capabilities due to the fact that the
turbulent length scale IQ has to be specified empirically. This is vir-
tually impossible to do in complex three-dimensional turbulent flows.
The turbulent length and time scales are not universal; they depend
strongly on the flow configuration under consideration.
It is thus argued that two-equation turbulence models - wherein
transport equations are solved for two independent turbulence quan-
tities that are directly related to IQ and to - represent the minimum
acceptable level of Reynolds stress closure. These models should
be formulated with a properly invariant anisotropic eddy viscosity
model that is nonlinear in the mean velocity gradients. The stan-
dard Boussinesq eddy viscosity hypothesis makes it impossible to
properly describe turbulent flows with:
(i) Body force effects arising from a system rotation or from stream-
line curvature;
(ii) Flow structures generated by normal Reynolds stress anisotro-
pies (e.g., secondary flows in non-circular ducts).
In the most common approach to two-equation models, 10 oc K3/2/e,
to oc K/e and modeled transport equations for K and £ are solved
(here e = ^f^ff1: is the turbulent dissipation rate). In lieu of e, the
inverse time scale e/K has also been used (see Wilcox 1993).
Limitations in computer capacity, and issues of numerical stiff-
ness, make second-order closure models - wherein transport equa-
tions are solved for the individual components of T,-J along with a
192 C. G. Speziale

scale equation - the highest level of closure that is currently feasible


for practical calculations.

2.5 Transport Equations of Turbulence


The transport equation for the fluctuating velocity U{ is obtained
by subtracting the Reynolds-averaged Navier-Stokes equation (10)
from its unaveraged form (1) which yields:

Eq. (13) can be written in operator notation as JV"ti,- = 0.


Reynolds Stress Transport Equation
This equation is obtained by constructing the second moment

which is given by (cf. Hinze 1975)

where

The transport equation for the turbulent kinetic energy (K = ^TH)


is obtained by a simple contraction of the Reynolds stress transport
equation (14):

where
Turbulent Transport Equations 193

Dissipation Rate Equation


We construct the ensemble mean

which yields the dissipation rate equation

Both two-equation models and second-order closure models are


obtained from the Reynolds stress transport equation as follows:

(i) Second-order closures are obtained by modeling the full in-


homogeneous Reynolds stress transport equation (14);
(ii) Two-equation turbulence models with an anisotropic eddy vis-
cosity are obtained algebraically from (14) by assuming that
the turbulence is locally homogeneous and in equilibrium.

2.6 Two-Equation Models

Representation for the Reynolds stress tensor


For homogeneous turbulence, the Reynolds stress transport equation
reduces to
194 C. G. Speziale

or

where (cf. Lumley 1978 and Reynolds 1987)

Since, from (3) - (4)

it follows that

Here,

which can be non-dimensionalized as follows:

It is assumed that an equilibrium state is reached where

achieve constant values that are independent of the initial condi-


tions. With no loss of generality, it can then be shown that for two-
dimensional mean turbulent flows in equilibrium (Speziale, Sarkar
Turbulent Transport Equations 195

and Gatski 1991):

where

For the moment, we will neglect the anisotropy of dissipation dij


and the quadratic return term containing Cj. In many studies, the
former assumption is justified by invoking the Kolmogorov hypothesis
of local isotropy for high Reynolds number turbulent flows. However,
this assumption is somewhat debatable as we will see later.
The coefficients C\ — €4 were obtained by Speziale, Sarkar and
Gatski (1991) based on a calibration with homogeneous shear flow
experiments (see Table 1).

Equilibrium
Values LRR Model SSG Model Experiments
(&ll)oo 0.158 0.204 0.201
(^22)00 -0.123 -0.148 -0.147
(£"12)00 -0.187 -0.156 -0.150
(StfA)oo 5.32 5.98 6.08

Table 1. Comparison of the predictions of the Launder, Reece and


Rodi (LRR) model and the Speziale, Sarkar and Gatski (SSG) model
with the experiments of Tavoularis and Corrsin (1981) on homoge-
neous turbulent shear flow.

For homogeneous turbulence in equilibrium, 6,-j achieves equilib-


196 C. G. Speziale

rium values that are independent of the initial conditions. Hence,

in equilibrium. Since,

it then follows from (24) that the Reynolds stress transport equation
has the equilibrium form

where (see Gatski and Speziale 1993):

By the use of integrity bases from linear algebra, Pope (1975) first
showed that the general solution to the implicit algebraic stress equa-
tion (26) is of the form:

where
Turbulent Transport Equations 197

are the integrity bases and {•} denotes the trace.


Three-Dimensional Solution
The resulting solution for G^ is given by (see Gatski and Speziale
1993):

where the denominator D is given by

and

Taulbee (1992) derived a simplified, but highly special, three-


dimensional solution for the degenerate case where €3 = 2. The
explicit solution for two-dimensional mean turbulent flows simpli-
fies considerably since only the first three integrity bases are linearly
independent in this cases (see Pope 1975 and Gatski and Speziale
1993):

where
198 C. G. Speziale

In more familiar terms, this expression is equivalent to

Here, 0:1,0:2 and 0:3 are not constants but rather are related to the
coefficients C\ — C$ and g. In mathematical terms, they are "projec-
tions" of the fixed points of Aij and Mijkt onto the fixed points of
bij - quantities that can vary from one flow to the next. However,
for two-dimensional mean turbulent flows, it appears that C\ — €4
can be approximated by constants due to the linear dependence on
b^ which allows us to use the principle of superposition. (For three-
dimensional turbulent flows, the general representation for Mijkf is
nonlinear in bij which leads to an inconsistency that will be discussed
later).
It should be noted that these models obtained via the algebraic
stress approximation need to be regularized before they are applied
to complex turbulent flows. This can be accomplished via a Fade
type approximation whereby we take (Gatski and Speziale 1993):

This constitutes an excellent approximation for turbulent flows that


are close to equilibrium with a form that is regular for turbulent
flows that are far from equilibrium (the original expression yields a
singular or negative eddy viscosity for sufficiently large values of 77 -
a feature that can lead to divergent computations).
It should be further noted that Yoshizawa (1984), Speziale (1987),
Rubinstein and Barton (1990), and Zhou et al. (1994) have derived
models of this general form - that are tensorially quadratic with con-
stant coefficients - via expansion techniques based, respectively, on
two-scale DIA, continuum mechanics, e-RNG (see Yakhot and Orszag
1986 and Yakhot et al. 1992), and recursion RNG techniques. How-
ever, due to the constant coefficients, these models are not consistent
with second-order closures and they have dispersive terms that, grow
unbounded with the strain rate - a feature that can destabilize com-
putations.
Turbulent Transport Equations 199

This explains why previous anisotropic corrections to eddy vis-


cosity models have not fully succeeded:

(i) The coefficients should depend nonlinearly on the invariants of


the rotational and irrotational strain rates.

(ii) Only the traditional algebraic stress models - based on the


solution of the implicit algebraic Eq. (26) such as the model
of Rodi (1976) - have had such a dependence and they are
ill-behaved!

Problem (ii) can now be overcome by a regularization procedure


based on a Fade type approximation as discussed above. In rotating
frames, Coriolis terms must be added to the r.h.s. of the Reynolds
stress transport equation (18). Gatski and Speziale (1993) showed
that this analysis exactly accounts for such non-inertial effects in
rotating frames if the extended definition of W*, is used:

where fim is the angular velocity of the reference frame.


If we have a separation of scales, then 77, £ <C 1 and, in the linear
limit, we recover the eddy viscosity model

which forms the basis for the standard K — e model of Launder and
Spalding (1974). However, in practical turbulent flows we do not
have a separation of scales; T] and £ are of O(l). Nonetheless, for
two-dimensional turbulent shear flows in equilibrium, the new model
of Gatski and Speziale (1993) yields

with

which is remarkably close to the value of CM = 0.09 used in the


standard K — £ model.
200 C. G. Spezi&le

The Dissipation Rate Model


In homogeneous turbulent flows, the turbulent kinetic energy is a
solution of the transport equation

Hence, closure is achieved once a transport equation is provided for


e in terms of T^J and dvi/dxj. The exact transport equation (17) for
e in homogeneous turbulence can be written in the form:

where

This equation can be rewritten as

where

For isotropic turbulence,

The major contributions to this integral occur at high wavenum-


bers where the energy spectrum E ( K , t ) oc E(Klk) given that Ik =
Turbulent Transport Equations 201

^3/4^1/4 jg ^e Kolmogorov length scale. This Kolmogorov scaling


yields (Bernard and Speziale 1992):

Hence,

We then invoke the standard high-Reynolds-number equilibrium hy-


pothesis whereby

Models for d^j and d^ can be obtained from an analysis of the


transport equation for the tensor dissipation which, for homogeneous
turbulence, is given by (see Durbin and Speziale 1991):

where

In physical terms,

Nij = Production by Vortex Stretching - Destruction


by Viscous Diffusion

A -If = Structure and Redistribution Term.


202 C. G. Spezia

Speziale and Gatski (1992) proposed the following models:

based on an expansion technique that makes use of tensor invariance,


symmetry properties and the fact that dij is small so that only linear
terms need to be maintained.
The standard equilibrium hypothesis is made where:

This leads to an algebraic system of equations analogous to that


obtained in the algebraic stress approximation discussed earlier. For
two-dimensional mean turbulent flows, the exact solution is:
Turbulent Transport Equations 203

where

(Speziale and Gatski 1992). The substitution of these algebraic equa-


tions into the contraction of the e;; transport equation yields the
scalar dissipation rate equation

where

It should be noted that the contraction of (40) yields dlydvi/dxj


— adijdvi/dxj which was used to obtain this result. The constants
03 and CE5 were evaluated using DNS results for homogeneous shear
flow (Rogers, Moin and Reynolds 1986). The constant C£2 can be
evaluated by an appeal to isotropic turbulence. This dissipation rate
model predicts that the turbulent kinetic energy decays according to
the standard power law in isotropic turbulence (cf. Speziale 1991):

Based on this, we have taken Cei — 1.83 which yields an exponent


of approximately 1.2 in agreement with the most cited experimental
data (see Comte-Bellot and Corrsin 1971).
204 C. G. Speziale

For two-dimensional turbulent shear flows that are in equilibrium,

which is remarkably close to the traditionally chosen constant value


of Cei = 1.44. For more general two-dimensional homogeneous tur-
bulent flows, this model takes the form:

where

In contrast to this result, the new dissipation rate model of Lumley


(1992) takes the form

for near equilibrium turbulent flows where C\ is a constant. The


model (49) obviously contains less physics than the model (48) de-
rived herein since it does not depend on rotational strains. It has long
been recognized that the dissipation rate is dramatically altered by
rotations. The results presented in this paper clearly show that this
effect can be incorporated by accounting for anisotropic dissipation.
To the best knowledge of the author, this constitutes the first sys-
tematic introduction of rotational effects into the scalar dissipation
rate equation. Previous attempts to account for rotational effects
(see Raj 1975; Hanjalic and Launder 1980; and Bardina, Ferziger
and Rogallo 1985) were largely ad hoc.

Applications
For weakly inhomogeneous turbulent flows that are near equilibrium,
we can extend the K and s transport equations by the addition of
gradient transport terms that are obtained by a formal expansion
technique:
Turbulent Transport Equations 205

where Uk and cre are constants that typically assume the values of
1.0 and 1.3, respectively. These forms are assumed to be valid ap-
proximations at high Reynolds numbers.
We will now consider several non-trivial applications of the two-
equation model derived herein which can be referred to as an explicit
algebraic stress model (ASM) based on the SSG second-order closure.
The first case that will be considered is homogeneous shear flow in a
rotating frame (see Figure 1). In this flow, an initially isotropic tur-
bulence (with turbulent kinetic energy KQ and turbulent dissipation
rate £Q) is suddenly subjected to a uniform shear (with constant shear
rate 5) in a reference frame rotating steadily with angular velocity
ft. In Figures 2(a)-2(c), the time evolution of the turbulent kinetic
energy predicted by this new two-equation model is compared with
the large-eddy simulations (LES) of Bardina, Ferziger and Reynolds
(1983), as well as with the predictions of the standard K — e model
and the full SSG second-order closure. From these results, it is clear
that the new two-equation model yields the correct growth rate for
pure shear flow (ft/5 = 0) and properly responds to the stabilizing
effect of the rotations ft/5 = 0.5 and ft/5 = —0.5. These results
are remarkably close to those obtained from the full SSG second-
order closure as shown in Figure 2. In contrast to these results,
the standard K — £ model overpredicts the growth rate of the turbu-
lent kinetic energy in pure shear flow (ft/5 = 0) and fails to predict
the stabilizing effect of the rotations illustrated in Figures 2(b)-2(c).
Since the standard K — £ model makes use of the Boussinesq eddy
viscosity hypothesis, it is oblivious to the application of a system
rotation (i.e., it yields the same solution for all values of ft/5). The
new two-equation model predicts unstable flow for the intermediate
band of rotation rates —0.09 < ft/5 < 0.53; this result is generally
consistent with linear stability theory that predicts unstable flow for
0 < ft/5 < 0.5.
In Figure 3, the prediction of the new two-equation model for
the mean velocity profile in rotating channel flow is compared with
the experimental data of Johnston, Halleen and Lezius (1972) for a
rotation number Ro — 0.068. It is clear from these results that the
model correctly predicts that the mean velocity profile is asymmetric
in line with the experimental results - an effect that arises from
Coriolis forces. In contrast to these results, the standard K — e
model incorrectly predicts a symmetric mean velocity profile identical
to that obtained in an inertial frame (the standard K — £ model is
206 C. G. Speziale

Figure 1. Schematic of homogeneous shear flow in a rotating frame.

oblivious to rotations of the reference frame, as alluded to earlier).


As demonstrated by Gatski and Speziale (1993), the results obtained
in Figure 3 with this new two-equation model are virtually as good
as those obtained from a full second-order closure. This is due to the
fact that a representation is used for the Reynolds stress tensor that
is formally derived from a second-order closure (the SSG model) in
the equilibrium limit. It is now clear that previous claims that two-
equation models cannot systematically account for rotational effects
were erroneous.
Two examples will now be presented that illustrate the enhanced
predictions that are obtained for turbulent flows exhibiting effects
arising from normal Reynolds stress differences. Here, we will show
results obtained from the nonlinear K — e model of Speziale (1987).
For turbulent shear flows that are predominately unidirectional, with
secondary flows or recirculation zones driven by small normal Rey-
Turbulent Transport Equations 207

Figure 2. Time evolution of the turbulent kinetic energy in rotating


homogeneous shear flow: Comparison of the model predictions with
the large-eddy simulations of Bardina et al. (1983). (a) ft/5 = 0,
(b) ft/5 = 0.5 and (c) ft/5 = -0.5 (from Gatski and Speziale 1993).

nolds stress differences, a quadratic appoximation of the anisotropic


eddy viscosity model discussed herein collapses to the nonlinear K—e
model (see Gatski and Speziale 1993). In Figure 4, it is demonstrated
that the nonlinear K — e model predicts an eight-vortex secondary
208 C. G. Speziale

Figure 3. Comparison of the mean velocity profile in rotating channel


flow predicted by the new explicit ASM of Gatski and Speziale (1993)
with the experimental data of Johnston, Halleen and Lezius (1972).

flow, in a square duct, in line with experimental observations; on the


other hand, the standard K — e model erroneously predicts that there
is no secondary flow. In order to be able to predict secondary flows
in non-circular ducts, the axial mean velocity vz must give rise to a
non-zero normal Reynolds stress difference rm — rxx (see Speziale and
Ngo 1988). This requires an anisotropic eddy viscosity (any isotro-
pic eddy viscosity, including that used in the standard K — £ model,
Turbulent Transport Equations 209

yields a vanishing normal Reynolds stress difference which makes it


impossible to describe these secondary flows).

Figure 4. Turbulent secondary flow in a rectangular duct: (a) exper-


iments, (b) standard K — e model, and (c) nonlinear K — e model of
Speziale (1987).

In Figure 5, results obtained from the nonlinear K — £ model


are compared with the experimental data of Kim, Kline and John-
210 C. G. Speziale

ston (1980) and Eaton and Johnston (1981) for turbulent flow past
a backward facing step. It is clear that these results are excellent:
reattachment is predicted at x/H sa 7.0 in close agreement with the
experimental data. In contrast to these results, the standard K — e
model predicts reattachment at x/H K 6.25 - an 11% underpredic-
tion (see Thangam and Speziale 1992). This predominantly results
from the inaccurate prediction of normal Reynolds stress differences
in the recirculation zone as discussed by Speziale and Ngo (1988).

2.7 Full Second-Order Closures

These more complex closures are based on the full Reynolds stress
transport equation with turbulent diffusion:

Full second-order closures are needed for turbulent flows with:


(i) Relaxation effects ;
(ii) Nonlocal effects arising from turbulent diffusion that can give
rise to counter-gradient transport.
In virtually all existing full second-order closures for inhomoge-
neous turbulent flows, II^ and peij are modeled by their homoge-
neous forms. The pressure-strain correlation !!,•_,• is modeled as

as discussed earlier. In Section 2.6, the equilibrium limit of the


Speziale, Sarkar and Gatski (SSG) model was provided. For tur-
bulent flows where there are mild departures from equilibrium, the
SSG model takes the form (see Speziale, Sarkar and Gatski 1991)
Turbulent Transport Equations 211

Figure 5. Turbulent flow past a backward facing step: comparison of


the predictions of the nonlinear K — e model with experiments, (a)
Streamlines and (b) turbulent shear stress profiles.

where

The Launder, Reece and Rodi (1975) model is recovered as a special


case of the SSG model when
212 C. G. Speziale

In most applications, at high Reynolds numbers, the Kolmogorov


assumption of local isotropy is typically invoked where

(then, £,-j = §£<% and a modeled transport equation for the scalar
dissipation rate £ is solved that is of the same general form as that
discussed in Section 2.6). However, this assumption is debatable as
discussed by Durbin and Speziale (1991). More generally, a repre-
sentation of the form

can be used where the algebraic model of Speziale and Gatski (1992)
discussed in Section 2.6 is implemented.
The only additional model that is needed for closure in high-
Reynolds-number inhomogeneous turbulent flows is a model for the
third-order diffusion correlation Cijk- This is typically modeled using
a gradient transport hypothesis:

Some examples of commonly used models are as follows:

Hanjalic and Launder (1972) Model

Mellor and Herring (1973) Model

where Cs is a constant (« 0.11). When these models are used in a


full second-order closure, counter-gradient transport effects can be
accounted for.
There is no question that, in principle, second-order closures ac-
count for more physics. This is quite apparent for turbulent flows
exhibiting relaxation effects. The return to isotropy problem is a
prime example where suddenly, at time t = 0, the mean strains in
Turbulent Transport Equations 213

a homogeneous turbulence are shut off; the flow then gradually re-
turns to isotropy (i.e., 6,-j —>• 0 as t —> oo). In Figure 6, results for the
Reynolds stress anisotropy tensor obtained from the Speziale, Sarkar
and Gatski (SSG) and Launder, Reece and Rodi (LRR) models are
compared with the experimental data of Choi and Lumley (1984) for
the return-to-isotropy from plane strain (here, T = Sot/Ko). It is
clear from these results that the models predict a gradual return to
isotropy in line with the experimental data. In contrast to these re-
sults, all two-equation models — including the more sophisticated one
based on an anisotropic eddy viscosity derived herein - erroneously
predict that at T = 0, fyj abruptly goes to zero.
It is worth noting at this point that while the SSG model was
derived and calibrated based on near equilibrium two-dimensional
mean turbulent flows, it performs remarkably well on certain three-
dimensional, homogeneously strained turbulent flows. The predic-
tions of the SSG and LRR models for the normal Reynolds stress
anisotropies are compared in Figure 7 with the direct simulations
of Lee and Reynolds (1985) for the axisymmetric expansion (here,
t* = Tt where F is the strain rate).

While the previous results are encouraging, it must be noted that


the Achilles heel of second-order closures is wall-bounded turbulent
flows:

(i) Ad hoc wall reflection terms are needed in many pressure-strain


models (that depend on the distance y from the wall) in order
to mask deficient predictions for the logarithmic region of a
turbulent boundary layer;

(ii) Near-wall models typically must be introduced that depend on


the unit normal to the wall - a feature that makes it virtually
impossible to systematically integrate second-order closures in
complex geometries (see So et al. 1991).

In regard to the former point, it is rather shocking as to what


the level of error is in many existing second-order closures for the
logarithmic region of an equilibrium turbulent boundary layer, when
no ad hoc wall reflection terms are used. This can be seen in Table 2
where the predictions of the Launder, Reece and Rodi (LRR), Shih
and Lumley (SL), Fu, Launder and Tselepidakis (FLT) and SSG
214 C. G. Speziale

Figure 6. Time evolution of the anisotropy tensor in the return to


isotropy problem. Comparison of the predictions of the SSG model
and Launder, Reece and Rodi (LRR) model with the experiment of
Choi and Lumley (1984) (from Speziale, Sarkar and Gatski 1991).

are compared with experimental data (Laufer 1951) for the log-layer
of turbulent channel flow. Most of the models yield errors rang-
ing from 30% to 100%. These models are then typically forced into
agreement with the experimental data by the addition of ad hoc wall
reflection terms that depend inversely on the distance from the wall
— an alteration that compromises the ability to apply the model in
complex geometries where the wall distance is not always uniquely
defined. Only the SSG model yields acceptable results for the log-
Turbulent Transport Equations 215

Figure 7. Time evolution of the anisotropy tensor in the axisymmet-


ric expansion for eo/TKo = 2.45. Comparison of the predictions of
the LRR model and SSG model with the direct simulations of Lee
and Reynolds (1985).

layer without a wall reflection term. This results from two factors:
(a) a careful and accurate calibration of homogeneous shear flow (see
Table 3) and (b) the use of a Rotta coefficient that is not too far re-
moved from one (see Abid and Speziale 1993). The significance of
these results is demonstrated in Figure 8 where full Reynolds stress
computations of turbulent channel flow are compared with the ex-
perimental data of Laufer (1951). It is clear that the same favorable
216 C. G. Speziale

trends are exhibited in these results as with those shown in Table 2


which were obtained by a simplified log-layer analysis.

CHANNEL FLOW
Equilibrium LRR SL FIT SSG Experimental
Values Model Model Model Model Data
611 0.129 0.079 0.141 0.201 0.22
&12 -0.178 -0.116 -0.162 -0.160 -0.16
&22 -0.101 -0.082 -0.099 -0.127 -0.15
&33 -0.028 0.003 -0.042 -0.074 -0.07
SK/e 2.80 4.30 3.09 3.12 3.1

Table 2. Comparison of the model predictions for the equilibrium


values in the log-layer (P/£ — 1) with the experimental data of Laufer
(1951) for channel flow.

HOMOGENEOUS SHEAR FLOW


Equilibrium LRR SL FLT SSG Experimental
Values Model Model Model Model Data
&n 0.152 0.120 0.196 0.218 0.21
bi2 -0.186 -0.121 -0.151 -0.164 -0.16
&22 -0.119 -0.122 -0.136 -0.145 -0.14
&33 -0.033 0.002 -0.060 -0.073 -0.07
SK/e 4.83 7.44 5.95 5.50 5.0

Table 3. Comparison of the model predictions for the equilibrium


values in homogeneous shear flow (P/e = 1.8) with the experimental
data of Tavoularis and Karnik (1989).

The near-wall problem largely arises from the use of homoge-


neous pressure-strain models of the form (53) that are only theo-
retically justified for near-equilibrium homogeneous turbulence. Re-
cently, Durbin (1993) developed an elliptic relaxation model that
accounts for wall blocking and introduces nonlocal effects in the
vicinity of walls - eliminating the need for ad hoc wall damping func-
Turbulent Transport Equations 217

Figure 8. Comparison of full Reynolds stress calculations of chan-


nel flow with the experimental data of Laufer (1951) (O) for Re =
61,600. SSG model; FLT model; - • - LRR model; and
SL model, (a) bn component and (b) &i 2 component (from Abid
and Speziale 1993).
218 C. G. Speziale

tions. While this is a promising new approach, it does not alleviate


the problems that the commonly used pressure-strain models have
in non-equilibrium homogeneous turbulence (the Durbin 1993 model
collapses to the standard hierarchy of pressure-strain models given
above in the limit of homogeneous turbulence). The failure of these
models in non-equilibrium homogeneous turbulence can best be illus-
trated by the simple example shown in Figure 9. This constitutes a
rapidly distorted turbulent flow that, initially, is far from equilibrium
since SKo/eo = 50 (the equilibrium value of SK/e is approximately
5). It is clear from these results that all of the models perform poorly
relative to the DNS of Lee et al. (1990). Even the SSG model, which
does extremely well for homogeneous shear flow that is not far from
equilibrium, dramatically overpredicts the growth rate of the turbu-
lent kinetic energy for this strongly non-equilibrium test case.
In the opinion of the author, it is a vacuous exercise to de-
velop more complex models of the form (53) using non-equilibrium
constraints such as Material Frame-Indifference (MFI) in the two-
dimensional limit (Speziale 1981, 1983) or realizability (Schumann
1977 and Lumley 1978). While these constraints are a rigorous con-
sequence of the Navier-Stokes equations, they typically deal with
flow situations that are far from equilibrium (two-dimensional tur-
bulence and one or two-component turbulence) where (53) would not
be expected to apply in the first place. Ristorcelli, Lumley and Abid
(1995) - following the earlier work by Haworth and Pope (1986) and
Speziale (1989) - developed a pressure-strain model of the form (53)
that satisfies MFI in the 2-D limit. Shih and Lumley (1985) at-
tempted to develop models of the form (53) that satisfy the strong
form of realizability of Schumann (1977). Reynolds (1987) has at-
tempted to develop models of this form which are consistent with
Rapid Distortion Theory (RDT). All of these models involve com-
plicated expressions for Mjjkt that are nonlinear in 6;r From its
definition,

which is linear in the energy spectrum tensor Eke(n,t). Since,


Turbulent Transport Equations 219

Figure 9. Comparison of the SSG, SL and FLT model predictions


for the time evolution of the turbulent kinetic energy with the DNS
results of Lee, Kim and Moin (1990) for homogeneous shear flow
(SKQ/£Q — 50) (from Speziale, Gatski and Sarkar 1992).

where

it follows that models for Mijke that are nonlinear in 6;j are also
nonlinear in E^j. This is a fundamental inconsistency that dooms
these models to failure. It is clear that it is impossible to describe a
range of RDTflows- which are linear - with nonlinear models (the
principle of superposition is violated). Furthermore, Shih and Lum-
ley (1985, 1993) unnecessarily introduce higher degree nonlinearities
and non-analyticity to satisfy realizability. In the process of doing
220 C. G. Speziale

so, they arrive at a model that is neither realizable nor capable of


describing even basic turbulent flows (see Speziale, Abid and Durbin
1994, Durbin and Speziale 1994 and Speziale and Gatski 1994).
Entirely new non-equilibrium models are needed for the pressure-
strain correlation and the dissipation rate tensor. The former should
contain nonlinear strain rate effects and the latter should account
for the effects of anisotropic dissipation and non-equilibrium vortex
stretching (see Bernard and Speziale 1992 and Speziale and Bernard
1992). Models of this type are currently under investigation.

3 COMPRESSIBLE TURBULENCE

The full equations of motion for an ideal gas will be considered:

Continuity

Momentum

Energy
Turbulent Transport Equations 221

where
p = mass density
Ui = velocity vector
p = thermodynamic pressure
[JL = dynamic viscosity
<Tj-j = viscous stress tensor
T = absolute temperature
K = thermal conductivity
R = ideal gas constant
Cv = specific heat at constant volume
$ = viscous dissipation
d
(\ \A*• -= dxi\( \)

3.1 Compressible Reynolds Averages

For any flow variable f, we can introduce the decomposition

where T is the standard ensemble mean, or the alternative decom-


position

where

is the Favre (or mass-weighted ) average. Here,

3.2 Compressible Reynolds-Averaged Equations

The Reynolds-aver aged equations of motion for an ideal gas are


given by (cf. Cebeci and Smith 1974 and Speziale and Sarkar 1991):

Continuity
222 C. G. Speziale

Momentum

Energy

where turbulent fluctuations in Cv have been neglected.

Molecular Diffusion Terms


In high-Reynolds-number turbulent flows, the molecular diffusion
terms are dominated by the turbulent transport terms except in a
thin sublayer near walls. If we assume in this region that fluctuations
in the viscosity, thermal conductivity and density can be neglected
we can then make the approximations:

The mean viscous dissipation is given by:

where £ = cr'-u^-fp is the turbulent dissipation rate ($ -> pe as


Re -> oo).
Hence, in order to achieve closure, models are needed for:

(i) The Favre-averaged Reynolds stress, rt-j


Turbulent Transport Equations 223

(ii) The Favre-averaged Reynolds heat flux, Qi

(Hi) The turbulent mass flux, u"

(iv) The turbulent dissipation rate, e

(v) The pressure-dilatation correlation, p'u'^.

This makes the problem of compressible turbulence modeling far


more difficult than its incompressible counterpart.

3.3 Compressible Reynolds Stress Transport Equation

The Reynolds stress transport equation for compressible turbu-


lent flows takes the form (cf. Speziale and Sarkar 1991)

where

are, respectively, the third-order turbulent diffusion correlation, the


deviatoric part of the pressure gradient-velocity correlation and the
dissipation rate tensor.
224 C. G. Speziale

Transport Terms
Speziale and Sarkar (1991) proposed the following models for the
turbulent transport terms:

where

More recently, from an analysis of the transport equation for the


turbulent mass flux, it has been argued that

to the lowest order when an equilibrium hypothesis of the boundary


layer type is invoked (see Zeman 1993 and Ristorcelli 1993). Here,
Mt EE (2K/~/RT)1/2 is the turbulent Mach number.

Dissipation rate model


The Kolmogorov assumption of local isotropy is typically used where

for turbulence Reynolds numbers Rt >• 1. The turbulent dissipation


rate can be decomposed into solenoidal and compressible parts:
Turbulent Transport Equations 225

respectively, as proposed independently by Zeman (1990) and Sarkar


et al. (1991) where

given that w,' is the fluctuating vorticity. Then, the Sarkar et al.
(1991) model can be used for ec yielding

where

The solenoidal dissipation es is a solution of the modeled dissi-


pation rate transport equation (see Speziale and Sarkar 1991):

where

and turbulent fluctuations in the molecular viscosity have been ne-


glected.

Pressure-Dilatation Model
For turbulent shear flows, Sarkar (1992) proposed the following model
for the pressure-dilatation correlation:

where

It should be noted that for turbulent flows with mean dilatation, a


term proportional to
226 C. G. Speziale

should be added (see Sarkar 1992).


Deviatoric Pressure Gradient- Velocity Correlation Iil}
A variable density extension of the pressure-strain model of Speziale,
Sarkar and Gatski (1991) (the SSG model) is used for II,-j:

where

With this model, we now have a complete second-order closure for


high-speed compressible turbulent flows.

3.4 Compressible Two-Equation Models

Two-equation models with an anisotropic eddy viscosity can be


systematically derived from the full compressible second-order clo-
sure presented in Section 3.3 via the same homogeneous equilibrium
hypothesis used earlier for incompressible flows. This, to the lowest
order, yields
Turbulent Transport Equations 227

where C*, C*D and C*E are variable density extensions of the incom-
pressible coefficients derived earlier, in Eq. (32), that depend on
both 77 and £.
This anisotropic eddy viscosity model can be solved in conjunc-
tion with modeled transport equations for K and es that take the
form

where

and the compressible terms e c , p'w''» and u" are modeled as before.

3.5 Illustrative Examples

A variety of examples of compressible homogeneous turbulent


flows will first be presented based on the full compressible second-
order closure provided in turbulent kinetic energy in compressible
isotropic turbulence are compared with the direct numerical simula-
tions of Sarkar et al. (1991) for a variety of turbulent Mach numbers.
The model is able to correctly capture the increase in the decay rate
of the turbulent kinetic energy that arises from compressible effects
- a feature that was traced to the compressible dissipation by Sarkar
et al. (1991).
228 C. G. Spezi&le

Figure 10. The decay of turbulent kinetic energy in compressible iso-


tropic turbulence for initial turbulence Mach numbers Mt,o = 0.1,0.3,
and 0.4. (a) Direct numerical simulations and (b) Model predictions
(from Speziale and Sarkar 1991 and Sarkar et al. 1991).
Turbulent Transport Equations 229

The second homogeneous flow that will be considered is the rapid


compression or expansion of isotropic turbulence. For this problem,
the mean velocity gradient tensor is given by

where F is the expansion/compression rate which is constant. The


model provided herein reduces approximately to the simple coupled
nonlinear ODE's:

for \T\K0/e0 > 1. The short-time solution to Eqs. (90) - (91) is


given by:

where A = li'3//2/e is the integral length scale. These are identical


to the results obtained by Reynolds (1987) based on Rapid Distor-
tion Theory (RDT). In contrast to these results, a variable density
extension of the commonly used second-order closures erroneously
predicts that

(see Reynolds 1987). According to (95), the integral length scale will
decrease under an expansion (F > 0) and increase under a compres-
sion (F < 0) - results that are clearly in error as first pointed out by
Reynolds (1987).
Finally, in regard to homogeneous turbulence, we will consider the
problem of compressible homogeneous shear flow. Here, an initially
isotropic turbulence is subjected to a uniform shear rate S with the
corresponding mean velocity gradient tensor
230 C. G. Speziale

The time evolution of the turbulent kinetic energy and turbulent dis-
sipation rate in compressible homogeneous shear flow obtained from
the SSG model with the dilatational models of Sarkar are compared
with the DNS results of BlaisdeU et al. (1991), for Mto = 0 and
Mt0 = 0.307, in Figure 11. It is clear from these results that the
model does an excellent job in reproducing the dramatic reduction
in the growth rate that arises from compressible effects (without
explicit dilatational terms, all existing second-order closures drasti-
cally overpredict the growth rate of the turbulent kinetic energy). In
Table 4, the equilibrium values obtained using a variety of pressure-
strain models - combined with the dilatational models of Sarkar -
are compared with the DNS results of BlaisdeU et al. (1991). Here,
the results are not so favorable: the normal Reynolds stress anisot-
ropies are drastically underpredicted and the Reynolds shear stress
anisotropy is overpredicted by more than 25%.

Equilibrium LRR SSG FLT DNS


Values Model Model Model Data
bu 0.166 0.230 0.189 0.424
b\i -0.187 -0.165 -0.148 -0.118
&22 -0.130 -0.148 -0.138 -0.236
&33 -0.036 -0.082 -0.051 -0.188
Mt 0.65 0.60 0.65 0.51

Table 4. Predicted equilibrium values for compressible homogeneous


shear flow using the Launder, Reece and Rodi (LRR), Speziale,
Sarkar and Gatski (SSG) and Fu, Launder and Tselepidakis (FLT)
pressure-strain models with the dilatational models of Sarkar. (Taken
from Speziale, Abid and Mansour 1994).

It can thus be concluded that with the addition of the newer


dilatational models, existing second-order closures properly predict
the reduced growth rate of the turbulent kinetic energy in compress-
ible homogeneous shear flow. However, the models are currently not
capable of predicting the equilibrium Reynolds stress anisotropies ac-
curately. This deficiency arises from the use of variable density ex-
tensions of incompressible pressure-strain models - a deficiency that
Turbulent Transport Equations 231

Figure 11. Time evolution of the turbulent fields in compressible


homogeneous shear flow using the SSG model with the dilatational
models of Sarkar for Mtfl = 0.307; Mt,0 = 0; O DNS of
Blaisdell et al. (1991). (a) Turbulent kinetic energy and (b) turbulent
dissipation rate (from Speziale, Abid and Mansour 1994).
232 C. G. Speziale

must be overcome in future research.


The same kind of improved predictions for the reduction in the
growth rate of compressible homogeneous shear flow are obtained for
the spreading rate in the supersonic mixing layer (see Figure 12) with
this type of compressible model. In Figure 13, the spreading rate
predicted by a simple second-order closure both with and without
the dilatational models of Sarkar are compared with experimental
data, as computed by Sarkar and Balakrishnan (1991). With
the addition of the dilatational terms, it now becomes possible to
predict the dramatic reduction in the spreading rate that arises from
high-speed compressible effects.

Figure 12. Schematic diagram of the supersonic mixing layer.


Turbulent Transport Equations 233

Figure 13. Normalized spreading rate as a function of the convective


Mach number Mc: Comparison of the model predictions with ex-
perimental data for the supersonic mixing layer. Second-order
closure with dilatational terms; Second-order closure without
dilatational terms; o Experimental Langley curve.

Interesting enough, good results are obtained for the compressible


flat plate boundary layer for Mach numbers M^ as large as 8 and
wall temperature ratios as low as 0.3 using properly calibrated two-
equation models with no explicit dilatation terms, as shown by Zhang
et al. (1993) (see Figure 14). In fact, it must be noted that the
compressible dissipation model of Sarkar causes a degradation of the
skin friction predictions; the model is not formally correct in the log-
layer of high-speed compressible boundary layers. This was probably
first pointed out by P. G. Huang (Huang 1990 and Huang et al.
1994). Consequently, more research is also needed on the modeling
of dilatational terms in wall-bounded turbulent flows.
234 C. G. Speziale

Figure 14. Comparison of the predictions of the K'—e model of Zhang


et al. (1993) for the compressible flat plate boundary layer with the
experimental data of Kussoy and Horstman (1991) for M^ — 8.18
and Tw/Taw — 0.3: (a) Mean velocity and (b) mean temperature.

4 CONCLUDING REMARKS

Some significant progress in turbulence modeling has been made


during the past two decades. The following general conclusions can
Turbulent Transport Equations 235

be drawn about the current status of incompressible turbulence mod-


els:

(1) There is a relatively sound theoretical basis for the existing


type of Reynolds stress models in two-dimensional mean tur-
bulent flows that are close to equilibrium and only weakly in-
homogeneous. A new generation of two-equation models and
second-order closures has emerged that provides a surprisingly
good description of these flows.

(2) More research is needed to place Reynolds stress models on


solid theoretical grounds for three-dimensional, non-equilibrium
turbulent flows. In addition, new methods for the integration
of second-order closures to a solid boundary in complex geome-
tries are required. Until such methods are more fully developed,
it is preferable to use two-equation models - with an anisotropic
eddy viscosity systematically obtained from a second-order clo-
sure - in complex wall-bounded turbulent flows. Fundamental
changes in the way second-order closures are formulated may
be needed to overcome this problem.

Compressible Reynolds stress models are not as well developed


as their incompressible counterparts. Significant progress has been
made in the modeling of turbulent dilatational terms, however, much
more research is still needed. Most importantly, fundamental new re-
search on the compressible modeling of the pressure-strain correlation
is needed. The modeling of turbulent dilatational terms in high-speed
compressible wall-bounded flows also needs more attention. Despite
these unresolved issues, there is still cause for optimism. With fur-
ther progress, we should start to see Reynolds stress models make a
major impact on the computation of the complex turbulent flows of
technological importance.

ACKNOWLEDGMENTS

Most of this work was funded by the National Aeronautics and


Space Administration under Contract NAS1-19480 while the author
was in residence at ICASE. Partial funding by the Office of Naval
Research under Grant N00014-94-1-0088 (ARI on Nonequilibrium
236 C. G. SpezJale

Turbulence, Dr. L. P. Purtell, Program Officer) is also gratefully


acknowledged.

5 REFERENCES
Abid, R. and Speziale, C. G., 1993. "Predicting equilibrium states
with Reynolds stress closures in channel flow and homogeneous
shear flow," Phys. Fluids A 5, pp. 1776-1782.

Baldwin, B. S. and Barth, T. J., 1990. "A one-equation turbu-


lence transport model for high Reynolds number wall-bounded
flows," NASA TM-102847.

Baldwin, B. S. and Lomax, H., 1978. "Thin-layer approximation


and algebraic model for separated turbulent flows," AIAA Pa-
per No. 78-257.

Bardina, J., Ferziger, J. H. and Reynolds, W. C., 1983. "Improved


turbulence models based on large-eddy simulation of homo-
geneous, incompressible turbulent flows," Stanford University
Technical Report No. TF-19.

Bardina, J., Ferziger, J. H. and Rogallo, R. S., 1985. "Effect of


rotation on isotropic turbulence: Computation and modeling,"
J. Fluid Mech. 154, pp. 321-336.

Bernard, P. S. and Speziale, C. G., 1992. "Bounded energy states


in homogeneous turbulent shear flow - An alternative view,"
ASME J. Fluids Eng. 114, pp. 29-39.

Blaisdell, G. A., Mansour, N. N. and Reynolds, W. C., 1991. "Nu-


merical simulations of compressible homogeneous turbulence,"
Stanford University Technical Report No. TF-50.

Cebeci, T. and Smith, A. M. O., 1974. Analysis of Turbulent Bound-


ary Layers, Academic Press, New York.

Choi, K. S. and Lumley, J. L., 1984. "Return to isotropy of ho-


mogeneous turbulence revisited," in Turbulence and Chaotic
Phenomena in Fluids (T. Tatsumi, ed.), pp. 267-272, North
Holland, New York.
Turbulent Transport Equations 237

Comte-Bellot, G. and Corrsin, S., 1971. "Simple Eulerian time


correlation of full- and narrow-band velocity signals in grid-
generated isotropic turbulence," J. Fluid Mech. 48, pp. 273-
337.

Durbin, P. A., 1993. "A Reynolds stress model for near-wall turbu-
lence," J. Fluid Mech. 249, pp. 465-498.

Durbin, P. A. and Speziale, C. G., 1991. "Local anisotropy in


strained turbulence at high Reynolds numbers," ASME J. Flu-
ids Eng. 113, pp. 707-709.

Durbin, P. A. and Speziale, C. G., 1994. "Realizability of second-


moment closure via stochastic analysis," J. Fluid Mech. 280,
pp. 395-407.

Eaton, J. K. and Johnston, J. P., 1980. "Turbulent flow reattach-


ment: An experimental study of the flow and structure behind
a backward facing step," Stanford University Report No. MD-
39.

Fu, S., Launder, B. E. and Tselepidakis, D. P., 1987. "Accommo-


dating the effects of high strain rates in modeling the pressure-
strain correlation," UMIST Technical Report No. TFD/87/5.

Gatski, T. B. and Speziale, C. G., 1993. "On explicit algebraic stress


models for complex turbulent flows," J. Fluid Mech. 254, pp.
59-78.

Hanjalic, K. and Launder, B. E., 1972. "A Reynolds stress model


of turbulence and its application to thin shear flows," J. Fluid
Mech. 52, pp. 609-638.

Hanjalic, K. and Launder, B. E., 1980. "Sensitizing the dissipation


equation to irrotational strains," ASME J. Fluids Eng. 102,
pp. 34-40.

Haworth, D. C. and Pope, S. B., 1986. "A generalized Langevin


model for turbulent flows," Phys. Fluids 29, pp. 387-405.

Hinze, J. O., 1975. Turbulence, 2nd ed., McGraw-Hill, New York.

Huang, P. G., 1990. Private Communication.


238 C. G. Speziale

Huang, P. G., Bradshaw, P. and Coakley, T. J., 1994. "Turbulence


models for compressible boundary layers," AIAA J. 32, pp.
735-740.
Johnston, J. P., Halleen, R. M. and Lezius, D. K., 1972. "Effects of
a spanwise rotation on the structure of two-dimensional fully-
developed turbulent channel flow," J. Fluid Mech. 56, pp. 533-
557.
Kim, J., Kline, S. J. and Johnston, J. P., 1980. "Investigation of a
reattaching turbulent shear layer: Flow over a backward facing
step," ASME J. Fluids Eng. 102, pp. 302-308.
Kolmogorov, A. N., 1942, "The equations of turbulent motion in
an incompressible fluid," Isv. Acad. Sci. USSR, Phys. 6, pp.
56-58.
Kussoy, M. I. and Horstman, C. C., 1991. "Documentation of two-
and three-dimensional shock wave/turbulent boundary layer
interaction flows at Mach 8.2," NASA TM-103838.

Laufer, J., 1951. "Investigation of turbulent flow in a two-dimen-


sional channel," NAG A TN 1053.
Launder, B. E. and Spalding, D. B., 1974. "The numerical com-
putation of turbulent flows," Comput. Methods Appl. Mech.
Eng. 3, pp. 269-289.
Launder, B. E., Reece, G. J. and Rodi, W., 1975. "Progress in the
development of a Reynolds stress turbulence closure," J. Fluid
Mech. 68, pp. 537-566.
Lee, M. J. and Reynolds, W. C., 1985, "Numerical experiments on
the structure of homogeneous turbulence," Stanford University
Technical Report TF-24.
Lee, M. J., Kim, J. and Moin, P., 1990. "Structure of turbulence
at high shear rate," J. Fluid Mech. 216, pp. 561-583.
Lumley, J. L., 1978. "Computational modeling of turbulent flows,"
Adv. Appl. Mech. 18, pp. 123-176.
Lumley, J. L., 1992. "Some comments on turbulence," Phys. Fluids
A 4, pp. 203-211.
Turbulent Transport Equations 239

Mellor, G. L. and Herring, H. J., 1973. "A survey of mean turbulent


field closure models," AIAA J. 11, pp. 590-599.

Pope, S. B., 1975. "A more general effective viscosity hypothesis,"


J. Fluid Mech. 72, pp. 331-340.

Prandtl, L., 1925. "Uber die ausgebildete turbulenz," ZAMM 5,


pp. 136-139.

Prandtl, L., 1945. "Uber ein neues formelsystem fur die ausge-
bildete turbulenz," Nachr. Akad. Wiss. Gottingen Math.
Phys. Kl 1945, pp. 6-19.
Raj, R., 1975. "Form of the turbulence dissipation equation as
applied to curved and rotating turbulent flows," Phys. Fluids
18, pp. 1241-1244.

Reynolds, W. C., 1987. "Fundamentals of turbulence for turbulence


modeling and simulation," in Lecture Notes for Von Kdrmdn
Institute, AGARD Lect. Ser. No. 86, pp. 1-66, NATO, New
York.
Ristorcelli, J. R., 1993. "A representation for the turbulent mass
flux contribution to Reynolds stress and two-equation closures
for compressible turbulence," ICASE Report No. 93-88, NASA
Langley Research Center.

Ristorcelli, J. R., Lumley, J. L. and Abid, R., 1995. "A rapid-


pressure correlation representation consistent with the Taylor-
Proudman theorem materially-frame-indifferent in the 2-D
limit," J. Fluid Mech. 292, pp. 111-152.

Rodi, W., 1976. "A new algebraic relation for calculating the Rey-
nolds stresses," ZAMM 56, pp. T219-221.
Rogers, M. M., Moin, P. and Reynolds, W. C., 1986. "The structure
and modeling of the hydrodynamic and passive scalar fields in
homogeneous turbulent shear flow," Stanford University Tech-
nical Report No. TF-25.
Rubinstein, R. and Barton, J. M., 1990. "Nonlinear Reynolds stress
models and the renorrnalization group," Phys. Fluids A 2, pp.
1472-1476.
240 C. G. Speziale

Sarkar, S., 1992. "The pressure-dilatation correlation in compress-


ible flows," Phys. Fluids A 4, pp. 2674-2682.

Sarkar, S. and Balakrislman, L., 1991. "Application of a Reynolds


stress turbulence model to the compressible shear layer," AIAA
J. 29, pp. 743-749.

Sarkar, S., Erlebacher, G., Hussaini, M. Y. and Kreiss, H. O., 1991.


"The analysis and modeling of dilatational terms in compress-
ible turbulence," J. Fluid Mech. 227, pp. 473-493.
Schumann, U., 1977. "Realizability of Reynolds stress turbulence
models," Phys. Fluids 20, pp. 721-725.
Shih, T. H. and Lumley, J. L., 1985. "Modeling of pressure cor-
relation terms in Reynolds stress and scalar flux equations,"
Cornell University Technical Report FDA-85-3.

Shih, T. H. and Lumley, J. L., 1993. "Critical comparison of second-


order closures with direct numerical simulations of homoge-
neous turbulence," AIAA J. 31, pp. 663-670.
Smagorinsky, J., 1963. "General circulation experiments with the
primitive equations," Mon. Weather Review 91, pp. 99-165.

So, R. M. C., Lai, Y. G. and Hwang, B. C., 1991. "Second-order


near-wall turbulence closures: A review," AIAA J. 29, pp.
1819-1835.
Spalart, P. R. and Allmaras, S. R., 1992. "A one-equation turbu-
lence model for aerodynamic flows," AIAA Paper No. 92-439.

Speziale, C. G., 1981. "Some interesting properties of two-dimen-


sional turbulence," Phys. Fluids 24, pp. 1425-1427.
Speziale, C. G., 1983. "Closure models for rotating two-dimensional
turbulence," Geophys. & Astrophys. Fluid Dyn. 23. pp. 69-
84.
Speziale, C. G., 1987. "On nonlinear K - i and K - e models of
turbulence," J. Fluid Mech. 178, pp. 459-475.
Speziale, C. G., 1989. "Turbulence modeling in noii-inertial frames
of reference," Theoret. & Comput. Fluid Dyn. 1, pp. 3-19.
Turbulent Transport Equations 241

Speziale, C. G., 1991. "Analytical methods for the development


of Reynolds-stress closures in turbulence," Ann. Rev. Fluid
Mech. 23, pp. 107-157.

Speziale, C. G. and Ngo, T., 1988. "Numerical solution of turbu-


lent flow past a backward facing step using a nonlinear K — £
model," Int. J. Eng. Sci. 26, pp. 1099-1112.

Speziale, C. G. and Sarkar. S., 1991. "Second-order closure models


for supersonic turbulent flows," AIAA Paper No. 91-0217.

Speziale, C. G. and Bernard, P. S., 1992. "The energy decay in


self-preserving isotropic turbulence revisited," J. Fluid Mech.
241, pp. 645-667.

Speziale, C. G. and Gatski, T. B., 1992. "Modeling anisotropies in


the dissipation rate of turbulence," Bull. Am. Phys. Soc. 37,
p. 1799.

Speziale, C. G. and Gatski, T. B., 1994. "Assessment of second-


order closure models in turbulent shear flows," AIAA J. 32,
pp. 2113-2115.

Speziale, C. G., Sarkar, S. and Gatski, T. B., 1991. "Modeling the


pressure-strain correlation of turbulence: An invariant dynam-
ical systems approach," J. Fluid Mech. 227, pp. 245-272.

Speziale, C. G., Abid, R. and Durbin, P. A., 1994. "On the read-
ability of Reynolds stress turbulence closures," J. Sci. Comp.
9, pp. 369-403.

Speziale, C. G., Abid, R. and Mansour, N. N., 1994. "Evaluation


of Reynolds stress turbulence closures in compressible homoge-
neous shear flow," ZAMP, to appear.

Taulbee, D. B., 1992. "An improved algebraic Reynolds stress


model and corresponding nonlinear stress model," Phys. Fluids
A 4, pp. 2555-2561.

Tavoularis, S. and Karnik, U., 1989. "Further experiments on the


evolution of turbulent stresses and scales in uniformly sheared
turbulence," J. Fluid Mech. 204, pp. 457-478.
242 C. G. Speziale

Thangam, S. and Speziale, C. G., 1992. "Turbulent flow past a


backward facing step: a critical evaluation of two-equation
models," AIAA J. 30, pp. 1314-1320.
Wilcox, D. C., 1993. Turbulence Modeling for CFD, DCW Indus-
tries and Griffin Printing, Glendale, CA.

Yakhot, V. and Orszag, S. A., 1986. "Renormalization group anal-


ysis of turbulence I. Basic theory," J. Sci. Comp. 1, pp. 3-51.
Yakhot, V., Orszag, S. A., Thangam, S., Gatski, T. B. and Speziale,
C. G., 1992. "Development of turbulence models for shear flows
by a double expansion technique," Phys. Fluids A 4, pp. 1510-
1520.
Yoshizawa, A., 1984. "Statistical analysis of the deviation of the
Reynolds stress from its eddy viscosity representation," Phys.
Fluids 27, pp. 1377-1387.

Zeman, 0., 1990. "Dilatational dissipation: The concept and appli-


cation in modeling compressible mixing layers," Phys. Fluids
A 2, pp. 178-188.

Zeman, O., 1993. "A new model for super/hypersonic turbulent


boundary layers," AIAA Paper No. 93-0897.

Zhang, H. S., So, R. M. C., Speziale, C. G. and Lai, Y. G., 1993.


"Near-wall two-equation model for compressible turbulent
flows," AIAA J. 31, pp. 196-199.
Zhou, Ye., Vahala, G., and Thangam, S., 1994. "Development of a
Turbulence Model Based on Recursion Renormalization Group
Theory," Physical Review E, 49, pp. 5195-5206.
Chapter 6

AN INTRODUCTION TO
SINGLE-POINT CLOSURE
METHODOLOGY

Brian E. Launder

1 INTRODUCTION
1.1 The Reynolds Equations
The material presented in Section 1 will replicate that covered
in earlier chapters. It may nevertheless be helpful to include it to
familiarize the reader both with the nomenclature adopted and the
assumptions being made.
The instantaneous velocity field in a turbulent flow is described
by the continuity and the Navier-Stokes equations which in conser-
vative form may be written:

Here /5,P,/i and F{ denote the density, pressure, viscosity and body
force per unit volume. The tildes appearing above all quantities serve
as a reminder that each, potentially, will display fluctuations due to
turbulence. Here, however, density and viscosity variations will be

243
244 B. E. Launder

assumed sufficiently small to have negligible direct effect on turbu-


lence. A further limitation will be to flows which are incompressible
in the mean, i.e. ones where dUi/ dx{ = 0. The resultant somewhat
simpler pair of equations may then be written

We now proceed to average each term in the above equation.


All except the second term on the left of eq (1.2b), the convective
transport of a;,- momentum, contain only a single fluctuating quantity;
so, the averaging results in the instantaneous value being replaced
by the mean of the variable in question. In the exceptional case

Thus after some rearrangement, the averaged equations of motion


are obtained as:

These equations, generally known as the Reynolds equations, dif-


fer from those describing a laminar flow only by the presence of the
term containing averaged products of fluctuating velocities. The pro-
cess it represents is the additional transfer rate of x,- momentum due
to turbulent fluctuations0. Habitually it is brought to the right side.
"For cases where density variations are large (requiring retention of density
fluctuations) the equations of motion are simplified by the adoption of mass
weighted velocities.
Single-Point Closures 245

Since the first term within the square brackets is the viscous stress,
the second, -puiuj, has, naturally, been re-interpreted as a turbulent
stress or, more usually, the Reynolds stress tensor. As the tensor
is symmetric there are six independent Reynolds stress components.
They are unknown elements in the averaged equations of motion and
the theme of these notes is that of obtaining a satisfactory approxi-
mation to their magnitude.

1.2 Mean Scalar Transport

We consider the transport of some scalar property by the turbu-


lent motion. Its instantaneous value 0 is assumed to be governed by
the equation:

where F is the appropriate molecular diffusivity and Sg is the rate of


creation of the property per unit volume.
Besides conventional sources or sinks it is convenient to imagine
the term Sg absorbing any terms which, for a particular transported
scalar property, do not fit elsewhere. For example, the molecular
diffusion process may be governed by a more elaborate law than
the simple gradient-diffusion relation supposed while, if © stands for
temperature, the time-dependent term in the equation should strictly
be multiplied by the specific heat at constant volume divided by that
at constant pressure, a difference that can be accounted for in Sg.
Applying an averaging to eq (1.5) leads to the following rate
equation for the mean level of the scalar
Thus

The quantity pvivj is known as the mass-weighted Reynolds stress. Notice that,
in contrast with conventional "volume" averaging, the number of turbulent cor-
relations arising from averaging the convective terms is just one - as in a uniform
density flow. This is the feature that has made the use of mass weighted quantities
in variable-density flow so popular.
246 B. E. Launder

An overbar has been placed on Sg to serve as a reminder that the


source term may be a non-linear one and, in that event, its mean
value may differ considerably from that obtained by inserting just
the mean values of the separate constituent terms. Having drawn
attention to this possibility, we shall not consider it further in the
present treatment, attention being limited to cases where the source
or sink term is effectively zero. Just as with the averaged momen-
tum equation, the additional convective flux of the scalar due to
the turbulent velocity fluctuations is conveniently interpreted as a
supplementary diffusional process, so the term puj0 is transferred
to the right side of the equation. Like the Reynolds stresses, this
turbulent scalar flux is an unknown and will correspondingly require
approximation.

1.3 The Modeling Framework

In all practically interesting problems the mean momentum and


continuity equations together, in many cases, with one or more equa-
tion of the type (1.6) for transported scalars are to be solved numer-
ically. The solving procedure will, in all probability, be of fairly gen-
eral construction designed to cope with some class of flow problems:
for example, two-dimensional thin shear flows, three-di- mensional,
confined, non-recirculating flows, etc. Ideally the turbulence model -
the scheme for evaluating v^uj and u$ ~ should enjoy a range of ap-
plicability comparable to that of the numerical procedure and should
fit comfortably within it.
Schemes discussed in this article are single-point closures. In such
approaches the only averaged products of fluctuating quantities that
appear are those in which the two or more quantities in question are
evaluated at the same point in space. Mathieu and Jeandel (1984)
and Leslie (1973) have contributed text books describing 'two-point'
or 'spectral' schemes. Models of this type are seen more as helping
to reveal the underlying physics of turbulence than as models for use
in engineering computational procedures.
Within the single-point framework there is a wide range of mo-
delling methodology. At present most applied computational work
on turbulent flows still adopts the idea that the turbulent fluxes and
Single-Point Closures 247

stresses can be represented in terms of effective turbulent diffusion


coefficients for momentum, heat, chemical species, etc. Approaches
of this type range from simple mean-field closures (where the tur-
bulent viscosity is expressed in terms of the mean velocity field and
flow topography) to the widely used two-equation models where the
effective diffusion coefficients are determined from local values of two
scalar properties of the turbulence (and which may or may not have
a directly measurable physical significance). These in turn are obtai-
ned from transport equations similar to those describing mean flow
quantities save that source and sink terms always play an important
role.
In some respects it makes sense to consider models based on the
idea of a turbulent viscosity first before proceeding to more advanced
second-moment treatments. On balance, however, it seems preferable
here to go directly to a more comprehensive treatment from which
'turbulent viscosity' models emerge as special cases under particular
circumstances. The term 'second-moment' applies to models based
on the exact transport equations for the second moments, i.e. for
U{Uj, UiO etc. These equations, while exact, are unclosed: they con-
tain correlations that are not exactly determinable and which must
therefore be approximated in terms of quantities that are.
Section 2 presents briefly the most popular current approaches to
closure at this level. Before considering closure questions; however,
it will be instructive to examine the exact second-moment equations
and, in particular, the processes causing these quantities to depart
from the levels found in isotropic turbulence. These are topics taken
up below.

1.4 Second-Moment Equations

An exact equation describing the transport of the kinematic Rey-


nolds stress u^Uj is formed by multiplying the momentum equation
for Ui (in which we now use k rather than j as the repeated suf-
fix) by Uj and averaging, then adding to it the mirror equation in
which suffices i and j are interchanged. After a certain amount of
manipulation, the resultant equation may be written:
248 B. E. Launder

The left side of the equation expresses the total rate of increase
of the correlation ujuj for a small identified packet of fluid. The rate
of change arises from an imbalance of the terms on the right. Here
the terms have been grouped, following well-established practice, so
as best to allow a physical interpretation of the processes. One line is
given to each process and, beneath each term, appears a shorthand
symbol for the process in question which we shall use to simplify later
equations. The first two processes represent rates of creation of UiUj,
in one case by the effects of mean strain, P,-j, and in the other by
body forces,G;j. The first of these, comprising products of Reynolds
stresses and mean velocity gradients, can clearly be treated exactly
in a second-moment closure. If the body force is linear, as when one
examines the flow in a rotating coordinate frame, that too can be
handled without further approximation.
The correlation between fluctuating pressure and fluctuating
strain, faj, is a very important one. We note that its trace is zero,
since by continuity

The term thus makes no contribution to the overall level of turbulence


energy but serves to redistribute energy among the normal stress
components (those for which i and j take the same value).
The terms comprising d{j are easily recognized as diffusive in
character since we see from integrating them across a thin shear flow
bounded by non-turbulent fluid that they make no contribution to
the average level ofuiUj at any section even though, within the shear-
flow, the correlations themselves are non-zero. Their effect is thus
Single-Point Closures 249

to promote a spatial redistribution. The last term of dij describes


diffusive transport due to molecular action; over all or nearly all the
flow it will be entirely negligible.
Finally, term e,-j represents (very nearly) the destruction rate of
UiUj by viscous action. Unlike viscous diffusion, the dissipation terms
cannot in general be ignored. We may see this is so by contracting
eq (1.7) to produce an equation for the transport of kinetic energy.
Then, for the thin shear flow discussed above, dkk makes no contri-
bution to the overall level of turbulence energy at any section while
<f>kk vanishes identically at all points. The term P^k will be positive
representing the continual extraction of energy from the mean flow
by the action of the Reynolds stress on the mean shear. Thus, if ekk
were negligible, there would be a limitless growth of the flow's tur-
bulent kinetic energy. Such a scenario is contrary to both intuition
and observation. The crucial difference between e,-j and the viscous
diffusion terms of d^ is that the former comprises correlations of
fluctuating velocity derivatives and, in the finest scales of motion
present, turbulent velocity derivatives are large.
Further consideration of <^-, d±j and Cjj is deferred to Section
2. It will, however, be instructive to examine specific forms that the
production tensors take in a few cases for, in most practical flows, the
production terms firmly stamp the character of the resultant turbu-
lent stress tensor. First, consider the case of simple shear, dUi/dx^
— A , where A is assumed positive. From the upper row of Table 1.1,
Pi2, the production rate of u\u^ is negative (since u\ is undoubtedly
positive); though short of a proof, the reader will accept the likeli-
hood that uiu-2 is consequently negative (that is, of the same sign
as its production rate) which gives in turn the "sensible" result that
the generation of u\ is positive. There is no direct production of
either u\ or M§. This does not mean that there will be no turbulent
fluctuations in the x-i and £3 directions for we have already noted
that the pressure-strain correlation fyj serves to redistribute energy
among the various normal stresses. Nevertheless, we_should expect
— and this is amply confirmed by experiment - that u\ would be the
largest of the normal stress components. The self-sustaining charac-
ter of turbulence in a simple shear is emphasized by the clockwise
rotating arrows connecting the stresses in Figure 1.1: turbulent ve-
locity fluctuations in the direction of mean-velocity gradient promote
a growth in shear stress which, in turn, serves to augment the inten-
250 B. E. Launder

sity of streamwise fluctuations. Pressure interactions deflect some


of this energy to fluctuations in the direction of the velocity gradi-
ent - and so the sequence repeats itself. It is what we might call
turbulence's eternal triangle.

Figure 1.1: Stress couplings in simple and curved shear flows:


Primary strain; _ Secondary(curvature) strain; ~*-~ Pressure-
strain effect.

The lower line in Table 1.1 shows the effects of superimposing on


the primary shear a weak secondary strain 6 (=dU2/dxi) which rep-
resents a curving of the mean-flow streamlines. The shear stress and
the normal stress in the direction of the velocity gradient are directly
modified and their effects reinforce one another. That is to say, if
6 is positive, the extra contribution to P\i will tend to enlarge the
(negative) magnitude of u\ui while that for P22 leads to an enhance-
ment of u| which in turn helps amplify u\U2 through the principal
contribution to P^. It is this mutual reinforcement property of P;J,
represented in Figure 1.1 by the broken lines, that makes turbulent
shear flows so sensitive to weak streamline curvature.
Buoyancy has an effect on turbulence generation that in some
respects is akin to streamline curvature. It is more complicated,
however, for it involves a coupling of the Reynolds stress and scalar
flux fields Ui& . The corresponding equation for u$ is obtained by
multiplying eq (1.5) by U{ then adding it to eq (1.2b) multiplied by
and 0 averaging. The result may be expressed as
Single-Point Closures 251

Pij Pll P"22 -Ps3 -Pi 2

Due to primary

shear, -lH^u^X 0 0 -u\X

^
9Z2 = A

Due to curvature

perturbation, 0 — 2uiu^6 0 —u\d

QUi _ g
9si

Table 1.1: Stress Production Rates due to Primary and Secondary


Shear

The emerging equation is similar in structure to that describing


the transport of Reynolds stress. The generation terms P,# comprise
products of second-moment correlations and mean-field quantities
and will not require approximation. Diffusive transport of u$ (dio} is
caused by velocity and pressure fluctuations and by molecular trans-
252 B. E. Launder

port while pressure fluctuations also play a non-dispersive role (<fog)


which we shall later see is of vital importance.
The unknown processes in the above equation will be approxi-
mated in Section 2. Here we consider briefly the form taken by the
generation terms under a simple temperature gradient dQ/dx^ in a
fluid moving in direction x\. If the background Reynolds stress field
is isotropic (ufUk = \&ikUmu,m) the only direction in which a heat
flux is generated is that of the temperature gradient x2 and the sign
of P-2B is opposite from that of the temperature gradient. This is
also the case in a non-isotropic stress field if the mean shear lies en-
tirely in a plane normal to the temperature gradient (last line Table
1.2). When the direction of mean velocity gradient coincides with
that of the temperature gradient; however, a stream wise scalar flux
is generated, contributions arising from both P{$\ and Pt'02- If we
accept the idea that the stresses and fluxes will be of the same sign
as the generation terms, we can see that the two contributions to the
generation reinforce one another, both being of the same sign as the
product

Pie Pie Pie Pze

Due to a shear ~uiu2^ ~U2§^ ®

dU!/dx2 -^"g-
Duetoashear 0 ~ulj^ °

dUt/dxa

Table 1.2: Heat flux generation rates due to a temperature gradient


8®/dx2

With the above information, it is now possible to infer the effect


of a buoyantly stable stratification of fluid on the stress generation
Single-Point Closures 253

rates. Turbulent variations in density,yO, give rise to a fluctuating


body force per unit pass —p' g/p in the vertical direction, g being
the gravitational acceleration. Thus, in the Reynolds stress equa-
tions (with #2 vertically up) we find non-zero buoyant generation
components as follows

If the fluid in question is a perfect gas p'/p — —6/Q (where the


origin on the 0 scale is absolute zero) and so the generation rates
may be re-written

In a stably stratified medium dQ/x^ is positive; to fix ideas let us


suppose dUi/dx? is also positive. So, P^Q and thus (we suppose) G^
is negative. Likewise, from the above discussion, Pie and hence G\e
will be positive. Thus, there is a double-edged effect of buoyancy
on the Reynolds shear stress. G-22 will tend to reduce u\ thus di-
minishing PI 2 (see Table 1.1) while Gi2, being of the opposite sign
from P12 will also act to suppress the vertical transport of streamwise
momentum by turbulence.

1.5 The WET Model of Turbulence

The foregoing sub-section has obtained the exact transport equa-


tion for UiUj and UiO and examined the form that the generation
terms in these equations take in a few situations. As we have seen,
these generation terms are exact.
Of course, before we can use the transport equations to find the
stress levels, models must be devised for the unknown processes - the
task of Section 2. To round off the Introduction; however, here is a
model to avoid modelling. It is based on the simple economic idea
that

As such it is a gross over simplification; yet it is still more true


than untrue that the more one earns the better off one is; and that
someone of 50, with his house paid for, will be wealthier than someone
254 B. E. Launder

of 25 just starting out on his career. Extrapolating this concept to


the second moments leads naturally to the proposition

Thus:

The choice of turbulent time scale Tt and how to compute it is de-


ferred to the next section. The basic idea can, however, readily be
tested by looking at ratios of the scalar fluxes. Thus:

In a mildly heated shear flow without body forces in which dQ/dx^


and dU\/dx2 are the only non-zero temperature and velocity com-
ponents, one finds from experiment that the left side of eq (1.11) is
approximately -1.3 in a free flow, the right side about -1.6. Near a
wall the ratio of the turbulent heat fluxes is larger, about -2.2, as is
correspondingly the ratio of the generation terms (about -2.1). Thus,
there does indeed seem to be more than a casual connection bet-
ween the left and right sides of eq (1.11). The success is particularly
striking when set against the background of results given by simple
eddy diffusivity models. Such schemes would predict that, because
there are no x\ gradients in mean temperature, the turbulent flux in
that direction would be zero!
When considering the Reynolds stresses we need to apply the
WET concept to departures from the isotropic state. (This is not
inconsistent with eq (1.10); the isotropic value of a vector is of course
zero). Accordingly:

In fact, this formula, arrived at by a much less direct route, has


formed the basis of many successful predictions of turbulent free
shear flows.
Single-Point Closures 255

2 CLOSURE AND SIMPLIFICATION OF THE


SECOND-MOMENT EQUATIONS

2.1 Some Basic Guidelines

Our aim is to mimic those processes not exactly determinable


at second-moment level in terms of mean and turbulence properties
that are. At the practical level there is pressure to adopt the simplest
closure consistent with achieving the desired width of applicability.
This naturally affects the importance that different workers have
attached to different closure principles.
Clearly, as a first requirement, any surrogate form must have the
same dimensions as the correlation it replaces; there is no controversy
on this point. Next, we would also insist that the mathematical
character of the model should conform in various respects with the
original. For example, if the process requiring approximation is a
second-rank, symmetric tensor with zero trace, the search for a model
should be limited to forms possessing these properties. Although
this principle is usually adhered to at second-moment level, it is
frequently ignored in modelling the third moments. This can be
regarded as the application of a further fundamental concept: the
principle of receding influence. Broadly, the idea is that the nth-
moment correlations have markedly less effect on the mean flow than
those of (n — l)th order. So, rules that are held inviolate for second
moments are sometimes dispensed with in the interests of algebraic
and computational convenience in dealing with third moments. It is
clearly a matter of taste and of the flows to be calculated how freely
one invokes this principle. Everyone developing models at this level
makes some use of it; however, for it is that idea which ultimately
legitimatizes second-moment closure.
Two further principles of mathematical physics have been in-
voked in determining modelling approaches. It is generally accepted
that the approximate forms should exhibit the same responses to
translations, accelerations and reflections of coordinate frame as the
real processes (e.g. Donaldson et a/., 1972). The second constraint
is that the modelled set of transport equations should be rendered
physically incapable of generating impossible or 'unrealizable' values
such as negative normal stresses or correlation coefficients (such as
uiuif^/(u\ ^2)) greater than unity, Schumann (1977). Here men-
tion should also be made of the work of Andre et al. (1976) who
256 B. E. Launder

devised a scheme for overwriting or 'clipping' the values of triple


moments whenever they reached physically unattainable values in
comparison with other double and triple moments. Although Schu-
mann suggested ways of securing 'realizability', the interested reader
is referred to the far more detailed treatment by Lumley (1978). Un-
fortunately, although the principle of realizability is a sound one, its
adoption adds considerably to the complexity of the turbulence clo-
sure and only the latest generation of modelling proposals employ
'realizable' forms in computations of inhomogeneous flows, e.g. Shih
and Lumley (1985), Fu et al (1987), Craft and Launder (1989).
Models which in principle are capable of generating impossible val-
ues may, in practice, only do so for flows which are of only academic
interest. Second-moment-closure studies of recirculating, swirling or
other complex flows have been made with forms that do not guaran-
tee realizability. Some would argue that to insist that all models be
fully realizable is rather like requiring that all new buildings in Paris
be as resistant to earthquake damage as those built in San Francisco.
Of course, it could be the case that building in 'earthquake-proof
technology brought other benefits to the overall design of the struc-
ture that justified the expense. There are some indications that
this might be the case provided a sufficiently capacious framework is
adopted. For example, the UMIST group has found that employing
fully realizable closures has enabled the observed reduction of tur-
bulent mixing at high strain rates to be accommodated - a feature
that consistently eluded the simpler models that were not.
The next concept is one that blurs into intuition. The turbulence
modeller should always try to ensure that the proposed form is a
physically plausible substitute for the real process. This statement
relates to the choice of physical quantities; for example, whether the
model should comprise exclusively turbulence correlations or terms
containing mean-field elements - and the way they are combined
together. It is often helpful to explore different ways of expressing
the correlation of interest; one form might give much more insight
than another.
A considerable simplification to the task of turbulence modelling
results from applying the high-Reynolds-number hypothesis. Simply
stated the proposal contains two complementary ideas:
Single-Point Closures 257

1. that the large-scale interactions predominantly responsible for


momentum and scalar transport are unaffected by the fluid's
viscosity;
2. that the fine-scale motions responsible for viscous dissipation
are unaware of the nature of the mean flow and the large-scale
turbulence. Their structure is similar to that found in isotropic
turbulence.
It is becoming increasingly evident that the fine-scale motion,
particularly in regard to higher moment correlations, is not exactly
like isotropic turbulence (e.g. Wyngaard and Sundararajan, 1979),
but nevertheless, if judiciously applied, both aspects of the high-
Reynolds-number hypothesis are very useful in turbulence modelling.

2.2 The Dissipative Correlations

As noted in Section 1, the dissipative correlations c^j and e,-# arise


from the fine-scale motion for it is in these eddies that gradients of
velocity, temperature, etc. are steepest. We assume, therefore, from
the high-Reynolds-number hypothesis, that the motions contributing
predominantly to e^ and e,-0 are isotropic. Now, in isotropic turbu-
lence dui/dxkdd/dxf; changes sign if the coordinate direction x; is
reversed; but the properties of isotropic turbulence are unaffected
by such reflections of axes. The only value that dui/dxkdO/dxk can
take, therefore, and be consistent with isotropic turbulence, is zero.
That is:

Likewise e tj must be expressible as proportional to the product


of the contraction e(=v(duij'5xj)2) and the isotropic unit tensor Sij.
Thus:

where the constant of proportionality is obtained by contracting each


side of the equation. An implication of (2.2) is that there is no viscous
sink of shear stress. The process </>,-j is thus the only mechanism for
preventing the unlimited growth of the off-diagonal components of
UiUj.
There is still little agreement in the literature on whether the
dissipative correlations can be adequately obtained from isotropic
258 B. E. Launder

relations. Direct measurements are not usually sufficiently accurate


to allow conclusions to be drawn. Direct numerical simulations by
the Stanford group and others show significant and systematic de-
partures from local isotropy (Reynolds, 1984) but inevitably these
simulations are at relatively low Reynolds number. In practice, in
applying second-moment closure, it is difficult to disentangle depar-
tures from equations (2.1) and (2.2) from errors in accounting for the
pressure-strain and pressure-scalar gradient processes. So, a common
practice (Lumley, 1978) is to adopt the isotropic relations for e,-j and
€{g and to absorb any departure from isotropy in the dissipation pro-
cesses into the turbulent parts of (f>ij and </>;# whose approximation
is now considered.

2.3 Non-Dispersive Pressure Interactions

By taking the divergence of the Navier-Stokes equations (1.14)


and subtracting its mean part, a Poisson equation is produced with
the fluctuating pressure as its subject:

We may regard terms on the right of eq (2.3) as sources of pres-


sure fluctuations. There are evidently three agencies: non-linear
interactions between fluctuating velocities p^y a process involving
mean velocity gradients, p/2y and one arising from fluctuating body
forces, p(3). It is to be expected, therefore, that a successful model
for the turbulent correlations involving pressure fluctuations (0,-j and
fag ) will contain terms corresponding to these different sources in
eq (2.3). Thus, the pressure-strain process may be written
Single-Point Closures 259

where the primes denote that the quantity in question is evaluated at


a distance r from where faj is determined. It is convenient to write
the above equation in the short-hand form:

An analogous equation for fag^ may likewise be obtained and the


result written as

In principle, approximations to each of the constituent processes


in (2.4) and (2.5) may be developed starting from the solution of the
Poisson equation for p. Analyses of this type have been presented
by Naot et al. (1973), Lin and Wolfshtein (1979), and others. A less
formal approach is more often favoured, however. General surrogate
forms are assumed for the different component parts of faj and fag.
Then, by insisting that the model possess certain symmetry and
contraction properties of the original process and/or that it should
comply with some other physical constraint or experimental data,
the various constants in the model are determined.

Mean-Strain Contributions fajzdiez


The strategy summarized in the above paragraph is well illus-
trated by the 'quasi-isotropic' (QI) model for faj2. This approach
assumes

where aikij is a fourth-rank tensor comprising Reynolds-stress ele-


ments. Note that the postulated model, like p/2\, depends linearly
on the mean velocity gradient. The form of a^ij is unknown at the
outset so one begins by writing the most general form and then uses
the desired symmetry properties to reduce the number of unknown
constant coefficients. If attention is restricted to terms linear in the
Reynolds stress there are five independent coefficients.
260 B. E. Launder

Moreover, reference to eq (2.4a) shows that continuity requires that


aiku = 0 while direct integration of the Poisson equation for the
contraction formed by setting j = k leads to the requirement that

Each of these constraints furnishes two relations that the coeffi-


cients a, /3, etc. must satisfy; for example, the latter requires that

So, the model may be organized in a form with just one undetermined
coefficient. After some algebra eq (2.6) may be expressed

where

Equation (2.7) has been invented and rediscovered many times


in the past twenty years e.g., Launder et al. (1973), Naot et al.
(1973), Lumley (1975, 1978), Reynolds (1984). Different strategies
have been adopted by different workers in fixing the final constant, 6.
The writer's group, like Naot et al. (1973), have chosen S to optimize
the relative stress levels found in simple shear (S = 2/55); Reynolds
(1984) took 6 = 10/77 in order that the resultant expression should
be formally independent of the mean vorticity while Lumley (1978)
takes 6 = -2/33. These vastly different values for S do not have
a major effect on the first of the three groups in eq (2.7) (which
is the major term in components where P,-j is large) but give large
variations in the coefficients of the other two.
The behaviour produced in shear flows by either Reynolds' or
Lumley's value of 6 is quite unacceptable. To compensate, these
workers introduce non-linear terms to a^ij (terms like u\u^ u,Uj/k
etc.). An advantage of this elaboration is that, because of the ad-
ditional unknown coefficients, one can satisfy the "two-component
Single-Point Closures 261

limit." At a wall or an interface, such as a free surface, turbulent


fluctuations normal to the surface vanish faster than the other com-
ponents. This requires that (with xa the coordinate normal to the
interface) <t>aa2 should vanish as u^, goes to zero. If one retains all
the possible quadratic products of Reynolds stress in a^,-;- one finds
that the number of assignable coefficients exactly equals the num-
ber of algebraic equations in satisfying the two-component limit as
well as the other constraints discussed above. The resultant form is
conveniently written (Shih and Lumley, 1985; Fu et al, 1987):

where P is the generation rate of turbulence energy, | Pkk-


Eq (2.7b) has an interesting appearance. The leading term is
the same as in eq (2.7a) with a somewhat smaller coefficient. As we
shall see later, the second term effectively introduces a dependence
on (P/e) into the return-to-isotropy coefficient of faji while the third
introduces more varied non-linear effects. This formulation, despite
the rigour and the respect for realizability implicit in the satisfaction
of the two-component limit, does not lead to acceptable relative stress
levels in simple shear flows: with U\ — E/i(x2) it produces u\ >
u| contrary to both experiment and direct simulations. Shih and
Lumley [4] (1985) overcame this problem by adding a correction term
^2

The parameter A is defined as [1 - 9/8 (A^ - AS)] where A^ and


AS are the second and third invariants of the anisotropic part of the
Reynolds stress:

Az = aijOfci ; As = aikakmUmi where aij = (uTuJ — l / 3 6 i j U k U k ) / k .

It equals unity when the stress field is isotropic but its most useful
property (Lumley, 1978) is that it vanishes in two-component tur-
bulence. It is thus tempting to refer to A as the "flatness factor",
262 B. E. Launder

despite the quite different meaning usually attached to that expres-


sion.
In place of equation (2.7c) one may alternatively include third-
rank products of Reynolds stress in aikij'. in that case there are in
total twenty independent groups. On applying the same kinematic
constraints as for the quadratic model one arrives, after considerable
algebra, at the following form (Fu et al., 1987):

where

The first two lines of this equation are identical to the quadratic
model while the coefficients c-2 and c'2 of the cubic terms are freely
assignable. The quantities that these coefficients multiply are evi-
dently of very different length, that associated with c'2 being particu-
larly long. It has been found, however, that the behaviour of simple
free shear flows can be very satisfactorily mimicked by setting c% to
0.6 and c'2 to zero, Fu et al. (1987) (see also Launder (1989)), thus
reducing the algebra to more manageable proportions.
Notice that

where fij/t is the vorticity tensor (dUj/dxk — dUk/dxj). Thus, if


c2 is taken as zero, the cubic terms are identically zero in an irrota-
tional deformation and they can be seen as purely a correction for
rotationality.
Single-Point Closures 263

Recently at UMIST there has been a tendency to include non-


zero values of c'2. In fact, in his original study of homogeneous shear
flows, Fu (1988) showed that the choice 02 = 0.55, c'2 = 0.6
led to an improved prediction of the stress tensor in strong simple
shears 6 . Li (1993) (see Launder and Li, 1993, 1994) found that when
these same values were adopted for near-wall inhomogeneous flows
they were more successful in reproducing the increasing anisotropy as
the wall was approached than the earlier combination (0.6 and 0.0).
Examples of predictions made with both sets are presented later.

The Isotropization of Production (IP) Model

In fact, to date, most computations of complex flows have been


made with a simpler model of 0^2 than any of the above. Sometimes
called the Isotropization of Production model, it supposes that the
net effect of the mean-strain part of the pressure strain process is
effectively to reduce the anisotropy of the production tensor. Thus
one postulates

</>i# = - C2 (Pij - \^Pkk). (2.8)

While the proposal was inspired by intuition (Naot et a/., 1970)


it is interesting to note that if, as is often the case, the value of c%
is taken as 0.6, in isotropic turbulence (uj/uj = 1/3^,-jupZfc"), one
obtains the exact result obtained by Crow (1968)

Equation (2.8) can be regarded as a simplification of equation


(2.7d) in which simply the leading term - indeed the only linear
term of that equation - is retained. It has been widely applied in a
diversity of flows and has been found to be conclusively better than
the superficially more general linear form, eq (2.7a), given earlier, at
least when used in conjunction with the simple linear model of faji
(discussed below).
The idea underlying the IP model is readily applied to the force
field part of </>,j and, moreover, to the mean-strain contribution of
the pressure-scalar gradient contribution fooi- Thus
b
The shear rate may be characterized by the dimensionless parameter
S = k ( d U i / d x 2 ) / e . Local equilibrium is attained for S ~ 3.0 ; 'strong'
shear relates to values of 5 of 5 or greater.
264 B. E. Launder

With the coefficient c2# set to about 0.5 this formula is distinctly more
successful than the linear quasi-isotropic form (i.e. the equivalent
of eq (2.7a) for 0jj) which in t his case is obtained with no free
coefficient:

The latter form (Launder, 1973) is incompatible with predicting the


correct ratio of heat fluxes down and at right angles to the mean
temperature gradient in a simple shear flow. The fact that it has nev-
ertheless been quite widely used in predicting boundary layer flows
reflects that turbulent heat fluxes in the streamwise direction are
usually not important (being far outweighed by mean convection).
The problems with eq (2.10) can be largely overcome by proceeding,
within the 'quasi-isotropic' analysis, to include terms up to quadratic
order in the Reynolds stresses. The approximation level is then for-
mally equivalent to that of eq (2.7d). The resultant form is of similar
length to that equation and is not reproduced here. Successful ap-
plications of the model have, however, been reported by the UMIST
team (Craft and Launder, 1989; Cresswell et al. 1989) in a range of
free shear flows.

Effects of Force Fields <pij3, 4*i93

Because of the considerable variety of possible 'force' fields and


of source or sink terms in the transport equation for U{ and 0, no
general statement can be made on the effect that they bring to (^3
and (f>ies- In engineering, the two most important force fields are
those due to buoyancy and to rotating the coordinate frame (Cori-
olis forces) to facilitate, for example, the study of flows in rotating
passages. In these cases, the QI model leads to a form identical to
that of the IP approach (Launder, 1975, 1976: Cousteix and Aupoix,
1981; Hah, 1981), namely

The QI analysis gives a value of the coefficient 03 of 0.3 for buoyant


flows though those adopting the IP approach have generally preferred
to choose a value akin to that used for mean strain effects; a value
of 0.5 to 0.6 is typical. Correspondingly, for 4>ig3 we find
Single-Point Closures 265

where c^g from a QI analysis equals 1/3 for buoyant flows while,
again, most applications have tended to chose a value of about 0.5.
The accuracy with which these simple models can hope to capture
the real processes hardly justifies a separate optimization for each
constant. A choice in the range 0.3 to 0.5 would nearly always be
appropriate.
In the case of Coriolis forces, there are grounds for taking the co-
efficient €3 to be half that adopted for c%, a topic discussed extensively
by Launder et al. (1987). To illustrate this point, an axisymmetric
swirling jet can be analysed perfectly well in a stationary reference
frame but it can equally well be examined with the coordinate frame
rotating about the jet's axis. If, however, one takes any ratio of 03 : c2
other than one half, the predicted results are dependent on the rate of
rotation of the coordinate system. Obviously, the predictions should
be independent of the observer's frame of reference.
In fact, there are various ways of ensuring independence of the
results from the observer's motion as discussed by Fu et al. (1987).
One is the practice noted above; another is to include the convec-
tive transport tensor DUiUj/Dt - which is hereafter denoted dj -
along with the production tensor and Coriolis tensors in applying
the isotropization of production approach:

where Fij denotes the Coriolis tensor

Fu et al. (1987) found the use of eq (2.13) brought great improve-


ments to their predictions of swirling jets while in non-swirling flows
the net contribution of terms containing dj was small.

The Turbulence-Turbulence Part of faj and fag

To model the parts of 4>ij and fag arising from p^ we seek forms
containing only turbulent quantities. Experiments indicate that grid
turbulence made strongly non-isotropic by passing it through a duct
of rapidly changing cross-sectional shape will revert towards isotropy
once the mean strain is removed. In the absence of any alternative
266 B. E. Laundei

process, we must conclude that <$>(j\ (which, like the other parts o:
(f>ij, is traceless) is the agency promoting this reversion. In mos1
computations of complex flows Rotta's (1951) linear return mode'
has been adopted:

More elaborate versions have been proposed (Lumley, 1978, Rey-


nolds 1984, Fu et a/., 1987) which are conveniently expressed through
the dimensionless anisotropic stress a,-j:

The second term in eq (2.15) produces an interesting asymmetry of


response. Suppose all the ajj's are zero except an = £ and a^i — —£
(so A? = 2<5 2 ). If c( is positive, </>m takes a larger negative value
than the (positive) value of ^221- Consequently u\ will tend to revert
more rapidly to isotropy from above than will u2. from below, which
in turn means that 0,33 will increase from its initial zero value.
The coefficients Cj and c" may be chosen so that the resultant
expression for </>,-ji exactly satisfies the two-component limit0. If, as
before, we take u\ — k(l + <5), u\ = k(l — <5), the requirement that
</>22i should vanish irrespective of the size of 6 leads to c^ = 3/2,
c" = —3/4. While we have experimented at UMIST with these
values, a problem that their use introduces is that, for moderate
levels of anisotropy (A 2 in the range 0.2 - 0.5, say), experiments and
computer simulations on the return to isotropy of highly anisotropic
stress fields indicate a faster proportionate rate of return than when
the stress field is only weakly anisotropic. The negative value of c'{
on its own produces the reverse of the desired effect. One can, in
principle, offset the consequences of a negative c!{ by causing c\ to
increase rapidly with A 2 . No firm proposals of this type have so far
been advanced, however. A simpler way of making </>,-ji vanish in
two-dimensional turbulence is to arrange that c\ should contain the
factor An. Lumley's group prefer this route and take the coefficients
c[ and c'{ as zero. At UMIST a non-zero value of c( is retained, the
presently adopted form being:
c
Clearly, Rotta's original form does not satisfy this limit since, if one compo-
nent of Reynolds normal stress vanishes, its value of ai} is — 2/3 ; so (juj for this
component equals +2/3 cie rather than zero.
Single-Point Closures 267

The corresponding process fagi in the U{6 equation is convention-


ally modelled, following Monin (1965), as:

More elaborate forms have been proposed (Samaaweera, 1978;


Lumley, 1978; Craft and Launder, 1989)

while several workers (Launder, 1976; Elghobashi and Launder, 1983;


Jones and Musonge, 1983) have suggested the partial or complete
replacement of the dynamic time scale k/e by the scalar time scale
1/2 ¥/Cg.

Optimum Choice of Coefficients in Basic Pressure-Strain Model

Equations (2.8) and (2.14) have been the basis of many differ-
ent proposals and have been incorporated into several commercial
software packages. For this reason the pair of equations (including,
where appropriate, wall-reflection terms as discussed below) is often
referred to as the basic model.
The question arises, however, as to what values should be as-
signed to the coefficients c\ and c 2 . Values proposed for c\ range
from 1 to 5 while recommendations for c2 cover the range from zero
to 0.8, Fig 2.1. The range of different choices suggests, at first glance,
that since such disparate values have been put forward the whole
approach is worthless. Looked at with an engineer's eye, however,
one might be tempted to fit a straight line through the 'data points'.
Now, in the case of a simple shear flow in local equilibrium (i.e. where
turbulence generation and dissipation processes are in balance) it is
readily shown that with this basic model the resultant stress ten-
sor depends not on the individual values of c\ and c2 but rather on
the single parameter (1 — c 2 )/ci. The line in Fig 2.1 is simply that
corresponding to (1 — c 2 )/ci = 0.23 which evidently does rather a
good job of fitting the various proposals. What we conclude is that,
for equilibrium simple shear flows, the very different pairs of ci and
c2 will lead to nearly the same results. In order to pick the 'best'
pairing one needs to look at non-equilibrium cases. We have already
268 B. E. Launder

Figure 2.1: Map of proposals for coefficients in Basic Model of <j>ij.

noted that a value of c2 of 0.6 exactly describes the case of isotropic


turbulence subjected to rapid distortion while direct simulations of
the return of anisotropic turbulence towards isotropy suggest a level
of GI of from 1.5 to 2.0 if the level of stress anisotropy is similar to
that found in a typical free shear flow. The pairing usually adopted
nowadays of 1.8 and 0.6 is fully compatible with these extreme cases
and is marked by a triangle in Fig 2.1. Before leaving this topic,
mention should be made of Younis' (1984) pairing since this is the
most recent entry in the table. His proposal arose from the failure to
predict swirling jets correctly when larger values of c2 were adopted.
However, as has been earlier noted, PZJ is not an objective tensor. If
the model of <f>ij is made objective (as it should be) by including the
stress convection tensor dj (eq (2.13)) then it is found that reason-
ably satisfactory predictions of the swirling jet are obtained with the
standard values of GI and cj, Fu et al. (1987).
Single-Point Closures 269

Wall Effects on ^

Although sometimes ignored, in near-wall flows pressure reflec-


tions from the surface have a significant effect on the pressure-strain
correlation. It isn't just rigid walls where surface effects on <j>ij and
(f>ig are important. In free-surface flows the resultant pressure "re-
flections" from the free surface are significant^ - indeed, it is their
effect that causes the maximum velocity in a river usually to be lo-
cated below the free surface. The main effect of the wall on faj is to
dampen the level of u\ (the normal stress perpendicular to the wall)
to about half the level found in a free shear flow. It is mainly this
reduction that makes the wall jet in stagnant surroundings spread at
only two thirds of the rate of the free jet.
Do these wall effects automatically get accounted for as one ap-
proaches the wall, effectively through the boundary conditions? A
great deal of research at present is aimed at developing models of fyj
that do automatically respond to the wall's proximity. However, the
'basic' model presented above requires rather substantial correction
close to a wall to return accurate relative levels of the stress tensor.
This is usually achieved by introducing the unit vector normal to the
wall (n/t) and applying a correction proportional to the turbulent
length scale, fc3/2/e, divided by the distance from the wall. The rec-
ommended correction of this type combines proposals of Shir (1973)
and Craft and Launder (1992) and may be written

where

^Strictly, the free surface is at a uniform pressure. What actually happens is


that agitations cause the free surface to be slightly "crinkled", i.e. the surface
topography is not quite plane and is continually changing. However, if we imagine
the free surface to be replaced by an undeformable but frictionless lid, as we
habitually do in free-surface studies, then pressure fluctuations will exist at the
lid surface.
270 B. E. Launder

In the above / = fc3/2/€ and y denotes the wall distance. Figure 2.2,

Figure 2.2: Behaviour of two near-wall flows a) Turbulence intensities


in flat plate boundary layer b) Component of turbulence normal to
wall on axis of axisyrnmetric impinging jet (from Craft et a/., 1993a)
Symbols: Experiment or DNS results; — Eq (2.18) wall reflection;
Gibson-Launder (1978) wall reflection.

from Craft and Launder (1992) shows the application of the basic
Single-Point Closures 271

model using the wall correction of eq (2.18) to compute the flat plate
boundary layer and the impinging turbulent jet. The latter flow is a
critical turbulence modelling test case: numerous models developed
by reference to flows parallel to walls give seriously incorrect pre-
diction of the radial wall jet development as the flow develops away
from the vicinity of the stagnation point (Craft et a/., 1993a). The
figure indicates that, for the impinging jet, the form recommended
for 4>™j (solid line) achieves much better agreement with experiment
than does the earlier (and more widely used) proposal of Gibson and
Launder (1978) shown by a broken line.
It would, however, seriously misrepresent the situation to leave
the impression that modelling of near-wall influences was in a satis-
factory state. Equation (2.18) and similar approaches may be ade-
quate if one is dealing with a single plane or mildly curved surface.
In most engineering applications, however, one needs to predict flows
within an enclosure or around bodies with several distinct faces. In
these cases the approach indicated by eq (2.18) is, at best, a scheme
that requires ad hoc, case-specific interpretation while in others it is
simply unworkable. In flow through a square duct, for example, there
are four vector directions normal to a wall and, correspondingly, four
wall-normal distances. Analogous problems arise in handling flow
through tube banks, within internal combustion engines or in turbo-
machinery flows. It is this geometrical complexity that has spurred
efforts to eliminate from the closure the very parameters on which
traditional 'wall-proximity' corrections depend, namely wall distance
and unit vectors.
What has come in their place? Firstly, as noted earlier, a benefit
of using the non-linear models for <$>{j\ and <^,-j2 is that there is a
much diminished wall-proximity correction. Secondly, there is the
recognition that, within the immediate wall vicinity, the turbulence
is varying so rapidly in the direction normal to the wall that some
explicit correction for inhomogeneity should be made to the pressure-
strain process, Bradshaw et al. (1987). For example in replacing the
exact expression for faji in eq (2.4a), there is the assumption that
dU'k/dx'j (that is, the mean velocity gradient evaluated at a distance r
from the point where (f>ij is to be determined) can be replaced by that
at the point itself. Launder and Tselepidakis (1991) first proposed
improving this assumption by adopting, instead, an effective mean
velocity gradient
272 B. E. Launder

In current work at UMIST the quantity I is taken as IA1/2. The


reason for introducing A 1 / 2 is pragmatic: the conventional length
scale / = k3' 2 /e exhibits a point of inflexion within the buffer layer
which causes dl/dxm to undergo excessive variations in magnitude
thus making it unsuitable. The smooth increase of A as one moves
away from the wall removes this problem. The quantity c/ has an
optimum magnitude of about 0.15.

Figure 2.3: Prediction of fully developed flow in a duct of square


cross section, Launder and Li (1994) a) Mean velocity contours b)
Wall shear stress distribution. Symbols: Experiment, Gessner and
Emery (1981) ; Cubic Model without wall correction;
Basic Model.

Figure 2.3 shows an application of a closure employing eq (2.7d)


(c2 = 0.55; c'2 = 0.6) together with an earlier form of eq (2.19)
to the problem of flow in a duct of square cross section. The an-
isotropy of the Reynolds stress field in the cross sectional plane of
the duct induces weak secondary motions which are responsible for
causing the mean velocity contours to bulge towards the corners.
The predicted flow is extremely sensitive to suppositions about wall
damping and it is encouraging that agreement is reasonably good.
If one applies the Basic Model with wall-reflection terms omitted no
Single-Point Closures 273

secondary flows are predicted. However, to include wall reflection,


by way of an equation like '(2.18), requires arbitrary decisions about
how to handle the abutting walls. In the calculations shown in Fig
2.3 it has been supposed that wall reflections associated with the four
walls can be linearly superimposed treating each alone as though it
was a plane surface of infinite extent. These assumptions cannot be
applied for ducts of arbitrary cross-section and even here, the quality
of agreement is much inferior to that using the non-linear pressure-
strain models with an inhomogeneity correction and no wall reflection
term. This latter approach is particularly successful in predicting the
distribution of shear stress around the perimeter.
The use of non-linear models, together with an inhomogeneity
correction to <pij2, is not fully satisfactory for handling flows im-
pinging orthogonally (or thereabouts) on a wall. In these cases the
mean velocity decreases virtually linearly to zero along the stagna-
tion streamline so eq (2.19) returns values negligibly different from
the local velocity gradient. At UMIST we are presently seeking al-
ternative approaches to achieving the desired damping including the
adoption of an inhomogeneity correction to (f>ij\, the turbulent part
of the pressure-strain process.

2.4 Diffusive Transport d,-j, dig

At least as much may be written about approximating diffusive


transport as about the non-dispersive action of pressure fluctuations
discussed above. Only a few brief remarks will be made, however,
partly because the processes are, in most engineering circumstances,
relatively uninfluential on the mean flow development, partly because
this article is intended to provide an introduction to turbulence mod-
elling, and partly because the quite elaborate models to be found in
the literature have by no means been extensively tested.
The most popular basis for representing diffusive transport in
second-moment closures is the generalized gradient diffusion hypoth-
esis, GGDH, which may be written:

Thus, if <p is the instantaneous Reynolds stress u,uy:


274 B. E. Launder

or, for the transport of scalar flux

One evident weakness of these forms is that while the indices i and k
on the left side of the equation can be interchanged without altering
the resultant product, such a rearrangement on the right intrinsi-
cally alters the form. No ambiguity arises because in the u^uj and
Ufd transport equations a d/dx^ operation is applied to the triple
moments; nevertheless, the difference in character between the exact
and the modelled form would appear to be a significant shortcoming.
In fact, forms very similar to (2.21) and (2.22) can be obtained by
making sweeping closure simplifications to the transport equations
for the triple moments (Hanjalic and Launder, 1972; Launder, 1976).
In that case, the model for u^UjUk consists of three terms identical to
the right side of (2.21) but with a permutation of the indices i, j and
k; likewise, that for u^u^Q also consists of three terms in which w/t, it,-,
0 successively occupy the position of u^ in (2.22). These somewhat
more elaborate and superficially more correct models do not, in prac-
tice, seem to bring better agreement when used in numerical solvers.
This could be due, at least partly, to the fact that those workers who
have adopted (2.21) and (2.22) for the triple moments have generally
not included any model for the pressure diffusion terms:

and

which appear in the u^Uj and u$ transport equations. In these pres-


sure terms the indices i and k are not interchangeable. Thus, perhaps
we should really regard (2.21) and (2.22) as models for the complete
turbulent diffusion of TZT/tZJ and w,-0 .
More rigorous and comprehensive models of triple moments have
been put forward by Lumley (1978), Andre el al. (1979), Reynolds
(1984) and Dekeyser and Launder (1984). While most approaches
have started from the exact triple-moment equations, Lumley's ap-
proach is based on the orthogonal decomposition theorem in which
approximations for the triple moments are developed from those
made for the second moments.
Single-Point Closures 275

2.5 Determining the Energy Dissipation Rate

The scalar 6 remains as an unknown in the models so far dis-


cussed; it is to be obtained by solving a transport equation. Before
discussing proposed forms for that equation; however, it may be help-
ful to address the following frequently recurring question: how, in
the absence of reliable measurements of e, can one tell whether poor
Reynolds stress and mean velocity predictions arise from errors in de-
termining the level of the energy dissipation rate or from weaknesses
in modelling the unknown processes in the ufuj equations, especially
</>{,•? While difficulties in discrimination do sometimes occur, it is
often possible to distinguish the source of the problem. Errors in e
will tend to give too high or too low energy levels; errors in fyj will
tend to give the wrong relative levels of individual Reynolds stresses.
Thus, Launder and Morse (1979) traced the failure to predict the
swirling free jet to the pressure-strain process, fajz, because the cor-
relation between streamwise and azimuthal velocity fluctuations was
of the wrong sign. For the (non-swirling) axisymmetric jet, however,
the relative stress levels are reasonably correct but the predicted level
of kinetic energy is too large - a result that points to the dissipation
rate equation as the main source of error.
Chou (1945) proposed the first equation for a variable propor-
tional to e but current concepts in modelling the energy dissipation
rate really spring from the work of Davidov (1961). Although an ex-
act transport equation for 6 may be derived from the Navier-Stokes
equations, this is by no means as useful as the corresponding equation
for UfUj. The reason for this is that the major terms in the equation
consist of fine-scale correlations describing the detailed mechanics of
the dissipation process. However - from the high-Reynolds-number
hypothesis — the fine-scale motions, so far as the Reynolds stress field
is concerned, are passive; they adjust in size as required to dissipate
energy at the rate dictated by a system of substantially larger eddies
whose structure is largely independent of viscosity. The conclusion
to be drawn from all this is that, in devising a transport equation
to mimic the spatial variation of e, one is essentially resorting to di-
mensional analysis and intuition. All proposals can be written in the
form :

where Tt denotes the net transport of c and EST stands for "extra
276 B. E. Launder

strain terms".
Originally, both coefficients cei and c £ 2 were taken as constant
with "standard" values of about 1.44 and 1.92 respectively. Lum-
ley (1975), however, argued that ct\ should be taken as zero and
c e 2 should become a function of the second invariant A^. The subse-
quent studies on buoyant diffusion undertaken by Zeman and Lumley
(1979), however, recommend the retention of both types of process.
Specifically they proposed

There are sound physical reasons for preferring to eliminate the


term containing c^ from the (. transport equation but Lumley's ex-
periences and those of Morse (1980) show that this is not possible.
The latter found that, in predicting free shear flows, taking cei as
zero and c £ 2 a function linearly proportional to AI led to poor profile
shapes. Worse, the loose coupling between the HiUj and e fields that
resulted produced strongly oscillatory rates of spread. Recent work
at UMIST has concluded that there are overall benefits to reducing
the value of cei from its "standard" value and compensating by al-
lowing c e2 to depend on the invariants A 2 and A. The form presently
recommended is

The presence of the terms marked EST is an indication that,


like the Reynolds stresses, the energy dissipation rate (and thus ef-
fective time and length scales) appear to be rather sensitive to small
secondary strains. Unfortunately while, in the u^Uj equations, the
production tensor P,-j shows clearly the origin and the approximate
sensitivity to particular strains, no equivalent help is available in the
case of the € equation. So, as remarked earlier, one proceeds by intu-
ition. Pope (1978), in a well argued proposal, suggests the inclusion
of an additional source term in the e equation proportional to

This term is zero in plane two-dimensional flows but not in an ax-


isymmetric flow. The size of the coefficient was thus tuned to give the
Single-Point Closures 277

correct rate of spread of the round jet in stagnant surroundings (us-


ing the k - e EVM). Huang (1986) has made further tests on Pope's
correction, re-optimizing the coefficient to suit the second-moment
closure he was using. When, however, he shifted from the round jet
in stagnant surroundings to data of coaxial jets, he found that these
flows were predicted better with the extra term deleted.
Hanjalic and Launder (1980) recommended the addition of a term
that is superficially similar in appearance to Pope's:

The action of the term is quite different, however. The coefficient of


proportionality is negative and, in a straight two-dimensional shear
flow, the main contributor to the term is —k(dUi/dx<2) . The term
thus acts to reduce the sensitivity of the e source to shear strain rela-
tive to normal strain. The additional term has been found to be help-
ful in boundary layers in strongly adverse pressure gradients as well
as in the round jet for it allows the term — 2(i^ — u^dUi/dxi (in P^)
to make a larger contribution (in comparison with —2uiU2 dU\/ dx-^)
to the generation rate of e. Since (dUi/dxi) is predominantly nega-
tive in these flows and u\ > u^ the result is that e becomes relatively
larger (length scales smaller), reducing the Reynolds stress levels.
The originators tested their proposals over several cases but unfor-
tunately not in curved flows. When that comparison was made the
modification that the new term brought to the e equation was sub-
stantial and its effect was to worsen agreement with experiment.
Leschziner and Rodi (1980) adopted a modification of Hanjalic
and Launder's proposal in which at every node the orientation of
the mean velocity was determined; this direction was designated as
Xi with x-2 orthogonal to it. With that choice made, they could
unambiguously discriminate between "normal" strains and "shear"
strains. The contributions of the former to P^k in eq (2.21) were
simply multiplied by a coefficient approximately three times as large
as c€i (which was retained as the coefficient of the shear strain).
With this re- interpretation, Leschziner and Rodi (1980) effectively
changed the sign of the sensitivity of the equation to streamline cur-
vature, though the effects the term produced in a second-moment
closure, were too great.
The above examples illustrate the difficulty with the f. equation.
The standard version is naively simple in form and has well known
278 B. E. Launder

failures. Yet, so far, all attempts at improving it have produced


modifications which, when subjected to wider testing, have produced
worse agreement than the original for some other flow. This applies
equally to the much publicized RNG form of the equation (Orszag,
1993).
It is now becoming popular to compute the scalar dissipation rate
€g from its own transport equation. It is a rapidly developing field
of research that lies outside the scope of this discussion. Interested
readers may refer to Craft and Launder (1989).

2.6 Simplifications to Second-Moment Closures

Although, at research level, full second-moment closures are now


being used in fairly complex recirculating flows, they place sufficient
demands on computer resource to make it desirable to devise simpler
schemes, at least for "production runs". This subsection therefore
describes briefly the main steps to simplification followed by different
workers.
A popular step in simplification is the so-called algebraic second-
moment (ASM) closure. In these schemes the transport (i.e. convec-
tion and diffusion) of the Reynolds stresses is approximated in terms
of the transport of their contraction, or rather the turbulence energy,
k. Only the transport terms contain stress gradients, so this step re-
duces the differential equations for u^Uj to a set of algebraic ones
(hence the name). The most widely used stress-transport hypothesis
is due to Rodi (1976):

where T^ means "net transport (convection minus diffusion) of </>" .


Note that in this notation the turbulence energy equation is simply

so finally we obtain

Insertion of eq (2.26) into the stress transport equation, closed


with the help of equations (2.2), (2.8) and (2.14), produces the fol-
lowing constitutive equation linking the turbulent stress and mean
strain fields:
Single-Point Closures 279

where P = 1/2P^. The equation is in fact identical to that given


by the WET model, eq (1.24), save that now the coefficient on the
right hand side is explicitly a function of P/e. Note also that the
time scale Tt emerges here as k/e.
Experience at UMIST suggests that, in wall-bounded flows where
transport effects are of secondary importance, the ASM approach
gives results very similar to that of a full-second-moment closure
for about 60% of the computational effort. The behavour is less
satisfactory in free flows; however, particularly axisymmetric flows
(Fu et al., 1987). Transport terms are then much more important
and the weakness of eq (2.26) has more serious consequences than in
wall flows. The situation becomes even worse in flows with strong
swirl. For this reason the user needs to be cautious when applying
the ASM approach to free flows.
Brief remarks will now be made about some other approaches to
avoid solving the full set of differential Reynolds stress equations.
Specifically for application to thin, simply-sheared flows, Bradshaw
et al (1967) and Hanjalic and Launder (1972,1976) have made other,
less general simplifications than the ASM route that allowed the num-
ber of stress transport equations to be reduced. Hanjalic and Laun-
der retain an equation for uiu? but express all the normal stresses
either as constant fractions of k or make use of the near constancy
1 /2
of the shear-stress correlation coefficient UiU2/(u\ ufy in a two-
dimensional boundary layer. Bradshaw et al. (1967) assume a simple
prescribable connection between the shear stress and the turbulence
energy and solve a transport equation only for the latter. For most
of the shear flows considered, a direct proportionality between the
shear stress and the turbulence energy was assumed:

the quantity c^ being generally taken as 0.09.


One can place most confidence in eq (2.28) when turbulence is in
local equilibrium. That is, when energy generation and dissipation
rates are in balance. Now, in a thin shear flow, the energy generation
rate is effectively equal to —uiu^dUi/dx^- Thus, in local equilibrium,
the kinetic energy equation reduces to
280 B. E. Launder

or, with the help of (2.28)

and finally,

This is the constitutive relation that is employed in the so-called


k — € (eddy viscosity) model (EVM). For the purpose of applying the
idea to curved and recirculating flows, eq (2.30) is generalized in an
obvious way to:

Now, eq (2.31) is a much simpler stress-strain connection than eq


(2.27) - but it is also a less general one. It does not mimic the subtle
responses that the turbulent stress field makes to small perturbations
in the strain field. Nevertheless, it often allows a predicted behaviour
in sufficiently close agreement with experiment for engineering pur-
poses.
In dealing with thin shear flows and, especially, boundary layers,
it is often the case that, instead of solving a transport equation for the
dissipation rate, e can be obtained as accurately from the formula:

where / is an algebraically prescribed function of position in the


boundary layer.
Thus, if one is working within the framework of an eddy viscosity
model, the expression for the turbulent viscosity becomes

The first such 1-equation EVM was proposed by Prandtl (1945) and
subsequently reinvented by Emmons (1954). It is not just in eddy
viscosity treatments that a prescribed length scale has been adopted.
Much of Donaldson's group's work with second-moment closures ( e . g .
Donaldson et a/., 1972) has used a prescribed profile of /, as has the
ASM study of turbulence driven secondary flows by Launder and
Single-Point Closures 281

Ying (1973) and Bradshaw!s "boundary-layer method" (Bradshaw


et al, 1967).
For historical completeness, one final stage of simplification must
be mentioned. The introduction of eq (2.30) into (2.29) and the
subsequent elimination of k and t with the help of equations (2.32)
and (2.33) leads to

or

Equation (2.34) is known as Prandtl's mixing-length hypothesis


in which the mixing length, lm, is equal to cj I . Prandtl (1925) con-
ceived the model (by arguments akin to those employed in the kinetic
theory of gases) to deal with simple shearing wherein dlli/dx^ was
the only significant strain. However, the local-equilibrium argument
presented in these notes enables a more general expression to be
obtained; for, in a multi-component strain field, eq (2.29) becomes:

or

This more elaborate form of the mixing length hypothesis has been
used with a surprising degree of success in the computation of flows
near spinning discs and cones where the velocity vector is strongly
skewed.

2.7 Non-Linear Eddy Viscosity Models

Eddy viscosity schemes do not, on the whole, cope well with


strong streamline curvature, whether this arises from flow over curved
surfaces or imparted swirling. The question arises whether, by adding
non-linear strain elements to the basic constitutive equation, the sen-
sitivity to streamline curvature, that is missing from linear EVM's,
can be captured. Models of this type have been proposed since the
282 B. E. Launder

early 1970 's but only relatively recently, with the prospect of applying
CFD to very complex flows, has an impetus developed to devise mod-
els with a width of applicability approaching that of second-moment
closure for a computational cost similar to a linear EVM. Such non-
linear EVMs have many similarities with ASM's but, for complex
flows, they require much less computational effort (typically, one
half to one quarter) due to their improved stability characteristics.
Recent contributors to models of this type include Speziale (1987),
Nisizima and Yoshizawa (1987), Rubinstein and Barton (1990), My-
ong and Kasagi (1990), Shin et al. (1993) and Taulbee et al. (1993).
Our experience at UMIST is that most of the weaknesses of the lin-
ear EVM cannot be rectified by introducing just quadratic terms to
the stress-strain relation. Only by proceeding to cubic level does one
find sufficient variety in the stress-strain couplings to achieve the
desired effects. The UMIST work thus adopts the following stress
strain relation:

where

Ci C2 C3 C4 Cs CQ C'j

-0.1 0.1 0.26 -1.0 0 -0.1 0.1


Single-Point Closures 283

Figures 2.4 and 2.5 show two recent applications of the above
model (Craft et a/., 1993b) to flows that defeat conventional lin-
ear models, namely fully developed swirling flow in a pipe and the
impinging jet. In the former the roughly parabolic variation of the
circumferential velocity with radius is correctly reproduced while any
linear eddy viscosity model gives rise to solid body rotation. Equally,
the dynamic field of the impinging jet is captured just as well with the
non-linear model as with any of the tested second-moment closures.

Figure 2.4: Non-linear Eddy Viscosity model applied to the predic-


tion of swirling flow in a pipe, Craft et al. (1993b): o Experiment,
Cheah et al. (1993) ; non-linear EVM- linear EVM.

Again it should be said that this is an area where rapid changes


are occurring and where further testing is needed to establish rigo-
rously the width of applicability of the proposed form.
284 B. E. Launder

Figure 2.5: Non-linear Eddy Viscosity model applied to the turbulent


impinging jet, Craft et al. (1993b): a) Mean velocity b) Shear
stress c) rms turbulence intensity normal to wall: o Experiment,
Cooper et al. (1993); non-linear EVM; linear EVM.

3 LOW REYNOLDS NUMBER TURBULENCE NEAR


WALLS

3.1 Introduction
The closure proposals made so far imply negligible effect of vis-
cosity on the energy containing motions and no influence of the
mean-strain field on the dissipative ones. While we may expect these
assumptions to be reasonably valid throughout most of a turbulent
flow, the no-slip condition at a rigid boundary ensures that over some
region of a turbulent wall layer, however thin, viscous effects on the
transport processes must be large. The present chapter provides a
brief account of turbulent transport in this viscosity-modified sub-
layer. A more extensive treatment, albeit less up to date, has been
provided by the writer elsewhere, Launder (1986).
In round terms, molecular influences may be expected to be in-
Single-Point Closures 285

fluential over a region extending from the surface to where the local
"eddy" Reynolds number, based on a typical eddy dimension normal
to the wall and the intensity of velocity fluctuation in that direction,
is of order 102. Turbulence models for this region may reasonably
contain elements that introduce appropriately these viscous influ-
ences. The region in question is thin (even in low-speed laboratory
studies rarely extending over more than 2mm), the processes are
highly complex and the acquisition of accurate experimental data is
greatly complicated not just by the thinness of this sublayer but by
wall-proximity influences of various kinds on the instruments them-
selves. It is thus no wonder that , despite the important contributions
being made by direct numerical simulations, our knowledge of this
important region of flow is still incomplete.
Turbulence models applied in this low-Reynolds-number region
have usually been developed from some high-Reynolds-number clo-
sure by introducing various viscosity-dependent terms and/or by
making the coefficients of existing terms functions of a turbulence
Reynolds number. These adaptations, while largely empirical, are
chosen to ensure that certain general kinematic constraints are satis-
fied. Models of this type, which we refer to as "low-Reynolds-number-
treatments", are discussed in Section 3.3
Although the extreme thinness of the viscosity-affected sublayer,
in some respects, complicates the task of model development, in an-
other it simplifies the problem; for, streamwise convective transport
within this sublayer is often sufficiently small compared with diffu-
sive or (in the case of properties of the turbulent field) source or
sink processes that it may be neglected. In cases where surface
transpiration is absent and where the influence of force fields (in-
cluding pressure gradients) on the region is negligible, the flow-field
properties, suitably normalized, are then functions of only a normal-
distance Reynolds number. For the case of the mean velocity, the
resultant distribution is known as the 'Law of the Wall'. This connec-
tion between velocity, friction velocity, \/TW/p , kinematic viscosity
and normal distance can be used in place of the no-slip boundary
condition to avoid the need to extend what may be a costly two-
or three-dimensional numerical solution to the wall itself. This ap-
proach is especially advantageous if the matching is applied outside
of the viscosity dependent region (though close enough to the wall
for streamwise transport to be negligible); for then the turbulence
model used for the numerical computation does not have to include
286 B. E. Launder

viscous effects and, moreover, one escapes the need for the especially
fine mesh that is inevitably required to resolve the viscous region
because there the curvature in the profiles of both the mean velocity
and turbulence properties is so high.
However, as CFD gains in maturity and computational power
per dollar continues to double every 18 months or so, an increas-
ingly large proportion of problems being tackled are ones where it is
not safe to treat the near-wall sublayer as though it is in its quasi-
equilibrium, 'universal' state. That is why, in the present article, a
detailed modelling of the sublayer is the only route considered.

3.2 Limiting Forms of Turbulence Correlations


in the Viscous Sublayer

We consider flow in the immediate vicinity of a wall lying in the


plane x% — 0. It follows from continuity that 8u2/dx2 = 0 at
x2 = 0 so, if we expand the fluctuating velocities in a Taylor series
about the wall, there follows:

where the a's, 6's and c's are functions of time whose mean value must
be zero, since ul = 0. The linear variation of u\ over a significant
sublayer region is well confirmed by experiment and, more recently,
by direct numerical simulation.
It follows from (3.1) that

This cubic variation of turbulent stress helps explain why the viscous
layer exhibits a distinct, if very thin, region where turbulent stress
is negligible compared with viscous stress which gives way, rather
rapidly, for small increase in X2, to a zone where molecular transport
is of only minor importance in comparison with that due to turbulent
exchange processes.
Likewise, the turbulent kinetic energy variation is given by:

Alternatively, casting the expression in dimensionless form:


Single-Point Closures 287

where

Patel, Rod! and Scheuerer (1985) concluded from the available


data that a+ lay in the range 0.025 < a+ < 0.05. The newer
direct-simulation data indicate a value of approximately 0.035.
From (3.1) we can also obtain the following expression for
e = v(du{ldxjf

or, in dimensionless form,

Thus, (. which, at the wall, equals the turbulence energy dissipa-


tion rate 6 , is non-zero there. The direct numerical simulations of
wall turbulence give values of e as the wall is approached consistent
with those indicated recently from turbulence energy measurements.
These simulations suggest unequivocally that the maximum level of
e occurs at the wall itself which is very much different from that pre-
dicted by any pre-1990 turbulence model (these typically produced
a peak level of e at x J w 12 with the wall value being less than half
the maximum value).
By comparing (3.3) and (3.4) we may readily deduce that the
approximation

is correct up to terms linear in x%. This result has been used in


different ways in modelling the limiting behaviour of e at the wall. A
popular strategy is to adopt e rather than e as the dissipative variable
(Jones and Launder, 1972) where:

a choice which evidently permits the dependent variable to be set to


zero at the wall.
e
The term v ( d m / d x ] ) ( d u ] / d x t ) , by which t differs from the true dissipation
rate, vanishes at the wall
288 B. E. Launder

It may be wondered how, in the absence of turbulence energy


generation (P, like u\u?, is proportional to cc^), a non-zero dissipa-
tion rate can be sustained at the wall. The answer is that viscous
action diffuses energy towards the wall (vd2k/dx^} at a rate which
just balances the dissipation rate there.
The corresponding component terms, €{j, defined as

may be treated in the same way as e. It is of interest to compare the


ratio (e;j/e) with u^Uj/k. In isotropic turbulence the terms are both
equal to 2/3 Sij. In the limiting case of wall turbulence, however,
the relative magnitudes of the stress and dissipation rate ratios dif-
fer from component to component. The question of modelling this
limiting behaviour is taken up in the following section.

3.3 Low Reynolds Number Modelling

Preamble

The present notes are written at a time when second-moment


closure strategy for low-Reynolds-number regions is in great ferment.
Models and ideas developed in the 70's and 80's are being cast aside
and new activity, aimed at replacing these outmoded approaches by
something more modern, is underway at several institutions. So far
as the writer's group is concerned, the change of emphasis has arisen
from finally accepting two important facts (which, in truth, have been
plain to see for several years) allied to an immensely important store
of new information about near-wall turbulence. The two essential
facts are:

• The variation of Reynolds stresses through the sub-layer region


and the associated damping of the effective transport coeffi-
cients seem to be affected rather little by viscosity.

• As the wall is approached, the stress field approaches a "two-


dimensional" state since the fluctuations normal to the wall
die out faster than those parallel to the surface. It is this
feature that is mainly responsible for the damping of near-wall
transport coefficients. It follows, therefore, that turbulence
Single-Point Closures 289

modelling in the sublayer should be consistent with the two-


dimensional limit especially as regards the pressure-strain and
dissipation processes.

Substantial support for these assertions_is offered by Fig 3.1 which


examines the experimental variation of u\jk across the near-wall
region. (In the figure u2 is denoted by v and x 2 by y). According to
the high-Reynolds-number free flow model of Gibson and Launder
(1978), in local equilibrium the connection between shear stress and
mean velocity gradient in a simple shear is simply

Figure 3.1: Variation oiu\/k across sublayer (Launder, 1986).

(This is also the form given by the generalized gradient diffusion


hypothesis GGDH). Now, low-Reynolds-number k — e models present
their constitutive stress-strain relation in the following form:

where C M normally takes the constant value 0.09 and /^ is the "vis-
cous" damping function employed in models of that type. On com-
paring these two equations it is evident that they would be identical if
290 B. E. Launder

u\lk were equal to 0.342 /M. Figure 3.1, from Launder (1986), shows
the variation of the quantity across the sublayer region as deduced by
Patel, Rodi and Scheuerer, from admittedly rather imprecise experi-
mental data (the uncertainty band attaching to /^ is probably larger
than that shown for u\/k}. The figure does show quite convincingly,
however, that the two parameters vary in essentially the same way
across the sublayer. The more recent direct simulation data lead
to essentially the same distribution of f ^ . Thus, provided, as the
wall is approached, the proper diminution of u\ can be modelled,
there would be little need for other "viscous" damping. The pro-
cesses that are most influential in determining the level of u\ are the
pressure-strain and dissipative correlations. There are good reasons
for supposing that the former is dependent on viscosity only within
the near wall sublayer while u^/k increases with distance from the
wall over a more extensive region. It implies, therefore, that the
requisite damping of u\/k should, to a large extent, be provided by
non-viscous parameters.
The "store of new information about near-wall turbulence" noted
above refers to the results of full simulations of turbulent shear flows
that have become available over the last five years. The first of these,
the flow between plane and slightly curved parallel surfaces by Kim
et al. (1987) and Moser and Moin (1984), present detailed budgets of
each Reynolds-stress component right up to the wall itself, including
the contributions made by the pressure-strain correlation. Thus, for
the first time, the modeller can make direct appeal to data of the
processes to be approximated — data generated not by experiment
but by computer simulation.

Component dissipation

The process in which it appears most important to include vis-


cous effects is the dissipative correlation. As remarked earlier, at high
Reynolds numbers, it is usually assumed that the fine scale motion
is isotropic and thus:

As the Reynolds number approaches zero the energy-containing and


dissipating range of motions are no longer distinct and the dissipa-
tion rate is then commonly approximated (for example Rotta, 1951,
Haiijalic and Launder, 1976) as:
Single-Point Closures 291

If one accepts eq (3.6) and (3.7) as asymptotic forms valid in the


limit of very high and very low Reynolds number, it is natural to
generalize the result as:

The function fa is usually assumed to be a function of the turbu-


lent Reynolds number (Rt = k2/z/e), its value going from unity to
zero as Rt varies from zero to infinity. The recommended form for
fs to have emerged from the computations of Hanjalic and Launder
(1976) was:

so that, by the start of the fully turbulent region (Rt « 150), the
dissipation is very nearly isotropic.
Equation (3.7), while attractively simple, is not an exact limiting
form. It may readily be deduced (Launder and Reynolds, 1983) from
the polynomial expansions for the velocity components that, while
it is correct if neither i nor j takes the value 2, if one of the indices
refers to the direction normal to the wall:

and, for the u\ component,

An invariant way of expressing these results is through the unit vector


normal to the wall Hk (Launder and Reynolds, 1983, Kebede et al.,
1985):

It is readily deduced that, if eq (3.12) is to satisfy equations (3.10)


and (3.11), the coefficients a and /? must be unityA A model for
the dissipative correlation that satisfies both this wall-adjacent limit
and the high Reynolds number, locally isotropic form can clearly be
y
For the case i = 1 , j = 1 arid i = 3 , j = 3, eq(3.12) gives
292 B. E. Launder

obtained by replacing euiUj/k in eq (3.8) by the right hand side of


eq (3.12). This has been the route followed in several closure studies
from the mid-eighties to the present. However, eq (3.12) cannot be
said to be a convenient equation for general application as it brings
in the unit vector normal to the wall (c.f. the discussion in Sec.
2 on the desirability of removing unit vectors and wall distance).
Recently the writer has discovered that the desired limiting values
may be obtained in a form that is free of unit vectors, namely:

The form of eq (3.13) was suggested by eq (3.5) which expresses the


limiting dissipation rate at the wall in terms of gradients of -\fk. The
expression for t*j only departs significantly from eu^Uj/k within the
viscous sublayer where ^-gradients are steep, Fig 3.2a.
The quantity c*- can only represent e^j in the immediate vicin-
ity of a wall. To obtain an expression for e^ of greater width of
applicability a composite expression of type (3.8) would need to be
adopted

The question arises as to what parameter fs should depend on. As


noted above, the turbulent Reynolds number is the usual choice but
Fig 3.3, which shows the DNS-computed variation of different quan-
tities (including Rt~) across the near-wall region of turbulent channel
flow, brings out a problem: for this case Rt reaches its maximum
value at x% ~ 20 whereas the dissipation ratios shift towards their
isotropic levels well beyond that. It is for this reason that present
studies at UMIST continue the practice of Launder and Tselepidakis
(1991) in employing one of the stress-invariant parameters as an ar-

where 'a' takes the value 1 or 3. In the immediate vicinity of the wall u\ is
negligible compared with u\ or v% giving the desired limiting behaviour.
Single-Point Closures 293

Figure 3.2: Variation of fij/e in near-wall turbulent flow a) e*,-/e


versusus normalized wall distance b) resultant profile of e,-j/e using
eq (3.13-3.15).

gument of fs. The form currently adopted is (J R Cho, personal


communication):
The quantity e*- can only represent c,-j in the immediate vicinity
of a

which produces, with eq (3.14), the component dissipation profiles


shown in Fig 3.2b.

Near-wall effects on pressure-strain process

As discussed in Section 2.3, to achieve a widely applicable for-


mulation it is important to adopt a model of <foj that keeps wall-
reflection effects to a minimum (and, if possible, eliminates them
entirely). This is because there seems no way at present to model
their impact in flows with highly irregular boundary surfaces. For
this reason we here discuss only schemes employing cubic versions
of faff because this formulation seems naturally to give rise to the
decrease of uiu^/k with increasing strain rate, such as occurs as one
294 B. E. Launder

Figure 3.3: Distribution of Rt, A<2, A$ and A across viscous sublayer


of low Reynolds number channel flow of Kim et al. (1987)(from
Launder and Tselipidakis, 1991): Rt] — A 2 ; A3;
A.

enters the viscosity-dependent sublayer. At present such models have


been applied only in fairly simple flows so it remains to be seen how
well they extrapolate to recirculating and impinging flow regions.
Simpler linear schemes have been extensively reviewed by Hanjalic
(1994). _
The principal task is to make u\/k fall to zero in the correct
manner as the wall is approached. Making the coefficient Cj of ^i
(viz eq (2.15)) vanish with A would appear to ensure that ^221 fell to
zero in the 2-component limit that applies at the wall, thus causing
u\jk to fall to zero. Figure 3.4, from early work of Launder and
Tselepidakis (1990) shows this is not the case, however. While eq
(2.15) is consistent with the 2-component limit it does not enforceit.
Thus u\/k remains virtually constant across the near-wall sublayer,
A does not vanish at the wall and neither does cj. Replacing € in the
equation for 4>iji by f. at least ensures that faji vanishes at the wall;
however, it still fails to reduce u\/k except in the immediate wall
vicinity. Apparently, the only way to ensure a proper decrease of
this stress ratio is to introduce damping via the turbulent Reynolds
number Rt as was done for the lowest curve in Fig 3.4. Of course,
agreement with the DNS results is still far from complete with that
Single-Point Closures 295

early model. The more recent results to emerge from the UMIST
group (to be shown below) have been obtained with the following
choice of the coefficient of <^,-ji:

Figure 3.4: Variation of u\/k across low-Reynolds-number channel


flow: A Kim at al (1987) - DNS data; Lines: Second-moment pre-
dictions of Launder and Tselipidakis (1990) without wall reflection:
: ci a function of A; : as before but e replaced by e; :
ci damped by turbulent Reynolds number.

The high-Reynolds-number, free-flow form of <^,-j2 is adapted in


two ways as one approaches the wall. The term with coefficient c2
gives rise to the following sink term in the shear-stress equation:

Both AW and (an - 022) increase rapidly as the wall is approached


leading to the annihilation of Uiu2. To limit the strength of this sink,
the coefficient c2 is amended to c2 = 7mn(c2co, A), where C2oo is the
value c2 would take in a free shear flow (that value depends slightly
on whether c2 takes a non-zero value; see the discussion in Section
2.3). The other amendment, discussed already in connection with eq
296 B. E. Launder

(2.19), is the need to correct eq (2.6) for the rapid spatial variation of
the mean velocity gradient. In a channel flow the correction, given by
eq (2.19) or similar forms, is of substantial importance in the buffer
region but becomes negligible further from the wall. This behaviour
is consistent with the conclusions reached by Bradshaw et al. (1987)
from a processing of the DNS data.

Diffusion

Serious study of stress diffusion in the near-wall sublayer is only


just beginning. Most computations have retained the GGDH repre-
sentation with the same value for the coefficient cs (0.22) as in fully
turbulent regions. It isn't that the formula is regarded as satisfac-
tory, but rather that, in the near-wall fully turbulent region, trans-
port is often of relatively little importance while, in the immediate
vicinity of the surface, where diffusive transport balances the loss
by dissipation, molecular transport outweighs diffusion by velocity
fluctuations, UiUjUk- A minor improvement to the standard GGDH
scheme would be to introduce e in place of e so that the turbulent
time scale remains finite as the wall is approached.
In fact, because non-zero pressure fluctuations occur on the wall,
pressure diffusion dies out less rapidly than the velocity diffusion as
the wall is approached. This may be seen by considering the stress
budget in the viscous sublayer. For example, with u\u<i = a^bix^
(c.f. eq (3.2)) the net viscous diffusion of UiH? is Qa^b^x^ while its
dissipation rate is ^aib^vx-^. Since the form of faji ensures a varia-
tion of this process to a power of x^ greater than unity, the imbalance
between viscous diffusion and dissipation can only be compensated
by a non-zero pressure diffusion amounting to — laib^vx^. Likewise,
in the u\ equation it is easily deduced that a pressure diffusion of
magnitude -\b\vx\ is required to offset the excess viscous diffusion
(\1b\vx\) over dissipation. By contrast, in the u\ and u\ budgets,
where no pressure-diffusion terms appear, the leading viscous diffu-
sion and dissipation terms do balance.
Tselepidakis (1992) devised a model of pressure-diffusion among
components thereby satisfying the budget in all components. The
formula adopted, however, employed the wall normal vector and for
this reason it is not reproduced here. A less restrictive form is
Single-Point Closures 297

This formula embodies the' same underlying idea but does not employ
the unit vector. However, while the imbalance in the u\ equation is
entirely eliminated with this form, that in the ~u\u^ equation is only
reduced by one half. Nevertheless, since Tselepidakis (1992) found
his formulation to have only a secondary effect on the stress profiles,
the above replacement — which removes about 75% of the (small)
problem - is probably satisfactory.

Low Reynolds number effects on e


In Section 2.5 no detailed consideration was given to the exact 6
equation because, at high Reynolds numbers, the scales of the con-
trolling eddies and the dissipative motions are quite distinct. That
distinction does not apply, however, in the viscous layer: there the
exact dissipation equation shows the mean strain to act as a source
of e principally through the term:

Having regard for the fact that, in a thin shear flow, the mean
velocity is significant only in directions parallel to the wall (so index
'i' denotes 1 or 3), the second term is evidently small compared with
the first since, within the sublayer, rates of change of the instanta-
neous turbulence field will be much larger in direction x% than in x\
or £3. The first term may, in fact, be expressed in terms of eq (3.12)
or (3.13). If direction x\ is aligned with the near-wall mean velocity
both lead to the result:

Equation (3.17) suggests therefore that at x^ = 0 the value of c£i


should be 2.0 rather than the value 1.44 commonly used at high Rey-
nolds numbers 9 . The changeover from the high-Reynolds-number
limit to equation (3.17) should consistently adopt the function fs.
Earlier work by the writer's group (Hanjalic and Launder, 1976) has
adopted a value for cf\ that was entirely independent of Reynolds
9
Or the value 1.0 adopted in recent UMIST work when cC2 is a function of AI
and A.
298 B. E. Launder

number. A viscosity dependent influence of mean strain was intro-


duced, however, via the source term

which may be regarded as the outcome of using the generalized


gradient-diffusion hypothesis to close a term containing d2Ui/dxjdxk
in the exact e equation. In this case the transportable fluctuating
quantity </> is d u { / d x j . The value 1.0 adopted for c'fl by Hanjalic
and Launder (1976) was, however, 3 - 4 times as large as is usu-
ally adopted; so, arguably, that choice was compensating for the fact
that c£i had been held fixed. In more recent studies c'el has usually
been set to considerably smaller values. For example, Launder and
Tselepidakis (1993) take c£l = 0.43.
We now shift attention to the decay term in the dissipation
rate equation which, at high-Reynolds-numbers, has been taken as
c&^/k- For high-Reynolds-numbers the coefficient c& was deter-
mined by reference to grid turbulence decay and it is natural, there-
fore, that one should turn to low-Reynolds-number grid turbulence
to decide if modifications are needed. The most complete turbulence
decay data are those of Batchelor and Townsend (1948); these sug-
gest that below a value of Rt of about 10 the decay exponent n in the
relation k oc t~n increases from the high-Reynolds-number limit of
about 1.2 to an asymptotic value of 2.5. If we assume the changeover
to be describable in terms of the local value of Rt, c e2 must be mul-
tiplied by some function of Reynolds-number, / £ , to give the desired
limiting values of the decay exponents. Hanjalic and Launder (1976)
took

There is no reason to suppose much similarity between the low-


Reynolds-number sublayer near a wall and the viscous decay of weak
grid turbulence. However, the exponential term in eq (3.18) falls
to zero so rapidly that /E is essentially unity over the whole of the
region near the wall where turbulence is important. There is, how-
ever, a more important modification to this sink term needed in wall
turbulence since, as noted in Section 3.1, e is non-zero at the wall
while k varies as x\. Even with the introduction of /£, therefore,
c £ 2/ e e 2 /fc tends to infinity as x 2 goes to zero. Accordingly, Hanjalic
and Launder (1976) replaced e2 in the sink term by ec where
Single-Point Closures 299

As noted in Section 3.2, e varies as x\ near the wall and thus the
term c^f^l/k tends to a constant value as the wall is approached.
Virtually all proposals for extending the e equation to the wall adopt
this form.
Finally, there remains the matter of viscous effects on the diffu-
sion of e to be accounted for. _The writer's group has retained the
high-Reynolds-number form, c^fc/e, as the appropriate diffusion co-
efficient in the low-Reynolds-number region while Prud'homme and
Elghobashi (1983) multiply the coefficient ce by a viscous damping
function, /^. All proposals have included the exact viscous trans-
port vdt/dxj as an addition to the high-Reynolds-number diffusion
model. There is also the question of pressure diffusion to be consid-
ered. In high-Reynolds-number turbulence no explicit accounting of
this process was attempted. An order-of-magnitude estimate of this
term in the exact e equation suggests, however, that as the wall is
approached the process becomes significant (Launder, 1986). This
suggests that a separate approximation ought to be incorporated. A
possible model for the additional term that is significant only in the
viscous region has been proposed in that article:

Tselepidakis and Launder (1993) have suggested a value of c£4 of


0.92.

3.4 Applications

Low-Reynolds-number models designed to match the wall-limiting


behaviour have so far been applied only in simple shear flows. This is
a little ironic because models of this type are really designed to han-
dle complex mean strain patterns and far-from-equilibrium states.
Nevertheless, one has to start with simple flows to calibrate any free
coefficients in the model. Figures 3.5 and 3.6 show applications from
two recent Ph.D studies at UMIST.
The first relates to plane channel flow and the figure compares the
computed variation of rms velocity fluctuations across the near-wall
300 B. E. Launder

Figure 3.5: Distribution of turbulent intensities and shear stress


across near-wall region of low-Reynolds-number channel flow; Sym-
bols: DNS results, Kim at al (1987) —: closure computations, Laun-
der and Li (1994) a) turbulence intensities normalized by friction
velocity b) shear stress.

sublayer with the DNS results of Kim et al. (1987). The adopted
model of <pij2 was the complete cubic form (c2 = 0.55; c'2 = 0.6 in eq
(2.7(d)) with an inhomogeneity correction (eq (2.19)) but no "wall-
reflection" term. The agreement of the closure computations with
the DNS results is reasonably satisfactory.
The second application, from Launder and Tselepidakis (1994),
also relates to channel flow but here the channel rotates about the
#3 axis, i.e. in orthogonal mode rotation. The resultant asymme-
try in the turbulence intensity profiles arises from the direct effect
of the Coriolis forces on the turbulent stresses in the x\ — x^ plane,
augmenting fluctuations normal to the wall on the pressure (left)
side of the channel which, in turn, increases the shear stress u\u-i
and thus of streamwise fluctuations too. The model in this case was
similar to that adopted in the previous example except that, because
the computations were carried out rather earlier, the simplified cubic
model of </>,-j2 was adopted (c2 = 0.6 ; c'2 = 0) and, in consequence a
small 'wall-reflection' contribution was added. Generally the closure
Single-Point Closures 301

Figure 3.6: Application to turbulent flow in plane channel rotating


about x3 axis (orthogonal mode rotation) a) Mean velocity profile;
b) turbulent shear stress profile 8 ; c) rms turbulent intensities Sym-
bols: DNS results, Kristoffersen and Andersson (1990) : Launder
and Tselipidakis (1994); : Launder and Shima (1989) (from
Kristoffersen and Andersson, 1990).

computations (continuous line) are quite successful in mimicking the


asymmetries found in the DNS data generated by Kristoffersen and
h
Note: UTo is friction velocity for non-rotating duct at same bulk Reynolds
number
302 B. E. Launder

Andersson (1990). For comparison, predictions obtained by Professor


Andersson's group using the older 2nd-moment closure of Launder
and Shima (1989) are also included; this model only partially com-
plied with asymptotic wall-limiting behaviour. While it too captures
broadly the asymmetry of the flow, it is also clearly not as successful
as the cubic form.

ACKNOWLEDGEMENTS

The recent developments described are based largely on the re-


search discoveries of former and present students and research assis-
tants in the UMIST CFD group. Particular contributions have been
made by Drs. J R Cho, T J Craft, S Fu, N Z Ince, S-P Li, Mr. K
Suga and Dr. D P Tselepidakis. The text of this article has been
prapared in TATgX format by Erdinc, Ozkan.

4 REFERENCES
Andre, J.C., de Moor, G., Lacarrere, P., Therry, G. and du Vachat,
R., 1979. "The clipping approximation and inhomogeneous
turbulence simulations," Turbulent Shear Flows - 1 (ed. F.J.
Durst et a/.), 307, Springer-Verlag, Heidelberg.
Andre, J.C., de Moor, G., Lacarrere, P. and du Vachat, R., 1976.
"Turbulence approximation for inhomogeneous flows: Part I:
The clipping approximation," J. Atmos. Sciences, 33, p.476.
Bradshaw, P., Ferriss, D.H. and Atwell, N.P., 1967. "Calculation of
boundary-layer development using the turbulent energy equa-
tion," J. Fluid Mech. 28, p.593.
Bradshaw, P., Mansour, N.N. and Piomelli, U., 1987. "On local
approximations of the pressure strain-term in turbulence mod-
elling," Proc. Summer Program, Center for Turbulence Re-
search, Stanford University, p.159.
Cheah, S.C., Cheng, L., Cooper, D. and Launder, B.E. 1993. "On
the structure of flow in spirally fluted tubes," 5th IAHR Con-
ference on Refined Flow Modelling and Turbulence Measure-
ments, pp.293-300, Presse Ponts et Chaussees, Paris.
Single-Point Closures 303

Chou, P.Y., 1945. "On velocity correlations and the solution of the
equations of turbulent fluctuation," Quart. J. Appl. Math. 3,
pp.38-54.

Cooper, D., Jackson, D.C., Launder, B.E. and Liao, G.X., 1993.
"Impinging jet studies for turbulence model assessment: Part
I: Flow-field experiments," Int. J. Heat Mass Transfer 36,
p.2675.

Cousteix, J. and Aupoix, B., 1981. "Modelisation des equations aux


tensions de Reynolds dans un repere en rotation," Unpublished
report, Dept. d'Aerothermique, ONERA/CERT.

Craft, T.J. and Launder, B.E., 1989. "A new model for the pres-
sure/scalar gradient correlation and its application to homoge-
neous and inhomogeneous flows," Proc. 7th Symposium Tur-
bulent Shear Flows, Stanford University, Paper 17-1.

Craft, T.J. and Launder, B.E., 1992. "New wall-reflection model


applied to the turbulent impinging jet," AIAA J. 30, p.2970.

Craft, T.J., Graham, L.W. and Launder, B.E., 1993a. "Impinging


jet studies for turbulence model assessment. Part 2: An exam-
ination of the performance of four turbulence models," Int. J.
Heat Mass Transfer 36, p. 2685.

Craft, T.J., Launder, B.E. and Suga, K., 1993b. "Extending the
applicability of eddy viscosity models through the use of defor-
mation invariants and non-linear elements," 5th IAHR Confer-
ence on Refined Flow Modelling and Turbulence Measurements,
Presse Fonts et Chaussees, Paris, pp. 125-132.

Cresswell, R.W., Haroutunian, V., Ince, N.Z., Launder, B.E, and


Szczepura, R.T., 1989. "Measurement and modelling of buoyancy-
modified elliptic shear flows," Proc. 7th Symp. Turbulent Shear
Flows, Stanford University, Paper 12-4.

Crow, S.C., 1968. "Visco-elastic properties of fine-grained incom-


pressible turbulence," J. Fluid Mech. 41, p.81.

Daly, B.J. and Harlow, F.H., 1970. "Transport equations of turbu-


lence," Phys. Fluids 13, p.2634.
304 B. E. Launder

Davidov, B.I., 1961. "On the statistical dynamics of an incompress-


ible turbulent fluid," Soviet Physics Doklady 6, p.10.

Deardorff, J.W., 1970. "A numerical study of 3-dimensional channel


flow at large Reynolds numbers," J. Fluid Mech. 41, p.453.

Dekeyser, I. and Launder, B.E., 1984. "A comparison of triple-


moment temperature-velocity correlations in the asymmetric
heated jet with alternative closure models," Turbulent Shear
Flows - 4 (ed. L.J.S. Bradbury et a/.), Springer, Berlin, p.102.

Donaldson, C. du P., Sullivan, R.D. and Rosenbaum, H., 1972. "A


theoretical study of the generation of atmospheric clear-air tur-
bulence," AIAA J. 10, p. 162.

Elghobashi, S.E. and Launder, B.E., 1983. "Turbulent time scales


and the dissipation rate of temperature variance in the thermal
mixing layer," Phys. Fluids 26, p.2415.

Emmons, H.W., 1954. "Shear Flow Turbulence," Proc. 2nd US


Cong. Appl. Mech., ASME.

Favre, A., 1965. "Equations de gaz turbulents compressibles," J.


Mecanique 4, pp.361-421.

Fu, S., 1988. "Computational modelling of turbulent swirling flows


with second-moment closure models," PhD Thesis, Faculty of
Technology, University of Manchester.

Fu, S., Huang, P.G., Launder, B.E. and Leschziner, M.A., 1986. "A
comparison of algebraic and differential second-moment clo-
sures for axisymmetric turbulent shear flows with and without
swirl," Mech. Eng. Dept., Rep. TFD/86/9, UMIST.

Fu, S., Launder, B.E. and Leschziner, M.A., 1987a. "Modelling


strongly swirling recirculating jet flow with Reynolds-stress
transport closures," Proc. 6th Symposium on Turbulent Shear
Flows, Toulouse.

Fu, S., Launder, B.E. and Tselepidakis, D.P., 1987b. "Accommo-


dating the effects of high strain rates in modelling the pressure-
strain correlation," Mech. Eng. Dept., Rep. TFD/87/5, UMIST.
Single-Point Closures 305

Gessner, F. and Emery, 1981. "The numerical prediction of devel-


oping turbulent flow in rectangular ducts," ASME J. Fluids
Eng. 103, p.445.

Gibson, M.M. and. Launder, B.E., 1978. "Ground effects on pressure


fluctuations in the atmospheric boundary layer," J. Fluid Mech.
86, p.491.

Hah, C., 1981. "Three-dimensional structures and turbulence clo-


sure of the wake developing in a wall shear layer," AIAA Paper
No. 81-1269, AIAA 14th Fluid and Plasma Dynamics Conf.,
Palo Alto, California.

Hanjalic, K., 1994. "Advanced turbulence closure models," Int. J.


Heat Fluid Flow, 15, pp.178-203.

Hanjalic, K. and Launder, B.E., 1972. "A Reynolds-stress model


of turbulence and its application to thin shear flows," J. Fluid
Mech. 51, p.301.

Hanjalic, K. and Launder, B.E., 1976. "Contribution towards a


Reynolds stress closure for low Reynolds number turbulence,"
J. Fluid Mech. 74, p.593.

Hanjalic, K. and Launder, B.E., 1980. "Sensitizing the dissipation


equation to secondary strains," ASME J. Fluids Eng. 102,
pp.3440.

Huang, P.G., 1986. "The computation of elliptic turbulent flows


with second-moment closure models," PhD Thesis, Faculty of
Technology, University of Manchester.

Jones, W.P. and Launder, B.E., 1972. "The prediction of laminar-


ization with a two- equation model of turbulence," Int. J. Heat
Mass Transfer 15, p.301.

Jones, W.P. and Musonge, P., 1983. "Modelling of scalar transport


in homogeneous turbulent flows," Proc. 4th Turbulent Shear
Flows Symp., Karlsruhe, 17.18.

Kebede, W., Launder, B.E. and Younis, B.A., 1985. "Large am-
plitude periodic flow: A second-moment closure study," Proc.
5th Turbulent Shear Flows Symp., Cornell University, Ithaca.
306 B. E. Launder

Kim, J., Moin, P. and Moser, R.D., 1987. "Turbulence statistics


in fully-developed channel flow at low Reynolds number," J.
Fluid Mech. 177, p.133.

Koo Sin Lin, M., Launder, B.E. and Sharma, B.I., 1974. "Prediction
of momentum, heat and mass transfer in swirling turbulent
boundary layers," ASME J. Heat Transfer 94C, p.204.

Kristoffersen, R. and Andersson, H.I., 1990. Proc. 3rd European


Turbulence Conference, Stockholm.

Launder, B.E., 1975. "On the effects of a gravitational field on the


transport of heat and momentum," J. Fluid Mech, 67, p.569.

Launder, B.E., 1976. "Heat and Mass Transfer", Chap. 6 Topics in


Applied Physics: 12 - Turbulence (ed. P. Bradshaw), Springer,
Berlin.

Launder, B.E., 1986. "Low-Reynolds-Number Turbulence Near


Walls," Mech. Eng. Dept. Rep. TFD/86/4, UMIST.

Launder, B.E., 1989. "Second-moment closure: present... and fu-


ture?" Int. J. Heat Fluid Flow 10, p.282.

Launder, B.E. and Li, S-P, 1993. "On the prediction of flow over
riblets via 2nd-moment closure," Proc. Conf. Near-Wall Tur-
bulence (ed. R.M.C. So et a/.) Elsevier, New York.

Launder, B.E. and Li, S-P., 1994. "On the elimination of wall
topography parameters from second-moment closure," Phys.
Fluids 6 (2), p.999.

Launder, B.E., Morse, A.P., Rodi. W. and Spalding, D.B., 1973.


"Prediction of free shear flows - a comparison of the perfor-
mance of six turbulence models," Proc. Langley Free Shear
Flows Conf., NASA, SP320.

Launder, B.E. and Morse, A.P., 1979. "Numerical prediction of


axisymmetric free shear flows with a Reynolds stress closure,"
Turbulent Shear Flows - 1, Springer-Verlag, Heidelberg, p.279.

Launder, B.E. and Reynolds, W.C., 1983. "Asymptotic near-wall


stress dissipation rates in a turbulent flow," Phys. Fluids 26.
Single-Point Closures 307

Launder, B.E. and Shima, N., 1989. "A second-moment closure for
the near-wall sublayer: development and application," AIAA
J. 27, p.1319.
Launder, B.E., Tselepidakis, D.P. and Younis, B.A., 1987. "A
second-moment closure study of rotating channel flow," J. Fluid
Mech. 183, p.63.

Launder, B.E. and Tselepidakis, D.P., 1990. "Contribution to the


second-moment modelling of sublayer turbulent transport" Near-
Wall Turbulence (ed S. J. Kline and N. H. Afgan), Hemisphere,
pp.818-833.

Launder, B.E. and Tselepidakis, D.P., 1991. "Directions in second-


moment modelling of near-wall turbulence," AIAA Paper 91-
0219.

Launder, B.E. and Tselepidakis, D.P., 1993. "Contribution to the


modelling of near-wall turbulence," Turbulent Shear Flows-8
(ed. F. Durst et a/.) Springer-Verlag, Heidelberg.

Launder, B.E, and Tselepidakis, D.P., 1994. "Application of a new


second-moment closure to turbulent channel flow rotating in
orthogonal mode," Int. J. Heat Fluid Flow 15, pp.210.

Launder, B.E. and Ying, W.M., 1973. "Prediction of flow and heat
transfer in ducts of square cross section," Proc. Inst. Mech.
Engrs. 187, p.455.

Leschziner, M.A. and Rodi, W., 1981. "Calculation of annular


and twin parallel jets using various discretization schemes and
turbulence-model variations," J. Fluid Eng'g. 103, p.352.

Leslie, D.C., 1973. Developments in the Theory of Turbulence, Ox-


ford.

Li, S-P, 1993. "Predicting riblet performance with engineering tur-


bulence models," PhD Thesis, Faculty of Technology, Univer-
sity of Manchester.

Lin, A. and Wolfshtein, M., 1979. "Theoretical study of the Rey-


nolds stress equation," Turbulent Shear Flows - 1, Springer-Verlag,
Heidelberg, p.327.
308 B. E. Launder

Lumley, J.L., 1975. "Prediction methods for turbulent flow," - Lec-


ture Series 76, von Karman Inst., Rhode-St.-Genese, Belgium.

Lumley, J.L., 1978. "Computational modelling of turbulent flows,"


Advances in Applied Mechanics 18, p.123.

Mathieu, J. and Jeandel, D., 1984. "La simulation des modeles de


turbulence et leurs applications," I, Eyrolles, Paris.

Monin, A.S., 1965. "On the symmetry of turbulence in the surface


layer of air," Isv. Atmos. Ocean. Phys. 1, p.45.

Morse, A.P., 1980. "Axisymmetric free shear flows with and with-
out swirl," PhD Thesis, Faculty of Engineering, University of
London.

Moser, R.D. and Moin, P., 1984. "Direct numerical simulations of


curved channel flow," NASA Rep. TM-85974.

Myong, H.K. and Kasagi, N., 1990. "Prediction of anisotropy of the


near-wall turbulence with an anisotropic low Reynolds number
k-e turbulence model," J. Fluids Engrg. 112, p.521.
Naot, D., Shavit, A. and Wolfshtein, M., 1970. Israel J. Technol.
8, p.259.
Naot, D., Shavit, A. and Wolfshtein, M., 1973. "Two-point correla-
tion model and the redistribution of Reynolds stresses," Phys.
Fluids 16, p. 738.

Newman, G., Launder, B.E. and Lumley, J.L., 1981. "Modelling the
behavior of homogeneous, scalar turbulence," J. Fluid Mech.
lll,p.217.
Nisizima, S. and Yoshizawa, A., 1987. "Turbulent channel and Cou-
ette flows using an anisotropic k-e model," AIAA J. 25, p.414.
Orszag, S., 1993. "Renormalization group modelling and turbu-
lence simulations," Near-Wall Turbulent Flows (Ed R.M.C. So
et a/.), Elsevier, Amsterdam.
Patel, V.C., Rodi, W. and Scheuercr, G., 1985. "Turbulence models
for near-wall and low Reynolds number flows: a review," AIAA
J. 23, p.1320.
Single-Point Closures 309

Pope, S.B., 1978. "An explanation of the round-jet/plane-jet anomaly,"


AIA J. 16, p.279.

Prandtl, L., 1925. "Bericht iiber Untersuchungen zur ausgebildeten


Turbulenz," ZAMM 5, p.136.

Prandtl, L., 1945. "Uber ein neues Formelsystem fur die ausge-
bildete Turbulenz," Nachrichten von der Akad der Wissenschaft,
Gottingen.

Prud'homme, M. and Elghobashi, S., 1983. "Prediction of wall-


bounded flows with an improved version of a Reynolds stress
model," Proc. J^th Turbulent Shear Flows Symp., Karlsruhe,
1.7.

Reynolds, 0., 1895. "On the dynamical theory of incompressible


viscous fluids and the determination of the criterion," Trans.
Proc. Roy. Soc. A. 186, p.123.

Reynolds, W.C., 1984. "Physical and analytical foundations, con-


cepts and new directions in turbulence modelling and simula-
tion," in Turbulence Models and their Applications II by B.E.
Launder, W.C. Reynolds and W. Rodi, Eyrolles, Paris.

Rodi, W., 1976. "A new algebraic relation for calculating the Rey-
nolds stresses," ZAMM 56, p.219.

Rodi, W. and Scheuerer, G., 1983. "Calculation of curved shear


layers with two equation turbulence models," Phys. Fluids 26,
pp.1422-1436.
Roshko, A., 1972. Open forum discussion, Proc. Langley, Free
Shear Flows Con/., NASA SP320, pp.629-635.

Rotta, J.C., 1951. "Statistiche Theorie nichthomogener Turbu-


lenz," Z. Phys. 129, p.547.
Rubinstein, R. and Barton, J.M., 1990. "Renormalization group
analysis of the stress transport equation," Phys Fluids A 2,
p.1472.
Samaraweera, D.S.A., 1978. "Turbulent heat transport in 2- and
3-dimensional temperature fields," PhD Thesis, Faculty of En-
gineering, University of London.
310 B. E. Launder

Schumann, U., 1977. "Realizability of Reynolds stress turbulence


models," Phys. Fluids 20, p.721.
Shih, T.H. and Lumley, J.L., 1985. "Modeling of pressure correla-
tion terms in Reynolds stress and scalar flux equations," Rep.
FDA-85-3, Sibley School of Mech. and Aerospace Eng., Cornell
University.

Speziale, C.G., 1987. "On non-linear k-L and k-e models of turbu-
lence," J. Fluid Mech. 178, p.459.
Taulbee, D.B., Sonnenmeier, J.R. and Wall, K.M., 1993. "Appli-
cation of a new non-linear stress-strain model to axisymmetric
swirling flow," Proc. 2nd Int. Symp. on Engrg. Turbulence
Modelling and Measurement (ed. W. Rodi and F. Martelli),
Elsevier, Amsterdam.
Tselepidakis, D.P., 1992. "Development and application of a new
second-moment closure for turbulent near- wall flows," PhD
Thesis, Faculty of Technology, University of Manchester.
Wyngaard, J.C. and Sundararajan, A., 1979. "The temperature
skewness budget in the lower atmosphere and its implications
for turbulence modelling," Turbulent Shear Flows- 1, Springer-
Verlag, Heidelberg, p.319.
Younis, B.A., 1984. "On modelling the effects of streamline curva-
ture on turbulent shear flows," PhD Thesis, Faculty of Engi-
neering, University of London.
Zeman, O. and Lumley, J.L., 1979. "Buoyancy effects on entrain-
ing turbulent boundary layers: a second order closure study,"
Turbulent Shear Flows - 1 (ed. F.J. Durst et a/.), Springer-
Verlag, Heidelberg, pp.295-306.
Index

algebraic dissipation rate model, Reynolds stress equation,


202 223
algebraic second-moment mod- Reynolds-average equations,
els, see algebraic stress 221
model Reynolds-averages, see Favre-
algebraic stress model, 197 averages
applications, 204-210 two-equation models, 226
anisotropic eddy viscosity model, Coriolis forces, 264
191,208
autocorrelation coefficient, 6 deformation work, 23
diffusive transport, 248, 273
backscatter term, 119 direct numerical simulation (DNS),
backstep flow, 210 79,115
buoyancy effects, 250 compressible, 100
homogeneous, 102
closure problem, 189 incompressible
coherent structure capturing, 147 homogeneous, 85
coherent structures, 29, 111, see inhomogeneous, 86
very large eddy simu- resolution requirements, 81
lation solution methods
large eddy simulation, 146— fast vortex, 97
148 remeshing, 96
modeling, 148 spectral element, 98
compressed (expanded) isotro- vector expansion, 88
pic turbulence, 229 viscous vortex, 91
compressible mixing layer dissipation
convective Mach number, dilatational, 26, 224
103, 233 dissipation anisotropy, 194
compressible turbulence dissipation rate, 249, 257, 275
equations of motion, 220 component, 290-293
homogeneous shear, 229 low Reynolds number, 297-
isotropic, 227 299

311
312 Index

dissipation wavemimber, 156 large eddy simulation (LES),


109
eddy damped quasi-normal Marko- a priori testing, 123
vian (EDQNM), 119, cross term, 118
135 cutoff filter, 127
eddy shocklets, 28 extra strain effects, 135-137
eddy viscosity model, 280 filtering, 116
eigenvectors, 25 numerical methods, 141-143
elliptic relaxation model, 216 subgrid-scale models, 118-
ensemble average, 188 119
equilibrium flows differential, 137-138
dissipation, 10 dynamic, 127-132
energy cascade, 6 scale similarity, 125-127
equilibrium, turbulence in, 194, shifted model, 140
267 Smagorinsky, 119-123
ergodic hypothesis, 188 spectral, 132-135
extra strain terms, 275 law of the wall, 285
Favre-aver ages, 221, 244 Leonard term, 118
force fields, 264 low Reynolds number turbulence,
friction factor, 81 284
modeling, 288
generalized gradient diffusion hy-
mass flux, 223
pothesis, 273
mass-weighted average, see Favre-
heat flux, 223, 250 averages
Helmholtz decomposition, 102 material frame-indifference, 218
homogeneous turbulence, 187 mean scalar transport, 245
mean-field closure, 247
integral length scale, 6 mean-field scalar transport, 247
integrity bases, 196
isotropic turbulence, 9, 103,121, Navier-Stokes equations, 187
169, 200,247,288 near-wall asymptotic behavior,
isotropization of production model, see viscous sublayer
263 non-equilibrium flows
energy cascade
Kolmogorov, 9 delay in crossing, 14
hypothesis of local isotropy, fading memory, 20
195 negative production, 19
scales, 9, 80, 201 spectral cascade, 13
universal spectrum, 156 nonlinear eddy viscosity model,
Kurarnoto-Sivashinsky, 68 281
Index 313

nonlocal effects, 210 statistical perturbation the-


ory, 159-161
one-equation models, 191 thermal noise, 166
outscatter term, see Leonard two-equation model, 169
term Reynolds stress anisotropy, 194
Reynolds-averaged equation, 243
power-law divergences, 162
Reynolds-averaged Navier Stokes
Prandtl mixing length, 190
(RANS), 112
pressure Poisson equation, 258
Reynolds-averages, 187
pressure reflections, 269
rotating channel flow, 205, 300
pressure-dilatation, 225
rotating homogeneous shear, 205
pressure-strain correlation, 210,
rotational strains, 204
248
near-wall, 293-296
second-moment closure, see Rey-
turbulence-turbulence part,
nolds stress
265
second-order closure, see Rey-
principle of receding influence,
nolds stress
255
single-point closure, 246
proper orthogonal decomposi-
solenoidal dissipation rate equa-
tion (POD), 33
tion, 225
eigenfunctions, 52-66
spectral scaling, 5
models using, 37-42
square duct, flow in, 208
wall region, 42-52
statistically steady turbulence,
quasi-isotropic model, 259 187
strain rate tensor, 195
realizability, 218 stress coupling, 249
regularization, 199 subgrid-scale Reynolds stress,
relaxation effects, 210 118
renormalization group (RNG), supersonic mixing layer, 232
155
e—expansion, 163 third-order diffusion, 212
applications turbulent eddies, 7
flapping-hydrofoil, 177- turbulent kinetic energy, 11
181 turbulent Mach number, 224
time-dependent turboma- turbulent pipe flow, 81
chinery, 174-177 turbulent scalar flux, 246
turbomachinery heat trans- turbulent transport equations
fer, 171-174 dissipation rate, 193
resummation of divergent kinetic energy, 192
series, 162-168 Reynolds stress, 192
314 Index

near-wall models, 213-220 wall models


large eddy simulation, 138-
van Driest damping, 122 141
very large-eddy simulation (VLES), wall shear velocity, 81
147,175 wavelets, 67
viscous sublayer, 286 WET concept, 253
Von Karman constant, 190
vorticity rate tensor, 195
zero-equation models, 190
wall effects, 269 Baldwin-Lomax model, 190
diffusion, 296 Smagorinsky model, 190

Вам также может понравиться