LES For Acoustics by Wei Shyy and Michael J. Rycroft

P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
This page intentionally left blank
ii
November 22, 2006
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
LARGE-EDDY SIMULATION FOR ACOUSTICS

Noise pollution around airports, trains, and industries increasingly attracts environmental concern and regulation. Designers and researchers have intensified the use of large-eddy simulation
(LES) for noise-reduced industrial design and acoustical research. This book, written by thirty
experts, presents the theoretical background of acoustics and LES followed by details about numerical methods such as discretization schemes, boundary conditions, and coupling aspects. Industrially relevant, hybrid Reynolds-averaged NavierStokes/LES techniques for acoustic source
predictions are discussed in detail. Many applications are featured ranging from simple geometries for mixing layers and jet flows to complex wing and car geometries. Selected applications
include recent scientific investigations at industrial and university research institutions. Presently
perfect solution methodologies that address all relevant applications do not exist; however, the
book presents a state-of-the-art collection of methods, tools, and evaluation methodologies. The
advantages and weaknesses of both the commercial and research methodologies are carefully
presented.
Claus Wagner received his Ph.D. in Fluid Dynamics in 1995 at the Technical University of
Munich, Germany. He is Honorary Professor for Industrial Aerodynamics, Ilmenau University
of Technology, Germany. Since 1998, he has been a scientist in and head of the Section of Numerical Simulations for Technical Flows of the Institute for Aerodynamics and Flow Technology,
German Aerospace Center, Gottingen, Germany. Dr. Wagners research includes experimental
investigations of the resonant control of nonlinear dynamic systems, theoretical and numerical
investigations of thermal convection in cylindrical containers, and direct numerical simulation
and large-eddy simulations of turbulent flows in different configurations. He has held visiting
positions in Gainsville, Florida, USA as well as in Bremen, Germany.
Thomas Huttl received his Ph.D. in Fluid Dynamics in 1999 at the Technical University of
Munich, Germany. His academic research included work on the direct numerical simulation of
turbulent flows in curved and coiled pipes, and direct and large-eddy simulations of boundary
layer flows with and without separation, in the framework of a FrenchGerman DFG-CNRSCooperation project. He has held visiting positions in Nantes, France and Bangalore, India.
Between 2000 and 2003, he was a senior engineer for aeroacoustics at MTU Aero Engines GmbH,
Germanys leading manufacturer of engine modules and components and of complete aero
engines. Dr. Huttl led MTUs contribution for the European research project TurboNoiseCFD and
contributed to the European research project SILENCER. After some years working as IT quality
manager and internal auditor, he is now Chief Privacy Officer for the entire MTU Aero Engines
concern. Among his many honors, Dr. Huttl was elected a member of the Senate of the DGLR,
Deutsche Gesellschaft fur Luft- und Raumfahrt - Lilienthal - Oberth e.V. in 2003 and 2006.
Pierre Sagaut received his DEA in Mechanics in 1991 and his Ph.D. in Fluid Mechanics in
1995 at Universite Pierre et Marie Curie Paris 6 (Paris, France). He worked as a research
engineer at ONERA (French National Aerospace Research Center) from 1995 to 2002. He has
been a professor in mechanics at University Pierre et Marie Curie Paris 6 since 2002. He is
also part-time Professor at Ecole

Polytechnique (France) and scientific consultant at ONERA,
IFP, and CERFACS (France). His main research interests are fluid mechanics, aeroacoustics,
numerical simulation of turbulent flows (both direct and large-eddy simulation), and numerical
methods. He is also involved in uncertainty modeling for computational fluid dynamics (CFD).
He has authored and coauthored more than sixty papers in peer-reviewed international journals
and 130 proceedings papers. He is the author of several books dealing with turbulence modeling
and simulation. He is member of several editorial boards: Theoretical and Computational Fluid
Dynamics, Journal of Scientific Computing, and Progress in CFD. He received the ONERA
award three times for the best publication and the John Green Prize (delivered by ICAS, 2002).
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
ii
November 22, 2006
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
Cambridge Aerospace Series

Editors
Wei Shyy
and
Michael J. Rycroft
1.
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
J. M. Rolfe and K. J. Staples (eds.): Flight Simulation

P. Berlin: The Geostationary Applications Satellite
M. J. T. Smith: Aircraft Noise
N. X. Vinh: Flight Mechanics of High-Performance Aircraft
W. A. Mair and D. L. Birdsall: Aircraft Performance
M. J. Abzug and E. E. Larrabee: Airplane Stability and Control
M. J. Sidi: Spacecraft Dynamics and Control
J. D. Anderson: A History of Aerodynamics
A. M. Cruise, J. A. Bowles, C. V. Goodall, and T. J. Patrick: Principles of Space
Instrument Design
G. A. Khoury and J. D. Gillett (eds.): Airship Technology
J. Fielding: Introduction to Aircraft Design
J. G. Leishman: Principles of Helicopter Aerodynamics, 2nd Edition
J. Katz and A. Plotkin: Low Speed Aerodynamics, 2nd Edition
M. J. Abzug and E. E. Larrabee: Airplane Stability and Control: A History of the
Technologies that Made Aviation Possible, 2nd Edition
D. H. Hodges and G. A. Pierce: Introduction to Structural Dynamics and
Aeroelasticity
W. Fehse: Automatic Rendezvous and Docking of Spacecraft
R. D. Flack: Fundamentals of Jet Propulsion with Applications
E. A. Baskharone: Principles of Turbomachinery in Air-Breathing Engines
Doyle D. Knight: Elements of Numerical Methods for High-Speed Flows
C. Wagner, T. Huttl, and P. Sagaut: Large-Eddy Simulation for Acoustics
iii
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
iv
November 22, 2006
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
Large-Eddy Simulation for

Acoustics
Edited by
CLAUS WAGNER
German Aerospace Center
THOMAS H UTTL
MTU Aero Engines GmbH
PIERRE SAGAUT
Universite Pierre et Marie Curie
November 22, 2006
4:57
CAMBRIDGE UNIVERSITY PRESS
Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo

Cambridge University Press
The Edinburgh Building, Cambridge CB2 8RU, UK
Published in the United States of America by Cambridge University Press, New York
www.cambridge.org
Information on this title: www.cambridge.org/9780521871440
Cambridge University Press 2007
This publication is in copyright. Subject to statutory exception and to the provision of
relevant collective licensing agreements, no reproduction of any part may take place
without the written permission of Cambridge University Press.
First published in print format 2006
eBook (EBL)
ISBN-13 978-0-511-29466-2
ISBN-10 0-511-29466-2
eBook (EBL)
hardback
ISBN-13 978-0-521-87144-0
hardback
ISBN-10 0-521-87144-1
Cambridge University Press has no responsibility for the persistence or accuracy of urls
for external or third-party internet websites referred to in this publication, and does not
guarantee that any content on such websites is, or will remain, accurate or appropriate.
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
4:57
Contents
List of Figures and Tables
page xiii
Contributors
xxi
Preface
xxv
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 The importance of acoustic research
Thomas Huttl
1.1.1
1.1.2
1.1.3
1.1.4
1.1.5
1.1.6
1.1.7
Health effects
Activity effects
Annoyance
Technical noise sources
Political and social reactions to noise
Reactions of industry
Research on acoustics by LES
1.2 Introduction to computational aeroacoustics

Manuel Keler
1.2.1
1.2.2
1.2.3
1.2.4
1.2.5
Definition
History
Aeroacoustics
Conceptual approaches
Remaining problem areas for sound computation
1.3 State of the art: LES for acoustics

Claus Wagner, Oliver Fleig, and Thomas Huttl
1.3.1 Broadband noise prediction in general

1.3.2 Broadband noise prediction based on LES
2
1
1
2
2
3
4
5
5
7
7
7
8
9
13
15
15
15
17
Theoretical Background: Aeroacoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

Avraham Hirschberg and Sjoerd Rienstra
2.1 Introduction to aeroacoustics
2.2 Fluid dynamics
2.2.1 Mass, momentum, and energy equations
2.2.2 Constitutive equations
24
25
25
28
vii
P1: JZZ
CUFX063-FM
viii
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
CONTENTS
2.2.3
Approximations and alternative forms of the basic

equations
2.3 Free-space acoustics of a quiescent fluid
2.3.1 Orders of magnitude
2.3.2 Wave equation and sources of sound
2.3.3 Greens function and integral formulation
2.3.4 Inverse problem and uniqueness of source
2.3.5 Elementary solutions of the wave equation
2.3.6 Acoustic energy and impedance
2.3.7 Free-space Greens function
2.3.8 Multipole expansion
2.3.9 Doppler effect
2.3.10 Uniform mean flow, plane waves, and edge diffraction
2.4 Aeroacoustic analogies
2.4.1 Lighthills analogy
2.4.2 Curles formulation
2.4.3 Ffowcs WilliamsHawkings formulation
2.4.4 Choice of aeroacoustic variable
2.4.5 Vortex sound
2.5 Confined flows
2.5.1 Wave propagation in a duct
2.5.2 Low-frequency Greens function in an infinitely long uniform duct
2.5.3 Low-frequency Greens function in a duct with a discontinuity
2.5.4 Aeroacoustics of an open pipe termination
3
30
33
33
35
36
38
39
44
47
47
49
52
55
55
59
60
62
65
68
68
72
74
76
Theoretical Background: Large-Eddy Simulation . . . . . . . . . . . . . . . . . . . . 89

Pierre Sagaut
3.1 Introduction to large-eddy simulation
3.1.1 General issues
3.1.2 Large-eddy simulation: Underlying assumptions
3.2 Mathematical models and governing equations
3.2.1 The NavierStokes equations
3.2.2 The filtering procedure
3.2.3 Governing equations for LES
3.2.4 Extension for compressible flows
3.2.5 Filtering on real-life computational grids
3.3 Basic numerical issues in large-eddy simulation
3.3.1 Grid resolution requirements
3.3.2 Numerical error: Analysis and consequences
3.3.3 Time advancement
3.4 Subgrid-scale modeling for the incompressible case
3.4.1 The closure problem
3.4.2 Functional modeling
3.4.3 Structural modeling
3.4.4 Linear combination models, full deconvolution, and Lerays
regularization
3.4.5 Extended deconvolution approach for arbitrary nonlinear terms
3.4.6 Multilevel closures
3.4.7 The dynamic procedure
89
89
90
91
91
92
95
98
100
105
105
109
112
112
112
113
118
119
120
121
121
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
ix
CONTENTS
3.5 Extension of subgrid models for the compressible case

3.5.1 Background
3.5.2 Extension of functional models
3.5.3 Extension of structural models
3.5.4 The MILES concept for compressible flows
4
125
125
125
126
126
Use of Hybrid RANSLES for Acoustic Source Predictions . . . . . . . . . . . . 128

Paul Batten, Philippe Spalart, and Marc Terracol
4.1 Introduction to hybrid RANSLES methods
4.2 Global hybrid approaches
4.2.1 The approach of Speziale
4.2.2 Detached-eddy simulation
4.2.3 LNS
4.2.4 The approach of Menter, Kunz, and Bender
4.2.5 Defining the filter width
4.2.6 Modeling the noise from unresolved scales
4.2.7 Synthetic reconstruction of turbulence
4.2.8 The NLAS approach of Batten, Goldberg,
and Chakravarthy
4.3 Zonal hybrid approaches
4.3.1 The approach of Quem
er
e and Sagaut
4.3.2 The approach of Labourasse and Sagaut
4.3.3 Zonal-interface boundary coupling
4.4 Examples using hybrid RANSLES formulations
4.4.1 Flow in the wake of a car wing mirror
4.4.2 Unsteady flow in the slat cove of a high-lift airfoil
4.5 Summary of hybrid RANSLES methods
128
130
130
131
133
138
140
142
143
145
148
150
150
152
154
154
158
163
Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

5.1 Spatial and temporal discretization schemes
Tim Broeckhoven, Jan Ramboer, Sergey Smirnov, and Chris Lacor
5.1.1
5.1.2
5.1.3
5.1.4
Introduction to discretization schemes

Dispersion and dissipation errors
Spatial discretization schemes
Temporal discretization schemes
5.2 Boundary conditions for LES

Michael Breuer
5.2.1
5.2.2
5.2.3
5.2.4
5.2.5
4:57
Outflow boundary conditions

Inflow boundary conditions
Boundary conditions for solid walls
Far-field boundary conditions for compressible flows
Final remark for discretization schemes
5.3 Boundary conditions: Acoustics

Fang Q. Hu
5.3.1 Characteristic nonreflecting boundary condition
5.3.2 Radiation boundary condition
5.3.3 Absorbing-zone techniques
167
167
168
170
197
201
203
204
208
214
215
216
217
218
218
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
CONTENTS
5.3.4 Perfectly matched layers

5.3.5 Summary of boundary conditions for acoustics
220
222
5.4 Some concepts of LESCAA coupling

Wolfgang Schroder
and Roland Ewert
222
5.4.1 LES inflow boundary

5.4.2 Silent embedded boundaries
225
232
Applications and Results of Large-Eddy Simulations for Acoustics . . . . . 238

6.1 Plane and axisymmetric mixing layers
Christophe Bogey and Christophe Bailly
6.1.1 Plane mixing layer
6.1.2 Axisymmetric mixing layers jets
6.1.3 Concluding remarks for mixing layer simulations
6.2 Far-field jet acoustics
Daniel J. Bodony and Sanjiva K. Lele
6.2.1
6.2.2
6.2.3
6.2.4
6.2.5
6.2.6
Introduction to jet acoustics

Numerics of jet simulations
Results for jet simulations
Future directions of jet acoustics
Conclusions for far-field jet acoustics
Acknowledgments
6.3 Cavity noise

Xavier Gloerfelt, Christophe Bogey, and Christophe Bailly
6.3.1
6.3.2
6.3.3
6.3.4
Introduction to cavity noise

Overview of cavity-flow simulations
Recent achievements using LES
Concluding remarks for cavity noise
6.4 Aeroelastic noise

Sandrine Vergne, Jean-Marc Auger, Fred Peri
e,
Andre Jacques,
and Dimitri Nicolopoulos
6.4.1
6.4.2
6.4.3
6.4.4
6.4.5
6.4.6
6.4.7
6.4.8
6.4.9
Introduction to aeroelastic noise

Fluidstructure interaction
Numerical simulation
Application
Simulation model
Numerical results
Mesh influence
Conclusions for aeroelastic noise prediction
Acknowledgment
6.5 Trailing-edge noise

Roland Ewert and Eric Manoha
6.5.1 Introduction to trailing-edge noise simulation using LES
6.5.2 Trailing-edge noise simulation using LES and APE
6.5.3 Trailing-edge noise simulation using LES, Euler equations,
and the Kirchhoff integral
6.5.4 Unsteady pressure-field analysis
238
239
241
244
245
245
247
252
259
262
262
262
262
263
266
272
272
272
274
276
278
279
283
290
293
293
293
293
296
315
320
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
xi
CONTENTS
6.6 Blunt bodies (cylinder, cars)

Franco Magagnato
6.6.1 Overview of blunt-body simulations
6.6.2 Circular cylinder
6.6.3 Car
6.7 Internal flows
Philippe Lafon, Fabien Crouzet, and Jean Paul Devos
6.7.1 Introduction to internal flows
6.7.2 Computation of acoustic fluctuations due to
turbulence-generated noise at low Mach number
6.7.3 Computation of flow acoustic coupling in low-Mach-number
ducted flows
6.7.4 Computation of aeroacoustic instabilities in high-Machnumber ducted flow
6.7.5 Conclusions for internal flow prediction
6.8 Industrial aeroacoustics analyses
Fred Mendonca
6.8.1
6.8.2
6.8.3
6.8.4
6.8.5
6.8.6
7
4:57
Introduction to industrial aeroacoustics analyses

Preliminary considerations
A two-step CFD modeling process (steady-state and transient)
Postprocessing through acoustic coupling
Conclusions for industrial aeroacoustics analyses
Acknowledgments
333
333
335
345
349
349
349
351
354
355
356
356
357
358
370
376
376
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 378
Claus Wagner, Pierre Sagaut, and Thomas Huttl
7.1 Governing equations and acoustic analogies
378
7.2 Numerical errors
384
7.3 Initial and boundary conditions
385
7.4 Examples
386
Appendix A. Nomenclature
389
Appendix B. Abbreviations
391
References
395
Index
429
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
xii
November 22, 2006
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
List of Figures and Tables
Figures
1.1
1.2
1.3
1.4
1.5
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.9
3.1
3.2
3.3
3.4
4.1
4.2
Magnetic levitation hover train project Transrapid

Noise prediction methods
Insulator of a high-speed trains pantograph
Sketch of the insulator of a high-speed trains pantograph
Far-field sound-pressure spectra
Monopole, dipole, and quadrupole generating waves on the surface
of the water around a boat
Sketch of scattered plane wave with mean flow
A potential flow through the vocal folds is silent
Straight duct of arbitrary cross section
(a) Method of images applied to a source at y = (y1 , y2 , y3 ) at a distance
y3 from a hard wall x3 = 0 has a Greens function: G(x ,t|y , ) =
G0 (x,t|y, ) + G0 (x,t|y , ) with y = (y1 , y2 , y3 ). (b) A source between
two parallel hard walls generates an infinite row of images. (c) A source
in a rectangular duct generates an array of sources
The end correction for no flow (M j = 0) and a little flow (M j = 0.01)
Plane-wave reflection coefficient |R| and end correction at jet
exhaust without coflow for M j = 0.01, 0.1, . . . , 0.6
Plane-wave reflection coefficient |R| and end correction at jet
exhaust with M j = 0.3 and coflow velocities Mo/M j = 0, 0.25,
0.5, 0.75, 1
Acoustic flow at a pipe outlet (a) for an unflanged pipe termination
and (b) for a horn
Schematic kinetic energy spectra of resolved and subgrid scales
with spectral overlap (Gaussian filter)
Streaks in the inner layer of the boundary layer
Schematic of kinetic energy transfer in isotropic turbulence
Schematic of the two-level filtering procedure and the Germano identity
Turbulence energy spectrum partitioned into resolvable
and unresolvable frequencies
Required near-wall mesh resolutions with DNS, traditional LES, global
hybrid RANSLES, and nonlinear acoustics solvers (NLAS) based on
disturbance equations
4
12
21
22
22
44
54
66
68
73
84
85
86
87
95
107
114
122
135
147
xiii
4:57
P1: JZZ
CUFX063-FM
xiv
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
LIST OF FIGURES AND TABLES
4.3
4.4
4.5
4.6
4.7
4.8
4.9
4.10
4.11
4.12
4.13
4.14
4.15
4.16
4.17
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
Initial startup transient at probe 111 (cylinder rear face) predicted by

unsteady RANS and hybrid RANSLES (LNS model) using identical
base model, mesh, and time step
Instantaneous streamwise vorticity contours (with streamwise velocity
shading) predicted by hybrid RANSLES model
Resolved and unresolved (synthetically generated) signals for probe
111 (cylinder rear face)
Sound-pressure levels determined by resolved, unresolved, and
composite signals for probe 111 (cylinder rear face)
Instantaneous streamwise vorticity contours (with streamwise velocity
shading) predicted by the nonlinear acoustics solver (NLAS)
Sound-pressure levels determined by NLAS method at probe 111
(cylinder rear face)
Location of the LES subdomain (displayed mesh shows every eighth
grid line)
Mean flow streamlines. Left: RANS computation, Right: Hybrid
RANSLES computation
Instantaneous Schlieren-like view
Isovalue contours of the dilatation field
Acoustic pressure spectrum at location S2
Acoustic pressure spectrum in the recirculation bubble (location S3)
Isovalue contours of the dilatation field
Acoustic pressure spectrum at the slats trailing edge (location S4)
Acoustic pressure spectrum in the slats wake
Resolution of different explicit and compact (implicit) schemes
Comparison of the resolution of Taylor and Fourier difference
schemes
Effect of f on a second- and an eighth-order filter
Perturbation pressure of an acoustic wave initiated by a Gaussian
pulse: comparison of solution with smooth and randomly perturbed
mesh
Example for a reasonable choice of the integration domain (inflow
and outflow boundaries) for the flow past a wing in a wind tunnel
Von Karm
an
vortex street past an inclined wing (NACA4415) at
Re = 20,000 and = 12 visualized by streaklines; four different time
instants of a shedding cycle in the vicinity of the outflow boundary
are shown
Sketch of Lund et al.s (1998) procedure for generating appropriate
inflow conditions for a boundary layer flow
Example for the generation of inflow data for a 90 bend using a
second simulation for a straight duct flow with periodic b.c.
Law of the wall u+ (y+ ) in a turbulent boundary layer without or with
only a weak pressure gradient (half-logarithmic plot)
Sketch of the computational domains to determine, for example,
trailing-edge noise with the hybrid approach
Sketch of the rescaling concept
Sketch of the flat-plate boundary layer domain (left) and the
trailing-edge domain (right). The procedure to provide the inlet
distribution to simulate trailing-edge flow is visualized
155
155
156
157
157
158
159
159
160
160
161
162
162
163
164
178
180
188
197
202
205
206
207
211
224
226
230
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
xv
5.13
5.14
5.15
5.16
6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
6.10
6.11
6.12
6.13
6.14
6.15
6.16
6.17
6.18
6.19
6.20
6.21
LES of a turbulent boundary layer for Re0 = 1400 and M = 0.4.

Skin-friction coefficient c f versus Re for different rescaling
formulations
Transfer function | F | as a function of wave number scaled by
damping zone thickness d
Spurious sound waves and velocity field generated at an artificial
boundary at x = 0
Pressure distribution on y = 35 in Figure 5.15 for several thickness
values d and BiotSavarts law (denoted as compensation)
Simulation of a 2D mixing layer. (a) Snapshot of the dilatation field
= .u on the whole calculation domain, levels in s1 . (b) View
of the pairing zone with the vorticity field in the mixing layer and the
dilatation field outside
LES of a ReD = 6.5 104 subsonic jet. Time evolution of the
fluctuating pressure p in Pa as a function of t = tu j /D, at x = 16r 0 ,
y = 8r 0 , and z = 0
LES of a ReD = 6.5 104 subsonic jet. Snapshot of the vorticity norm
in the flow field and of the fluctuating pressure outside in the plane
z = 0 at t = 7.5
LES of a ReD = 6.5 104 subsonic jet. Snapshots of the vorticity norm
in the plane z = 0 at times: (a) t = 2.2, (b) t = 3.5
LES of a ReD = 4 105 subsonic jet. Snapshot of the vorticity norm in
the flow field and of the fluctuating pressure outside in the plane z = 0
LES of a ReD = 4 105 subsonic jet. (a) Pressure spectra at

(x = 11r 0 , r = 15r 0 ). (b) Profiles of vrms
/u j in the shear layer for r = r 0 .
Different simulations: LESac (- -), LESampl (. . . . . . ), LESshear (- - -),
LESmode ( )
Schematic of a turbulent jet issuing into a still fluid
OASPL directivity at a distance of 30D j from the unheated, Mach 0.9
jet exit
Centerline distribution of streamwise root-mean-square fluctuations
Centerline distribution of density root-mean-square fluctuations
normalized by the difference ( j )
Far-field OASPL taken at a distance of 30D j from the nozzle exit
Integral Lagrangian time scale of streamwise fluctuations near the end
of the potential core
One-thirdoctave spectral comparisons of an unheated, Mach 0.5 jet
of Bodony and Lele (2004) with the 195 m/s data of Lush (1971)
Azimuthal correlation of the far-field sound field of LES data from
Bogey, Bailly, and Juve (2003) ( ) and Bodony and Lele (2004) ()
Far-field OASPL taken at a distance of 30D j from the nozzle exit
Far-field acoustic spectra taken at a distance of 30D j from the
nozzle exit
Transition toward a wake mode for large L/ ratio (2D DNS of the
flow over an L/D = 4 cavity, at M = 0.5, and ReD = 4800)
LES of a deep cavity.
Influence of the boundary layer turbulence on cavity noise
Mode switching
The plate model
4:57
232
235
236
237
240
242
242
243
243
244
247
249
254
255
256
257
258
258
260
260
265
268
270
271
275
P1: JZZ
CUFX063-FM
xvi
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
6.22
6.23
6.24
6.25
6.26
6.27
6.28
6.29
6.30
6.31
6.32
6.33
6.34
6.35
6.36
6.37
6.38
6.39
6.40
6.41
6.42
6.43
6.44
6.45
6.46
6.47
6.48
6.49
6.50
6.51
Sketch of the experimental setup

Wall-pressure transducer positions on the elastic steel plate
Vibroacoustics measurements: (+) accelerometers and ()
microphones positions
General mesh dimensions
General sizes of the mesh
Detailed mesh around the obstacle
Contours of the streamwise mean velocity U1 /U0 near the ruler in the
median plane
Streamwise velocity fluctuations urms /U0 near the ruler in the
median plane
Velocity profiles at the origin of the reference: (- -) measurement;
(+ + +) simulation
Pressure coefficient
Wall-pressure fluctuation rms levels calculated at various positions
along the median plane
Comparison of calculated ( ) and measured (+ + +) wall-pressure
fluctuation rms levels on the plate
PSD of wall-pressure fluctuations measured on the flat plate
Coherence 2 of wall-pressure fluctuations measured on the flat plate
for streamwise points using the same notation as in Figure 6.34
Coherence 2 of wall-pressure fluctuations measured on the flat plate
for spanwise points using the same notation as in Figure 6.34
Phase velocity Up of wall-pressure fluctuations measured on the flat
plate for streamwise points using the same notation as in Figure 6.34
Acceleration PSD aa channel 26
Acoustic pressure PSD pp channel 30 (Pref = 2 105 Pa) using the
same notation as in Figure 6.38
First mesh: detailed mesh around the obstacle
Comparison of the two simulations wall-pressure fluctuations DSP
on the plate
Comparison of the two simulations acoustic pressure in cavity
using the same notation as in Figure 6.41
Numerical simulation of airfoil aerodynamic noise: possible hybrid
strategies
Sketch of the computational domains to determine trailing-edge noise
with the hybrid approach
Coordinate system and nomenclature used to determine corrections
for a 2D acoustic simulation
LES subdomain at the trailing edge (left) and horizontal weighting
function (right)
Damping function |F (d )| over vortical wavenumber in x1 -direction
= 2/v scaled with the damping zone on- and offset width
d = (l 2 l 1 )/2
LES grid with partially resolved plate and every second grid point
shown
Acoustic grid scaled with the plate length l
Enlargement of the leading- (left) and trailing-edge region (right)
Visualization of the instantaneous flow field
278
279
280
280
281
281
283
283
284
285
285
286
287
288
289
290
291
291
292
292
293
295
297
301
304
305
305
306
307
307
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
xvii
6.52
6.53
6.54
6.55
6.56
6.57
6.58
6.59
6.60
6.61
6.62
6.63
6.64
6.65
6.66
6.67
6.68
6.69
6.70
6.71
6.72
6.73
6.74
6.75
6.76
APE source terms L = [

u] = (L x , L y )T , L x (left), L y (right), and
CAA grid
Pressure contours of the trailing-edge problem M = 0.15,
Re = 7 105 , at time level T = 3.0 with APE solution
equations (6.32, 6.33)
Pressure contours of the trailing-edge problem M = 0.15,
Re = 7 105 at time level T = 3.0
Directivity 1/2 (, r, f ) with
= nondimensional power spectral

2
2 2

, i.e., 0 (, r, f )df = p 2 (, r )/
density (PSD) of p (, r )/ c

2 2
c
Comparison of the trailing-edge noise directivities 1/2 (, r, f ) for
r = 1.5 applying Equations (6.32, 6.33) ( =

PSD of p 2 ) and Mohrings
2
(1999) acoustic analogy ( =
PSD of B ), Equations (6.43. 6.44)
Sound-pressure level (SPL) versus frequency for a receiving point in
r = 1.5 above the trailing edge for various directions (see Figure 6.45)
Generation of a periodical source term via window weighting
CAA grid
A harmonic test source over the trailing edge (left) and directivities
obtained for M = 0, 0.088 with LEE and APE (right) applied
APE source term (
u) y (left) and sound radiation from the trailing
edge (right)
Computational grid
Contours of instantaneous Mach-number isovalues
Flow streamlines at the trailing edge: instantaneous (above) and
time-averaged (below)
Wall-pressure fluctuations on pressure and suction sides near the TE
Evolution of the wall-pressure PSD along the chord
Evolution of the wall-pressure PSD along the chord
Wave-numberfrequency spectrum of wall-pressure fluctuations
on the airfoil suction side at x/C = 0.9
Wave-numberfrequency spectrum of wall-pressure fluctuations
on the airfoil suction side at x/C = 0.5
Instantaneous isovalues of pressure fluctuations obtained from
LES data
Evolution of the wall-pressure spectra along the vertical grid
line x = C (starting from the TE upper corner) with respect to the
vertical distance z
Spanwise evolution of the coherence of the surface-pressure field on
the suction side near the TE (x/C = 0.9958). The frequency
bandwidths are integrated
Spanwise evolution of the coherence of the pressure field at distance
z0 = 37.9 mm from the TE at x/C = 1
Final problem-adapted acoustic grid
Final problem-adapted acoustic grid (closer view)
Isovalue contours of instantaneous pressure fluctuation field
(range 2 Pa, black and white) computed from (i) LES inside the
injection interface and (ii) E3P (from LES data injection) outside the
injection interface
308
309
310
310
311
311
313
313
314
315
319
320
321
322
323
324
324
325
326
326
327
328
328
330
331
4:57
P1: JZZ
CUFX063-FM
xviii
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
6.77
6.78
6.79
6.80
6.81
6.82
6.83
6.84
6.85
6.86
6.87
6.88
6.89
6.90
6.91
6.92
6.93
6.94
6.95
6.96
6.97
6.98
6.99
6.100
6.101
6.102
6.103
6.104
6.105
6.106
6.107
6.108
6.109
6.110
Isovalue contours of instantaneous pressure fluctuation field

(range 2 Pa, black and white) computed from (i) LES inside the
injection interface and (ii) E3P (from LES data injection) outside the
injection interface (closer view)
Isovalue contours (range 3 Pa black and white) of instantaneous
pressure fluctuation field computed from (i) LES data inside the
injection interface, (ii) Euler data (from LES data injection) between the
injection interface and the Kirchhoff control surface, and (iii) from
Kirchhoff integration data beyond the Kirchhoff control surface
Numerical mesh (2D plane)
Sound-pressure level of LES
Sound-pressure level of 3D URANS
Experimental setup
Instantaneous streamlines in 2D URANS simulation
Instantaneous streamlines in LES with adaptive model
Lift and drag coefficients of the cylinder in the third-finest grid
Lift and drag coefficients of the cylinder in the second-finest grid
Lift and drag coefficients of the cylinder in the finest grid
Acoustic density fluctuations of 2D URANS simulation in the finest grid
Sound-pressure level of 2D URANS simulation in finest grid
Acoustic density fluctuations of LES with adaptive model in
second-finest grid
Sound-pressure level of LES with adaptive model in second-finest grid
Acoustic density fluctuations of 3D LES simulation in finest grid
Sound pressure level of LES with adaptive model in finest grid
Sound pressure level of LES in finest grid with Smagorinsky and Lilley
model
The Ahmed body from Ahmed et al. (1984) (isosurface of zero
streamwise velocity from Kapadia et al. 2003)
Mesh for the CFD model from Volkswagen
Streamlines in the wake of the CFD model
Vorticity in the wake of the CFD model
Pressure coefficient in the symmetry plane of the CFD model
Sound-pressure level at an observer point of x = 10 m, y = 10 m,
and z = 1 m
Aerodynamic computational domain
Acoustic computational domain
LES velocity field; longitudinal component (U = 14 ms1 ,
t = 6.6 102 s)
Acoustic results: acoustic power radiated by the diaphragm with
respect to the mean velocity (top) and acoustic power spectrum for
U = 14 ms1 (bottom)
Geometry of the cavity duct system
Snapshots of the pressure in the duct (left) and the vorticity in the
cavity (right) during a period of the oscillation
Geometry of the sudden enlargment
Snapshots of the Mach number for 1 = 5.5
Snapshots of the Mach number for 2 = 2.65
Opel 2004 Astra
331
333
335
337
337
338
339
340
340
341
341
341
342
342
343
343
343
344
345
346
347
347
348
348
351
351
352
352
353
354
355
355
356
360
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
xix
6.111 Lilley turbulence shear-source distribution illustrated by isosurfaces for

the idealized wing-mirror example of Siegert et al. (1999)
6.112 Mesh frequency cutoff (MFC) estimate; idealized wing-mirror example
of Siegert et al. (1999)
6.113 Predicted versus measured pressure spectra; idealized wing-mirror
example of Siegert et al. (1999)
6.114 Mach 0.85 cavity: symmetry-plane snapshot at t = 0.3 s; DES/k
(top) and URANS/k (bottom)
6.115 Overall (a) and band-limited (b) Prms along cavity ceiling centerline
6.116 PSD (kPa2 /Hz) at location x/L = 0.45
6.117 Diesel injector primary liquid spray breakup; liquid-free surface with
synthetic inlet perturbation (top) and without inlet perturbation (bottom)
6.118 Audi A2 full-vehicle geometry with localized domain shown in dark
6.119 Instantaneous velocity magnitude field
6.120 SPL against frequency at Microphone 4
6.121 Experimental (top) and simulated (bottom) pressure trace to 0.3 s at
x/L = 0.95
6.122 Sampling effects on overall Prms (kPa) along the cavity ceiling for
M219 experimental data
6.123 Sampling effects on overall Prms (kPa) along the cavity ceiling for
CFD data
6.124 Resonator geometry: application challenge from BEHR GmbH in the
DESTINY-AAC project
6.125 Velocity contours (top), pressuretime traces (bottom left), and spectral
magnitude (bottom right) at three bulk velocities (4, 8, and 12 m/s)
taken at the neck of the resonator
6.126 Acoustic response for 8 m/s case at the microphone
6.127 Experimental prototype with inlet cylinder and outlet filter removed
and locations of far-field monitors
6.128 Comparison of steady-state RANS and snapshots from the DES
calculation
6.129 Surface acoustic pressure (Pa) on the exterior model (a) and acoustic
pressure in the far-field (b)
6.130 Computed and measured dB(A) levels at the nine microphone
locations at the blade-passing frequency (BPF)
361
362
363
365
366
367
367
368
368
369
369
370
370
372
373
374
374
375
375
376
Tables
1.1
3.1
3.2
3.3
3.4
3.5
3.6
Sound-pressure levels for common sounds

Examples of usual spatial convolution filters
Various decompositions for the nonlinear terms
Resolution requirements referred to Kolmogorov length scale used in
DNS based on spectral methods of some incompressible
homogeneous and wall-bounded flows
Typical mesh size (expressed in wall units) for DNS and LES of
boundary layer flow
Definition of simulation types for compressible flows
Modified wave-number analysis of some classical centered finite
difference schemes
2
94
97
105
107
109
109
4:57
P1: JZZ
CUFX063-FM
xx
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
5.1
5.2
5.3
5.4
6.1
Coefficients for the DRP scheme of Equation (5.27)

Coefficients for filter formula 5.72
Optimized coefficients of the amplification factor for the LDDRK schemes
Typical mesh sizes (expressed in wall units) for a boundary layer flow
using DNS, wall-resolved LES, and LES with an appropriate wall model
LES of a ReD = 4 105 subsonic jet. Sideline sound levels and

-maxima in the shear layer for the different simulations
vrms
174
188
200
209
244
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
Contributors
Jean-Marc Auger
PSA Peugeot Citroen
1 route de Gisy
F-78943 Velizy-Villacoublay Cedex
FRANCE
e-mail: jeanmarc.auger@mpsa.com
Dr. Christophe Bailly
Centre Acoustique LMFA & UMR CNRS
5509
Ecole Centrale de Lyon
36, avenue Guy de Collongue
F-69134 Ecully
FRANCE
e-mail: christophe.bailly@ec-lyon.fr
Dr. Paul Batten
Metacomp Technologies, Inc.
28632-B Roadside Drive Suite 255
Agoura Hills, CA 91301
USA
e-mail: batten@metacomptech.com
Dr. Daniel J. Bodony
Stanford University
Department of Aeronautics and
Astronautics
Stanford, CA 94305-4035
USA
e-mail: bodony@stanford.edu
Dr. Christophe Bogey

Centre Acoustique LMFA & UMR CNRS
5509
36, avenue Guy de Collongue
F-69134 Ecully
FRANCE
e-mail: christophe.bogey@ec-lyon.fr
Priv.-Doz. Dr.-Ing. Michael Breuer
Lehrstuhl fur Stromungsmechanik
(LSTM)
Universitat Erlangen-Nurnberg
Cauerstr. 4
D-91058 Erlangen
GERMANY
e-mail: breuer@lstm.uni-erlangen.de
Ir. Tim Broeckhoven
Vrije Universiteit Brussel Department of
Mechanical Engineering
Fluid Mechanics and Thermodynamics
Research Group
Pleinlaan 2
B-1050 Brussels
BELGIUM
e-mail: tim@stro.vub.ac.be
xxi
4:57
P1: JZZ
CUFX063-FM
xxii
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
CONTRIBUTORS
Fabien Crouzet
Electricite de France
Analysis in Mechanics and Acoustics
Department
1 avenue du General de Gaulle
F-92141 Clamart Cedex
FRANCE
e-mail: Fabien.Crouzet@edf.fr
Jean Paul Devos
Electricite de France
Analysis in Mechanics and Acoustics
Department
FRANCE
e-mail: Jean-Paul.devos@edf.fr
Dr.-Ing. Roland Ewert
DLR Institute of Aerodynamics and Flow
Technology
Technical Acoustics
Lilienthalplatz 7
D-38108 Braunschweig
GERMANY
e-mail: roland.ewert@dlr.de
Oliver Fleig
The University of Tokyo
Graduate School of Engineering
Department of Mechanical Engineering
7-3-1 Hongo, Bunkyo-Ku
Tokyo 113-8656
JAPAN
e-mail: oliverfleig@hotmail.com
Xavier Gloerfelt
Laboratoire SINUMEF
ENSAM (Ecole Nationale Superieure
dArts et Metiers)
151, boulevard de lHopital
F-75013 Paris
FRANCE
e-mail: Xavier.Gloerfelt@paris.ensam.fr
Prof. Dr. Ir. Avraham Hirschberg
Faculteit Technische Natuurkunde
Technische Universiteit Eindhoven
CC 2.24, Postbus 513
NL-5600 MB Eindhoven
THE NETHERLANDS
e-mail: a.hirschberg@tue.nl
Prof. Fang Q. Hu
Department of Mathematics and Statistics
Old Dominion University
Norfolk, VA 23529
USA
e-mail: fhu@odu.edu
Dr. Thomas Huttl
MTU Aero Engines GmbH

Dachauer Str. 665
D-80995 Munchen
GERMANY
e-mail: Thomas.Huettl@muc.mtu.de
Andre Jacques
Mcube
54, Rue Montgrand - BP 232
F-13178 Marseille Cedex 20
FRANCE
e-mail: andre@mcube.fr
Dr. Manuel Keler, Dipl.-Phys.
Institut fur Aerodynamik und Gasdynamik
Universitat Stuttgart
Pfaffenwaldring 21
D-70550 Stuttgart
GERMANY
e-mail: kessler@iag.uni-stuttgart.de
Prof. Dr. Ir. Chris Lacor
Vrije Universiteit Brussel
Research Group
Pleinlaan 2
B-1050 Brussels
BELGIUM
e-mail: chris.lacor@vub.ac.be
Philippe Lafon
Laboratoire de Mecanique des Structures
Industrielles Durables
(LaMSID)
UMR CNRS EDF 2832
FRANCE
e-mail: Philippe.Lafon@edf.fr
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
xxiii
CONTRIBUTORS
Sanjiva K. Lele
Stanford University
Department of Aeronautics and
Astronautics & Department of
Mechanical Engineering
Stanford, CA 94305-4035
USA
e-mail: lele@Stanford.edu
Ir. Jan Ramboer

Research Group
Pleinlaan 2
B-1050 Brussels
BELGIUM
e-mail: jan@stro.vub.ac.be
Dr.-Ing. Franco Magagnato

Fachgebiet Stromungsmaschinen
Universitat Karlsruhe (TH)
Kaiserstrasse 12
D-76128 Karlsruhe
GERMANY
e-mail: magagnato@mach.uni-karlsruhe.de
Dr. Sjoerd W. Rienstra

Department of Mathematics
and Computer Science
Eindhoven University of Technology
P.O. Box 513
NL-5600 MB Eindhoven
THE NETHERLANDS
e-mail: S.W.Rienstra@tue.nl
Eric Manoha
ONERA/DSNA/BREC
BP 72
F-92322 Chatillon Cedex
FRANCE
e-mail: eric.manoha@onera.fr
Fred G. Mendonca
CDadapco
CFD Engineering Services Manager,
London
CD adapco Group, UK
200 Shepherds Bush Road
London W6 7NY
ENGLAND
e-mail: fred@uk.cd-adapco.com
Dimitri Nicolopoulos
Mcube
FRANCE
e-mail: dimitri@mcube.fr
Fred Perie
Mcube
FRANCE
e-mail: fred@mcube.fr
Prof. Pierre Sagaut

LMM - UPMC/CNRS
Laboratoire de modelisation en mecanique
Universite Pierre et Marie Curie
Boite 162, 4 place Jussieu
F-75252 Paris Cedex 05
FRANCE
e-mail: sagaut@lmm.jussieu.fr
Univ.-Prof. Dr.-Ing. Wolfgang Schroder
Lehrstuhl fur Stromungslehre und
Aerodynamisches Institut
RWTH Aachen
Wullnerstr. zw. 5 u. 7
D-52062 Aachen
GERMANY
e-mail: office@aia.rwth-aachen.de
Ir. Sergey Smirnov
Research Group
Pleinlaan 2
B-1050 Brussels
BELGIUM
e-mail: serg@stro.vub.ac.be
Philippe Spalart
Boeing Commercial Airplanes
P.O. Box 3707
Seattle, WA 98124-2207
USA
e-mail: Philippe.r.spalart@boeing.com
4:57
P1: JZZ
CUFX063-FM
xxiv
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
CONTRIBUTORS
Marc Terracol
ONERA
29 avenue de la Division Leclerc
F-92320 Chatillon
FRANCE
e-mail: marc.terracol@onera.fr
Sandrine Vergne
PSA Peugeot Citroen
2 route de Gisy
F-78943 Velizy-Villacoublay Cedex
FRANCE
e-mail: sandrine.vergne@mpsa.com
Dr. Claus Wagner

Deutsches Zentrum fur Luft- und
Raumfahrt e.V., DLR
Institut fur Aerodynamik und
Stromungstechnik
Abt. Technische Stromungen
Bunsenstrae 10
D-37073 Gottingen
GERMANY
e-mail: Claus.Wagner@dlr.de
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
Preface
Two branches of the same tree are growing together: Acoustics and the large-eddy
simulation (LES) technique are based on the same fundamental equations of fluid
dynamics. In the past, both scientific disciplines developed independently from each
other. Acoustics is one of the classical disciplines of mechanics, having its roots in Greek
and Roman times. LES is a comparatively young field of research that has benefited from
the exponential growth in computational possibilities over the last few decades. Each
scientific community has developed its own methods, definitions, and conventions, and
it sometimes seems that experts and scientists in acoustics and LES techniques speak
different languages. During the last few years, the LES and the acoustics communities
realized that LES can be a comprehensive tool for acoustical research and design and
intensified its use.
This book presents the current state of the art for LES used in acoustical investigations and comprises 19 contributions from 30 authors, each an expert in his field
of research. A general introduction to the subject is followed by descriptions of the
theoretical background of acoustics and of LES. A chapter on hybrid RANSLES for
acoustic source predictions follows. More details are given for numerical methods,
such as discretization schemes, boundary conditions, and coupling aspects. Numerous
applications are discussed ranging from simple geometries for mixing layers and jet
flows to complex wing or car geometries. The selected applications deal with recent
scientific investigations at universities and research institutes as well as applied studies
at industrial companies. Side areas of LES for acoustics are addressed in a contribution
on vibroacoustics.
The book is a collection of different methods, tools, and evaluation methodologies.
Currently it is not possible to offer a perfect solution methodology that generally covers
all possible applications. Although interesting results of several commercial codes
are presented, a recommendation for any specific solver cannot be made because a
benchmark of the codes has not been established and several other codes have not
been considered yet. Each method, both scientific and commercial, has its individual
advantages and weaknesses. It was also not our intention to harmonize the definitions
xxv
4:57
P1: JZZ
CUFX063-FM
xxvi
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 22, 2006
PREFACE
and conventions in acoustics and LES computing. Therefore, the same nomenclature
is not used by all authors.
The book is intended to be used by students, researchers, engineers, and code
developers willing to become more familiar with the use of the LES technique for
acoustical studies. The limitations of the method have been outlined as well as its
requirements. The reader should acquire an impression of possible and appropriate
applications for this methodology. The editors would welcome any initiatives motivated
by this book for international cooperation in the development or application of LES for
acoustics.
The idea for this book came from Eric Willner, the former commissioning editor for
engineering at Cambridge University Press, when he read the first call for papers for
the International Workshop on LES for Acoustics organized by Thomas Huttl, Claus
Wagner, and Jan Delfs in Gottingen, 2002. At this time, Cambridge University Press
was actively seeking a book on LES for acoustics for its aerospace series. Thomas Huttl
and Claus Wagner agreed to edit a scientific book based on the contributions of the
workshop in Gottingen. Several speakers and participants of the workshop and other
experts promised to contribute to the book, which was conceived as more of a scientific
handbook than a simple workshop proceedings. Pierre Sagaut separately developed the
idea of a book on LES for acoustics and joined the team of editors.
The book would not exist without the contributions from each of the authors. The
editors are not only grateful for these contributions but also for valuable review comments from several authors during two book reviews as well as interesting scientific
discussions of review comments and proposals. We would also like to thank Peter
Gordon, Senior Editor of Engineering at Cambridge University Press, for his help in
preparing the book but also for enthusiasm, patience, and confidence during the last 2
years when the progress of the book was sometimes slow but never stopped.
Thomas Huttl gratefully acknowledges the advice and comments of MTU Aero
Engines aeroacoustics specialist Fritz Kennepohl, who introduced him to the secrets of
acoustics during the TurboNoiseCFD research project.
Pierre Sagaut, Thomas Huttl, and Claus Wagner
Europe, May 2006
International ERCOFTAC-DGLR-DLR-Workshop on LES for Acoustics organized by T. Huttl, C.

Wagner and J. Delfs, German Aerospace Center (DLR), Gottingen, Germany, 78 October 2002.
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
LARGE-EDDY SIMULATION FOR ACOUSTICS
xxvii
November 22, 2006
4:57
P1: JZZ
CUFX063-FM
CUFX063/Wagner
0 521 87144 1
printer: sheridan
xxviii
November 22, 2006
4:57
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
23:37
Introduction
1.1 The importance of acoustic research

Thomas Huttl
There are some aspects of acoustics that significantly affect the quality of our daily lives.
By speaking, we transfer information and knowledge from one person to another. The
sound of rain, of wind or anything else gives us orientation and aids optical perceptions.
Music can fascinate us and stimulate our emotions and moods. Pleasant sounds and
music positively affect health by their calming character.
The negative side of acoustics is noise. Noise is the most commonly cited form
of environmental pollution. Noise is easily detected by the human hearing system. Its
effects can be cumulative, and it influences our work environment as well as our leisure.
Even the quality of our sleep is reduced if we are exposed to noise. In recent decades,
the effects of noise on people have been studied intensively.
1.1.1 Health effects
There is no doubt that noise has an impact on health. Very loud sounds are clearly highly
injurious to people as well as animals. Table 1.1 shows sound-pressure levels (dB(A))
for common sounds. At sound-pressure levels of 160165 dB(20 kHz) flies die when
exposed only for a short time. With these exposures levels, human beings become tired,
may experience facial pain, and may develop burned skin. When the sound pressure
is lowered, reactions to the sound decrease. Long-term exposure to high noise levels
of about 90 dB(A) can result in permanent hearing loss. Even for a steady daily noise
level of 75 dB(A) for 8 hours per day, there is a risk of permanent hearing damage after
40 years of exposure. Apart from the auditory effects of noise, nonauditory effects on
health are well known. Noise can induce a range of physiological response reactions
such as increases in blood pressure, heart rate, and breathing, and these reactions are not
confined to high noise levels and sudden noise events but are also true for noise levels
commonly experienced in noisy environments such as busy streets (Nelson 1987).
1
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
Table 1.1. Sound-pressure levels for common sounds

(dB(A))
Common sounds
30
50
60
70
80
90
100
110
120
130
140
180
whisper
rainfall, quiet office, refrigerator
dishwasher, normal conversation
traffic, vacuum cleaner, restaurant
alarm clock, subway, factory noise
electric razor, lawnmower, heavy truck or road drill at 7 m
garbage truck, chain saw, stereo system set above halfway mark
rock concert, power saw
jet takeoff, nightclub, thunder
jack hammer
shotgun, air raid system
rocket-launching pad
Exposure to excessive noise during pregnancy may result in high-frequency hearing

loss in newborns and may be associated with prematurity and intrauterine growth
retardation (American Academy of Pediatrics 1997).
1.1.2 Activity effects
Noise influences human activities such as sleep, communication, and peoples general
performance. Studies have shown that noise can affect sleep in many ways. Noise may
shorten the length of the sleep period and increase the number or frequency of awakenings, and it may affect the duration of the various stages of sleep. Given the importance
of sleep for individual health, a perturbation of sleep is usually not tolerated by people.
First- and second-grade school children chronically exposed to aircraft noise have
significant deficits in reading as indexed by standardized reading tests administered
under quiet conditions. Chronic noise may also lead to deficits in childrens speech acquisition (Evans and Maxwell 1997). Other studies show the impact of noise from
aircraft, road traffic, and trains on long-term recall and recognition (Hygge 1993;
Groll-Knapp and Stidl 1999). Mood and behavioral abnormalities can also be related
to noise exposure.
1.1.3 Annoyance
In addition to the direct effects of noise on sleep, communication, and performance there
are also indirect consequences of annoyance or disturbance that are related to the way
a person feels about noise. Annoyance is a very complex psychological phenomenon,
and it is scarcely possible to define or measure it here. For physically loud sounds the
perceived annoyance is often equivalent to the perceived loudness. For physically soft
sounds (rustling papers at the movies, people talking while watching television), the
perceived annoyance deviates greatly from perceived loudness. Thus, for physically soft
sounds, more attention is paid to other aspects of the sound than to the loudness-related
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1.1 THE IMPORTANCE OF ACOUSTIC RESEARCH
sound-pressure level (Berglund, Preis, and Rankin 1990). Nevertheless, annoyance

effects can also influence mood and behavior and can ultimately cause even severe
health problems.
1.1.4 Technical noise sources
Many noise sources are man-made especially transportation noise from road traffic,
aircraft, and trains. Other technical noise sources can also be annoying such as wind
turbines or cooling and climate systems. Noise can be produced by several physical
interaction mechanisms:
r Solid-body friction noise (e.g., gearbox),
r Solid-body vibration (e.g., unbalanced rotating machines),
r Combustion noise (e.g., piston engines),
r Shocks (e.g., explosions or pneumatic hammer), and
r Aerodynamic noise (e.g., vortexstructure interaction)
In this book we concentrate on aerodynamic, flow-induced noise.

Aerodynamic noise is one of the major contributors to external vehicle noise emission as well as of internal vehicle noise due to the transmission of the externally
generated noise through structure and window surfaces into the cabin. Aerodynamic
noise becomes dominant at driving speeds exceeding 100 km/h when compared with
structure-borne, power train, and tire noise, for which substantial noise reduction has
been achieved. The interaction of the flow with the geometrical singularities of the vehicle body produces unsteady turbulent flows often detached resulting in an increased
aerodynamic noise radiation (Vergne et al. 2002).
Aircraft noise is dominant for residents near airports when planes fly at low altitudes
such as during departure and landing. The engines, especially their free-jet flow but
also flaps, wings, airbrakes, landing gear, or openings contribute significantly to the
total sound emission. When an aircraft is flying at cruising altitude, the aerodynamic
noise of the aircraft body and propeller, or engine noise, can be annoying for passengers
and crew.
For trains, the bow collector and the wheelrail interaction are dominant noise
sources. For high-speed trains, especially unconventional concepts like the magnetic
levitation hover train project Transrapid (Figure 1.1), the relevance of aerodynamic
noise is increased.
Wind turbines in operation emit aerodynamic noise that can be perceived by humans.
With respect to large wind turbines with low rotational speed, the main contribution to
aerodynamic noise is the narrow-band noise caused by bladetower interaction. Smaller,
fast-rotating wind turbines, on the other hand, emit broadband noise, which is mostly
caused by vortexblade interaction (Fleig, Iida, and Arakawa 2002).
Cooling and climate systems are designed to regulate temperature and filter out pollution. Their power units often emit tonal noise at constant frequency that is transported
through the duct system. Furthermore, tonal noise due to aeroacoustic resonance can
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
Figure 1.1. Magnetic levitation hover train project Transrapid.
be activated (Hein, Hohage, and Koch 2004). Although the noise level of cooling and
climate systems is usually very low, it is annoying and should be considered by the
designer.
1.1.5 Political and social reactions to noise
Noise is becoming generally accepted as an environmental and even health hazard to

the population. Public reaction to noise problems is causing governments to adopt laws,
regulations, and guidelines for the certification of noise-emitting vehicles and machines
as well as for temporal or spatial limitations on their use. Aircraft and jet engine manufacturers face increasingly stringent noise requirements for near-airport operations
worldwide (Bodony and Lele 2002a). Old-fashioned and noisy aircraft may not be operated from an increasing number of airports or their operators must pay additional fees for
noise emission. Some aircraft may not be operated during the night if their noise emission is too high. Airlines have to consider the noise-related airport fees in their operating
costs or reduce the noise emissions by adding new or additional noise-reducing devices
to their planes and jet engines. A sociocultural aspect of traffic noise is that the prices
for houses and apartments are lower if they are close to highways and roads, railways,
or flying routes near the airports. Costly noise-reduction measures like fences, walls,
and special windows are required to compensate for the increased traffic noise.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1.1 THE IMPORTANCE OF ACOUSTIC RESEARCH
Butterworth-Hayes (2004) describes the European Union (EU) plans to overtake

the United States as the largest aerospace power in the world by 2020. The 2001 report European Aeronautics: A Vision for 2020, which was drawn up by a group of
wise men to advise the European Commission (EC) on the future of the continents
aerospace industry, laid out a series of research objectives that Europes civil aircraft
manufacturers must achieve to ensure dominance of the market. One of the key objectives of Vision 2020 research is the halving of perceived aircraft noise. This means,
in particular, reducing external noise by 45 dB and 10 dB per operation in the short
and long terms, respectively. For rotorcraft, the objective is to reduce the noise footprint area by 50% and external noise by 6 dB and 10 dB over the short and long terms,
respectively.
1.1.6 Reactions of industry
The aircraft industry projects a growth in passenger kilometers of 100% or more in the
next 15 years. Satisfying the resulting demand for larger or faster airplanes, or both,
requires that new airplanes be designed. The same is true for future high-speed trains,
which, in order to be able to compete with air traffic, have to become faster as well.
This would boost the emitted noise levels tremendously if these trains were built based
on todays technology because the emitted aerodynamic noise increases approximately
with the fifth to sixth power of the vehicle speed (Schreiber 1995).
After having achieved significant progress in reducing the level of other primary
noise sources such as piston engine noise or road-contact noise, the automotive industry
is now contending with the major problem of interior and exterior noise related to
aerodynamic effects. Aeroacoustics is becomimg increasingly important in many other
fields such as the energy industry or personal computer manufacturing industry.
Sensitivity to noise is increasing all over the world. Therefore, significant noise
reduction in new airplanes, trains, and cars is mandatory if this growth in the transportation system is to be accepted by the population and their political representatives.
This noise reduction can only be realized if the design process is guided by robust and
fast computational aeroacoustic methods. Owing to the lack of commercial software,
many companies and research institutes are using their specialized, self-developed,
in-house codes for solving engineering problems. Remarkable progress has also been
made by designers of commercial codes for aeroacoustic applications, although these
codes are currently still under development (Wagner and Huttl 2002).
1.1.7 Research on acoustics by LES
Aeroacoustics is the scientific discipline between fluid mechanics and classical acoustics. It considers sound generated by aerodynamic forces or motions originating from
(turbulent) flows (Ihme and Breuer 2002). Initially, experimental investigations were
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
used to derive some empirical relations for estimating the noise emission of new technical products. However, owing to strongly increased computer performance, the numerical simulation of acoustic fields generated by fluid flow, called computational
aeroacoustics (CAA), has become very attractive. At this time no unique solution procedure exists for all acoustic problems. Instead, various strategies have been developed,
each with individual advantages and disadvantages (Ihme and Breuer 2002).
For applications with complex, inhomogeneous flows and flow-induced noise
radiation, the most promising and commonly used numerical technique is to adopt
a hybrid approach. In such an approach, the sound-generation and sound-propagation
processes are considered separately. A nonlinear aerodynamic near field in which the
aerodynamic perturbations are generating the sound is matched to a linear acoustic
far field in which no flow or homogeneous flow exists and the sound waves are only
propagating. The underlying assumption is that there is no feedback of the acoustic
waves on the flow.
Coupled or hybrid approaches are currently being created by several research groups
and code developers. Such efforts have also been the focus of two recently finished
European research projects: Application of Large-Eddy Simulation to the Solution of
Industrial Problems (ALESSIA) and TurboNoiseCFD. The primary aim of the EU
ALESSIA project has been to develop software tools for the simulation of fluctuating
flows by large-eddy simulation (LES), with a particular focus on flow-induced acoustics
(Montavon 2002). Commercial computational fluid dynamics (CFD) and CAA codes
have been interfaced within ALESSIA.
The aim of the TurboNoiseCFD project was to contribute to the objective of a
10-dB reduction in 10 years in aircraft external perceived noise through new design
technologies. To achieve this objective, the aircraft engine manufacturing industry
significantly enforces to reduce engine noise levels at the source. In response to this
challenge, new methods have been created that will aid in the design of low-noise
turbomachinery components based on the adaptation of existing CFD software and
its integration with propagation and radiation models. Besides other techniques, an
LES methodology has been tested for aeroacoustical evaluations of broadband noise
(Boudet, Grosjean, and Jacob 2003; Jacob et al. 2005).
After the successful funding of noise reserach programs under the Fifth Research
Framework Program (19972001), the EC continues to invest money in research
projects aimed at reducing the environmental impact of aviation (fuel consumption,
noise pollution, and emissions of carbon dioxide (CO2 ), nitrous oxides (NOX ), and
other chemical pollutants) in its Sixth Research Framework Program (20022006).
Other coupling methods are also called hybrid methods such as the coupling of the Reynolds-averaged
NavierStokes (RANS) approach and large-eddy simulation (LES) for detached-eddy simulation
(DES); see Chapter 4.
Turbomachinery noise-source CFD models for low-noise aircraft designs. This project was funded by
the EC within the GROWTH Fifth Framework Program 19982002.
ALESSIA is an ESPRIT Project funded at 50% by the EC.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1.2 INTRODUCTION TO COMPUTATIONAL AEROACOUSTICS
1.2 Introduction to computational aeroacoustics

Manuel Keler
Computational fluid dynamics (CFD) has reached a level of maturity that permits many
industrially relevant problems to be solved routinely using commercially available tools,
although some difficult problems are still out of reach. Consequently, the research interest in the fluid mechanics community has shifted slightly, and there is now a large
and still growing group of experts engaged in the field of acoustics. For a long time
their work was mostly based on analytical and experimental studies, but the astonishing advances in computer technology have made a numerical approach feasible. That
approach is called computational aeroacoustics (CAA).
1.2.1 Definition
Numerous definitions exist for CAA reflecting the many people who have been
attached to this subject. We understand CAA here in the broadest possible sense
that is, as a process using some kind of numerical computation to produce acoustical information for aerodynamic phenomena. That obviously includes all flavors of
acoustical transport techniques (Lighthills acoustic analogy, the Kirchhoff method, the
Ffowcs WilliamsHawkings equation), linearized Euler approaches, combined procedures with CFD, semiempirical treatments like stochastic noise generation (SNGR),
and even compressible direct numerical simulation (DNS) admittedly very rarely
used for acoustic analysis these days.
Experimentors evidently use numerical computations for data processing and evaluation as well. However, although experimental studies are of substantial significance
in the efforts to find appropriate models, tune constants, and validate computations,
they do not constitute CAA.
1.2.2 History
Even though sufficient computing power has become available only recently, CAA has
quite a long tradition in the fluid mechanics community. Among the earliest efforts was
the paper by Gutin (1948) published first in Russia in 1936. However, modern CAA
rests mostly on the shoulders of Sir James Lighthill (1952, 1954), who published what
are certainly the most influential and therefore the most cited papers in aeroacoustics.
He introduced the idea of representing sound as the difference between the actual flow
and a reference flow usually a quiescent medium at rest. Because sound pressure and
velocities are in general small perturbations around a background flow, approximations
are possible to simplify the problem. In the late 1960s, his acoustic analogy approach
was extended by Ffowcs Williams and Hawkings (1969) to the case of moving surfaces
immersed in the flow (and acoustic) field, which is second in citations only to the work
of Lighthill. Another classic of its time was Goldsteins (1976) book.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
Nevertheless, owing to the highly abstract and mathematical presentation of the

subject and the lack of appropriate aerodynamic simulation data, progress was slow
during the next two decades. Only sound emitted from jets received considerable interest
in the early days of CAA mostly for their simple geometry (because there are no solid
walls) as well as the importance of this sound source for the noise levels of aircraft
developed at that time. Many discoveries of basic sound-generation mechanisms and
scaling laws date from this time.
Computers have only recently become powerful enough to tackle other, more difficult aeroacoustical problems one way or another with sufficient accuracy to provide
results valuable for industry in the design process. Associated with increasing societal
interest in noise reduction, another golden age in aeroacoustics dawned in the late 1980s
and early 1990s. CAA gained momentum since then, and there is still no slowdown in
sight.
1.2.3 Aeroacoustics
Aerodynamic noise occurs because of two basically different phenomena. The first
one is impulsive noise, which is a result of moving surfaces or surfaces in nonuniform
flow conditions. The displacement effect of an immersed body in motion and the nonstationary aerodynamic loads on the body surface generate pressure fluctuations that
are radiated as sound. This kind of noise is deterministic and relatively easy to extract from aerodynamic simulations because the required resolution in space and time
to predict the acoustics is similar to the demands from the aerodynamic computation.
Aerodynamic noise arises primarily from rotating systems (e.g., helicopter rotors, wind
turbines, turbine engine fans, and ventilators). If the surfaces move at speeds comparable to the speed of sound or there is an interaction between a rotor and a stator wake,
these tonal noise components can be dominant.
The other noise mechanism is the result of turbulence and therefore arises in nearly
every engineering application. Turbulence is by its very nature stochastic and therefore
has a broad frequency spectrum. Interestingly enough, turbulent energy is converted
into acoustic energy most efficiently in the vicinity of sharp edges (e.g., at the trailing
edge of an aircraft wing). In this case the uncorrelated turbulent eddies flowing over the
upper and lower sides of the edge have to relax with each other, generating locally very
strong equalizing flows that result in highly nonstationary pressure spikes. Another
major source of turbulence sound is jet flows, in which the shear layer in the mixing
zone again radiates into the far field.
A third but here neglected phenomenon is the case of combustion noise, which is
a result of the chemical reactions and the subsequent introduction of energy into the flow.
As previously stated, turbulence noise almost always exists, and as a consequence
aerodynamic noise is usually a broadband noise sometimes augmented by narrow-band
tonal components coming from impulsive noise sources. Impulsive noise can usually
be derived from nonstationary aerodynamic calculations. Like CFD, turbulence noise is
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
much more difficult to simulate because the turbulence has to be either simulated fully,
as in DNS, or modeled as in the Reynolds-averaged NavierStokes (RANS) approach,
or something in between, as in large-eddy simulation (LES) or hybrid computations.
1.2.4 Conceptual approaches
In the CFD area, several tools have been developed to a very high level of maturity that
makes them not only useful as a scientific research instrument but also as an industrial
design tool. Some of these tools have reached sufficient reliability to make them helpful
for users not considered experts in their field. However, pushing tools to their limits
can lead to disaster and often to nonobvious paths.
Because CAA is a more recent domain of activity, the situation is much less favorable. There is not yet a clear path to follow for reliable acoustical information for
each and every application, which may consist of such different things as a cooling fan
for a personal computer or a supersonic jet driving an airplane. Consequently, many
different techniques exist nowadays, each working well in one area and failing totally
in another. We try to classify a few of them in the following paragraphs.
Direct methods can be considered the most exact technology for CAA and are
comparable to the DNS in the CFD field. The complete, fully coupled compressible
Euler or NavierStokes equations are solved in the domain of interest for the unsteady
combined flow and acoustic field from the aerodynamic effective area down to the
far-field observer. They do not include any modeling of the sound (besides, possibly, a
turbulence model) and thus do not suffer from modeling or approximation errors. Of
course, they require tremendous computational resources because especially in the
case of small-Mach-number flows flow and acoustics represent a multiscale problem
with its inherent difficulties. The difficulty here is that the small acoustic perturbations
are not drowned out by numerical errors of the much larger aerodynamic forces. Space
and time resolution requirements for the aerodynamic data combined with the large
distances up to an observer in the far field give rise to ridiculously high numbers of cells
and time steps. Even if the necessary computer power were available, the discretization
schemes well known from CFD do not work very well in CAA applications because
they have dispersion and diffusion errors that are much too high. A plane wave is usually
severely distorted and dampened after being transported for just a few wavelengths,
which is clearly too short for the common case of an observer in the far field.
Direct methods are deceptively attractive because well-known and well-understood
CFD packages promise to provide aerodynamic and aeroacoustic data at the same
time. Sometimes they do work surprisingly well mostly in those cases in which the
differences between aerodynamics and aeroacoustics are negligible as in transonic
problems. However, for many or most other problems they do not because the basic
requirements of CFD and CAA are just too different. Several CFD schemes are tuned
specifically to suppress spurious acoustic waves, which of course is a bad idea when one
is interested in the acoustic properties of a problem at hand. CFD usually is designed
23:37
P1: JzG
0521871441c01
10
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
to solve a near-field problem because the perturbations from the mean flow vanish
quickly. Furthermore, the flow in this region is usually highly nonlinear but basically
stationary, or at least changing only slowly (aside from turbulent motions). Acoustics,
on the other hand, is clearly a far-field problem in which sound is generated locally
in the aerodynamic active area and passively radiated outside to an observer with a
smaller exponent of decrease. Outside the aerodynamic active area where the sound
is generated the perturbations are small, and a linear description is usually sufficient.
However, noise is inherently unsteady with time scales quite comparable to turbulent
eddies even if the spatial wavelengths are large compared with aerodynamic ones by
an order of the reciprocal of the Mach number.
These different properties (linear versus nonlinear, far-field versus near-field, timedependent versus (quasi) stationary, large versus small spatial scale) obviously necessitate different tradeoffs regarding computational schemes. No currently known schemes
score best on all possible requirements of CFD and CAA; therefore, methods especially adapted for the specific demands of an application will always be superior to
more general ones that necessarily have to balance all their characteristics carefully.
Most of the computational aeroacoustic tools in successful use nowadays are therefore of the hybrid type in which sound generation due to aerodynamics is more or less
decoupled from the acoustic transport process to the far field, making it possible for
tailored algorithms to be used for both tasks. This decoupling leads straightforwardly
to an arbitrary combination of a sound generation method with another sound transport
method.
CFD Sources: On the sound-generation side some kind of CFD tool is primarily
in place. If a direct coupling mechanism to the transport method is used it has
to provide sound data in the coupling region (surface or volume). In this case
aeroacoustic applicability is very important; that is, dispersion and diffusion
errors must be at the lowest possible levels. However, the demands are not as high
as for direct methods because an undistorted transport has to be sustained only up
to the coupling regions, which are seldom farther away than a few wavelengths.
A small error in phase and amplitude is therefore acceptable.
Semiempirical Sources: Alternatively, the sound sources can be reconstructed
semiempirically from CFD data derived mostly from turbulence quantities. A
straightforward, steady RANS computation provides information about turbulent length and time scales that translate by empirical relations into sound-source
spectra. These spectra are then radiated by one of the transport methods described
in the next paragraph. Of course, this process depends heavily on the soundness of
the empirics and the validation data used to calibrate them. However, the methods
can be very fast and reliable for obtaining a judgment between two close configurations (e.g., in the acoustic optimization of an airfoil shape). In such situations,
with a small and well-defined application domain close to the calibration data,
they are fairly useful.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
After identification of the sound sources in the aerodynamic active area, the noise
generated in this near field has to be transported outside to an observer. Again we may
choose between two alternatives:
Computational transport methods are similar to a CFD computation, in the sense
that they solve some partial differential equations(s) in the entire field up to the
observer. In contrast to direct methods, they do not simulate the aerodynamic
area as well but concentrate on the solely acoustic domain. Therefore, several
assumptions and, consequently, approximations are allowable to achieve accurate solutions efficiently. Usually computational transport methods solve simple
equations such as the linearized Euler equations (LEE) or simply the wave equation. Because these are both linear, many problems encountered with the full
set of equations do not occur, and consequently the discretization schemes can
be highly tuned to reach the desired low level of dispersion and diffusion errors. On the boundary between the CAA domain and the CFD computation, the
CFD solution is used as a boundary condition for the CAA simulation. Owing
to the change in discretization, resolution, and even equations the definition of
proper transmission conditions is a highly nontrivial task. The computational
cost can grow tremendously if the observer is far away because all the space in
between has to be solved for. However, if an entire noise carpet computation is
required in contrast to a single-observer spectrum, this approach may be beneficial inasmuch as the cost is the same regardless of the number of observer
points.
Analytical transport techniques employ an integrated form of the relevant acoustic propagation equation either Kirchhoffs surface integral or the Ffowcs
WilliamsHawkings (FWH) equation. In this case the sound pressure at an
observer at a specific point in time is computed by an integration-of-source term
along a surface either a physical one or surrounding the aerodynamic area and
possibly additional volume integrals outside the surface in the case of the FWH
equation. Owing to the finite speed of sound and the deterministic relationship
between emission and observer time of a signal, there has to be some kind of
interpolation of the data at least on one side. In the case of parts of the integration surface or volume moving at transonic speeds, the integrals become highly
singular because of the Doppler effect, which leads to difficulties regarding the
numerical stability of the procedure.
Taking all these different techniques together, we end up with a general map of noise
prediction methods shown in Figure 1.2. Given all the different numerical formulations
and specific implementations of each method, there is a plethora of variants to choose
from. In this book we want to delve in more detail into the promising part of this tree
emphasized in bold.
11
23:37
12
Wave
equation
Computational
transport
Acoustic
analogy
(FWH)
Analytical
transport
Kirchhoff
integral
printer: sheridan
Figure 1.2. Noise prediction methods.
Linearised
Euler
equation
Acoustic
perturbation
equation (APE)
Transport
method
0 521 87144 1
Reconstructed sources
(RANS+SNGR/SATIN)
Sound
generation
Hybrid methods
CUFX063/Wagner
Resolved sources
(DNS,LES,DES)
CFD
analysis
Direct methods
CAA
P1: JzG
0521871441c01
November 16, 2006
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1.2.5 Remaining problem areas for sound computation
Despite the progress made in recent years and the substantial number of researchers in
the field, there are still several unsolved issues involving the numerical prediction of
noise. Because CFD still has to struggle with (at least) turbulence modeling, transition
to turbulence, and separation (even more so with relaminarization and reattachment),
the CAA community has not yet reached consensus about the proper treatment of sound
computations in most application areas. Consequently, some remarks concerning these
issues seem to be in order.
Sound is a fundamental nonstationary process. In physical time, the sound waves
generated by some aerodynamic processes in the active region have to be transported
for at least some distance. So that the sound generated (e.g., at a trailing edge) will not
be totally destroyed, it is of paramount importance to perform this transport process as
accurately as possible. This holds especially for two basic features of sound: the wave
speed and the amplitude. Only specifically tuned methods can guarantee that waves will
be supported for a time and distance long enough to be of practical interest, and even
fewer methods can do this for the correct phase speed. Short wavelengths at the edge
of resolvability by the spatial discretization are particularly prone to numerical errors.
They usually encounter totally nonphysical propagation properties possibly up to the
point that the acoustic solution is totally spoiled.
1.2.5.1 Discretization
The main problem of CAA is the disparity of energy, length, and time scales between
the aerodynamics and the aeroacoustics especially at smaller Mach numbers. The
ratio of noise energy to mechanical energy is on the order of Pnoise /Pmech 104 M5 .
For a low-speed case of M = 0.1 we obtain a ratio of 109 , and even for an airliner
at M = 0.7 the number is only 105 . In an aerodynamic simulation we of course introduce numerical errors. To be at the same level as the acoustics we are interested
in, these errors have to be five to nine orders of magnitude smaller than the intended
physical values. To obtain acceptable signal-to-noise ratios for sound levels, we have
to add at least one more order of magnitude. In this sense, basically every CFD simulation is very loud, counting just the numerical errors that introduce numerical noise.
Common CFD schemes are adapted to stationary simulations and therefore just suppress acoustic waves, and so for the aerodynamic community the problem seems to be
solved.
For the purpose of aeroacoustics, on the other hand, the discretization schemes in
both space and time must be specifically tuned. Diffusion and dispersion errors have
to be reduced to the lowest possible level. This noble goal is achieved only partly,
but progress has been made to develop highly accurate discretization techniques to
reach the desired levels. Another problem is the introduction of artifical noise sources
caused by numerical errors that may overwhelm the physical sound sources and thus
generate entirely unreliable noise information. Finally, if we use dimensional splitting
13
23:37
P1: JzG
0521871441c01
14
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
techniques, discontinuities in the metrics of grid lines (e.g., at corners and edges) may
cause spurious noise reflections.
1.2.5.2 Boundary conditions
At the very heart of CAA lies the proper handling of boundary conditions, which is a
situation not very different from that of CFD. Because acoustics is a radiation problem,
basically all the sound energy will sooner or later try to leave the computational domain.
On solid walls, we obviously get reflections, which we can handle straightforwardly just
as in CFD. However, on artificial far-field boundaries, the physics dictates a straight
pass-through without any spurious reflections. Although this requirement seems to
be obvious, it is indeed very hard to fulfill sufficiently. Several concepts have been
developed for this specific problem and optimized for one application or another, but
the downside is that a generally accepted solution does not yet exist.
Interior boundaries can be problematic as well. If we couple a CFD tool to a computational transport method, there is a difference between the models, and we may
therefore obtain spurious reflections at the interface of the boundaries. The same holds
for other interior interfaces, even in the same code (e.g., in the interpolation region
between Chimera-type overlapping grids). Any difference in modeling, discretization,
or resolution is a potential cause for most unwanted reflections.
1.2.5.3 Coupling
Owing to the fundamental differences between aerodynamics and aeroacoustics in many

applications, a hybrid approach seems to be a sensible way to tackle the sound prediction
problem, as stated before. However, in this case the problem of coupling between the
two different algorithms inevitably arises. There is no theory of backscattering (i.e.,
a feedback from the acoustics to the aerodynamics) because this is not relevant in
any engineering application. Even the sole direction from a flow solution to a sound
computation, however, is difficult enough to handle properly in order to guarantee
continuity, conservation, and, again, the avoidance of artificial reflections at the coupling
while admitting physical acoustic waves. This is especially true when the two schemes
operate in different spaces (e.g., the flow solver in physical time and the acoustic
prediction in the frequency domain).
1.2.5.4 Conclusions
After having looked at some of the issues and partly unsolved problems in computational aeroacoustics, we can legitimately claim that CAA will be a demanding and
exciting field for surely some time to come. So let us step forward to the collective
wisdom of the experts in the hope of overcoming some of the obstacles in order to help
the researcher and later on the engineer with his or her daily work on the subject of
design to noise.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1.3 STATE OF THE ART: LES FOR ACOUSTICS
1.3 State of the art: LES for acoustics

Claus Wagner, Oliver Fleig, and Thomas Huttl
1.3.1 Broadband noise prediction in general
Peoples perception of noise is changing and their sensitivity toward it is increasing all
over the world. Significant noise reduction to accommodate this development has been
realized by controlling or damping tonal noise sources, but a further increase in noise
pollution is expected. The aircraft industry, for example, expects a growth in passenger
kilometers of 100% or more in the next 15 years. At the same time new high-speed
trains are under development that will allow railways to compete with air traffic on
short-distance routes.
Further noise reduction is mandatory if this growth in the transportation system
is to be accepted by people and their political representatives. To achieve this, the
design process needs support from computer-based noise prediction tools. To provide
the needed robust and efficient methods, research on reliable CFD and CAA methods
has increased tremendously in the last several years. The currently applied numerical
approaches range from extremely costly DNS to hybrid approaches that solve timeaveraged governing equations to obtain the near flow field and a wave propagation
equation fed by a synthetic turbulent flow field.
During the last 20 years, CFD has made impressive progress. There is no doubt
that CFD methods will play an indispensable role in the industrial design process,
but for several reasons the use of CFD methods for aeroacoustic predictions faces
many difficulties. One of the reasons is that the methods solve the Reynolds-averaged
(time-averaged) NavierStokes (RANS) equations, which are generally time independent. Aeroacoustics problems though, are, by definition, highly time dependent. This is
the reason more sophisticated methods must be developed and investigated to predict
broadband noise generation and propagation reliably.
The straightforward method for predicting broadband noise is to solve the compressible NavierStokes equations using direct numerical simulations. Initial work on
this approach by Mitchell, Lele, and Moin (1997) and Mankbadi et al. (1998) focused
on supersonic jets. Resolving all scales of a turbulent flow in a DNS properly requires
that the discretized equations be solved on extremely fine grids because the size of
the smallest turbulent scales decreases with increasing Reynolds numbers. For threedimensional flows, this results in computing times that scale with the third power of
the Reynolds number. Because most technically relevant flows are characterized by
high Reynolds numbers, it can easily be shown that it will be impossible to use DNS
for applied turbulent flow problems in the near future. Additionally, a huge computational domain has to be chosen to simulate the far-field acoustics. To fulfill these two
conditions at the same time will be a challenge for some generations of researchers
to come. Nevertheless, direct numerical calculations of noise have been performed
for some simple geometries. Examples are the two-dimensional driven cavity flow, for
15
23:37
P1: JzG
0521871441c01
16
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
which direct simulations of the flow and the noise have been conducted by Gloerfelt,
Bailly, and Juve (2001); Colonius, Basu, and Rowley (1999); and Shieh and Morris
(1999b).
A more promising approach for technically relevant aeroacoustic problems is to
apply hybrid methods. With this approach, the near-field aerodynamics are computed
to obtain velocity and pressure fluctuations that form the acoustic source terms for a
separate computation of the far-field acoustics. The reason for splitting off aerodynamics from aeroacoustics is the great disparity of levels and length scales between the flow
and aeroacoustic fields. For both the simulation of the flow and the computation of the
sound waves, a variety of methods exist that differ in accuracy and in their demand
for computational resources. As mentioned before, flow simulations can be conducted
in DNS with the restriction to flows with low Reynolds numbers and rather simple
geometries. On the other hand RANS simulations in CFD are applicable for complex
geometries and high Reynolds numbers. To make predictions of the radiated noise it is
necessary to describe the properties of the turbulence. Because a complete simulation
of turbulence and its generated and radiated noise are computationally too expensive
for high-Reynolds-number flows, noise predictions are often based on the solution
of RANS equations using a two-equation turbulence model like the (k) turbulence
model. Such solutions provide an estimate of the amplitude of turbulent velocity fluctuations, but because this approach provides only time-averaged properties it is necessary
to make assumptions about the statistical characteristics of the turbulence. In particular,
two-point cross correlation of the turbulent noise sources must be approximated.
One of the earliest attempts to couple estimates of the statistical properties from
steady-flow predictions with a noise model based on acoustic analogy was the approach
developed by Balsa and Gliebe (1997). More recent extensions have used RANS solutions for the flow field based on a (k) turbulence model. Tam and Auriault (1999),
Page et al. (2001), and Bailly, Candel, and Lafon (1996), for example, performed RANS
computations of jet flows to predict the radiated noise based on such a coupled method.
The perfectly suited method to compute the large-scale fluctuations, which are
known to contribute most to the noise generated in many problems, is the LES technique.
To investigate this, Avital, Sandham, and Luo (1996) conducted DNS of temporally
evolving mixing layers with transonic and supersonic velocities and used these to
estimate the sound emitted by corresponding spatially evolving mixing layers. The
sound predictions were obtained applying Lighthills analogy, and a large-scale model
for the implementation of Phillips analogy (Phillips 1960). For high Reynolds numbers,
this method allows prediction of the dynamics of the large turbulent scales, whereas the
effect of the fine scales is modeled using a subgrid-scale model. When one considers
turbulent flows in the absence of walls like jet flows or isotropic turbulence, it is easy
to estimate the spatial resolution required to capture a wave number at the lower limit
of the inertial range in a LES. Assuming a Reynolds number of Re = 8.7 106 for a
supersonic turbulent jet, Tam (1998b) estimated that a mesh with 2.1 109 grid points
is required to accurately perform a LES of the noise of a supersonic jet. Although
currently very few research institutions have the computer capability to conduct such
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
simulations, the use of LES as a tool for noise research appears to be feasible. However,
practical use of LES for noise prediction and design is still some years away.
The most general methodology for the prediction of far-field jet noise is to compute
the near-field unsteady flow field using a DNS or LES technique in conjunction with an
acoustic analogy. Following the flow computation, an acoustic source term is determined
from the large-scale fluctuating flow field. The far-field acoustics are then calculated
in a larger domain on grids, which are generally characterized by their coarser spatial
resolution. As determined by the aeroacoustic problem considered, different methods
can be applied. The most common approaches for noise predictions are discussed in
the next section.
1.3.2 Broadband noise prediction based on LES
In recent years a variety of large-eddy simulations were performed with the objective of
generating the source terms for wave propagation equations. In spite of this, the investigated configurations are limited. We want to give some examples for the application of
LES to acoustics and classify them with respect to the investigated configuration. Note
that the work listed below is not complete. Our main objective is to demonstrate that the
hybrid approach has been applied for noise predictions in different configurations with
a variety of combinations that used subgrid-scale models and noise prediction methods.
1.3.2.1 Decaying and forced isotropic turbulence
Seror et al. (2000) addressed the problem of evaluating and modeling the contribution
of the unresolved scales to the radiated noise production when Lighthills analogy is
employed together with LES. To achieve this, they split the Lighthill tensor into three
parts: a high-frequency part that is not resolved in LES, the filtered Lighthill tensor
computed from filtered LES variables, and a subgrid-scale tensor. They performed DNS
and LES of decaying isotropic turbulence to determine the three parts of the Lighthill
tensor for both approaches and evaluated the subgrid and high-frequency contributions
to the complete Lighthill tensor. This work showed that the high-frequency part of the
Lighthill tensor does not contribute significantly to noise production if cutoff wave
numbers of the usual values are employed, but this work also indicated that the subgridscale contribution cannot be neglected.
With a similar objective Seror and Sagaut (2000) and Seror et al. (2001) computed
forced isotropic turbulence by means of DNS and LES together with a noise prediction
based on Lighthills analogy. DNS was performed on a grid with 1923 points, whereas
large-eddy simulations with the structure function model were conducted on two grids
with 323 and 643 points, respectively. The authors investigated whether noise radiation
generated in forced isotropic turbulence can be estimated using a hybrid LESLighthill
analogy approach. They confirmed that the high-frequency part of Lighthills tensor does
not contribute significantly to the overall noise production and showed that the subgridscale intensity and the subgrid-scale fluctuating pressure can not be neglected. Again
17
23:37
P1: JzG
0521871441c01
18
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
they addressed the parametrization of subgrid-scale effects based on a scale-similarity

model and demonstrated the efficiency of the latter in a priori and a posteriori tests.
1.3.2.2 Jet flows
The prediction of high-speed jet noise is one important problem that largely benefits from recent developments in LES. Jet noise is an important component of noise
emission from civilian aircraft. For engines having low-to-medium bypass ratios, jet
noise reduction can be achieved by internal mixing of the core and the bypass streams;
however, for the high-bypass streams of engines having high bypass ratios, the noise
reduction benefit from internal mixing is small. Therefore, jet noise is the most prominent source at full power takeoff conditions. To design nozzles for which a significant
jet noise reduction can be achieved for high bypass ratios, many research groups have
conducted applied and fundamental research based on numerical simulations of the
turbulent flow field and the radiated noise. In the initial work on jet noise, Mankbadi
et al. (1994, 1998) and Mitchell et al. (1997) concentrated on supersonic jet noise.
Mankbadi et al. (1994) and Morris, Wang et al. (1997) performed large-eddy simulations of supersonic jet flow and noise using the Smagorinsky subgrid-scale eddy viscosity model. Mankbadi et al. (1994, 1998) restricted themselves to two-dimensional
simulations, and Morris, Wang et al. (1997) conducted three-dimensional computations
for the case of a Mach 2.0 jet and presented numerical results in qualitative agreement
with measurements. Kolbe et al. (1996) and Chyczewski and Long (1996) carried out
LES of rectangular jets, but they relied on numerical and artificial damping to dissipate
the turbulence energy.
Mitchell, Lele, and Moin (1999) performed direct numerical simulations for both
the near-field flow and the far-field sound radiated from subsonic and supersonic twodimensional, axisymmetric jets. The predicted acoustic far field was found to agree
with predictions from Lighthills acoustic analogy. Freund (1999) performed threedimensional DNS of a randomly forced round jet at Re = 3.6 103 and M = 0.9
using 25 million grid points. Predicting the far-field acoustic pressure by solving a
wave equation within the near-field pressure data from the DNS, they also obtained
good agreement with experimental data.
More work on jet flow and its noise generation was performed by Boersma and Lele
(1999). They conducted three-dimensional LES of a round jet under the same conditions as Freund (1999) using a dynamic Smagorinsky model, but no sound radiation
predictions were reported. Additionally, Mankbadi, Shih et al. (2000) studied supersonic jet noise using two-dimensional LES with a nondynamic Smagorinsky model,
but the computational domain did not include part of the acoustic far field. Bogey,
Bailly, and Juve (2000a) conducted LES of a jet for a Reynolds number Re = 6.5 104
and Mach number M = 0.9 with the standard Smagorinsky model. On the basis of the
obtained unsteady results they directly computed the aerodynamic noise. The mean
flow and turbulence intensities, as well as sound radiation directivity and sound levels,
were in good agreement with experimental data.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Direct numerical simulations of the near field of a two-dimensional, axisymmetric,

supersonic hot jet were performed by Gamet and Estivalezes (1998) for Reynolds
number Re = 1 103 and Mach number M = 2. They also conducted a LES on the
near field of a three-dimensional round jet at the same Mach number but for Re =
3 104 . Kirchhoffs (1877) method has been used to predict the directivity of the
radiated sound field. The comparison of the results with experimental data from the
literature showed that three-dimensional LES predictions yield better agreement than
two-dimensional axisymmetric DNS predictions.
The more complex case of simulating the near-field flow and far-field sound radiation
of subsonic jets is the subject of Zhao, Frankel, and Mongeau (2001a). They performed
LES for a Reynolds number, based on the jet nozzle diameter, of Re = 5 103 with
a dynamic Smagorinsky and a dynamic mixed subgrid-scale turbulence model. LES
predictions obtained with these two turbulence models were similar, although the mixed
model resulted in higher turbulence and sound levels. Both the near-field velocity
statistics and the far-field sound directivity, as determined with Kirchhoff s (1877)
method, were found to be in good agreement with experimental data and DNS results
from the literature.
1.3.2.3 Flow around airfoils and cylinders
Wang and Moin (2000) performed LES computations of the flow past an asymmetrically
beveled trailing edge of a flat strut at a chord Reynolds number of Re = 2.15 106 .
Their computed mean and fluctuating velocity profiles compare reasonably well with
experimental measurements. The far-field acoustic calculation was facilitated by an
integral-form solution to the Lighthill equation according to Ffowcs Williams and
Hawkings (1969). The authors concluded that, for an accurate predicted noise radiation
using the partial source field included in the LES domain, the size of the spanwise
domain must be larger than the coherence length of the source field in this direction.
The incompressible flow around the blunt trailing edge of a thick plate was investigated by Manoha, Troff, and Sagaut (2000), who performed LES using a gradient
diffusion eddy viscosity model. The simulations were conducted for a Reynolds number
based on the far-field flow velocity and the plate thickness H of Re H = 1000 on a grid
with 161 points in streamwise, 51 points in spanwise, and 201 points in lateral directions, respectively. To determine the far-field noise, they applied the theory of Ffowcs
Williams and Hawkings (1969). Although they compared their results with experimental data obtained for a different Reynolds number, they reported good agreement of
computed and measured radiated acoustic fields.
The unsteady incompressible flow around an blunt trailing edge has also been the
subject of LES of the turbulent flow around a NACA0012 airfoil for a Reynolds number based on the free-stream velocity and the plate thickness of Re = 1000 by
Troff, Manoha, and Sagaut (1997). They solved the filtered NavierStokes equation
using an eddy viscosity subgrid-scale model on a nonstaggered grid with a secondorder accurate hybrid finite differencefinite volume method. Comparing wall-pressure
19
23:37
P1: JzG
0521871441c01
20
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
fluctuations and the far-field noise applying Ffowcs WilliamsHawkings theory, they
observed good agreement with airfoil noise measurements.
More work on trailing-edge flow was reported by Schroder et al. (2001), who conducted LES of a trailing-edge flow in conjunction with noise prediction based on
solutions of the linearized Eulers equation.
The unsteady flow around a high-lift configuration was calculated via LES by
Manoha, Delahay, Redonnet et al. (2001) using the selective mixed-scale model. The
integral methods for the far-field noise prediction of Kirchhoff (1877) and Ffowcs
Williams and Hawkings (1969) were used.
In Japan, first attempts to apply LES for the prediction of aerodynamic sound were
made by Kobayashi and Satake (1991) and Kato et al. (1991) for a flat plate and a
circular cylinder, respectively. Kobayashi and Satake (1991) numerically predicted the
aerodynamic sound generated by a flat plate in a low-Mach-number turbulent flow. The
flow field was first calculated by LES on staggered grids by assuming incompressibility
and making use of a wall model and the Smagorinsky subgrid-scale model. A total of
195,000 grid points were used to simulate the flow for a Reynolds number based
on the length of the flat plate and the average flow speed of Re = 89,290. Using the
pressure fluctuations obtained by LES, the generated aerodynamic sound was predicted
by Lighthills equation.
Kato et al. (1991) computed the aerodynamic sound radiated from the low-Machnumber turbulent wake of a circular cylinder for a Reynolds number based on the
cylinder diameter of Re = 10,000. The unsteady flow field around the circular cylinder was solved by LES using an incompressible streamline upwind finite-element
method. The sound pressure was computed by LighthillCurles equation using the
fluctuating surface pressure obtained from LES. The numerical results were compared
with the measured data obtained in a low-noise wind tunnel. Lift and drag coefficients
as well as vortex shedding frequency agreed well with experimental measurements.
Simulation results were compared against velocity fluctuation spectra obtained by hotwire measurements. The computational frequency spectra reached an inertial 5/3
range slope. In the low-frequency domain, good agreement with experimental measurements was observed. However, in the high-frequency domain, the power levels
were underestimated in the simulation most likely because of insufficient grid resolution. Further, the sound-pressure levels were overestimated by the LES simulation in
the high-frequency domain. This was attributed to the fact that the span of the simulated
cylinder was 2 diameters whereas the experimental cylinder had a span of 50 cylinder
diameters.
Miyake, Bando, and Hori (1993) performed a numerical sound source analysis for
aerodynamic sound generated by uniform flow over a square cylinder at a Reynolds
number of Re = 10,000. The incompressible, time-dependent turbulent flow was first
calculated by LES, and the sound generation in the far field was then obtained by a
surface integral on the square cylinder using Curles equation modified for low-Machnumber flows. The sound-source intensity maps on the square cylinder thus obtained
could be used to detect dominant noise origins.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Figure 1.3. Insulator of a high-speed trains pantograph.
1.3.2.4 Internal flows
Coupling LES with a linearized Euler equation, Crouzet et al. (2002) investigated
the flow and the sound radiation in a duct with a diaphragm. They performed two
computations for different upstream velocities and project source terms which were
calculated on the LES grid and on the acoustic grid. The results are in good agreement
with experimental ones.
A correction of the Lighthill tensor by adding a subgrid-scale tensor based on
properties simulated in LES of channel flow was proposed by Piomelli, Streett, and
Sarkar (1997).
1.3.2.5 Applied flow problems
The aerodynamic and acoustic behavior of an exhaust diffuser was investigated by

Jayatunga et al. (2001) by means of LES and the Ffowcs Williams and Hawkings
version of the acoustic analogy.
Considerable research on LES and acoustics has been also carried out in Japan
especially with respect to practical applications of LES to acoustics for very complex
geometries at high Reynolds numbers.
In this respect Kato et al. (2000) used LES to analyze aerodynamic noise emitted
from the insulator of a high-speed trains pantograph (see Figures 1.3 and 1.4) and to
identify the sound-generating vortical structures with the aim of finding noise-reducing
concepts. The insulator is composed of a main circular cylinder and several circular
disks. Its unique geometry is responsible for the occurrence of large separation regions
21
23:37
P1: JzG
0521871441c01
22
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
INTRODUCTION
Figure 1.4. Sketch of the insulator of a high-speed

trains pantograph.
and thus aerodynamic noise, preventing a further speedup of high-speed trains. The authors simulated the turbulent near wake and resulting far-field sound by decomposing
the flow field and the resulting acoustic field with low-Mach-number flow assumption.
The Reynolds number based on the uniform wind velocity and the representative diameter of the insulator was 140,000. They solved the unsteady flow field using LES
with the Smagorinsky model. The complex shape was treated using overset grids to
allow enhanced grid resolution in the vicinity of the wall. The resulting far-field sound
pressure was calculated by LighthillCurles equation using the surface-pressure fluctuations obtained by LES. As seen in Figure 1.5, the predicted sound-pressure levels
quantitatively agree up to around 2.5 kHz with the ones measured in a low-noise-level
wind tunnel for the fine mesh using 6 million elements.
Figure 1.5. Far-field sound-pressure spectra. From Kato et al. (2000); reprinted by permission of the
American Institute of Aeronatics and Astronautics, Inc.
23:37
P1: JzG
0521871441c01
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
The authors found that the predicted sound-pressure levels were strongly affected
by the grid resolution near the object. They investigated the origin of the sound sources
and found that the longitudinal vortices generated by the circular disks are primarily
responsible for the generation of sound from this flow. Through the same numerical
methods, but with moving overset grids, Kato et al. (2002) performed a numerical
simulation of the aerodynamic noise radiated from an engine-cooling propeller fan
using a total of 4 million finite elements. The far-field sound was predicted by feeding
the surface pressure fluctuations obtained by LES into the Ffowcs WilliamsHawkings
equation. The computed static pressure rise of the fan quantitatively agreed with the
experimentally measured values, indicating that the flow around the fan blades was
accurately computed by LES. The predicted trend of the overall sound-pressure levels
was found to be consistent with those measured for a propeller fan in general.
23
23:37
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Theoretical Background: Aeroacoustics

Avraham Hirschberg and Sjoerd Rienstra
2.1 Introduction to aeroacoustics

Owing to the nonlinearity of the governing equations it is very difficult to predict the
sound production of fluid flows. This sound production occurs typically at high-speed
flows, for which nonlinear inertial terms in the equation of motion are much larger than
the viscous terms (high Reynolds numbers). Because sound production represents only
a minute fraction of the energy in the flow, the direct prediction of sound generation is
very difficult. This is particularly dramatic in free space and at low subsonic speeds.
The fact that the sound field is in some sense a small perturbation of the flow can,
however, be used to obtain approximate solutions.
Aeroacoustics provides such approximations and at the same time a definition of the
acoustical field as an extrapolation of an ideal reference flow. The difference between
the actual flow and the reference flow is identified as a source of sound. This idea was
introduced by Lighthill (1952, 1954), who called this an analogy. A second key idea of
Lighthills (1954) is the use of integral equations as a formal solution. The sound field
is obtained as a convolution of the Greens function and the sound source. The Greens
function is the linear response of the reference flow, used to define the acoustical field, to
an impulsive point source. A great advantage of this formulation is that random errors
in the sound source are averaged out by the integration. Because the source also depends
on the sound field, this expression is not yet a solution of the problem. However, under
free-field conditions one can often neglect this feedback from the acoustical field to the
flow. In that case the integral formulation provides a solution.
When the flow is confined, the acoustical energy can accumulate into resonant
modes. Because the acoustical particle displacement velocity can attain the same order
of magnitude as the main flow velocity, the feedback from the acoustical field to the
sound sources can be very significant. This leads to self-sustained oscillations, which
we call whistling. Despite the back reaction, the ideas of the analogy appear to remain
useful.
Because linear acoustics is used to determine a suitable Greens function, it is important to obtain basic insight into properties of elementary wave equation solutions.
24
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
2.2 FLUID DYNAMICS
We will focus here on the wave equation describing the propagation of pressure perturbations in a uniform stagnant (quiescent) fluid.
Although in acoustics of quiescent media it is rather indifferent whether we consider
a wave equation for the pressure or the density, we will see that in aeroacoustics the
choice of a different variable corresponds to a different choice of the reference flow
and hence to another analogy. It seems paradoxical that analogies are not equivalent
inasmuch as they are all reformulations of the basic equations of fluid dynamics. The
reason is that the analogy is used as an approximation. Such an approximation is
based on some intuition and usually empirical observations. An example of such an
approximation was already mentioned above. In free-field conditions we often neglect
the influence of the acoustical feedback on the sound sources.
Even though Lighthills analogy is very general and useful for order-of-magnitude
estimates, it is less convenient when used to predict sound production by numerical
simulations. One of the problems is that the sound source deduced from Lighthills
analogy is spatially rather extended, leading to slowly converging integrals. For lowMach-number isothermal flow we will see that aerodynamic sound production is entirely
due to mean flow velocity fluctuations, which may be described directly in terms of the
underlying vortex dynamics. This is more convenient because vorticity is in general limited to a much smaller region in space than the corresponding velocity field (Lighthills
sound sources). This leads to the idea of using an irrotational flow as reference flow.
The result is called vortex sound theory. Vortex sound theory is not only numerically
efficient but also allows us to translate the very efficient vortex-dynamical description
of elementary flows directly into sound production properties of these flows.
We present here only a short summary of the elements of acoustics and aeroacoustics.
The structure of this chapter is inspired by the books of Dowling and Ffowcs Williams
(1983) and Crighton et al. (1992). A more advanced discussion is provided in textbooks
by Pierce (1981), Temkin (2001), Morse and Ingard (1968), Goldstein (1976), Blake
(1986), Crighton et al. (1992), Hubbard (1995), Howe (1998, 2002), and Rienstra
and Hirschberg (2001). The influence of wall vibration is discussed in, among others,
Cremer and Heckl (1988), Junger and Feit (1986), and Norton (1989).
In the sections of this chapter we will consider the following:
r Some fluid dynamics (Section 2.2),
r Free-space acoustics (Section 2.3),
r Aeroacoustic analogies (Section 2.4), and
r Aeroacoustics of confined flows (Section 2.5),
2.2 Fluid dynamics

2.2.1 Mass, momentum, and energy equations
We consider the motion of fluids in the continuum approximation. This means that
quantities such as the velocity v and the density are smooth functions of space and
25
14:12
P1: JZZ
0521871441c02
26
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
THEORETICAL BACKGROUND: AEROACOUSTICS
time coordinates (x, t) as considered in Prandtl and Tietjens (1957), Batchelor (1976),
Landau and Lifshitz (1987), Schlichting (1968), Kundu (1990), Paterson (1983), and
Durand (1963). We consider the fundamental equations of mass, momentum, and energy
applied to an infinitesimally small fluid particle of volume V . We call this a material
element. We define the density of the material element equal to , and the mass is
therefore simply V . Because the mass is conserved, as denoted by
d(V ) = dV + V d = 0,
(2.1)
the rate of change of the density observed while moving with the fluid velocity v is
equal to minus the dilatation rate
1 DV
1 D
=
= v,
(2.2)
Dt
V Dt
where denotes a symbolic vector of partial derivatives used to describe gradient,
divergence, curl, and convective derivative. In Cartesian coordinates it is given by
= ( x1 , x2 , x3 ). The Lagrangian time derivative D/Dt is related to the Eulerian
time derivative /t by
D
=
+ (v ).
(2.3)
Dt
t
For a Cartesian coordinate system x = (x1 , x2 , x3 ), we can write this in the index notation as follows:
,
=
+ vi
(2.4)
Dt
t
xi
where vi
= v1
+ v2
+ v3
.
xi
x1
x2
x3
According to the convention of Einstein, the repetition of the index i implies a summation over this dead index. Substitution of definition (2.4) into Equation (2.3) yields the
mass conservation law applied to a fixed infinitesimal volume element:
vi
+ (v) = 0, or
+
= 0.
t
t
xi
(2.5)
We call this the conservation form of the mass equation. For convenience one can
introduce a mass source term Q m in this equation:
vi
= Qm .
+
t
xi
(2.6)
In a nonrelativistic approximation such a mass source term is of course zero and is only
introduced to represent the influence on the flow of a complex phenomenon (such as
combustion) within the framework of a model that ignores the details of this process.
Therefore, there is some ambiguity in the definition of Q m . We should actually specify
whether the injected mass has momentum and whether it has a different thermodynamic
state than the surrounding fluid.
In agreement with the nonrelativistic approximation, we apply the second law of
Newton to a fluid particle:
Dv
= P + f,
Dt
(2.7)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
27
2.2 FLUID DYNAMICS
where f is the density of the force field acting on the bulk of the fluid, and P is
the net force acting on the surface of the infinitesimal volume element. This force is
expressed in terms of a stress tensor P. Using the mass conservation law (2.5) without
mass source term (Q m = 0), we obtain the momentum equation in conservation form:
Pi j
vi v j
v
vi
=
+ fi .
+ (P + vv) = f, or
+
t
t
x j
x j
(2.8)
The isotropic part pi j of this tensor corresponds to the effect of the hydrodynamic
pressure p = Pii /3:
Pi j = pi j i j ,
(2.9)
where i j = 0 for i = j and i j = 1 for i = j. The deviation i j from the hydrostatic

behavior corresponds in a simple fluid to the effect of viscosity. We define a simple
fluid as a fluid for which i j is symmetrical (Batchelor 1976).
The energy equation applied to a material element is

D
e + 12 v 2 = q (P v) + f v + Q w ,
Dt
(2.10)
where e is the internal energy per unit of mass, v = v, q the heat flux, and Q w the
heat production per unit of volume. In conservation form this equation becomes

e + 12 v 2 + v(e + 12 v 2 ) = q (P v) + f v + Q w ,
t
(2.11a)
or, in index notation,

Pi j v j

qi

+ f i vi + Q w .
e + 12 v 2 +
vi e + 12 v 2 =
t
xi
xi
xi
(2.11b)
The mass, momentum, and energy conservation laws in differential form are only
valid when the derivatives of the flow variables are defined. When those laws are applied
to a finite volume V , one obtains integral formulations that are also valid in the presence
of discontinuities such as shock waves. For an arbitrary volume V enclosed by a surface
S with outer normal n, we have

d
dV + (v b) n dS = 0,
(2.12a)
dt V
S

d
v dV + v(v b) n dS = P n dS +
f dV
(2.12b)
dt V
S
S
V

d
1 2
e + 2 v dV + e + 12 v 2 (v b) n dS
dt V
S

f v dV,
(2.12c)
= q n dS (P v) n dS +
S
where b is the velocity of the control surface S. For a material control volume we have
vn = bn. For a fixed control volume we have b = 0.
14:12
P1: JZZ
0521871441c02
28
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
2.2.2 Constitutive equations
The mass, momentum, and energy equations (2.5), (2.8), and (2.11) involve many more
unknowns than equations. The additional information needed to obtain a complete set
of equations is provided by empirical information in the form of constitutive equations. An excellent approximation is obtained by assuming the fluid to be locally in
thermodynamic equilibrium that is, within a material element (van Kuiken 1995).
This implies, for a homogeneous fluid, that two intrinsic state variables fully determine
the state of the fluid. For acoustics it is convenient to choose the density of mass
and the specific entropy (i.e., per unit of mass) s as variables. All other intrinsic state
variables are a function of and s. Hence, the specific energy e is completely defined
by a relation
e = e(, s).
(2.13)
This is what we call a thermal equation of state. This equation is determined empirically.
Variations of e may therefore be written as

e
e
d +
ds.
(2.14)
de =
s
s
Comparison with the fundamental equation of thermodynamics,
de = T ds pd 1 ,
provides thermodynamic equations for the temperature T and the pressure p:

e
T =
,
s
and

p=
(2.15)
(2.16)

.
(2.17)
Because p is also a function of and s, we have

p
p
d +
ds.
dp =
s
s
(2.18)
Because sound is defined as isentropic (ds = 0) pressure-density perturbations, the

speed of sound c = c(, s) is defined by

p
.
(2.19)
c=
s
An extensive discussion of the speed of sound in air and water is provided by Pierce
(1981). In many applications the fluid considered is air at ambient pressure and temperature. Under such conditions we can assume the ideal gas law to be valid:
p = RT,
(2.20)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
29
2.2 FLUID DYNAMICS
where R is the specific gas constant, which is the ratio R = kB /m w of the constant of
Boltzmann, kB , and of the mass of a molecule, m w . By definition, for such an ideal gas
the energy density only depends on T , e = e(T ), and we have
p
(2.21)
c = = RT ,
where = cp /cv is the Poisson ratio of the specific (i.e., per unit of mass) heat capacities
at, respectively, constant volume,

e
,
(2.22)
cv =
T
and constant pressure,

cp =
i
T

,
(2.23)
where i is the enthalpy per unit of mass defined by

i =e+
p
.
(2.24)
For an ideal gas, we have cp cv = R. An ideal gas with constant specific heats is
called a perfect gas.
As we consider local thermodynamic equilibrium, it is reasonable (Schlichting
1968; van Kuiken 1995) to assume that transport processes are determined by linear
functions of the gradients of the flow-state variables. This corresponds to a Newtonian
fluid behavior

(2.25)
i j = 2 Di j 13 Dkk i j + v Dkk i j ,
where the rate-of-strain tensor Di j is defined by

v j
1 vi
.
+
Di j =
2 x j
xi
(2.26)
Note that Dkk = v takes into account the effect of dilatation. In thermodynamic
equilibrium, according to the hypothesis of Stokes, one assumes that the bulk viscosity
v vanishes. The dynamic viscosity is a function of the thermodynamic state of
the fluid. For an ideal gas, is a function of the temperature only. Although the assumption of vanishing bulk viscosity v is initially an excellent approximation, one
observes significant effects of the bulk viscosity in acoustical applications such as
propagation over large distances (Pierce 1981). This deviation from local thermodynamic equilibrium is due, in air, to the finite relaxation time of rotational degrees of
freedom of molecules. The corresponding approximation for the heat flux q is the law
of Fourier,
qi = K
T
,
xi
(2.27)
14:12
P1: JZZ
0521871441c02
30
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where K is the heat conductivity. For an ideal gas, K is a function of the temperature
only. It is convenient to introduce the kinematic viscosity and the heat diffusivity a:
=
(2.28)
and
a=
K
.
cp
(2.29)
The kinematic viscosity and the heat diffusivity are diffusion coefficients for momentum
and heat transfer, respectively. For an ideal gas, both transfer processes are determined
by the same molecular velocities and similar free molecular path. This explains why the
Prandtl number Pr = /a is of order unity. For air at ambient pressure and temperature,
Pr = 0.72.
2.2.3 Approximations and alternative forms of the basic equations
Starting from the energy equation (2.11) and using the thermodynamic law (2.15), one
can derive an equation for the entropy:
T
Ds
= q + :v + Q w ,
Dt
(2.30)
where the scalar product : of two tensors (or the double contraction of their product)
is defined in Cartesian coordinates by
: =
3
3

i j ji .
i=1 j=1
If heat transfer and viscous dissipation are negligible and there are no heat sources, the
entropy equation reduces to
Ds
= 0.
Dt
(2.31)
Hence, the entropy of a material element remains constant and the flow is isentropic.
When the entropy is uniform, we call the flow homentropic, and thus s = 0. An
isentropic flow originating from a reservoir with uniform state is homentropic.
When there is no source of entropy, the sound generation is dominated by the fluctuations of the Reynolds stress vi v j . Therefore, sound generation often corresponds
to conditions for which the term |vi v j / x j | in the momentum equation (2.8) is large
compared with |i j / x j |. Assuming that both gradients scale with the same length D
while the velocity scales with U0 (a main flow velocity), we find Re = U0 D/ 1,
In this chapter, the tensor vi v j includes the linear and the steady component of the full Reynolds
stress and not only the turbulent component as commonly used in the computational fluid dynamics
community.
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
31
2.2 FLUID DYNAMICS
where Re is the Reynolds number. In such a case one can also show that the dissipation
is limited to thin boundary layers near the wall and that, for time scales of the order of
U0 /D, the bulk of the flow can be considered as isentropic. Note that the demonstration of this statement in aeroacoustics has been the subject of research for a long time
(Morfey 1976, 2003; Obermeier 1985; Verzicco et al. 1997). It is not a trivial statement. Actually, a turbulent flow is essentially dissipative. On the time scales relevant to
sound production dissipation is negligible outside the viscous boundary layers at walls
(Morfey 2003). Similarly, we often assume that heat transfer is limited to thin boundary
layers at the wall and that the bulk of the flow is essentially isothermal. We will see in
Section 2.4, that, when the entropy of the flow is not uniform, the convection of those
inhomogeneities is an important source of sound. We have now discussed the problem
of dissipation and heat transfer in the source region. We will later consider the effect
of friction and heat transfer on wave propagation.
In a frictionless flow, the momentum equation (2.8) reduces to the equation of Euler
as follows:
Dv
= p + f.
Dt
(2.32)
Using the definition of enthalpy (2.24), i = e + p/, combined with the fundamental
equation (2.15), T ds = de + pd(1/), we find
Dv
f
= i + T s + .
Dt
(2.33)
The acceleration Dv/Dt can be split up into an effect of the time dependence of the
flow, v/t, an acceleration in the direction of the streamlines, ( 12 v 2 ), and a Coriolis
acceleration due to the rotation, = v, of the fluid as follows:

v
Dv
=
+ 12 v 2 + v.
Dt
t
(2.34)
Substitution of Equations (2.34) and (2.33) in Eulers equation (2.32) yields

v
f
+ B = v + T s + ,
t
(2.35)
where the total enthalpy or Bernoulli constant B is defined by

B = i + 12 v 2 .
(2.36)
In general, the flow velocity field v can be expressed in terms of a scalar potential
and a vector stream function :
v = + .
(2.37)
There is an ambiguity in this definition that may be removed by some additional condition. One can, for example, impose = 0. In most of the problems considered,
14:12
P1: JZZ
0521871441c02
32
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
the ambiguity is removed by boundary conditions imposed on and . Although the

scalar potential is related to the dilatation rate
v = 2 ,
(2.38)
because () = 0, the vector stream function is related to the vorticity
= v = ()
(2.39)
because = 0. This will be used as an argument to introduce the unsteady

component of the potential velocity field as a definition for the acoustical field within
the framework of vortex sound theory (see Section 2.4.5).
For a homentropic (s = 0) potential flow (v = ) without external forces (f =
0), the momentum equation (2.32) can be integrated to obtain the equation of Bernoulli:
+ B = g(t),
t
(2.40)
where the function g(t) can be absorbed into the definition of the potential without
any loss of generality.
In the vortex sound theory (Section 2.4.5), where the reference flow (acoustic field)
corresponds to the unsteady component of the potential flow , the source is the
difference between the actual flow and the potential flow. Therefore, the sources are
directly related to the vorticity . In a homentropic flow the density is a function
= ( p) of the pressure p only (barotropic fluid). In such a case we can eliminate the
pressure from the equation of Euler by taking the curl of this equation (2.32). We obtain
an equation for the vorticity (Saffman 1992) as follows:

f
D
= v v +
.
(2.41)
Dt
In the absence of external forces, the equation reduces to a purely kinematic equation.
Solving this equation yields, with = v, the velocity field. This approach is most
effective for two-dimensional plane flows v = (v1 (x1 , x2 ), v2 (x1 , x2 ), 0). In that case
the vorticity equation reduces to D3 /Dt = 0. The study of such flows provides much
insight into the behavior of vorticity near sharp edges.
Mathematically, the assumed absent viscosity yields a set of equations and boundary
conditions that have no unique solution. By adding the empirically observed condition
that no vorticity is produced anywhere, we have again a unique solution. This, however,
is not exactly true near sharp edges. As determined by the Reynolds number and
the (dimensionless) frequency and amplitude, a certain amount of vorticity is shed
from a sharp edge. For high enough Reynolds number and low enough frequency and
amplitude, the amount of shed vorticity is just enough to remove the singularity of
the potential flow around the edge. This is the so-called Kutta condition (Prandtl and
Tietjens 1957; Landau and Lifshitz 1987; Rienstra 1981a; Paterson 1983; Crighton
1985).
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
33
2.3 FREE-SPACE ACOUSTICS OF A QUIESCENT FLUID
When the flow is nearly incompressible (such as in acoustical waves), we can approximate the enthalpy by

p
dp
.
(2.42)
i=
0
Under these circumstances, the equation of Bernoulli (2.40) reduces to
p
+ 12 v 2 +
= 0.
t
0
(2.43)
When considering acoustical waves propagating in a uniform stagnant medium, we

may neglect the quadratically small term 12 v 2 , which yields the linearized equation of
Bernoulli:
p
+
= 0.
t
0
(2.44)
2.3 Free-space acoustics of a quiescent fluid

2.3.1 Orders of magnitude
In acoustics one considers small perturbations of a flow. This will allow us to linearize
the conservation laws and constitutive equations described in the previous section (2.2).
We will focus here on acoustic perturbations of a uniform stagnant (quiescent) fluid.
For that particular case we will now discuss orders of magnitude of various effects.
This will justify the approximations that we use further on.
We will focus on the pressure perturbations p that propagate as waves and that can
be detected by the human ear. For harmonic pressure fluctuations, the audio range is
20 Hz f 20 kHz.
The sound-pressure level (SPL) measured in decibels (dB) is defined by

prms
,
SPL = 20 log10
pref
(2.45)
(2.46)
where pref = 2 105 Pa for sound propagating in gases and pref = 106 Pa for propagation in other media. The sound intensity I
= In
is defined as the time-averaged
energy flux associated to the acoustic wave, propagating in direction n. The intensity
level (IL) measured in decibels (dB) is given by

I
,
(2.47)
IL = 10 log10
Iref
where in air Iref = 1012 Wm2 . The reference intensity level Iref is related to the
reference pressure pref by the relationship valid for propagating plane waves,
I
=
p 2
,
0 c0
(2.48)
14:12
P1: JZZ
0521871441c02
34
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
because in air at ambient conditions 0 c0 400 kg m2 s1 . The time-averaged power

P
generated by a sound source is the flux integral of the intensity I
over a surface
enclosing the source. The sound power level (PWL) measured in decibels (dB) is defined
by

P
,
(2.49)
PWL = 10 log10
Pref
where Pref = 1012 W corresponds to the power flowing through a surface of 1 m2
surface area with an intensity I
= Iref .
The threshold of hearing (for good ears) at 1 kHz is typically around SPL = 0 dB.
This corresponds physically to the thermal fluctuations in the flux of molecules colliding
with our eardrum. In order to detect 1 kHz, we can at most integrate the signal over
about 0.5 ms. At ambient conditions, this corresponds to the collision of N 1020
molecules with our eardrum.The thermal fluctuations in the measured pressure are
therefore of the order of p0 / N = 105 Pa, with p0 the atmospheric pressure. The
maximum sensitivity of the ear is around 3 kHz (pitch of a policemans whistle), which
is due to the quarter-wave-length resonance of our outer ear a channel of about 2.5-cm
depth. The threshold of pain is around SPL = 140 dB. Even at such high levels we have
pressure fluctuations only of the order p / p0 = O(103 ). The corresponding density
fluctuations are
p

=
,
(2.50)
0
0 c02
which are also of the order of 103 because in air 0 c02 / p0 = = c p /cv 1.4. This
justifies the linearization of the equations. Note that in a liquid the condition for linearization /0 1 does not imply a small value of the pressure fluctuations because
p / p0 = (0 c02 / p0 )( /0 ) whereas 0 c02 p0 . In water 0 c02 = 2 109 Pa. We should
note that, when considering wave propagation over large distances, nonlinear wave
steepening will play a significant role. In a pipe this can easily result in the formation of
shock waves. This explains the occurrence of brassy sound in trombones at fortissimo
levels (Hirschberg 1995; Hirschberg et al. 1996). Also, in sound generated by aircraft,
nonlinear wave distortion significantly contributes to the spectral distribution (Crighton
et al. 1992).
For a propagating acoustic plane wave the pressure fluctuations p are associated to
the velocity u of fluid particles in the direction of propagation. We will see later that
u =
p
.
0 c0
(2.51)
The amplitude of the fluid particle displacement is for a harmonic wave with circular frequency given by = |u |/. At f = 1 kHz, the threshold of hearing (0 dB)
corresponds to = 1011 m. At the threshold of pain we find = 104 m. Such small
displacements also justify the use of a linearized theory. When the acoustical displacement attains the same order of magnitude as the radius of curvature of a wall, one
will observe acoustical flow separation and the formation of vortices. In a pipe, when
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
35
approaches the pipe cross-sectional radius, one will observe acoustical streaming. At
the pipe outlet this will result in periodic vortex shedding (Ingard and Labate 1950;
Ingard and Ising 1967; Disselhorst and van Wijngaarden 1980; Peters and Hirschberg
1993; Peters et al. 1993; Dalmont et al. 2002). In woodwind musical instruments and
bass-reflex ports of loudspeaker boxes this is a common phenomenon (Hirschberg 1995;
Roozen et al. 1998; Dalmont et al. 2002).
When deriving a wave equation in the next section, we will not only linearize the
basic equations but will also neglect friction and heat transfer. This corresponds to the
assumption that, in an acoustical wave with wavelength = c/ f , the unsteady Reynolds
number

| u
|
2 f
t
(2.52)
=O
Re unst =
2
| xu2 |
is very large. For air = 1.5 105 m2 s1 , and thus for f = 1 kHz we find Re unst =
O(107 ). We therefore expect that viscosity only plays a role on very large distances.
Because the Prandtl number is of order unity Pr = O(1) in a gas, we also expect heat
transfer to be negligible. At high frequencies, however, we observe a much stronger
attenuation due to nonequilibrium effects (bulk viscosity). This results in a strong
absorption of these high frequencies when we listen to aircraft at large distances. Furthermore, in the presence of walls, viscothermal dissipation will also be much larger.
The amplitude of a plane wave traveling along a tube of cross-sectional radius R will
attenuate exponentially exp(x) with the distance x. The attenuation coefficient is
given for typical audio conditions by Pierce (1981) and Scheichl (2004):

f
1
.
(2.53)
1+
=
Rc0
Pr
In most woodwind musical instruments at low pitches the viscothermal dissipation
losses are larger than the sound radiation power (Fletcher and Rossing 1998).
2.3.2 Wave equation and sources of sound
We consider the propagation of pressure perturbations p in an otherwise quiescent fluid.

The perturbations of the uniform constant reference state p0 , 0 , s0 , v0 are defined by
p = p p0 ,
= 0 ,
s = s s0 ,
v = v v0 ,
(2.54)
where, for a quiescent fluid v0 = 0. We assume that f, Q w and the perturbations p / p0 ,

/0 , . . . are small so that we can linearize the basic equations. Furthermore we neglect
heat transfer and viscous effects. The equations of motion (2.5, 2.8, and 2.30) reduce
to

+ 0 v = 0,
t
v
+ p = f,
t
0 T0
s
= Qw,
t
(2.55)
14:12
P1: JZZ
0521871441c02
36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
and the constitutive equation (2.19) becomes

p
s.
p = c02 +
s
(2.56)
Subtracting the divergence of the linearized momentum equation from the time derivative of the linearized mass conservation law yields
2
2 p = f.
t 2
(2.57)
Combining the entropy equation with the constitutive equation yields

2
( p/s) Q w
2 p
2
.
=
c
+
0
t 2
t 2
0 T0
t
(2.58)
Elimination of the density fluctuations from Equations (2.57) and (2.58) yields a
nonhomogeneous wave equation:
1 2 p
2 p = q,
c02 t 2
q=
(2.59)
( p/s) Q w
f.
0 c02 T0 t
The first source term corresponds to the dilatation of the fluid as a result of heat
production in processes such as unsteady combustion or condensation. This type
of sound-generation mechanism has been discussed in detail by Morfey (1973) and
Dowling and Ffowcs Williams (1983). The second term describes the sound production by a nonuniform, unsteady external force field. When one considers a moving body,
the reaction of the body to the force exerted by the fluid can be represented by such a
force field. An example of this is a model of the sound radiated by a rotor calculated by
concentrating the lift force of each wing into a point force. This model will be discussed
in Section 2.4.2 and corresponds to the first theory of propeller sound generation as
formulated by Gutin (1948) and commonly used in many applications (Brouwer 1992;
Roger 2004).
We introduced q(x, t) as shorthand notation for the source term in the wave equation.
In the absence of a source term, q = 0, the sound field is due to initial perturbations
or boundary conditions. In the next section we present a general solution of the wave
equation.
2.3.3 Greens function and integral formulation
Using Greens theorem (Morse and Feshbach 1953), we can obtain an integral equation that includes the effects of the sources, the boundary conditions, and the initial
conditions on the acoustic field. The Greens function G(x, t|y, ) is defined as the
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
37
response of the flow to an impulsive point source represented by delta functions of

space and time:
1 2G
2 G = (x y)(t ),
c02 t 2
(2.60)
where (x y) = (x1 y1 )(x2 y2 )(x3 y3 ). The delta function (t) is not a common function with a pointwise meaning but a generalized function (Crighton et al. 1992)
formally defined by its filter property:

F(t)(t) dt = F(0)
(2.61)
for any well-behaving function F(t).

The definition of the Greens function G is completed by specifying boundary conditions at a surface S with outer normal n that encloses both the source placed at position
y and the observer placed at position x. Because this is an acoustical phenomenon, we
follow Crightons (1992) suggestion to call the observer a listener. A quite general linear
boundary condition is a linear relationship between the value of the Greens function G
at the surface S and the (history of the) gradient n G at the same point. If this relationship is a property of the surface and independent of G, we describe the surface as
locally reacting. Such a boundary condition is usually expressed in the Fourier domain
in terms of an impedance Z () of the surface S that is, the ratio between pressure
and normal velocity component as follows:
G
i
Z () =
,
0
n x G
where G is the Fourier-transformed Greens function defined by

1
G(x, t|y, ) eit dt

G(x,
|y, ) =
2
and its inverse

G(x, t|y, ) =
G(x,
|y, ) e it d.
(2.62)
(2.63)
(2.64)
(Always check the sign convention in the exponential! Here, we used exp(+it).
This is not essential as long as the same convention is used throughout!) A problem
when using Fourier analysis is that the causality of the solution is not self-evident. We
need to impose restrictions on the functional dependence of Z and 1/Z on the frequency
(Rienstra 1988; Rienstra and Hirschberg 2001).
Causality implies that there is no response before the pulse (x y)(t ) has
been released, and so
G(x, t|y, ) = 0
for t < .
(2.65)
14:12
P1: JZZ
0521871441c02
38
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Consider a Greens function G not necessarily satisfying the actual boundary condition
prevailing on S and a source q not necessarily vanishing before some time t0 . For the
wave equation (2.59) we then find the formal solution
t

q(y, )G(x, t|y, )dVy d
p (x, t) =
t0
t
t0
1
c02

G(x, t|y, )y p p (y, t)y G n dSy d

G(x, t|y, )
V
G
p
p (y, )

=t0
dVy ,
(2.66)
where dVy = dy1 dy2 dy3 . The first integral is the convolution of the source q with
the pulse response G, the Greens function. The second integral represents the effect
of differences between the actual physical boundary conditions on the surface S and
the conditions applied to the Greens function. When the Greens function satisfies
the same locally reacting linear boundary conditions as the actual field, this surface
integral vanishes. In that case we say that the Greens function is tailored. The last
integral represents the contribution of the initial conditions at t0 to the acoustic field.
If q = 0 and p = 0 before some time, we can choose t0 = and leave this term
out.
Note that in the derivation of the integral equation (2.66) we have made use of the
reciprocity relation for the Greens function (Morse and Feshbach 1953):
G(x, t|y, ) = G(y, |x, t).
(2.67)
Owing to the symmetry of the wave operator considered, the acoustical response measured in x at time t of a source placed in y fired at time is equal to the response
measured in y at time of a source placed in x fired at time t. The change of sign of
the time t and t is necessary to respect causality. The reciprocity relation
will be used later to determine the low-frequency approximation of a tailored Greens
function. This method is extensively used by Howe (1998, 2002). It is a particularly
powerful method for flow near a discontinuity at a wall. In many cases, however, it is
more convenient to use a very simple Greens function such as the free-space Greens
function G 0 . We will introduce this Greens function after we have obtained some
elementary solutions of the homogeneous wave equation in free space.
2.3.4 Inverse problem and uniqueness of source
It can be shown that for given boundary conditions and sources q(x, t) the wave equation
has a unique solution (Morse and Feshbach 1953). However, different sources can
produce the same acoustical field. A good audio system is able to produce a music
performance that is just as realistic as the original. Mathematically the nonuniqueness
of the source is demonstrated by the following enlightening example of Dowling and
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
39
Ffowcs Williams (1983). Let us assume that p (x, t) is a solution of the nonhomogeneous
wave equation
1 2 p
2 p = q(x, t)
c02 t 2
(2.68)
in which q(x, t) = 0 in a limited volume V . Outside V the source vanishes, and thus,
/ V . However, p + q satisfies the
q(x, t) = 0. As a result, p + q = p for any x
equation
1 2q
1 2 ( p + q)
2
(
p
+
q)
=
q(x,
t)
+
2 q,
2
t 2
c0
c02 t 2
(2.69)
which has in general a different source term than Equation (2.68).

To determine the source from any measured acoustical field outside the source
region, we need a physical model of the source. This is typical of any inverse problem in
which the solution is not unique. When using microphone arrays to determine the sound
sources responsible for aircraft noise, one usually assumes that the sound field is built up
of so-called monopole sound sources (Sijtsma, Oerlemans, and Holthusen, 2001). We
will see later that the sound sources are more accurately described in terms of dipoles
or quadrupoles (Section 2.4.1). Under such circumstances it is hazardous to extrapolate
such a monopole model to angles outside the measuring range of the microphone array
or to the field from flow Mach numbers other than used in the experiments.
2.3.5 Elementary solutions of the wave equation
We consider two elementary solutions of the homogeneous wave equation (q = 0)

1 2 p
2 p = 0
c02 t 2
(2.70)
that will be used as building blocks to obtain more complex solutions:

r the plane wave and
r the spherical symmetric wave.
We assume in both cases that these waves have been generated by some boundary
condition or initial condition. We consider their propagation through an in-all-directions
infinitely large quiescent fluid we call free space.
We first consider plane waves. These are uniform in any plane normal to the direction
of propagation. Let us assume that the waves propagate in the x1 direction, in which
case p = p (x1 , t) and the wave equation reduces to
1 2 p
2 p
= 0.
c02 t 2
x12
(2.71)
14:12
P1: JZZ
0521871441c02
40
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
This one-dimensional wave equation has the solution of dAlembert,

x1
x1

p =F t
+G t +
,
c0
c0
(2.72)
where F represents a wave traveling in the positive x1 direction and G travels in

the opposite direction. This result is easily verified by applying the chain rule for
differentiation. The functions F and G are determined by the initial and boundary
conditions.
Consider for example the acoustic field generated by an infinite plane wall oscillating around x1 = 0 with a velocity u 0 (t) in the x1 direction. In linear approximation
v1 (0, t) = u 0 (t); that is, the acoustical velocity at x1 = 0 is assumed to be equal to the
wall velocity. It is furthermore implicitly assumed in the definition of free space that
no waves are generated at infinity. Therefore, we have for x1 > 0 that G = 0. Using the
linearized equation of motion (2.55) in the absence of external force field f = 0,
0
v1
p
= 1,
t
x1
(2.73)
p = 0 c0 v1 .
(2.74)
we find
We call 0 c0 the specific acoustical impedance of the fluid. Using the boundary condition
v1 (0, t) = u 0 (t) and p (x1 , t) = F(t x1 /c0 ), we find
p = 0 c0 u 0 (t x1 /c0 )
(2.75)
as a solution for x1 > 0. This equation states that perturbations, observed at time t at
position x1 , are generated at the wall x1 = 0 at time t x1 /c0 . The time te = t x1 /c0
is called the emission time or retarded time. In a similar way, we find
p = 0 c0 u 0 (t + x1 /c0 )
(2.76)
for x1 < 0 if the wall is of zero thickness and perturbs the fluid at either side.
By analogy of Equation (2.72), we easily find for a plane-wave solution propagating
in a direction given by the unit vector n the most general form

nx

.
(2.77)
p =F t
c0
For the particular case of harmonic waves the plane-wave solution is written in complex
notation as
p = p e itik x ,
(2.78)
where k = kn is the wave vector, k = /c0 is the wave number, and p is the amplitude.
The complex notation is a shorthand notation for
cos(t k x) ( p)
sin(t k x).
p = ( p e itik x ) = ( p)
(2.79)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
41
By means of Fourier analysis in time, an arbitrary time dependence can be represented by

a sum or integral of harmonics functions. In a similar way, general spatial distributions
can be developed in terms of plane waves.
Another important elementary solution of the homogeneous wave equation (2.70)
is the spherically symmetric wave. In that case the pressure is only a function p (r, t)
of time and the distance r to the origin.
2 F 1 2r F
r r = r r 2 , we find that the wave equaBy identifying 2 F(r ) = r12 r

tion (2.70) reduces for r > 0 to
2 pr
1 2 pr
= 0.
r 2
c02 t 2
(2.80)
Note that at r = 0 the equation is singular. As we will see this will correspond to a
possible point source. Equation (2.80) implies that the product p r of the pressure p
and the radius r satisfies the 1D wave equation and may be expressed as a solution of
dAlembert,
p =
1
[F(t r/c0 ) + G(t + r/c0 )] ,
r
(2.81)
in which F represents outgoing waves and G denotes incoming waves. In many applications we will assume that there are no incoming waves G = 0. We call these free-field
conditions. We now focus on the behavior of outgoing harmonic waves:
p =
A itikr
e
,
r
(2.82)
where A is the amplitude and k = /c0 the wave number. The radial fluid particle
velocity vr associated with the wave can be calculated by using the radial component
of the momentum equation (2.55):
0
p
vr
=
.
t
r
We find the following:

vr
p
=
0 c0
i
1
kr
(2.83)

.
(2.84)
At distances r large compared with the wavelength = 2/k (kr = 2r/ 1) we

find the same behavior as for a plane wave (2.74). The spatial variation due to the
harmonic wave motion dominates over the effect of the radial expansion. We call this
the far-field behavior. In contrast to this, we have for kr 1 the near-field behavior in
which the velocity vr is inversely proportional to the square of the distance r . This is
indeed the expected incompressible flow behavior. Over small distances the speed of
sound is effectively infinite because any perturbation arrives without delay in time. As
a result, the mass flux is conserved and vr r 2 is constant. All this can be understood by
the observation that |( 2 p /t 2 )/[c02 ( 2 p / 2r 2 )] (kr )2 , and thus the wave equation
reduces to the equation of Laplace 2 p = 0 for kr 0.
14:12
P1: JZZ
0521871441c02
42
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Outgoing spherical symmetric waves correspond to what is commonly called a

monopole sound field. Such a field can be generated by a harmonically pulsating rigid
sphere with radius a:
a = a0 + a e it .
(2.85)
0 ) we have
In linear approximation (in a/a
vr (a0 ) = ia.
(2.86)
Combining this boundary condition with Equations (2.82) and (2.84), we obtain the
amplitude A of the wave:
p =
0 2 a0 a a0 itik(r a0 )
e
.
1 + ika0 r
(2.87)
In the low-frequency limit ka0 1 we see that the amplitude of the radiated sound
field decreases with the frequency. If the volume flux V = 4a02 vr (a0 ) = 4 ia02 a
generated at the surface of the sphere is kept fixed, the sound pressure p decreases
linearly with decreasing frequency:
p =
i0 V itik(r a0 )
e
.
4r
(2.88)
A monopole field can, for example, be generated by unsteady combustion, which corresponds to the entropy source term in the wave equation. This will occur in particular for
a spherically symmetric combustion. In general the monopole field will be dominant
when the source region is small compared with the acoustic wavelength ka0 1. We
call a region that is small compared with the wavelength a compact region. We have
seen that a compact pulsating sphere is a rather inefficient source of sound under freefield conditions. More formally, a monopole source corresponds to a localized volume
source or point source placed at position y:
q(x, t) =
V
(x y).
t
(2.89)
We will discuss this approach more in detail later. Note the time derivative in the source
term of Equation (2.89). It reflects the fact that a steady flow does not produce any
sound.
Using the monopole solution (2.82), we can build more complex solutions. If p0
is a solution of the wave equation (2.70), any spatial derivatives p0 / xi are also
solutions because the wave equation has constant coefficients and the derivatives may
be interchanged. A first-order spatial derivative of the monopole field is called a dipole
field. Second-order spatial derivatives correspond to quadrupole fields.
An example of a dipole field is the acoustic field generated by a rigid sphere translating harmonically in a certain direction x1 with a velocity vs = vs e it . The radial
velocity vr (a0 , ) on the surface of the sphere is given by
vr (a0 , ) = vs cos ,
(2.90)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
43
where is the angle between the position vector on the sphere and the translation
direction x1 . Because we have the identity

r
x1
=
x12 + x22 + x32 =
= cos ,
(2.91)
x1
x1
r
we can write for the dipole field
p = A
x1
eikr
r
= A cos
r
eikr
r

.
Substitution of Equation (2.92) into the momentum equation (2.83) yields

2 eikr
i0 vr = A cos 2
.
r
r
We apply this equation at r = a0 . Comparison with Equation (2.90) yields
a 2
i0 vs a0 cos
0
eik(r a0 ) .
(1
+
ikr
)
p =
2 + 2ika0 (ka0 )2
r
(2.92)
(2.93)
(2.94)
Another example is the calculation of the field p generated by an unsteady, nonuniform force field f = ( f 1 , f 2 , f 3 ). Following Equation (2.59), we have
1 2 p
2 p = f.
c02 t 2
(2.95)
Let us assume that we have obtained a solution F1 of the wave equation in free space,
thus satisfying
1 2 F1
2 F1 = f 1 .
c02 t 2
(2.96)
Then we may find the solution p of Equation (2.95) in free space by taking the space
derivative of F1 as follows:
p =
F1
.
x1
(2.97)
This indicates that the dipole field is related to forces exerted on the flow.
Another way to deduce the relationship between forces and dipole fields is to consider the dipole as the field obtained by placing two opposite monopole sources of
amplitude V at a distance y1 from each other. Taking the limit of y1 0 while
we keep V y1 constant yields a dipole field. Because in free space changes in source
position y are equivalent to changes in listener position x, it is obvious that this limit
relates to the spatial derivative of the monopole field. If we consider now the two oscillating volume sources forming the dipole, there will be a mass flow V from one
source to the other. Such an unsteady mass flow is associated with an unsteady momentum flux. This unsteady momentum flux must, following Newton, be produced by
an external force acting on the flow (Prandtl and Tietjens 1957; Durand 1963). Hence
we see that a dipole is not possible without the action of a force. This idea is illustrated
in Figure 2.1 in which we consider waves generated by a boat on the water surface.
14:12
P1: JZZ
0521871441c02
44
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 2.1. Monopole, dipole, and quadrupole

generating waves on the surface of the water
around a boat.
When a person jumps up and down in the boat, he or she produces an unsteady
volume injection, and this generates a monopole wave field around the boat. When two
persons on the boat play with a ball, they will exert a force on the boat each time they
throw or catch the ball. Exchanging the ball results in an oscillating force on the boat.
This will make the boat translate, and this in turn generates a dipole wave field. We could
say that two individuals fighting with each other is a reasonable model for a quadrupole.
This indicates that quadrupoles are in general much less efficient in producing waves
then monopoles or dipoles. This indeed appears to be the case.
It is often stated that Lighthill (1954) has demonstrated that the sound produced
by a free, turbulent isentropic flow has the character of a quadrupole. A better way of
putting it is that, because there is no net volume injection due to entropy production
nor any external force field, in such flows the sound field can at most be a quadrupole
field (Hirschberg and Schram 1995). Therefore, Lighthills statement is actually that
we should ignore any monopole or dipole emerging from a poor description of the flow.
We will consider this in more detail in Section 2.4.1.
2.3.6 Acoustic energy and impedance
The definition of acoustical energy is not obvious when we define the acoustic field on
the basis of linearized equations. The energy is essentially quadratic in the perturbations. We may anticipate therefore that there is some arbitrariness in the definition of
acoustical energy. This problem has been the subject of many discussions in the literature (Goldstein 1976; Pierce 1981; Myers 1986a, 1986b, 1991; Landau and Lifshitz
1987; Jenvey 1989; Mohring 2001). In the particular case of the acoustics of a quiescent
fluid, the approach proposed by Kirchhoff (Landau and Lifshitz 1987), starting from
the linearized equations (2.55), appears to be equivalent to the result obtained by expanding the energy equation (2.11) up to the second order (Landau and Lifshitz 1987).
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
45
After elimination of the density by using the constitutive equation, we can write the
linearized mass conservation in the form

1 p
s
1 p
v
=
+
(2.98)
0
c02 t
c02 s t
and the momentum equation in the form
v
+ p = f.
t
(2.99)
We multiply the first equation (2.98) by p /0 and add the result to the scalar product
of the second equation (2.99) with v to obtain the acoustic energy equation
E
+ I = D,
t
(2.100)
where we have defined the acoustic energy E by

E=
1
1 p 2
2
.
0 v +
2
2 0 c02
(2.101)
The intensity I, defined as

I = p v ,
(2.102)
is identified as the flux of acoustic energy. The dissipation D is the power per unit
volume delivered by the acoustical field to the sources

p
1
s
D=
f v .
p
(2.103)
0 c02 s t
From the mass conservation law (2.98), we see that the source term
( p/s) /(0 c02 )(s /t)
in the dissipation corresponds to the dilatation rate induced by the source. This allows
us to relate the first term in the dissipation to the work of the acoustical field due to the
change in volume (dW = p dV ).
For harmonically oscillating fields p = p e it , v = v e it , the time-averaged E
of the acoustic energy is (of course) independent of time:

2/
E
=
E dt;
(2.104)
2 0
hence, the energy equation (2.100) reduces to
I
= D
.
(2.105)
By integration of this equation over a volume enclosing the sources, we find the source
power

(2.106)
P
= D
dV = I
n dS,
V
14:12
P1: JZZ
0521871441c02
46
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where n is the outer normal to the control surface S. If we assume an impedance

boundary condition on the surface S,
Z () =
p
,
v n
(2.107)
we have
I
n = 12 (Z )|v n|2 .
(2.108)
We see that the real part (Z ) of the impedance Z is associated with the transport of
acoustic energy through the surface S. The imaginary part is associated with pressure
differences induced by the inertia of the flow.
We can now easily verify, by using Equation (2.106), that the spherically symmetric
wave solution (2.82) satisfies the acoustic energy conservation law. The r 1 dependence
of the pressure (Equation (2.81)) in a simple outgoing wave results in a conserved value
of 4r 2 I
n.
To illustrate this we consider the impedance of a pulsating sphere of radius a0 . From
Equation (2.87) we find for the impedance Z of the surface of the sphere that

0 c0
i
p
Z=
=
1
+
.
(2.109)
vr
ka0
1 + (ka10 )2
The real part is given by
(Z ) = 0 c0
(ka0 )2
.
1 + (ka0 )2
(2.110)
We see that for a large sphere, ka0 1, the impedance is equal to 0 c0 , which is the
impedance experienced by a plane wave (Equation (2.74)) of any plane control surface.
For a compact sphere ka0 1, we see that (Z ) 0 c0 (ka0 )2 , which implies very
little energy transfer and thus a very inefficient sound source. The imaginary part (Z )
of the impedance of the sphere, given by
(Z ) = 0 c0
ka0
,
1 + (ka0 )2
(2.111)
vanishes for ka0 . For a compact sphere, ka0 1, it corresponds to the pressure calculated by means of the linearized equation of Bernoulli (2.44) if we assume
0 /r ) around the sphere. (Note that (a0 ) =
an incompressible flow vr = ia(a

a0 vr dr = iaa0 .)
Furthermore, we note that, in order to deliver acoustical energy, a volume source
needs to be surrounded by a field of high pressure. This occurs when it is surrounded
by a surface of which the real part of the impedance is large. A force field needs a large
velocity fluctuation to produce acoustical energy efficiently. This corresponds to a large
real part of the acoustical admittance Y = 1/Z .
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
47
2.3.7 Free-space Greens function
The free-space Greens function G 0 is the acoustical field generated at the observers
position x at time t by a pulse (x y)(t ) released in y at time . To calculate the
free-space Greens function G 0 , we will make use of the Fourier transform (2.63, 2.64).
We seek a spherically symmetric wave solution (2.82) of the form
A
(2.112)
G 0 = eikr ,
r
where r = x y. To determine the amplitude A, we integrate the wave equation (2.60) over a compact sphere of radius a0 around y. Making use of the properties of the delta function we find that

ei
= (k 2 G 0 + 2 G 0 ) dV
2
V

G 0
2
2 G0

G 0 dV =
.
(2.113)
dS = 4a0
r
V
S r
r =a0
Using the near-field approximation ( G 0 /r )r =a0 A/a02 , we can calculate the amplitude A. We find
1
ei ir/c0 ,
8 2r
which leads by (generalized) inverse Fourier transformation to
G 0 =
(2.114)
1
(2.115)
(t r/c0 ).
4r
We observe at time t at a distance r from the source a pulse corresponding to the
impulsion delivered at the emission time:
r
(2.116)
te = t .
c0
G0 =
Because G 0 depends only on r = x y rather than on the individual values of x and

y, the free-space Greens function not only satisfies the reciprocity relation (2.67) but
also the symmetry relation:
G 0 r
G 0 r
G 0
G 0
=
=
=
.
xi
r xi
r yi
yi
(2.117)
Approaching the source by the listener has the same effect as approaching the listener
by the source r/ xi = r/ yi .
2.3.8 Multipole expansion
We can use the free-space Greens function G 0 to obtain a more formal definition of
monopoles, dipoles, quadrupoles, and so on. As we will see, this corresponds to the use
of a Taylor expansion of the free-space Greens function. We will consider the far-field
14:12
P1: JZZ
0521871441c02
48
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
p in free space of a compact source distribution q(x, t). To derive the general multipole
expansion, we will first consider the field at a single frequency. By using the free-field
Greens function
G 0 (x|y) =
eikr
,
4r
(2.118)
e it in a
we find the acoustic field for a given time-harmonic source distribution q(x)
finite volume V to be given by

eikr

G 0 (x|y) dVy =
dVy .
(2.119)
q(y)
q(y)
p =
4r
V
V
Suppose the origin is chosen inside V . We are interested in the far field (i.e., x is
large) and a compact source (i.e., k L is small, where L is the typical diameter of V ).
This double limit can be taken in several ways. Because we are interested in the radiation
properties of the source, which corresponds to kx O(1), we will keep kx fixed. In
that case, the limit of small k is the same as small y, and we can expand in a Taylor
series around y = 0:

2
xy
y2
(xy)2
2 1/2
= x 1
+
+ ...
r = x 2(xy) + y
x2
2x2
2x4

1 y2 2
y
sin
+
.
.
.
,
cos +
= x 1
x
2 x2
where is the angle between x and y, and

3

eikx
eikr
1
=
1 + (1 + ikx)
x j yj + . . .
r
x
x2 j=1
l+m+n

eikr
y1l y2m y3n

.
=
l! m! n! y1l y2m y3n r
y1 =y2 =y3 =0
l,m,n=0
(2.120)
Utilizing the symmetry of r as a function of x and y, we find this is equivalent to

ikx

e
eikr
(1)l+m+n l m n l+m+n
=
y1 y2 y3 l m n
.
(2.121)
r
l! m! n!
x
x1 x2 x3
l,m,n=0
The acoustic field is then given by
ikx

e
1
l+m+n
(1)l+m+n

l m n
dy l m n
.
p =
y1 y2 y3 q(y)
4 l,m,n=0 l! m! n! V
x
x1 x2 x3
(2.122)
Because each term in the expansion is by itself a solution of the reduced wave equation,
this series yields a representation in which the source is replaced by a sum of elementary
sources (monopole, dipoles, quadrupoles in other words, multipoles) placed at the
origin (y = 0). Expression (2.122) is the multipole expansion of a field from a finite
source in the Fourier domain. From this result, we can obtain the corresponding expansion in the time domain.
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
49
From the integral formulation (2.66), we have the acoustic field from a source q(x, t):

q(y, t r/c0 )
(t r/c0 )
p =
dyd =
dy.
(2.123)
q(y, )
4r
4r
V
V
If the dominating frequencies in the spectrum of q(x, t) are low, such that L/c0 is
small, we obtain by Fourier synthesis of Equation (2.122) the multipole expansion in
the time domain (see Goldstein 1976):

1
1
(1)l+m+n l+m+n
l m n
y
y
y
q(y,
t
)
dy
p =
e
4 l,m,n=0 l! m! n! x1l x2m x3n x V 1 2 3

l+m+n
(1)l+m+n
(2.124)
lmn (te ) ,
=
l
m
n
4 x
l,m,n=0 x 1 x 2 x 3
where te = t x/c0 is the emission time and lmn (t) is defined by
l m n
y1 y2 y3
lmn (t) =
q(y, t) dy.
V l! m! n!
(2.125)
The (lmn)-th term of the expansion (2.124) is called a multipole of order 2l+m+n . The
20 -order term corresponds to a monopole, a concentrated volume source at y = 0 with

source strength 000 = V q(y, t)dVy , which is called the monopole strength.
Because each term is a function of x only, the partial derivatives to xi can be
rewritten into expressions containing derivatives to x. In general, these expressions
are rather complicated, and so we will not try to give the general formulas here.
For very large x, each multipole further simplifies because

(te )
(te ) xl
1
(te ) =
xl x
c0 x
x2 x

xl
(te ) xl
=
(te ).
c0 x x
c0 x2 t
(2.126)
This leads to
p
x1l x2m x3n

l+m+n
lmn (te ),
4 (c0 x)l+m+n x t l+m+n
l,m,n=0
(x ).
(2.127)
Most results in Section 2.3.9 will be presented in this far-field approximation.
2.3.9 Doppler effect
We can use the Greens function formalism to determine the effect of the movement of
a source on the radiated sound field. We consider a point source localized at the point
xs (t):
q(x, t) = Q(t)(x xs (t)).
(2.128)
14:12
P1: JZZ
0521871441c02
50
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
For free-field conditions, using Equation (2.115), we find

y x
Q( )(y xs ( ))

t
dVy d.
p (x, t) =
4 y x
c0
V
(2.129)
After integration over space, using the property (2.61) of the delta function, we obtain

R
Q( )

p (x, t) =
t
d,
(2.130)
c0
4 R
where
R = R.
R(, x) = x xs ( ),
(2.131)
The contributions of this integral are limited to the zeros of the argument of the function. In other words, this is an integral of the type

tn +
d
F( )(g( )) d =
F( )(( tn ) d
g(tn )) d
tn
F(tn )
d
| d
g(tn )|
(2.132)
where = tn corresponds to the roots of g( ). In the present application, we have

g( ) = t
R(, x)
,
c0
(2.133)
and so
dg
Rvs
= 1 +
= 1 + Mr ,
d
Rc0
where vs =
dxs
d
(2.134)
and Mr is the component of the source velocity vs in the direction of the listener scaled
by the sound speed c0 . We call this the relative Mach number of the source. It is positive
for a source approaching the observer and negative for a source receding from the
observer. It can be shown that, for subsonic source velocities |Mr | < 1, the equation
g(te ) = 0 or
c0 (t te ) = R(te , x)
(2.135)
has a single root, which is to be identified as the emission time te . Hence, we find for
the acoustic field the LienardWiechert potential (Jackson 1999)
p (x, t) =
Q(te )
.
4 R(1 Mr )
(2.136)
When the source moves supersonically along a curve, multiple solution te can occur.
This may lead to a focusing of the sound into certain regions of space, leading to the
so-called super bang phenomenon.
The increase (when approaching) or decrease (when receding) of the amplitude is
called the Doppler amplification, and the factor (1 Mr )1 is called the Doppler factor.
This Doppler factor is best known from its occurrence in the increased or decreased pitch
of the sound experienced by the listener. For a sound source harmonically oscillating
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
51
with frequency that is high compared with the typical sound source velocity variations,
the listener experiences at time t a frequency
d(te )
.
=
dt
1 Mr
(2.137)
The right-hand side is obtained by implicit differentiation of Equation (2.135).

Hence, the observed frequency is the emitted one multiplied by the Doppler factor.
In this discussion we ignored the physical character of the source. If, for example,
we consider a monopole source with a volume injection rate V (t), the source is
given by
q(x, t) = 0
(V (t)(x xs (t))) .

t
The corresponding sound field is

0 V (te )
p (x, t) =
t 4 R(te , x)(1 Mr (te ))

0 V (te )
1
.
=
1 Mr (te ) te 4 R(te , x)(1 Mr (te ))
(2.138)
(2.139)
Although this is for an arbitrary source V and path xs an extremely complex solution, it
is interesting to note that, even for a constant volume flux V , there is sound production
when the source velocity vs is nonuniform.
In a similar way, we may consider the sound field generated by a point force F(t):
q(x, t) = [F(t) (x xs (t))].
The produced sound field is given in a far-field approximation by

R
.
p (x, t) =
c0 R(1 Mr ) te 4 R(1 Mr )
(2.140)
(2.141)
Even when the source flies at constant velocity dtd xs this solution involves high powers
of the Doppler factor.
An interesting application of this theory is the sound production by a rotating blade.
The blade can be represented by a point force (mainly the lift force concentrated in a
point) and a compact moving body of constant volume Vb (the blade). The lift noise (the
contribution of the lift force) can be calculated by means of Equation (2.141). Note that
in practice this lift is not only the steady thrust of the rotating blade but also contains
the unsteady component due to interaction of the blades with obstacles like supports
or with a turbulent or nonuniform inflow. The effect of the volume of the blade can be
shown to be given by the second time derivative

0 Vb
2

.
(2.142)
p (x, t) = 2
t 4 R(1 Mr )
Even though the blade volume remains constant, the displacement of air by the rotating
blade induces a sound production. This so-called thickness noise depends on a higher
14:12
P1: JZZ
0521871441c02
52
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
power of the Doppler factor than the lift noise. This implies that, at low Mach numbers
such as prevail for a ventilation fan, the lift noise will be dominant. At high-source
Mach numbers, such as aircraft propellers, the lift noise will still dominate at takeoff
when the required thrust is high, but at cruise conditions the thickness noise becomes
comparable because the tip Mach number is close to unity (or even higher).
We will see in the next section that the sound produced by turbulence in a free jet has
the character of the field of a quadrupole distribution. Because the vortices that produce
the sound are convected with the main flow, there will be a very significant Doppler
effect at high Mach numbers. This results in a radiation field mainly directed about
the flow direction. Owing to convective effects on the wave propagation, the sound is,
however, deflected in the shear layers of the jet. This explains why, along the jet axis,
there is a so-called cone of silence (Goldstein 1976; Hubbard 1995).
2.3.10 Uniform mean flow, plane waves, and edge diffraction
The problem of a source, observer, and scattering objects moving together steadily in
a uniform stagnant medium is the same as the problem of a fixed source, observer,
and objects in a uniform mean flow. If the mean flow is in the x direction and the
perturbations are small and irrotational, we have for potential , pressure p, density ,
and velocity v the problem given by

2
2
2
1
2
+ 2 + 2 2
= 0,
+ U0
x2
y
z
x
c0 t

+ U0
, p = c02 , v = ,
p = 0
t
x
(2.143)
where U0 , 0 , and c0 denote the mean flow velocity, density, and sound speed, respectively. We assume in the following that |U0 | < c0 . The equation for is known as the
convected wave equation.
2.3.10.1 Lorentz or PrandtlGlauert transformation
By the transformation (in an aerodynamic context named after Prandtl and Glauert but
qua form originally due to Lorentz)
X=
x
,
T = t +
M
X,
c0
M=
U0
,
c0

1 M 2,
(2.144)
the convected wave equation may be associated with a stationary problem with solution
(x, y, z, t) = (X, y, z, T ) satisfying
2
2
1 2
2
+
+
= 0,
X2
y2
z 2
c02 T 2
0
p=
+ U0
T
X

.
(2.145)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
53
For a time-harmonic field e it (x, y, z) = e iT (X, y, z) or (x, y, z) =

e iK M X (X, y, z), where = /, k = /c0 , and K = k/, we have
2
2
2
+
+
+ K 2 = 0.
X2
y2
z 2
(2.146)
The pressure may be obtained from , but because p satisfies the convected wave equation too, we may also associate the pressure field directly by the same transformation
with a corresponding stationary pressure field. The results are not equivalent, however,
and especially when the field contains singularities some care is in order. The pressure
obtained directly is no more singular than the pressure of the stationary problem, but the
pressure obtained via the potential is one order more singular owing to the convected
derivative. See the example in Sections 2.3.10.2 and 2.3.10.3.
2.3.10.2 Plane waves
A plane wave (in x, y plane) may be given by

x cos n + y sin n
i = exp ik
1 + M cos n

cos( n )
= exp ikr
,
1 + M cos n
(2.147)
where n is the direction of the normal to the phase plane and x = r cos , y = r sin .
This is physically not the most natural form, however, because n is due to the mean
flow, not the direction of propagation. By comparison with a point source field far away
or from the intensity vector

e y
I = (0 v + v0 )(c02 /0 + v0 v) 2 ik M e x +
x
y
(M + cos n )e x + sin n e y ,
(2.148)
we can learn that s , the direction of propagation (the direction of any shadows), is
given by
M + cos n
,
cos s =
1 + 2M cos n + M 2
sin s =
sin n
1 + 2M cos n + M 2
(2.149)
By introducing the transformed angle s

cos s
M + cos n
cos s =
=
,
1 + M cos n
1 M 2 sin2 s
(2.150)
sin s
sin n
=
sin s =
2
2
1 + M cos n
1 M sin s
(2.151)
and the transformed polar coordinates X = R cos , y = R sin , we obtain the plane
wave
i = exp (i K M X i K R cos( s )) .
(2.152)
14:12
P1: JZZ
0521871441c02
54
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 2.2. Sketch of scattered plane wave with mean flow.
2.3.10.3 Half-plane diffraction problem
By using the foregoing transformation, we obtain from the classical Sommerfeld solution for the half-plane diffraction problem (see Jones 1964) of a plane wave (Equation (2.152)) incident on a solid half-plane along y = 0, x < 0 (Figure 2.2), the following solution (see Rienstra 1981a):

(2.153)
(x, y) = exp (i K M X i K R) F(s ) + F( s ) ,
where
ei/4 2
F(z) = e iz
eit dt
2
(2.154)
and
s , s = (2K R)1/2 sin 12 ( s ).
(2.155)
An interesting feature of this solution is the following. When we derive the corresponding pressure

p(x, y) = exp (i K M X i K R) F(s ) + F( s )
ei/4 M cos 12 s
exp (i K M X i K R) sin 12
+
1 M cos s
2
KR
1/2
,
(2.156)
we see immediately that the first part is just a multiple of the solution of the potential,
and so the second part has to be a solution too. Furthermore, the first part is regular
like , whereas this second part is singular at the scattering edge. Because the second
part decays for any R , it does not describe the incident plane wave, and so it may
be dropped if we do not accept the singularity in p at the edge. By considering this
solution,
sin 1
pv (x, y) = exp (i K M X i K R) 2 ,
KR
(2.157)
a bit deeper, we find that it has no continuous potential that decays to zero for large |y|
(see Jones 1972; Rienstra 1981a). This solution corresponds to the field of vorticity (in
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
55
2.4 AEROACOUSTIC ANALOGIES
the form of a vortex sheet) that is being shed from the edge. This may be more clear if
we construct the corresponding potential v for large x, which is

v sign(y) exp |y| i x ,

U0
U0
pv 0.
(2.158)
In conclusion, we obtain the continuous-potential, singular solution by transforming

the no-flow solution in potential form and the discontinuous-potential, regular solution
from the no-flow solution in pressure form. The difference between both is the field of
the shed vortex sheet.
The assumption that just enough vorticity is shed that the pressure field is no longer
singular is known as the unsteady Kutta condition. Physically, the amount of vortex
shedding is controlled by the viscous boundary layer thickness compared with the
acoustic wavelength and the amplitude (and the Mach number for high-speed flow).
These effects are not included in the present acoustic model; therefore, they have to be
included by an additional edge condition (e.g., the Kutta condition). Because vorticity
can only be shed from a trailing edge, a regular solution is only possible if M > 0. If
M < 0, the edge is a leading edge and we have to leave the singular behavior as it is.
Note that the same physical phenomenon seems to occur at a transition from a
hard to a soft wall in the presence of mean flow. Normally, at the edge there will be
a singularity. When the soft wall allows a surface wave of particular type accounting
for a modulated vortex sheet along the line surface to form, a Kutta condition can be
applied that removes the singularity possibly at the expense of a (spatial) instability
(Quinn and Howe 1984; Rienstra 1986, 2002). It should be emphasized that this results
from a linear model and that an instability that is too severe may not be acceptable in a
fully nonlinear model.
2.4 Aeroacoustic analogies

2.4.1 Lighthills analogy
Until now we have considered the acoustic field generated in a quiescent fluid by an
imposed external force field f or by heat production Q w . We have further assumed that
the sources induce linear perturbations of the reference quiescent fluid state. Lighthill
(1952) proposed a generalization of this approach in the case of an arbitrary source
region surrounded by a quiescent fluid. Hence, we no longer assume that the flow in
the source region is a linear perturbation of the reference state. We assume only that the
listener is surrounded by a quiescent reference fluid ( p0 , 0 , s0 , c0 uniformly constant
and v0 = 0) in which the small acoustic perturbations are accurately described by the
homogeneous linear wave equation (2.70). Lighthills key idea was to derive from the
exact equations of mass conservation (2.6) with Q m = 0 and momentum conservation
(2.8) a nonhomogeneous wave equation that reduces to the homogeneous wave equation
(2.70) in a region surrounding the listener.
14:12
P1: JZZ
0521871441c02
56
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
By taking the time derivative of the mass conservation law (2.6) and subtracting the
divergence of the momentum equation (2.8), we obtain
2
fi
2
=
(Pi j + vi v j )
.
t 2
xi x j
xi
(2.159)

By adding the term c02 t
2 p to both sides and by making use of definition (2.9) (i.e.,
i j = pi j Pi j ), we can write Equation (2.159) as

1 2 p
p
2 p
2
fi
2

(2.160)
=
(v
v
i j
ij
xi x j
xi
t 2 c02
c02 t 2
xi2
2
in which the perturbations p and are defined by

p = p p0
and
= 0 .
(2.161)
This equation is called the analogy of Lighthill. Note that neither /0 nor p / p0
is necessarily small in the so-called source region. In fact, Equation (2.160) is exact.
2
Furthermore, because we obtained this equation by adding the term c02 t
2 p to both
sides, this equation is valid for any value of the velocity c0 . In fact, we could have chosen
c0 = 3 108 ms1 (the speed of light in vacuum) or c0 = 1 mm/century. Of course,
the equation would be quite meaningless then. By choosing for c0 and p0 the values
of the reference quiescent fluid surrounding the listener, we recover the homogeneous
wave equation (2.70) whenever the right-hand side of Equation (2.160) is negligible.
Hence, Equation (2.160) is a generalization of Equation (2.59), which was derived for
linear perturbations of a quiescent fluid.
We should now realize that we did not introduce any approximation to Equation
(2.160), and so this is exact. Therefore, this equation does not provide any new information that was not already contained in the equations of mass conservation (2.6) and
of momentum (2.8). In fact, we have lost some information. We started with four exact
Equations (2.6, 2.8) and eleven unknowns (vi , p, , i j ). We are now left with one
Equation (2.160) and still eleven unknowns. Obviously, without additional information
and approximations we have not gotten any closer to a solution for the acoustic flow.
The first step in making Lighthills analogy useful was already described above.
We have identified a listener around which the flow behaves like linear acoustic perturbations and is described by the homogenous wave equation (2.70). This assumption
is valid in many applications. When we listen, under normal circumstances, to a flute
player we have conditions that are quite reasonably close to these assumptions. At this
stage the most important contribution of Lighthills analogy is that it generalizes (2.161)
the equations for the fluctuations and p to the entire space even in a highly nonlinear
source region. The next steps are the introduction of approximations to estimate the
source terms that is, the right-hand side of Equation (2.160). We recognize in the
2 p

right-hand side of Equation (2.160) the term t
2 ( 2 ), which is a generalization of
c0
the entropy production term in Equation (2.59).
As shown by Morfey (1973) (see Crighton et al. 1992), this term includes complex
effects due to the convection of entropy nonuniformities. The effect of external forces f
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
is the same in both Equation (2.70) and Lighthills analogy (2.160). We have, however,
now removed the condition that the force should only induce a small perturbation
to the reference state. An arbitrary force is allowed as long as we take into account
any additional effects it may have on the other terms at the right-hand side of Equation (2.160).
We observe additional terms due to the viscous stress i j and the Reynolds stress
vi v j . The viscous stress i j is induced by molecular transport of momentum, whereas
vi v j takes into account the nonlinear convection of momentum. One of Lighthills
(1952) key ideas is that, when the entropy term and the external forces are negligible,
the flow will only produce sound at high velocities, corresponding to high Reynolds
numbers. He therefore assumed that viscous effects are negligible and reduce the
sound source to the nonlinear convective effects ( 2 vi v j / xi x j ). It is worth noting
that a confirmation of this assumptions reasonableness was provided qualitatively by
Morfey (1976, 2003) and Obermeier (1985) only about 30 years after Lighthills original publication. A quantitative discussion for noise produced by vortex pairing was
provided by Verzicco et al. (1997). An additional assumption, commonly used, is that
feedback from the acoustic field to the source is negligible. Hence, we can calculate the
source term from a numerical simulation that ignores any acoustic wave propagation
and subsequently predict the sound production outside the flow. In extreme cases of
low-Mach-number flow, a locally incompressible flow simulation of the source region
can be used to predict the (essentially compressible!) sound field.
Equation (2.160) can formally be solved by an integral formulation of the type
(2.66). This will have the additional benefit of reducing the effect of random errors
in the source flow on the predicted acoustic field. One can state that such an integral formulation combined with Lighthills analogy allows to be obtained a maximum
of information concerning the sound production for a given information about the
flow field. A spectacular example of this is Lighthills prediction that the power radiated to free space by a free turbulent isothermal jet scales as the eighth power, U08 ,
of the jet velocity. This result is obtained from the formal solution for free-space
conditions:

r
2
c
0
dVy d
vi v j
p (x, t) =
xi x j V
4r

vi v j
2
=
dVy ,
(2.162)
xi x j V 4r =te
where r = x y and te = t r/c0 . Assuming that the sound is produced mainly by
the large turbulent structures with a typical length scale of the width D of the jet, we
estimate the dominating frequency to be f = U0 /D, where U0 is the jet velocity at the
exit of the nozzle. Hence, the ratio of jet diameter to the acoustical wavelength D f /c0 =
U0 /c0 = M. This implies that, at low Mach numbers, we can neglect variations of the
retarded time te if the source region is limited to a few pipe diameters. Because the
acoustic power decreases very quickly with decreasing flow velocity, we have indeed
57
14:12
P1: JZZ
0521871441c02
58
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
a source volume of the order of D 3 . Inasmuch as v U0 and 0 , we may assume

that vi v j 0 U02 . From the far-field approximation / xi = /(c0 t) 2 f /c0 ,
we find
p (x, t) 0 U02 M 2
D
,
x
(2.163)
where we ignore the effect of convection on the sound production (Goldstein 1976;
Hubbard 1995). The radiated power is thus found to be
I
4x2 p vr
0 U03 M 5 D 2 U08 .
(2.164)
This scaling law appears indeed to be quite accurate for subsonic isothermal free jets.
Note that this law implies that we can achieve a dramatic reduction of aircraft jet noise
production by decreasing the flow Mach number. In order to retain the necessary thrust,
the jet cross section has to be increased. This is exactly what happened in the 1960s and
1970s when the high-bypass turbofan aeroengines replaced the older turbofan engines
without or with little bypass flow. This is known as the turbofan revolution (Smith
1989; Hubbard 1995).
As stated by Crighton et al. (1992) and Powell (1990), Lighthills theory is rather
unique in its prediction of a physical phenomenon before experiments were accurate
enough to verify it. This made Lighthills analogy famous. Note that we have actually
discarded any contribution from entropy fluctuations or external forces. This means
that, if we use as input for the analogy data obtained from a numerical simulation that
include significant viscous dissipation and spurious forces, we still would predict the
same scaling law.
Because the amplitude of the acoustic pressure generated by a compact monopole
and dipole would scale in free-field conditions as M 2 and M 3 , respectively, these
spurious sources easily dominate the predicted sound amplitude. This is one reason
why most of the direct numerical simulations of sound production by a subsonic flow
in free-field conditions are carried out at high Mach numbers (M 0.9).
In the presence of walls, the sound radiation by turbulence can be dramatically
enhanced. In the next section, we will see that compact bodies will radiate a dipole
sound field associated with the force they exert on the flow as a reaction to the hydrodynamic force of the flow applied to them. Sharp edges are particularly efficient
radiators. This corresponds to our common experience with the production of sounds
like the consonant /s/. The interaction of the sharp edges of our teeth is essential. In
free-field conditions, much attention has been devoted to the so-called trailing-edge
noise of aircraft wings (Goldstein 1976; Blake 1986; Hubbard 1995; Howe 1998). The
scaling rule for this sound field is p M 5/2 , which is just between a monopole and
a dipole. For confined subsonic flows at low frequencies, the scaling rules are quite
different. A compact turbulent flow in an infinitely long straight duct will produce
an acoustic field that scales according to p M 6 (Ffowcs Williams 1969; Michalke
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
59
1989; Rienstra and Hirschberg 2001). The sound field from a monopole and an axially
aligned dipole will both scale with M 2 , whereas transversal dipoles will not radiate any
sound. We will consider elements of the acoustics and aeroacoustics of confined flows
in Section 2.5. The scaling rules for supersonic free jets are obscured by the temperature
difference between the flow and the environment (Tam 2001). Globally, however, one
expects a power that is limited by a U03 law because an extrapolation of the scaling
rule U08 would imply that the acoustical power generated by the flow would soon become larger than the kinetic energy flux of the flow, which scales as 12 0 U03 . At high Mach
numbers, the theory has to be modified to take intrinsic temperature differences into
account.
2.4.2 Curles formulation
The integral formulation of Lighthills analogy can be generalized for flows in the presence of walls. We will discuss later the use of tailored Greens functions (Section 2.5.3).
We now consider Curles (1955) approach. We use the free-space Greens function G 0
(2.116). Instead of the pressure p as the aeroacoustical variable we use the density .
Subtracting from both sides of Equation (2.159) the term c02 ( 2 / xi2 ), we obtain the
analogy of Lighthill for the density,
2 Ti j
2
2
fi
c02 2 =
,
2
t
xi x j
xi
xi
(2.165)
in which the stress tensor of Lighthill, Ti j , is defined by

Ti j = Pi j + vi v j c02 i j .
(2.166)
We further assume that f = 0 and focus on the other sound sources. We have selected
here the density as the dependent variable because this was Lighthills (1952) original
choice. We will later discuss the implications of this choice (Section 2.4.4).
We consider a fixed surface S with outer normal n and apply Greens theorem (2.66)
to the volume V outside of S. Note that n is chosen toward the interior of V , and so the
sign convention of n is opposite to the sign convention used in Equation (2.66). By means
of partial integration and by utilizing the symmetry properties G 0 / xi = G 0 / yi
(2.123) and G 0 / = G 0 /t of the Greens function G 0 , we obtain (Goldstein
1976; Rienstra and Hirschberg 2001)
p (x, t) = c02 (x, t) =
2
xi x j
x j

V

S
Ti j
4r

=te
dVy +
Pi j + vi v j
4r

=te

vi
n i dS
S 4r =te
n i dS,
(2.167)
14:12
P1: JZZ
0521871441c02
60
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
which is clearly a generalization of Equation (2.162). In this equation we used the

assumption that, at the listeners position, p = c02 . In the far-field approximation
(2.126), we find

1
vi
dVy +
n i dS
4 t S r =te
V
=te

Pi j + vi v j
xj
n i dS.
(2.168)
+
4 xc0 t S
r
=te
xi x j
2
p (x, t)
4 x2 c02 t 2

Ti j
r
For a compact body we can neglect the variations of te over the surface and can write
r = x if we choose the origin y = 0 inside or near the body. Assuming that Ti j decays
fast enough, we have in that case
xi x j
2
p (x, t)
4 x3 c02 t 2

xj
+
2
4 x t

1
[Ti j ] =tedVy +
4 x t
[Pi j + vi v j ] =ten i dS,

S
[vi ] =ten i dS
(2.169)
and te = t x/c0 . The second integral corresponds to the monopole sound field
generated by the mass flux through the surface S. The third integral corresponds to
the dipole sound field generated by the instantaneous force F j of the surface to

the surrounding fluid. This is the reaction of the surface to the force F j = S (Pi j +
vi v j )n i dS of the flow on the surface. This result is a generalization of Gutins principle
(Gutin 1948; Goldstein 1976).
Using this theory, we easily understand that a rotor blade moving in a nonuniform
flow field will generate sound owing to the unsteady hydrodynamic forces on the blade.
At low Mach number, this will easily dominate the Doppler effect due to the rotation.
Wind rotors placed downwind of the supporting mast are cheap because they are hydrodynamically stable. There is no need for a feedback system to keep them in the wind.
However, the interaction of the wake of the mast with the rotor blades causes dramatic
noise problems (Hubbard and Shepherd 1991; Wagner, Bareiss, and Guidati 1996).
2.4.3 Ffowcs WilliamsHawkings formulation
Although Curles formulation discussed in the previous section assumes a fixed control
surface S, the formulation of Ffowcs Williams and Hawkings allows the use of a
moving control surface S(t). The key idea is to include the effect of the surface in the
differential equation (2.165) as described by Ffowcs Williams and Hawkings (1969),
Goldstein (1976), and Crighton et al. (1992). This is achieved by a most elegant and
efficient utilization of generalized functions (viz. so-called surface distributions).
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
61
We assume that the volume B(t) enclosed by the surface S(t) and this surface are
sufficiently smooth to allow the definition of a smooth function h(x, t) such that
h(x, t) < 0 if x B(t),
(2.170)
h(x, t) = 0 if x S(t),
h(x, t) > 0 outside B(t).

Consider any physical quantity, like , defined outside B(t) and extend its definition
to all space by giving it a value equal to zero inside B(t). This is efficiently done by
assuming smoothly defined everywhere. Then by multiplying it by the Heaviside
function H (h) we create a new variable H (h) that vanishes within the volume B(t)
(where H (h) 0) and is equal to outside B(t) (where H (h) 1). The next step
will be to extend the prevailing equations to the whole space by adding suitable surface
sources. To achieve this we need the normal n to the surface S(t) given by

h
.
(2.171)
n=
h h=0
We assume that the surface S(t) is parameterized in time and space by the coordinates
(t; , ). A point xs (t) S(t) with fixed parameters and is moving with the velocity
b. Hence, we have h(xs , t) = 0 and
h
= b h = (bn)h.
(2.172)
t
After multiplying the mass conservation equation (2.6) and the momentum equation (2.8) by H (h) and reordering terms, we obtain the following equations valid everywhere:

[ H ] + [vH ] = [0 b + (v b)] H,
t
(2.173a)
[vH ] + [(P + vv)H ] = [P + v(v b)] H,

(2.173b)
t
where H stands for H (h). Because H = (h)h, the equations can be interpreted
as generalizations of the mass and momentum equations with surface sources at S.
Using the relations above and following Lighthills procedure for acoustic variable
p = p p0 , we find
1 2
[ p H ] 2 [ p H ] = [[(vv )H ]] [fH ]
c02 t 2

2
p

[(0 b + (v b)) H ]
+ 2
H +
2
t
t
c0
[( p I + v(v b)) H ],
(2.174)
When S(t) is the surface of a solid and undeformable body, it is natural to assume a spatial parametrization that is materially attached to the surface. Like the auxiliary function h, this parametrization is not
unique.
14:12
P1: JZZ
0521871441c02
62
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where (I)i j = i j and p0 is the uniform reference value of the pressure. Note that
(( p0 I H )) = ( p0 I H ). For a solid surface, vn = bn. In that case, by
applying Greens theorem and using the free-space Greens function, we find
p (x, t) =

(vi v j i j )H
2
dVy
x i x j R3
4r
=te

2
fH
( p /c0 )H
dV
+
dVy
y
xi R3 4r =te
t 2 R3
4r
=te

0 bn
+
dS
t S(te ) 4r (1 Mr ) =te

p n i i j n j
dS,
(2.175)
xi S(te ) 4r (1 Mr ) =te
where r = x y and Mr = b(x y)/r c0 , and we use the following generalizations

of Equation (2.132):

g(x)
dS,
(2.176a)
g(x)(h(x)) dx =
3
R
S |h|

g(x)
H
(h(x))
dx
=
(g h)(h) dx
3
3
R
R

g h
=
(2.176b)
dS = g(x)n(x) dS.
S |h|
S
The first three integrals correspond to the contribution of the flow around the surface
S(t), and the last two integrals represent generalizations of the thickness noise and sound
generated by the surface forces we have discussed earlier. A reduced form, widely used
for subsonic propeller and fan noise when volume sources and surface stresses are
negligible, is thus (Farassat 1981)

1
0 bn

dS
p (x, t) =
4 t S(te ) r (1 Mr ) =te

p ni
1
dS.
(2.177)
4 xi S(te ) r (1 Mr ) =te
2.4.4 Choice of aeroacoustic variable
In the previous discussion we used p as the dependent aeroacoustical variable to

introduce the analogy of Lighthill (Section 2.4.1). We then used to introduce the
formulation of Curle describing the effects of stationary boundaries (Section 2.4.2).
Finally, for the formulation of Ffowcs Williams and Hawkings describing the effect of
moving boundaries, we retuned to p . In acoustics this would have been indifferent
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
63
because the two variables are related by the equation of state p = c02 . Inasmuch as
we assume linear acoustics of quiescent fluids to be valid around the listener, we made
use of this relationship in Equation (2.167). Actually, in aeroacoustics there is a subtle
difference that appears when we compare the source terms of the two wave equations

p
2 p
2
fi
2
1 2 p

(2.160)
=
(v
v
i j
ij
xi x j
xi
t 2 c02
c02 t 2
xi2
and
2
2
fi
2
2
2
c
=
(v
v
+
( p c02 ).
i j
ij
0
t 2
xi x j
xi
xi2
xi2
(2.178)
Without further approximation, these two forms are both equivalent. However, considered as an analogy, the right-hand sides are assumed to be known and act as given
source distribution. In that case we see that when p is used as the aeroacoustical variable,
the effect of entropy fluctuations 2 (( p /c02 ) )/t 2 has the character of a monopole
sound source. On the other hand, when is used the apparently same effects produce a
quadrupole distribution 2 ( p c02 )/ xi2 that is qualitatively different. Of course,
there is no difference if we consider the exact equations, but if we do not introduce any
approximation the analogy is just a reformulation of basic equations without much use.
Clearly, we have to be careful in selecting the aeroacoustic variable.
When considering sound production by subsonic flames, we should choose p as the
aeroacoustical variable because most of the sound is produced by the volume changes
associated with the combustion. When we neglect convection effects, we have exactly the source term 2 (( p /c02 ) )/t 2 . As shown by Morfey (1973) and Crighton
et al. (1992), this term includes complex effects due to the convection of entropy
nonuniformities. They become explicit when we use the equation of state (2.17) as
applied to a material element,

Ds
p
Dp
2 D
=c
+
,
(2.179)
Dt
Dt
s Dt
in combination with Equation (2.14):
Ds
De
D
T
=
+p
Dt
Dt
Dt

1
.
Morfey (1973) obtains the following result:

2

e D
2 e
c
2 T
Ds
1+
2 =
+ 2
+ (e v) ,
t
t
Dt
c02
c0 s Dt
(2.180)
(2.181)
where the excess density e is defined by e = ( p /c02 ). The first term vanishes in
a subsonic free jet of an ideal gas with constant heat capacity. The second term is the
entropy production (combustion) term corrected for convective effects. The last term
corresponds to the force exerted by a patch of fluid with a different density on its surroundings in an accelerating flow. This may be compared with a buoyancy (Archimedes)
14:12
P1: JZZ
0521871441c02
64
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
effect. It is induced by the fact that, owing to the difference of density between the particle and its surroundings, the pressure gradient imposed by the surroundings of a particle
does not match its acceleration.
The question arises whether could be a useful choice too. If we consider a
turbulent isentropic flow in a region with a speed of sound c that differs strongly
from the speed of sound c0 at the listener position, the source term can be rewritten
as 2 p (1 (c0 /c)2 )/ xi2 . In a subsonic turbulent flow we expect the local pressure
fluctuations p to scale with 12 U02 . Hence, the analogy indicates that, when the speed
of sound c0 at the listener position is much higher than the speed of sound c in the
source region, there will be a strong enhancement of the sound production compared
with a similar flow in a uniform fluid. Such a spectacular effect does indeed occur when
we consider a flow in a waterair mixture such as obtained by directing the nozzle of
the shower toward the surface of the water. As the result of air entrainment there will
be a volume fraction of air in the water flow. The density of the mixture will be
= (1 )water + air .
(2.182)
If we assume a quasi-static response of the bubbles and neglect dissolution of air in the
water, the compressibility 1/(c2 ) of the mixture will be the sum of the compressibilities
of both phases (van Wijngaarden 1972; Crighton et al. 1992):
1
1
=
+
.
2
2
2
c
water cwater
air cair
(2.183)
In the case of waterair mixtures for not-too-small nor too-large values of , the density
is mainly determined by the water phase, whereas the air determines the compressibility.
Hence, we find for the speed of sound that
c2
2
air cair
.
(1 )water
(2.184)
2
Typical values are = 0.5, air cair
= 1.4 105 Pa , water = 103 kg m3 , and cwater =
3
1
1.5 10 ms . We find c 2 101 ms1 , and, if we have our head underwater,
(c0 /c)2 5 103 . This should result in an enhancement in SPL on the order of 60 dB.
Indeed the flow is much noisier than when we avoid air entrainment by putting the
shower nozzle underwater.
Other choices of aeroacoustic variables lead to different analogies. In many cases
such analogies tend to avoid the problem induced by the fact that the analogy of Lighthill
does not distinguish between propagation and production of sound waves in a strongly
nonuniform flow, which induces refraction. This becomes very important in supersonic
flow. In such cases the source is not compact. Simple results like Equation (2.166) are
not valid anymore. Alternative analogies that attempt to overcome such problems are
described in the literature (Goldstein 1976; Hubbard 1995; Lilley 2004).
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
2.4.5 Vortex sound
One of the drawbacks of the analogy of Lighthill (2.160) is that the source term is spatially quite extended. As observed by Powell (1964), the sound production in subsonic
homentropic flows is associated to the dynamics of vortices. Vorticity = v appears to be spatially less extended than the Reynolds stress vi v j . The reason for this
is that around vortices there is a large region of potential flow that actually does not
produce any sound. In its original form, the vortex sound theory (Powell 1964) was
applied to free-field conditions at low Mach numbers. It was only a special form of
Lighthills analogy stressing the role of vorticity. In the case of free-field conditions, the
vortex sound theory greatly enhances the predicted sound field. Various modifications
of the theory of Powell have been proposed that more explicitly impose the conservation
of momentum and energy on the flow in the source region (Mohring 1978; Hirschberg
and Schram 1995; Schram and Hirschberg 2003). This also improves the performance
of the theory.
Howe (1975, 1998, 2002) has generalized the theory of Powell (1964) to allow
its application to confined flows and conditions in which the listener is placed into a
potential flow rather than a quiescent fluid. In this theory the fluctuations of the total
enthalpy B = ( p /0 ) + v v0 appear as a natural aeroacoustical variable (Howe 1975;
Doak 1989, 1995; Musafir 1997). In general form the vortex sound theory can be
applied to arbitrary Mach numbers. A similar analogy was derived by Mohring (2001)
on the basis of acoustical energy considerations. However, such analogies become quite
obscure. They do not provide many intuitive insights and can only be used numerically
as proposed by Ewert and Schroder (2004).
We consider now the case of low subsonic flows in that Howe (1980) proposed a
very nice energy corollary that provides much insight into the role of vorticity in sound
production. We propose here an intuitive approach to this theory. The key idea of the
theory is that a potential flow is silent. This is illustrated in Figure 2.3 in which we
consider a sketch of human vocal folds. The oscillation of the vocal folds results in a
variable volume flow from our lungs into the vocal tract. This variable volume flux is
the source of sound. Seeking a simplified model for the flow through the glottis, we
consider the Reynolds number Re and the Strouhal number St. For a typical flow velocity
U0 = 30 ms1 , a length scale on the glottis on the order of L = 2 mm, a frequency
f = 102 Hz, and a kinematic viscosity = 1.5 105 m2 s1 , we have Re = U0 L/ =
O(103 ) and St = f L/U0 = O(102 ). A quasi-steady, frictionless approximation seems
promising. Because the Mach number is low, M = O(101 ), and the flow is compact,
L/ = MSt = O(103 ), we assume that the flow is locally incompressible. Under such
circumstances, the equation of Bernoulli in the form p + 12 0 v 2 = pt with pt a constant
should be valid. About one diameter upstream and downstream of the vocal folds we
expect a uniform flow velocity. Assuming the flow channel upstream of the vocal folds
(trachea) to have the same cross section as the channel downstream (vocal tract), we
find that the velocities should be equal. As a consequence we conclude that there is no
pressure difference across the vocal folds. The volume flux is therefore not controlled
65
14:12
P1: JZZ
0521871441c02
66
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 2.3. A potential flow through the vocal folds is silent. Sound
is produced by the volume-flows-control associated with flow separation and formation of a free jet. This implies vorticity injection in
the main flow.
by the opening of the folds! This implies that they cannot produce any sound. This
corresponds to the paradox of dAlembert, which is solved by realizing that we can
never neglect the effect of friction at the walls. Even at high Reynolds numbers there
are always thin viscous boundary layers near the walls. In these boundary layers the
pressure is essentially equal to that of the main flow, but the velocity decreases to
zero at the wall. This implies that the stagnation pressure pt is lower than its value in
the main flow. Hence, the fluid in the boundary layers cannot flow against the strong
adverse pressure gradient in the diverging part of the glottis. This results in separation
of the boundary layers and the formation of a free jet. Turbulent dissipation in the free
jet explains the flow control by the oscillating vocal folds and the pressure difference
across the glottis. Of course, the fluid from the boundary layers injected into the main
flow has vorticity. Hence, we see that vorticity injection into the main flow is associated
with the production of sound. This illustrates the statement of Muller and Obermeier
(1988) that vortices are the voice of the flow.
As explained in Section 2.2.3, the velocity field can be separated into a potential
and a vortical flow. The potential part of the flow is associated with the dilatation
rate v = 2 of fluid particles in the flow. Because the acoustical flow is essentially
compressible and unsteady, Howe (1980) proposed to define the acoustic field as the
unsteady component of the potential flow uac ,
uac = ,
(2.185)
in which = 0 is the time-dependent part of the potential . For a homentropic

flow, we can write the equation of Euler (2.35) in the form
v
fc
+ B =
t
(2.186)
in which fc = ( v) is the density of the Coriolis force associated with the vorticity of the flow. When = 0, we have a potential flow. Hence, we identify fc / as the
source of sound. At low Mach number for compact flows, we can neglect the variation
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
67
of the density so that fc is the source of sound. Under such circumstances we can, in
first approximation, apply the energy equation (2.107) in the form

(2.187)
P
= fc uac
dV,
V
where P
is the time-averaged acoustical power generated by the vortices and V is the
volume in which the vorticity is confined.
The success of this equation is due to the fact that, even when using a highly
simplified vortex model, it provides a fair prediction of the sound production. A first
explanation for this is that it is an integral formulation that smooths out random errors. A
second aspect is that, when considering models of isolated vortices with time-dependent
induced
circulation , it ignores the effect of the spurious pressure difference t
by the model along the feeding line of the vortex. This corresponds to the force
necessary for the transport of vorticity toward the free vortex. Hence, a model that
does not satisfy the momentum equation still gives reasonable predictions because we
ignore the effect of spurious forces.
The theory of Howe provides much insight. In particular, Equation (2.187) clarifies
the essential role of sharp edges at which the potential flow is singular and vortex
shedding occurs. In Section 2.5.4 we consider examples of the use of this theory.
For the sake of completeness we now provide a more formal form of the analogy
of Howe (2002). Starting from the divergence of the momentum conservation equation
(2.186),
v
+ 2 B = ( v),
t
(2.188)
and eliminating v by means of the mass conservation law (2.5)

v =
1 D
,
Dt
(2.189)
we obtain for an isentropic flow

1 D20 B
1 D20 B
1 Di
2
,
B
=
(
v)
+
c2 Dt 2
c2 Dt 2
t c2 Dt
(2.190)
B
D0 B
=
+ (0 ) B ,
Dt
t
(2.191)
where
and we make use of the equation of state for isentropic flows:

1 Dp
1 Di
1 D
= 2
= 2 .
Dt
c Dt
c Dt
(2.192)
As in Howe (1998), the first source term ( v) in Equation (2.190) is dominant

at low Mach numbers. Although the resulting wave equation (2.190) looks relatively
simple, for an arbitrary steady potential flow 0 it can only be solved numerically.
14:12
P1: JZZ
0521871441c02
68
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 2.4. Straight duct of arbitrary cross section.
2.5 Confined flows

2.5.1 Wave propagation in a duct
For a narrow, hard-walled duct, the only waves that propagate are of the form given by
Equation (2.72), which is for a time-harmonic sound field in complex notation the pair
of plane waves
p (x, t) = A e itikx +B e it+ikx ,
(2.193)
where k = /c0 and the medium is uniform and stagnant. Note that each plane wave is
self-similar in x apart from a phase change. This solution can be generalized for wider
ducts or higher frequencies as follows.
The time-harmonic sound field in a duct of constant cross section with linear boundary conditions that are independent of the axial coordinate may be described by an
infinite sum of special solutions modes that retain their shape when traveling down
the duct. They consist of an exponential term multiplied by an eigenfunction of the
Laplace eigenvalue problem on a duct cross section.
Consider the two-dimensional area A in the (y, z) plane with a smooth boundary
A and an externally directed unit normal n. By shifting A in the x direction, we obtain
the duct D given by
D = {(x, y, z)|(0, y, z) A}
(2.194)
with axial cross sections being copies of A and for which the normal vectors n are the
same for all x (Figure 2.4). In the usual complex notation (with +itsign convention),
the acoustic field
p (x, t) p (x, ) e it ,
v (x, t) v (x, ) e it
(2.195)
satisfies, in the duct (x D), the equations

2 p + k 2 p = 0,
i0 v + p = 0.
(2.196)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
69
2.5 CONFINED FLOWS
At the duct wall we assume the impedance boundary condition

p = Z (v n) for x D.
(2.197)
Hard walls correspond to Z = . The solution of this problem may be given by

p (x, y, z) =
Cn n (y, z) ein x ,
(2.198)
n=1
where n are the eigenfunctions of the Laplace operator reduced to A, that is, solutions
of

2
2
+ 2 = 2 for (y, z) A,
y2
z
(2.199)
i0 = (n )Z for (y, z) A,
where 2 is the corresponding
eigenvalue. The axial wave number n is given by one
of the square roots n = k 2 n2 (+ for right and for left running), whereas the
branch of the square root is to be taken such that (n ) 0 and (n ) 0. Therefore,
the left-running and right-running modes are usually explicitly given as
p (x, y, z) =

n (y, z) An ein x +Bn e in x .
(2.200)
n=1
Each term in the series expansion that is, n (y, z) ein x is called a duct mode. For
hard walls, the eigenvalues n2 are real and positive, except for the first one, which is
1 = 0. In the hard-wall case we have the important distinction between k > n , where
n is real and the mode is propagating, and k < n , where n is imaginary and the mode
is evanescent (i.e., exponentially) decaying. Propagating modes are described as being
cut on, and evanescent modes are cut off. The frequency = c0 n is called the
cutoff frequency of the mode. For low frequencies, the only cut-on mode is the plane
wave (with cutoff frequency zero):
1 = 1,
1 = 0,
1 = k.
(2.201)
The next eigenvalue 2 is typically of the order of a number between 3 and 4 divided
by theduct diameter D. As a result, any higher modes of a sound field with frequency
< 10 c0 /D decay faster than exp(x/D). At any distance more than, say, two
diameters away from a source, the field is well described by just plane waves.
If the duct cross section is circular or rectangular and the boundary condition is
uniform everywhere, the solutions of the eigenvalue problem are relatively simple and
may be found by separation of variables. These eigensolutions consist of combinations
of trigonometric and Bessel functions in the circular case or combinations of trigonometric functions in the rectangular case. In particular, for a cylindrical duct we have in
polar coordinates (r, ) the spiraling modes
= Jm (mr ) eim ,
m Z,
(2.202)
14:12
P1: JZZ
0521871441c02
70
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where Jm is a Bessel function of integer order m. Note that Jm = ()m Jm . Positive

m correspond to counterclockwise-rotating modes (phase t m = constant) and
negative m with clockwise-rotating modes. With a fixed source it is sometimes clearer
to distinguish between symmetric and antisymmetric modes by using sin(m) and
cos(m ) rather than eim . For a rectangular duct we have

sin( y) sin( y)
=
,
(2.203)
cos( y) cos( y)
2
where
= 2 + 2 . Note that, owing to the symmetry, in both geometries the eigenvalue has multiplicity 2 that is, for each eigenvalue there are two eigenfunctions.
This is not the case for an arbitrary cross section. Some other geometries, like ellipses,
do also allow explicit solutions, but only in special cases such as with hard walls.
For other geometries one has to fall back on numerical methods for the eigenvalue
problem.
Without mean flow, the problem is symmetric, and to each eigenvalue there corresponds a right-running and a left-running mode because both n and n can occur. The
modes form a complete set of basis functions for the solutions to the wave equation.
These modes are not exactly orthogonal to each other, but the complex conjugated
modes (more precisely, the solutions of the adjoint which is here the complex
conjugated problem) are mutually orthogonal or biorthogonal with n . It follows
that, if we specify n = n (the asterisk denotes the complex conjugate), we have

def
n m d = nm
n2 d = nm (n , n ),
(2.204)
(n , m ) =
A
nm = 1 if n = m and is zero otherwise. Note that hard-walled modes are real, and so
n = n ; thus, hard-walled modes are orthogonal. The biorthogonality can be used to
determine the coefficients of an expansion. Assume we have the pressure given along
cross section x = 0 at the right side of a source such that only right-running modes
occur. We have
p (0, y, z) = F(y, z) =
An n (y, z).
(2.205)
n=1
After multiplication of the left- and right-hand sides by m and integration across A,
we obtain
Am =
(F, m )
.
(m , m )
(2.206)
In a similar way we can determine an expansion of the Greens function G for a duct
(Morse and Feshbach 1953) as follows:
G(x,
y, z|, , ) = 12 i

n (y, z)n (, )
n=1
n (n , n )
ein |x | .
(2.207)
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
71
2.5 CONFINED FLOWS
This result can also be obtained, albeit in a more laborious way, by means of Fourier
transformation to x.
Solutions within more complex geometries consisting of piecewise constant pipe
elements can be obtained by matching two series of expansions in modes of adjacent
pipe segments. This procedure involves an approximation by truncation of the mode
expansion. This is quite successfully applied to stepwise changes in pipe cross section (Morse and Feshbach 1953; Kergomard, Garcia, and Taguy 1989; Rienstra and
Hirschberg 2001), the diaphragm (Morse and Feshbach 1953; Rienstra and Hirschberg
2001), the elbow (Miles 1947), the bend (Cabelli 1980), the T-joint (Redmore and
Mulholland 1982), and the tone-hole of woodwinds (Dubos et al. 1999). A subtle point
is that the physical singularity of the sound field at the sharp edges is related to the
convergence rate of the series expansion. It is therefore important, when using such
approximations, to tune the number of modes carefully in the two series expansions
(Rienstra and Hirschberg 2001).
With mean flow, the problem is not symmetric anymore, but for uniform mean
flow in a hard-walled duct this is only a minor obstacle because the problem can be
transformed to an equivalent no-flow problem as we described in Section 2.3.10.1. This
is not possible in the combination of mean flow with soft walls at least when we
use the IngardMyers boundary condition (Ingard 1959; Myers 1980) for inviscid flow
along an impedance wall, where the eigenvalue appears in the boundary condition. We
have different eigenvalues for the left- and right-running modes, and we now write the
following instead of Equation (2.200):
p (x, y, z) =
An n+ (y, z) ein x +Bn n (y, z) ein x .
(2.208)
n=1
Orthogonality cannot be used to determine the amplitudes, but a Greens function can
still be determined by Fourier transformation to the axial coordinate. Furthermore, there
are indications that the system may be unstable for certain impedances with the result
that one seemingly upstream-running mode is to be interpreted as really a downstreamrunning instability (Rienstra 2002).
Modes can also be obtained in more complex situations such as pipes with a main
flow (Eversman 1993; Pagneux and Froelich 2001; Rienstra 2002) or nonrigid walls
(Rienstra 1985). In many applications the prediction of mode propagation allows a
significant insight into sound radiation problems. An example of such an application
is the Tyler and Sofrin (1962) rules for the design of aircraft turbines, minimizing the
number of excited cut-on modes and thus the radiation of sound generated by rotor
stator interaction.
When the duct and its possible mean flow vary in the axial direction, modes that are
strictly self-similar in x are in general not possible. However, if the duct varies slowly
as, for example, in an aeroengine flow duct, we can still identify approximate, almost
modes that retain their shape but have a slowly varying amplitude, eigenvalue, and axial
wave number (Rienstra 1999, 2003). (The approximation is known as a variant of the
14:12
P1: JZZ
0521871441c02
72
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
WKB approximation.) If the duct and mean flow variations scale on a so-called slow
variable X = x, where is small, we can describe a slowly varying mode by
p (x, y, z) = Nn (X )n (y, z; X ) ei
x
n ( )d
(2.209)
where n with n is a mode with wave number at cross section A(X ). The variable
X acts as no more than a parameter. An adiabatic invariant can be identified yielding
the varying amplitude. If the modes are normalized such that (n , n ) = 1 and the
mean flow is nearly uniform (i.e., modulated by the duct), the amplitude Nn is given
for hard-walled ducts by

1/2
= constant,
0 Nn2 k 2 (1 M 2 )n2
(2.210)
where 0 , k = /c0 , and M = U0 /c0 depend on X and eigenvalue n corresponds to the

mode n at position X . For soft walls the expression is similar. When the mode passes
a turning point that is, where k = (1 M 2 )1/2 n the solution breaks down because
the incident mode couples with its back-running counterpart (the mode reflects). A
local analysis is possible to describe this effect (Ovenden 2002, 2005; Rienstra 2003;
Ovenden et al. 2004).
2.5.2 Low-frequency Greens function in an infinitely long uniform duct
At frequencies below the cutoff frequency for higher-order modes, the acoustical field
in a duct is, at some distance from the source, dominated by the plane-wave mode. We
expect, therefore, the Greens function to be independent of the transverse coordinate of
the listeners position. If we use the principle of reciprocity (2.67), this implies also that
the Greens function G should not depend on the transverse coordinate of the source.
Hence, we can use a one-dimensional Greens function g defined by
1 2g 2g
2 = (x3 y3 )(t ).
c02 t 2
x3
(2.211)
A solution of this equation will be nonzero only for t > and of the form f (t x3 /c0 )
for x3 > y3 and g(t + x3 /c0 ) for x3 < y3 . So we try (or see Kanwal 1998) the auxiliary
function

x
x
+ H (x)g t +
,
(2.212)
(x, t) = H (x) f t
c
c
where H denotes the Heaviside step function, and note that it has the properties

x
x
2

t
t
+
+
H
(x)g
,
=
H
(x)
f
t 2
c
c

1
x 1
x
= (x) f (t) (x)g(t) H (x) f t

+ H (x)g t +
,
x
c
c
c
c
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
73
2.5 CONFINED FLOWS
Figure 2.5. (a) A method of images applied to a source at y = (y1 , y2 , y3 ) at a distance y3 from a hard
wall x3 = 0 has a Greens function: G(x ,t|y , ) = G0 (x,t|y, ) + G0 (x,t|y , ) with y = (y1 ,y2 ,y3 ). (b) A
source between two parallel hard walls generates an infinite row of images. (c) A source in a rectangular
duct generates an array of sources.
where we introduced (x) f (t x/c) = (x) f (t) and similarly for g,

2
1
= (x)( f (t) g(t)) (x)( f (t) + g (t))
2
x
c

1
x
x
1
+ 2 H (x)g t +
,
+ 2 H (x) f t
c
c
c
c
and hence satisfies
2
1
1 2
= (x)( f (t) g(t)) + (x)( f (t) + g (t))

c2 t 2
x2
c
= (t)(x)
(2.213)
if f (t) = g(t) = 12 cH (t). By a simple coordinate transformation we find henceforth

that

|x3 y3 |
1
g(x3 , t|y3 , ) = 2 c0 H t
.
(2.214)
c0
A more generic approach may be the use of Fourier transformation in t and x (see
Rienstra and Hirschberg 2001).
This result can also be understood by using the method of images (Pierce 1981).
The acoustical field of a point source placed at a distance d from an infinite plane hard
wall will generate the same acoustical field as the original source and its image placed
in free space (Figure 2.5a). The images position is along the normal to the plane at
the distance d from the plane opposite to the source. The image takes into account the
effect of the waves reflection on the plane.
A source placed between two hard plane walls will generate an infinite row of images
(Figure 2.5b). A source placed in a rectangular duct at y = (y1 , y2 , y3 ) has the same
effect as an array of sources in the (y1 , y2 ) plane, as shown in Figure 2.5c. Each source of
the array generates a pulse G 0 = (t te )/(4r ) with te = t r/c0 . For a given listener
coordinate (0, 0, x3 ) in the duct, we have r 2 = y12 + y22 + (x3 y3 )2 and dte = dr/c0 .
14:12
P1: JZZ
0521871441c02
74
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
The number of images corresponding to an emission time between te and te + dte scales
for fixed t, x, and y3 because d(y12 + y22 ) = dr 2 = 2r dr = 2 c0r dte . Because G 0
scales as r 1 , the increase in number of sources with r exactly compensates the decrease
of G 0 . This results in Heaviside step function behavior for a delta function source
behavior.
Note that, for a source placed between two rigid planes, we obtain a two-dimensional
response intermediate between the delta pulse and the Heaviside step. One observes
a long decay scaling with t 1/2 . Thunder acts qualitatively as a line source with a
two-dimensional acoustical field. This phenomenon explains the long decay time of
thunder sound. This makes acoustics in two dimensions quite complex and different from three- and one-dimensional acoustics (Dowling and Ffowcs Williams 1983).
A two-dimensional acoustical field does not have a simple near-field behavior as a
three-dimensional acoustical field. This essential difference between three- and twodimensional acoustics is a major problem when one considers the sound production by
a two-dimensional model of a three-dimensional flow. Extending the two-dimensional
model to the acoustical far field will dramatically exaggerate radiation losses. Placing
a radiation condition at a finite distance will provide results that depend on the distance
between the boundary and the flow.
In an analogous way, we may find that, in the presence of a uniform subsonic main
flow U0 in the duct, the Greens function, satisfying

2
1
2g
+
U
g 2 = (x3 y3 )(t ),
(2.215)
0
2
x
x
c0 t
is given by

x3 y3
g(x3 , t|y3 , ) =
y3 )H t
c0 + U 0

x3 y3
1
.
+ 2 c0 H (y3 x3 )H t +
c0 U 0

1
c H (x3
2 0
(2.216)
Note that this Greens function satisfies the reverse-flow reciprocity principle (Howe
1975) rather than the reciprocity principle (2.67). When we exchange the source and
listeners positions, we should also reverse the main flow in order to keep the travel time
r/(c0 U0 ) of the waves between the source and the listener constant.
2.5.3 Low-frequency Greens function in a duct with a discontinuity
We want to introduce the concept of low-frequency Greens function by considering the

effect of a discontinuity in an infinitely long duct. We will use the reciprocity principle
(2.67) to determine the Greens function. This implies that, for a source placed at
the discontinuity, we can deduce the Greens function by considering the sound field
generated by a source placed at the position of the listener. For a source placed far
from the discontinuity and a listener placed at the same side of the discontinuity as the
source, the Greens function will be built up of the waves generated at the source plus
the reflection of one of the waves at the discontinuity. A listener placed beyond the
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
75
2.5 CONFINED FLOWS
discontinuity will only be reached by the waves transmitted through the discontinuity.
In other words the problem reduces to the determination of the reflected and transmitted
waves at the discontinuity. We limit ourselves to the case of a compact transition region
x I x3 x I I of cross-sectional area S(x3 ) between two semi-infinite ducts of cross
section S I at the side of the source and S I I at the opposite side. In the transition region we
will now for simplicity assume a quasione-dimensional, incompressible potential flow
v3 (x3 , t). Applying the integral (2.12a) mass conservation law across the discontinuity,
we have for x I x3 x I I
S(x3 )v3 (x3 , t) = S I v3 (x I , t) = S I I v3 (x I I , t).
(2.217)
The equation of Bernoulli (2.44) applied in x I x3 x I I between x I and x3 yields

x3

p I p (x3 , t) = 0 [(x3 , t) I ] = 0
v3 dx3
t
t x I
x3

SI
v3 (x I , t)
= 0
dx3
.
(2.218)
S(x
)
t
3
xI
ikx3
ikx3
ikx3
For harmonic waves p I = p +
+ p
and p I I = p +
we have
I e
I e
II e

ikx I
ikx I
ikx I I
SI p+
p
= SI I p+
I e
I e
II e
and

ikx I
ikx I
ikx I I
ikx I
ikx I
p+
+ p
p+
= ik0 L eff p +
p
I e
I e
II e
I e
I e
with the effective length L eff defined by

L eff =
xI I
xI
SI
dx3 .
S(x3 )
(2.219)
(2.220)
We find for the reflection coefficient R and the transmission coefficient T that
R=
ikx I
p
S I S I I (1 ik L eff )
I e
+ ikx I =
S I + S I I (1 + ik L eff )
PI e
T =
ikx I I
p+
II e
+ ikx I .
pI e
and
(2.221)
These results reduce to the well-known result R = [(S I S I I )/(S I + S I I )] and T =

2S I /(S I + S I I ) in the limit of k L eff 0. For an ideal open pipe end at x I = 0 we find
R = 1 by taking the limit S I /S I I 0 and x I = 0 (Morse and Ingard 1968; Dowling
and Ffowcs Williams 1983). For a closed pipe at x I = 0, we find R = 1 by taking the
limit S I /S I I . When S I = S I I we recover the result for a diaphragm in a pipe
R = ik L eff /(2 + ik L eff ) and T = 2/(2 + ik L eff ).
In the limit k L eff 0 the Greens function g we are looking for is, for x I > y3 ,

c0
|x3 y3 |
g = H t +
2
c0

x3 + y3 2x I
c0
for x3 < x I ,
+ R H t +
2
c0
14:12
P1: JZZ
0521871441c02
76
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
and

x3 y3
c0
g = T H t
2
c0
for
x I < x3 .
(2.222)
Considering the case x I > y3 while we neglect the effect of reflections R = 0 and
T = 1, we can write for k L eff 1 along x I < x3 < x I I

x I + xeff y3
c0
,
(2.223)
g = H t
2
c0
where

xeff =
x3
xI
SI
dx3 .
S(x3 )
(2.224)
The effective coordinate xeff corresponds to the potential difference induced between
x3 and x I by a flow having at x I a unit velocity v I (x I ) = 1.
We can apply the same equations to the problem of an arbitrarily shaped discontinuity
in a pipe if we replace the definition (2.224) of the effective position xeff by the more
general definition
xeff =
(x3 ) I
,
v I
(2.225)
where we do not assume a quasione-dimensional flow at the discontinuity. This corresponds to the low-frequency Greens function proposed by Howe (1975) for a discontinuity in a pipe. This generalization allows one to calculate the effect of a thin diaphragm
in a pipe (Morse and Ingard 1968; Pierce 1981) or the reflection at an elbow (Bruggeman 1987) by using potential theory for incompressible flows. This approach is limited
to compact regions but has the great advantage of providing a detailed model of the
acoustical flow at such a discontinuity even in the neighborhood of sharp edges. Note
that near a sharp edge an expansion of the solution in pipe modes fails to converge. Such
theories are first-order approximations in a matched asymptotic expansion procedure
(Lesser and Crighton 1975). The matched asymptotic expansion procedure permits a
systematic approximation that can be extended to higher order in the small parameter
k L eff . In the present case we have neglected terms of order (k L eff )2 . We take the inertia
of the flow at the discontinuity into account but neglect the effect of compressibility.
The same kind of approximation leads to the low-frequency Greens function for a
compact rigid body in free space proposed by Howe (1998, 2002).
2.5.4 Aeroacoustics of an open pipe termination
2.5.4.1 Introduction to open pipe termination acoustics
Modeling of the aeroacoustical behavior of internal flow involves in many cases an

open pipe termination. This is not only a boundary condition for the calculations of the
internal acoustical field but also a source of sound for the external acoustical field. In the
presence of a main flow or at high amplitudes this involves a complex, unsteady flow at
the pipe termination. This problem is often oversimplified and underestimated. At low
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
77
2.5 CONFINED FLOWS
frequencies, for example, the boundary condition p = 0 is often used. In the presence
of a main flow this corresponds to a quasi-steady response of the free jet outside the
pipe. This model is quite reasonable to describe the reflection of acoustical waves.
When, however, we consider the convection of vorticity, such a model induces spurious
sounds and flow phenomena because vortices cannot be transported through such a
boundary. Also such a model does not predict the spectacular effect of the shape of the
nozzle edges on the aeroacoustical behavior of the pipe termination. Under particular
flow conditions and for specific nozzle shapes, the pipe termination can be a source of
sound. Coupling with acoustical resonances of the pipe system results in whistling. We
give here a short survey of the aeroacoustical behavior of open pipe terminations.
2.5.4.2 Low-frequency linear behavior without main flow
We consider plane harmonic waves propagating in the x1 direction along a duct of

uniform cross-sectional area A:
p = p e it = ( p + eikx1 + p e ikx1 ) e it .
(2.226)
These waves induce a volume flow at the open pipe termination in x1 with an amplitude
V = Av1 = A
p+ p
.
0 c0
(2.227)
The boundary condition at the open pipe termination x1 = 0 for the internal acoustical field can be expressed in terms of a radiation impedance as follows:

p+ + p
p
= 0 c 0 +
.
(2.228)
Zp =
v1 x1 =0
p p
The acoustical power P
radiated at the pipe termination is given by
P
= ( p v1 )x1 =0
A = 12 v1 v1 A(Z p ),
(2.229)
where the asterisk denotes the complex conjugate. The real part of this impedance can
be determined by applying the conservation of acoustical energy (Equation (2.101))
between the pipe exit and the far field outside the pipe. We consider first a thin-walled,
unflanged pipe emerging into free space. In that case the far field will be dominated by
the monopole radiation (Equation (2.84)):
i0 V itikr
i0 Av1 itikr .
e
e
=
(2.230)
4r
4r
The corresponding radial velocity becomes, for kr 1, equal to vr p /(0 c0 ).
Together with Equation (2.230) this gives, for the power P
radiated in the far field,
p
P
= p vr
4r 2
1
1
2
0 c0 k 2 A2 v1 v1 .
p
4r 2 =
0 c0
8
(2.231)
Combining Equations (2.229) and (2.231) with the conservation of acoustical energy
yields
(Z p ) = 0 c0
k2A
.
4
(2.232)
14:12
P1: JZZ
0521871441c02
78
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
When we consider a pipe emerging from a hard wall into a half-infinite free space
(flanged pipe), the radiation energy is distributed over a surface 2r 2 rather than 4r 2 .
Furthermore, the acoustical amplitude is increased by a factor of two owing to reflection of waves at the wall. The combination of both effects results in an increase of
(Z p ) by a factor of two compared with the unflanged pipe termination case (Equation (2.232)). Further confinement of the pipe outlet will further increase the radiation
impedance.
The imaginary part (Z p ) of the radiation impedance takes the inertia of the acoustical flow outside the pipe into account. This effect is often expressed in terms of an
end correction :
(Z p ) = 0 c0 k.
(2.233)
This is the length of a pipe segment that has the same inertia as the outer flow. In first
order approximation in the small parameter k A/ , the reflected wave p seems to be
generated from a reflection of the incoming wave p + , without phase shift, from an ideal
open pipe termination with p = 0 placed at x1 = . The pipe behaves acoustically as
if it were longer by a length if we assume that it is terminated by ideal open ends. In
contrast to (Z p ), which is determined by global flow properties, the end correction
of an open pipe termination is sensitive to the details of the local flow around the
pipe outlet. For a pipe with circular cross section A = a 2 and for infinitely thin walls,
Levine and Schwinger (1948) found in the low-frequency limit 0.61a. For thick
walls the end correction increases gradually from = 0.61a up to the flanged pipe
limit = 0.82 (Ando 1969; Peters et al. 1993; Nederveen 1998; Dalmont, Nederveen,
and Joly 2001). Please note that the order of magnitude of the radiation impedance
of an unflanged pipe is found by considering the low-frequency limit of the radiation
impedance Z s of a compact sphere of radius a. Following Equations (2.82) and (2.83),
we have
Z s = 0 c0
ika + (ka)2
0 c0 (ika + (ka)2 + . . . ).
1 + (ka)2
(2.234)
The simple behavior just discussed is typical for a radiation into a three-dimensional
free field. Confinement will drastically affect the radiation impedance. If we consider,
for example, the radiation of a pipe termination emerging between two closely spaced
parallel hard walls, the radiation field will be two-dimensional. The radiation impedance
has a complex behavior in such a case (Lesser and Lewis 1972). One can not, for
example, identify a locally incompressible near field of the pipe termination. In general
the radiation impedance will be much larger than in a three-dimensional case. It is
quite tempting to use a two-dimensional flow simulation to describe the oscillating
grazing flow along a wall cavity. It is, however, essential to realize that, with this twodimensional calculation, we will dramatically overestimate the radiation losses of the
flow. The results of the calculations will actually depend on the size of the calculation
domain.
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
79
2.5 CONFINED FLOWS
2.5.4.3 High-frequency linear behavior without main flow
For high frequencies, any incident mode of amplitude A, frequency , circumferential

order m, and radial order is reflected in several propagating radial modes. Owing to
cylindrical symmetry, no mode is reflected in another circumferential order. Outside
the pipe we have in the far field
pm (x, r ) 0 c02 ADm ( )
eik
k
(k ),
(2.235)
where x = cos , r = sin , and Dm ( ) is called the directivity function, and

|Dm ( )| is the radiation pattern. Because each mode has its own spiraling phaseplane orientation, the radiated pattern consists of lobes interlaced by zeros. Each zero
is found at the propagation direction of a reflected mode (that is, at = arc sin(m /k)),
except for the direction of the incident mode (that is, at m = arc sin(m /k)). As may
be expected, here we find the radiation maximum or main lobe.
The field inside may be written as

2
itim
im x
im x
+
Rm Jm (m r ) e
Jm (mr ) e
, (2.236)
p(x, r, , t) = 0 c0 A e
=1
where the branch of the square root

2
m
=

2
k 2 m
(2.237)
is chosen such that m is either positive real or negative imaginary. The matrix Rm =
{Rm } is called the reflection matrix. Explicit analytical expressions (in the form
of complex contour integrals) are found by application of the WienerHopf method
(Levine and Schwinger 1948; Crighton et al. 1992).
A subtlety in the solution of the scattering problem is the duct edge. Here, the boundary condition is, technically speaking, not applicable, and without further condition we
have no unique solution owing to possible point or line sources hiding at the edge.
So we have to add the so-called edge condition of finite energy in any neighborhood
of the edge. This amounts to an integrable squared velocity field | p|2 . It transpires
that p varies near the edge like the square root of the distance.
2.5.4.4 Influence of main flow on linear behavior at low frequencies
Flow has a dramatic effect on the radiation impedance Z p of an open pipe termination.
We consider the effect of a uniform subsonic flow velocity U0 in the pipe. In this section
we limit our discussion to low frequencies. By low frequencies we do not only mean that
we limit ourselves to plane-wave propagation in the pipe, but we also assume that the
flow at the pipe termination is locally quasi-steady. This corresponds to a low Strouhal
number limit St = a/U0 1. Because the Strouhal number is related to the Mach
number M = U0 /c0 by St = ka/M, the combination of St 1 with M < 1 implies
that we consider very long wavelength ka 1.
14:12
P1: JZZ
0521871441c02
80
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
The acoustic field in the pipe is described by the convective wave equation

2
2 p
+ U0
p c02 2 = 0.
(2.238)
t
x1
x1
The solution of this wave equation is
p = p + e itik
x1
+ p e it+ik
x1
(2.239)
with
k =
.
c0 U0
(2.240)
The fluctuations v1 in the flow velocity are obtained by applying the linearized momen
+ U0 x1 )v1 = x1 p :
tum equation 0 ( t
v1 =
1 + itik + x1

p e it+ik x1 .
p e
0 c0
(2.241)
Following the vortex sound theory (Howe 2002), the acoustic field is most efficiently
described in terms of fluctuation B = U0 v1 + p /0 of the total enthalpy B = i + v12 /2:
B = B + e itik
with
B =p
x1
+ B e it+ik
U0
1
c0
x1
(2.242)

.
(2.243)
We consider a flow leaving the pipe. As the result of flow separation at the pipe outlet,
a free jet will be formed. It can be demonstrated that the only possible subsonic jet flow
is one in which the pressure is uniform and equal to the pressure in the surroundings:
p = patm (Shapiro 1953). Following this model, we have the boundary condition p = 0
at the open pipe termination x1 = 0. This implies a pressure reflection coefficient
p
= 1
p+
(2.244)
and a total enthalpy reflection coefficient

p c0 U 0
c0 U0
B
=
=
.
+
+
B
p c0 + U 0
c0 + U 0
(2.245)
This result implies a loss of acoustical energy. In the free jet the kinetic energy of the
flow will be dissipated without any pressure recovery. Because an acoustical modulation
of the flow at the pipe termination implies a modulation of the kinetic energy in the
jet, one can expect an absorption of acoustical energy. The vorticity in the shear layers
of the free jet acts as a source of sound. The end correction appears to be determined
by details of the flow just outside the pipe. As we will see in the next section, in the
limit of very low Strouhal numbers for an unflanged circular pipe, 0.2a (Rienstra
1981b, 1983; Peters et al. 1993) which is a rather unexpected result.
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
81
2.5 CONFINED FLOWS
The preceding model corresponds to the very simple boundary condition p = 0 at

the pipe outlet x1 = 0. Although this model appears to be quite accurate in the limit
St 1, it should be used with care. A vortex with a dimension of the order of the
pipe diameter corresponds to a perturbation with a Strouhal number of order unity. A
boundary condition of uniform pressure at a pipe termination will not allow such a
vortex to flow out of the pipe! This is of course in contradiction to the experimental
observations that vortices do leave the pipe through an open pipe termination.
In the case of inflow, we can have a potential flow. In such a case the radiation
impedance for low frequencies will not be strongly affected by the flow. If flow separation occurs, one can consider the use of a quasi-steady model (van Wijngaarden
1968).
It is interesting to note that a diaphragm placed at the pipe outlet can be used as a
nonreflecting pipe termination at a critical Mach number M = U0 /c0 . This anechoic
pipe termination behavior was predicted by Bechert (1980) using a quasi-steady flow
model. Applying the quasi-steady equation of Bernoulli for an incompressible flow
between the pipe termination x1 = 0 with cross section A and the free jet of cross
section A j formed at the diaphragm, we have
1
(U0
2 0
+ v1 )2 + p(0, t) = 12 0 (U0 + v1 )2
A
,
Aj
(2.246)
where we assume that in the jet the pressure remains constant, p j = 0. Substitution of
p = p + + p and v1 = ( p + p )/(0 c0 ) yields after linearization

U 0 A2
1
1
+
2
c0
A
p
,
2j
=
(2.247)
A
p
1 + Uc00 A
2 1
j
which vanishes for U0 /c0 = ((A/A j )2 1)1 . Generalization of this result to arbitrary
subsonic Mach numbers is discussed by Hofmans et al. (2000). Application to the
design of silencers is discussed by Durrieu et al. (2001).
When the Mach number is increased, we will eventually reach a critical flow at
the diaphragm (choked flow, unit Mach number at the diaphragm). In this case the
Mach number of a steady flow in the pipe is imposed by the ratio A/A j of pipe and
diaphragm cross sections. It will in a quasi-steady theory remain constant, independent
of any acoustical perturbation. The condition
c
v1
=
U0
c0
(2.248)
combined with the assumption of an ideal gas behavior ( p / p0 = /0 and c02 =

p0 /0 ) yields (Anthoine, Buchlin, and Hirschberg 2002)
p
2
.
= 0 c0
v1
M( 1)
(2.249)
A more elaborate discussion of the acoustical response of a choked nozzle is provided

by Marble and Candel (1977).
14:12
P1: JZZ
0521871441c02
82
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
2.5.4.5 Influence of main flow on linear behavior at high frequencies
The most complete model available of sound radiation from a flow duct with jet that is
analytically tractable is the following. It is a semi-infinite hard-walled duct with walls
of zero thickness containing inviscid uniform mean flow, density, and sound speed, the
values of which may differ inside and outside the duct or jet. The interface between jet
and outer medium is modeled as a vortex sheet. Its analytically exact but mathematically
rather involved solution of the WienerHopf type was given by Munt (1977, 1990). The
solution contains all the elements of the no-flow solution such as reflection into radial
modes and a radiation pattern with mode-related lobes, but the presence of flow adds
several particular features. Of course, there are geometrical effects like a redirecting
of the radiation pattern and refraction across the jet interface, but there are also two
effects due to coupling with the mean flow.
First, for a velocity different across the vortex sheet, the jet is unstable. This instability is mathematically recognizable by means of a causality analysis in the complex
frequency plane. Second, there is the edge condition, which is even more subtle than
without flow (see Section 2.3.10.3). The condition of finite energy is still necessary to
select physically possible solutions but is now not enough for a unique solution. We still
have a choice corresponding to the amount of acoustic vorticity shed from the trailing
edge. The possibility of vortex shedding is included in the model, but its amplitude
is not yet fixed because it is determined by viscous and nonlinear processes that are
not included. We have to add an extra condition. One such condition is full regularity of the field at the trailing edge, the Kutta condition, which corresponds physically
to the maximum amount of vortex shedding possible. Any other non-Kuttacondition
solution will be singular at the edge, but only one corresponds to no vortex shedding.
Because it is the shed vorticity that excites the jet instability, the strict absence of the
instability is a way to apply the condition of no vortex shedding for jets. If the mean
flow is uniform everywhere, the absence of vortex shedding is most easily typified by
a continuous potential. An alternative way to characterize the Kutta condition is via
the streamline emanating from the edge given by r = a + (h(x) exp(it im)).
Without the Kutta condition, its shape for x 0 is like h(x) = O(x 1/2 ). With the Kutta
condition, it is like h(x) = O(x 3/2 ).
The Kutta condition seems to be the relevant condition for relatively low frequencies and acoustic amplitudes and high Reynolds numbers (Crighton 1985). Therefore,
it will be the condition that is supposed here throughout. As governed by the mean flow
Reynolds number, dimensionless frequency, and amplitude, other conditions (generalized Kutta conditions) are also possible (Rienstra 1981a, 1984).
Physically, the shed vorticity, while moving near the edge of the solid duct wall,
produces some additional sound and thus adds some acoustic energy to the sound field.
At the same time a certain amount of acoustic energy and a certain amount of mean
flow energy are needed for the creation of vorticity. Usually the net profit of energy is
negative (acoustic energy is lost to the hydrodynamic vorticity), but this is not necessary
(Rienstra 1981a).
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
83
2.5 CONFINED FLOWS
Very interesting effects of vortex shedding occur for low frequencies, and thus we
will consider results of Munts solution in more detail in the next section.
2.5.4.6 Frequency dependence of the effect of flow on the
radiation impedance
The experiments by Bechert, Michel, and Pfizenmaier (1978) showed for the first time
a dramatic loss of acoustic energy when a long plane wave from upstream of a jet
exhaust was partially reflected and partially transmitted at the exit. Although only a
part was reflected, practically nothing was recovered in the radiated field outside. This
was explained by Howe (1979) for low Mach numbers by showing that, owing to the
presence of mean flow, the sound field sheds vorticity from the edge in such way that
acoustic energy is converted into the acoustically undetected hydrodynamic energy.
Cargill (1982) then showed that the Munt model (Munt 1977) includes all the reported
effects for any Mach number. This makes the Munt model extremely interesting for
studying various aspects of sound field mean flow coupling.
Let us consider the predictions of the Munt model for plane waves of relatively low
wave number k in a duct of radius a with Mach jet number M j . The feature that is directly
related to the energy absorption is the modulus of the pressure reflection coefficient |R|.
With the Kutta condition, it tends to unity for low Helmholtz number ka, which yields
for the transmitted power in the duct PT (1 + M j )2 |R|2 (1 M j )2 a finite value
in contrast to the vanishing radiated value of O((ka)2 ). The difference is the energy lost
in the vortices.
Apart from the modulus reflection coefficient (Munt 1990), few results are reported
from the Munt solution pertaining to the low-frequency range. Therefore, we present
here some recent results obtained by int Panhuis (2003).
A rather remarkable feature is the behavior of the end correction for a jet without
coflow. We define the end correction as the virtual point just outside the duct exit
against which the plane wave appears to reflect with a free-field boundary condition:
the modulus of the plane-wave pressure | p| attains a minimum. If we write R = r ei ,
then the plane-wave pressure is given by
i
p(x) e
and so
kx
1+M j
r e
kx
i 1+M
i
j

2k
M
x
j
,
| p(x)|2 1 + r 2 2r cos
1 M 2j
(2.250)
which attains its minimum for x = , where

1 M 2j
=
.
a
2ka
(2.251)
(2.252)
A marked difference is the nonuniform limit of the end correction for ka 0 and
M j 0. As long as the Strouhal number ka/M j is large, it converges to the no-flow
14:12
P1: JZZ
0521871441c02
84
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
ka
Figure 2.6. The end correction for no flow (M j = 0) and a little flow (M j = 0.01). Note the nonuniform
behavior for ka 0, M j 0.
value 0.6127, but once ka M j , the Strouhal number becomes small and the behavior changes completely: the final value will be 0.2554. Therefore, results for low
Mach number and low Helmholtz number are usually better presented on the basis of
the Strouhal number. See Figure 2.6, where Munts solution of very low Mach number
is compared with the Levine and Schwinger (1948) results for no flow. This nonuniform
limit was predicted by Rienstra (1981b, 1983) and actually experimentally confirmed
more than 10 years later by Peters et al. (1993).
A pressure reflection coefficient larger than 1 for ka between 0 and, say, 1 was
already reported by Munt (1990). In Figure 2.7 the reflection coefficients together with
the corresponding end corrections are given for various Mach numbers (no flow outside,
no density or sound speed difference). Note that the limiting values for ka 0 of |R|
are unity for any Mach number; the end corrections converge to 0.2554(1 M 2j )1/2 .
Finally, in Figure 2.8 the effect of coflow is shown. When the outer Mach number
Mo varies from zero to the jet Mach number M j , both reflection coefficient and end
correction return to a behavior similar to the no-flow case.
2.5.4.7 Whistling
As determined by the shape of the nozzle, a pipe termination can become a source of
sound. This is a familiar phenomenon. By adjusting our lips and blowing at a critical velocity, we can whistle. Sound production by such a smoothly curved pipe outlet
was first studied in laboratory experiments by Blake and Powell (1986) and by Wilson
et al. (1971). We propose a qualitative model of this phenomenon in terms of vortex
sound theory. The same model has been used to explain the behavior of a whistler-nozzle
(Hill and Greene 1977; Hirschberg et al. 1989; Howe 1998). The model has also been
applied to other configurations in which self-sustained oscillations of the flow can be
described in terms of discrete vortex shedding. Examples of such phenomena are
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
85
2.5 CONFINED FLOWS
1.2
1.1
1
0.1
0.01
0.2
0.3
0.4
0.5
0.6
0.9
0.8
0.7
0.6
0.5
0.4
0.5
1.5
0.01
0.6
0.1
0.5
0.2
0.3
0.4
0.4
0.5
0.6
0.3
0.2
0.1
0.5
1.5
Figure 2.7. Plane-wave reflection coefficient |R| and end correction at jet exhaust without coflow for
M j = 0.01, 0.1, . . . , 0.6.
the edge-tone (Holger, Wilson, and Beavers 1977), the Helmholtz resonator (Nelson,
Halliwell, and Doak 1983; Mongeau et al. 1997; Ricot et al. 2001; Dequand, Luo
et al. 2003), the splitter plate (Welsh and Stokes 1984), the closed side branches
(Bruggeman et al. 1991; Ziada and Buhlmann 1992), the flute (Verge et al. 1997a,
1997b; Dequand, Willems et al., 2003), the wall cavities (Thompson, Hourigan, and
Welsh 1992; Howe 1998), the diffusors (van Lier et al. 2001), and the solid-propellant
rocket engine (Anthoine et al. 2002). Other examples of related self-sustained flow
oscillations are described in many review papers and textbooks: Rockwell (1983),
Rockwell and Naudascher (1978), Blake (1986), and Blevins (1990).
Let us start our discussion by reconsidering the acoustical response of an unflanged
pipe termination in the case of a uniform main flow U0 directed toward the pipe outlet.
The pipe cross section has a radius a. Flow separation at the sharp edges of the pipe
14:12
P1: JZZ
0521871441c02
86
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
1.2
1.1
0.5
0.75
0
0.25
0.9
0.8
0.7
0.6
0.5
0.4
0.5
0.5
1.5
0.65
0.6
1
0.55
0.75
0.5
0.50
0.45
0.25
0.4
0.35
0
0.3
0.25
0.2
1.5
Figure 2.8. Plane-wave reflection coefficient |R| and end correction at jet exhaust with M j = 0.3 and
coflow velocities Mo/M j = 0, 0.25, 0.5,0.75, 1.
outlet results in the formation of a free jet. We consider the reflection of harmonic plane
waves with a frequency f traveling down the pipe in the direction of the main flow.
The acoustical flow uac is defined following the vortex sound theory (Howe 1980) as
the unsteady part of the potential flow component of the flow field. This potential flow
bends around the sharp edges of the pipe outlet, as illustrated in Figure 2.9.
In a potential flow, the centripetal force on the fluid particles, allowing a curved
stream line, is provided by the pressure gradient. The corresponding acceleration is
|uac |2 /R, where R is the radius of curvature of the stream line. This implies that
the pressure decreases toward the interior of the bend. As a consequence the velocity
should increase. This is easily verified in a quasi-steady approximation by applying
the equation of Bernoulli. At a sharp edge, the radius of curvature of the streamlines
14:12
P1: JZZ
0521871441c02
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
87
2.5 CONFINED FLOWS
Figure 2.9. Acoustic flow at a pipe outlet (a) for an unflanged pipe termination and (b) for a horn.
in the potential flow vanishes. This implies a singular behavior. The pressure becomes
infinitely low and the velocity infinitely large.
The singular behavior of the flow at a sharp edge implies that viscous forces can
never be neglected at such an edge. The result is that flow separation occurs, which
induces the shedding of vorticity. The vorticity is such that it compensates the potential
flow singularity. Within the framework of a frictionless theory, this is the so-called
Kutta condition (Crighton 1985). At moderate acoustical amplitudes |uac |/U0 1,
the amount of vorticity generated by the acoustic field will be negligible compared
with the vorticity shedding induced by the main flow. The acoustic field does, however,
trigger an instability of the shear layers of the free jet at the pipe exit. This instability
results in the concentration of the vorticity into coherent structures that can be described
as vortex rings. From experiments it appears that a new vortex starts to be formed at the
pipe outlet at the beginning of each acoustical oscillation period, when the acoustical
velocity turns from pipe inwards to pipe outwards. We call this t = 0. The vortices
travel along the pipe axis with a convective velocity Uc , which is about half the main
flow velocity U0 . The circulation of the vortex ring increases almost linearly in time
as it accumulates the vorticity shed at the edge (Nelson et al. 1983; Bruggeman 1987;
Dequand, Willems et al. 2003). After one oscillation period t = T , a new vortex is shed
and the vortex ring travels farther downstream. Following the theory of Howe (1980),
the acoustical power generated by the vorticity is given by
fc uac = (v)uac .
(2.253)
This theory predicts that the vorticity field will initially absorb energy from the acoustic
field. This seems logical inasmuch as the acoustic field is perturbing the shear layer.
This initial absorption is strong because, near the edge of the pipe exit, the acoustical
velocity is normal to the convective velocity v = (Uc , 0, 0). Furthermore, the acoustical
velocity is large because of the singular behavior near the sharp edge. The theory of
Howe (1980) predicts that, after half an oscillation period T /2 < t < T , the vortex will
produce sound because the acoustic flow is reversed whereas the signs of the convective
velocity and the vorticity do not change. In spite of the growth of the vortex, the sound
production will be weaker than the initial absorption that occurred for 0 < t < T /2.
14:12
P1: JZZ
0521871441c02
88
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
This occurs because the vortex moves away from the edge singularity. Both the magnitude of the acoustical velocity and the angle between the vortex path and the acoustical
streamlines decrease. This discussion has provided a qualitative understanding of the
sound absorption predicted by linear theory in the previous sections.
An important feature of this discussion is that it has stressed that the sound absorption is a balance between the initial absorption for 0 < t < T /2 and the sound
production for T /2 < t < T . This implies that it should be possible to obtain a net
sound production by reducing the initial absorption. This is done by using a pipe with
rounded edges like our lips or a horn. The singular behavior at the flow separation point
is suppressed. Furthermore the convective velocity near the flow separation point is
almost parallel to the potential flow lines of the acoustic field. This reduces the initial
absorption so much that at critical flow velocities sound production is observed. Typically one finds an optimal sound production when the travel time of the vortices along
the lips corresponds to an oscillation period of the acoustic field. Experimentally, one
observes this for a Strouhal number based on the radius of curvature R of the lips given
by St R = f R/U0 = 0.2.
When acoustical energy can accumulate in a resonant mode of the pipe system
upstream of the pipe termination, one can observe self-sustained oscillation, which
is referred to as whistling. This occurs when the sound production is large enough to
compensate for viscothermal losses and radiation losses of the resonator. It is important
to realize that in such a case the oscillation amplitude is limited by nonlinear effects.
This can be an increase of losses at high amplitudes or a saturation of the source. In
the particular case of whistling, the main nonlinear-amplitude-limiting effect is the
saturation of the circulation of the vortex rings. Once the vortex has accumulated all
the vorticity available in the shear layer, its circulation has reached a maximum value.
In the case of a flute, the main nonlinear-amplitude-limiting mechanism is additional
shedding of vorticity at sharp edges such as the labium (Verge et al. 1997a). The
present example also illustrates that sound production does not necessarily involve
impingement of a vortex or shear layer on a sharp edge.
This qualitative discussion indicates that, for Strouhal number of order unity, a
numerical flow simulation of a pipe system should include a model of the dynamical
response of the free jet formed at a pipe outlet. Simple boundary conditions such as
assuming a constant outlet pressure are only reasonable at low Strouhal numbers.
14:12
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Theoretical Background: Large-Eddy Simulation

Pierre Sagaut
3.1 Introduction to large-eddy simulation

3.1.1 General issues
This chapter is devoted to large-eddy simulation (LES) of turbulent flows. The framework is restricted to Newtonian, single-phase, nonreactive fluids without external forcing or coupling as in magnetohydrodynamics. The primary approximations for unsteady
simulations of turbulent flows are the following:
r Direct numerical simulation (DNS), which is based on the direct resolution of the
full, unsteady NavierStokes equations without any additional physical assumptions

or models. To get reliable results, one must represent all the dynamically active scales
of motion in the simulation. This means that the grid spacing x and the time step
t must be fine enough to capture the dynamics of the smallest scales of the flow
down to the Kolmogorov scale, referred to as , and that the computational domain
must be large enough to represent the largest scales. These criteria lead to a high
computational cost, which is responsible for the fact that DNS is nowadays almost
only used for theoretical analysis and accurate understanding of flow dynamics and
is not a brute force engineering tool.
r Averaged or filtered simulations: To reduce the complexity of the simulation (and
then lower the computational effort), a classical technique is to apply an averaging
or filtering procedure to the NavierStokes equations, yielding new equations for a
variable that is smoother than the original solution of the NavierStokes equations
because the averaging or filtering procedure removes the small scales or high frequencies of the solution. Because it is smoother, the smallest scales are no longer
of the order of the Kolmogorov scale but are comparable with a cutoff length scale
> (to be discussed in Section 3.2.2), resulting in a lower complexity. As a
consequence, the averaged or filtered simulation can be interpreted as a DNS of a
new fluid flow with a modified constitutive law. The high frequencies are no longer
captured by the computation, but their action on the resolved scales (through the
nonlinear terms appearing in the equations) must be taken into account via the use of
89
23:56
P1: IBE
0521871441c03
90
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
THEORETICAL BACKGROUND: LARGE-EDDY SIMULATION
a statistical model. The most popular averaging- and filtering-related methods are the
following:
Reynolds-averaged NavierStokes (RANS), which relies on a statistical average.
The ergodic theorem says that procedure can be asymptotically interpreted as
a time-averaging procedure (if the solution fulfills some given requirements),
leading to steady computations in the general case. Unsteady RANS can also
be obtained when the statistical average is related to a conditional or phaseaveraging procedure, or both. Note that the RANS approach does not permit
explicit control of the complexity of the simulation because the cutoff frequency
can not be specified during the averaging procedure.
Large-eddy simulation (LES), which is based on a filtering operator, leading to
unsteady 3D computations. The filtering procedure (to be detailed in Section 3.2)
can be explicit (i.e., associated with the application of a convolution filter to the
DNS solution) or implicit (i.e., imposed by numerical errors, the computational
mesh, or modeling errors) or even a blending of these two possibilities.
3.1.2 Large-eddy simulation: Underlying assumptions
Large-eddy simulation (Rogallo and Moin 1984; Sagaut 2002) is based on a scale
separation, the smallest scales of the exact solution being parametrized via the use of
a statistical model referred to as a subgrid-scale model. The removal of the highest
frequencies is carried out considering the following assumptions, which are derived
from the local isotropy hypothesis formulated by Kolmogorov (1941):
r Large scales of the flow
characterize the flow (i.e., drive its dynamics). In particular, the driving physical
mechanisms, which are responsible for transition to turbulence and production
of the turbulent kinetic energy, are carried by the large scales.
are sensitive to boundary conditions and so are anisotropic.
contain the main part of the total fluctuating kinetic energy (a minimum threshold
of 80 to 90% is commonly agreed).
r Small (subgrid) scales of the flow
have a universal character and are isotropic. This corresponds to the local isotropy
hypothesis of Kolmogorov.
are only responsible for the viscous dissipation. No driving mechanism should
be associated to that range of scales.
are weak. They contain only a few percent of the total kinetic energy.
If these hypotheses are verified, one can expect that retaining only the large scales of
the flow will make it possible to reduce the cost of the computation while still capturing
the desired features of the flow. It is important to note that these features of large and
small scales can be seen as a physical definition of these two families of scales of
motion. This definition is to be compared with the previous one, which is based on the
definition of an arbitrary cutoff length . These two definitions are not contradictory
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
91
3.2 MATHEMATICAL MODELS AND GOVERNING EQUATIONS
in practice because experience shows that LES yields satisfactory results when is
chosen so that it is compatible with the physical criteria.
3.2 Mathematical models and governing equations

3.2.1 The NavierStokes equations
The discussion here is restricted to the flows of compressible and incompressible Newtonian fluids governed by the NavierStokes equations. Effects of stratification, buoyancy,
combustion, or magnetic forces introduce new physical phenomena and characteristic
scales, which will not be considered here. It is worth noting, however, that they increase the simulation complexity in degree rather than kind. For compressible flows,
the resulting mathematical model is
+ (u) = 0
t
(3.1)
u
+ (u u + p) = v
(3.2)
t
E
(3.3)
+ (u(E + p)) = (uv ) qT ,
t
where , u, and E are the density, the momentum, and the total energy, respectively.
The symbol refers to the tensorial product that is, (u v)i j u i v j . The viscous
stress tensor and the heat flux are referred to as v and qT , respectively. The system is
closed assuming that the fluid is a perfect gas: p = RT , T being the temperature and
R the perfect gas constant (R = 287.6 J kg1 K1 for air). If temperature-dependent
viscosity and diffusivity are considered that is, = (T ) and = (T ) then the
viscous stresses are nonlinear functions of the variables. Viscous stresses and heat flux
are expressed as follows using the Stokes assumption:

2
(3.4)
v = u + T u ( u)Id
3
qT = T,
(3.5)
where Id is the identity tensor. Temperature-dependent viscosity and diffusivity are

usually computed using the Sutherland empirical law:

T 1 + S/273.15
(3.6)
(T ) = (273.15)
273.15 1 + S/T
with (273.15) = 1.711 105 kg m1 s1 and S = 110.4 K (for air). For temperatures
lower than 120 K, the Sutherland law is extended as
T
.
(3.7)
120
Diffusivity (T ) is obtained assuming the molecular Prandtl number Pr constant. Its
value is 0.7 for air at ambient temperature.
(T ) = (120)
23:56
P1: IBE
0521871441c03
92
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
In the incompressible case, the system (3.1)(3.3) simplifies as

u =0
u
+ (u u) = p + 2 u.
t
(3.8)
(3.9)
The energy is totally decoupled from the continuity and momentum equation, and
thus the temperature can be treated as a passive scalar if stratification and dilatational
effects are neglected:
T
+ (uT ) = 2 T.
t
(3.10)
3.2.2 The filtering procedure

3.2.2.1 Definition
As previously mentioned in Section 3.1.1, the effective filter observed in practical LES
computations originates from very different sources. This effective filter appears as a
mixture of the following elements:
r The grid filter: the choice of a computational grid with mesh size x induces the
definition of the maximum resolution of the computation. No scale smaller than the
mesh size can be captured. According to the Nyquist theorem, the corresponding
cutoff wave number is kc = /x on uniform grids.
r The theoretical filter, which is the filter to be applied to the exact solution of
the NavierStokes equations to smooth it. This filter has a characteristic cutoff
length .
r The numerical filter: the numerical error, which is not uniformly distributed over
the resolved frequencies, can also be interpreted as a filter. When local numerical
methods such as finite element, finite volume, or finite difference methods are used
to solve the governing equations, the numerical error is observed to be an increasing
function of the wave number. Consequently, the dynamics of the highest frequencies
resolved on the computational grid are only poorly captured, and these scales can
be considered as being filtered.
r The subgrid model filter (Magnient, Sagaut, and Deville 2001; Mason and Brown
1994; Mason and Callen 1986; Muchinsky 1996): During the computation, the only
term that contains information related to the convolution filter is the subgrid model
(through the use of the filter length scale to define the subgrid viscosity, for
example). But the exact filtered velocity is the solution of the filtered equations
including the exact subgrid tensor, and we have to account for the fact that subgrid
models are not exact, and introduce new errors, which modify the original filter.
The theoretical filter is then transformed into a subgrid model filter.
Very little is known at the present time about the numerical filter and the subgrid
model filter. The emphasis is put here on the theoretical filter applied to the continuous
NavierStokes equations. Leonard (1974) proposes to model it as the application of a
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
93
convolution filter to the exact solution. The filtered part u of the variable u is defined
by the following convolution operator (denoted by the symbol
hereafter):
t +
u(x, t) =
G(, , |x x |, t t )u(x , t )d x dt
= G(, )
u(x, t),
(3.11)
where G(, , |x x |, t t ) is the kernel of the filter. The two arbitrary parameters
and are the cutoff length and the cutoff time, respectively. It can be proven that
G must depend on the distance |x x | to preserve certain symmetries of the Navier
Stokes equations, such as the Galilean invariance (Fureby and Tabor 1997; Ghosal
1999; Oberlack 1997). The small-scale subgrid part u is then defined as
u (x, t) = u(x, t) u(x, t).
(3.12)
From the definition of u, one can see that u will be different for each filter kernel.
For the sake of simplicity, we restrict ourselves here to the case of an isotropic filter.
We recall that a filter is said to be isotropic if both its mathematical expression and
cutoff scales do not depend on the orientation of the Cartesian axes and the location in
space. This implies that the domain is not bounded and that and are constant in
space and time. The extension for space-dependent cutoff scales and bounded domains
is presented in Section 3.2.5.
Almost all authors have considered spatial filtering only, resulting in G = G(, |x

x |). Eulerian time-domain filtering has recently been revisited by Pruett (2000), who
considered causal filters of the form G = G( , t t ).
3.2.2.2 Properties
To obtain tractable governing equations for LES, we assume that the filtering operator
satisfies the following conditions:
r Linearity:
u + v = u + v.
(3.13)
This property is automatically satisfied because we are considering the convolution

filter (see Equation (3.11)).
r Preservation of the constant:
+
a = a
G(, |x x |)d x = 1, x.
(3.14)
r Commutation with derivatives:
, G
= 0,
s
s = x, t,
where the commutator is defined as [a, b](u) = a b(u) b a(u).
(3.15)
23:56
P1: IBE
0521871441c03
94
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Table 3.1. Examples of usual spatial convolution filters

Name
Gaussian filter
Sharp cutoff filter
Box/top-hat filter
G(, |x x |)

2
2
6/ exp(6|x x |2 / )
sin(|x x |kc )/(|x x |kc )
1/ if |x x | /2
0
otherwise

G(k)
2
exp(k 2 /24)
1 if k kc
0 otherwise
sin(k/2)/(k/2)
kc = /
A difference with the RANS approach appears here such filters, in the general case,
are not Reynolds operators (i.e., they do not belong to the family of the projector
operators):
u = u,
u = 0.
(3.16)
Note that projective filters can not be inverted because they are associated with an
irreversible loss of information, whereas some filters verifying relation (3.16) can be
inverted. In the latter case, the filtering procedure is nothing but a change of variables.
The associated filtering operator in Fourier space can be written as
u (k, t),

u (k, t) = G(k)
(3.17)
are the Fourier transforms of u and G (i.e., the transfer function of

where
u and G(k)
the filter), respectively.
Expressions for three filters of common use are given in Table 3.1. The Gaussian
filter possesses the two interesting properties of being quickly decaying and positive
in both physical and Fourier space. The Gaussian and the box filter are called smooth
filters because they do not lead to a sharp scale separation between resolved and subgrid
scales in the Fourier space, resulting in a spectral overlap. This property will be of great
importance for certain modeling strategies. The sharp cutoff filter is a projective filter,
whereas the Gaussian and the box filters satisfy Equation (3.16). Typical spectra of the
resolved and subgrid scales are displayed in Figure 3.1.
3.2.2.3 Other filtering procedures
Other filtering procedures, based on mixed spatial and time-domain filters (Dakhoul
and Bedford 1986a, 1986b) or the projection of the exact solution on a finite basis
of continuous functions have also been proposed (Hughes et al. 1998) but will not be
detailed here because their use is less common. We will only mention the case of the
scale selection through projection of the exact continuous solution onto a finite basis
{i (x, t)}, i = 1, . . . , N because such a projection is the cornerstone of all discretization
methods based on Galerkin-type procedures. If we restrict ourselves to the common
It is worth noticing that Osborne Reynolds, in his original 1884 paper, used a volume average that is
equivalent to the so-called box filter in the LES framework.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
95
Figure 3.1. Schematic kinetic energy spectra

of resolved and subgrid scales with spectral
overlap (Gaussian filter).
case of time-independent basis functions, the exact solution u is transformed into

u i (t)i (x),
(3.18)
u d (x, t) = d (u)(x) =
i=1,N
where u i (t) are the coefficients of the decomposition and functions i (x) may be chosen
among a wide class of functions, ranging from piecewise constant functions (finite
volume approach, finite element) to trigonometric functions (Fourier method) and highorder polynomials (pseudospectral methods) and proper orthogonal disposition (POD)
modes (PODGalerkin methods). The filtering procedure here is associated with the
choice of the projection basis. The transfer function of this implicit filter is formally
written as

=
l (x), e2ikx ,
(3.19)
G(k)
l=1,N
where the , is related to an ad hoc scalar product in the domain . The dimension of
the projection basis being finite, we have |G(k)| 0 when |k| , enlightening
the existence of a filter.
3.2.3 Governing equations for LES

3.2.3.1 First example: Generic conservation law
We first present the case of a generic nonlinear conservation law to show how LES
governing equations are derived from their original unfiltered counterpart. Let us consider the conservation law
u
+ F(u, u) = 0,
t
(3.20)
where the nonlinear flux function F is assumed to exhibit a quadratic behavior with
respect to u. Applying the convolution filter to Equation (3.20), one obtains
u
+ F(u, u) = 0.
t
(3.21)
23:56
P1: IBE
0521871441c03
96
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Using the commutation property (3.15), we find that this relation simplifies as
u
+ F(u, u) = 0.
t
(3.22)
The last step consists in rewriting the filtered nonlinear flux F(u, u) as a function of
the new filtered unknown, u. This is achieved writing u = u + u and inserting this
decomposition into the nonlinear term, yielding

u
(3.23)
+ F(u, u) = F(u, u ) + F(u , u) + F(u , u ) ,
t
where all the terms that can not be directly computed from the known quantity u have
been grouped in the right-hand side. These terms can not be exactly computed because
they explicitely depend on u , which is not contained in the LES problem inasmuch as it
corresponds to scales that have been eliminated. Thus, the terms will be approximated
through the use of a subgrid model, which is a function of u.
3.2.3.2 Incompressible flows
Applying a convolution filter to the NavierStokes equations for incompressible flows,

and taking into account all the properties of the filter, one obtains the governing equations for LES:
u
+ (u u) = p + 2 u
t
u = 0.
(3.24)
(3.25)
This set of equations can not be directly used because the nonlinear term u u must
first be decomposed as a function of the only acceptable variables, which are u and u .
Two decompositions of the nonlinear term are now presented.
3.2.3.3 Leonards decomposition
A first decomposition was proposed by Leonard (1974). It is obtained by inserting the

decomposition u = u + u into the nonlinear term, yielding
u u = (u + u ) (u + u )
= u
u + u u
+ u u +
resolved
C:Cross terms
(3.26)
u
u.
(3.27)
R:Reynolds stresses
The resolved term can be further decomposed as

u u = u
u + (u u u u) .

new resolved
(3.28)
L:Leonard stress tensor
The decomposition procedure makes three tensors appear as follows:

r The Leonard stress tensor L, which corresponds to the fluctuations of the interac-
tions between resolved scales (zero for RANS).
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
97
Table 3.2. Various decompositions for the nonlinear

terms
Form of the resolved convection term
Expression for
uu
uu
C+R
L +C + R
r The cross-stress tensor C, which accounts for direct interactions between resolved
and unresolved scales (zero for RANS).
r The subgrid Reynolds stress tensor R, which is associated to the action of subgrid
scales on the resolved field (Reynolds tensor for RANS).

Two possibilities arise for the definition of the subgrid-scale tensor that depend on
the choice of the formulation of the resolved convection term (see Table 3.2). These two
decompositions can be used, but they introduce some interesting conceptual problems.
Consider the basic philosophy of LES: filtered equations are derived, and classical
wisdom tells us that all the terms appearing in the equations must be filtered terms
(i.e., appear as the filtered part of something). Only the first decomposition = R + C
satisfies that condition. This is especially true when the filtering operator is associated
with the definition of a computational grid (and overbar just means defined on the
grid): the convection term is computed on the same grid as the filtered variables and
then should be written as u u.
3.2.3.4 Germanos consistent decomposition
Another possibility is to build a decomposition based on the generalized central moments of Germano (1986, 1987, 1992). The generalized central moment of two quantities u and v, for the filter kernel G, is defined as
{u, v}G = G
(uv) (G
u)(G
v) = u v u v.
(3.29)
Using that definition, we can write generalized subgrid tensors as

L = {u, u}G ,

C = u, u G + u , u G ,

R = u, u G .
(3.30)
It can be easily verified that

L + C + R L + C + R.
(3.31)
Germano named this decomposition the consistent decomposition because all the terms
can be recast into the generalized central moments formalism. One can verify that each
term of that decomposition is Galilean invariant.
23:56
P1: IBE
0521871441c03
98
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
3.2.4 Extension for compressible flows

3.2.4.1 Definition of the filtered variables
We now discuss the extension of the filtering procedure, as previously defined for incompressible isochoric flows, to variable density and compressible flows. The definition
of the filtered set of equations results from three preliminary choices:
r The original set of unfiltered variables or equations.
r The filter (same low-pass as in the incompressible case).
r The set of filtered variables.
The first problem arises because a very large number of formulations of the compressible NavierStokes equations can be found that rely on the choice of basic variables for describing the flow. Velocity is usually described using velocity or momentum,
but mass conservation and energy equations can be expressed using two independent
variables among a large set: density, entropy, total energy, internal energy, enthalpy,
and so on. The last point, which is the most difficult one when dealing with compressible flows, will be detailed now. Let us consider the set of the conservative variables (, u, E). A straightforward use of the filtering procedure for incompressible
flows leads to the definition of the filtered variables (, u, E). The term u can be
rewritten as
u
u,
(3.32)
where
u = u/ is the mass-weighted filtered velocity. It is worthwhile noting that the
operator can not be interpreted as a convolution-like filtering operator, and it does
not commute with derivatives. It corresponds only to a change of variable.
Here appears the problem of the choice or definition of the basic set of variables
for the LES of compressible flows. It arises because u, E, and p are nonlinear functions of the other variables and then can be decomposed. On the basis of that choice,
different subgrid terms will arise from the filtered equations, inducing a need for specific subgrid modeling. There is an infinite number of possibilities for the definition
of the variables of the filtered problem, starting from (, u, E). Two examples are as
follows:
r System I: as proposed by Comte and Lesieur (Ducros, Comte, and Lesieur 1996)
u , E) supplemented by the
and Vreman (1995), the set of filtered variables is (,
macropressure and the macrotemperature . These two quantities are defined as
1
= p + kk ,
3
2
+ ( 1)M kk ,
=T
2
(3.33)
It is worth noting that this mass-weighting approach was proposed by Osborne Reynolds in his 1884
paper!
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
99
where M is the Mach number. The associated filtered equation of state is

3 5
(3.34)
kk .
6
The uncomputable term kk requires a priori an adequate model. But some authors
2
p, where Msgs is
propose to neglect it, arguing that it can be recast as kk = Msgs
the Mach number based on the characteristic velocity scale of the subgrid motion,
which is generally very small compared with the mean flow Mach number M. Comte
and Lesieur also argue that = 5/3 for monoatomic gases like argon or helium,
leading to a vanishing contribution of the subgrid term in the filtered equation of
state.
r System II: as proposed in references (Sreedhar and Ragab 1994; Vreman 1995;
suppleu , E)
Vreman, Geurts, and Kuerten 1997), the selected variables are (,
.
mented by the filtered pressure p and the mass-weighted filtered temperature T
is defined as
The synthetic total energy E
= R
=
E
p
u2
+
1
(3.35)
and corresponds to the computable part of the total energy. Using this set of variables,
we find that no subgrid contribution appears in the filtered equation of state.
The choice of a system of governing equations is seen to change the closure problem
by modifying the subgrid terms. The total complexity of the system being conserved,
the choice depends essentially on the preference of the practioners. In the previous
examples, system I will not give direct access to the pressure and the temperature,
whereas system II does. Consequently, system II will be preferred for applications
dealing with aeroacoustics and aero-optics. But system I is known to yield simpler
energy equations and is more easily implemented.
3.2.4.2 Example of filtered equations
We now illustrate the problem by writing the set of filtered equations corresponding to
system II. Simple algebra yields
u ) = 0,
+ (
t
(3.36)

u
+ (
u
u) + p
v = A 1 + A 2 ,
t
(3.37)

E
+ p)) (
u
v ) +
qT
+ (
u( E
t
= B1 B2 B3 + B4 + B5 + B6 B7 ,
(3.38)
where
v and
qT are the viscous stresses and the heat flux computed from the resolved variables and are equal to their respective filtered counterparts if the viscosity
23:56
P1: IBE
0521871441c03
100
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
and the diffusivity are assumed to be constant. The subgrid terms A1 , A2 , B1 ,. . . , B7

are defined as
u
u
u ),
(u
B1 =
A1 = ,
( pu p
u ),
B3 = (
u ),
(3.39)
B2 = p u p
u,
B4 =
u,
u
v
u ),
B6 = (v
A2 = (v
v ),
B5 = v u v
u,
B7 = (qT
qT ).
(3.40)
Terms A2 , B6 , and B7 arise from the possible dependency of molecular viscosity and
diffusivity on the temperature and correspond to a new nonconvective source of nonlinearity. The use of the mass-weighthed filtered variable
u instead of the filtered velocity
u prevents the appearance of subgrid terms in the continuity equation.
3.2.5 Filtering on real-life computational grids
Only isotropic filters have been considered in Section 3.2.2. These filters can be defined,
only on unbounded domains; on bounded domains, the filter kernel must be modified
near the boundaries in order to remain definite, resulting in a violation of the homogeneity property. Another problem is that isotropic filters have a constant cutoff length
scale and are not appropriate to describe what is occuring on general curvilinear grids.
This lack of homogeneity will result in the loss of the property of commutation with the
spatial derivatives, and the filtered NavierStokes equations, as written at the beginning
of this chapter, are no longer valid.
The general form of the commutation error for space-dependent cutoff length filters
in a bounded fluid domain can be written in the following compact form (Fureby and
Tabor 1997):

G
, G
() =
+
G((x), x x )(x )n(x )ds,
(3.41)
x
x

where (x) is a dummy function and n(x) is the outward unit normal vector to the
boundary of . The first term on the right-hand side arises from the fact that varies
in space, and the second term comes from the boundedness of the domain.
Two possibilities arise to derive the filtered NavierStokes equations on general
meshes:
(i) Apply the filter to the basic equations written in Cartesian coordinates.
(ii) Rewrite the filtered equations in general coordinates to solve them on general grids.
or
(i) Write the NavierStokes equations in general coordinates.
(ii) Apply the filter.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
101
These two possibilities yield very different sets of governing equations for LES.
They are described in the following section.
3.2.5.1 Filtering the equations in Cartesian coordinates
This solution was first proposed by Ghosal and Moin (1995), who addressed the problem
of the definition of a filtering procedure well suited for the case of a variable cutoff length
scale filter on a bounded domain [a, b]. In this section, Greek symbols are related to
coordinates in an unbounded reference domain [, +] (with uniform mesh spacing
), whereas Roman symbols refer to coordinates in the physical (bounded) domain.
Let f be the coordinate transformation
= f (x).
(3.42)
This function is monotone and differentiable (with f d f /d x), and

f (a) = ,
f (b) = +.
(3.43)
Using that change of variable, one can associate a variable cutoff length scale (x)
(defined on the physical domain) to the constant length scale (defined in the reference
space):

.
(3.44)
f (x)
Let (x), x [a, b] be the function to be filtered. The next step is the definition
of the filtering procedure on the physical domain. This is done using the following
four-step algorithm:
(x) =
(i) Redefine the convolution filter (in the isotropic case, i.e., on the reference
space) as

1 +
( ) =
G
()d.
(3.45)

(ii) Operate the inverse change of variable x = f 1 ( ) and define ( ) = ( f 1 ( )).
(iii) Apply the isotropic filtering procedure (3.45) to ( ):

1 +

(x) ( ) =
G
()d.
(3.46)

(iv) Come back to the physical space (operate the inverse change of variable):

1 b
f (x) f (y)
(y) f (y)dy.
(x) =
G
(3.47)
a

This procedure leads to the definition of second-order commuting filters (SOCF),
2
whose commutation error with spatial derivatives is O( ). It means that the classical,
filtered NavierStokes equations, which were given in the previous sections, obtained
using the commutation property, are only a second-order accurate approximation of
the true filtered problem. One can demonstrate that a straightforward extension of the
classical filtering procedure yields a zeroth-order accurate method.
23:56
P1: IBE
0521871441c03
102
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Vasilyev, Lund, and Moin (1998) have recently demonstrated that this method can
be extended to arbitrary orders of accuracy by choosing a filter kernel G with an ad hoc
number of vanishing moments. More precisely, the commutation error will be of order
n
O( ) if
(G, 0) = 1,
(G, k) = 0,
(3.48)
1 k n,
(3.49)
(G, k) < , k n
with
(3.50)

(G, p) =
p G( )d.
(3.51)
3.2.5.2 Filtering the equations in general coordinates
The incompressible NavierStokes equations written in generalized coordinates read

as follows:
(J 1 ik u i ) = 0,
k
1
1
(J u i ) + k (U k u i ) = k (J 1 ik p) +
t
Re k
(3.52)
J 1 G kl
(u
)
,
i
l
(3.53)
where k are the coordinate directions in the transformed space, ik = k / xi , J 1 is

the Jacobian of the transformation, G kl = ik il denotes the contravariant metric tensor,
u i the Cartesian components of the velocity field, and U k = J 1 ik u i the contravariant
flux across the k = constant plane.
Jordan (1999) proposed to operate the filtering in the transformed plane, leading to
the following set of governing equations for LES:
(J 1 ik u i ) = 0,
k
1
1
(J u i ) + k (U k u i ) = k (J 1 ik p) +
t
Re k
(3.54)

J 1 G kl l (u i ) ,
( k )
k i
(3.55)
where the contravariant counterpart of the subgrid tensor is defined as

ik = J 1 kj u i u j J 1 kj u j u i = U k u i U k u i .
(3.56)
The metrics being computed by a discrete approximation in practice, they can be considered as filtered quantities, yielding
U k = J 1 kj u j J 1 kj u j .
(3.57)
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
103
Jordan (1999) proved that the commutation error vanishes if the variables are first
transformed into computational space before being filtered because a uniform kernel
can be defined in this space.
For compressible flows, one obtains the following set of equations (it is assumed
that metrics are filtered quantities):

1
J + k J 1 kj
u j = 0,
(3.58)
t

1

u j + pi j
J
u i + k J 1 kj
ui
v,i j = A1 + A2 ,
(3.59)
t

1

+ p)
J E + k J 1 kj
ui
v,i j +
qT, j
u j (E
t
= B1 B2 B3 + B4 + B5 + B6 B7 .
The subgrid terms A1 , A2 , B1 , . . . , B7 are now defined as

v,i j ,
A1 = k J 1 kj i j , A2 = k J 1 kj v,i j

u

u
1
1 k
j
j
B1 =
J j pu j p
u j , B2 = J 1 kj p k p k ,
1 k

u

u
1 k
j
j
u i k J j i j , B5 = J 1 kj v,i j k v,i j k ,
B3 + B4 =

ui
u i , B7 = k J 1 kj qT, j
B6 = k J 1 kj v,i j
v,i j
qT, j .
(3.60)
(3.61)
(3.62)
(3.63)
(3.64)
We observe that the set of filtered equations in generalized coordinates is formally

similar to the previous one, the main difference being that metrics are now intrinsically
included in the definition of the problem. As a consequence, the derivation of physical
subgrid models is more difficult (What is the physics of contravariant velocity components?), and the subgrid models used in practice are derived from subgrid models built
using the set of filtered NavierStokes equations in Cartesian coordinates.
3.2.5.3 What is the filtering length scale on general meshes?
Many subgrid models explicitly account for the filtering length scale . In most of the
computations, is computed as a function of the mesh topology.
For Cartesian grids, with relatively isotropic grids, the most popular formula is
= (xyz)1/3 .
For more anisotropic grids, other empirical values are

= Max(x, y, z) or = (x 2 + y 2 + z 2 )/3.
(3.65)
(3.66)
For curvilinear meshes, a commonly used rough estimate is

= (V )1/3
(3.67)
23:56
P1: IBE
0521871441c03
104
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
for finite volume methods (where V is the volume of the control volume) and
= (J )1/3
(3.68)
for finite difference methods (where J is the Jacobian, and , , the grid spacing
in the reference space).
There is no general theory for defining the best approximation, and the choice is
very often made to render the implementation as simple as possible. The optimization
of the LES solution is then carried out by changing the value of the constant in the
subgrid model and redesigning the computational grid (i.e., making it as isotropic as
possible).
3.2.5.4 Discrete test filters
Many techniques and models related to LES involve the use of a second filtering level,
which is applied to the resolved field. During the computation, that filter, referred to as
the test filter, is approximated using some weighted linear combinations of the values
of the resolved field.
The two most popular one-dimensional discrete test filters are the following:
1
(k+1 + 2k + k1 ),
4
1
k = (k+1 + 4k + k1 ).
6
k =
(3.69)
(3.70)
Two- and three-dimensional filters are then obtained using a one-dimensional test
filter in eachdirection. Note that the characteristic length scales of these filters are not
equal: = 6x for the first one, and = 2x for the second one. This difference
must be accounted for when using the dynamic procedure, the best results being obtained
are the cutoff lengths of the LES
= 2 (Najjar and Tafti 1996), where and
with
filter and the discrete test filter, respectively.
The numerical experiments have shown that the subgrid models that rely on a test
filter are almost all very sensitive to its discrete features. Another point when one deals
with the dynamic procedure is that the best results are generally obtained when using
test filters restricted to the homogeneous directions of the flow (but if at least two
periodic directions are available). The use of three-dimensional test filters may reduce
the efficiency of the dynamic procedure.
The two filters presented above can be used on structured meshes. Similar filters can
be designed on unstructured meshes by discretizing the differential operator associated
with the targeted filter or by solving a linear system to compute the weighting coefficients
of the neighboring points. Optimized discrete filters can also be derived, which yield
the best least-squares approximation of a required transfer function over a frequency
band (Sagaut and Grohens 1999; Vasilyev et al. 1998).
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
3.3 BASIC NUMERICAL ISSUES IN LARGE-EDDY SIMULATION
Table 3.3. Resolution requirements referred to Kolmogorov

length scale used in DNS based on spectral methods of some
incompressible homogeneous and wall-bounded flows
Flow
Resolution in a
Boundary layer
Homogeneous shear
Isotropic turbulence
x 15, y 0.33, z 5b
x 8, y 4, z 4
x 4.5, y 4.5, z 4.5
at the wall is used for the wall-bounded flows.

x, y, and z refer to the streamwise, wall-normal, and spanwise direction,
respectively.
Source: Adapetd from Moin and Mahesh (1998).
3.3 Basic numerical issues in large-eddy simulation

Both DNS and LES have been implemented using all known numerical approaches,
including vortex methods and Lattice Boltzmann methods. This section is devoted to
basic issues that apply for all simulations based on finite difference, finite element,
finite volume, and spectral (including spectral element) methods.
This section is devoted to the presentation of two very important issues: the required
grid resolution for obtaining reliable results, and the effects of numerical errors. In both
cases, the results dealing with DNS will be recalled.
3.3.1 Grid resolution requirements

3.3.1.1 Homogeneous and free-shear flows
As mentioned in Section 3.1.1, the main constraint dealing with resolution in DNStype computations is that all relevant scales of the flow must be directly captured.
This means that the computational domain must be large enough to accommodate the
largest scales of the flow and that the mesh size must be small enough to represent
the smallest dynamically active scales. In practice, the former condition means that
the two-point correlation of the fluctuations must vanish within the domain in each
direction of space. The latter corresponds to the fact that the viscous dissipation of
kinetic energy must be captured. The viscous dissipation takes place in the wave-number
band0.1 k 1, corresponding to a length-scale band of about 6 to 60, where
= 3 / is the Kolmogorov scale. Here, is the viscous dissipation rate of kinetic
energy. As a consequence, the resulting criterion for DNS of turbulent flows is x .
This is also equivalent to requiring that the wormlike structures observed in isotropic
turbulence, whose typical diameter is 410, must be correctly represented on the grid.
Examples of grid resolution for incompressible flows are displayed in Table 3.3.
The computational effort required for DNS computations can be estimated from a
priori knowledge about the physics of the flow.
For isotropic turbulence, the ratio between the integral scale (characteristic of large
3/4
scales), L, and the Kolmogorov scale evolves as L/ Re L , where Re L = LU/ is
105
23:56
P1: IBE
0521871441c03
106
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006

the Reynolds number based on L and U = 12 u i u i . Thus, the number of grid points,
N , required to perform DNS of isotropic turbulence in a box of volume L 3 scales as
9/4
Re L . Let TL be the integral time scale of turbulence and T = / the Kolmogorov
time scale. The resolution requirement is t T . Thus, the minimal number of time
1/2
steps is proportional to TL /T Re L . The minimum complexity of the simulation
that is, the number of times the NavierStokes equations should be solved is then of
11/4
the order of Re L . Similar scaling laws can be derived for free-shear flows without
shocks.
The basic philosophy of LES is to use coarser grids than DNS while still correctly
capturing the large-scale dynamics. As mentioned in Section 3.1.2, large scales can
be physically defined as scales responsible for turbulent fluctuation mechanisms. Most
subgrid models are efficient in representing the turbulent energy cascade only. Thus,
the LES computational grid must be fine enough to resolve production of turbulence
and transition to turbulence directly.
For free-shear flows and separated-shear layers, this constraint is weakened because
turbulence production is associated with scales whose size is of the order of the shear
layer thickness . Consequently, a typical mesh size is /100 x /10, irrespective
of the Reynolds-number value.
3.3.1.2 Wall-bounded flows
Wall-bounded flows induce more drastic constraints (Moin and Mahesh 1998; Piomelli
and Balaras 2002; Rogallo and Moin 1984). Large scales are of the order of the boundary
layer thickness in the outer layer of the boundary layer, and the relevant scaling for the
turbulent motion responsible for the turbulence production in the inner layer is the
viscous length /u , where u = w / is the friction velocity. In incompressible

boundary layers, channel flows, and pipe flows we have u U C f and C f Re ,
where C f = 2w /U 2 and U are the skin friction coefficient and the outer velocity
scale of the flow, respectively. Experimental data yield 0.2 0.25. The streaky
structures (Figure 3.2) observed in the inner layer have constant spatial characteristics
in wall units, and therefore the grid spacing must be of constant size in this coordinate
system. Therefore, the number of grid points in each direction scales as
/2
N=
L
LU Re L
L
=
+
x
x /u
1/2
Re L
3(1/2)
(3.71)
Re2.625
Re2.7
The total number of grid points is then proportional to Re L
L
L .
The size of the time steps t and the characteristic time of large events scale as
1/2
and L/U , respectively. The estimated number of time steps is then
(L/u )Re L
+1/2
3.5/2
. The complexity of the simulation is then Re L
Re3.4
of the order Re L
L .
Recommended values for the mesh size (expressed in wall units) are displayed in the
first column of Table 3.4.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Table 3.4. Typical mesh size (expressed in wall units) for DNS
and LES of boundary layer flow
x (streamwise)
y + (spanwise)
Min(z + ) (wall-normal)
Number of points in 0 < z + < 10
DNS
Wall-resolving LES
1015
5
1
3
50100
1020
1
3
For wall-bounded flows, the size of the events responsible for the turbulent production is a function of the Reynolds number, and the mesh size for LES obeys the same
scaling law as for DNS. Such LES simulations with a direct resolution of the dynamics of the inner layer are referred to as wall-resolving LES. Note, however, that LES
makes it possible to achieve considerable savings by coarsening the mesh size: Values
like x + 50 (streamwise direction) and y + 15 (spanwise direction) are usual
and must be compared with their DNS counterparts. A more significant cost reduction
might be achieved if a specific model for the inner layer, referred to as a wall stress
model, were used. The wall-modeling LES approach is discussed in Section 5.2.3.
3.3.1.3 Sound waves
Compressible flows introduce new characteristic scales. The typical wavelength of

acoustic waves in turbulent flows of interest is generally larger than the Kolmogorov
scale, and thus the aerodynamic mesh is well adapted. This result is quickly recovered
by considering a fluctuation with frequency f in time (the associated period is noted
as T f = 2/ f ). If it is a turbulent hydrodynamic fluctuation, we have T f = L t /u t ,
where L t and u t are the corresponding length and velocity scale, respectively. Assuming that the considered turbulent fluctuation is located within the inertial range
Figure 3.2. Streaks in the inner layer of the boundary layer.
107
23:56
P1: IBE
0521871441c03
108
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
of thekinetic energy spectrum and that the Kolmogorov hypotheses hold, we have
u t = 2 E(2/L t )/L t , yielding

2/3
L t = T f u t = T f 2 E(2/L t )
.
(3.72)
Now, considering an acoustic fluctuation, we have L a = T f (U + c), where U and
c are the local mean flow velocity and the sound speed, respectively. The ratio of these
two lengths is
U +c
U
1
La
=
=
+
,
Lt
uf
uf
Mf
(3.73)
where M f = u f /c is the Mach number associated with the fluctuation. In practical

flows, we have M f 1 and |U | |u f |, showing that the acoustic waves will be
directly resolved on the LES grid.
3.3.1.4 Shock waves
Another problem associated with compressible flows is the presence of shocks. In this
case, even the definition of DNS must be revised. The shock introduces a new length
scale associated with its thickness referred to as shock . The problem induced by the
presence of a shock is twofold.
First, the NavierStokes equations are relevant to describe the internal dynamics
of the shock in the low-supersonic regime only (for Mach numbers M 1.21.3).
Second, the thickness of the shock is usually much smaller than the Kolmogorov scale
. Moin and Mahesh (1998) estimate their ratio as
M 1
0.13
Re ,
(3.74)
shock
Mt
where Mt is the turbulent Mach number based on the square root of the turbulent
kinetic
energy and Re the Reynolds number based on the Taylor microscale = 2 u 2 /
(brackets stand for the Reynolds average). For most realistic applications, one has
/shock 10.
As determined by the respective size of these length scales, several approximations
can be defined, which are summarized in Table 3.5. It is observed that both quasi-DNS
and LES are based on shock capturing. Here again, new problems arise: the control of
the numerical dissipation (to be discussed in Section 3.2.2) and the required mesh size
to capture shockturbulence interaction. For shockisotropic turbulence interaction,
DNS performed by Lee, Lele, and Moin (1993, 1997) has shown that the mesh must
be fine enough to capture the shock corrugation caused by turbulence to recover good
results. This is in agreement with results of linear interaction approximation, which
shows that the shock deformation plays an important role (Mahesh et al. 1995, 1997).
An empirical criterion is x/y u rms /U , where x and y are the mesh size
in the shock-normal and shock-parallel directions, respectively. Velocity scales U and
u rms are taken in the shock-normal direction. Published DNS results correspond to
u rms /U = 0.10.2.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
109
Table 3.5. Definition of simulation types for compressible flows

Casea
Shock treatment
Turbulence treatment
Simulation type
x < shock <

shock < x <
shock < < x
shock-resolving
shock-capturing
shock-capturing
direct simulation
direct simulation
subgrid modeling
DNS
quasi-DNS
LES
a ,
shock , and x are the Kolmogorov scale, the shock thickness, and the mesh size, respectively.
The same restriction applies for shockturbulence interaction (Ducros et al. 1999;
Garnier, Sagaut, and Deville 2001, 2002): shock corrugation must be directly represented, and the LES mesh is very similar to the DNS one in the vicinity of the shock.
3.3.2 Numerical error: Analysis and consequences

3.3.2.1 Direct numerical simulation
The preceding discussion deals with scaling laws, and the remaining question is, How
many grid points are required to represent an event (eddy, wave) of characteristic size
l? The answer is obviously a function of the numerical method. Several error types
are present that contribute to the global efficiency of the method. The differentiation
error is measured using the Fourier analysis and the modified wave-number approach.
Considering a test solution in one dimension f (x) = eikx , we find its exact derivative is
ik f (x), whereas a discrete scheme yields an approximate solution of the form ik f (x),
where k is the modified wave number. Some algebra shows that k is a function of k
and x. The Fourier spectral method leads to exact differentiation and k = k. Using
the modified wave-number analysis, one can evaluate the minimal value of the ratio
l/x required to obtain an arbitrary level of accuracy. This value is also referred to as
the required number of points per wavelength (PPW). Typical results are displayed in
Table 3.6.
The second main source of error is aliasing: when functions are represented using
a finite number of basis functions, nonlinear terms require more degrees of freedom
and thus can not be exactly represented using the same set of basis function. The
Table 3.6. Modified wave-number analysis of some classical centered finite
difference schemes
Scheme
PPW ( = 0.1)a
PPW ( = 0.01)
PPW ( = 0.001)
Second-order, explicit
Fourth-order, explicit
Sixth-order, explicit
Fourth-order, implicit
Sixth-order, implicit
8
4.54
3.70
3.38
2.85
25
8.69
5.71
5.71
4
100
15.38
8.69
10
5.71
Points per wavelength (PPW) required for an arbitrary level of accuracy; parameter =
|k k|/k measures the error.
Source:Adapted From Table III of Lele (1992).
23:56
P1: IBE
0521871441c03
110
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
resulting error will pollute all resolved frequencies, the highest ones being the most
corrupted. This error is reported to result either in numerical instability or excessive
damping. The differentiation error is known to lower the effect of the aliasing error, and
so the higher the accuracy of the scheme, the larger the aliasing error. Aliasing error
is also a function of the form of the convection terms: although conservative quasilinear forms exhibit similar properties, the skew-symmetric form 12 [uu + (u
u)] significantly reduces the aliasing error in both incompressible and compressible
flows (Blaisdell, Spyropoulos and Qin 1996). Ducros et al. (2000) proposed discrete
schemes for the conservative form that mimic properties of the skew-symmetric form.
Compressible flow computations introduce a new nonlinearity through the equation of
state, whose influence is less known. For spectral methods, dealiasing is performed
using the so-called 2/3 rule: for a total number of Fourier modes equal to N , all modes
k such that k > 2N /3 are set to zero at each time steps. High-order upwind biaised
schemes, because they damp the highest resolved frequencies, are also observed to
partly cure the aliasing problem.
Quasi-DNS relies on the use of shock-capturing schemes. These schemes being
very dissipative, most authors used hybrid schemes, which consist of a blending of the
usual DNS scheme far from the shock and high-order shock-capturing scheme in the
vicinity of the shock.
3.3.2.2 Large-eddy simulation
The requirement of controlling the numerical error appears more stringent for LES than
for DNS because the LES cutoff is supposed to occur within scales that are much more
energetic than for DNS, leading to a much higher level of numerical error. This may
become very problematic if the numerical scheme introduces some artificial dissipation (artificial viscosity, upwind scheme, filter, etc.) because the amount of numerical
nonphysical dissipation may overwhelm the physical drain of resolved kinetic energy
associated to the energy cascade. The possible situations depending on the relative values of the numerical dissipation num , and the subgrid dissipation, sgs are summarized
as follows:
r Controlled LES: num sgs . The physical model is dominant, and explicit subgrid-
scale modeling is required.
r Intermediary case: num sgs . Some specific stabilized numerical schemes may
serve as subgrid model (MILES approach), and the use of explicit subgrid models
is not justified.
r Uncontrolled LES: num sgs . No subgrid model is needed, and results are strongly
case-dependent.
Practical experience reveals that the best results are obtained using centered, nondissipative schemes to make it possible to capture a much broader resolved band of scales
than dissipative methods. However, because they do not provide stabilization, these
methods require the use of much finer grids to ensure numerical stability.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
111
A particular case is the monotonically integrated large-eddy simulation (MILES)

approach (Boris et al. 1992), which relies on the use of dissipative schemes that mimic
the energy cascade. This approach is detailed in Section 3.5.4. Various shock-capturing
schemes and stabilized methods have been analyzed, leading to possibly satisfactory
results especially for free-shear flows or massively separated flows. Garnier
et al. (1999) showed that higher-order statistics of turbulence may be affected when the
MILES approach is used. The two main reasons for this are that numerical dissipation
is most of the time much higher than the physical drain of resolved kinetic energy and
that the spectral and spatial distributions of these two dissipations are not the same.
The use of stabilized methods is sometimes necessary to enforce physical behavior
of the solution. Examples are flows with shocks, flows at very high Reynolds number
on coarse grids, and flows with additional advected variables such as passive scalar
or chemical species. In these cases, the usual subgrid models are known to fail to
enforce entropic or realizability constraints, and numerical stabilization must be used.
The problem is then to keep the level of numerical dissipation as low as possible. For
the shock problem, an efficient solution proposed by Ducros et al. (1999) is to introduce
the following shock sensor in front of the dissipative part of the scheme:
| u|
.
| u| + | u|
(3.75)
To reduce the numerical dissipation provided by usual stabilized methods for compressible flows without shocks, new methods based on entropy splitting (Sandham, Li,
and Yee 2002) or wiggle detection (Mary and Sagaut 2002) have recently been used
successfully. Mary and Sagaut (2002) defined the wiggle detector i as (in one dimension at grid point number i)
i = sign [(i+2 i+1 )(i+1 i )] ,
Wk =
1 if
0 otherwise
i1
i
(ik + i+1
k ) < 0 or (k + k ) < 0
i = Maxk (Wk ),
(3.76)
,
(3.77)
(3.78)
where k is the vector of conservative variables.

Centered schemes introduce dispersive errors instead of dissipative errors. They do
not appear directly in the budget equation for the mean resolved kinetic energy and
thus are not explicit competitors of subgrid models for the energy cascade. The level
of error was shown by Ghosal (1996) to be much larger than the amplitude of the true
subgrid terms even for eighth-order accurate schemes but, owing to their dispersive
character, centered schemes are known to yield good results when they are applicable.
A possible way to minimize the numerical error with regard to the subgrid model is
to employ the prefiltering technique (Ghosal 1996). That technique consists of choosing
a cutoff length scale for the LES filter larger than the mesh spacing x. Then, while
keeping the same value for , the dominance of the subgrid terms can be recovered
by decreasing x (i.e., refining the discretization grid). Theoretical works have shown
23:56
P1: IBE
0521871441c03
112
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
that when a centered, fourth-order accurate scheme having x < /2 is used, the
subgrid contribution is expected to be dominant. The use of this technique, however,
induces a large increase in the cost of the computation (by a factor of (/x)4 if
the computation is performed using the same Courant number; i.e., U t/x = const,
where U is a characteristic convection speed and t is the time step). This is the reason
why it is not popular among LES practitioners.
3.3.3 Time advancement
The time advancement scheme and time step are chosen so as to enforce numerical
stability and accurate description of the resolved scales of motion. The Courant number
ut/x based on the velocity of the fluid, u, is usually taken of the order of unity.
In compressible flows, the fluid velocity must be replaced by the highest velocity
of acoustic waves, |u| + c. These constraints make the use of explicit schemes very
attractive.
Implicit methods can be used for convective terms in the subsonic regime if acoustic
modes are not important. In this case, the relevant velocity scale is still the fluid velocity.
Diffusive terms are also often treated implicitly especially in wall-bounded flows in
which the use of very small meshes in the vicinity of the wall may yield a very stringent
time-step restriction if explicit time integration is carried out.
3.4 Subgrid-scale modeling for the incompressible case

3.4.1 The closure problem
To obtain a tractable set of governing equations, one must close the problem (i.e., the
subgrid tensor must be expressed as a function of the unknowns of the filtered problem
u). As mentioned in Section 3.2.2 the effective filter in an LES computation appears to
be a blending of theoretical and numerical filters. Thus, subgrid scales can be grouped
into scales smaller than the effective cutoff length but larger than the Nyquist length
2x, and scales smaller than 2x. The former are sometimes called subfilter scales,
whereas the latter are referred to as subgrid scales. The former can be reconstructed
on the grid if the filter is invertible, but the latter can not.
Two philosophies can be identified for subgrid model derivation, which lead to very
different solutions (Sagaut 2002):
r Functional modeling: the idea is to introduce a new term in the filtered equations
that has the same effect (dispersive, dissipative) on the resolved scales as the subgrid
scales but that can not be considered as a true model for the tensor . It should rather
be interpreted as a model for the term .
r Structural modeling: on the opposite side, structural modeling will aim at building a
model for the subgrid tensor itself, whatever the nature of the interaction between
resolved and subgrid scales may be. Multilevel simulations, in which subgrid scales
are evaluated on an auxiliary grid, are a special case of structural models.
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
3.4 SUBGRID-SCALE MODELING FOR THE INCOMPRESSIBLE CASE
Reconstruction of interactions between scales that can be captured on the computational grid is generally carried out using structural models, whereas representation
of interaction between resolved scales and scales smaller than the Nyquist frequency
kc = /x is performed using dissipative functional models. This point is more precisely discussed in Section 3.4.4.
As guidelines for the derivation of subgrid models, we can think of the following
constraints:
r Physical constraints:
Subgrid models should have the same symmetry properties as the true subgrid
terms. It is interesting, however, to note that most of the time the discretized
NavierStokes equations do not have the same properties.
Subgrid models should be consistent (i.e., return a zero contribution when the
flow is fully resolved), leading to an automatic switch to DNS.
Subgrid models should have the same effect on the resolved scales as the true
subgrid scales (dispersion, dissipation, diffusion).
r Numerical constraints (it is important to remember that subgrid models are developed to perform simulations on computers):
Subgrid models must be realistic from the computational point of view: The extra
cost for their evaluation must be acceptable (if you really can afford a large extra
cost, then it is sometimes better to use a finer grid with a simpler subgrid model!).
As a consequence, subgrid models should be local in space and time.
Subgrid models should not induce numerical instability.
Subgrid models should be numerically robust that is, once discretized, the
effect of the model should remain the same as in the continuous case (no spurious
effects).
These constraints seem to be very simple, but a careful analysis shows that many
existing subgrid models do not satisfy all of them. A large number of models have
been proposed but will not be described here. See the monograph by Sagaut for a more
exhaustive presentation (Sagaut 2002).
3.4.2 Functional modeling
Functional modeling uses the dynamics of the isotropic turbulence as a basis. Both
theoretical and numerical studies show that the net effect of the subgrid scales is a drain
of kinetic energy from the resolved scales (see Chollet and Lesieur 1981; Domaradzki,
Liu, and Brachet 1993; Kraichnan 1971, 1976; Lesieur 1997; Leslie and Quarini 1979;
Zhou 1993). That process is directly linked to the energy cascade from large to small
scales (see Figure 3.3). A weaker inverse energy cascade is also detected (i.e., transfer
of kinetic energy from small subgrid scales to large resolved scales), but that secondary
process, commonly referred to as backscatter, is neglected by almost all the LES users.
There are two reasons for this: (1) Several models have been proposed, but none of them
has been found to be satisfactory. (2) When accounting for inverse cascade, energy is
113
23:56
P1: IBE
0521871441c03
114
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Figure 3.3. Schematic of kinetic energy transfer in isotropic turbulence.
injected into the resolved field, yielding possible numerical instabilities. See, however,
Carati, Ghosal, and Moin (1995) and Leith (1990) for backscatter models in physical
space. A possible use for this backscatter model could be to cure the overdissipation
provided in most MILES approaches.
A simple way to account for that net drain of energy is to parameterize it as an
additional dissipation. This is done by defining an eddy viscosity in the same spirit that
is done in RANS modeling. One can then write
D = 2t S, S =
1
(u + t u).
2
(3.79)
It is important to note that only the deviatoric part D = (kk /3)Id is taken into
account because S is a traceless tensor in the incompressible case. The isotropic part
(kk /3)Id is added to the filtered pressure, leading to the definition of the pseudopressure
= p + kk /3.
The closure problem now consists in the definition of the eddy viscosity t . Before
describing the most popular eddy viscosity models, let us note that the use of a scalar
viscosity is a simplification a simple tensorial analysis shows that t should be a
fourth-order tensor:
i j = t i jkl S kl .
(3.80)
This kind of eddy viscosity model is local in space and time (and is then very simple
to use), but theoretical analysis demonstrates that it should be nonlocal in space and
time (Yoshizawa 1979, 1984). The locality is recovered only when the existence of
a spectral gap between resolved and subgrid scales is assumed. An equivalent statement is that the characteristic scales of the subgrid modes are much smaller than
those of the resolved field. With L 0 and l0 being the characteristic length scale of the
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
115
resolved and subgrid fields, and T0 and t0 the associated time scales, this condition
reads
l0 L 0 , t0 T0 .
(3.81)
It is worth noting that this condition is fulfilled for RANS modeling, but not for LES,
because the kinetic energy spectrum is a continuous function and the cutoff is assumed
to occur in the inertial range; there is no spectral gap.
3.4.2.1 Smagorinsky model
The most popular subgrid model is certainly the Smagorinsky model, which is also
the oldest one (Smagorinsky 1963). It is obtained by performing a simple dimensional
analysis. Considering that
t l02 t01 ,
(3.82)
we find that the problem is now to evaluate the two characteristic scales l0 and t0 .
Assuming that the cutoff length scale is representative of the subgrid modes, one can
write
l0 = Cs ,
(3.83)
where Cs is a constant to be evaluated. The evaluation of the time scale t0 is a bit more
complex and requires new assumptions on the dynamics. We first assume that the local
equilibrium hypothesis is satisfied that is, the production rate of kinetic energy is equal
to the transfer rate across the cutoff, which is equal to the dissipation rate by the viscous
effect, resulting in an automatic adaptation of subgrid scales to the resolved ones (i.e.,
the information propagates at an infinite speed along the spectrum). The characteristic
time scale of the subgrid modes is then equal to that of the resolved scales, which is
assumed to be the turnover time defined as

(3.84)
T0 = 1/ 2S i j S i j = t0 ,
leading to

t = (C S ) 2S i j S i j .
2
(3.85)
The constant of the model remains to be evaluated. A theoretical value of the constant
can be derived under the assumptions that the spectrum is a Kolmogorov spectrum that
is, E(k) = K 0 2/3 k 5/3 , with K 0 1.4 and that filter is a sharp cutoff filter, yielding
C S 0.18. Note that this spectrum is a nonphysical one because it corresponds to an

infinite value of the turbulent kinetic energy 0 E(k)dk.
That value is not a universal one and depends a priori on the energy spectrum.
Lower values should be used for shear flows (typically 0.1 for channel flow; Deardorff
1970) in order to account for the contribution of the mean shear into the evaluation
23:56
P1: IBE
0521871441c03
116
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
of the time scale T0 . Another possibility when dealing with shear flows is to use the
splitting technique of Schumann (1975):
(1)
(2)
iDj = 2t (S i j S i j ) 2t S i j
with
(1)

= (C1 )2 2(S i j S i j )(S i j S i j ),

(2)
t = (C2 )2 2 S i j S i j ,
(3.86)
(3.87)
(3.88)
where is the statistical average operator (here, average over time/homogeneous

directions). That splitting yields an improvement of the results but is restricted to
configurations in which the statistical average can be performed.
3.4.2.2 WALE model
The wall-adapting local eddy viscosity (WALE) model was designed by Nicoud and
Ducros (1993) to recover the correct asymptotic behavior of the subgrid viscosity at
the wall in zero-pressure gradient incompressible boundary layers:

3/2
Sidj Sidj
(3.89)
t = (Cw )2
5/2 d d 5/4
Si j Si j
+ Si j Si j
with

1
Sidj = S ik S k j + ik k j i j S mn S mn mn mn ,
3
=
1
(u t u).
2
(3.90)
(3.91)
3.4.2.3 Mixed-scale model
Another popular eddy viscosity subgrid model is the mixed-scale model (Sagaut 2002),
which reads
t = Cm
3/2 1/4
qc (2S i j S i j )1/2 ,
(3.92)
where the constant Cm is equal to 0.06. The quantity qc is the kinetic energy of the
highest resolved frequencies, which is assumed to be a good surrogate for the subgrid
kinetic energy. It is evaluated during the simulation by applying a second filter with
> , referred to as the test filter (symbol ), to the resolved field u:
cutoff length
1
u)2 .
(3.93)
(u
2
In practice, a three-point discrete test filter is used to compute qc (see Section 3.2.5.4
for a description of discrete filters). This model appears to have different properties
when compared with the two previous ones: although the Smagorinsky and the structure
function models are only sensitive to the gradient of the solution and then return nonzero
qc =
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
117
values for the laminar uniform shear flow, the mixed model is also sensitive to the
2
u 2 u.
second-order derivatives of the solution; a three-point test filter yields u
This is a good illustration of the problem of defining subgrid models by looking only
at the resolved scale features.
3.4.2.4 Estimation of the subgrid-scale kinetic energy
As mentioned in Section 3.4.2 subgrid viscosity models for incompressible flows are
relevant models for the deviatoric part of the subgrid tensor only. As a consequence,
the subgrid kinetic energy, which is tied to the trace of the subgrid tensor, is not a direct
output of these models and must be modeled separately when it is required for some
physical analysis (but it is important to note that its knowledge is not necessary for
integrating the closed, filtered equations).
First, we emphasize that the trace of the subgrid tensor is not equal to the filtered
subgrid kinetic energy q sgs = 12 u i u i . The exact relation is (Knaepen, Debliquy, and
Carati 2002; Sagaut 2002)

1
1
u l u l u l u l = q sgs + u l u l = k.
ll =
2
2
(3.94)
It is seen that the equality is recovered only if the filter is a Reynolds operator. Some
models relying on an evolution equation for q sgs have been proposed but will not be
described here. We will restrict ourselves to algebraic models.
Assuming that the Kolmogorov spectrum holds for all wave numbers and introduc = , > 1, one obtains
ing a secondary filter (noted by a tilde) with cutoff length
the simple expression
+
1 L ii
q sgs =
E(k)dk =
u i
ui .
(3.95)
, L ii = u i u i
2 2/3 1
/
The two filters are assumed here to be sharp spectral cutoff filters. The idea of
introducing an auxiliary filter can also be used to obtain a self-adaptive version of the
Yoshizawa model, leading to
2
q sgs = Cq |S|2 ,
Cq =
L ii
( 2 |
S|

|S|)
(3.96)
A last, improved expression was proposed by Knaepen et al. (2002) to account for
viscous effects at very high wave numbers:
q sgs =
1 L ii
(1 2/3 ),
2 2/3 1
(3.97)
where

3/4
2i j S i j
kc
4/3
,
= , kc = , k = kc
k
3 A
A=
2/3 L ii
3
2/3
(1 2/3 )
(3.98)
23:56
P1: IBE
0521871441c03
118
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
3.4.3 Structural modeling
Many ways of reconstructing and approximating subgrid scales have been proposed
that rely on very different approaches: approximate deconvolution, scale-similaritybased models, deterministic vortex models, and so on. The emphasis is put here on the
first two classes of models because they are by far the most popular ones and can be
interpreted as different implementations of the same fundamental idea.
The key idea is to approximate the subgrid tensor as
i j u i u j u i u j u i
u
j u
i u
j ,
(3.99)
where u
is a synthetic velocity field reconstructed from u, which approximates the exact
unfiltered solution u. Computing u
is equivalent to searching for inverting the filtering
procedure. This is possible, at least theoretically, for nonprojective filters. Because exact
deconvolution is impossible in practice, an approximate estimate is used.
A first way to compute u
is proposed by Stolz and Adams (1999), Adams and Stolz
(2001), and Stolz, Adams, and Kleiser (2001a, 2001b). If the filter kernel G has an
inverse G 1 , the latter can be obtained using the van Cittert iterative procedure:

(Id G)l = lim G 1
(3.100)
G 1 = (Id (Id G))1 = lim
p .
p
l=0, p
An approximate defiltered field is then obtained using a truncated series expansion:

u
= G 1
p
u
(3.101)
with p = 5 in practice. The scale-similarity model proposed by Bardina (Bardina,

Ferziger, and Reynolds 1983; Meneveau and Katz 2000) is recovered by taking p = 0,
yielding u
= u and
i j = u i u j u i u j .
(3.102)
Another way to compute u

is to approximate the filtering operator as a differential
operator. Writing the Taylor series expansion of the velocity field at point x as
u(y) = u(x) +
(y x)l l u
(x)
l!
xl
l=1,
(3.103)
and inserting it into Equation (3.11), we obtain

u(x) =
(1)k k
ku
Mk (x) k (x),
k!
x
k=0,
(3.104)
where Mk is the kth-order moment of the kernel G (the flow field is now assumed to
be 2 -periodic in the three directions of space):

Mk (x) = (1)
(+x)/
( x)/
G(, x, y)y k dy.
(3.105)
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
119
The approximate inverse operator is computed as the inverse of the approximate

filter operator obtained considering a finite Taylor series expansion, leading to

1
(1)k k
k
Mk (x) k
u(x, t).
(3.106)
u (x, t) =
k!
x
k=0, p
Several practical implementations have been proposed: direct inversion of the matrix
associated to the implicit discretized problem associated with Equation (3.106) (Fischer
and Iliescu 2001) or explicit approximation of its solution. The latter is obtained by considering as a small parameter and expanding relation (3.106) using the well-known
relation (1 )1 = 1 + 2 . . . . Carati, Winckelmans, and Jeanmart (1999) derived a general expression for the resulting model, which is valid for a large family of
filter kernels, including all symmetric kernels:
i j =

l,m=0,;(l,m)=(0,0)
Clm
l ui m u j
,
xl xm
(3.107)
where Clm are filter-dependent coefficients. This general expression shows that models
like Clarks gradient model (Clark, Ferziger, and Reynolds 1979) and the tensor eddy
diffusivity model (Winckelmans et al. 2001) are nothing but particular members of a
more general family.
These models can be further improved by adjusting the constant that appears in
front of them. This makes it possible to account for the fact that the best low-order
approximation coefficients are not those of the infinite Taylor series expansions (e.g.,
Wagner 2001, 2003).
3.4.4 Linear combination models, full deconvolution, and
Lerays regularization
The deconvolution models presented in Section 3.4.3 are not able to reconstruct scales
smaller than the grid spacing x and so are not able to account for nonlocal energy
transfer across the cutoff. As a consequence, they must be supplemented by another
model specially designed for this purpose. A simple way to do this is to use functional
models of the eddy viscosity type based on the description of the kinetic energy transfers
associated with nonlocal interactions.
Thus, a full deconvolution model is obtained by operating a linear combination of
a deconvolution-type structural model with a functional model of the eddy viscosity
type.
The use of these linear combination models is encouraged because many theoretical
arguments and practical observation prove it is beneficial. The main reasons are as
follows:
r The structural models presented in Section 3.4.3 are known to dissipate no or very
small amounts of resolved kinetic energy, but they are very good at predicting the
anisotropy of the subgrid scales. They exhibit very good correlation coefficients with
23:56
P1: IBE
0521871441c03
120
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
exact subgrid stresses. Subgrid viscosity models provide much more dissipation but
do not properly represent the subgrid tensor.
r Shao, Sarkar, and Pantano (1999) did show that structural models represent the
rapid part of the subgrid tensor very well (i.e., the part associated to the mean
flow gradient), whereas functional models are adequate to account for the slow
part. The rapid part is important when production of kinetic energy is much larger
than dissipation or when filter length is of the same order as the integral scale of
turbulence. It is responsible for the anisotropy production of the subgrid stresses.
It is worth noting that eddy viscosity can be replaced by other resolved turbulent
kinetic energy sinks. Possible ways are the use of numerical dissipation (MILES approach) or the implementation of a penalization term as proposed by Adams and Stolz
(2001):
1
(Id G 1
p
G)
u,
(3.108)
where is a relaxation time.

Another approach for deriving stable deconvolution models, based on mathematical
results dealing with functional analysis of NavierStokes-like equations, was proposed
by Geurts and Holm (2003). Considering Lerays analysis on regularized NavierStokes
equations as a starting point, these authors suggest using the following asymmetric
model:
i j u i u j u i u j u i u
j u i u j .
(3.109)
3.4.5 Extended deconvolution approach for arbitrary nonlinear terms
A very interesting feature of the deconvolution approach is that it can be used to close
subgrid terms arising from any kind of nonlinearity because it is a general approach that
relies not on any assumption on the nonlinear term but on the filter. A general closure
procedure for arbitrary nonlinear functions of a scalar Z was proposed by Pantano
and Sarkar (2001). Given a nonlinear function f , the associated subgrid term arising
through the filtering procedure is
f (Z ) = f (Z ) f (Z ).
(3.110)
The key idea, as in classical deconvolution models, is to approximate f (Z ) by f (Z

),
with Z
= G 1
p
Z , yielding
f (Z ) f (Z
) f (Z ).
(3.111)
Expanding relation (3.100), one can write the approximate defiltered variable as
Z
= Z + c0 (Z Z ) + c1 (Z 2Z + Z ) + . . . ,
(3.112)
where all scalar coefficients ci are equal to 1. The new idea is to optimize values of these
coefficients to enforce new constraints on Z
. Pantano and Sarkar (2001) considered
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
121
the first-order reconstruction (i.e., ci = 0, i 2) and searched for a value of c0 such

that the exact and modeled fields have averaged filtered moments of order 2 that are
equal:

(Z 2 (x) Z (x))d x =

2
Z
2 (x) Z
(x) d x.
(3.113)
Given the analytical expression for both the transfer function of G and the spectrum of
Z , an exact expression for c0 can be found.
3.4.6 Multilevel closures
Another way to evaluate the subgrid tensor is to compute the subgrid scales explicitly
using a multilevel approach. This approach can be seen as an extension of the deconvolution procedure based on direct evaluation instead of filter inversion. With u on a mesh
of size x, the idea is known to evaluate u on an auxiliary mesh of size x < x
(typically, x = x/2). The approximate unfiltered field on mesh x is u
= u + u .
The nonlinear term (u u) is computed on mesh x and then restricted on mesh x.
In order to save time with respect to DNS, the evaluation of the fluctuating field

u on the secondary grid must involve some simplifications. A first possibility is to
compute it using random functions, low-order chaotic dynamic systems or kinematic
models rather than solving partial differential equations fractal interpolation procedure
(Scotti and Meneveau 1997), subgrid-scale estimation procedure (Domaradzki and Liu
1995), . . . . Another approach is to use simplified models derived from the Navier
Stokes equations instead of the full NavierStokes equations (modified subgrid-scale
estimation procedure (Domaradzki and Yee 2000), local Galerkin approximation (McDonough and Bywater 1986). A last way of reducing the complexity is to use the
NavierStokes equations but on a grid such that all subgrid scales are not captured
(variational multiscale method (Hughes, Mazzei, and Jansen 2000; Hughes, Mazzei et
al. 2001; Hughes, Oberai, and Mazzei 2001), Terracols multilevel algorithm (Terracol,
Sagaut, and Basdevant 2001)). Here, u is solution of a new LES problem. Further
saving can be obtained defining a cycling strategy between the two grids and eventually
using more than two resolution levels (Terracol et al. 2001).
3.4.7 The dynamic procedure
All the eddy viscosity models presented in Section 3.4 exhibit a constant that was set
looking at the isotropic turbulence case. An idea to minimize modeling errors is to
adjust that constant at each point and at each time step in order to get the best possible
adaptation of the selected subgrid model to the local state of the resolved field. This
can be done using the dynamic procedure, which relies on the Germano identity.
23:56
P1: IBE
0521871441c03
122
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Figure 3.4. Schematic of the two-level filtering procedure and the Germano identity.
3.4.7.1 Germano identity
Germano derived the following exact identity (Germano et al. 1991) that establishes a
.
link between the subgrid tensors obtained at the two different filtering levels . and
:
with respective filtering length scales and

(u
i u j u i u j ) = (u i u j u i u j ) (u i u j u i u j ) ,

L imj
Ti j
(3.114)

i j
where L m is a resolved tensor at the first filtering level (bar symbol), which can be
directly computed, and T and are, respectively, the subgrid tensors obtained at the
second and first filtering level (see Figure 3.4).
3.4.7.2 Computation of the subgrid model constant
The Germano identity can be used to compute the optimal value of the constant of
subgrid models (Germano et al. 1991). Formally rewriting the closure as
i j = C f i j (u, ),
(3.115)
where C is the constant to be computed and f i j (u, ) the subgrid model itself (either of
functional or structural type), and assuming that the same model with the same constant
can be employed to close the problem at the second filtering level, we obtain

u, ).
Ti j = C f i j (
(3.116)
Inserting Equations (3.115) and (3.116) into Equation (3.114), we can define the residual
Ri j :
+ C
u, )
f i j (u, ).
Ri j = L imj C f i j (
(3.117)
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
123
To enable a simple evaluation of C, we rewrite the residual as (Ronchi, Ypma, and

Canuto 1992)

Ri j = L imj C( f i j (
u, )
f i j (u, )).

(3.118)
Mi j
The constant is now computed to minimize a given norm of the residual. The most used
solution is to operate a least-squares minimization of the residual (Lilly 1992) that is,
find C such that
Ri j Ri j
= 0,
C
(3.119)
yielding
C=
L imj Mi j
Mi j Mi j
(3.120)
The use of that relation at each grid point and each time step allows us to adapt the
subgrid model to the local state of the flow. Subgrid models whose constant(s) are
computed using this procedure are referred to as dynamic models. The main interesting
properties of dynamic models are as follows:
r They vanish automatically in the near-wall region (so there is no need for a wall-
damping function).
r They vanish in fully resolved regions of the flow (automatic switch to DNS).
r They are able to capture the transition process.
Numerical experiments show that the dynamic procedure also has some drawbacks.
First, the constant C may happen to have some infinite values. Second, when used
together with an eddy viscosity model, the dynamic constant happens to take negative
values corresponding to a negative subgrid viscosity. The physical meaning of this
phenomenon is not proven, and it has deleterious effects on the numerical stability. This
is why several stabilization procedures have been proposed. The first one, referred to as
the clipping procedure (Zang, Street, and Koseff 1993), is to bound the dynamic constant
so that the total viscosity ( + t ) remains nonnegative. The second one is to average
the numerator and denominator of Equation (3.120). Averaging can be performed over
homogeneous direction(s) (Germano et al. 1991; Zhao and Voke 1996), over time (Zang
et al. 1993), along streamlines (Meneveau, Lund, and Cabot 1996), or over neighboring
grid points (Zang et al. 1993). Another possibility is to remove the highest resolved
frequency from the velocity field before computing the constant.
3.4.7.3 Extension to multiparameter models
The dynamic procedure can also be used to compute the constants of multiparameter
models (such as the linear combination models). We now consider the following model:
i j = C1 f i j (u, ) + C2 gi j (u, ).
(3.121)
23:56
P1: IBE
0521871441c03
124
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
Following the same procedure, one can evaluate the two constants by solving the
two problems
Ri j Ri j
= 0,
C2
Ri j Ri j
= 0,
C1
(3.122)
which leads to the inversion of a 2 2 matrix. The residual is now defined as

g u, . (3.123)
G

Ri j = L i j C1 f i j (u, ) f i j (u, ) C2 gi j
u,
ij
Numerical experiments have shown that direct use of this two-parameter procedure
can lead to numerical instability when it is applied to the linear combination of a subgrid
viscosity and a soft deconvolution model. This instability occurs because the subgrid
viscosity constant appears as the scaling factor of a residual term (instead of a firstorder contribution to the modeling of the subgrid tensor), the soft deconvolution models
exhibiting a very high correlation level with the true subgrid viscosity tensor. It was
also proven by Kobayashi and Shimomura (2003) that the soft deconvolution part may
induce a strong antidissipative effect in boundary layers.
To cure this problem, it is recommended to first evaluate the constant of the subgrid viscosity part (without taking into account the soft deconvolution part) and then
dynamically adjust the constant of the second part of the composite model.
3.4.7.4 Accounting for numerical errors
Previous developments dealing with the dynamic procedure were all devoted to the
adjustment of the subgrid model constant with the purpose of best approximating the
subgrid tensor. This procedure does not take into account the discretization errors that
occur when computing the divergence of the subgrid tensor (i.e., the subgrid force that
is the true term appearing in the momentum equation).
The dynamic procedure can also be used to account for the numerical errors by
using it at the vector level and not at the tensor level (Morinishi and Vasilyev 2002;
Sagaut 2002). A solution is to write
i j = C
f i j (u, ),
x j
x j

Ti j = C
f i j (
u, ),
x j
x j
(3.124)
leading to the following definition of the vectorial residual:

Ri =

L i j C(
f i j (
u, )
f i j (u, )).
x j
x j
x j

(3.125)
Ni
The least-squares minimization procedure yields the new expression for the
constant C:
C=
Ni L imj / x j
Ni Ni
(3.126)
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
3.5 EXTENSION OF SUBGRID MODELS FOR THE COMPRESSIBLE CASE
3.5 Extension of subgrid models for the compressible case

3.5.1 Background
The extension of the subgrid models to the compressible case can mostly be interpreted
as a variable-density extension of the incompressible models rather than the development of true compressible subgrid models relying on the physics of compressible
turbulence. The main reasons are as follows:
r The subgrid Mach number Msgs is generally small, and thus compressible effects
at the subgrid motion level are expected to be small. Erlerbacher et al. (1992)
indicate that subgrid compressiblity effects can be neglected for subgrid Mach
number less than or equal to 0.4, and, because the subgrid Mach number is smaller
than the turbulent Mach number Mt , they propose to neglect these effects in flows
with a turbulent Mach number smaller than 0.6. An example is the boundary layer
flow without strong cooling or heating effects in which the Morkovin hypothesis
applies up to Mach number M 5 and subgrid modes are nearly incompressible.
Another point is that a priori evaluation of subgrid terms using DNS data reveals
that some subgrid terms can be neglected for a large class of flows. Examination
of DNS data shows that subgrid term A2 (see Equation (3.39)) in the momentum
equations can be neglected in all cases. The most significant terms in the energy
equation (as defined in Equations (3.39)(3.40)) are B1 through B5 , whereas B6 and
B7 are about one to two orders of magnitude smaller.
r The main compressible effects are expected to be taken into account by the resolved
modes.
Thus, functional models will be extended with variable-density extensions of incompressible subgrid models for the momentum equations, and models for the passive
scalar case will be used as a basis for the energy equation. Structural models will also
be extended but are not so restrictive because they do not rely on assumptions about
the dynamics.
3.5.2 Extension of functional models
Functional models are extended in a very simple way to compressible flows. Still
considering the splitting of the subgrid tensor into two parts
1
(3.127)
= D + kk Id ,
3
we find that the deviatoric part is parameterized using the following variable-density
extension of the incompressible eddy viscosity model:

1
D

(3.128)
= 2t S Skk Id ,
3
where t is evaluated using one of the previously discussed subgrid models. A few
subgrid models for the isotropic part kk have been proposed (Erlerbacher et al. 1990,
125
23:56
P1: IBE
0521871441c03
126
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
1992; Salvetti and Banerjee 1994), but it appears that in most of the existing computations it is very small in front of the filtered pressure and then is often neglected (see
also Section 3.4.2.4).
In the filtered energy equation discussed in Section 3.2.4.2, terms B1 and B2 still
need to be parameterized. Vreman et al. (1997) and many other authors who have been
dealing with similar formulations have employed a simple Boussinesq-like model:
B1 + B2 = (t T ),
(3.129)
where the eddy diffusivity is linked to the eddy viscosity through a turbulent Prandtl
number Prt = t /t . Usual values for this Prandtl number are Prt [0.3, 0.6]. It can
also be computed using a dynamic procedure.
3.5.3 Extension of structural models
The extension of the structural subgrid models is straightforward and will not be discussed here. The procedure is the same as in compressible flows considering massweighted filtered variables. The subgrid tensor is approximated as

(u)i
(u)
j
(u i )(u j )
1
1
u i u j

(u)
i (u)
j . (3.130)
i j =
Subgrid models for compressible flows are derived applying the deconvolution
procedure (iterative or explicit) to the filtered variables. An interesting point is that these
extensions do not rely on any assumption on the compressible interscale dynamics, and
numerical experiments have shown that they were able to account for compressible
effects on the subgrid dynamics. Although explicit deconvolution based on the Taylor
series expansion technique may yield some problem near shocks or interfaces (Garnier
et al. 2002), Adams recently proved that the van Cittert iterative procedure has some
shock-capturing capability (Adams 2002).
3.5.4 The MILES concept for compressible flows
It was shown in Section 3.3.2 that numerical errors may occur to overwhelm the subgrid
model, leading to uncontrolled simulations. However, some amount of numerical dissipation is necessary to guarantee that some realizability constraints will be fulfilled:
positive of specific mass fraction, entropic solution near shock waves, temperature
physical bound preservation, and so on.
A huge amount of work is devoted to the development of numerical methods that are
able to enforce these realizability constraints. A common feature of all these methods
is that they rely on the introduction of numerical dissipation. Thus, the problem of
producing realiable LES with stabilized numerical schemes arises.
To solve this problem, Boris et al. (1992) proposed a new concept referred to as
monotonically integrated large-eddy simulation (MILES), which is based on the idea
that numerical schemes can be designed that are capable of enforcing the realizability
23:56
P1: IBE
0521871441c03
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 16, 2006
127
3.5 EXTENSION OF SUBGRID MODELS FOR THE COMPRESSIBLE CASE
constraints while providing a subgridviscosity-like dissipation in turbulent flows. With

such a scheme, usual explicit subgrid models are no longer needed, and LES-like results
can be obtained by just solving the NavierStokes equations.
The MILES approach was further investigated by Fureby and Grinstein (1999,
2002), who considered the case of flux-limited finite volume schemes. If it is assumed
that the total numerical convective flux function v f is evaluated as

v f = v Hf (1 ) v Hf v Lf ,
(3.131)
where v Hf , v Lf , and are a high-order flux function, a low-order (dissipative) flux function, and the limiter function, respectively, the NavierStokes equations supplemented
by the leading numerical error terms are
+ (u) = 0
t
u
+ (u u + p) v = (E 1M + E 2M )
t
E
+ (u(E + p)) (uv ) + qT = (E 1e + E 2e )
t
E 1M = (CT u + uC T ),
E 1e = Ce,
E 2M = 2 (ud ud),
E 2e = 2 (e d)ud,
(3.132)
(3.133)
(3.134)
(3.135)
(3.136)
where d is the topology vector connecting the neigboring cells, e = E/, is a scalar
function that depends on the flux limiter , and C = (u d). Looking at these error
terms, we see that E 1M and E 1e can be interpreted as tensorial subgrid viscosity models,
whereas E 2M and E 2e are close to scale-similarity models. As a consequence, the MILES
approach can yield relevant results if is adequatly defined.
23:56
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
Use of Hybrid RANSLES for Acoustic

Source Predictions
Paul Batten, Philippe Spalart, and Marc Terracol
4.1 Introduction to hybrid RANSLES methods

The title of this book implies the current or impending feasibility of large-eddy simulation (LES) for use in noise prediction problems. The potential accuracy of LES is
generally well regarded; however, the cost of such calculations requires a specifically
focused algorithmic and computational effort in order to make LES affordable for practical engineering problems. At high Reynolds numbers, such problems can defeat LES
by raising the range of scales beyond affordability owing to the extremely small size of
the eddies in the viscous sublayer or even the size of the dominant eddies in the bulk
of the boundary layer (Spalart et al. 1997). This chapter outlines a variety of recently
developed hybrid methodologies that specifically address the issue of computational
cost and make the simulation of large-scale, sound-generating flow structures tractable
with existing computer resources. The hybrid nature of the methods discussed here
involves the simultaneous use of (or blending between) statistical Reynolds-averaged
NavierStokes (RANS) and traditional LES within the noise-source region. This hybrid
character is distinct from the methods used in traditional acoustic analogies for far-field
propagation of sound. Those issues remain here; however, this chapter specifically considers methods offering an affordable prediction of both the mean flow and pressure
fluctuations in the near field.
A common premise throughout the methods described in this chapter is that routine
engineering predictions of acoustic disturbances (or simply turbulence) arising from
realistic Reynolds-number flows must lead to a cost that is substantially less than
that of traditional, full-domain LES. Thus, the methods discussed herein consider the
application of LES to only those regions containing the most significant and geometrydependent noise-generating structures. The emphasis is on methods that are able to
contain the cost of establishing both coherent and broadband noise sources without
losing all accuracy. Additionally, because by nature hybrid RANSLES methods resolve
a narrower fraction of the frequency spectrum than traditional (full-domain) LES, this
chapter also considers the possibility of extracting the missing noise that would have
originated from the unresolved scales in the flow.
128
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
4.1 INTRODUCTION TO HYBRID RANSLES METHODS
The use of full-domain LES (involving a single subgrid model) may appear to be a
natural choice for aeroacoustics problems because of the weak modeling dependency
and its inherently time-dependent nature. Historically, however, cost has been a decisive
factor favoring the use of semianalytic models capable of extracting noise from either
unsteady RANS predictions or directly from statistical data. Cost will continue to be a
major consideration in future use of LES for acoustics problems, but its application is
made practical here by carefully confining the role of LES to the most critical noisesource regions. The finest scale structures are still expected to be modeled statistically,
and far-field propagation is expected to be handled by some alternative approach. In
general, the intention is to limit the use of LES or hybrid RANSLES to the (nonlinear) noise-source generation region with subsequent propagation handled via some
linear propagation method, which would ideally be analytic and not even involve a
computational mesh.
The statistical modeling employed in the hybrid methods discussed here has more
impact than simple subgrid-scale (SGS) modeling in the sense that it may model even
the dominant eddies, but then only within some specific RANS regions. The recent
developments in hybrid RANSLES are natural candidates for engineering aeroacoustic
calculations because the statistical modeling employed is restricted to the finest scales
with the larger, more energetic and more significant scales of motion (determined via a
choice of mesh and time step) being simulated directly. The weaker reliance on empirical
modeling, relative to unsteady RANS, is particularly welcome for problems involving
fluid resonance or discrete tones arising from geometry-related scales. Although empirical modeling of tonal noise has had some success for very specific geometries (such
as the square cavity analysis of Rossiter (1964) and Heller, Holmes, and Covert (1971),
in general, the prediction of fluid-dynamic or fluid-resonant instabilities associated
with the pockets of separated flow over, for example, base regions, wings, and cavities,
requires accurate numerical modeling of the dominant, large-scale, coherent structures.
General experience with unsteady RANS has been mixed because the outcome
of such predictions has been dependent on the particular flow and type of RANS
model employed. Some progress has been reported using unsteady RANS to predict
the largest scales of motion (e.g., Zhang, Bachman, and Fasel 2000b; and Iaccarino and
Durbin 2000) even though the RANS equations do not achieve the correct limit (direct
numerical simulation) as the mesh spacing 0. Although unsteady RANS may give
useful predictions for certain flows, its use in areas related to aeroacoustics is regarded
by many as an unsafe practice because RANS imposes hard empirical limits on the
representation of the flow physics limits that cannot be shifted by increased temporal
and spatial resolution, as would be the case in LES. There is not even a consensus
over whether such unsteady RANS simulations should be strictly three-dimensional
(i.e., even when the geometry is only two-dimensional). This creates a need for an
intermediate strategy that is richer than RANS but avoids the cost of a full-domain
LES.
One of the earliest mentions of hybrid RANSLES was in a paper by Schumann
(1975). However, Schumanns proposal applied only to the lower section of the boundary
17:51
129
P1: IBE
CUFX063-04
CUFX063/Wagner
130
0 521 87144 1
printer: sheridan
October 23, 2006
USE OF HYBRID RANSLES FOR ACOUSTIC SOURCE PREDICTIONS
layer; more recent methods typically envision treating the entire boundary layer with
a RANS model. Little attention was paid to the subject until the more recent papers
of Speziale (1996) and Spalart et al. (1997). These hybrid RANSLES methods and
more recent developments inspired by them have shown that statistical modeling can
be successfully blended with LES, yielding enhanced predictions of both turbulence
statistics and unsteady flow dynamics at a fraction of the cost of traditional LES (e.g.,
Spalart et al. 1997; Germano 1999; Batten, Goldberg, and Chakravarthy 2000, 2002b;
Strelets 2001; Menter, Kuntz, and Bender 2003; Labourasse 2002; Labourasse and
Sagaut 2002; Spalart and Squires 2003). A brief background on these hybrid methods
is presented in Sections 4.2 and 4.3, and examples are considered in Section 4.4.
An emerging strategy that is considered to show promise for acoustics predictions
is the use of nonlinear disturbance equations (Morris, Long et al. 1997) in which statistical modeling is used to determine an initial mean field about which fluctuations are
synthesized or allowed to develop naturally in resolvable scales. Recent developments
in this area, which we refer to as nonlinear acoustics solvers, are also discussed along
with their connection to the existing range of hybrid RANSLES methods.
Many of the methods discussed in this chapter are relatively new and the associated
terminology is still evolving, but these techniques can be broadly categorized into two
major classes (discussed in Sections 4.2 and 4.3) corresponding to global and zonal
hybrid methods.
4.2 Global hybrid approaches

Global (or nonzonal) hybrid RANSLES methods rely on a single set of model equations and a continuous treatment that blends between RANS and LES approaches.
Such methods can be considered as a form of very large eddy simulation with subgrid
stresses that are designed to reach RANS levels in certain limits of coarse or highly
stretched meshes. As in traditional LES, the subgrid model is designed to account for
the shorter-wavelength, higher-frequency velocity fluctuations in a statistical sense;
hence, full-spectrum predictions require an extended hybrid strategy in which resolved
scales are simulated directly with additional modeling employed to extract the missing
components of the noise from the remaining statistics.
4.2.1 The approach of Speziale
M
Speziale (1996) presented a hybrid framework in which the stress tensor u i u j provided
by a (conventional RANS) Reynolds-stress transport model was damped via
u i u j = u i u j
(4.1)
in which
= [1 exp( L /L k )]n ,
(4.2)
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
17:51
131
4.2 GLOBAL HYBRID APPROACHES
where and n are some (unspecified) parameters, L is some representative mesh

spacing, and L k is the Kolmogorov length scale,
3
4.
L k = 4 /
(4.3)
In Speziales approach, a regular RANS closure is recovered when L is much

larger than L k , whereas the subgrid stresses vanish completely as L 0. However,
several issues were never completely specified by Speziale (1996) such as the definition
of the parameters , n in Equation (4.2), and the definition of the model equations used
to derive the undamped Reynolds stresses. Further, more complete, developments of
Speziales Kolmogorov length-scale-dependent blending have since been pursued at
the University of Arizona (e.g., von Terzi and Fasel 2002) and CRAFT Technologies
(Arunajatesan and Sinha 2001). Knowledge of the Kolmogorov length scale is, however,
not strictly needed to ensure that the correct DNS behavior is reached in the limit of
vanishing mesh spacing, and there are many possible choices of such blending functions
(partly indicated by the free parameters in Equation (4.2)). In addition, properly reaching
both the DNS and RANS limits is no guarantee that a concept provides a viable LES
mode in between.
Speziales precise intention for the use of Equation (4.1) may never be known, but
there are (at least) two valid interpretations, both of which are represented in subsequent methods reviewed in this chapter. The first interpretation is that the modified
M
(undamped) stresses u i u j be computed via a coupled set of time-dependent transport equations evolving with the equations for the resolved scales. This interpretation
is consistent with the DES and LNS methods described in Sections 4.2.2 and 4.2.3,
M
respectively. The second interpretation is that u i u j be computed in advance (or in a
separate, time-dependent computation), with an independent set of equations solved
subsequently (or concomitantly) for the resolved-scale fluctuations. This interpretation
is closer to the NLAS and LabourasseSagaut methods described in Sections 4.2.8 and
Labourasse and Sagaut (2002), respectively.
4.2.2 Detached-eddy simulation
The detached-eddy simulation (DES) method was the most widely used hybrid method
in the 20002004 time frame thanks to clear objectives, open publications, resources
for detailed testing (Strelets 2001), commonality with a known RANS model, and a
simple and complete formulation (Spalart et al. 1997). Its impetus came from estimates
that, for a wing and similar flow problems, LES will have to wait until the year 2045,
even assuming that wall modeling has been achieved. It is then obvious that for now
any practical method will treat the regions of thin, unchallenging boundary layer with
RANS. Driven toward the simplest, nonzonal concept, the reasoning arrived at the use
of a one-equation RANS model sensitized to the filter width in the LES region, which
creates a plausible SGS model. From the onset, the gray area in which the turbulence
P1: IBE
CUFX063-04
CUFX063/Wagner
132
0 521 87144 1
printer: sheridan
October 23, 2006
needs to convert from fully modeled (in the boundary layer) to mostly resolved (in
massive separation) was recognized as potentially delicate.
The SpalartAllmaras (SA) model does not need to be introduced here. The specific
DES modification concerns its destruction term. In RANS, it represents blocking by
a nearby wall and scales with the wall distance. In LES (within DES), it represents a
limit on the mixing length and scales with the filter width. DES takes the classical view
that the filter width is best tied to the grid spacing in order to make the best use of the
available resolution. This is discussed further in Section 4.2.5. Therefore, the length
scale injected into the model is
d min(d, CDES ),
(4.4)
where d is the traditional wall distance. In a prototypical DES, the entire boundary
layer uses the d = d branch and the SA RANS model is fully active. The separated
region, given an adequate grid, uses the d = CDES branch, and the SGS model has
many similarities with the Smagorinsky model certainly its scaling with grid spacing
and dissipation rate. The effect of Equation (4.4) is very close to that of computing
a RANS model and an SGS model separately and injecting the weaker of the two
effective viscosities into the momentum equations; this describes LNS quite well (see
Section 4.2.3). However, the DES transport equation for eddy viscosity involves a
finite rate of dissipation, resulting in an eddy viscosity field that tends to be smoother
than that obtained via a pure Smagorinsky model (equivalent levels being reached
only when the DES production and dissipation terms become balanced). Whether this
smoothness has a beneficial effect on predictions within the LES region has yet to be
established.
The objectives of DES overlap with those of Speziale (1996), but the focus on
treating the entire boundary layer with RANS appears to represent a clear difference. In
addition, DES practice has benefited from the routine control of transition in the SA
model for which there is no better demonstration than the circular cylinder case (Travin
et al. 2000). LES is many years away from a serious attempt at its drag crisis, and yet
DES reproduced the major quantitative features of this flow in addition to nontrivial
qualitative features such as three-dimensional chaos and intense modulations of the
vortex shedding. The frontier for this flow resides in the intermediate Reynolds-number
range (which is unfortunately very wide: roughly 105 to 107 !). The physics mingles
separation and transition, which SA is not capable of, and thus, the DES misses both
the very low drag values and the spontaneous lifting cases.
As of 2004, DES enjoys an international user base and a publication stream that
is often successful and sometimes critical (some representative calculations are presented in Section 6.8). The principal cause for concern has been in cases in which the
grid is progressively refined until the d = CDES branch of the minimum in Equation
(4.4) intrudes inside the boundary layer. The result is a weakened eddy viscosity but
one that is not weak enough to allow LES eddies to form; as a result, the separation
line moves too far forward. This phenomenon needs careful monitoring, and a permanent solution is yet to be found. DES is now appearing in vendor computational fluid
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
dynamics (CFD) codes in which it has been demonstrated with very uneven levels of
understanding. Educating users to the complexities of designing a DES or even an LES
is a serious challenge. The statement that DES can function with a coarser grid than
LES is too often made to sound general, but in fact it is correct only in the boundary
layer.
The coupling of DES with sound prediction methods has not progressed far, but the
confidence in its capabilities to reproduce the dominant eddies of a separated flow is
good enough to be optimistic. Most probably, the coupling will be through Kirchhoff
or Ffowcs WilliamsHawkings (FWH) approaches for several reasons. Such methods
allow most of the grid points to be placed in the turbulent or nonlinear flow region.
Surface integrals can then be used to determine far-field propagation by exploiting the
pressure disturbances generated directly on a near-field-containing surface via the DES
model. The instantaneous effects of the Lighthill stress tensor (which would need to be
included in any volume-integral solution to Lighthills equation) are negligible if the
surfaces surrounding the turbulent region are designed well. A notable challenge for
most flows arises when turbulence is transported downstream and will cross the Kirchhoff or FWH surface before it has become negligible; as a result, careful corrections
and tests are indispensable.
4.2.3 LNS
The limited numerical scales (LNS) concept was presented by Batten et al. (2000) as a
means of closure for Speziales (1996) approach and (in 1999) was the first implementation of a hybrid RANSLES method in a commercial CFD solver. From its inception,
the LNS method has been used exclusively within a nonlinear eddy viscosity framework that is more general, but also slightly more complex, than that of existing DES
formulations.
In its basic mode of operation, LNS exhibits hybrid RANSLES characteristics
similar to DES, which will involve a RANS treatment of the entire boundary layer if
the near-wall region is meshed with similarly high-aspect-ratio cells. However, LNS
achieves this behavior through an instantaneous limit on the Reynolds stresses (and
hence production term) rather than through an elevated rate of dissipation, as in DES.
The two-equation, nonlinear eddy viscosity framework is not an essential ingredient
for LNS, although its use has some theoretical advantages in situations in which the
underlying RANS model is used to predict the primary separation location.
The use of a nonlinear framework is more significant in recent LNS developments
that require a realizability property of the local stress tensor. This is needed to allow
a reconstruction and transfer of data between different modes (statistical and directly
resolved representations) in a manner that aims to preserve the total shear stress. That
process becomes considerably more challenging with a single-equation framework that
provides no description of the turbulence kinetic energy. However, recent extensions of
DES to a two-equation framework (e.g., Strelets 2001) might form suitable candidates
for use with the data transfer procedures outlined in Section 4.2.7.
17:51
133
P1: IBE
CUFX063-04
CUFX063/Wagner
134
0 521 87144 1
printer: sheridan
October 23, 2006
The primary aim of LNS is to identify resolvable and unresolvable fractions of

the turbulence kinetic energy using an unambiguous expression for Speziales latency
parameter (4.2). A key distinguishing feature of LNS is that it contains no empirical
constants beyond those appearing in the baseline RANS and LES models. Whereas in
traditional turbulence closures models are calibrated as a complete entity, in LNS the key
parameter the turbulence shear stress in an axis aligned with the mean shear strain
is the only factor in determining the model blending. Thus, the LNS framework is
immediately able to exploit improvements in either RANS or LES modeling by directly
inserting any enhanced component model. No model recalibration is required.
LNS achieves this goal by redefining the latency factor in Speziales method via the
ratio of effective-viscosity norms

min ( f L .V )LES , ( f L .V )RANS
(4.5)
=
( f L .V )RANS
in which L and V denote length and velocity scales, respectively. The ( f L .V )LES
and ( f L .V )RANS products denote norms for the effective viscosity arising from the
component LES and RANS models, respectively. If consistent models are assumed for
both LES and RANS stress tensors, then the latency factor above simply selects the
shear stress of minimum magnitude (in an isotropic, Boussinesq model, such as the
Smagorinsky, this is the only relevant parameter). The use of the weaker of the two
effective viscosities is closely reminiscent of DES and likely to be a common feature
of RANSLES hybrids.
In current LNS computations, the same low-Reynolds-number damping function f
is used in both the RANS and LES component models. The conventional Smagorinsky
model has an eddy viscosity defined as

(4.6)
t = Cs (L i )2 Skl Skl
in which Cs is the Smagorinsky coefficient (here taken as 0.05) and L is a length-scale
parameter defined for an arbitrary, unstructured control volume such as
rc rk |),
L i = 4 max (|
k=0...n
(4.7)
where n is the number of faces forming cell i, rc is the centroid of cell i, and rk
is the midpoint of face k. For certain cell types (prismatic, tetrahedral), a preferred
definition may be the recent (unpublished) recommendation of Spalart to use twice
the cell diameter. This guarantees that wavelengths of L would be resolvable at any
orientation on the mesh. (The subject of defining this filter width L is discussed
further in Section 4.2.5.) The Smagorinsky SGS model requires some form of nearwall damping because, for any finite mesh spacing L , the nonvanishing strain rate
would otherwise imply a nonvanishing turbulent shear stress at the wall. Commonly
employed forms of near-wall damping for the Smagorinsky model are of the van Driest
type; for example,

b
(4.8)
f = 1 exp(y + /25)a .
17:51
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
135
Energy
production
Inertial
range
Figure 4.1. Turbulence energy spectrum partitioned

into resolvable and unresolvable frequencies.
Dissipation
Resolvable
kinetic energy
2 /
Unresolvable
kinetic energy
Wave number
The ad hoc nature of this damping is felt to justify the use of a single f function for
both RANS and LES component models. The use of a single low-Reynolds-number
damping function for both RANS and LES components is, in fact, a common feature
of both LNS and DES hybrid formulations. If the near wall were resolved with a
fine isotropic mesh, the use of a separate f for the SGS model might be justified
particularly if the Smagorinsky SGS model is to be replaced by a self-tuning or dynamic
SGS model that might require less damping. However, the reader should note that
the low-Reynolds-number damping of the LES subgrid stresses is not expected to be
significant in any of the hybrid calculations presented here because the near-wall layer
is intended to be treated via conventional RANS modeling, which is forced by the use
of high-aspect-ratio near-wall cells.
For the purpose of hybridization with the Smagorinsky model, the use of a single
f function simplifies the definition of as follows:
=
17:51
Energy
P1: IBE
CUFX063-04
min [(L .V )LES , (L .V )RANS ]

.
(L .V )RANS
(4.9)
The exact formulation for the current component RANS and LES models is given later
in Equation (4.19). The resulting equations then behave as RANS if = 1, or LES if
< 1. Thus, in LNS, the gray zone between pure RANS and pure LES relates only
to the data on which the equations operate.
In LNS, the energy fraction k is interpreted as unresolvable subgrid turbulence
kinetic energy that can only be modeled. The (1 )k component is interpreted as
resolvable turbulence kinetic energy, which, given the local grid resolution, could be
represented directly (see Figure 4.1). The sum total of k in LNS represents the current
level of statistically represented turbulence energy, which is not the same as the k that
would be achieved using a conventional RANS model.
Similarly, the quantity is interpreted as the dissipation that applies to the unresolvable scales, and the quantity (1 ) is interpreted as the dissipation or transfer
that applies to the resolvable scales. For a linear Boussinesq closure, the preceding definitions imply t = tM (i.e., the eddy viscosity simply gets multiplied by a number
P1: IBE
CUFX063-04
CUFX063/Wagner
136
0 521 87144 1
printer: sheridan
October 23, 2006
between 0 and 1). Although there is no specific requirement on the baseline RANS
model, in the present work the LNS model equations are based on a nonlinear k
model in which the Reynolds stress tensor is defined via a tensorial expansion cubic in
the mean strain and vorticity tensors:

k
k
u i

=
+
t
xi
xi

u i

=
+
t
xi
xi
t
+

t
+
k

k

+ Pk
xi

+ E)Tt1 .
+ (C1 Pk C2
xi
(4.10)
(4.11)
In the preceding equation,

ui

,
Pk = u
i uj
x j
(4.12)
and

2
1
k
t

Sik Sk j Skl Skl i j
u i u j = ki j t Si j + c1
3

3

k
t
ik Skj + jk Ski

k
t
1
ik jk lk lk i j
+ c3

3
+ c2

k2
t

+
S

S
Skl
l
j
li
ki
k
j

2

2
k2
t
+ c5 2
mn nl i j
il lm Sm j + Sil lm m j Slm

3
+ c4
+ c6
k2
k2
t
t
S
S
S
+
c
S kl kl
7
i
j
kl
kl

2

2 i j
with
Sij

uj

ui
+
x j
xi

uj

ui
,
i j =
x j
xi

k 1
S=
S S ,

2 ij ij

k 1
=
i j i j .

2

uk
2
i j ,
3 xk
(4.13)
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
137
Note that the expression above for the anisotropy tensor is currently used throughout
(i.e., in both RANS and LES components models). A realizable turbulence time scale
is defined as

(4.14)
Tt = k max 1, 1 ,
k 2 /(
) and C = 2. The eddy viscosity is defined as
where = Rt /C with Rt =
k 2 /
,
t = C f
(4.15)
with
2/3
,
A1 + S + 0.9
3/4
15/4
c1 =
, c2 =
,
3
(1000 + S )C
(1000 + S 3 )C
C =
c3 =
19/4
,
(1000 + S 3 )C
c5 = 0,
and
17:51
c4 = 10C2 ,
c6 = 2C2 , c7 = c6 ,

1
1 eA Rt
max 1,
,
f =
1 e Rt
1

1
k 2 , (
) 4

Tt ,
E = A E max

k

k
,0 , = .
= max
x j x j

(4.16)
(4.17)
(4.18)
From Equation (4.9), the latency parameter is defined for the current choice of
RANS and LES component models as

2
Skl Skl /2
Cs L i
,1
(4.19)
= min
k 2 /
C
+
with some small parameter O(1020 ) to allow 1 without singularities in lowReynolds-number regions. The remaining model constants are
A1 = 1.25, C1 = 1.44, C2 = 1.92,
k = 1.0, = 1.3, A = 0.01,
A E = 0.15, Cs = 0.05.
When fine isotropic grid regions are encountered by the LNS method, the scaling of
the predicted Reynolds stress tensor by causes the effective viscosity to be reduced
instantly to the levels implied by the underlying LES subgrid model, and the local
flow also experiences a decreased rate of production owing to the reduced magnitude of the stress tensor components. In the basic version of LNS, nothing is done
P1: IBE
CUFX063-04
CUFX063/Wagner
138
0 521 87144 1
printer: sheridan
October 23, 2006
with the resolvable-scale (1 )k energy fraction, which continues to decay according

to the local dissipation rate. More recent efforts aim to convert this statistical data into
resolved-scale fluctuations (Batten, Goldberg, and Chakravarthy 2004). The use of realizable stress tensors to generate a synthetic reconstruction of RANS data is discussed
separately in Section 4.2.7.
4.2.4 The approach of Menter, Kunz, and Bender
A potential problem can arise with the previously discussed hybrid RANSLES methods
when flow is transported (either through mean convection or turbulent mixing) across
regions of varying mesh resolution. Because the predicted subgrid stresses in these
methods are intentionally sensitized to the local mesh resolution, the model equations
will change with L . However, a switch from RANS to LES mode (or vice versa)
does not imply an instantaneous transformation of the data (such as the velocity field)
on which those equations operate. This can be problematic if the grid experiences a
significant clustering (with decreasing aspect ratio) within the boundary layer (or indeed
in any region of significant turbulent and mean-flow interaction) because a sudden
switch to LES would reduce the unresolved stresses and the approaching statistically
steady flow would have no opportunity to initiate or sustain the fluctuations that would
be required to compensate by an increase in the resolved stresses. For separated flows
around sufficiently simple geometries, suitable mesh design can eliminate this problem
by maintaining uniformly high-aspect-ratio cells in the near-wall layer. However, for
realistic geometries involving abrupt surface curvatures, it can become difficult to
satisfy such stringent meshing constraints, which would also preclude or limit the use
of local adaptive mesh refinement.
Although these problems are now reasonably well understood, there have been few
proposed solutions. A recent extension of the LNS method attempts to address this issue
by automatically extracting statistically represented kinetic energy and injecting this into
the resolvable scales (Batten, Goldberg, and Chakravarthy 2004). This approach relies
on a tensorial extension of Kraichnans (1969) method described in Section 4.2.7. However, calculations by Batten, Goldberg, and Chakravarthy (2004) on a plane-channel
flow with streamwise grid clustering showed that several channel half-heights were required to switch to a fully developed hybrid RANSLES boundary layer a distance that
would not be of any practical help in regions containing abrupt clusterings owing to surface curvature. For sufficiently simple geometries, the recycling approach of Lund, Wu,
and Squires (1998) can be used to maintain unsteady inlet data; however, its generalization presents some challenges, and, again, its use is not practical for localized mesh clusterings. Furthermore, rescaling only accounts for the effects of data transported through
mean convection and does not address the problem in hybrid RANSLES of turbulent
transport of kinetic energy through the boundary layer. If the near-wall layer is described
only in terms of statistical data, this presents the unsteady outer flow with a nonphysical
quasi-steady layer that can cause a (resolved-scale) component laminarization and, as
a result, a nonphysical description of the total shear stress. Rescaling can always be
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
139
applied to match statistics at a certain inlet plane, but this does not guarantee that those
statistics will be preserved downstream even as far as the end of the recycling zone.
A recent proposal by Menter et al. (2003) offers an intriguing alternative. The Menter
et al. (2003) approach is based on a single transport equation for the eddy viscosity:

t 2
t t
(t ) u j t
= c1 t S c2
+
+
.
(4.20)
t
x j
LvK
x j x j
For a simple boundary layer, the von Karman length scale L vK is defined as

u/ y

.
L vK = 2
u/ y 2
(4.21)
For general multidimensional calculations, Menter et al. (2003) suggest two alternative
invariant formulations for L vK (note the missing square roots in the original paper of
Menter et al. 2003):

(S.S)

L vK =
(4.22)
S S
.
x j x j
or
L vK
17:51

=
u i u i
x j x j
2 ul 2 ul
xm2 xn2
(4.23)
In an earlier paper on the derivation of a single-equation, eddy-viscosity transport

model, Menter (1997) modified the dissipation based on the preceding length scale in
order to avoid the singularity caused by a vanishing mean-strain rate:

!

t 2
t 2 t t
c2 min
,
.
(4.24)
c2
LvK
LvK
x j x j
By removing the second argument on the right-hand side of Equation (4.24) and instead
using some grid-related tolerance on L vK , the dissipation in the resulting hybrid RANS
LES model becomes sensitized to the local flow structures. It is curious to note that
this expression for a dissipation length scale arose naturally in the derivation of a
single-equation model for the k eddy viscosity (Menter 1997). The claim of Menter
et al. (2003) that their approach contains no explicit grid dependency is not strictly
correct, however, because some fraction of is still explicitly required as a tolerance
on L vK in order to allow reasonable agreement with isotropic decaying turbulence
(Menter et al. 2003). This suggests that the numerically determined L vK underestimates
the required dissipation length scale an effect that is probably required if the largerscale eddies are to be allowed to follow their physical breakdown to the Kolmogorov
scale but one that defeats the objective of entirely removing the dependency from
the model. Nevertheless, variants of this method may help to preserve the near-wall
RANS layer in situations in which mesh refinement is used (locally) to help represent
complex surface topography. This issue is considered further in the following section.
P1: IBE
CUFX063-04
CUFX063/Wagner
140
0 521 87144 1
printer: sheridan
October 23, 2006
4.2.5 Defining the filter width
The definition of a characteristic length scale is a key factor in the design of nonzonal
hybrid RANSLES models (just as in pure LES) because this allows the approach
to distinguish between unresolvable and resolvable (or at least potentially resolvable)
scales of motion. Because the orientation of local flow structures cannot be known in
advance (except perhaps in the vicinity of a wall), a safe requirement is to choose the
smallest wavelength that can be supported at any orientation to the local mesh. This
leads us to a measure for the local Nyquist grid wavelength, given by expression (4.7),
which, for a Cartesian mesh, reduces to
L = 2 max(x, y, z)
(4.25)
with the additional factor of 2 (often absorbed into the Smagorinsky constant) accounting for the wavelength corresponding to the grid Nyquist frequency. A similar measure
of filter length is used in both LNS and DES methods. However, there is no universal
agreement on this definition, and it is still common for many practitioners to employ
measures such as the cube root of the volume (as in the approach of Labourasse and
Sagaut 2002), or even to replace the maximum in Equation (4.25) with a minimum.
Such definitions reach the wrong limit for extreme cell shapes such as a chopstick,
but this has not been recognized in typical LES of free-shear flows, with their nearly
cubic cells. Hybrid methods do make the issue more critical in that they lead to cells
that are far from cubic.
Additional filter-length controls are currently under investigation within the LNS
method. For example, the insistence that the time step be small enough to resolve the
convective transport of any grid-supportable structures accurately leads to a Courant
condition based on the local fluid velocity (which defines a basic condition for accuracy
in nonlinear flow problems). This can be used to provide an additional safety factor in
the local length-scale definition:

L = 2 max x, y, z, u i2 t ,
(4.26)
where u i is the local fluid velocity relative to the mesh. This ensures that, irrespective
of the local spatial resolution, the RANS solution is recovered when t becomes
large. This additional constraint can help in regions of sharp surface curvature (such
as corners) where the mesh resolution may become uniformly fine but the temporal
resolution would not permit those very-fine-scale structures to be predicted accurately
(these structures would be on the order of the smallest boundary layer spacing in all
directions whereas the global time step is typically set to match the larger cell spacings
away from the immediate near-wall region). However, the modified length scale (4.26)
is recognized as being incomplete. In stagnation and reattachment regions, Equation (4.26) will not augment the local filter width because of the vanishing mean kinetic
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
141
energy. Additional length-scale constraints have been proposed such as the following
suggestion of Mani and Paynter (2002):

(4.27)
L = 2 max x, y, z, u i2 t, kt .
The approach above has not been widely tested but is also considered potentially
problematic because large initial values for k (such as those from a RANS solution)
could perpetuate local RANS behavior. Another alternative would be to introduce a
solution-dependent length scale such as the von Karman length scale (4.22) used by
Menter et al. (2003) in their recently proposed hybrid approach. A disadvantage of
using the Menter et al. (2003) method directly is that velocity profile curvature in
any one direction can force adoption of a small length scale (analogous to choosing
min[x, y, z] in LNS or DES), and thus it becomes impossible to guarantee a
near-wall RANS layer (note that the computations presented in Menter et al. (2003)
were performed with wall functions). A safer approach would be to introduce the von
Karman length scale into Equation (4.25), giving
L = 2 max[x, y, z, L vK ].
17:51
(4.28)
The preceding approach ensures that near-wall RANS behavior will be maintained
on stretched, high-aspect cells irrespective of the normal-to-wall component of velocity
curvature. The constraint above is being tested as a modification to the LNS method,
denoted LNSSAS , but could equally be employed in a modified DESSAS method by using
Equation (4.28) in place of the usual grid scale (4.25). Current experience suggests that
modification (4.28) can be unpredictable in some cases allowing the length scale to
diminish in situations in which a preserved RANS layer was desired, and in other cases
preventing the length scale from diminishing in flows displaying only weak instabilities.
In the latter case, LNSSAS or DESSAS methods may never breed any unsteady motions,
and hence such modifications make a priori identification of resolved and unresolved
noise sources difficult. Furthermore, this approach puts the interface region between
RANS and LES at the most critical location in the flow the very point at which the
primary separations are initiated.
A simpler approach, which has proven effective in certain LNS calculations, is to
explicitly enforce a minimum filter width

(4.29)
L = 2 max x, y, z, L
min
in which L
min is a user-supplied length scale corresponding to the size of the largeeddy resolving cells (typically located away from the near wall). If suitably chosen, this
can prevent the rapid switching from RANS LES RANS in regions of abrupt
near-surface mesh clustering, thereby preserving the near-wall RANS layer.
For many problems, however, a more durable solution would be to initiate the transfer
of data to the LES mode some distance upstream of the interaction region, which would
require the use of recycling techniques, LES databases, or synthetic reconstruction or
P1: IBE
CUFX063-04
CUFX063/Wagner
142
0 521 87144 1
printer: sheridan
October 23, 2006
forcing using available RANS statistics. Research in the area of synthetic turbulence is
growing rapidly, and a brief review is provided in Section 4.2.7.
4.2.6 Modeling the noise from unresolved scales
The aim of time-dependent techniques such as LES or hybrid RANSLES is to account

only for the largest scales of motion and hence, by implication, the most significant
noise-generating scales. There remains, inevitably, an upper bound on the frequency
that can be predicted directly via such a time-dependent approach as a result of finite
spatial and temporal resolution. The objective of the global hybrid RANSLES methods
is to reduce the overwhelming computer cost of traditional LES by resorting to RANS
in certain locations (such as the near-wall region). Thus, the range of scales predicted
by a global hybrid RANSLES technique is likely to be less than that of a full-domain
or zonal LES (if it is assumed these were computationally affordable); hence, we may
expect the problem of missing noise from unresolved scales to be exacerbated by the
use of global hybrid RANSLES. The coarse mesh limit of hybrid RANSLES may
still adequately reproduce turbulence statistics but will leave the corresponding scales
of motion unresolved. It is therefore necessary to provide some mechanism to extract
noise from the unresolved fraction of hybrid RANSLES simulations.
To facilitate this, we consider the noise sources to be split into resolvable and
unresolvable components with all resolvable components assumed to be accounted for
via the hybrid RANSLES approach. This step is trivial with a constant filter length
but can be complicated by the use of ambiguous or solution-dependent (and hence
time-varying) filter lengths, as in the approach of Menter et al. (2003).
Because analytic methods are needed to generate and propagate unresolved noise
sources, Lighthills equation is considered written in tensor form as follows:
2
2 Ti j
2
2
c
=
,
2
2
t
xi x j
xi
in which

2
Ti j = u i u j i j + p c
i j .
(4.30)
(4.31)
The usual assumption (in what is otherwise an exact mathematical manipulation of the
NavierStokes equations) is that the right-hand-side terms are known and independent
of the left-hand side, which then simply represents a wave-propagation operator. A
solution to this equation was proposed by Curle (1955) and a more general extension
to moving surfaces by Ffowcs Williams and Hawkings (1969). These two approaches
are essentially equivalent for fixed (nonmoving surfaces). Curles approach is based on
a solution to the inhomogeneous wave equation (4.30) on a finite domain:
"""
1
1 2 Ti j

(x j , t) =
dV
2
4 c
r yi y j
""
1
1 r
1 r
1
+ 2 +
d S,
(4.32)
4
r n r n
cr n
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
where n defines the local surface normal. A solution to Curles equation, which is valid
for arbitrary r , can be written as (e.g., Larsson 2002)

"""
li l j 2 Ti j
3li l j i j
3li l j i j Ti j
1

+
+
Ti j d V
p (xi , t) =
2 r t 2
4
c
c r 2
t
r3

""
pi j i j
i j
1
p
1
+
i j
+
d S.
(4.33)
li n j
4
cr t
t
r2
Given a suitable mathematical model for the subgrid velocity fluctuations, Equation (4.33) can be used to completely determine the acoustic pressure generated and
transmitted from the unresolved scales. Representative results presented later in this
chapter were generated using the modeled, synthetic reconstruction of Batten, Goldberg,
and Chakravarthy (2004), which is presented in detail in Section 4.2.7.
4.2.7 Synthetic reconstruction of turbulence
This section considers some recent proposals for generating a field of synthetic, or
artificial, velocity fluctuations from a given set of turbulence statistics typical of those
obtained from a classical RANS method. Synthetic turbulence is required in several
areas critical to LES and acoustics:
(i) The generation of unsteady, turbulent inlet velocity fields particularly for inhomogeneous (e.g., spatially developing) flows.
(ii) The stimulation of large-scale (resolvable) eddy structures at coarsefine interfaces
or mesh clusterings in hybrid RANSLES.
(iii) The generation of full- or part-spectrum synthetic noise sources for use with
numerical or analytic acoustic wave propagation methods.
The prescription of an unsteady inlet condition for an LES is probably the most
pressing of the issues above. Obtaining quality unsteady inlet data (consistent with the
local numerical resolution and NavierStokes equations) is a problem that is becoming
increasingly important because of recent emphasis on hybrid RANSLES methods. For
LES or hybrid RANSLES, the imposition of a steady or quasi-steady inlet velocity field
is always incorrect for a fully turbulent flow; however, the resulting errors may not be
significant in situations in which a strong, inherent instability exists downstream and is
able to overwhelm any upstream disturbances. Much of the early work on hybrid RANS
LES concentrated on exactly this type of flow, but, because the range of applications
has widened, the issue of unsteady boundary conditions has become more acute.
A synthetic reconstruction of turbulence is attractive because it offers an inexpensive
method of generating an unsteady flow field from any localized set of RANS statistics.
For inlet boundary conditions, the main alternative is the use of a separate, auxiliary
section of LES, which may contain more physics than a synthetic model but adds a
significant overhead to the calculation and creates a large burden for the end user in
specifying the separate LES zone, the positioning of any overlapping or recycled inlet
17:51
143
P1: IBE
CUFX063-04
CUFX063/Wagner
144
0 521 87144 1
printer: sheridan
October 23, 2006
planes, and the interpolation and rescaling of mean (and possibly fluctuating) quantities
to account for developing boundary layers or other inhomogeneities.
The synthetic method outlined in Equation (4.34) was developed by Batten,
Goldberg, and Chakravarthy (2004) as an anisotropic extension to Kraichnans (1969)
method. This method allows a reconstruction of spatially and temporally coherent
turbulent velocity fluctuations from a given dissipation rate and set of second moments. This formulation follows from a simplification of the tensor scaling proposed
by Smirnov, Shi, and Celik (2001) in which the need to compute similarity transformations is avoided:

u i (x j , t) = aik
N

2 #
pkn cos d nj x j + n t + qkn sin d nj x j + n t
N n=1
(4.34)
V
x j = 2 x j /L , t = 2 t/, dnj = d nj n , V = L/,
c

3 n n n n
cn =
u u d d /d d , pin = i jk nj dkn , qin = i jk nj dkn ,
2 l m l m k k

1
.
in , in = N (0, 1), n = N (1, 1), din = N 0,
2
In the preceding equation, ai j is the Cholesky decomposition of u i u j . The Cholesky
decomposition is the lower triangular matrix ai j , with aik akTj = u i u j . For a symmetric,
positive, definite Reynolds stress tensor, u i u j , ai j can be determined as

u1u1
0
0

2
(4.35)
ai j = u 1 u 2 /a11
.
u 2 u 2 a21
0

2
2
u 1 u 3 /a11 (u 2 u 3 a21 a31 )/a22
u 3 u 3 a31
a32
Note that all elements of the Cholesky decomposition are real if the target stress tensor
is realizable.
In the synthetic model Equation (4.34) above, the time and space correlations are
represented by scaling the local time and distance coordinates in the Fourier model by
the local turbulence time scale and (anisotropic) velocity scale cn , which is a tensorially invariant measure in the direction of the modal wave vector din . This feature tends to
elongate the synthetic eddies preferentially in the direction of the strongest correlations,
producing spherical-shaped eddies in which the turbulence is closer to isotropy (typically, away from walls) and flatter eddies near solid surfaces where the normal-to-wall
velocity tends to be preferentially damped by inviscid blocking effects. In the long-time
average, the synthesized time-dependent flow field will reproduce the given length and
time correlations and all second moments. The storage requirements for this method
are dictated only by the number of modes and are independent of the grid size.
Because each mode in this Fourier reconstruction has an associated wavelength
(the wave-vector modulus), it is possible to reconstruct fluctuations over the full spectrum, or, by selectively summing over each mode (e.g., if L/|d n | < L ), it is possible
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
145
to reconstruct only that portion of the spectrum that would remain unresolved in an
LES or hybrid RANSLES. It is, of course, equally possible to reconstruct only those
components that are resolvable on the local mesh for example, in hybrid RANSLES
methods in which the finer scales are treated via a subgrid model. The latter situation
is particularly relevant to hybrid RANSLES calculations in regions in which the filter
width suddenly (or gradually) diminishes, implying a potential switch from RANS to
LES mode. Another useful feature of this mathematical description is that the modeled function is continuously (and easily) differentiable, making it simple to compute
the temporal or spatial derivatives, or both, needed for analytic acoustic propagation
methods such as Equation (4.33).
An important practical consideration for inflow conditions is the subsequent spatial
evolution of the synthetic turbulence before its arrival at the interaction region and, in
particular, the preservation and transmission of the imposed target statistics. Synthetic
inlet data can be expected to undergo a transient period during which physical structures develop. Experience suggests that the turbulence energy and shear stresses can
sometimes undergo an initial decay during this transient. This decay may be partly due
to the presence of nonphysical structures but also results from numerical and (modeled) subgrid viscosity, which may damp signals at wavelengths that the reconstruction
criteria in the synthetic turbulence expected to be resolvable (as determined by the
SGS model and numerical scheme employed, the smallest supported structures may be
significantly larger than the filter width L ).
Recent work by Spille-Kohoff and Kaltenbach (2001) and Keating et al. (2004)
suggests that an imposed forcing at a sequence of control planes downstream of the
inlet can help to mitigate unwanted decays in the shear stress from synthetically reconstructed inlet data. The controlled forcing approach currently lacks some generality
and burdens the user with the task of imposing appropriate data at the various control
planes; however, the method is much less expensive than using an auxiliary LES. Future
research is expected to improve the quality of synthetic turbulence data, which should
reduce or eliminate the need for additional forcing terms.
4.2.8 The NLAS approach of Batten, Goldberg, and Chakravarthy
An alternative form of hybrid method was presented by Batten et al. (2002c) and Batten,
Ribaldone et al. (2004) as a means of reducing the diffusive effect of the subgrid model
and directly accounting for the generation of acoustic waves from SGS structures. The
derivation of this NLAS begins with the NavierStokes equations:
Fiv
Fi
Q
= 0,
+
t
xi
xi
Q = u j ,
e
u i
Fi = u i u j + pi j ,
u i (e + p)
17:51
(4.36)
0
.
Fiv =
i j
i + u k ki
P1: IBE
CUFX063-04
CUFX063/Wagner
146
0 521 87144 1
printer: sheridan
October 23, 2006
In aeroacoustics, manipulations of these equations are usually made in order to

generate simplified wave equations that are amenable to analytic solutions (given a
separate hypothesis and model for the noise sources). Here, we consider a system of
equations that is formally identical to the NavierStokes equations (i.e., containing
no simplifying assumptions or elimination of variables) but is derived by considering
each primitive variable as being split into mean and fluctuating quantities, = + .
Substituting into the NavierStokes equations and rearranging for fluctuation and mean
quantities gives a system of perturbation equations referred to as nonlinear disturbance
equations (NLDE):

v
Fiv
Fi
Q
Q
Fi
Fi
=
+
,
(4.37)
+
t
xi
xi
t
xi
xi
Q = uj ,
e
ui
F i = u i u j + pi j ,
u i (e + p)

Q = u j + u j + u j ,
e

v
Fi
v
Fi
0
,
=
ij
i + u k ki
0
,
ij
=

i + u k ki + u k ki
u i
u i + u i
Fi = u i u j + u i u j + u i u j + p i j + u i u j + u i u j + u i u j + u i u j .
u i (e + p) + u i (e + p )
u i (e + p )
Neglecting density fluctuations and taking time averages of the preceding system
of equations cause the evolution terms and all flux terms linear in the perturbations to
vanish, resulting in
LHS = RHS =
in which
Ri
xi
(4.38)
u i u j
Ri =
.
1

c p T u i + u i u k u k + 2 u k u k u i + u k ki
The equations above correspond to the standard Reynolds stress tensor and heat flux
terms. The key step is to obtain these unknown terms in advance from a classical RANS
method. The unresolvable (short-wavelength) contribution to these terms can then be
generated using a synthetic model of turbulence such as that described in Section 4.2.7.
With these mean quantities and statistics established, time-dependent computations can
then be made to determine the perturbations about this mean using Equation (4.37).
Naturally, the time-dependent computation may reveal more physics than was available
in the precursor (steady-state) RANS calculation; hence, the time-averaged fluctuations
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
147
DNS
LES
Hybrid
RANSLES
NLAS
NLDE
Figure 4.2. Required near-wall mesh resolutions with DNS, traditional LES, global hybrid RANSLES,
and nonlinear acoustics solvers (NLAS) based on disturbance equations.
from the second, time-dependent calculation are not expected to agree fully with the
initial RANS mean data.
The corresponding inviscid system of disturbance equations was first proposed by
Morris, Long et al. (1997), and there may be many situations in which inviscid disturbance equations are sufficient particularly when using separate acoustics meshes
in which cell sizes are too large to properly resolve viscous layers. A grid-converged
Reynolds stress tensor can still be imposed through the background (RANS) field, but
the resulting stresses (in addition to cell-size-related numerical viscosity), are likely to
dominate the effects of the physical viscosity on such meshes, making the retention
of viscous terms unnecessary. However, more recent computations involving disturbance equations (e.g., Batten et al. 2002c; Batten, Ribaldone et al. 2004; Labourasse
and Sagaut 2002) have favored the more general approach of retaining viscous
terms.
If the local mesh is unable to support any resolved-scale fluctuations (e.g., in coarse
or high-aspect-ratio cells), the NLAS method of Batten et al. (2002c) can still access a set
of (grid-converged) RANS statistics to provide a description of the subgrid. In the NLAS
method, these data are provided in the form of a synthetic reconstruction of the RANS
statistics (see Section 4.2.7). This NLAS method offers several interesting advantages
over both traditional LES and more conventional hybrid RANSLES methods such
as DES (Section 4.2.2) or LNS (Section 4.2.3). NLAS provides a more sophisticated
subgrid treatment that allows the extraction of acoustic sources from the temporal
variation within the (modeled) subgrid structures; grid requirements can be relaxed in
the near-wall region during the NLAS transient calculation (the quasi-steady near-wall
RANS solution being obtained a priori; see Figure 4.2); the dissipative effects of a
subgrid eddy viscosity model are avoided; thus, on coarser meshes, NLAS proves less
diffusive. In terms of cost, NLAS is competitive with linearized Euler equation (LEE)
and acoustic-perturbation equation (APE) methods (e.g., Ewert, Meinke, and Schroder
2001b) having very similar mesh resolution and flux evaluation requirements, but the
17:51
If the only interest were the noise emission from the initial RANS solution, the NLDE mean could be
forced to agree with the RANS mean by subtracting out an averaged source term at each time step. In
general, however, the objective is also to use the available information on the unsteady flow physics to
improve the mean-flow predictions.
P1: IBE
CUFX063-04
CUFX063/Wagner
148
0 521 87144 1
printer: sheridan
October 23, 2006
synthetic source reconstruction is less costly in NLAS owing to the imposed filtering.
In addition, NLAS has the important advantages over LEE- and APE-based methods
in that it can account both for broadband, turbulence-related noise and discrete tones
arising from coherent structures or resonance, and (like traditional LES) its empirical
modeling content decreases with increasing spatial resolution.
The use of background RANS statistics also has the potential to reduce the far-field
mesh requirements. In the commercial CFD++ code (see Peroomian et al. 1998), NLAS
can operate on a subdomain with absorbing far-field boundaries defined by a truncated
subset of the original RANS calculation, which provides not only turbulence statistics
but also spatially varying mean-field data that can be used to place the outer boundary
conditions much closer to the regions of interest.
Experience with methods based on nonlinear disturbance equations and a priori
RANS calculations is less extensive than with DES (Section 4.2.2) or LNS (Section 4.2.3), and some issues have yet to be resolved. For example, the near-wall treatment might be brought into question if RANS and NLDE mean flows were substantially
different in separation or reattachment locations. It may also prove difficult to produce
a statistically steady RANS solution for flows that have an inherent, violent unsteadiness (although conventional hybrid RANSLES methods are likely to be well suited
to this category of flow). Nonlinear disturbance equations are, however, expected to
outperform more traditional hybrid RANSLES methods on flows with weak shear
instabilities.
4.3 Zonal hybrid approaches

Zonal hybrid methods are based on a purely discontinuous treatment between RANS
and LES approaches. This is achieved by means of a domain decomposition, with
each domain being solved in either pure RANS or pure LES mode, as dictated in
advance by the user. The previous global approaches assume a continuous treatment
between the RANS and LES modes and thus introduce an acknowledged gray area
in which the approach may act as neither RANS nor LES. The use of pure LES and
pure RANS zones may remove the issue of these gray areas from within the zones;
however, it also creates a clear requirement to account for the discontinuous boundary
condition between the two distinct descriptions of the flow given by RANS and LES.
This required two-way interaction typically satisfies conservation only in the long-time
average because the RANS zones require boundary conditions comprising averaged
(statistical) data, whereas the LES zones require boundary condition data containing
the full range of unsteady motion, which must ultimately be consistent with the RANStransported statistics, including local scales of length and time. Although such internal
boundary conditions are a clear requirement for zonal methods, similar treatments are
also urgently needed in certain global hybrid RANSLES applications. For example,
in the case of an abrupt mesh refinement (i.e., if the global hybrid approaches were
deployed on the same zonal meshes) the two approaches would exactly coincide because
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
17:51
149
4.3 ZONAL HYBRID APPROACHES
the finer mesh in the LES zones would force a switch to LES behavior within the global
RANSLES hybrids.
Pure zonal methods, if sufficiently well resolved, allow the possibility of directly
generating a locally accurate description of the unsteady noise sources in the LES
zones with less (or perhaps no) modeling being required to extract the full spectrum of
near-field noise (far-field radiation will still require additional procedures). However,
because of the differing near-wall resolution requirements, and therefore t are
expected to be much smaller in zonal calculations than in calculations using global
hybrid RANSLES approaches. Finer resolution in the LES zones would mitigate the
requirement for additional modeling to extract noise from unresolved scales but would
again require much larger (possibly prohibitive) computing resources.
The difficult (and, as yet, not fully solved) problem of the coupling between RANS
and LES regions currently reduces the range of applicability for both global and zonal
methods. The simplest possible coupling relies on a direct injection of the RANS
values at the boundaries of the LES zone and an average (or simply a direct injection;
Georgiadis, Alexander, and Reshotko 2003) of the LES field to create a boundary
condition for the RANS region. However, such a simple coupling would be of limited
practical use. Failing to take into account the intrinsically different natures of the RANS
and LES descriptions of the flow field would prevent the transmission of turbulence
data (in either resolved or statistical form) from one region to the other. An LES data
representation will generally contain much more information than the corresponding
RANS data, and thus the RANS field cannot be used to generate data for the LES
region without additional modeling, which must include further assumptions on local
length scales, time scales, and energy distributions. This topic is discussed further in
Section 4.3.3.
Recent work carried out by Quemere and Sagaut (2002) and Labourasse and Sagaut
(2002) involves zonal approaches that consider discontinuous boundaries between
RANS and LES regions. Both these methods also adopt the use of NLDE, which
can be viewed as an extension of the multilevelmultiresolution formalism of Harten
(1994, 1996) and Terracol et al. (2001) in which two levels of solution resolution are
introduced by the respective RANS and LES data representations. Each LES-filtered
(resolved-scale) flow quantity is decomposed into
filtered = + ,
(4.39)
where the overbar denotes the Reynolds average and is the frequency complement
between the RANS and the LES representation of . This last quantity is often referred
to as the detail in the multilevel representation. The complete field can be considered
as being decomposed using a triple decomposition:
complete = + + SGS ,
(4.40)
where SGS refers to the additional (unknown) detail modeled by the subgrid terms.
P1: IBE
CUFX063-04
CUFX063/Wagner
150
0 521 87144 1
printer: sheridan
October 23, 2006
4.3.1 The approach of Quem

er
e and Sagaut
The approach proposed by Quemere and Sagaut (2002) was one of the first attempts to derive a consistent discontinuous coupling between the two (RANS and
LES) approaches. The authors consider a global RANS domain in which a small
overlapping region is devoted to LES. In this region, a two-way coupling exists as
follows:
r The information provided by the RANS solver is used to derive some boundary
conditions for the LES domain. To take into account the discontinuity between
the two fields, the authors have proposed two procedures to synthesize turbulent
fluctuations. The first one relies on an extrapolation of the details ( ) computed in
the LES region in the ghost cells of the LES domain. In these cells, a reconstruction
of is then performed with the averaged field given by RANS. The authors
indicate that such a treatment is only valuable in the case of lateral or outflow
boundaries. For inflow boundaries, the authors have proposed to use a predictor
simulation and to apply the rescaling suggested by Lund et al. (1998) (see the
description at end of this section).
r A feedback from the LES region to the steady RANS or unsteady Reynolds-averaged
NavierStokes (URANS) region also exists. This is achieved by averaging the LES
field to provide data for the RANS field boundary conditions in the overlap region.
In this region, traditional RANS transport equations are used to determine the
quantities k, , and t (Quemere and Sagaut 2002).
The potential of this method was first demonstrated on plane-channel flow calculations
in cases in which the LES regions corresponded to the near-wall layer or to the core
region of the channel. In the latter case, the underlying idea of the method is very close
to DES (in which RANSLES regions are effectively fixed a priori by the wall-distance
function and choices of local cell sizes). Finally, the authors have applied their approach
to the simulation of a base flow with a small LES region located at the trailing edge of
a flat plate computed by URANS.
4.3.2 The approach of Labourasse and Sagaut
The approach proposed by Labourasse and Sagaut (2002), is again based on the solution
of the NLDE in certain zones, and an initial (or separate) RANS simulation is used
to provide the global mean-field and boundary conditions for the LES domains. In the
terminology used by Labourasse and Sagaut (2002), the NLDE read
Fiv

Fi
Q
Fi Fiv
+
Q +Q
Q + Q SGS =
Q
Q RANS .
t
xi
xi
xi
xi
(4.41)
The two terms SGS and RANS denote, respectively, unknown terms in the LES
and RANS equations that both require modeling. The system of equations (4.41),
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
151
which describes the temporal evolution of the details or difference between the LES
and RANS solutions, is then solved in the LES region instead of the classical LES
equations. This minimizes the sensitivity of the solution to numerical errors or to
the use of relatively coarse discretization grids, or both. In addition, the perturbation
field Q can sometimes display less sensitivity to the numerical treatment at boundaries than the total field values Q + Q (because any errors are then committed on a
smaller fraction of the total field value). Indeed, the correlation lengths and times associated with the perturbations are smaller than those of the full-field variables, which
can lead to a more rapid damping of the numerical errors at boundaries. However,
the authors have shown (Sagaut et al. 2003) that a careful treatment is still needed to
prevent strong reflections at the interfaces of the LES-like (viscous NLDE) domain
for both the unsteady turbulent structures and acoustic waves. The authors have thus
proposed to use an extension of existing characteristic theory applied to the perturbation equations. In the nonconservative form, the NavierStokes equations can be
written
V
V
V
V
+A
+B
+C
= VIS
t
x
y
z
(4.42)
where V = (, u 1 , u 2 , u 3 , p)T is the primitive variable vector, VIS accounts for the
viscosity, and SGS terms and A, B, and C are the classical convective-flux Jacobian
matrices. V can be divided into three parts according to the NLDE decomposition:
V = V + V + VSGS .
(4.43)
The first part deals with the mean field, whereas the second part is connected with the
fluctuating field V = ( , u 1 , u 2 , u 3 , p )T .
As in all other works dealing with the subject, the contribution of the SGS term
VSGS is neglected in this characteristic analysis. This leads to the following equation
for the fluctuating field:
V
t
VIS
+ A Vx + B Vy + C zV =
V
t
17:51

+ A Vx + B Vy + C zV VIS ,
(4.44)
where VIS and VIS account, respectively, for the fluctuating and mean part of the
viscosity terms.
By analogy with the extension of the characteristic theory to viscous flows (e.g.,
Poinsot and Lele 1992), only the convection terms of the fluctuations are considered,
which leads us to neglect the right-hand side of Equation (4.44).
The resulting equation is
V
V
V
V
+A
+B
+C
= 0.
t
x
y
z
(4.45)
P1: IBE
CUFX063-04
CUFX063/Wagner
152
0 521 87144 1
printer: sheridan
October 23, 2006
The matrix E = An x + Bn y + Cn z (where n = (nx , ny , nz )T is the vector normal to

the considered boundary) has to be diagonalized to set conditions on the characteristic
variables of the equation:
= L E L 1 = diag(1 , 2 , 3 , 4 , 5 )
= diag(u n , u n , u n , u n + c, u n c),
(4.46)
where u n is the projection of the velocity vector on the normal to the boundary, c is the
sound velocity, and L is the matrix of the left eigenvectors. The temporal evolutions of
the primitive variables (V ) and those of the characteristic variables are then linked by
W = LV or V = L 1 W ,
(4.47)
where W is the vector of characteristic variables.

A nonreflecting boundary condition for the fluctuating field can then finally be
derived following Thomson (1987) as
0
if i 0 ,
W
(4.48)
W = W n+1 W n
t
if i > 0 .
t
The approach proposed by Labourasse and Sagaut (2002) has been applied successfully in both academic and more realistic configurations. Initial computations were
performed on a plane-channel flow configuration and revealed the ability of the method
to handle coarser meshes than regular LES and to work with reduced LES domains
(Labourasse and Sagaut 2002). Several computations have also been performed in
the case of a low-pressure T106 turbine blade (Labourasse and Sagaut 2002; Sagaut
et al. 2003), where a small LESNLDE computational domain was localized at the
blades trailing edge while the overall configuration was treated by the RANS approach. These computations have allowed the authors to perform a critical analysis of
the boundary conditions used for the LESNLDE domain. The result of this analysis is
that a nonreflective boundary treatment, such as the one described in Equation (4.48),
is necessary to allow the correct transmission through the boundary of both turbulent
vortical structures and acoustic waves. The approach has also been applied more recently to the case of the unsteady flow in the slat cove of a high-lift airfoil (Terracol
et al. 2003). The computations performed in this case have enabled an analysis of the
various flow phenomena responsible for acoustic wave emission. Some of the results
obtained thus are reported in Section 4.4.
4.3.3 Zonal-interface boundary coupling
Despite the reported success of the zonal RANSLES approaches described in this chapter, the coupling between the two (RANS and LES) approaches at the domain interfaces
remains an open problem. Indeed, the numerical treatment proposed by Labourasse and
Sagaut (2002) can only be applied to lateral or outflow interfaces (where the mean flow
is parallel to the interface or passes from the LES to the RANS region). However, the
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
numerical treatment of inflow sections (where fully turbulent flow enters the LES region) remains a more challenging problem. The zonal approach proposed by Labourasse
and Sagaut (2002) has been successfully applied to application cases in which the flow
at the inflow boundary of the LES domain was quasi-laminar but would fail to accurately reproduce flows with turbulent inflow conditions in the LES region . To handle
turbulent flows at this interface, the method needs to incorporate a more detailed data
transfer procedure such as a synthetic reconstruction from the given RANS statistics.
Several approaches are possible for generating these inflow conditions for the LES
region. The simplest approach involves a superposition of white noise on the mean velocity profile. A further elaboration would be a white noise signal that would reproduce
the single-point statistics. However, most previously reported attempts in this direction
have failed owing to the rapid dissipation of Nyquist-frequency signals by the subgrid
treatment. Important requirements in this synthetic reconstruction are realistic spatial
and temporal correlations that allow the maintenance, regeneration, or both of physical
turbulent structures (Chung and Sung 1997).
Section 4.2.7 presents a method for synthesizing turbulent (RANS) data that reproduces the second moments as well as the two-point correlations. This method is already being applied to RANSLES interfaces (e.g., Batten, Goldberg, and Chakravarthy
2004).
For sufficiently simple geometries, an alternative strategy relies on the use of a
periodic recycling and rescaling such as that proposed by Lund et al. (1998) to generate
simulation inflow data for incompressible turbulent boundary layers. This strategy has
been extended to compressible flows by several authors (Sagaut et al. 2003; Urbin and
Knight 2001; Stolz and Adams 2001; Schroder et al. 2001) and has also been retained
by Quemere in his zonal RANSLES approach. The idea of this treatment is to extract
the turbulent fluctuations in a plane downstream of the inflow interface location and
to inject them at the inflow after a rescaling based on the ratio of the skin-friction
velocity and thickness at the two different locations. The rescaling approach is harder
to implement in general (i.e., in complex geometries) and, as determined by the scaling,
may not exactly reproduce the intended second moments, length scales, and time scales
but has the advantage (relative to most existing synthetic models) of requiring a shorter
transient period to reestablish self-sustaining physical disturbances.
The methods discussed in this section for imposing inflow LES boundary conditions
are relatively new and have had little application in the context of acoustic source
prediction. However, there is every justification to expect reasonable results. A good
synthetic model of the turbulent fluctuations at the inflow interface of the LES domain
should allow equivalently good predictions both of the flow downstream and of the
direct noise emission from that region of synthetic turbulence. It is imperative, however,
Treatment of LES inflow sections is also a critical issue for global hybrid RANSLES methods. The
hybrid RANSLES predictions of Fan et al. (2002, 2003) failed dramatically without an appropriate
imposition of the unsteady approach boundary layer.
Another solution is to use a second simulation with a periodicity condition and a forcing term in the
streamwise direction to provide the fluctuations. See Quemere and Sagaut (2002) for example.
17:51
153
P1: IBE
CUFX063-04
CUFX063/Wagner
154
0 521 87144 1
printer: sheridan
October 23, 2006
to ensure that these fluctuations are not additive that is, the underlying data must
be represented either as RANS (with no fluctuations) or as LES with synthesized
fluctuations (and appropriately diminished levels of statistical turbulence energy k)
and that any synthesized disturbances are then allowed to develop naturally as part of the
solution to the time-dependent NavierStokes equations. Reconstruction of unsteady,
synthetic turbulence fields is discussed further in Section 4.2.7.
4.4 Examples using hybrid RANSLES formulations

This section presents selected examples of calculations using some of the described
hybrid RANSLES formulations. The first example considers the application of global
hybrid RANSLES approaches to the wake of a car wing mirror together with the use of
a synthetic reconstruction to account for the noise from unresolved scales. The second
example presents results from a zonal NLDE approach to the simulation of the flow in
a slat cove of a high-lift wing device.
4.4.1 Flow in the wake of a car wing mirror
This example illustrates the use of hybrid RANSLES in predicting resolved-scale noise
sources with a synthetic reconstruction being used in conjunction with an analytic wave
propagation approach to account for noise from the unresolved scales. A second set of
results demonstrates the application of Batten, Ribaldone et al.s (2004) NLAS. The
problem consists of a low-speed (55.5 m/s) air flow over a hemispherical half-cylinder
mounted on a flat plate representative of a generic car wing mirror. Predictions are
compared with experimental measurements taken on the downstream surface of the
mirror. The hybrid approaches used here are the LNS and NLAS methods described in
Sections 4.2.3 and 4.2.8, respectively.
The mesh used for this case was intentionally coarse, consisting of approximately
400,000 hexahedral elements with wall functions employed on all solid surfaces. Three
layers of absorbing cells were used adjacent to all far-field boundaries by the introduction
of a source term of the form K (U U ), where K is some damping coefficient and
U corresponds to the far-field data used to define the boundary condition. For the
computation of the unresolved scales (used in conjunction with LNS), the synthetic
reconstruction described in Section 4.2.7 was carried out on a reduced mesh generated
by cutting away cells outside an isosurface containing 99% of the total mean turbulence
energy, as predicted by an a priori steady-state RANS calculation (refer to Section 6.8
for alternative methods of identifying dominant sound-source regions).
Figure 4.3 shows the initial startup transient computed by the LNS hybrid method
compared with its baseline (cubic) RANS model at a probe location (111) on the downstream face of the cylinder. Both hybrid and conventional RANS calculations used
the same mesh, the same time step, and the same spatial and temporal integration
17:51
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
4.4 EXAMPLES USING HYBRID RANSLES FORMULATIONS
250
Resolved signal (hybrid RANSLES)
Resolved signal (URANS)
200
150
100
Preusser Perturbation (Pa)
P1: IBE
CUFX063-04
50
0
50
100
150
200
250
0.1
0.2
0.3
Time (s)
0.4
0.5
Figure 4.3. Initial startup transient at probe 111 (cylinder rear face) predicted by unsteady RANS and
hybrid RANSLES (LNS model) using identical base model, mesh, and time step.
procedures. Nevertheless, Figure 4.3 shows a large difference in the predicted amplitudes at this probe location. Specifically, the unsteady RANS solution predicts very
regular oscillations but of such low amplitude they are not visible to the eye when
plotted on a scale comparable with the LNS output.
Figure 4.4 shows an instantaneous isosurface of the streamwise vorticity shaded
with the streamwise velocity. The coarseness of the mesh is evident in the exclusively
Figure 4.4. Instantaneous streamwise vorticity contours (with streamwise velocity shading) predicted
by hybrid RANSLES model.
17:51
155
CUFX063/Wagner
156
0 521 87144 1
printer: sheridan
October 23, 2006
300
Unresolved (synthesized) signal
Resolved (hybrid RANSLES) signal
200
Pressure Perturbation (Pa)
P1: IBE
CUFX063-04
100
100
200
0.1
0.2
0.3
0.4
0.5
Time (s)
Figure 4.5. Resolved and unresolved (synthetically generated) signals for probe 111 (cylinder rear
face).
large-scale structures predicted. As a result, higher-frequency components will be missing from the LNS predicted data; hence, a reduced domain was used to synthesize the
missing high-frequency noise components as discussed in Section 4.2.6.
Figure 4.5 shows the resolved and synthesized noise components at probe 111.
Their respective Fourier transforms are shown in terms of sound-pressure levels in
Figure 4.6. This figure also illustrates the signal-processed output from the RANS and
LNS simulations in isolation. The isolated LNS predictions fall slightly under the data at
higher frequencies as expected. The unsteady RANS predictions show a single discrete
tone with a significant underestimation of the broadband noise levels. The composite
signal shows the closest agreement with data, and the discrete nature of the energy
distribution in the synthetic reconstruction model is evident in the higher-frequency
components.
Figure 4.7 shows an instantaneous isosurface of the streamwise vorticity computed
using the NLAS method. This calculation was performed on the same mesh (but with
the outer, absorbing layer boundaries slightly truncated) and using the same choice
of time step as the hybrid RANSLES calculation. A comparison of Figures 4.7 and
4.4 shows the increased resolution obtained with the acoustics solver relative to the
hybrid RANSLES. For any finite mesh resolution, one can expect a high-frequency
cutoff for both traditional hybrid RANSLES and NLAS methods; however, at these
low flow speeds (approximately Mach 0.1), that limit is an order of magnitude less
17:51
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
150
Composite signal
Resolved signal (hybrid RANSLES)
Resolved signal (URANS)
Exp. - Probe 111
140
130
120
-5
SPL (dB), Pref = 2e Pa
P1: IBE
CUFX063-04
110
100
90
80
70
60
50
10
100
Frequency (Hz)
1000
Figure 4.6. Sound-pressure levels determined by resolved, unresolved, and composite signals for
probe 111 (cylinder rear face).
restrictive with NLAS, which contains a model for the time-dependent variation of the
noise-generating flow structures at subgrid scales.
Figure 4.8 shows the NLAS-predicted sound-pressure levels at probe 111. The levels
compare quite favorably with the basic hybrid RANSLES predictions, showing less
suppression of the higher frequencies.
Figure 4.7. Instantaneous streamwise vorticity contours (with streamwise velocity shading) predicted
by the nonlinear acoustics solver (NLAS).
17:51
157
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
150
140
Exp. - probe 111

NLAS - probe 111
130
-5
158
SPL (dB), Pref = 2e Pa
P1: IBE
CUFX063-04
120
110
100
90
80
70
10
100
Frequency (Hz)
1000
Figure 4.8. Sound-pressure levels determined by NLAS method at probe 111 (cylinder rear face).
4.4.2 Unsteady flow in the slat cove of a high-lift airfoil
This section describes the unsteady numerical simulation in the slat cove of a high-lift
wing. A three-element high-lift wing with deployed slat and flap is considered with the
following physical parameters: the wing chord (in cruise configuration) was equal to
0.61 m, and the upstream velocity was 65.5 m/s, leading to a chord Reynolds number
of 2.5 million.
A steady, two-dimensional RANS computation was first carried out over the full
high-lift configuration using the SA model on a 411,664-point grid. The instantaneous three-dimensional turbulent fluctuation field was then computed using the zonal
RANSLES method of Labourasse and Sagaut (2002) (described in Section 4.3.2)
in a small region encompassing the slat cove (see Figure 4.9). It is worth noting that
the LES subdomain is, itself, decomposed into several computational grid blocks.
The hybrid RANSLES simulations were carried out on three-dimensional meshes.
The grid distribution was the same as for the RANS simulation in the (x, z) plane
(with 110,000 points per plane). Several simulations were performed corresponding
to different spanwise extents (2.7 and 27% of the slat chord), spanwise mesh sizes
(32 or 52 points), or both. The fluctuating field was assumed to be periodic in the
spanwise direction. As mentioned previously, both explicit and implicit simulations
were performed to get a broadband description of the flow physics. In this chapter, only
the results with the 27% chord extent with 52 mesh points in the spanwise direction are
presented. Simulations with the reduced value of a chord length were found to remain
quasi-two-dimensional, and thus the span was clearly not large enough to represent the
transverse unstable modes.
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
Figure 4.9. Location of the LES subdomain (displayed mesh shows every eighth grid line).
Firstly, one can note that the hybrid RANSLES computation adjusts the mean field
in comparison with the quasi-steady RANS approach, as can be seen in Figure 4.10,
which illustrates mean-flow streamlines inside the slat cove. Two additional separation
zones are present in the hybrid RANSLES result: the first is just behind the slats
trailing edge; the second, more significant separation zone is found at the leading edge
of the main wing body. This highlights the difficulties of accurately simulating this type
of flow using traditional RANS methods. Indeed, Khorrami, Singer, and David (2002)
Figure 4.10. Mean flow streamlines. Left: RANS computation, Right: Hybrid RANSLES computation.
Taken from Terracol et al. (2005) (Fig. 27, p. 221) with kind permission of Springer Science and Business
Media.
17:51
159
P1: IBE
CUFX063-04
CUFX063/Wagner
160
0 521 87144 1
printer: sheridan
October 23, 2006
Figure 4.11. Instantaneous Schlieren-like view. Taken from

Terracol et al. (2005) (Fig. 28, p. 221) with kind permission of
Springer Science and Business Media.
found it necessary to switch off the model effects in the leading-edge region in order
to restore this second separation zone in their URANS computations.
An instantaneous, Schlieren-like view of the field in an (x, z) plane is shown in
Figure 4.11, and the dilatation field = u is displayed in Figure 4.12. This last
Figure 4.12. Isovalue contours of the dilatation field . Taken

from Terracol et al. (2005) (Fig. 29, p. 221) with kind permission
of Springer Science and Business Media.
17:51
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
110
100
90
80
70
1000
2000
3000
F (Hz)
4000
5000
17:51
161
p (db)
P1: IBE
CUFX063-04
6000
Figure 4.13. Acoustic pressure spectrum at location S2. Taken from Terracol et al. (2005) (Fig. 31,
p. 222) with kind permission of Springer Science and Business Media.
quantity is proportional to the acoustic pressure time derivative and permits observation
of very-low-frequency oscillations. These two figures make it possible to investigate
the flow structure and the resulting acoustic wave propagation. Looking at these figures,
one can observe several basic flow features associated with noise emission related to
specific frequencies in the pressure spectra:
r A mixing layer develops in the shear region that bounds the main recirculation
bubble on the suction side of the slat (S2). Owing to strong streamline curvature
effects, eddies of opposing spanwise vorticity are not symmetric. Pairing is also
observed. Figure 4.13 shows the acoustic spectrum obtained just behind the slats
trailing edge with an associated main frequency of 1.5 kHz, having subharmonics
(due to vortex pairing) and higher-order harmonics.
r A secondary recirculation bubble is observed on the slat suction side (S1) that
interacts with the coherent vortices swept along the slat surface by the main recirculation. The frequency spectrum observed at this location (figure not shown)
was extremely similar to that obtained in the mixing layer S2, suggesting a strong
interaction between the two flow features S1 and S2.
r An important secondary turbulent recirculation bubble is observed in the gap on the
main body surface (S3). Two different phenomena occur here: (1) a mixing layer
subjected to a KelvinHelmholtz instability with a main vortex shedding frequency
of about 10 kHz, and (2) an erratic, lower-frequency vortex shedding is detected
that could be explained by the interaction of the classic breathing mode of the
recirculation bubble with the chaotic advection of coherent vortices coming from
CUFX063/Wagner
162
0 521 87144 1
printer: sheridan
October 23, 2006
110
100
p (db)
P1: IBE
CUFX063-04
90
80
70
10
20
F (kHz)
30
40
50 60
Figure 4.14. Acoustic pressure spectrum in the recirculation bubble (location S3). Taken from Terracol
et al. (2005) (Fig. 33, p. 223) with kind permission of Springer Science and Business Media.
the mixing layer through the gap. Figure 4.14 presents the sound-pressure spectrum
obtained in the center of the recirculation bubble, which appears to be a full spectrum
owing to the high level of turbulence observed in the gap.
r The last phenomenon noted occurs in the slat wake (S4). Recalling that the slat
has a blunt trailing edge, we note that a coherent vortex shedding with a main
frequency of about 30 kHz is detected. A more precise analysis of this isolated flow
Figure 4.15. Isovalue contours of the dilatation field . Taken from Terracol et al. (2005) (Fig. 34,
17:51
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
110
100
90
80
70
60
50
40
30
20
10
10
F (Hz)
17:51
163
4.5 SUMMARY OF HYBRID RANSLES METHODS
p (db)
P1: IBE
CUFX063-04
10
Figure 4.16. Acoustic pressure spectrum at the slats trailing edge (location S4). Taken from Terracol
et al. (2005) (Fig. 35, p. 224) with kind permission of Springer Science and Business Media.
feature has also been carried out by means of a numerical simulation using a smaller
computational domain localized around the slats trailing edge. This simulation took
into account the boundary layer on the upper side of the slat, and the influence of
the turbulent bubble in the gap was removed. A global view of the dilatation field
obtained in this case on the full computational domain considered is shown in
Figure 4.15. An acoustic wave pattern is clearly emitted at the trailing edge with a
frequency of about 30 kHz, as mentioned previously. This frequency is associated
with the vortex shedding and recovered in the acoustic spectrum at the trailing
edge (Figure 4.16). In the wake, the main frequency quickly becomes about 15 kHz
(Figure 4.17), which is explained by the vortex pairing in the wake.
4.5 Summary of hybrid RANSLES methods

In this chapter we have attempted to summarize several hybrid RANSLES approaches that show promise for computational aeroacoustics problems. Various new
approaches have been described, ranging from global approaches that solve a single
set of equations throughout the entire domain to zonal methods that explicitly impose pure RANS or pure LES in individual zones according to some initial domain
decomposition.
For the task of simulating the turbulence and near-field noise, all methods presented
here are considered good potential candidates, with accuracy that is expected to be
superior to unsteady RANS and a cost anticipated to be significantly less than that of
CUFX063/Wagner
164
0 521 87144 1
printer: sheridan
October 23, 2006
80
70
60
50
p (db)
P1: IBE
CUFX063-04
40
30
20
10
0
-10
10
10
F (Hz)
10
Figure 4.17. Acoustic pressure spectrum in the slats wake. Taken from Terracol et al. (2005) (Fig. 35,
a traditional, full-domain LES. The question of accuracy relative to full-domain LES

remains to be assessed.
In comparison with traditional LES, hybrid RANSLES methods are intended to
allow larger mesh spacings (at least in certain regions) and hence larger time steps; thus,
they can result in direct resolution of a smaller portion of the total frequency spectrum.
This is not a weakness of the hybrid models themselves but simply an expected result
of exploiting the coarser spatial and temporal resolutions with which hybrid methods
are able to operate. The issue of unresolvable noise sources has been discussed, and
two possible approaches to help extend the range of predicted frequencies have been
described. Both approaches involve synthetic reconstruction of an unsteady velocity
field that reproduces key properties of the underlying (statistically represented) turbulence. One approach considered was a nonlinear acoustics solver that uses a stochastic
subgrid model; the other involves augmenting the frequency spectra of traditional LES
or hybrid RANSLES in order to separately model the generation and transmission of
the noise from unresolvable scales.
Of the hybrid methods discussed, the DES approach currently has the widest experience base. The simplicity of the DES formulation is clearly appealing and is likely to
make this the first method of choice for hybrid RANSLES implementations in new
or existing CFD codes. The LNS method has an increasing user base in the commercial CFD++ code, where the slight increase in model complexity is not an issue for
the end user. Differences between these two methods are expected to be small when
used as explicitly intended by the DES. The current challenge is to extend the range of
applicability of these models to handle impinging flow, thin separation, and automatic
17:51
P1: IBE
CUFX063-04
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
4.5 SUMMARY OF HYBRID RANSLES METHODS
initiation of LES fluctuations as RANS data are transported from coarse-mesh into
fine-mesh regions. This data exchange between statistical (RANS) and resolved (LES)
region is likely to be a pivotal issue in the near future because users will increasingly
expect CFD codes to allow the use of embedded LES regions, akin to the use of zonal
hybrid RANSLES but determined automatically through mesh refinement in regions
where detailed knowledge of the local flow structures and sound sources is required.
With the major challenge already facing companies to train CFD engineers adequately
in this increasingly multidisciplinary field, it is unlikely that manual intervention at
complex internal boundaries would be practical or even possible. Therefore, the CFD
code itself will be expected to automatically handle the (abrupt or gradual) interface
between the RANS and LES regions. The possibility of treating these zonal interfaces automatically is something we expect to see emerging soon from commercial
CFD vendors. The transfer from LES to RANS data requires nothing more than an
averaging process. The transfer from RANS to LES data is slightly more complex,
but procedures are described in this chapter that can create a synthetic field of velocity fluctuations that achieves the desired second moments as well as the scales of
time and length given a set of RANS statistics. Such a process requires knowledge of
realizable covariances of the velocity fluctuations information that is directly available within the LNS model. Similar developments are also under way within the DES
framework.
The concept of using local structure size estimates for the dissipation rate in hybrid RANSLES models has been introduced with reference to the recent Menter
et al. (2003) proposal. This approach has not yet been widely tested, but its aim and
philosophy seem well grounded and may lead (for certain cases) to a simpler alternative than the use of synthetic reconstruction. Further developments are required, but
the ideas presented by Menter et al. (2003) are expected to find their way into other
hybrid RANSLES approaches as the problems associated with handling increasingly
complex geometries and general meshes become more prevalent.
The use of nonlinear perturbation or disturbance equations, such as the system
proposed by Morris, Long et al. (1997) and its viscous extension, has been discussed.
Although such methods are not yet very popular, they have the potential to be more
accurate in the computational treatment of disturbances that are small relative to their
solution mean. Several advantages of using these equations are highlighted in the work
of Batten et al. (2002c), Batten, Ribaldone et al. (2004), and Labourasse and Sagaut
(2002). The use of background (statistically steady) RANS mean fields is an area that
shows potential for reducing the resolution requirements for a partially resolved LES.
Existing hybrid RANSLES formulations that solve a single set of equations require
sufficient near-wall resolution in the direction normal to the wall, whereas a priori
RANS calculations with subsequent interpolation to a more uniform acoustics mesh
can mitigate this requirement because a set of grid-converged background statistics can
always be obtained from the a priori RANS solution. In the hybrid RANSLES approach
of Labourasse and Sagaut (2002) and the NLAS approach of Batten et al. (2002c), a
grid-converged set of background statistics (synthetically reconstructed in the latter
17:51
165
P1: IBE
CUFX063-04
CUFX063/Wagner
166
0 521 87144 1
printer: sheridan
October 23, 2006
case) are always available wherever the mesh resolution is unable to support the local
flow structures.
Finally, we note that the choice of both numerical method and SGS closure is
expected to continue to play a crucial role in LES and hybrid RANSLES. In most
(if not all) existing LES or hybrid RANSLES codes, the SGS model and numerical
flux treatments are oblivious to one another, typically resulting in a mixing that is too
strong. Some developments in viscous limiter functions were presented by Toro (1992)
for simplified model equations, but further developments would be welcome progress
for use with LES of the full NavierStokes equations.
17:51
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Numerical Methods
5.1 Spatial and temporal discretization schemes

Tim Broeckhoven, Jan Ramboer, Sergey Smirnov,
and Chris Lacor
5.1.1 Introduction to discretization schemes
In contrast to standard computational fluid dynamics (CFD) applications, for which

second-order accuracy in space is sufficient for engineering purposes, the requirements
on the schemes are much more stringent in the computation of aeroacoustical applications. There are several reasons for this.
First, the amplitudes of acoustic waves are several orders of magnitude smaller
than the average aerodynamic field amplitudes. In addition, the length scales of acoustic waves, typically the principal acoustic wavelengths, are some orders of magnitude
larger than the dimensions of the sound-generating perturbations (vortices and turbulent eddies). Thirdly, sound generated by turbulence is broadband noise with often
three orders of magnitude difference between the largest and the smallest acoustic
wavelengths. Finally, acoustic waves propagate at the speed of sound (which is not
necessarily comparable to the mean flow velocity) over large distances in all spatial
directions, whereas aerodynamic perturbations are only convected by the mean flow.
Moreover, one is usually interested in the noise level at the far field, implying that the
waves have to be traced accurately over long distances.
This requires numerical methods with higher accuracy than routinely applied in CFD
codes. In particular, the discretization of the convective operator (i.e., the first-order
derivative) requires special attention. In this respect, the actual order of the scheme,
which can be determined based on a Taylor expansion analysis, is not the primary
concern. Although this order gives an indication of the error reduction when one goes
to finer grids, it gives no information about the dispersive and dissipative behavior
of the scheme. This information, which is related to the resolution of the scheme as
opposed to the accuracy (Adams and Shariff 1996) can be obtained from a Fourier mode
analysis. The resolution is then the ability of the scheme to represent Fourier modes
167
0:9
P1: IBE
0521871441c05
168
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
of increasing wave numbers accurately. The dispersive errors of the scheme cause an
error in the phase of the Fourier waves as compared with the exact solution from the
differential equation, whereas the dissipative errors cause an error in the amplitude of
the Fourier waves (e.g., Hirsch 1988).
Similar arguments apply to the temporal discretization. The acoustic waves have
to be tracked accurately in time. Again, more important than the temporal order of
accuracy of the time integration method are its dispersive and dissipative behavior. For
the solution to be accurate, these errors have to be minimized as much as possible.
This chapter is organized as follows. First, the notion of dispersive and dissipative
errors is briefly reviewed, both for spatial and temporal discretization schemes. Next an
overview of spatial discretization strategies for use in acoustical simulations is given.
The final section is devoted to temporal discretization schemes.
5.1.2 Dispersion and dissipation errors
Consider the scalar convection equation

u
u
+a
=0
t
x
(5.1)
u = F(t)e I kx .
(5.2)
and a Fourier wave
Substitution of Equation (5.2) in Equation (5.1) gives the following equation for F:
dF
= I akt.
F
(5.3)
F = eI akt ,
(5.4)
Solution of this equation gives
and the Fourier wave of type (5.2) satisfying Equation (5.1) is given exactly by
u ex = e I k(xat) .
(5.5)
5.1.2.1 Dispersion and dissipation errors of the spatial scheme
We again consider Equation (5.1), but semidiscretized in the sense that a spatial discretization scheme is used for the spatial derivative. Equation (5.1), written in point i,
becomes
u i
= R(u i , u i1 , u i+1 , u i2 , u i+2 , . . .)
(5.6)
t
i
. The letter R is used to indicate that this is
with R the discretization stencil of a u
x
the so-called residual. As dictated by the scheme used, the value of u in point i, u i , as
well as in surrounding nodes (u i1 , u i+1 , u i2 , u i+2 , ...) may be used.
Substitution of the Fourier wave (5.2) leads to the following equation for F:
dF
= Rt,
F
(5.7)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
169
5.1 SPATIAL AND TEMPORAL DISCRETIZATION SCHEMES
where R is the so-called Fourier footprint of the residual, which can be decomposed in
a real part, Rr , and an imaginary part, Ri :
R = Rr + I Ri .
(5.8)
F = eRr t e I Ri t .
(5.9)
The solution of Equation (5.7) is
Combining Equations (5.2) and (5.9), we find that the solution of the semidiscretized
equation is therefore
u = eRr t e I k(x+
Ri
k
t)
(5.10)
Comparing Equation (5.5) with Equation (5.10), one notes that two errors arise because
of the discretization:
r An error on the amplitude of the wave: The exact amplitude is 1, whereas, with
spatial discretization, an amplitude eRr t is obtained. This is the dissipation error.
r An error on the propagation speed of the wave: The exact speed is a (for all wave
numbers k), whereas, with spatial discretization, the wave speed depends on the
wave number and is given by Rk i . This is the dispersion error.
Some authors define a numerical wave number k as
R I ak .
(5.11)
The solution of the semidiscretized Equation (5.9) becomes

k
u = e I k(x k at) .
(5.12)
Hence, if the difference between numerical and actual wave speed describes the error the
real part of k corresponds to dispersive errors, whereas the imaginary part corresponds
to a dissipative error.
As an example, consider the second-order central scheme
u i+1 u i1
u i
+a
= 0.
t
2x
(5.13)
The Fourier footprint is given by

R = I a
sin(kx)
.
x
(5.14)
The Fourier footprint is purely imaginary, and hence the central scheme does not introduce a dissipation error. The dispersion error disp , defined as the ratio of numerical
wave speed and exact wave speed, is given by
disp =
sin(kx)
.
kx
(5.15)
Note that disp is also the ratio of numerical and actual wave speed with the numerical
wave speed as defined in Equation (5.11).
0:9
P1: IBE
0521871441c05
170
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
5.1.2.2 Dispersion and dissipation errors of the temporal scheme
We consider the semidiscretized Equation (5.6) and introduce a temporal discretization.

To fix thoughts, consider the Euler explicit scheme
u in+1 u in
n
n
n
n
= R(u in , u i1
, u i+1
, u i2
, u i+2
, . . .),
t
(5.16)
where the subscript indicates the time level of the solution. Substitution of the Fourier
wave (5.2) gives
F n+1 F n
= RF n
t
(5.17)
F n+1
= 1 + tR.
Fn
(5.18)
and
n+1

To study the stability of the temporal discretization it suffices to check that FF n 1.
For a study of dispersion and dissipation properties, the correct evolution of F in time
has to be considered. This is given by Equation (5.9), and hence
n+1
F
= eRt .
(5.19)
F n ex
Comparison of Equations (5.18) and (5.19) shows that, for a given spatial discretization
scheme, the temporal discretization in general introduces both an error in the module
of F (the dissipation error) and in the phase of F (the dispersion error).
This analysis is easily extendable to other discretization schemes. For example, for
RungeKutta schemes, Equation (5.18) becomes
F n+1
= G(tR),
Fn
(5.20)
where G is a polynomial of degree q with q the number of RungeKutta stages.

Expressed in terms of the numerical wave speed, Equation (5.11), one has to compare
G(I tak ) with eI tak .

5.1.3 Spatial discretization schemes
The equations to be solved in computational aeroacoustics (CAA) applications are the

NavierStokes equations, either in their full form or in a simplified form, by omitting viscous terms (Euler equations) or by linearizing (e.g., linearized Euler equations
(LEE)). This system can compactly be written as
F
G
1 V
W
U
+
+
=
+
t
x
y
Re x
y
(5.21)
with U the vector of unknowns; F, G the inviscid fluxes that depend on U; and V, W the
viscous fluxes depending both on U and its derivatives U . The inviscid fluxes describe
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
171
the convective transport, and, as such, their discretization requires special attention and
accurate schemes.
In the following sections we will at first consider a simplified scalar convection
equation of the form
u
u
+
=0
t
x
(5.22)
and focus on the discretization of the ux term. Where needed, the extension to the
system above will be made and further explained.
5.1.3.1 Classical central- and upwind-type schemes
Central schemes
The most straigthforward approach is to use central-type schemes that is, the derivative
of u is discretized as

M
u
1
a j u i+ j ,
(5.23)
x i
x j=M
where M defines the stencil width. To determine the coefficients a j and the order of the
scheme, one can decompose the u i+ j in a Taylor expansion around i and equate leftand right-hand sides of Equation (5.23) to the maximum possible order. One finds that
a0 = 0 and a j = a j and that the order is 2M. The simplest scheme is the well-known
second-order central scheme (M = 1):

1
u
(u i+1 u i1 ) .
(5.24)
x i
2x
Although this scheme is routinely used in standard CFD applications, where it is usually
supplemented by some artificial dissipation terms (Jameson, Schmidt, and Turkel 1981),
its accuracy seems insufficient for CAA applications and large-eddy simulation (LES)
especially for the accurate simulation of sound wave propagation. In the context of a
hybrid approach, however, this scheme has been used by several authors in the near
field for the determination of the sound sources (e.g., Biedron et al. 2001; Manoha
et al. 2000; Boersma 2002).
Also note that this scheme is often used in LES. One of the reasons is that, in its
finite volume form, it can be shown (Ducros et al. 2000) to be in the so-called skewsymmetric form. As recalled by Kravchenko and Moin (1997), such schemes have a
built-in dealiasing property that makes them attractive for use in LES.
However, in most CAA applications higher-order variants of the central scheme are
oruk and Long 1996a, 1996b, 1997; Ahuja, Ozy
oruk, and Long 2000), where
used (Ozy
a fourth-order accurate scheme is used for applications of ducted turbofan noise. This
latter scheme corresponds to M = 2 and is given by

1
1
1
u
u i+2 + 2u i+1 2u i1 + u i2 .
(5.25)
x i
3x
4
4
This fourth-order accurate scheme is also popular in LES simulations. Ducros
et al. (2000) show that it can also be put in skew-symmetric form.
0:9
P1: IBE
0521871441c05
172
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Chyczewski, Morris, and Long (2000b) use a sixth-order finite difference scheme
in combination with the nonlinear disturbance equation (NLD) approach of Morris,
Long et al. (1997). Applications, however, are restricted to a fully developed turbulent
boundary layer and the channel flow without noise predictions. In the NLD approach,
the unknowns are split in a mean value, which is assumed independent of time, and
a perturbation. The mean solution is obtained as a Reynolds-averaged NavierStokes
solution with usually a simple algebraic turbulence model. Substitution of the full
solution in the NavierStokes system leads then to the NLD equations, which contain
both linear and nonlinear fluctuation terms, and a mean-flow source term independent of
the fluctuations. To simplify the system, we assume that the disturbances are essentially
inviscid in nature so that viscous perturbation terms can be neglected.
Note that the preceding schemes only guarantee high-order accuracy on Cartesian
and uniform grids. This poses a problem because almost all applications require nonuniform grids (i.e., grids with clustering and, in most cases, also non-Cartesian grids). In
the finite difference context, this problem is circumvented by applying the discretization
above in computational space, where the grid is uniform and Cartesian. However, one
has to make sure then that the errors of the Jacobian transformation do not reduce the
accuracy of the scheme (cf. Gamet et al. 1999; see also Section 5.1.3.3).
Upwind schemes
In the central schemes of the previous section, the discretization stencil is symmetric
with respect to the point where the derivative is evaluated. In upwind-type schemes,
the discretization stencil is upwind (or downwind) biased that is, upwind (downwind) points have a higher contribution weight. For fully upwind (downwind) schemes
only upwind (downwind) points contribute to the discretization stencil. In the present
context upwind is used as a generic term indicating both upwind and downwind
schemes.
Several authors use similar upwind-type schemes as developed for CFD (and not
CAA) applications. In Smith et al. (2000) a Roe-type scheme with monotone upstreamcentered schemes for conservation laws (MUSCL) extrapolation and different limiters
is proposed. The Roe-type scheme, or flux difference splitting scheme, was developed in
the 1980s by Roe (1981) and is a Godunov-type scheme in the sense that the numerical
flux formulation is based on the solution of an approximate Riemann problem. Hence,
the numerical flux needs information from the cells left and right of the interface. In
the first-order accurate version, only the cells sharing the interface are used. With the
MUSCL approach, the accuracy of the scheme is increased by also involving cells
farther away from the interface. This allows an increase in the accuracy of the scheme
to a second-order or even third-order one with the so-called -schemes (Van Leer
1979). Limiters are used to restore the monotonicity of the scheme by incorporating
some nonlinearity in accordance with the Godunov theorem that states that monotone linear schemes for convection problems can at most be first-order accurate. In
Smith et al. (2000), the Roe scheme is applied for cavity acoustics at supersonic
freestream in the framework of LES simulations.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
173
Ewert et al. (2002) and El-Askary, Ewert, and Schroder (2002) use the secondorder advection upstream splitting method (AUSM) (Liou and Steffen 1993) for the
LES simulation of the near field of the flow around a sharp leading edge and around
a cylinder. The radiation to the far field is calculated via a linearized Euler approach
using a dispersion-relation-preserving (DRP) scheme (cf. Section 5.1.3.2) with artifical
selective damping (ASD) (cf. Section 5.1.3.2).
Ekaterinaris (1999b) extends the stencil used in the determination of the dissipative
term of the Roe scheme to obtain, respectively, a third- and fifth-order accurate scheme.
He shows results for the computation of the sound generation from a corotating vortex
pair using the HardinPope two-step approach (Hardin and Pope 1994). In this two-step
approach the Euler equations are linearized around the mean flow of an incompressible
computation. Because pressure and density variations obtained by this means are no
longer isentropically related, a further decomposition of the density is used.
Nance, Viswanathan, and Sankar (1997) improve the extrapolation, typically used
in MUSCL-type schemes by incorporating ideas of the DRP (cf. Section 5.1.3.2)
L
, the following formula is
schemes. To obtain the left value on cell face i + 1/2, u i+1/2
used:
L
= a2 u i2 + a1 u i1 + a0 u i + a1 u i+1 + a2 u i+2 .
u i+1/2
(5.26)
Four of the five coefficients a j are determined by Taylor expansions of the righthand side. This fixes the order of the extrapolation. The remaining degree of freedom
is used to improve the resolution. This is achieved by taking the Fourier transform of
Equation (5.26) and minimizing the dispersion error. As in the DRP schemes, the error
is expressed as an integral of the resolution error (cf. Section 5.1.3.2), and optimization
is only achieved for wave numbers with wavelength larger than 4x (i.e., the integral
ranges in between /2 and /2).
Casper and Meadows (1995) propose the use of essentially nonoscillatory (ENO)
schemes for CAA problems involving shockwaves. Weighted essentially nonoscillatory
(WENO) schemes are used by Seror and Sagaut (2002) in aeroacoustic simulations of
a supersonic rectangular jet with a hybrid LESKirchhoff or Lighthill method. Several
authors have also proposed upwind-type schemes in the context of finite elements (e.g.,
Baggag et al. 1999; Lemaire, Marquez, and Jansen 1999; Kato et al. 2000).
5.1.3.2 DRP schemes
Symmetric DRP schemes
The DRP schemes have been developed by Tam and Webb (1993). The basic idea is to
use explicit, high-bandwidth, finite difference stencils for the first-order derivatives:
f
x
M
1
a j f i+ j .
x j=N
(5.27)
In contrast to the central schemes of Section 5.1.3.1, only part of the available coefficients a j of the stencil are used to fix the order of the scheme. The remaining coefficients
0:9
P1: IBE
0521871441c05
174
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Table 5.1. Coefficients for the DRP scheme of Equation (5.27)

Source
a0
a1 = a1
a2 = a2
a3 = a3
Tam and Webb

Tam and Shen
0
0
0.79926643
0.77088238
0.18941314
0.16670590
0.02651995
0.20843142
Source: Modified from Tam and Shen (1993).
are tuned so as to minimize the resolution error. This is the difference between the physical wave number k and the numerical wave number k which is given by
M
I
a j e I jkx
k =
x j=N
(5.28)
with I 2 = 1. In Tam and Webb (1993) only symmetric schemes are considered (i.e.,
N = M). This implies that k is real and the errors are only dispersive. The error to be
optimized is expressed as

2
E=
|kx kx|
d(kx).
(5.29)
The value of determining the range of the integral was originally chosen as /2
by Tam and Webb (1993). This means that optimization is performed only for waves
with a wavelength larger than 4x. The reason for not optimizing over the complete
wave-number range (, + ) is that, irrespective of the scheme used, the numerical
wave number goes to zero for kx = . It therefore makes no sense to try to optimize
up to the highest wave numbers (see also Kim and Lee (1996) for optimization in the
framework of compact schemes). According to Lockard, Brentner, and Atkins (1994),
the choice = /2 already places a very stringent requirement on the minimization
that can only be met by schemes with some significant overshoots (the numerical wave
number is larger than the actual one) in the lower wave-number range of the wavenumber diagram. The better resolution for high wave numbers is then at the cost of
a reduced resolution for lower wave numbers. This was also recognized by Tam and
Shen (1993) in later work.
Table 5.1 shows the coefficients a j of Equation (5.27) as obtained by Tam and Webb
(1993) and by Tam and Shen (1993). In both cases, a seven-point stencil was used (i.e.,
N = M = 3 in Equation (5.27)). Fourth-order accuracy in space was imposed, leaving
one degree of freedom for the optimization.
In Lockard et al. (1994) the integral therefore extends only up to waves with wavelengths of 7x or larger (i.e., = 2/7). The DRP scheme is often used in combination with the LEE approach (e.g., Agarwal and Morris 2000) for acoustic scattering by a fuselage and Fleig et al. (2002) for the aeroacoustics of a wind turbine
blade.
The central-type DRP schemes do not have any inherent dissipation mechanism
to eliminate spurious short-wavelength numerical waves. As a result, these schemes
allow a pileup of energy in the high-wave-number range that might render the scheme
unstable. In order to dissipate this energy, some dissipative terms can be added.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
175
One of the first schemes to incorporate such dissipation terms was the Jameson
SchmidtTurkel scheme (Jameson et al. 1981), which nowadays is widely used in CFD
codes. For CAA applications this type of dissipation is not well adapted as shown by
Kim and Lee (2000), who combine it with compact schemes (cf. Section 5.1.3.3).
Morris, Long, and Scheidegger (1999), Morris, Scheidegger, and Long (2000), and
Agarwal and Morris (2000) use a standard sixth-order artificial dissipation. Originally
they also include a second-order dissipation in some applications (Morris et al. 1999),
but in a later paper this term is omitted (Morris et al. 2000). In Morris et al. (1999,
2000), applications are noise generation of both circular and rectangular supersonic
jets. The NLD approach (cf. Section 5.1.3.1) is based on an LES formulation (Morris,
Long et al. 1997). Both a direct as well as a hybrid approach are used. In the latter case,
a Kirchhoff or a Ffowcs WilliamsHawkings method is used for the far-field radiation.
Tam, Webb, and Dong (1993) propose ASD. A damping term of the following form
is added to the right-hand side of the equation:
D=
3
a
d j f i+ j .
(x)2 j=3
(5.30)
For a Fourier wave F(k) e I kx with k the wave number, the resulting damping is
a
D(kx) =
(x)2
3

dje
I jkx
F(k).
(5.31)
j=3
The coefficients d j can now be optimized in order to tune this damping so that it is
essentially active in the high-wave-number range (Tam et al. 1993). Note that the original
ASD proposal is not in conservative form. Extension to conservation form is presented
by Kim and Lee (2000); see also Section 5.1.3.3. Bogey, Bailly, and Juve (1999, 2000a)
use the DRP scheme with ASD in their LES solver and show results for mixing layer
noise (Bogey et al. 1999) and the sound radiated by a circular subsonic jet (Bogey
et al. 2000a).
Asymmetric DRP schemes
Dissipation terms can also be directly incorporated in the scheme by deviating from
a central-type formulation. Lockard et al. (1994) therefore relax the requirement of
symmetric schemes. This implies that k now contains an imaginery part, leading to
dissipative errors. The integral to be optimized contains, then, a weighted contribution
from both types of errors:

E=
2
2
+ ( 1) Im(kx)
Re(kx kx)
d(kx)
(5.32)
with a weighting function that allows the emphasis to be put on either the real or
0, in order to avoid
the imaginary part. Because one has to be careful that I m(k)
0:9
P1: IBE
0521871441c05
176
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
growing modes, a modified integral is also proposed to have more flexibility to force
the imaginary part to be negative:

+ ( 1) Im(kx)
sin (kx/2) d(kx)
Re(kx kx)
E=
(5.33)
with parameter always negative and a positive integer. Lockard et al. (1994) derive, then, optimized third-order accurate schemes with different stencils to be used
in their ENO code. A similar optimization in the context of compact-upwind schemes
(cf. Section 5.1.3.4) is formulated by Adams and Shariff (1996).
5.1.3.3 Compact central schemes
In recent years, so-called compact (or Pade-type) schemes (Lele 1992; Kim and
Lee 1996) have gained increasing popularity in applications such as DNS (Gamet
et al. 1999), LES (Boersma and Lele 1999; Meinke et al. 2002), and CAA (Gamet
et al. 1999; Koutsavdis, Blaisdell, and Lyrintzis 2000; Povitsky and Morris 2000) as
an alternative to spectral methods. The main advantage of these schemes is that, while
providing a better representation of the shorter length scales of the solution as compared with classical finite difference and finite volume schemes, they allow the use
more complex mesh geometries than the spectral methods, which are limited to applications in simple domains and with simple boundary conditions. Here, it should
be mentioned that combinations of spectral methods with h-type methods, using local
polynomial expansions in the cells of the mesh, do not suffer from the preceding limitations. This is the case for the so-called spectral/hp element methods (Karniadakis and
Sherwin 1999) and the discontinuous Galerkin method, which is further described in
Section 5.1.3.6.
Originally, compact schemes were defined in the finite difference context, in one
dimension and on uniform grids. One of the first applications of compact finite differencing to solve differential problems can be found in the work of Collatz (1966,
1972).
However, it is only recently, in view of the applications mentioned above, that there
has been a renewed interest in these types of schemes.
1D formulation
Finite difference formulations
The main idea of a Pade-type finite difference scheme is to construct the approximation
of the differential equation to be solved with not only node values being unknowns but
also the derivatives. For the first-order derivative of any scalar quantity u, the Pade-type
approximation can be written as follows (Lele 1992):

u i+2 + u i+1 + u i + u i1 + u i2
=C
u i+3 u i3
u i+2 u i2
u i+1 u i1
+ B
+A
,
6h
4h
2h
(5.34)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
177
where u denotes the first-order derivative, h is the grid spacing (in the assumption
of uniform mesh), and , , A, B, and C are unknown coefficients that determine
the accuracy of the approximation. These coefficients can be obtained by developing
all the variables in a Taylor series about point i and requiring the coefficients of the
resulting expansion to vanish up to some definite term. The first nonzero coefficient
will determine the formal truncation error. Note that the scheme of Equation (5.34) is
purely central and therefore will only introduce dispersive errors.
It is also possible to construct compact approximations for the higher-order derivatives (e.g., Lele 1992), where also second-order derivatives are considered. However, in
the present context, the focus is on the convective terms, and schemes for higher-order
derivatives will therefore not be discussed.
If and do not both equal zero, Equation (5.34) defines the derivative in an implicit
way. Note that if = = 0, the standard central schemes, with different stencil widths,
as determined by the coefficients A, B, and C, are retrieved.
Using the notation of Adams and Shariff (1996), one can write Equation (5.34)
symbolically as

(5.35)
Lu i = (Ru)i ,
where L and R are compact notations for the stencils on the left- and right-hand sides,
respectively. Note that in theory these stencils can even be wider than the ones appearing
in Equation (5.34).
In case = 0, the resulting system is tridiagonal. Lele (1992) shows that the maximum order of accuracy is then 8. If, in addition, the choice C = 0 is made, a oneparameter family of fourth-order schemes results. The most compact scheme within
this family corresponds to the choice B = 0 and is given by
1
3 u i+1 u i1
1
u i+1 + u i + u i1
.
=
4
4
2
2h
(5.36)
For = 0, the system becomes pentadiagonal and the order of accuracy can be increased
to 10.
Lele investigated the resolution of the compact schemes (of implicit type; i.e.,
and not both equal to zero) and found that they have much better resolution
properties as compared with the explicit schemes ( ==0) with a comparable order of
accuracy. Figure 5.1 illustrates this; it shows the numerical wave number k x (cf. Section 5.1.2.1) as a function of the exact wave number for different schemes. The diagonal
line corresponds to no error. It is observed that the bandwidth of wave numbers for
which dispersion errors are small grows with the order of the scheme. The fourth-order
explicit scheme, which is basically scheme (5.25), gives an important improvement
compared with the second-order scheme of Equation (5.24). However, the implicit
fourth-order scheme is clearly superior to its explicit counterpart.
This is also confirmed by Wilson, Demuren, and Carpenter (1998). They introduce
a maximal numerical wave number kc that defines the region of acceptable accuracy;
that is,
|kx
kx| < 0.01
for
0 < k < kc .
(5.37)
0:9
P1: IBE
0521871441c05
178
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
3.14
exact
second-order
fourth-order explicit
fourth-order implicit
sixth-order implicit
eighth-order implicit
2.09
Modified
Wave number
1.05
0.00
0.00
1.05
2.09
Wave number
3.14
Figure 5.1. Resolution of different explicit and compact (implicit) schemes.
The number of grid points per wavelength (PPW) for this maximal wavelength kc is
given by
P PW =
2
.
kc x
(5.38)
Wilson et al. (1998) found that PPW decreases from 28.6 for the second-order
accurate scheme to 4.1 for the sixth-order compact scheme.
Lele (1992) also proposesd to further optimize the schemes by imposing that, for
certain wave numbers, the numerical wave number equals the exact wave number.
For example, if one uses scheme (5.34) and imposes only fourth-order accuracy, three
degrees of freedom result, which can be used to equal numerical and exact wave number
in three points. The choice of these wave numbers is somewhat empirical, although
Lele argues that the result is quite insensitive to the choice made. For this fourth-order
scheme, Lele chooses
kx = 2.2, 2.3, 2.4
(5.39)
as points for optimization. The resulting scheme clearly outperforms the tenth-order
accurate scheme in terms of dispersive behavior.
Note that the DRP scheme of Tam and Webb (1993) is based on a more refined
idea of optimization. Their approach was adapted to compact schemes by Kim and
Lee (1996). In the latter paper a weighting function is introduced to (1) make the
integral analytically integrable and (2) weight the integrated error more in the highwave-number range. Note that the first reason for using a weighting function does not
apply to explicit-type schemes ( = = 0) such as the DRP schemes. Concerning
the choice of (cf. Equation (5.29)), Kim and Lee (1996) mention they obtain good
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
179
schemes for values up to = 0.9 . They adapt the range, as determined by the scheme,
to obtain maximum resolution properties. Yet, some of their schemes also produce an
overshoot (i.e., the numerical wave number exceeds the physical one), which, according
to Lockard et al. (1994), is not a desirable feature (cf. also Section 5.1.3.2).
A result of this optimization procedure is that better resolution properties can sometimes be obtained for lower-order accurate schemes because more degrees of freedom
are available. For instance, for tridiagonal schemes of type (5.34) ( = 0), a fourthorder accurate scheme offers two degrees of freedom (the variables and C can be
tuned to minimize the integral), but a second-order accurate scheme has three degrees
of freedom (in addition B can be tuned). This results in a second-order accurate scheme
with better resolution properties than the fourth-order accurate one (Kim and Lee 1996).
An alternative way for optimization is proposed by Tang and Baeder (1997). Their
optimization is based on the description of the compact schemes via Hermitian interpolation. Their approach is illustrated below for a three-point scheme involving u i1 ,
u i , u i+1 .
First, consider a Lagrangian interpolation, which will lead to standard central
schemes. The Lagrangian interpolation function q(x) for a three-point scheme is a
polynomial of power 2:
q(x) = a0 + a1 (x xi ) + a2 (x xi )2 .
(5.40)
Equating q(x) in the nodal points i 1, i, and i + 1 to, u i1 , u i , and u i+1 , respectively,
allows us to determine the coefficients ai . The discretization formula for u i is then
obtained because u i q (xi ) = a1 . One obtains the standard central scheme
u i
u i+1 u i+1
.
2x
(5.41)
The compact scheme (5.36) is obtained via Hermitian interpolation, and q(x) is now a
polynomial of order 4:
q(x) = a0 + a1 (x xi ) + a2 (x xi )2 + a3 (x xi )3 + a4 (x xi )4 .
(5.42)
Two additional conditions are needed to determine ai as follows:

,
q (xi1 ) = u i1
(5.43)

.
q (xi+1 ) = u i+1
(5.44)
The discretization formula for u i is again obtained because u i q (xi ) = a1 . It can

easily be checked that Equation (5.36) is retrieved.
Tang and Baeder (1997) propose the use of trigonometric functions instead of power
polynomials for the interpolation, leading to so-called Fourier difference schemes as
opposed to Taylor difference schemes. For the same example as above, the interpolation
function becomes

(x xi )
(x xi )
+ a2 sin
,
(5.45)
q(x) = a0 + a1 cos
2x
2x
0:9
P1: IBE
0521871441c05
180
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Figure 5.2. Comparison of the resolution of Taylor and Fourier difference schemes.
where the half-wavelength of the cosine and sine functions corresponds to the stencil
width of 2x. Equating q(x) in the nodal points i 1, i, and i + 1 to u i1 , u i , and,
respectively, u i+1 allows us to determine the coefficients ai . The discretization formula
a2 . One obtains
for u i is then obtained because u i q (xi ) = 2x
u i
u i+1 u i+1
.
2
2x
(5.46)
Comparing the numerical wave number of this scheme with the one of Equation (5.41),
one sees that the numerical wave number is multiplied with 2 . As a result the dispersion
errors are more uniformly distributed in the frequency domain, but there is also a large
accuracy contamination in the low-frequency range. This is illustrated in Figure 5.2,
where Fourier difference schemes of second and sixth order are compared with their
Taylor counterparts.
As a result, this low-order trigonometric interpolation does not lead to a reasonable
scheme. Although the sixth-order Fourier difference scheme seems superior to the sixthorder Taylor scheme, it has to be noted that the Fourier difference scheme always has a
zero-order truncation error, which is of course not acceptable from a mesh refinement
point of view.
Tang and Baeder (1997) therefore propose to combine the Fourier difference approach with the Taylor difference approach by replacing only the higher-order power
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
181
polynomials of the Hermitian interpolation with higher-order trigonometric functions.

For the compact scheme above, the interpolation function (5.42) is replaced by

x xi
x xi
2
+ a4 sin l
q(x) = a0 + a1 (x xi ) + a2 (x xi ) + a3 cos l
L
L
(5.47)
with L the stencil width (i.e., 2x) and l a free parameter. The expression for u i is
obtained in exactly the same way as for the compact scheme (5.36). By varying l, the dispersive behavior of the scheme can be improved. Note that l = 2 corresponds to the
maximum wave number visible on the present stencil (wavelength of 2x). Experience
shows that l must be well below this maximum wave-number value in order to produce
useful optimized schemes. Tang and Baeder (1997) also further refine this strategy by
including trigonometric functions of two different orders, l1 and l2 , in the interpolation.
Results shown in Tang and Baeder (1997) are limited, however, to 1D convection of a
Gaussian wave.
Finally, note that the compactness of the schemes can further be improved by use of
some prefactorization as proposed by Hixon (2000). This allows us to replace the tridiagonal (respectively, pentadiagonal) systems by products of two bidiagonal (respectively,
tridiagonal) systems. Hixon shows results for some CAA benchmark problems.
If the grid is nonuniform, the formulations above can be applied in computational
space (where the grid is uniform). However, this may lead to some loss of accuracy
related to the Jacobian transformation; see Section 5.1.3.3.
Finite volume formulations
The finite volume (FV) method is inherently conservative, which is not the case for
the finite difference (FD) approach in which special attention must be paid to the
conservation properties of the schemes. However, because the FV method is formulated
in physical space the construction of compact schemes is less straightforward (especially
on nonuniform grids and in multidimensions), and only few papers on the subject can
be found in the literature.
To our knowledge, one of the first papers dealing with a formulation of compact
schemes within the FV context is by Gaitonde and Shang (1997). Their scheme is
based on an implicit reconstruction step relating cell face values to cell averaged values.
Gaitonde and Shang (1997) discuss a 1D formulation and consider the extension into
multidimensions only for linear equations and uniform grids. Applications are shown
for the 1D propagation of a transverse electromagnetic wave and a 2D spherical dipole
field.
A similar approach, but restricted to 1D equations, was proposed more recently by
Kobayashi (1999). This approach was later extended to the multidimensional Navier
Stokes system by Pereira, Kobayashi, and Pereira (2000). Although results are only
shown on Cartesian grids, the proposed formulation can also be used for non-Cartesian
grids by working in computational rather than physical space. In that case, errors
in the calculation of the derivatives of the transformation might reduce the global
0:9
P1: IBE
0521871441c05
182
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
accuracy of the scheme (cf. Section 5.1.3.3). Lacor, Smirnov, and Baelmans (2000,
2004), and Smirnov, Lacor, and Baelmans (2001a) independently developed a similar
approach that is also applicable to NavierStokes on arbitrary, structured, non-Cartesian
grids.
In 1D all the different approaches are equivalent. Consider the following generic
1D advection equation:
f (u)
u
+
= 0.
t
x
(5.48)
Integration over mesh cell i, with cell faces i + 1/2 and i + 1/2, leads to
u i
x + f i+1/2 f i1/2 = 0
t
with u i the cell-averaged value of cell i:
u i =
1
x
(5.49)
xi+1/2
u(x)d x.
(5.50)
xi1/2
Note that relation (5.49) is still exact, and the order of the scheme will be determined
by the order with which the fluxes f i+1/2 and f i1/2 are evaluated. Because the relation
between f and u is known, this reduces to the problem of evaluating u i+1/2 and u i1/2
(i.e., the value of the unknowns at the cell faces). An approach based on an implicit
interpolation formula is proposed:
u i3/2 + u i1/2 + u i+1/2 + u i+3/2 + u i+5/2
=a
u i+1 + u i
u i+2 + u i1
u i+3 + u i2
+b
+c
.
2
2
2
(5.51)
This implicit step is equivalent to the implicit procedure used in the FD schemes
(cf. Equation (5.34)) and allows us to achieve high orders of accuracy with relatively
small stencils. Taylor expansions can be used to determine the interpolation coefficients
to fix the order of the interpolation and hence the order of the scheme. If the mesh is
nonuniform, the grid metrics will appear in the expressions for the coefficients, which
can then be stored.
Extension to multidimensions and NavierStokes
Because compact schemes define the derivatives in an implicit way, the extension to
arbitrary grids and the NavierStokes system is not always straightforward. In the
FD context however, the formulations of Lele (1992) can be used directly in Navier
Stokes codes. The idea is to apply the 1D formulation of Equation (5.34) in each
of the three grid directions in computational (as opposed to physical) space. This
approach is used by many authors (e.g., Visbal and Gaitonde 1998, 2002; Gaitonde
and Visbal 1999; Ekaterinaris 1999b). Visbal and Gaitonde (1998, 2002) use basically
a tridiagonal fourth-order accurate and a pentadiagonal sixth-order accurate scheme.
Near the boundaries they use a one-sided formulation for the point on the boundary
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
183
(usually fourth-order accurate) and the tridiagonal fourth-order scheme for the point
next to the boundary.
Ekaterinaris (1999b) uses tridiagonal fourth- and sixth-order accurate schemes. He
mentions that the cost of using these schemes is comparable to that of standard, explicit
schemes provided the inverses of the implicit matrices are stored.
Ekaterinaris (1999b) also uses the fourth-order compact scheme in the implicit
operator of the BeamWarming scheme (Beam and Warming 1976) to replace the standard central scheme for approximating the Jacobian derivatives. This reduces the need
for many subiterations to eliminate linearization and factorization errors. Also, the diagonalized version of Pulliam and Chaussee (1981) is extended using this approach.
Ekaterinaris (1999a) also proposes a compact formulation when an upwind implicit
operator is used. The standard first-order upwind discretization of the derivatives of
the split Jacobians is replaced by compact upwind formulations. When using a fully
upwind formulation such as the Cockburn and Shu scheme (Equations (5.73) through
(5.75)), this leads to pentadiagonal systems. An alternative is to use upwind-biased
central stencils as in the schemes proposed by Adams and Shariff (1996) (cf. Section 5.1.3.4). By keeping only three points (instead of five in the Adams and Shariff
schemes) in the left- and right-hand sides, such as in the schemes proposed by Zhong
(1998), the tridiagonal structure can be kept.
The authors above all make use of the Jacobian transformation to define the compact
schemes in computational space. This is a straightforward procedure that is briefly
explained below.
Consider the nondimensionalized NavierStokes system (below the 2D system is
presented for compactness; extension to 3D is straightforward):
F
G
1 V
W
U
+
+
=
+
t
x
y
Re x
y
(5.52)
with F, G the inviscid flux vectors and V , W the viscous flux vectors. If the partial
derivatives are denoted as a subscript, the preceding equation becomes
Ut + Fx + G y =
Vx + W y .
Re
(5.53)
We consider a transformation from physical time-space (x, y, t) to computational timespace (, , ) with

= t,
= (x, y, t),
= (x, y, t).
(5.54)
Application of the chain rule to Equation (5.53) gives

U + t U + t U + x F + x F + y G + y G
=
x V + x V + y W + y W .
Re
(5.55)
0:9
P1: IBE
0521871441c05
184
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
This equation is in weak conservation form because the metrics terms appear as
coefficients in the differential equation. It can be put in so-called strong conservation
form by multiplying first with J 1 x y x y , the inverse Jacobian of the transformation, and using the chain rule on all terms such as
J

x F
x
F
J
x
J
(5.56)
If terms are regrouped, this leads to

t U + x F + y G
t U + x F + y G
+
+
J
J

x V + y W
x V + y W
1
=
+
+ T.
Re
J
J
U
J
(5.57)
The extra term T is given by

t
x
1
t
x
+
+
+
+F
T =U
J
J
J
J
J

y
x
1
y
x
+
+
V
+G
J
J
Re
J
J

y
1
y
W
+
.
Re
J
J
(5.58)
Equation (5.58) contains the so-called invariants of the transformation (between brackets). It can be shown analytically that these invariants are identically zero. However,
when evaluated numerically the invariants might be nonzero because the numerical
derivatives do not necessarily commute. This implies that, for a uniform flow, Equation (5.57) is not satisfied: all terms will disappear numerically, except T . Hence, special
care must be taken to make the invariants also numerically equal to zero. Once this is
satisfied, the resulting equation, now in strong conservation form, reads
U
J
G
1
F
+
=
+
Re
V
W
+

(5.59)
with
u
1 u u + x p
,
F =
J v u + y p
E + p) t p
u(
(5.60)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
185
v
1 u v + x p
,
G =
J v v + y p
v(
E + p) t p
(5.61)
x V + y W ,
V =
J
(5.62)
x V + y W ,
W =
J
v the contravariant velocities, defined as
and u,
(5.63)
u = t + x u + y v,
(5.64)
v = t + x u + y v.
(5.65)
Because in most cases the transformation from physical to computational space is

not known analytically, the metrics (t , x , y , t , x , y ) are calculated numerically on
a cell-by-cell basis. A convenient way of establishing the coordinate transformation
is by making use of isoparametric elements. The transformation of the element in
physical space to an element in the curvilinear coordinates of computational space is
then described via the following relation:
x =
4

x i Ni (, )
(5.66)
i=1
with x i as the coordinates of the nodes of the physical elements and Ni the shape
functions, function of the curvilinear coordinates , . The preceding relation allows us
to calculate x , x , y , y immediately. These are related to x , y , x , y via the inverse
relations:
x = J y ; y = J x ; x = J y ; y = J x .
(5.67)
As mentioned by Gamet et al. (1999), the Jacobian transformation can lead to large
errors in case of nonsmoothly varying mesh spacings. In the general case, the theoretical
accuracy of the compact schemes (as derived in 1D on uniform grids) can be maintained
only for FD schemes that take into account the stretching of the grid. The critical point
of the Jacobian transformation is whether the Jacobians, which contain first and second
derivatives of the transformation, can be defined and calculated without appreciable
loss in overall accuracy. The loss of accuracy on curvilinear grids is also discussed by
Gaitonde and Visbal (1999).
Gamet et al. (1999) therefore take the nonuniformity of the mesh directly into account by adapting the coefficients of the compact stencils. The coefficients in Equation
(5.34) now depend on the index i. The conditions for a given order of accuracy (resulting from Taylor expansions) will now involve also the mesh spacing and lead to
appropriate expressions for the coefficients. The important gain in accuracy using such
0:9
P1: IBE
0521871441c05
186
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
an approach, as compared with a Jacobian transformation, is clearly quantified in Li,

Ma, and Fu (2000). However, the method of Gamet et al. (1999) only deals with nonuniform but Cartesian grids. If the grid is non-Cartesian, a Jacobian tranformation must
still be used.
Freund (2001) and Freund, Lele, and Moin (1998) use a sixth-order compact scheme
in DNS simulations of the noise of a circular subsonic and supersonic jet, respectively.
LES results of this subsonic jet using the same compact scheme but supplemented with
a sixth-order filter to remove energy from the poorly resolved wave numbers, where the
subgrid-scale model is ineffectual, are shown by Bodony and Lele (2002a).
Mitchell et al. (1999) use the fourth-order scheme in their DNS simulations of
an axisymmetric jet. The same scheme is used by Povitsky (2001) for aeroacoustics of
spherical pulse propagation in 3D stagnation flows and the flow over a circular cylinder.
Colonius, Lele, and Moin (1997) use both fourth- and sixth-order accurate schemes
in the DNS simulation of sound generation in a mixing layer.
Mankbadi, Hixon, and Povinelli (2000) apply the prefactored compact scheme of
Hixon to a very large eddy simulation (VLES) simulation of the noise generation of a
supersonic, heated round jet.
Below, we restrict ourselves to 2D for the sake of clarity. Consider the 2D Euler equations
u t + f x + g y = 0.
Integrating over a cell, one obtains
i j u i, j +
t
(5.68)

(gd x + f dy) = 0
(5.69)
with i j the volume (area) of the cell i j (with i and j the mesh directions) and u i, j
the cell-averaged value; that is,

1
u(x, y)d xd y.
(5.70)
u i, j
i, j
i, j
The accuracy of the scheme is determined by the accuracy with which the line integrals

(second term of Equation (5.69)) are determined. Consider the flux f dy, which is
over the cell faces transverse to the i-direction; if the flux is linear (e.g., f u), as is the

case for a linear 2D convection problem, the integral udy can be determined with a
formula similar to Equation (5.51). In the left-hand side the line integrals on consecutive
cell faces in the i-direction appear, and in the right-hand side the cell-averaged values
in the same direction, keeping the j-index constant. However, on arbitrary grids this
1D-like approach, as proposed in Pereira et al. (2000) and Pereira, Kobayashi, and
Pereira (2001), cannot guarantee a high order of accuracy anymore. Pereira et al. (2000,
2001) therefore use a transformation to a Cartesian grid as in FD methods. An alternative, avoiding transformations, is to use a multidimensional stencil. In Lacor et al. (2000)
this is achieved by extending the stencil in the right-hand side by including cell-averaged
values with a different j-index.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
187
For Euler and NavierStokes applications, the cell-face integrals are calculated

for all the primitive variables. The nonlinear parts of the fluxes (e.g., uvdy) can be
obtained by combining the integrals of the primitive variables
1
y
1
uvdy =
y
1
udy
y

vdy + O(h 2 ),
(5.71)
where y is the length of the cell face. Note that the approximation above is only secondorder accurate, automatically reducing the order of the scheme to 2. Pereira et al. (2000)
propose a correction to the relation above valid for Cartesian grids, making it fourthorder accurate. In Lacor et al. (2000) and Smirnov, Lacor, and Baelmans (2001) this
relation is extended for use on arbitrary grids.
Applications with these FV compact schemes are shown for convection-based model
problems (Pereira et al. 2000; Lacor et al. 2000); a laminar flat-plate calculation and
LES of channel flow (Lacor et al. 2004; Smirnov, Lacor, and Baelmans 2001).
Stabilization using filters or artificial dissipation
Because the compact schemes formulated in the previous sections are purely central,
they do not guarantee monotonicity. Also, because Fourier waves are not damped, the
nonlinear character of the NavierStokes system may lead to a pileup of energy in the
high-wave-number range, which can lead to numerical instability unless there is some
dissipation or filter to remove this energy.
Different filters, both explicit and implicit, have been investigated by Lele (1992).
The order of the filter should of course be at least of the same order of accuracy as the
compact scheme.
The simplest filters are those of explicit type. Their use in the present context is,
however, limited (e.g., Pruett et al. 1995).
Compact, implicit filters are more accurate than the equivalent explicit filters and
therefore more popular. They also allow us to achieve a high-order accuracy on a
narrower stencil than the explicit filters. The two main disadvantages are that (1) a
matrix must be inverted and (2) the boundary stencil has a large effect on the interior
accuracy of the filter (Hixon 1999).
Visbal and Gaitonde (1998, 1999, 2002), Gaitonde and Visbal (1999), Koutsavdis,
Blaisdell, and Lyrintzis (1999), and Ekaterinaris (1999a, 1999b) use such implicit, lowpass filters (based on the suggestions of Lele (1992)) to remove unwanted oscillations
and waves from the solutions and to suppress numerical instabilities arising from all
sources, including those associated with mesh nonuniformities, boundary conditions,
or low local grid resolution.
one
Denoting a component of the solution vector as and its filtered value as ,
can describe the filter as (cf. Visbal and Gaitonde 2002)
f i1 + i + f i+1 =
N

an
n=0
(i+n + in ) .
(5.72)
0:9
P1: IBE
0521871441c05
188
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Table 5.2. Coefficients for filter formula (5.72)

Order
a0
a1
a2
10
1
+ f
2
3
5
+ 4f
8
5
11
+ 8f
16
70
93
+ 128f
128
126
193
+ 256 f
256
1
+ f
2
1
+ f
2
17
15
+ 16 f
32
18
7
+ 16 f
16
302
105
+ 256 f
256
1
+ 4f
8
3
3
+ 8f
16
14
7
+ 32 f
32
30
15
+ 64 f
64
Order
a3
a4
a5
1
16f
32
1
8f
16
90
45
512f
512
1
128
5
256
4
6
8
2
4
6
8
10
+
+
f
64
10 f
256
1
256
2 f
256
Source: After Visbal and Gaitonde (2002).
The coefficients a j are derived in terms of f with Taylor- and Fourier-series analyses,
and f is an adjustable parameter in the range 0.5 < f 0.5 with higher values
corresponding to a less dissipative filter. Table 5.2 gives the values of a j for filters of
orders varying from 2 to 10. Figure 5.3 shows the effect of f on a second- and an
eighth-order filter.
The bottleneck with filtering is the procedure near the boundaries where the same
stencil as for inner points cannot be used. One possibility is to use a filter of lower
Figure 5.3. Effect of f on a second- and an eighth-order filter.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
189
order (with reduced stencil); see Visbal and Gaitonde (1998). In a later paper the same
authors propose an improved approach based on one-sided, but still high-order, filters
(Gaitonde and Visbal 2000).
According to Visbal and Gaitonde (1998), filtering seems superior to artificial dissipation. In Koutsavdis et al. (1999), the application for CAA is investigated in the
framework of LEE on Cartesian, nonuniform grids. The conclusion is that filtering is
very useful because it removes unwanted q-waves; thus, grid refinement to suppress
these waves is not needed. This allows us to obtain solutions with an accuracy that
could only be realized on much finer grids if the filtering is not used.
An alternative way to stabilize the scheme is through the use of some artificial
dissipation either standard Jameson dissipation (Jameson et al. 1981) or, in view of
CAA applications, an ASD procedure, as proposed by Tam (1995) (cf. Section 5.1.3.2).
Shim, Kim, and Lee (1999) and Kim and Lee (2000) recently proposed a conservative
implementation of this approach and combined it with their optimized compact FD
schemes of Kim and Lee (1996). They also added a second-order dissipation term to be
able to deal with nonlinear discontinuities. Comparison on a bell-shaped linear pulse
propagation shows the same accuracy as with the standard ASD scheme but a significant
improvement compared with the result obtained with Jameson-type dissipation. In Shim
et al. (1999) this scheme is applied to the study of radiation of multiple pure tone noise
from an aircraft engine inlet. The use of ASD in combination with compact schemes
is also discussed in Broeckhoven et al. (2003).
5.1.3.4 Compact upwind schemes
In contrast to the central compact schemes, upwind compact schemes have an inherent dissipation. As for the dispersive errors, it is important that the dissipative errors
mainly act on the higher-frequency waves. This ensures that the resolution of the lowerfrequency waves will not be spoiled by dissipation. On the other hand, energy aliasing
back from high wave numbers will be damped efficiently.
1D formulation
The general notation used in Equation (5.35) means that the operators L and R are
not of the central type anymore. Different basic schemes can be attributed to different
authors.
A third-order upwind scheme is proposed by Cockburn and Shu (1994) in which the
implicit operator L is of the central type but the explicit operator R is upwind biased:
Lu
1
1
u i1
+ u i u i+1
5
5
(5.73)
and
(Ru)i
3
(3u i 4u i1 + u i2 ) .
10x
(5.74)
0:9
P1: IBE
0521871441c05
190
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
In the downwind version, the R operator becomes

3
(3u i + 4u i+1 u i+2 ) .
(5.75)
10x
Upwind schemes with both the left- and right-hand-side operators upwind biased
have been formulated by Tolstykh and Shirobokov (1995) (third-order) and Fu and Ma
(1997) (third- and fifth-order accurate). The Tolstykh scheme is defined as
(Ru)i
Lu
5
1
,
u + u i u i+1
8 i1
8
3
(u i u i1 ) ,
2x
and the schemes of Fu and Ma (1997) as
(Ru)i
(Ru)i
(Ru)i
(5.76)
(5.77)
1
u + u i ,
2 i1
(5.78)
1
(u i+1 + 4u i 5u i1 ) ,
4x
(5.79)
Lu
2
u + u i ,
3 i1
(5.80)
1
(u i+2 + 12u i+1 + u i 44u i1 3u i2 ) .
36x
(5.81)
Lu
In Li et al. (2000), the Fu and Ma schemes are extended for nonuniform grids using
an approach similar to that of Gamet et al. (1999) (cf. Section 5.1.3.3).
Adams and Shariff (1996) use an optimization procedure (see also Section 5.1.3.2)
to derive two upwind-type schemes based on a five-point stencil both in left- and
right-hand-side form. The first scheme is described as being compact upwind with high
dissipation (CUHD) and is designed to be about as dissipative at nonresolved wave
numbers as a noncompact upwind scheme while giving a much better representation
of the dispersion relation. It is given by (numbers are only approximate; see appendix
of Adams and Shariff 1996 for the exact numbers)

+ 0.734u i1
+ u i 0.169u i+1
0.061u i+2
, (5.82)
Lu i 0.036u i2
1
(0.176u i2 1.137u i1 + 1.063u i + 0.452u i+1 0.201u i+2 ) . (5.83)
x
The second scheme, denoted as compact upwind with low dissipation (CULD), is
designed to be less dissipative (about one order of magnitude lower dissipation than
CUHD). It is defined as (numbers are only approximate; see appendix of Adams and
Shariff 1996 for the exact numbers)

+ 0.456u i1
+ u i + 0.477u i+1
+ 0.039u i+2
,
(5.84)
Lu i 0.028u i2
(Ru)i
(Ru)i
1
(0.119u i2 0.748u i1 + 0.0005u i + 0.718u i+1 + 0.148u i+2 ). (5.85)
x
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
191
The schemes above can easily be used for 1D linear convection. However, for 1D
Euler and NavierStokes applications, one needs a mechanism to choose between the
upwind- or downwind-biased versions. Such mechanisms were already developed for
standard upwind-type schemes and can be combined with the compact upwind schemes.
Cockburn and Shu (1994) combine the 1D Euler equation

u
+ f i = 0
(5.86)
t i
in different points to bring the implicit operator in front of the time derivative. Using a
compact scheme of the form (5.35) leads to
u i + (R f )i = 0
t
(5.87)
u i (Lu)i .
(5.88)
with
This can be put in conservation form by noticing that R f can be rewritten as

(R f )i = (N f )i+1/2 (N f )i1/2 .
(5.89)
For example, for the fourth-order central compact scheme one has
(N f )i+1/2
3
( f i + f i+1 ) .
4x
(5.90)
The numerical flux f N f can be split into positive and negative parts for easier
upwinding as follows:
f = f+ + f = N f + + N f .
(5.91)
Different splittings are possible; Cockburn and Shu (1994) propose a LaxFriedrichs
splitting.
Note that the calculation of the (splitted) fluxes requires the u i values, whereas only
u i is known from solving the equations. This requires a reconstruction step. The +/
numerical fluxes on the cell faces i + 1/2 are compared with the +/ fluxes in neighboring points i for purposes of limiting. According to the theory of upwind schemes,
the f + ( f ) fluxes should be obtained, respectively, by upwind- or downwind-biased
formulas. This calls for a compact formula (5.35) where the operator R is, respectively,
upwind or downwind. Cockburn and Shu (1994) also use the central scheme of Equation (5.36). This works well for smooth flows. On flows with discontinuities, however,
the results are inaccurate, although the scheme remains stable as a result of the limiters
action.
A related approach is proposed in Ravichandran (1997). In contrast to Cockburn
and Shu (1994), the flux is split before applying the compact scheme:

u
+
+ f i + f i = 0.
(5.92)
t i
0:9
P1: IBE
0521871441c05
192
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Ravichandran (1997) uses kinetic flux vector splitting (Mandal and Deshpande 1994)
for its robustness as compared with Van Leer or StegerWarming flux vector splitting.
Upwind- (downwind-)biased compact formulas of type (5.35) are used for f i+ ( f i ).
The corresponding operators are denoted with a + or sign, that is,

+
(5.93)
L + f = R+ f + i ,
i

L f = R f i .
i
(5.94)
Formally one can therefore write

+
+
f i = L 1
,
+ R+ f
i
(5.95)

.
f i = L 1
R f
i
(5.96)
This can be substituted in Equation (5.92) but is first put in conservation form similar
to that of Equation (5.89); for example,

(5.97)
R+ f + i = N+ f + i+1/2 N+ f + i1/2
and similar for R f . This leads to

u
+
1
+ L 1
+ N+ f + L N f

i+1/2
t i

L + N+ f + + L 1
= 0.
N f
i+1/2
(5.98)
The numerical flux f is thus given by

+
1
f = L 1
+ N+ f + L N f .
(5.99)
The +/ numerical fluxes are compared with the +/ fluxes in neighboring points
for purposes of limiting. The compact schemes used for the split fluxes are those
of Cockburn and Shu (1994; cf. Equations (5.73)(5.75)), Tolstykh and Shirobokov
(1995), and Fu and Ma (1997).
Deng and Maekawa (1997) use a cell-centered version of Equation (5.34), which,
according to Lele (1992), has a better dispersive behavior than the unstaggered formula (5.34). Equation (5.34) is, for example, appropriate for FD schemes, where the
solution is stored in the mesh modes; the cell-centered version uses the derivatives in
the cell centers in the left-hand side whereas in the right-hand side the fluxes on the
cell interfaces appear. The expression for the flux derivative of the 1D Euler equation
becomes

+ f i + f i+1
=
f i1
f i+1/2 f i1/2 +
f i+3/2 f i3/2 .
h
h
(5.100)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
193
Deng and Maekawa (1997) replace the flux in the right-hand side with expressions for
the numerical flux f from upwind-type schemes such as Roes flux difference splitting
scheme or flux vctor splitting. For example, for the Roe scheme one has
L 1
R

1

R
L
f u i+1/2 + f u i+1/2
|a| u i+1/2 u i+1/2
.
(5.101)
fi+1/2 =
2
2
The u L and u R values are obtained via implicit tridiagonal, fifth-order interpolation
upwind and downwind biased, respectively. In case discontinuities are present, interpolation across the discontinuity is avoided (for reasons of accuracy) by switching locally
to bidiagonal interpolation, either upwind or downwind, as determined by the location
of the discontinuity. The choice of the appropriate interpolation is based on ideas similar
to those used in ENO schemes by measuring the smoothness of the different stencils
using first- and second-order differences. In a subsequent paper, Deng and Mao (1997)
combine the three interpolations in a weighted formulation. The eight coefficients are
chosen so that, in smooth regions, the interpolation is fifth-order accurate and third
order near discontinuities. In the latter case, the weights are such that no interpolation
across the discontinuity is guaranteed. The determination of the weights also uses the
first- and second-order differences as in Deng and Maekawa (1997). In a recent paper
by Deng and Mao (2001), extension to multidimensional NavierStokes is proposed.
Also, dissipative terms are added in the right-hand side of Equation (5.100), leading
to so-called dissipative, weighted compact schemes. The dissipative terms are based
on split fluxes and are therefore only used in combination with flux vector splitting
schemes. In Deng and Mao (2001), only results with the explicit version of the schemes
are shown; that is, 0 in Equation (5.100).
A systematic analysis of upwind-type compact schemes was done by Ramboer, Smirnov,

and Lacor (2002) and Ramboer and Lacor (2002). The approach is similar to that of
the 1D central compact scheme, but Equation (5.51) is replaced with an upwind-biased
formulation. Three types of schemes can be considered: those with both L and R operator upwind biased (LURU); schemes with L operator upwind biased and R operator
central; and schemes with L operator central and R operator upwind biased. Through
symbolic manipulation, the properties (order of accuracy, dispersive and dissipative
error) of schemes of each category have been investigated, thereby restricting them to
at most a five-point implicit operator and a six-point explicit operator. The conclusion is
that the LURU schemes are preferable. Two promising schemes are a bidiagonal thirdorder accurate scheme and a tridiagonal fifth-order scheme. The following interpolation
(cf. Equation (5.51)) is used:
5
1
1
u i1/2 + u i+1/2 = u i + u i+1
2
4
4
for the third-order scheme and
1
1
19
10
1
u i1/2 + u i+1/2 + u i+3/2 =
u i1 + u i + u i+1
2
6
18
18
18
(5.102)
(5.103)
0:9
P1: IBE
0521871441c05
194
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
for the fifth-order scheme. The expressions above correspond to the upwind-biased
formulation. The downwind-biased formulation follows immediately from analogy.
On nonuniform grids, the coefficients will depend on the grid metrics as is the case for
the central-type schemes (cf. Section 5.1.3.3).
When applied to the Euler/NavierStokes system, Roes flux difference splitting
scheme is used to distinguish between positive and negative eigenvalues. The cell face
flux is given by
f i+1/2 =

1
L
f + f R |A|(u R u L ),
2
2
(5.104)
where the R (L) values are obtained with, respectively, downwind (upwind) interpolation. The parameter (which equals 1 in the original Roe scheme) allows to reduce the
dissipation; see also Lin, Yu, and Shih (1997) and Bui (1999).
Extension to multidimensions and NavierStokes
Cockburn and Shu (1994) also extend their method to multidimensions for Cartesian
grids. The operators A and B now become L x , Rx for x-derivatives and L y , R y for the
y-derivatives. Starting from the 2D Euler system
u
f
g
+
+
= 0,
t
x
y
(5.105)
we find that the resulting equation is
u i + L y Rx f i + L x R y g i = 0
t
(5.106)

u i L y L x u i .
(5.107)
with
The numerical fluxes are now defined as

fi+1/2, j = L y N x f + + N x f i+1/2, j ,
(5.108)

fi, j+1/2 = L x N y f + + N y f i, j+1/2 .
(5.109)
The approach of Ravichandran (1997) is easily extended to multidimensions by

applying the same strategy in each direction of the computational space.
The upwind schemes defined in Section 5.1.3.4 can be extended in much the same
way as in the central compact schemes of Section 5.1.3.3. Instead of determining the

flux udy, two fluxes are determined namely, u L dy and u R dy. The formulas
for these fluxes are again implicit but respectively upwind and downwind biased. In
Ramboer et al. (2002) and Ramboer, Smirnov, and Lacor (2003) results are shown for,
respectively, a vortex and a pressure pulse superposed on a uniform 2D Euler flow. By
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
195
tuning the dissipation with the parameter of Equation (5.104), one can obtain very
accurate results especially with the fifth-order scheme.
5.1.3.5 Predictor-corrector schemes
Other high-order accurate schemes have been formulated based on a predictor-corrector

approach. Gottlieb and Turkel (1976) modified the McCormack scheme (McCormack
1969) to increase the accuracy to fourth order. Bayliss (1985) proposed a sixth-order
accurate scheme. In Snyder and Scott (1999) different predictor-corrector schemes are
tested for CAA applications, using category 2 benchmark problems. The Gottlieb and
Turkel scheme outperforms the DRP scheme of Tam and Webb (1993) and is competitive
with the Bayliss predictor-corrector scheme.
In Hixon (2000), compact McCormack-type schemes are proposed. A prefactorization method is used to reduce a nondissipative central difference stencil to two
lower-order biased stencils that have easily solved reduced matrices. However, these
high-accuracy McCormack schemes have some drawbacks. If inherent damping in the
biased stencil is too large, it can damp out waves the original central difference scheme
could have propagated accurately; conversely, if the inherent damping is too small,
nonlinear waves may generate unresolved high-frequency waves that will destroy the
solution accuracy (Hixon 1998).
5.1.3.6 Discontinuous Galerkin methods
The discontinuous Galerkin (DG) method is a compact finite element projection method
that provides a practical framework for the development of high-order methods using
unstructured grids. The higher-order accuracy is obtained by representing the solution
in each cell as a high-degree polynomial whose time evolution is governed by a local
Galerkin projection. In contrast to standard finite element methods, the DG methods
enforce the conservation law only locally. This allows them to have a mass matrix that
can easily be made to be the identity and that therefore does not necessitate lumping
or matrix inversion while being highly accurate and nonlinearly stable. Because of the
local character, the solution may be discontinuous across different elements. As a result,
the basis for the polynomials is essentially unrestricted. However, for computational
efficiency, an orthogonal basis is chosen relative to the inner product. Note, however,
that, even if the polynomials are not orthogonal, one still only needs to invert a small
mass matrix and there is never a global mass matrix as in a typical finite element method.
The DG method can also be interpreted as an extension of a FV method, incorporating
notions such as approximate Riemann solvers, numerical fluxes, and slope limiters
into a finite element framework. However, instead of only one degree of freedom per
cell as in a finite volume method (viz., the cell average of the solution), there are, in
2D, (n + 1)(n + 2)/2 degrees of freedom for a polynomial of order n. These degrees of
freedom are chosen as the coefficients of the polynomial when expanded in a local basis.
Consider, for instance, the Euler system
U
+ F = 0.
t
(5.110)
0:9
P1: IBE
0521871441c05
196
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
In each cell, U is expanded in terms of the polynomials bk , k = 1, . . . , N :

U=
N

Uk bk .
(5.111)
k=1
The discretized equation is obtained by multiplying Equation (5.110) by b j and integrating over the cell:

U
Fb j d V = 0.
(5.112)
bjdV +
V t
V
Using the expansion (5.111) and applying Greens theorem, one obtains

N

Uk
bk b j d V
Fb j d V + Fb j d S = 0.
t V
V
k=1
(5.113)
This equation shows that the derivative of Uk is retrieved by inverting a matrix M with

components Mk j V bk b j d V . The surface integral requires an expression for the flux
on the cell faces. Because the solution is discontinuous across cell faces (as in FV
methods), numerical fluxes based on Riemann solvers and incorporating limiters can
be used. One can show that, for polynomials of degree p, the order of accuracy is at
least p + 12 (Johnson and Pitkarata 1986). Equation (5.113) is usually discretized by
evaluating the integrals using quadrature formulas of the required order, which is 2 p for
the volume integral and 2 p + 1 for the edge integral (Cockburn, Hou, and Shu 1990).
This limits the usefulness of the method because the number of terms in the quadrature
summations significantly exceeds the number of unknowns, making the method computationally expensive. However, Atkins and Shu (1996, 1998) described a quadrature-free
formulation.
The application of the DG method to hyperbolic systems in combination with
RungeKutta methods, the so-called RungeKutta discontinuous Galerkin (RKDG)
method, was thoroughly investigated by Cockburn (1999) and Cockburn, Karniadakis,
and Shu (2000) for a detailed survey.
Hu, Hussaini, and Rasetarinera (1999) studied the wave propagation properties of the
DG method. Based on an analysis of scalar convection, they found that in a formulation
with upwind fluxes the dissipative errors are dominant, whereas the dispersive error is
negligible for wave numbers up to a value equal to the order of the method. In a centered
flux formulation, there is no dissipative error, but the range of wave numbers for which
the dispersion error is negligible is smaller than with the upwind flux. Because DG
methods are especially useful on unstructured grids, the influence of the mesh was also
investigated. From an analysis of the 2D wave equation, it is found that an unstructuredlike triangular mesh has better dispersion and dissipation properties than a structured
quadrilateral grid or than triangular grids derived from such a grid. Also, the properties
are less prone to anisotropy (i.e., vary less with the orientation of the Fourier waves).
Atkins and Lockard (1999) used the RKDG method to solve the linearized Euler equations (LEE) to study acoustic scattering from a two-dimensional slat and a
three-dimensional blended-wing-body combination. In addition, they showed, for a
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Figure 5.4. Perturbation pressure of an acoustic wave initiated by a Gaussian pulse: comparison of
solution with smooth and randomly perturbed mesh.
three-dimensional wave propagation in a cube, that the RKDG method is insensitive

to the mesh smoothness: on a smooth tetrahedral mesh and the same mesh with 20%
random perturbation of each grid point, the two solutions are almost indistinguishable
as shown in Figure 5.4.
Stanescu, Hussaini, and Farassat (2002) use the spectral element implementation of
DG of Kopriva, Woodruff, and Hussaini (2000) for the computation of sound radiation
from aircraft engine sources to the far field. The nonlinear Euler equations are solved,
and only the radiation of inlet noise into a quiescent fluid is modeled. Both nacelle alone
and fuselage-nacelle and fuselage-wing-nacelle configurations are considered. In a later
paper by Xu et al. (2003), the same methodology is applied to an actual two-engine
jet aircraft and compared with a spectral element solution in the frequency domain.
The results show that trends of the noise field are well predicted by both methods. In
Baggag et al. (1999) the DG method is applied to a benchmark problem of the 1997
CAA workshop.
5.1.4 Temporal discretization schemes

5.1.4.1 RungeKutta schemes
By far the most popular schemes in CAA are the RungeKutta schemes. They are
explicit and can be formulated up to an arbitrary order of accuracy. Many authors use
the classical RungeKutta scheme, whether or not in a low-storage form to reduce the
usage of memory. Recently optimized RungeKutta schemes have been developed with
better dissipative and dispersive behavior.
197
0:9
P1: IBE
0521871441c05
198
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Classical RungeKutta schemes
The NavierStokes equations, Euler equations, or linearized Euler equations, which are
usually solved in CAA, consist of a set of coupled partial differential equations. By
spatially discretizing these equations, one obtains a set of coupled ordinary differential
equations. Therefore, the time integration is concerned with solving the initial value
problem
dU
dt
= F (t, U (t)) ; U (t = 0) = U0
(5.114)
with U the vector of unknowns. The most general q-stage explicit RungeKutta scheme
(Kennedy, Carpenter, and Lewis 1999) to integrate from time level t n to time level t n+1
can be written as
U(i) = U(n) + t
U(n+1) = U(n) + t
i1

j=1
q

ai j F( j) ,
(5.115)
b j F( j) ,
(5.116)
j=1
where

F(i) = F t (i) , U(i) ,

t
(i)
=t
ci =
(n)
s

(5.117)
+ ci t,
(5.118)
ai j
(5.119)
j=1
and the stage number i runs from 1 to q. The formulation above requires considerable
memory because, at stage i, all F( j) on the previous stages need to be known. The
RungeKutta schemes can be implemented in a low-storage form, given by
U(i) = U(n) + ti F(i1) ,
q

U(n+1) = U(n) + t
i F(i) .
(5.120)
(5.121)
i=1
A classical fourth-order RungeKutta scheme corresponds to the choice

1 = 12 , 2 = 12 , 3 = 1,
(5.122)
1 = 16 , 2 = 13 , 3 = 13 , 6 = 16 .
(5.123)
The classical RungeKutta scheme given by Equations (5.122) and (5.123) is often
used in CAA in combination with DNS or LES for the calculation of the near-field
sources. Examples are LES simulations of jet flows by Zhao, Frankel, and Mongeau
(2000), Morris et al. (1999), and Bogey et al. (2000a) and DNS simulations of jet flows
by Freund (2001) and Mitchell et al. (1999) as well as DNS of sound generation in a
mixing layer by Colonius et al. (1997).
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
199
The same scheme has also been used for solving the LEE by, for example Bailly,
Bogey, and Juve (2000), Agarwal and Morris (2000), Povitsky (2001), and Ekaterinaris
(1999c). In Yu, Hsieh, and Tsai (1995) the dissipation and dispersion errors for this
RungeKutta scheme are studied in combination with several higher-order spatial discretizations and results of vortex calculations based on the nonlinear Euler equations
are shown.
Kim and Lee (1996) compared classical RungeKutta schemes of different orders
in combination with a fourth-order spatial discretization. Going from a second-order
RungeKutta scheme to a fourth-order RungeKutta scheme significantly improved
their results. However, going to even higher-order RungeKutta schemes did not bring
much improvement. This might explain why the classical forth-order RungeKutta
scheme is so popular.
As a last remark, it is important to mention that another four-stage RungeKutta
scheme commonly used in CFD applications
U(i) = U(n) + ti F(i1) ,
U(n+1) = U(n) + tF(q) ,
(5.124)
(5.125)
is only fourth-order accurate in time for linear advection problems with constant speed.
In case of nonlinearity, this scheme has a maximum second-order accuracy in time.
5.1.4.2 Optimized schemes
Time integration schemes can be optimized for different properties such as linear and
nonlinear stability, accuracy efficiency, error control reliability, and dissipation and dispersion accuracy (Kennedy et al. 1999). In CAA applications, where one is especially
concerned about the dissipative and dispersive behavior of discretized equations, optimization of dissipation and dispersion properties of the time integration scheme is of
paramount importance.
However, until now only a few authors have considered optimizing time integration
schemes.
A very commonly used optimized time integration scheme is the one by Hu,
Hussaini, and Manthey (1996). Several RungeKutta schemes are optimized for their
dispersion and dissipation behavior and are baptized as low-dissipation and lowdispersion RungeKutta schemes (LDDRK). The optimization is achieved by minimizing the difference between the numerical amplification factor, G(I tak ), and
the exact amplification factor, eI tak , as explained in Section 5.1.2.2.

The following integral is therefore minimized:

|G(I ) eI |2 d,
(5.126)
0
where = tak and specifies the range of in the optimization. In Hu et al. (1996)
the coefficients for a four-, five-, and six-stage optimized RungeKutta scheme
(LDDRK4, LDDRK5, LDDRK6) are given.
0:9
P1: IBE
0521871441c05
200
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Table 5.3. Optimized coefficients of the amplification factor for the LDDRK
schemes
Stages
c3
c4
c4
c4
4
5
6
0.162997
0.166558
1/3!
0.0407574
0.0395041
1/4!
0.00781071
0.00781005
0.00132141
Source: Hu et al. (1996).
The coefficients of the amplification factor for these schemes are given in Table 5.3.
All of these are kept second-order accurate in time (c1 = 1, c2 = 12 ) while minimizing
the dispersion and dissipation errors. A second method in Hu et al. (1996) to optimize time integration schemes is the use of different coefficients for the RungeKutta
schemes in two alternating time steps. This leads to a further reduction of the dispersion and dissipation errors while maintaining a higher order of accuracy. To do this, the
product of the amplification factors of the two different schemes is compared with the
square of the exact amplification factor. The coefficients are determined such that
the following integral is minimized:

|G 1 (I )G 2 (I ) eI 2 |2 d
(5.127)
with G 1 , G 2 the polynomials corresponding to the two RungeKutta schemes.

Two fourth-order optimized schemes are given, a foursix alternating scheme
(LDDRK46), and a fivesix alternating scheme (LDDRK56). Note that Hu et al.
(1996) only consider central spatial schemes, ensuring that the numerical wave number
and hence are real.
In his extensive study on different numerical methods used in CAA, Goodrich
(1999) concludes that the six-stage optimized LDDRK scheme by Hu et al. (1996) in
combination with at least a sixth-order spatial differencing can provide between one
and two orders of magnitude decrease in error at a given grid density. However, when
the fourth-order central scheme or the DRP scheme of Tam and Webb (1993) is applied,
no significant improved is found by using an optimized RungeKutta scheme. Other
applications of the schemes formulated by Hu et al. (1996) can be found in Ewert
et al. (2001a), who use LDDRK56 in combination with the fourth-order spatial DRP
scheme, as well as Bogey et al. (1999), Bogey, Bailly, and Juve (2000b), and Morris,
Long et al. (1997), who both use the LDDRK4 scheme.
A nonlinear extension of the LDDRK56 is used by Mankbadi, Hixon, and Povinelli
(2000) for a VLES application. Further details of this nonlinear extension are described
in Stanescu and Habashi (1998).
Recently, Bogey and Bailly (2004) presented an optimization of RungeKutta
schemes based on ideas similar to those of Hu et al. (1996). The integrals to be optimized
are defined somewhat differently, leading to different schemes. Optimized versions of
two five-stage and six-stage RungeKutta schemes are presented. The stability of these
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
201
5.2 BOUNDARY CONDITIONS FOR LES
optimized schemes is better than that of the five- and six-stage algorithms proposed by
Hu et al. (1996).
Tam and Webb (1993) are among the few authors who optimized a four-level timeintegration scheme. The procedure is comparable to the one followed by Hu et al. (1996).
By taking the Laplace transform of the four-level time-integration scheme given by

3

dU (n j)
bj
,
(5.128)
U(n+1) U(n)
t
dt
j=0
it is easy to find the effective angular frequency of the time discretization:

i eit 1
=
.

t 3j=0 b j ei jt
(5.129)
The scheme is optimized by minimizing the integral error E 1

0.5

[ (t t)]2 + (1 ) [ (t t)]2 d (t) (5.130)
E1 =
0.5
with , denoting the real and the imaginary part of their argument, respectively. The
real part of describes the dispersive behavior and the imaginary part the dissipative
behavior. For = 0.36, the following coefficients are obtained:
b0 = 2.30255809,
b2 = 1.57434093,
b1 = 2.49100760,
b3 = 0.38589142.
(5.131)
5.1.4.3 Other time-integration schemes
Most of the predictor-corrector schemes used in CAA, like the modified McCormack
scheme by Gottlieb and Turkel (1976) discussed in Section 5.1.3.5, are second-order
accurate in time. This scheme is used in Gamet and Estivalezes (1998), Mankbadi
et al. (1994), Mankbadi, Hixon et al. (1995), and Mankbadi, Shih et al. (1995) for jet
noise applications. Another predictor-corrector scheme, proposed by Bayliss (1985),
is also second-order in time. Manoha et al. (2000) use a second-order accurate semiimplicit AdamsBashforth/CrankNicolson scheme for LES calculation of trailingedge noise.
5.2 Boundary conditions for LES

Michael Breuer
The filtered NavierStokes equations used for LES predictions are in general elliptic in space and parabolic in time. Hence, to solve these equations for a specific
flow configuration under consideration, boundary conditions at all borders of the integration domain and initial conditions for all flow variables in the entire field are
required.
0:9
P1: IBE
0521871441c05
202
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Nozzle
Tu = 0
Airfoil / Wing
x1
feasible
inflow boundary
ui , p, T = f (x1, y, z, t)
feasible
outflow
boundary
Figure 5.5. Example for a reasonable choice of the integration domain (inflow and outflow boundaries)
for the flow past a wing in a wind tunnel.
With a few exceptions, such as predictions of the transition process or fundamental

turbulence investigations, in most cases the initial conditions play a subsidiary role
because the statistically steady-state flow status is of major concern, which should be
reached independently from the initial conditions. An example of an exception is the
decay of a homogeneous isotropic flow often used for basic investigations in turbulence
research. For this purpose a cubical integration domain with periodic boundary conditions in all directions is chosen, requiring the initialization of the flow field. Based on
the scalar energy spectrum of von Karman and Pao (see Hinze 1975), a homogeneous
isotropic and divergence-free vector field is generated in spectral space and then transformed back to the physical space by fast Fourier transformation. This procedure leads
to reasonable initial conditions with prescribed values for the turbulent kinetic energy
k0 , the dissipation rate 0 , and the smallest and largest wave numbers of the flow field.
On these initial conditions the basis of basic investigations on the decaying turbulence
can be carried out using LES or DNS predictions (e.g., Domaradzki and Rogallo 1990;
Rogallo 1981). In most practical LES applications, however, initial conditions are not
of prior importance. Here appropriate initial conditions can be chosen to shorten the
simulation time until statistically steady state is achieved. For that purpose a variety
of tricks are available such as symmetry breaking, superposition of perturbations, and
successive grid refinement (Breuer 2002).
Compared with initial conditions, boundary conditions (b.c.) play a dominant role
for LES especially for spatially inhomogeneous flows and will therefore be discussed
in detail. On the basis of their mathematical character, b.c. are typically classified into
Neumann b.c., Dirichlet b.c., combinations of Neumann and Dirichlet b.c., and periodic b.c. Physically, we can distinguish between two kinds of situations. Either a
physical boundary such as a solid wall or an artificial boundary of the integration domain has to be modeled. The latter appears if the solution domain constitutes only a
part of the total flow field typically encountered for predictions of external flows. A
typical example is shown in Figure 5.5, which depicts an experimental investigation of
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
the flow around an inclined wing. To simulate the corresponding flow numerically,
a reasonable integration domain has to be chosen not encompassing the entire wind
tunnel. This leads to artificial inflow and outflow boundaries that require physically
meaningful approximations of the flow situation. These are often difficult to formulate.
The quality of these b.c., however, determines the extension of the computational domain required for a high-quality LES and hence strongly influences the efficiency of
the prediction. Unlike at physical boundaries, natural b.c. of the physical model can be
applied in most cases.
In the following sections, the most important kinds of b.c. for LES are discussed
(i.e., outflow b.c., inflow b.c., and b.c. for solid walls). For the sake of simplicity, first
the flow field without additional scalars is considered solely. At the end some hints for
the extension toward nonisothermal flows and far-field b.c. for compressible flows are
given.
5.2.1 Outflow boundary conditions
The integration of the (filtered) NavierStokes equations requires a finite solution domain. Hence, a chosen outflow boundary can always be interpreted as an artificial cut
through the flow field (see Figure 5.5). Survey, for example, the flow around a wing,
where the flow in the vicinity of the airfoil and the resulting forces and noise sources
are of interest whereas the far wake is of less importance. Consequently, the grid points
will be clustered in the vicinity of the wing. The outflow margin of the computational
domain will be chosen as close as possible to the body in order to restrict the total size
of the domain. The flow variables at this artificial boundary have to be approximated
in a physically meaningful manner in order not to influence the solution of the conservation equations within the internal field by upstream-traveling perturbations. In order
to guarantee this condition, on the one hand the outflow boundary has to be shifted
downstream as far as possible which conflicts with the condition above. Furthermore,
this downstream shift is strongly restricted by the number of grid points acceptable and
the grid stretching allowed. It is very important for the location of the boundary that no
reverse flow exist on the entire plane and that the streamlines be more or less parallel.
Consequently, a compromise for the location of the outflow margin has to be found
taking all these aspects seriously into account.
On the other hand, the formulation of the boundary conditions itself plays a major
role. For stationary laminar or Reynolds-averaged flow predictions very often simple
extrapolations based on zero-, first-, or second-order polynomials along streamlines or
grid lines are applied. Especially, the condition of a fully developed flow in the main
flow direction (e.g., x, y, or z) is most often used. It leads to the extremely simple
Neumann boundary condition u i /|outflow = 0, which does a good job for steadystate flows at high Reynolds numbers. The reason is the weak upstream propagation of
perturbations under these flow conditions.
LES predictions are, however, inherently unsteady and dominated by vortical structures. Under these conditions, the reflecting boundary conditions of zero or any higher
order mentioned above lead to unsatisfactory results. This is not surprising because they
203
0:9
P1: IBE
0521871441c05
204
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
are physically not assured. The problem of upstream-traveling perturbations triggered

by outflow boundary conditions is a long-lasting topic of general interest for CFD.
Thus, a variety of so-called nonreflecting boundary conditions (see Section 5.2.4) have
been proposed in the literature especially for compressible subsonic and supersonic
flows (e.g., Hirsch 1990 or Engquist and Majda 1977). Appropriate outflow boundary
conditions for LES have to ensure that vortices can approach and pass the outflow
boundary without significant disturbances or reflections into the inner domain. For this
purpose, a convective boundary condition (Orlanski 1976) of the type

u i
u i
= 0
(5.132)
+ Uconv
t
outflow
has proven to work extremely well (cf. Breuer 1998, 2000, 2002). Equation (5.132)
represents nothing other than a simplified and linearized one-dimensional transport
equation in the main flow direction namely the momentum equation for the filtered
velocity component u i , and Uconv denotes a constant mean convection velocity in the
-direction. Typically, Uconv is independent of the position at the outflow plane and has
to be adjusted with respect to the flow simulated. One criterion for the appropriate choice
of Uconv may be the mass flow rate at the outlet. For incompressible flows this mass flow
rate has to balance the inflow mass flux (i.e., satisfy the global mass conservation). The
gradient u i / |outflow substitutionally denotes the gradient in the x-, y-, or z-direction
as determined by the direction of the mean flow. Numerically, the gradient in Equation
(5.132) is approximated by one-sided differences. This convective boundary condition
has proven to minimize the problem of pressure waves traveling upstream from the
outlet and hence to avoid the propagation of errors from the outflow boundary into the
computational domain. Consequently, it is widely used for LES predictions of external
as well as internal flows.
In Figure 5.6, the application of this boundary condition is demonstrated for an
LES prediction of the flow past an inclined wing at Re = 20,000 and = 12 (Breuer
and Jovicic 2001; Breuer 2002). Only a part of the entire three-dimensional integration
domain is shown. However, the right boundary depicted in each subfigure is in fact a
part of the actual outflow boundary. Hence, the figure clearly reveals that the large-scale
vortices are passing the outflow boundary and are leaving the computational domain
without large perturbations reflected back to the internal flow domain.
5.2.2 Inflow boundary conditions
Like the outlet boundary the inflow boundary typically represents an artificial cut
through the flow field. If, for example, the sample flow configuration installed in a
wind tunnel is numerically simulated from above, it is impossible and impractical to
take the flow in the entire experimental setup into account. Instead a reasonably chosen
subdomain composed of the wing is typically considered. As a direct consequence LES
predictions require appropriate inflow data of the Dirichlet type that adequately represent the flow field at or near the nozzle of the wind tunnel. For the prediction of laminar
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Figure 5.6. Von Karm

an
vortex street past an inclined wing (NACA4415) at Re = 20,000 and = 12
visualized by streaklines; four different time instants of a shedding cycle in the vicinity of the outflow
boundary are shown.
flows the specification of a steady velocity (or pressure) field at the inlet is sufficient
in most cases. For turbulent flow predictions based on the Reynolds-averaged Navier
Stokes (RANS) equations using, for example, two-equation closure models such as
the standard k model, additional specifications for the level of the turbulent kinetic
energy k and its dissipation rate are required. Typically, k and are expressed in terms
of the turbulence intensity Tu and the length scales l of the energy-carrying vortices.
For LES partially resolving the spectrum of turbulent length scales in the flow,
appropriate b.c. that mimic the unsteady vortical flow at the inlet are essential. These
have to be physically meaningful instantaneous data for the entire inflow plane satisfying characteristic autocorrelations and cross-correlations of the velocity components
between each other. Because the inflow b.c. can have a strong influence not only in the
vicinity of the inlet but also for the entire flow development, the specification has to
be made with care. Examples can be found in Klein, Sadiki, and Janicka (2003), Lund
et al. (1998), and Stanley and Sarkar (2000) for DNS of spatially developing flows and
where in principle no differences exist between LES and DNS concerning the challenge
of appropriate b.c.
Three special cases leading to simple resorts should be discussed first as follows:
(i) Laminar inflow:
In the first case, transition to turbulence appears far away from the inflow plane and
consequently takes place within the computational domain as a part of the solution.
Hence, the inflow data can be represented by laminar steady-state flow. An example
is the subcritical flow past a cylinder where transition takes place in the freeshear layers. This case is reasonably simulated by LES (Breuer 1998, 2000, 2002)
using a constant velocity profile (without any perturbations at the inlet) mimicking
an ideal wind tunnel or free-flight experiment with zero turbulence intensity.
205
0:9
P1: IBE
0521871441c05
206
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
rescaling
arbitrary
position
Auxiliary
Simulation
inflow
or
outflow
Main
Simulation
inflow
outflow
Figure 5.7. Sketch of Lund et al.s (1998) procedure for generating appropriate inflow conditions for a
boundary layer flow.
(ii) Periodic b.c. with or without rescaling:

The second way out of the dilemma is the use of periodic b.c. that avoid inflow and
outflow b.c. completely. However, the applicability of this most elegant boundary condition is restricted to flow configurations that are indeed periodic owing
to their geometry, such as turbine blades, or statistically homogeneous in one or
more coordinate directions such as channel, pipe, or duct flows. In the second case
the length of the integration domain L has to be chosen such that the two-point
correlations of the fluctuating quantities drop to zero along the homogeneous direction within L/2. Spalart and Leonard (1987) and Spalart (1988) extended the
area of application of periodic b.c. to turbulent boundary layers by using a coordinate transformation. Because of the complexity of Spalart and Leonards approach
and its restriction to flows with small mean streamwise variation compared with
the transverse variation, Lund et al. (1998) proposed an enhanced version. Their
method is based on an auxiliary simulation applying a much simpler rescaling
procedure than Spalart and Leonards for the velocity field from a downstream
location in order to generate its own inflow conditions. For this simplified transformation the periodicity of the b.c. is no longer valid. However, this is not a
critical issue because, in contrast to Spalart and Leonards method (Spalart and
Leonard 1987; Spalart 1988), the rescaled b.c. are applied to the additional simulation only, which is solely carried out in order to produce appropriate inflow data
for the main simulation. This main LES prediction directly employs one plane
of the auxiliary simulation as inflow data. The entire procedure is depicted in
Figure 5.7.
It is worth noting that the rescaling method of Lund et al. (1998) can be extended
to compressible flows in order to take Mach number and temperature effects into
account. The corresponding procedure was proposed by El-Askary, Meinke, and
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
207
Storage of data
or on-the-fly
simulation
Inflow
plane
Periodic
Duct
90 bend
Outflow
plane
p
Periodic
boundary conditions
Figure 5.8. Example for the generation of inflow data for a 90 bend using a second simulation for a
straight duct flow with periodic b.c. (Breuer 2002).
Schroder (2001). Following Spalart and Leonard (1987) and Spalart (1988), all
flow variables are first decomposed into a mean and a fluctuating part. Then the
temperature as well as the velocity components are rescaled using appropriate
scaling laws for the inner and the outer layer separately. For a detailed description
of this method, refer to Section 5.4.1 or El-Askary et al. (2001).
(iii) Reusage of data from auxiliary simulations:
The technique of Lund et al. (1998) also fits to the third special resort that is,
the reusage of data obtained by another LES or DNS prediction applying periodic
b.c. Examples are fully inhomogeneous flows such as the flow through a 90 bend
(see Figure 5.8) or around an obstacle mounted at the bottom wall of a plane
channel (Breuer 2002). To generate appropriate inflow conditions for these cases,
additional simulations for a straight duct or a plane channel with the same cross
sections have to be carried out applying periodic b.c. in the homogeneous flow
directions. The instantaneous data from one plane of these auxiliary simulations
can then be applied as inflow b.c. for the inhomogeneous flows. The inflow data
from the additional simulation can either be stored in a preprocessing step to
the intrinsic simulation, leading to an enormous amount of data to be stored, or
generated on the fly, running both simulations in parallel. The second option is
favorably used on parallel computers, directly exchanging the data between the
processors by a communications library such as message-passing interface. In
order to reduce the additional computational efforts with respect to CPU time or
disk space, the successive reusage of a previously generated time series or the
application of a frozen spatial turbulence field making use of Taylors hypothesis
can be taken into account under certain conditions.
All three special cases discussed above lead to physically reasonable inflow conditions of high quality. However, their application is restricted to specific flows and hence
is not practicable in general. An alternative is the generation of artificial inflow data
purely based on the knowledge of the flow geometry and eventually some experimental
data. The first step of this procedure is to split up the (filtered) velocity field u i into a
0:9
P1: IBE
0521871441c05
208
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS

steady, constant mean value u i and a fluctuating component u i according to Reynolds

approach:

u i (xi , t)|inflow = ui (xi )|inflow + u i (xi , t)
.
(5.133)
inflow
Typically the mean values are known from experiments or theory. As a consequence,
the generation of artificial data can be restricted to the fluctuating components. A
variety of methods have been proposed for that purpose in the literature. Here only
some important techniques can be mentioned.
The simplest but also the worst technique to generate time series for the fluctuating
velocities is the application of random-number generators producing white noise. The
only free parameter to adjust is the root-mean-square (rms) values of the fluctuations. A
serious drawback of this method is that there is absolutely no way to take any spatial or
temporal correlations into account. Consequently, the inflow data produced have nothing
in common with the physical situation. Furthermore, time series produced by randomnumber generators mainly consist of high-frequency components that are immediately
damped out by the numerical methods used to solve the filtered conservation equations
especially when coarse grids are used in the vicinity of the inflow plane as is often the
case. Thus, the results of these inflow conditions are more or less identical to laminar
inflow data.
In order to improve the situation by taking low wave numbers into account, Lee, Lele,
and Moin (1992) suggested a method based on stochastic fluctuations with a prescribed
energy spectrum. In general, the idea of all methods proposed is to produce inflow data
that satisfy certain statistical properties such as rms values, cross-correlations, higherorder moments, length and time scales, or energy spectra. In a recent publication, Klein
et al. (2003) provided a good overview of these techniques, typically consisting of two
steps. The first is the generation of a provisional three-dimensional signal for each
velocity component which possesses a prescribed two-point statistic. In the second step
the cross-correlations between different velocity components are taken into account
by using a method proposed by Lund et al. (1998). For the first step two methods
are applicable. The first was provided by Lee et al. (1992) and is based on an inverse
Fourier transform. The second was recently suggested by Klein et al. (2003) and is
based on digital filtering of random data. This method allows a prescribed second-order
(one-point) statistic as well as autocorrelation functions to be reproduced. Compared
with inverse Fourier transforms (Lund et al. 1998), the new procedure possesses several
advantages such as simplicity, flexibility, and accuracy. In Klein et al. (2003) these were
impressively demonstrated based on two different test cases; however, the technique is
generally applicable.
5.2.3 Boundary conditions for solid walls

5.2.3.1 Introduction to solid-wall boundary conditions
A key technology for the application of LES to high-Reynolds-number flows is an

appropriate wall-modeling strategy. Of course, Stokess no-slip boundary condition and
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
209
Table 5.4. Typical mesh sizes (expressed in wall units) for a boundary
layer flow using DNS, wall-resolved LES, and LES with an appropriate
wall model
Streamwise
Spanwise
Wall-normal
x +
z +
min(y + )
Number of points in
0 < y + < 10
DNS
Wall-resolved
LES
LES with
wall model
1015
5
1
50150
1040
1
100600
100300
30150
35
35
Source: Sagaut (2004).
the impermeability condition are also valid for turbulent flows at solid walls. However,
extremely large velocity gradients are found in the near-wall region that are mainly
responsible for the production of turbulent kinetic energy. A prediction based on the
(filtered) NavierStokes equations has to resolve these velocity gradients adequately in
order to determine the level of turbulence production and hence the Reynolds stresses
and wall friction reasonably. This solely requires several grid points in the thin viscous
sublayer (0 y + 5; y + denotes the wall-normal coordinate in wall units). In contrast
to a RANS prediction, an extremely fine grid is required not only in the wall-normal
direction but in all spatial directions to resolve the near-wall turbulence, including
coherent structures such as the well-known high- and low-speed streaks.
Typical mesh sizes for DNS and LES prediction of an attached boundary layer are
displayed in Table 5.4. These values are common, but the quality of the results exhibits
a strong dependence on the size of the mesh especially in LES. As mentioned above
and validated by numerical experiments, the quality of a wall-resolved LES prediction
strongly depends on a good resolution of the inner layer (see Table 5.4). Furthermore,
the resolutions in the streamwise (x + ) and the spanwise (z + ) directions are very
important parameters that also govern the quality of the solution (Sagaut 2004):
r High-resolution LES (i.e., x + 50 and z + 12):
good agreement of the predicted skin friction in plane channel compared with
DNS or experiments when nondissipative numerical methods are used.
r Medium-resolution LES (i.e., x + 100 and z + 30):
thicker and shorter streaks, error on the skin friction.
r Poor-resolution LES (i.e., x + 100 and z + 30):
unphysical streaks, large error on the skin friction.
Physically these considerations reflect the fact that, in an attached turbulent boundary layer, the smallest eddies approximately scale with the distance from the wall, leading to a DNS-like resolution requirement for wall-resolved LES (Spalart et al. 1997).
Piomelli and Balaras (2002) estimated the computational costs for this case to be proportional to Re2.4 . For practically relevant high-Re flows such an extremely fine resolution
is not achievable and workable. For a flow around an airfoil at cruise flight condition
[Re = O(107 )] Spalart et al. (1997) estimated the number of required grid points and
0:9
P1: IBE
0521871441c05
210
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
time steps to 1011 and 5 106 , respectively. Thus, such computations will not be feasible
during the next decades.
5.2.3.2 Classical wall models
As a remedy for this dilemma, wall models (sometimes called wall functions) bridging
the near-wall region and allowing placement of the first grid point in the logarithmic
part of the velocity profile (30 y + 500) have been developed. Consequently, the
near-wall behavior is not resolved in detail, which also leads to reduced requirements
with respect to the grid characteristics in streamwise and spanwise directions. In total,
the resolution requirements are drastically reduced and the computational costs scale
only with Re0.5 , allowing high-Reynolds-number flows to be tackled.
One of the first wall models developed and applied to flows in plane channels and
annuli was the pioneering one by Schumann (1975). It is based on the phase coincidence
of the instantaneous wall shear stress 12,w and the tangential velocity component u at
the grid point nearest the wall. The wall-normal velocity component v is defined by the
impermeability condition. For a flat wall parallel to the xz plane, the wall model for
a colocated grid in nondimensional form reads
12,w (x, z, t) =
w
u(x, y2 , z, t) ,
u(y2 )
22,w (x, z, t) = 0
32,w (x, z, t) =
with
(5.134)
v w (x, z, t) = 0 ,
(5.135)
1 w(x, y2 , z, t)
.
Re
y2
(5.136)
The notation ... represents statistical Reynolds averaging. If homogeneous flow

directions are present, additional averaging along lines or planes is applicable. The
expression y2 denotes the wall-normal distance of the first off-the-wall grid point
and w the wall-normal velocity difference of the spanwise component.
The coincidence assumed between the wall shear stress 12,w and the velocity component u has been experimentally verified. The relationship between these two quantities
is established based on the law of the wall for the averaged flow (e.g., the relation between 12,w and u). As determined by the wall distance y2 , a corresponding law
for the viscous sublayer (VS), the logarithmic buffer layer (BL), and the logarithmic
outer layer (OL) is assumed (see Figure 5.9):
u + = y +
for:
0 y+ < 5
(VS)
(5.137)
u + = a2 ln (y + ) + b2
for:
5 y + < 30
(BL)
(5.138)
u + = a3 ln (y + ) + b3
for:
30 y + 500 (OL),
(5.139)
where y + = y u Re , u + = u/u , and u = w /. The constants are

a2 = 5.0, b2 = 3.05, a3 = 2.5 = 1/, and b3 = 5.05.2.
Based on Schumanns approach, Piomelli et al. (1989) developed a slightly enhanced
version that takes the inclination of the near-wall structures and the resulting temporal
delay between the tangential velocity and the wall shear stress into account (shifted
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
211
30
u+
u+= y +
20
Figure 5.9. Law of the wall u+ (y+ ) in a turbulent
boundary layer without or with only a weak pressure
gradient (half-logarithmic plot).
+ 1
+
u = __
ln y + C
10
0
1
10
100 + 1000
y
logarithmic buffer
viscous
layer (BL)
sublayer (VS)
10000
logarithmic outer
layer (OL)
boundary condition). For the most important wall shear-stress component 12,w , this
model reads
12,w (x, z, t) =
w
u(x + s, y2 , z, t) ,
u(y2 )
(5.140)
where the spatial delay is determined by s = |y2 | cot ( ) with opt 813 .
Another wall model suggested by Piomelli et al. (1989) is the ejection boundary
condition, which is based on the observation that the near-wall dynamic is dominated by
sweeps and ejections. Both models lead to slightly improved results for the plane channel
flow compared with Schumanns original formulation. A major drawback of all these
models is that they are difficult to assign to complex, statistically three-dimensional
flows because they require the determination of the averaged wall-shear-stress and
velocity. Additionally, the use of the customarily applied laws of the wall is highly
questionable for flows with large pressure gradients or even local separation and recirculation regions.
A further extension of Schumanns and Piomellis models was provided by Bagwell
et al. (1993). On the basis of a linear stochastic estimation approach, they use the
entire velocity field in a plane parallel to the wall in order to determine the wall shear
stress, which a priori requires a two-point correlation tensor generally not available.
Following Schumanns concept, Werner and Wengle (1993) suggested a wall model
that is also based on the phase coincidence mentioned above but that applies the laws
of the wall (viscous sublayer and power law) directly to the instantaneous velocity field.
This simplifies the determination of the wall shear stress and allows the use of their
model in flows with separations. However, the application of the laws of the wall to
instantaneous velocities and in separated flows is theoretically not justified.
Hoffmann and Benocci (1995) derived an analytical expression for the local wall
shear stress. They analytically integrated the boundary layer equations coupled with an
0:9
P1: IBE
0521871441c05
212
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
algebraic eddy-viscosity model. Neglecting both advective terms and approximating the
unsteady term, they finally ended up with an expression for the wall shear stress leading
to satisfactory results for plane-channel flow and rotating-channel flow. An improved
model was also derived by Manhart (2001) that takes the local instantaneous pressure
gradient in the streamwise direction into account. Actually, the search for appropriate
wall models is the ongoing. Abel, Stojkovic, and Breuer (2003, 2004) applied advanced
statistical techniques such as nonlinear stochastic estimation to instantaneous DNS
and wall-resolved LES data to identify the most important model terms and to derive
appropriate wall models for LES. In a first attempt this technique was successfully
applied to attached flows but the extension to separated flows is actually carried out.
5.2.3.3 Zonal and nonzonal approaches
Another possibility for wall modeling is to use so-called zonal or nonzonal approaches
based on the explicit solution of a different set of equations in the inner and outer layers
(Balaras and Benocci 1994; Balaras, Benocci, and Piomelli 1996; Cabot and Moin
1999; Piomelli and Balaras 2002). The basic assumption is that the interaction between
the near-wall region and the outer region is weak.
For modeling the inner near-wall region, simplified governing equations such as
the two- and three-dimensional thin boundary layer equations or the RANS equations
with a statistical turbulence model are taken into account. These equations are solved
on an embedded inner grid in the direct vicinity of the wall, whereas the original
LES prediction is carried out on an outer grid not resolving the near-wall region. This
leads to a two-layer model as proposed by Balaras and Benocci (1994) and extensively tested by Balaras et al. (1996). The boundary layer equations for the inner layer
read

u i
u i
+
( + t )
,
(5.141)
(u n u i ) =
+
t
xi
xi
xn
xn
where n denotes the wall-normal direction y and i = 1 and 3 if we assume a wall
parallel to the xz plane. The wall-normal velocity component u n is predicted based
on the mass conservation equation for the inner layer:
y
un =
0
u i
dy .
xi
(5.142)
The system is closed by setting the wall boundary condition for the inner layer to the
no-slip condition u i = 0, and at the upper boundary the velocity distribution is obtained
from the outer-flow LES prediction as a freestream condition (Piomelli and Balaras
2002). Furthermore, the pressure gradient p/ xi in Equations (5.141) is assumed to
be independent of y in the inner layer and thus taken from the outer-flow prediction.
Consequently, no Poisson equation for the pressure has to be solved and the costs for
solving the two momentum equations in the inner layer are only marginally higher than
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
using equilibrium boundary conditions such as in the wall models of Schumann (1975)
and Piomelli et al. (1989).
The quality of the results depends on the choice of the model for the eddy viscosity t .
The most commonly used model for t , originally applied by Balaras and Benocci (1994)
and Balaras et al. (1996), is a simple mixing-length model with near-wall damping. It
reads

3
y+
2
+
+
,
(5.143)
t = ( y) |S| D(y ) with D(y ) = 1 exp +
A
where denotes the von Karman constant, |S| the magnitude of the strain rate, y + the
distance from the wall in wall units, and A+ = 25, a constant. Finally, the wall-stress
components obtained from the integration of Equations (5.141) in the inner layer are
used as boundary conditions for the outer LES prediction.
The two-layer model was successfully applied to channel flow with and without extra
rotation and backward-facing step flow. In the latter case, however, the boundary layer
equations used within that model are no longer valid in the vicinity of the separation
region. If the full RANS equations are applied instead, a hybrid LESRANS approach
is achieved. Because the regions for RANS and LES are defined in advance, this method
is called a zonal approach (see also Chapter 4 for more details).
The counterparts to the zonal techniques discussed above are nonzonal hybrid LES
RANS approaches. Here a gradual transition between both methods takes place based on
an automated switch, ideally removing the need for user-defined information. Speziales
(1998) formulation belongs to this latter class. Continuing the idea of Speziale, Zhang,
Bachman, and Fasel (2000a) have demonstrated the first successful applications of this
hybrid concept for a flat-plate boundary layer with and without separation. Conceptually,
their approach called flow simulation methodology is very similar to the detached-eddy
simulation (DES) proposed by Spalart et al. (1997) and Spalart (2000), which is more
widely known. The DES approach can still be considered a zonal method because the
two domains (LESRANS) are fully determined by the grid topology and the segmentation is independent of the flow solution. In DES, attached flow regions (attached eddies)
are distinguished from separated flow regions (detached eddies). The former are properly predicted based on RANS with statistical turbulence models, whereas the latter,
including large-scale unsteady vortices, are computed more reasonably by LES. Thus,
DES could be represented as a natural hybrid method combining RANS and LES. This
means that, near solid boundaries, the governing equations work in the RANS mode
(i.e., all turbulent stresses should be modeled). Furthermore, pressure and velocity fields
are time averaged, and unsteady vortical structures should not be resolved directly. Far
away from solid boundaries, the method switches to the LES mode. Hence, the basic
concept is to combine the advantages of both methods, yielding an optimal solution at
least for a special class of flows, and to afford predictions of high-Reynolds-number
flows with reasonable computational effort. For further details, refer to Chapter 4.
However, a variety of open issues need to be addressed before one can rely on
hybrid methods. These include, in particular, the demand for appropriate coupling
213
0:9
P1: IBE
0521871441c05
214
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
techniques between LES and RANS, adaptive control mechanisms, and proper SGS
RANS models. Because approximate wall-boundary conditions partially rely on RANS
modeling and the classical wall functions discussed above are a limiting case of a RANS
model in the near-wall region, both topics are strongly interconnected. For a complete
review of wall-modeling strategies, refer to Piomelli and Balaras (2002).
5.2.3.4 Pressure and temperature wall-boundary conditions
The previous section solely concentrated on the velocity field. Appropriate boundary
conditions for the pressure field (if required at all) are not so critical. They depend on
the flow problem considered and on the numerical methodology applied. For example,
the boundary layer over a flat plate at rest exhibits a zero wall-normal pressure gradient
at the wall, which can be discretized by a Neumann boundary condition. For a more
general case of a flow over a curved, moving, or rotating surface or when external
forces such as buoyancy or centrifugal forces are present, large pressure gradients may
appear in the vicinity of the surface. In that case it is advisable to determine the pressure
gradient based on a simplified momentum equation in the wall-normal direction using
one-sided finite differences or to extrapolate the pressure at the wall from the internal
region to the boundary. Because these techniques used in the context of LES do not
deviate from standard applications for laminar or RANS flow predictions, we refer to
the basic literature for CFD.
One important issue remains. In the case of nonisothermal flows in which an additional energy equation (e.g., for the temperature T or the enthalpy h) has to be solved,
a situation similar to that of the velocity field arises. In principle, a wall heat flux qw or
wall temperature Tw can be prescribed and discretized without further approximations.
However, like the viscous wall layer for the velocity field, this measure requires the
conductive wall layer to be resolved. As determined by the molecular Prandtl number
Pr = c p / = / of the fluid, which describes the ratio of diffusivities for momentum and heat, this layer might be even thinner than the viscous layer (e.g., for water,
<
<
<
1.7 <
Pr 13.7, and for oil, 50 Pr 100,000). If a fine near-wall resolution is
not possible or desired, suitable time-dependent formulations of wall models for the
temperature equation have to be applied. These models relate the instantaneous local
heat fluxes to the temperature fluctuations at the grid point nearest the wall by using
time-averaged wall laws (Grotzbach 1981, 1987; Grotzbach and Worner 1999). Consequently, these models are basically analogous to the wall models for the velocity field
such as the models of Schumann (1975) and Piomelli et al. (1989) described above.
Furthermore, any additional scalar transport equation (e.g., for species) can be treated
in much the same way as the temperature equation.
5.2.4 Far-field boundary conditions for compressible flows
Most of the boundary conditions discussed above in the context of incompressible flows
can easily be extended to compressible flows (e.g., the physical boundary conditions at
solid walls). However, at artificial boundaries arising owing to finite integration domains
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
215
for the prediction of external compressible flows, appropriate far-field boundary conditions are required. This issue is directly related to the boundary conditions applied to the
acoustic simulation explained in detail in Section 5.3. Hence, we only address this topic
briefly and refer to this section and the cited references therein for further details.
For the derivation, all viscous effects are neglected in the far field, leading to the
Euler equations. Although this assumption is in general questionable for vortical flows,
it is a reasonable condition under practical aspects such as a large extension of the
domain and a highly stretched grid with a coarse resolution in the vicinity of the farfield boundaries. The three-dimensional Euler equations for a compressible fluid are
a hyperbolic system of five equations with five real eigenvalues. These eigenvalues
define the directions along which information is transported. Owing to different signs
of the eigenvalues, subsonic flows (Mach number M = u/a < 1) and supersonic flows
(M > 1) have to be distinguished. If a positive mean flow direction (u > 0) for a
supersonic flow is assumed, all eigenvalues are positive. This is the most uncritical case
because the transport of information takes place only in a positive coordinate direction.
For an inflow boundary, this means
bc = , u bc = u , vbc = v , wbc = w , pbc = p ,
(5.144)
that is, the values of all variables at the boundary (bc) are prescribed by the undisturbed
inflow (). For an outflow boundary, this means
bc = int , u bc = u int , vbc = vint , wbc = wint , pbc = pint ,
(5.145)
that is, the values of all variables at the boundary are set equal to the values of the first
grid point of the internal domain (int).
For a subsonic flow (M < 1), four eigenvalues are positive and only one eigenvalue
is negative. In contrast to the supersonic case, this means that four variables have to be
prescribed by the undisturbed flow at the inlet and one variable has to be extrapolated
from the internal region to the boundary (e.g., by a Neumann boundary condition).
Correspondingly, at the outlet, four variables at the boundary are given by the internal
field, whereas one variable is defined from outside. The latter may lead to reflections. As
shown in Section 5.3, the Euler equations can be transformed to a characteristic form that
builds the base for nonreflecting boundary conditions. For that purpose the Jacobian
matrix A in the Euler equations is reconstructed to the new matrix A+ , using only
positive eigenvalues. On the basis of this modified system of equations nonreflecting
boundary conditions were derived (Hirsch 1990; Engquist and Majda 1977). More
details, including the corresponding reflection coefficients, are presented in Section 5.3.
5.2.5 Final remark for discretization schemes
In conclusion, boundary conditions play a dominant role for LES predictions of high
quality, and the derivation of appropriate b.c. is not a trivial task. Especially for inflow
0:9
P1: IBE
0521871441c05
216
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
boundaries and solid walls, intensive research activities are ongoing to widen the applicability of LES, partially leading to hybrid LESRANS approaches. This will result
in exciting new techniques and prospects for LES.
5.3 Boundary conditions: Acoustics

Fang Q. Hu
With current computing capabilities, it remains difficult to compute the turbulence
and its acoustic radiation at the same time in most practical problems. One apparent
compromise is to compute the turbulence and the acoustic field separately. An LES
or unsteady RANS calculation, which captures the main physics of the underlying
noise-generating flow, is conducted first. Then a second calculation is carried out for the
acoustic radiation of the turbulent flow to the far field, using the result of the turbulence
calculation as the input. The acoustic calculation is often done by solving the Ffowcs
WilliamsHawkings equation or the linearized Euler equations. This section deals with
the implementation of far-field nonreflecting boundary conditions for the linearized
Euler equations. Issues related to the boundary conditions for the LES calculations
have been discussed in the previous section.
At the far-field boundary, it is possible to use the linearized Euler equation because
the viscous and nonlinear effects are often negligible. For convenience, we will use the
following linearized Euler equation as the model equation:
u
u
u
u
+A
+B
+C
+ Du = 0,
(5.146)
t
x
y
z
where

0 0
0 0
U
0 0 0

1
0 0
0
u
0
0 0
U
0 0

0
0 1 ,
u = v ,A = 0
0
U 0 0 , B = 0 0

0 0
0
w
0
0 0
0
0 U 0
p
0 0 P0 0 0
0 P0 0 0 U
0
0
C = 0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
P0
0
0
0
0
0 , D = 0
1
0

0
0
0
0
0
0
0
d
dy
d U
dy
0
0
0
0
0
0
0
0
0
0
,
0
0
0
and where U , ,
and P0 are the mean velocity, density, and pressure, respectively, and
is the specific heats ratio. We have assumed a parallel mean flow that varies only in
the y-direction. In many practical situations, the far field has a uniform mean flow.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
217
5.3 BOUNDARY CONDITIONS: ACOUSTICS
We will review the most commonly used nonreflecting boundary conditions, including the characteristic method, radiation conditions, and absorbing zones as well as
the recently emerged perfectly matched layer technique.
5.3.1 Characteristic nonreflecting boundary condition
The characteristic nonreflecting boundary condition is based on the characteristic splitting of the Jacobian matrices A, B, or C in Equation (5.146) at a boundary where the
x-, y- , or z-coordinate is constant, respectively (Giles 1990; Thompson 1990). Extensions to viscous NavierStokes equations can be found in Poinsot and Lele (1992). For
example, consider a nonreflecting boundary at an outflow boundary x = x0 . Let the
characteristic decomposition of A be
A = EE1 ,
(5.147)
where is the eigenvalue diagonal matrix and E is the eigenvector matrix. Then, for
boundary grids at the nonreflecting boundary, the differential equation is modified to
be
u
u
u
u
+ A+
+B
+C
+ Du = 0,
(5.148)
t
x
y
z
where A+ is reconstructed from Equation (5.147) using only the positive eigenvalues,
namely, A+ = E+ E1 . The spatial derivative in the direction normal to the boundary,
the x derivative in this example, can now be evaluated using backward differences.
Note that the characteristic nonreflecting boundary condition is not exact because
matrices A, B, and C in Equation (5.146) are not simultaneously diagonalizable. In fact,
for a uniform mean flow of Mach number M, the reflection coefficient for an outgoing
acoustic wave is found to be
1 cos i
,
(5.149)
Racoustic =
1 cos r
where i and r are the angles of the outgoing and reflected waves, respectively (Hu
and Atkins 2003). Here the reflected angle r is related to i as
sin i
sin r
=
.
1 + M cos r
1 + M cos i
(5.150)
The reflection coefficient for an outgoing vorticity wave is found to be

Rvorticity =
sin i
,
1 cos r
(5.151)
where the incident and reflected angles are related by

sin r
tan i
=
.
1 + M cos r
M
(5.152)
Equations (5.149) and (5.151) indicate that the characteristic nonreflecting boundary
condition is most effective for outgoing waves with a nearly normal incident angle.
0:9
P1: IBE
0521871441c05
218
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Several techniques are available in the literature that can be used to circumvent this
limitation partially (Atkins and Casper 1994; Higdon 1987).
5.3.2 Radiation boundary condition
The radiation boundary condition is based on the assumption that certain asymptotic
forms can be developed for the solutions at the far field (Bayliss and Turkel 1982a).
The asymptotic expansion is usually written in ascending powers of 1/r , where r is the
distance from the boundary point to the source of sound. Once the asymptotic expansion is known, differential equations are derived such that they can be satisfied by the
asymptotic expansion up to a certain order of 1/r . These differential equations constitute the radiation boundary condition. In this way, the radiation boundary condition
is approximately matched to the Euler equation. The radiation boundary conditions
have only outward characteristics; thus, they can be easily discretized with backward
differences or without any information from the domain exterior to the computational
domain.
A far-field radiation boundary condition that lets out acoustic waves in two dimensions has been given by Tam (1998a):
u
1 u u
+
+
= 0,
V () t
r
2r
(5.153)
where V () = M cos + 1 M 2 cos2 . Here M is the mean flow Mach number and
the angle of the boundary grid point measured from the direction of the flow.
For 3D acoustic waves, Bayliss and Turkel (1982a) give a series of radiation conditions with increasing order of accuracy. The leading approximation yields the radiation
condition
p
p p
+
+ =0
(5.154)
t
r
r
for the pressure component.
Because the radiation boundary condition is based on the asymptotic expansion of
the solution, it works best if the nonreflecting boundary is far away from the source of
the sound.
5.3.3 Absorbing-zone techniques
These methods are variously referred to as absorbing zones, sponge layers, exit
zones, or buffer zones in the literature. The basic strategy of these methods is to
introduce additional zones of grid points, or layers, to surround the physical domain
so that outgoing disturbances are attenuated in the added zones and thus minimize the
reflections. The numerical solutions inside absorbing zones need not be physical as
long as the use of the zone does not cause significant reflection back into the physical
domain and the zone is numerically stable. Various ways of constructing the absorbing
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
219
zones have been proposed in the literature. Some of these approaches will be reviewed
here. We note that these approaches are often implemented in a combined fashion. In
some cases, the absorbing zone is terminated with a characteristics-based, nonreflecting
boundary condition at the exit end of the zone.
5.3.3.1 Artificial dissipation and damping
In this method, an absorbing zone is created and appended to the physical computational
domain in which the governing equations are modified to mimic a physical dissipation
mechanism (Israeli and Orszag 1981).
For the Euler and NavierStokes equations, the artificial damping term can easily
be introduced into the governing equations as follows:
u
= L(u) (u u0 ),
t
(5.155)
where u is the solution vector and L(u) denotes the spatial operators of the equations.
The damping coefficient assumes a positive value and should be increased slowly
inside the zone. Here, u0 is the time-independent mean value in the absorbing zone
(Freund 1997). Kosloff and Kosloff (1986) analyzed a system similar to Equation
(5.155) for the two-dimensional wave equation in which, in particular, a reflection
coefficient of a multilayer absorbing zone was calculated.
5.3.3.2 Grid stretching and numerical filtering
Attenuation of the solution may also be achieved by using purely numerical means. One
such approach, used by Rai and Moin (1991) and Colonius, Lele, and Moin (1993),
is to create a sponge layer, or exit zone, in which the grids are stretched and
coarsened. When an outgoing wave enters the sponge layer, it becomes underresolved
in the coarsened grid. Because most numerical schemes have a built-in mechanism of
dissipating the disturbances in unresolved scales, the numerical solutions inside the
sponge layer are attenuated through numerical dissipation.
Computationally, grid stretching is equivalent to modifying spatial derivative terms.
For example, if the grids in the x-direction are to be stretched, one can replace the x
derivative by
1
,
x
(x) x
(5.156)
where (x) 1 is an increasing function with = 1 at the start of the added zone.
Then, inside the sponge, spatial derivative terms can be discretized in the same way as
those in the interior region.
The stretching has to be done very gradually, and (x) should be a slowly varying
function. A sudden increase in grid spacing can cause numerical reflection (Hu and
Atkins 2003; Vichnevetsky 1981). A model of smoothly stretching the grid has been
discussed by Colonius et al. (1993).
0:9
P1: IBE
0521871441c05
220
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
The attenuation of the numerical solution can be enhanced by applying low-pass

numerical filters inside the added zone, thus reducing the length of the exit zone.
Colonius et al. (1993) enforced a five-point explicit filter operation in the sponge layer.
Visbal and Gaitonde (2001) used implicit high-order filters. Visbal and Gaitonde (2001)
noted that, because grid stretching can cause high-frequency grid-to-grid oscillations,
high-order filters should be applied to the solutions in the physical domain as well.
Examples were presented by Visbal and Gaitonde (2001) in which grid stretching could
even be done nonsmoothly when combined with high-order filters. Liu and Liu (1993)
combined a grid stretching in the streamwise direction with an increase in viscous
dissipation. The performance of the boundary condition depends on the gradualness
with which the mean flow is increased.
5.3.3.3 Modification of convective mean velocity
Streett and Macaraeg (1989) and Taasan and Nark (1995) modified the mean flow inside
the buffer zone for the compressible Euler equations so that it increased gradually and
eventually became supersonic at the end of the buffer domain. At that point, because
the flow is supersonic and outward, termination of the grid will not cause any reflection.
Freund (1997) combined this approach with additional damping terms similar to those
used in a sponge layer.
The formulation and implementation of the buffer zone appears to be relatively simple. It essentially involves an addition to the governing equations of artificial convective
terms of the form
U0 (x, y)
u
u
+ V0 (x, y) ,
x
y
(5.157)
where U0 and V0 are the artificial velocities that are zero at the start of the buffer zone
and gradually increase to become supersonic in the outward direction.
5.3.4 Perfectly matched layers
The perfectly matched layer (PML) technique employs the same strategy as that used
in the absorbing-zone techniques reviewed in the previous section. The main difference
is that the PML equations are constructed in such a way that they match perfectly to the
governing equations of the physical domain while being absorbing for all disturbances
that enter the PML zone. The match is perfect when the interface of the physical
and PML domains is reflectionless for waves of any frequency and angle. As such,
PML zones are usually much shorter compared with other absorbing zones and are less
sensitive to parametric variations of the zone. The first PML formulation was introduced
by Berenger (1994) for the Maxwell equation in computational electromagnetics. A
numerically stable PML formulation for the linearized Euler equation was recently
given by Hu (2001).
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
221
The PML technique can be viewed as a complex change of variable in the frequency domain (Collino and Monk 1998; Gedney 1996; Petropoulos 2000; Zhao and
Cangellaris 1996). For instance, a change of variable
x
i
x d x,
(5.158)
x x + x , where x =
x0
would be used for an x-layer starting at x = x0 , where x is the absorption coefficient
and can be any positive function of x or a constant. For the linearized Euler equation
with a convective mean flow, it has been pointed out by Hu (2001) that a Prandtl
Glauert-type, or a Lorentz-type, transformation should be applied to the Euler equation
before employing the complex change of variable (5.158) to avoid causing instability
in the PML domain.
In three-space dimensions, the PML equation for Equation (5.146) with a uniform
mean flow can be expressed as follows:
u yz
ux z
ux y
u
+A
+B
+C
+ u + x Au yz = 0,
(5.159)
t
x
y
z
where
U
(5.160)
= 2
a U 2
and u yz , ux z , ux y , ux , and u denote
u yz = u + ( y + z )q1 + y z q2 ,
(5.161)
ux z = u + (x + z )q1 + x z q2 ,
(5.162)
ux y = u + (x + y )q1 + x y q2 , ux = u + x q1 ,
(5.163)
u = (x + y + z )u + ( y z + x z + x y )q1 + x y z q2 .
(5.164)
Here q1 and q2 are auxiliary variables and are computed as

q1
= u,
t
(5.165)
q2
= q1 .
(5.166)
t
The absorption coefficients x , y , and z are functions of x, y, and z, respectively,
with the Euler equation as the special case of x = y = z = 0. The auxiliary variables
q1 and q2 are needed only inside the PML zones. Equation (5.159) can be further
simplified wherever any of the absorption coefficients is zero. In particular, q2 is not
needed in a region where any two of the absorption coefficients are zero.
In two dimensions, we have a simpler version,
u
u y
ux
+A
+B
+ u + x Au y = 0,
t
x
y
(5.167)
where
ux = u + x q1 , u y = u + y q1 , u = (x + y )u + x y q1 .
(5.168)
0:9
P1: IBE
0521871441c05
222
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Further details on the numerical implementation of PML equations are available

in Hu (1996a, 1996b, 2001, 2002). PMLs for nonuniform mean flows are given by
Hu (2004).
5.3.5 Summary of boundary conditions for acoustics
As far as accuracy is concerned, the PML technique appears to be the most accurate
among all the methods reviewed here. However, the PML equations are not yet available
for every type of nonreflecting boundaries to be encountered in practical situations. So
far, this technique has only been developed for a parallel mean flow in a direction
aligned with one of the coordinates. The technique is still very much under active
development (Abarbanel, Gottlieb, and Hesthaven 1999; Hagstrom and Nazarov 2003;
Hu and Atkins 2003; Hu 2004, 2005). In this regard, the absorbing-zone techniques can
be applied to a wider class of problems and are often coupled with the characteristic
boundary conditions. For problems with a centralized and compact noise source, the
radiation condition offers an effective alternative.
5.4 Some concepts of LESCAA coupling

Wolfgang Schroder
and Roland Ewert

It has already been stated in Chapter 1 that community noise is one of the major
problems to be addressed by the aircraft industry, to improve the quality of life in
the neighborhood of airports, and to ensure the current growth of passenger numbers.
Theoretically, radiated jet noise as well as airframe noise can be determined by solving
the unsteady, compressible NavierStokes equations. However, this straightforward
method to predict the acoustic field of technically relevant subsonic turbulent flow
problems significantly exceeds todays computational capabilities.
To substantiate this statement, let us briefly approximate the computational effort
for the direct approach of the sound field. If the characteristic length is denoted by L
and is the Kolmogorov length scale to describe the smallest eddies, the total number
of grid points of a uniform mesh is
3
L
.
(5.169)
Nuni
Considering L to be a multiple of the integral length scale of turbulence (i.e.,

3/4
L = c f ) and isotropic turbulence with / O (Ret ), we obtain
9/4
N f c3f Ret ,
(5.170)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
223
5.4 SOME CONCEPTS OF LESCAA COUPLING
%
where the turbulent Reynolds number Ret = u 2 / is used and c f > 1 represents
a constant.
To take into account the acoustic length scale, we consider the Strouhal number Sr =
L /u that can be interpreted to be proportional to the ratio of the characteristic length L
to the length scale of the vortices u/ (i.e., Sr L/ ). The Helmholtz number
H e = L /c relates L to the acoustic wavelength a c/ such that the vortical and
the acoustic structures are connected via H e = Sr M, where M is the Mach number
characteristic to the flow problem. More precisely, in the case of low-speed flows the
acoustic length is M 1 larger than the vortical length a M 1 . Using the relation
between L and a via the Helmholtz number L ca a and expressing the product
M 1 by the integral length scale and the turbulent Mach number M 1 Mt1
such that L ca Mt1 with ca > 1 holds, we obtain for the grid resolution on a
uniform mesh
N f a ca3 Mt3 Ret .
9/4
(5.171)
Equation (5.170) approximates the grid resolution of the turbulent flow field, and
Equation (5.171) estimates the number of cells to resolve the flow field plus the acoustic
field. If we drop the dependence on the constants c f and ca , it is evident that N f a
scales with Mt3 compared with N f . The turbulent Mach number Mt is in the range of
O(102 ) for subsonic jets and boundary layers. Therefore, the increased requirement
in storage resources is extremely severe in low-speed flows. Moreover, besides this
memory problem there is the issue of accuracy. Because acoustic pressures can be almost
neglected compared with hydrodynamic pressures, the numerical approximations in
CAA have to be more accurate than in classical CFD methods.
Given the strong discrepancy in the characteristic length of the flow field and that of
the sound field, it is more or less natural to separate the computation of both problems
that is, to apply the hybrid approach to the analysis of low-speed flow and noise problems. In the first step, the flow field is determined by solving the NavierStokes equations via a traditional CFD method, and in the second step, the sound field is computed
by a numerical solution of an appropriate system of acoustic equations. Rather than
using approximate models to describe the sound sources occurring in the acoustic
equations, we base our discussion on the assumption that large-eddy simulations are
performed to predict the acoustic source terms accurately. That is, an LES is conducted just in the region where the noise is generated, as indicated by the inner domain in Figure 5.10 for the trailing-edge noise problem, and the LES data are used as
acoustic input to determine the sound propagation over the complete outer domain in
Figure 5.10.
Numerous sound propagation equations exist among other formulations various
acoustic analogies, the most famous of which is Lighthills acoustic analogy (Lighthill
1952), and some equations that are based, generally speaking, on some form of the
linearized Euler equations (Hardin and Pope 1994; Bailly et al. 2000). In our analysis
0:9
P1: IBE
0521871441c05
224
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
l
x
Figure 5.10. Sketch of the computational domains to determine, for example, trailing-edge noise
with the hybrid approach. The LES domain encompasses the vicinity of the trailing edge, whereas the
computational aeroacoustics domain includes the whole airfoil owing to the less stringent demands
concerning the grid resolution (xacoustic xLES /M). Thus, scattering at the leading edge captured
and directivities predicted.
we use the acoustic perturbation equations that were derived in Ewert and Schroder
(2003):

+ ( u + u ) = 0,
t

p
u
+ u u +
= qm ,
t
p
) = qe
+ c2 ( u + u
t
(5.172)
(5.173)
(5.174)
with the source terms determined by the LES

qm = ( u + u ) + T s s T ,
qe =
p s
,
c p t
(5.175)
(5.176)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
where the customary notation has been used (i.e., u is the velocity, the vorticity,
the density, T the temperature, p the pressure, s the entropy, and = c p /cv the
ratio of specific heats). These equations describe mean flow convection effects, but
unlike other systems for the acoustic field they do not possess instabilities for any
nonuniform mean flow field with arbitrary density gradients (Ewert and Schroder 2003).
The coupling between the LES and the CAA solutions is based on the source terms in
Equations (5.175) and (5.176), which are cyclically fed into the acoustic simulation.
Over the time interval, the windowing concept, which is used in signal processing to
provide a smooth periodic distribution, is applied to the source data to avoid artificial
discontinuities between the data samples.
When the formulations for the acoustical field (like those presented in Hardin and
Pope 1994; Bailly et al. 2000; Shen and Srensen 2001) are used, the coupling procedure
is more or less the same. Only different source terms occur that might require various input data from the solution. Because a detailed comparison of those source formulations
is beyond the scope of this section, the interested reader is referred to Ewert, Meinke,
and Schroder (2001a) for a more in-depth discussion of the differences. Furthermore,
to make this contribution concise and yet thorough, we will discuss neither the impact
of different interpolation strategies between the LES and the CAA domain, nor the
influence of different windowing formulations of the source terms, nor the pros and
cons of various weak or strong coupling concepts, nor the bearings of the interaction on
the acoustical and the flow field with respect to phenomena such as receptivity. However, because the efficiency of the hybrid approach, which is the basis for the coupling
discussion, strongly depends on the locality of the LES solution and on an acoustically
accurate formulation of the boundary conditions on the boundary between the flow domain and the acoustic domain (Figure 5.10), the following discussion of the LESCAA
coupling focuses on two issues that in general do not receive sufficient attention in the
context of the two-step approach. That is, we consider a concept to prescribe inflow
distributions for an LES locally (El-Askary et al. 2001, 2003) and a silent condition on
the LESCAA boundary (Ewert, Meinke, and Schroder 2002). The fundamentals of
large-eddy simulations, sound-propagation equations, and appropriate discretizations
are described at length in Chapters 2, 3, and 4 and Section 5.1, respectively.
5.4.1 LES inflow boundary
Although the boundary conditions of local large-eddy simulations were already discussed in Section 5.2, we will address once again in the following section the issue
of inflow formulations with emphasis on compressible flows with and without weak
pressure gradients. The incorporation of varying pressure distributions is of major
importance. For example, in the vicinity of the trailing edge of an airfoil, which is
one of the most significant noise sources during takeoff and landing, the academic
condition of a constant pressure distribution is not fulfilled. Because the impact of the
upstream condition persists for large distances downstream, it defines the size of the
LES domain. The better the inflow condition, the smaller the streamwise extension of
225
0:9
226
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Rescaling
Rescaled Turbulent Velocity Profile
Velocity Profile to be Rescaled
Inflow
Outflow
P1: IBE
0521871441c05
y
x
( in )
( re )
Figure 5.11. Sketch of the rescaling concept.
the LES region and as such the more efficient the hybrid approach no matter what kind
of description is used for the acoustic step. In this sense the turbulent inflow boundary
conditions in spatially developing compressible flows are extremely significant for the
hybrid concept.
A simple inlet formulation is based on periodic boundary conditions, which unfortunately are restricted to a few simple geometries like channel flow. Spalart and Leonard
(1987) were able to extend the application of periodic formulations by introducing a
transformation, which is defined via the minimum streamwise inhomogeneity of the
boundary layer. Thus, their approach is restricted to spatially evolving boundary layers,
the mean streamwise variation of which is small compared with the transverse alteration. It goes without saying that periodic boundary conditions are very appealing, but
the formulation of Spalart and Leonard (1987) suffers from several additional intricate
terms whose evaluation is very costly. Therefore, Lund et al. (1998) reformulated the
concept discussed in Spalart and Leonard (1987). Unlike Spalart and Leonard (1987),
they transform only the boundary conditions by using standard scaling laws for an
equilibrium turbulent boundary layer. Their formulation is much simpler; however, the
periodicity of the boundary conditions is no longer valid. An auxiliary simulation, which
produces its own inflow distribution by rescaling the velocity field from a downstream
location and reintroducing it at the inlet, is used to extract instantaneous distributions of
velocity. A schematic of the procedure, which yields so-called semiperiodic or scaled
boundary conditions, is shown in Figure 5.11. In the following, the rescaling method is
described. First, the discussion focuses on compressible flows at zero pressure gradient;
then, a formulation for a weakly variable pressure gradient in the mainstream direction
is introduced, which was already successfully used in Meinke et al. (2004) to simulate
airfoil trailing-edge flow.
To be able to account for Mach number and temperature effects across the boundary
layer, El-Askary et al. (2001, 2003) extended the rescaling method of Lund et al. (1998)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
227
to compressible fluids. Following Spalart and Leonard (1987), all flow components are
decomposed into a mean and fluctuating part, and then an appropriate scaling law is
applied to each quantity separately for the inner and outer layer. Consequently, the
velocity fluctuation u reads
u = u(x, y, z, t) U (x, y),
(5.177)
where u = (u, v, w)T is the instantaneous velocity and U = (U, V, W )T the mean
velocity. Note, however, that the spanwise component vanishes such that no rescaling
in the spanwise direction is necessary. The mean streamwise velocity component at the
downstream station is rescaled (index re) and linked with that at the inlet (index in)
inner +
(yin ) = s Uvd,re (yin+ ),
Uvd,in
(5.178)
outer
Uvd,in
(in ) = s Uvd,re (in ) + (1 s )Uvd, .
(5.179)
In these equations the inner law coordinate y + = u y/ is defined via the friction
velocity u = w /, represents the outer coordinate, and the subscript vd refers

to the van Driest transformation, which is applied only to the streamwise component

U (x, y)
1
Uvd (x, y)
1
sin
B
(5.180)
=
U
B
U
with
&
'
'
B=(
1
2 r
M
2
2
+ 1
M
2
(5.181)
where r represents the recovery factor and M is the freestream Mach number. The
velocity fluctuations are analogously formulated:
inner
+
yin , z, t = s u re yin+ , z, t ,
u in
(5.182)
inner

u in (in , z, t) = s u re (in , z, t) ,
(5.183)
where the rescaling factor s is defined as

u ,in
s =
> 1.
u ,re
(5.184)
The superscripts inner and outer denote the inner and outer layer components
of the variables, respectively. To account for compressibility effects the temperature is
rescaled, too. For the mean static temperature T , we locally use the equation of Walz
(1969),

T
|U |2
=1+ A 1 2
,
(5.185)
T
U
0:9
P1: IBE
0521871441c05
228
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
where
2
A = 0.5 ( 1) r M
(5.186)
with the ratio of specific heats, . Furthermore, the temperature fluctuation T can
be calculated if one assumes a negligible total temperature fluctuation Tt compared
with the static temperature fluctuation. This assumption is valid at Mach numbers in
the range M 3.0 (Bradshaw 1977). Then, by focusing on the streamwise velocity
component and by neglecting the higher-order velocity fluctuations, we find that the
static temperature fluctuation reads T = U u /c p or, in rewritten form,
T (y, z, t)
u (y, z, t)
= ( 1) M 2
,
T (y, z)
U (x, z)
(5.187)
where M is the local Mach number determined by the mean velocity and the mean
temperature. Using Equations (5.185) and (5.187) and following the rescaling process,
we find that the equations for the rescaled temperature and its fluctuations read
Tininner (yin+ ) = s2 Tre (yin+ ) + C1 T ,
(5.188)

Tin,inner yin+ , z, t = s2 Tre (yin+ , z, t),
(5.189)
Tinouter (in ) = s2 Tre (in ) C2

T in
outer
Ure (in )
T + C3 T ,
U
(5.190)
(u )re (in , z, t)
T
U
(5.191)
(in , z, t) = s2 Tre (in , z, t) C2
with
C1 = (1 + A) (1 s2 ),
C2 = 2 A s (1 s ),
C3 = (1 s ) [1 + s + 2 A s ].
We now discuss the impact of an adverse pressure gradient. The equilibrium-type
similarity analysis for boundary layers with pressure gradient is explained in Tennekes
and Lumley (1972) and Castillo and Walker (2002). For the characterization of these
equilibrium turbulent boundary layers, it is sufficient that a single pressure gradient
parameter is constant. In the similarity solution discussed in Castillo and Walker (2002),
the pressure gradient parameter m
m=
dU
dx
d
U d x
= constant
(5.192)
is constant, where U is the velocity at the edge of the boundary layer. The self-similar
turbulent boundary layer is obtained if U x m . This relation can be inferred from
keeping u /U = constant. For nonzero values of m, integration of the m-equation
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
229
yields U 1/m . In other words, x and d/d x = constant. Castillo and Walker (2002)
use in their similarity solution for equilibrium boundary layers the relation
u 2 U 2
d
.
dx
(5.193)
If x is inserted in this relation, u U can be concluded.

This means, for the inner layer of an equilibrium turbulent boundary layer, that the
pressure gradient has hardly any effect on its similarity. Thus, we can apply the same
inflow condition as derived for the inner layer of the zero-pressure-gradient boundary
layer. In the outer layer, the experimental results by Castillo and Walker (2002) show,
for a wide range of Reynolds numbers, that the equation
U U
= f (y/)
U
(5.194)
gives similar velocity profiles in the outer part of the boundary layer. If this relation is
applied in the inlet and the rescaling cross section, we obtain
Uinouter = apg Ure (in ),
(5.195)
where apg = U in /U re . In Equations (5.196) and (5.197) below, the velocity fluctuations are assumed to obey the same relation:
+
+

(u )inner
in (yin , z, t) = apg (u )re (yin , z, t),
(5.196)

(u )outer
in (in , z, t) = apg (u )re (in , z, t).
(5.197)
The temperature fluctuations in the inner and outer layer are conjectured to satisfy
Morkovins hypothesis, which is the basis for the Bradshaw (1977) relation (5.187),
such that they can be expressed by
T in
2
= apg
T re (yin+ , z, t),
(5.198)
T in
2
= apg
T re (in , z, t).
(5.199)
inner
outer
The mean temperature distribution is determined by Equation (5.185) of Walz

(1969).
A smooth distribution of the profile over the entire boundary layer with and without
pressure gradient is obtained by forming a weighted average of the inner and outer
profile
W f (in ),
()in = ininner [1 W f (in )] + ()outer
in
(5.200)
inner
= ()
+ ( )inner
()inner
in
in
in ,
(5.201)
outer
()outer
= ()
+ ( )outer
in
in
in ,
(5.202)
where
0:9
(in)
inner layer
outer layer
buffer layer
log layer
(re)
Outflow
Velocity Profile to be
Rescaled to Inflow
Turbulent
Streamwise
Velocity Profile
L1
X =0
X>0
L2
Sponge Layer
0 521 87144 1
printer: sheridan
Figure 5.12. Sketch of the flat-plate boundary layer domain (left) and the trailing-edge domain (right). The procedure to provide the
inlet distribution to simulate trailing-edge flow is visualized.
Inflow
Rescaling
Rescaled
Turbulent Velocity
Profile
230
CUFX063/Wagner
P1: IBE
0521871441c05
November 17, 2006
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
231
is the rescaled time-averaged part of the variable and its rescaled fluctuation
in which

is . The weighting function W f () ensures a smooth transition between the inner
and outer layer:
W f () =
1
2

1+
tanh [ ( b)/(1 2b) + b]

tanh( )

(5.203)
with w = 4 and b = 0.2. The parameter w controls the width of the region, in which
the function varies from W f ( = 0) = 0 to W f ( = 1) = 1. A linear interpolation is
used to determine the corresponding locations in the inlet and rescaling planes.
The rescaling operation requires the friction velocity u and the boundary layer
thickness at the rescaling station and the inlet. At the rescaling station these quantities
are determined by the solution of the mean velocity profile, whereas at the inlet they
must be specified. It is sufficient to fix at the inlet, but u is evaluated via
)

1
wall,re
1re [2(n1)]
u ,in = u ,re
,
(5.204)
wall,in
1in
where 1 is the displacement thickness and wall is the local density at the wall; the
exponent is n = 5. Equation (5.204) can be derived using the customary power law
1/n
1/n
approximations, c f Rex , 1 /x Rex , where c f is the skin-friction coefficient.
Unlike Lund et al. (1998), who use the momentum thickness, the displacement thickness
is incorporated in the rescaling formulation to reduce the nonlinearity effects of the
momentum thickness. The integrand of the displacement thickness is a linear function
of the velocity, whereas that of the momentum thickness is a quadratic function. The
latter accumulates more inaccuracies from the spanwise average such that a slightly
less benign behavior at the boundaries results.
We return now to the inflow boundary condition of the LES domain in Figure
5.10; the boundary layer solution is coupled with the LES that provides the acoustic
sources by using the vector of solution of the boundary layer computation at time steps
t = tn = n t in the cross section x = xct with xin < xct < xre as inflow distribution at
t = tn , as shown in Figure 5.12. That is, the flow computation in the first step of the
hybrid approach is based on m + 1 large-eddy simulations performed simultaneously,
where m corresponds to the number of inlet boundaries of the acoustic source region
above a solid surface (i.e., for the trailing-edge noise problem of a flat plate sketched
in Figure 5.12, m = 2 holds).
The importance of the thorough formulation of the rescaling method is shown by
the skin-friction distribution in Figure 5.13 for the flow over an adiabatic flat plate at
M = 0.4 and Reo = 1400, where o is the momentum thickness at the inlet. The
comparison of the distributions computed using a simple temperature rescaling, which
assumes the mean and fluctuation temperature to be determined by the corresponding
velocity quantities, and the aforementioned rescaling method evidence the superiority
of the latter as far as the agreement with data from the literature (Murlis, Tsai, and
Bradshaw 1982; Falkner 1943; Fernholz 1971) is concerned.
0:9
232
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
0.0045
0.004
0.0035
0.003
cf
P1: IBE
0521871441c05
0.0025
0.002
Falkner (1943), c f=0.012/Re1/6

sponge layer
simple temperature rescaling
semiempirical law of Fernholz (1971)
(Exp.) Murlis et al. (1982)
0.0015
0.001
1400 1500 1600 1700 1800 1900 2000 2100
Re
Figure 5.13. LES of a turbulent boundary layer for Re0 = 1400 and M = 0.4. Skin-friction coefficient
c f versus Re for different rescaling formulations.
5.4.2 Silent embedded boundaries
Aeroacoustics deals with the minutest energy levels compared with the entire flow.
Even when the computational analysis is based on a smooth one-block mesh, an approximation especially designed to simulate the acoustic field has to be used to ensure
that the acoustics is not determined by the numerical procedure but by the acoustic
source mechanisms contained in the flow field. It goes without saying that particular attention has to be paid to the formulation of boundary conditions because their accurate
definition may have unexpectedly large effects on the sound field. This is, for example, known from the broadband amplification phenomenon (Bechert and Pfizenmaier
1975; Moore 1977). This extreme sensitivity of the sound field against perturbations
at the boundaries holds not only at outer but also at inner boundaries such as those
encountered in the hybrid approach in which different matching blocks covering the
LES and the acoustic domain possess artificial boundaries (Figure 5.10). Owing to the
discontinuity in the vorticity distribution, numerical noise is generated at these embedded boundaries that can, remarkably, falsify the acoustic field. This issue has been
briefly discussed by various authors (Crighton 1993; Mitchell, Lele, and Moin 1995b;
Wang and Moin 2000; Kalitzin and Wilde 2000) in the context of airframe noise and jet
noise. To remedy this problem Mitchell et al. (1995b) and Kalitzin and Wilde (2000)
implement a controlled decay of the acoustic source term at the boundary between the
flow domain and the acoustic domain.
To reach a better insight into spurious sound that is generated by the mismatch of the
vorticity distribution on the artificial boundary between the inner LES subdomain and
the outer acoustic domain and to get an idea of how to find formulations to avoid this
phenomenon, we start by considering the following formulation for the source term q m
in Equation (5.175) of the momentum equation of the acoustic perturbation equations
(5.172, 5.173, 5.174):
qm = H ( f ) L.
(5.205)
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
233
u is a perturbed form of the Lamb vector. This

The quantity L = u +
vortex term represents the major acoustic source when, for example, trailing-edge noise
is considered. The function f (x) is f (x) < 0 in the acoustic domain and f (x) > 0 in
the LES domain such that the Heaviside function is H ( f ) = 1 within the source region
and H ( f ) = 0 outside. The gradient of H is nonzero only on the boundary between
the LES and the acoustic domain on which the gradient of f is in the direction of the
normal vector on the boundary n pointing into the interior of the acoustic region f =n.
Equation (5.205) limits the vorticity distribution just to the acoustic source region. This
truncation leads to a discontinuity on the boundary, which generates a spurious velocity
field with vorticity confined to the boundary that encloses the acoustic region. In the
following the discussion of the generation of vortical perturbations due to the modified
source defined in Equation (5.205) will be based on the momentum equation (5.173).
Using H ( f ) = ( f ) f , where is the Dirac delta function, and taking the curl
of the velocity field = u with the truncated source in Equation (5.205) result in

= H ( f )L,
t
(5.206)

= [H ( f )( L) + ( f )( f L)],
t
(5.207)

= [H ( f )( L) + ( f )(n L)].
(5.208)
t
Note that only the second term on the right-hand side represents a vorticity source
on the boundary. That is, it is exactly this expression that excites spurious perturbations
on the boundary. In the following, it will be shown that at a perpendicular flow over the
boundary only rotational velocities will be generated by the modified source. For this
reason, a potential and a vector potential are defined to decompose the source qm
into an irrotational and a solenoidal part :
qm = + .
(5.209)
Taking div qm eliminates the -term and leads to the following formulation for the
irrotational velocities:
2 = H ( f ) ( L) + ( f )(n L).
(5.210)
The source term on the boundary ( f )(n L) vanishes because, for an orthogonal
flow across the boundary, the Lamb vector L and the normal vector n are perpendicular
to each other. That is, no spurious irrotational velocities are generated. Considering
curl qm removes the potential such that the right-hand side of Equation (5.208) is
completely determined by the vector potential . Hence, it can be concluded from
the analysis that, on the boundary, only spurious vortical perturbations are excited by
( f )(n L); that is,
svp
= ( f ) (n L),
t
where the subscript svp denotes spurious vortical perturbations.
(5.211)
0:9
P1: IBE
0521871441c05
234
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
Having derived an equation for spurious perturbations generated at embedded

boundaries, we turn now to the discussion of how to suppress this spurious vorticity. Two methods will be described; the first is based on the BiotSavart formula. We
consider a two-dimensional formulation. Let S be the boundary, where the source is
located, (x , y )T the source point, (x, y)T a point in the domain, and r the distance between (x , y )T and (x, y)T ; we will use the basic properties of the Dirac delta function.
Then, the term containing the perturbation velocities can be computed by the following
integral:

u svp
(n L)z
y y x x T
=
,
d S = qcorr ,
(5.212)
t
2 r
r
r
S
where the subscript z denotes the component of the source term orthogonal to the
xy plane. This correction term qcorr , when added to the acoustic source L, will
prevent the occurrence of spurious vorticity, and thus a silent artificial boundary without
any spurious sound is determined. The computation of the compensation expression
hardly impairs the efficiency of the hybrid approach because qcorr can be simultaneously
calculated with the acoustic source L.
In the second method, to avoid spurious sound generated by artificial boundaries,
a damping zone of finite thickness d is imposed at the boundary x = 0 that is,
| x | < d/2. As far as the mathematical formulation of the modified boundary source in
Equation (5.205) is concerned, this means a smooth filter function h(x); for instance,

h(x) = 12 1 + sin dx
for x d2 ,
for
d
2
for x
< x < d2 ,
(5.213)
d
2
is substituted for the Heaviside function qm = h (x) L, yielding for the spurious perturbation on the boundary
svp
t
h
(n L).
x
(5.214)
Under the assumption of passively convected vorticity in the positive x-direction

y)
that is, the vorticity is simply shifted by x = u t such that L(x, y, t) = L(x x,
the boundary equation reads
svp
t
h
q (x x)
(5.215)
with q = n L. Note that the y-dependence has been dropped for convenience. This
equation expresses that a spurious vorticity source occurs at any x within the zone
between the acoustic and the LES domain, which is the origin of a numerically induced
acoustic wave. To be able to determine the overall impact of all sources via a simple
integration over the thickness d of the damping zone, we assume compactness (i.e.,
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
235
Figure 5.14. Transfer function | F | as a function of

wave number scaled by damping zone thickness d.
d is much smaller than the acoustic wavelength). Taking into account the vanishing
gradient of h(x) outside the zone |x| d/2, we obtain the total influence via

svp, tot
h
h
dx =
q (x x)
q;
(5.216)
t
x
x
that is, the entire impact corresponds to the spurious sound source convoluted by the
gradient of the filter function h. Because the Fourier transform of the convolution is
equal to the product of the Fourier transforms of the functions, the preceding equation
reads in Fourier space
. svp
.
h
q,
(5.217)
t
x
where the equal sign has been substituted for the approximate sign. From Equation
(5.217) it is evident that in wave-number space the spurious vorticity source q is filtered
x. To analyze the effect of wave-number-dependent filtering, consider the filter
by h/
function defined in Equation (5.213). The gradient is given by
/
cos ( dx ) for | x | < d2
h
2d
(5.218)
=
F(x) =
x
0
otherwise
=
and the Fourier transform of F(x) reads

d
1
1
F()
= cos
2
2
d +d
(5.219)
with the wave number of the vortical perturbation = 2/. The graph of | F()|
in Figure 5.14 evidences a low-pass distribution (i.e., perturbations with wavelength
d or d 1 are hardly modified, whereas small disturbances satisfying < d or
d 1 are effectively filtered).
To show the efficiency of both methods to suppress spurious sound generated at the
flow-acoustic boundary, the BiotSavart and the damping zone approach, we consider
the artificial noise caused at a boundary located at x = 0 in Figure 5.15. The boundary
represents the case encountered in a hybrid computational aeroacoustics approach,
0:9
236
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
NUMERICAL METHODS
P1: IBE
0521871441c05
x
Figure 5.15. Spurious sound waves and velocity field generated at an artificial boundary at x = 0.
where the vorticity distribution of the LES subdomain, assumed to be located at x > 0,
does not exactly correspond to that of the outer domain, which is described by x < 0.
That is, the outer velocity field does not coincide with the inner velocity field.
The impact of such an unbalanced distribution is evidenced when the effect of a
convecting point source being turned on within an interval located in the immediate
vicinity of x = 0 in Figure 5.15 is simulated. In general, the point source acts like a force
in the y-direction (Figure 5.15), which produces a doublet-like velocity field, the axis of
which points normal to the mainstream direction. The convection of the source causes
the velocity field of a downstream-generated doublet to balance the velocity distribution
of the upstream doublet. In other words, the remaining velocity field describes a silent
vortex convecting with the mean velocity. If, however, the initial velocity field is not
properly prescribed (e.g., by suddenly turning on the point source), the generated vortex
is not balanced by the velocity distribution. Therefore, a steady counterrotating vortex
compared to the convecting vortex occurs at the initial source position. Because, as was
shown in Ewert and Schroder (2003), the acoustic perturbation equations do not possess
the convection property of vorticity, the vortex dynamics is completely determined by
the source term. From the KuttaZhukhovski theorem it is known that, in a mean flow
field, a steady vortex is related to a pressure field. Hence, under zero initial conditions
for the pressure field, the steady-state pressure distributions of the initial vortex are
compensated by opposite-pressure fluctuations that propagate as acoustic waves.
The behavior of the convecting point source is visualized in Figure 5.15. A vortex
convects to the right, and a steady counterclockwise rotating vortex, whose origin is at
x = 0, y = 0, produces propagating acoustic waves leaving the computational domain.
0:9
P1: IBE
0521871441c05
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Figure 5.16. Pressure distribution on y = 35 in Figure 5.15 for several thickness values d and
BiotSavarts law (denoted as compensation).
The impact of the suppression concept is evidenced in Figure 5.16 by the distribution of
the spurious pressure on the line y = 35. Without any compensation, a strong overshoot
occurs that is drastically reduced when the filter function (5.213) is turned on within the
damping zone | x | d/2. The larger the thickness of this zone (i.e., a wider range of
wave numbers is filtered), the smaller the amplitudes of the spurious pressure waves. In
fact, at d = 20 the suppression is as effective as it is when the BiotSavart law (5.212)
is used.
It can be concluded from the findings of the numerical analysis of the convecting vortex problem that both methods can be applied to avoid spurious sound waves
induced by boundaries embedded in the overall computational domain. However, the
more straightforward implementation makes the damping zone approach the method
of choice when the acoustic field of technically relevant problems is to be computed.
237
0:9
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Applications and Results of Large-Eddy

Simulations for Acoustics
6.1 Plane and axisymmetric mixing layers

Christophe Bogey and Christophe Bailly
Velocity-gradient regions between two streams are found in numerous flows such as
plane or axisymmetric jets. These simple flow configurations, usually referred to as
shear layers or mixing layers, have been extensively investigated as reported in the review of Ho and Huerre (1984). These studies have allowed better understanding of the
transition to turbulence occurring in initially laminar or transitional shear layers thanks
notably to the development of instability theories (Michalke 1984) and to the observations of coherent structures. Among the experiments supporting the latter observations,
the famous one conducted by Brown and Roshko (1974) shows clearly that large-scale
coherent structures are intrinsic features of mixing layers even at high Reynolds numbers. These structures, through their interactions such as vortex mergings, have been
recognized to be appreciably responsible for the spreading of the shear layers or for
noise generation, as stated by Winant and Browand (1974).
From the preceding considerations, the large-eddy simulation (LES) approach appears especially well suited to mixing-layer computations because with LES all the
scales larger than the grid size, and consequently a large part of the coherent structures, are calculated. The earlier large-eddy simulations of practical flows have thus
often involved mixing layers. They have permitted testing of the LES methodology for
simple transitional flows. For instance, Vreman et al. (1997) investigated the effects
of several subgrid modelings for a temporal mixing layer, and Doris, Tenaud, and Ta
Phuoc (2000) studied the influence of different numerical schemes and inlet conditions
for a spatially developing mixing layer. LES has also been applied with the aim of improving the description of the turbulent structures in mixing layers. An example is given
by Comte, Silvestrini, and Begou (1998), who, using LES, investigated the formation of
streamwise vortices in mixing layers developing spatially. As a further step, LES seems
appropriate for determining the noise radiated by these large structures, which are likely
238
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.1 PLANE AND AXISYMMETRIC MIXING LAYERS
to be dominant in transitional shear layers (Winant and Browand 1974), although their
contribution in fully turbulent shear layers where no pairing occurs is questionable, as
claimed, for instance, by Bridges and Hussain (1987).
The first simulations of free-shear flows to compute their radiated noise from the
unsteady compressible flow equations were, however, performed using direct numerical
simulation (DNS). These simulations were DNS of bidimensional and axisymmetric
mixing layers and were carried out in the Stanford group by Colonius et al. (1997) and
by Mitchell et al. (1999), respectively. They showed the feasibility of the direct noise
computation (DNC) by directly calculating the noise generated by vortex pairings
in mixing layers. Afterward, DNC has been naturally applied using DNS for threedimensional mixing layers. A clear illustration is provided by Fortune, Lamballais, and
Gervais (2001), who studied by DNS the effects of temperature on the sound generated
by temporally evolving mixing layers. Because DNS is limited to low-Reynolds-number
flows, LES must be used to perform DNC at high Reynolds numbers (i.e., for flows
with realistic turbulence).
In what follows, some recent works of the authors are presented to illustrate the
use of LES to investigate the sound generated in mixing layers. Three configurations
are shown: the first one involves a bidimensional mixing layer, and the two others
involve circular subsonic jets at different Reynolds numbers with initially transional
shear layers. In the three cases, the noise radiated by the turbulence is obtained directly
from LES, and the sound-generation mechanisms are discussed.
6.1.1 Plane mixing layer
As in DNS, with the work of Colonius et al. (1997), the first DNC using LES has often
been performed for simple bidimensional mixing layers. The first illustration in this
chapter therefore deals with the computation of the sound generated by vortex pairing
in a 2D mixing layer using large-scale simulation and the Smagorinsky model for the
subgrid modeling. All the details of the simulation can be found in Bogey et al. (2000b).
The inflow of the mixing layer is defined by an hyperbolic-tangent velocity profile
as

2y
,
(6.1)
u 1 (y) = Um 1 + Ru tanh
(0)
where Um = 0.3 c0 and Ru = U/(2Um ) with U = U2 U1 = 0.48 c0 0.12 c0
(c0 is the mean speed of sound) and (0) is the initial vorticity thickness of the shear
layer. The Reynolds number based on the initial vorticity thickness is equal to Re =
(0)U/ = 12, 800 ( is the molecular viscosity).
To control the first vortex pairing, the mixing layer is classically forced at its
fundamental frequency f 0 and its subharmonic f 0 /2. The fundamental frequency
f 0 0.132 Um / (0) corresponds to the most amplified instability in the initial shear
layer predicted by the linear instability theory (Michalke 1984). This kind of forcing
239
13:36
0 521 87144 1
printer: sheridan
November 20, 2006
APPLICATIONS AND RESULTS OF LARGE-EDDY SIMULATIONS FOR ACOUSTICS
240
CUFX063/Wagner
P1: IBE
0521871441c06
Figure 6.1. Simulation of a 2D mixing layer. (a) Snapshot of the dilatation field = .u on the whole
calculation domain, levels in s1 . (b) View of the pairing zone with the vorticity field in the mixing layer
and the dilatation field outside.
allows fixing the location of the vortex pairings around x = 70 (0) in the downstream
direction and their frequency as f p = f 0 /2.
The dilatation field = .u obtained directly from the simulation on the whole
computational domain is displayed in Figure 6.1(a). Dilatation is indeed related to the
acoustic pressure in the uniform streams by

p
p
1
+ Ui
with i [1, 2].
(6.2)
= .u =
x
0 c02 t
Wave fronts originating clearly from the zone of vortex pairings are observed.
As expected, the acoustic wavelength corresponds exactly to the pairing frequency.
This wavelength appears to be modulated by the effects of the uniform streams on
sound propagation. The wave fronts have typical oval forms especially in the upper
high-velocity stream. The noise directivities in the high- and low-velocity streams are
also appreciably altered by the mean flow effects on sound waves. Note that these
interactions between the flow and the acoustic waves have been investigated using
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
two-step approaches with the DNC solution as reference solution. Both Lighthills
analogy in Bogey, Gloerfelt, and Bailly (2003) and the approach based on the linearized
Euler equations in Bogey, Bailly, and Juve (2002) have thus been applied.
Finally, to study the sound source, a view of the pairing zone is presented in Figure 6.1(b) with the vorticity field in the shear layer and the dilatation field outside.
A double spiral structure with four lobes corresponding to a rotating quadrupole is
observed. This kind of acoustic source was described for a corotating vortex pair in the
analytical works of Powell (1964) and later computed through DNS (Mitchell, Lele,
and Moin 1995a). The present simulation has thus allowed to reveal the presence of
such a source in a mixing layer.
6.1.2 Axisymmetric mixing layers jets
This section reports some recent investigations of the noise generated in the axisymmetric shear layers of circular jets computed by LES. Links between the shear layer
turbulence and the noise radiated in the sideline direction are shown for two jets at
different Reynolds numbers.
6.1.2.1 Moderate-Reynolds-number jet
The first three-dimensional flow simulated by LES in our group to compute its radiated
noise was a circular jet at Mach number M = u j /c0 = 0.9 and Reynolds number Re D =
u j D/ = 6.5 104 , where u j is the jet inflow velocity and D its initial diameter. In
this simulation, the inflow shear layer was forced with random velocity disturbances to
seed the turbulence, and the Smagorinsky model was used. Please refer to Bogey, Bailly,
and Juve (2003) for all the simulation parameters. The flow and acoustic results have
been successfully compared with corresponding measurements to show the feasibility
of directly computing jet noise using LES. Connections between the dynamics of the
turbulent structures in the jet and the radiated waves have then been tracked, as was
done in experiments by Hileman and Samimy (2001) for instance. In this way, a strong
link has been exhibited between the intrusion of turbulent structures in the jet core and
the noise radiated in the downstream direction.
To detect possible noise-generation mechanisms occurring in the shear layer, we
have recorded the pressure at a point defining a wide angle from the flow direction.
Experiments such as those of Zaman (1986) have indeed shown that jet noise in the
sideline direction can be related to the turbulence in the shear layer. The observation
point has been chosen to be at x = 16r0 , y = 8r0 , and z = 0. This point is at 9.4r0
from the end of the potential core at x = 11r0 and y = z = 0, where the sound sources
are classically assumed to be located, at an angle of about 60o with respect to the
downstream direction. The signal of pressure at this point is displayed in Figure 6.2
during a nondimensional time period T = T u j /D = 40.
An attempt has then been made to connect the peak of highest amplitude observed
for t = 6.9 with the interactions between the turbulent structures in the jet. The sound
wave corresponding to the pressure peak is shown in Figure 6.3. It is located at about
241
13:36
P1: IBE
0521871441c06
242
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.2. LES of an ReD = 6.5 104 subsonic jet. Time

evolution of the fluctuating pressure p in Pa as a function of
t = tu j /D, at x = 16r 0 , y = 8r 0 , and z = 0. Reprinted from
Bogey, Bailly, and Juve (2003) with the kind permission of
Springer Science and Business Media.
x = 18r0 and y = 9r0 . Its directivity is marked for wide angles between 30o and 70o ,
which supports that it be generated in the shear layer.
To find the turbulent event likely to generate the present acoustic wave, we evaluated
the time delay between its emission and its arrival at the observation point by assuming a propagation at the mean sound speed. A delay of t = 4.2 has been obtained,
which implies that the generation occurs for t 2.7. The vorticity in the plane z = 0
has therefore been depicted at t = 2.2 and t = 3.5 in Figure 6.4. An isolated large
structure is observed in the upper shear layer in Figure 6.4(a) at x 11r0 and y r0 .
It interacts with other structures to merge with the vortical field located downstream
in Figure 6.4(b). This kind of vortex pairing is conjectured to be responsible for the
acoustic wave of Figure 6.3. More generally, this observation supports that the interactions between coherent structures in the shear layer contribute to jet noise at moderate
Reynolds numbers.
6.1.2.2 High-Reynolds-number jet
Subsequent to the simulation reported just above, the LES of a circular jet at the same
Mach number M = 0.9 but at higher Reynolds number Re D = 4 105 was carried out.
Figure 6.3. LES of a ReD = 6.5 104 subsonic jet. Snapshot of the vorticity norm in
the flow field and of the fluctuating pressure outside in the plane z = 0 at t = 7.5.
Reprinted from Bogey, Bailly, and Juve
(2003) with the kind permission of Springer
Science and Business Media.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
243
3
6
(a)
y/r0
y/r0
12
3
6
15
(b)
x/r0
12
15
x/r0
Figure 6.4. LES of a ReD = 6.5 104 subsonic jet. Snapshots of the vorticity norm in the plane z = 0
at times (a) t = 2.2, (b) t = 3.5. Reprinted from Bogey, Bailly, and Juve (2003) with kind permission
of Springer Science and Business Media.
This simulation, referred to as LESac, has been performed using a selective filtering
alone for modeling the dissipative effects of the unresolved scales. Flow and sound
features have been shown in Bogey and Bailly (2006) to correspond well to what is
observed for high-Reynolds-number jets particularly regarding the changes in the
acoustic field according to the observation angle. Further, large-eddy simulations have
then permitted investigations of the effects of the inflow conditions and of the subgrid
modelings on the jet properties in Bogey and Bailly (2005b) and Bogey and Bailly
(2005a), respectively. In the latter study, the results with the dynamic Smagorinsky
model and with the filtering alone designed not to affect the resolved scales (Bogey and
Bailly 2004) have been discussed.
A view of the vorticity field in the jet and of the pressure field outside is represented
in Figure 6.5 for the simulation LESac. Two kinds of acoustic waves are visible: a lowfrequency wave with high amplitude located near x = 20r0 and y = 6r0 predominant
in the downstream direction, and waves characterized by higher frequencies, for wider
angles, appearing to come mainly from the turbulent axisymmetric shear layer around
x = 8r0 .
16
12
8
y/rc
Figure 6.5. LES of a ReD = 4 105 subsonic jet. Snapshot of the vorticity norm in
the flow field and of the fluctuating pressure
outside in the plane z = 0.
20
12
x/rc
16
24
13:36
0 521 87144 1
printer: sheridan
November 20, 2006
Table 6.1. LES of a Re D = 4 105 subsonic jet. Sideline

-maxima in the shear layer for the
sound levels and vrms
different simulations
OASPL in dB at
LESshear
LESampl
LESac
LESmode
(x = 11r0 , r = 15r0 )
max
125.2
124.7
124.1
121.8
0.196
0.190
0.186
0.168
(r =r )
vrms
0
uj
To exhibit links between the shear-layer turbulence and the sideline noise, we now
report some flow and sound results provided by four simulations using various inflow
conditions for the jet in Bogey and Bailly (2005b). The sound spectra at x = 11r0
and r = 15r0 obtained from the different large-eddy simulations are thus provided in
Figure 6.6(a). Following Zaman (1986), who demonstrated that the noise sources can
be approximated by the maxima of turbulence intensities, the profiles at r = r0 of the
root-mean-square (rms) values of the radial fluctuating velocity v are also plotted in
Figure 6.6(b). From the four simulations, it appears clearly that the sideline noise levels

peak values.
are arranged in the same order as the vrms
This correspondence is emphasized in Table 6.1: the amplitudes of the sideline

peak values in the shear layer. The present
sound levels vary accurately with the vrms
LES results thus provide a new indication that sideline jet noise is predominantly
generated by the turbulence developing in the axisymmetric shear layer. Further studies
are, however, required to clearly identify the generation mechanisms involved at high
Reynolds numbers.
6.1.3 Concluding remarks for mixing-layer simulations
Direct noise computation is an outstanding approach for studying the sound generated
by free-shear flows because it permits us to correlate turbulence events with the sound
1500
0.24
1200
0.2
0.16
900
vrms/uj
244
CUFX063/Wagner
SPL(Pa /St)
P1: IBE
0521871441c06
600
0.08
300
0
0.12
0.04
0
0.3
0.6
0.9
St
1.2
1.5
6
x/r0
12
Figure 6.6. LES of a ReD = 4 105 subsonic jet. (a) Pressure spectra at x = 11r 0 , r = 15r 0 ; (b) Profiles
/u in the shear layer for r =r . Different simulations: LESac (), LESampl (.......), LESshear
of vrms
j
0
(- - -), LESmode (-. -. -. ).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.2 FAR-FIELD JET ACOUSTICS
far field. In this context, LES seems the better tool to clarify Reynolds-number effects
especially for subsonic jet noise. The dependence on Reynolds number is particularly
expected for the noise-generation mechanisms observed in mixing layers whose initial
state, transitional or turbulent, is fixed by the Reynolds-number value.
In this section, only the noise radiated in the sideline direction from the jet axis has
been discussed. It has been shown that its generation likely takes place in the developing
mixing layer. Note also that a similar analysis has been conducted for the downstream
noise. This jet noise component was correlated in Bogey, Bailly, and Juve (2003) with
the periodic intrusion of vortical structures into the potential core. Its level was recently
demonstrated in Bogey and Bailly (2005b) to vary as the maximum of turbulence
intensities on the jet centerline.
6.2 Far-field jet acoustics

Daniel J. Bodony and Sanjiva K. Lele
6.2.1 Introduction to jet acoustics
It can be said that most of the current research on the generation of sound by turbulent
flows has had its origin in the prediction of jet noise (in particular the noise from jet
exhaust plumes) by Lighthill (1952, 1954) in his two papers On Sound Generated
Aerodynamically. Lighthills U 8j scaling of the radiated noise intensity, where U j
is the jet exit velocity, gave engine manufacturers an initial tool by which to design
quieter engines and led, in part, to the development of internal and external mixers and
the turbofan engine. In the early 1970s, it became clear that a more detailed knowledge of the noise sources and the resulting radiation field would be needed to further
reduce jet noise. Lilleys (1974) inhomogeneous convective wave equation, based on
the work of Phillips (1960), gave further insight into the noise sources and attempted to
separate out the sound-generation and sound-propagation processes that are present in
the jet.
The numerical prediction of jet noise for design applications came with the introduction of the MGB jet prediction tool (Balsa et al. 1978). The MGB code uses a
Reynolds-averaged NavierStokes (RANS) mean flow solution to define local length
scale, time scale, and source strength parameters for a semiempirical source model.
The source model is based on a simplified Lighthill quadrupole source term with an
It should, however, be noted that the importance of reducing jet noise for commercial air transport use
had been recognized prior to Lighthills work in the experiments of Morley (1939), Westley and Lilley
(1952), and others.
It is beyond the scope of this section to discuss the acoustic analogies of Lighthill (1952, 1954),
Lilley (1974), Mohring (1999), Fedorchenko (2000), and others. Many of these analogies have been
formulated to directly address the prediction of jet noise.
The name MGB stems from its primary authors last names: ManiGliebeBalsa.
245
13:36
P1: IBE
0521871441c06
246
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
approximate high-frequency solution of Lilleys equation for propagation to the far

field. The noise source modeling of the original MGB formulation continues to be
refined (e.g., see Khavaran, Bridges, and Freund 2002; Khavaran and Bridges 2004
and the references within), and the more recent versions of the code are referred to as
MGBK.
Tam and Auriault (1998, 1999) have proposed a RANS-based prediction scheme for
the fine-scale mixing noise component of jet noise. In their formulation, a modified
k turbulence model provides parameters for a semiempirically based spacetime
correlation function of the fluctuation turbulence kinetic energy which, in contrast to
the source term used in the MGBK code, is not based on Lilleys (1974) form of
Lighthills (1952, 1954) analogy. Instead, in the spirit of kinetic theory, they postulate a
relationship between the turbulence kinetic energy and fluctuating pressure. An adjoint
field is used to project the near-field source onto the far field. Tams reformulation,
like MGBK, retains the strong dependence on a RANS solution with a calibrated,
semiempirical noise-source model.
The fundamental limitation of the aforementioned RANS-based approaches is that
the flow unsteadiness is modeled semiempirically. Observations of the two-point, timeseparated statistical correlation u i (x, t)u j (x + , t + ) as a function of and suggest using the functional form of the noise model (Goldstein 1976). However, statistics
of the noise-source terms themselves have not yet been measured in the laboratory
owing primarily to the difficulty in experimentally measuring the necessary products.
Freund (2003) has recently detailed statistics of the Lighthill source term using DNS
data in a low-Reynolds-number jet.
In a real flow it is the temporal evolution of a particular turbulent eddy within its
immediate environment that generates sound, which suggests that source models solely
using the RANS mean flow to provide values for model constants will be limited in their
ability to predict the radiated noise accurately. These methods perform reasonably well
for generic axisymmetric or rectangular nozzles, but they are often unable (Lord and
Feng 2000) to account for subtle nozzle design changes such as the addition of nozzle
chevrons or tabs. Moreover, they have difficulties at the low- and high-frequency ends
of the acoustic spectrum (NASA Glenn Research Center 2001).
It is believed that noise-prediction tools that do not model the flow unsteadiness and
involve fewer a priori assumptions will better be able to represent the noise sources.
In this regard, LES of jet flows is seen to represent the next step in jet noise predictions because some of the unsteadiness naturally present in the exhaust plume is
retained.
The remainder of this section discusses the use of LES in jet noise prediction with an
emphasis on the far-field sound. We further confine our discussion to those simulations
and associated techniques that directly compute the turbulence-generated sound. That
is, the simulations of incompressible jets from which the sound is extracted via a sound
source assumption (usually that of Lighthill 1952, 1954) will not be addressed here,
but the reader is encouraged to visit the work of Zhao, Frankel, and Mongeau (2001b),
Boersma (2002, 2003), or Rembold and Kleiser (2003), for example, for analogy-based
jet noise predictions.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.7. Schematic of a turbulent jet issuing into a still fluid. The jet near field, where the LES
equations are solved, is situated within an exterior domain into which the sound is propagated. Not to
scale.
6.2.2 Numerics of jet simulations
A schematic of the computational domain used in the simulations of Gamet and

Estivalezes (1998), Zhao et al. (2001a), Bodony and Lele (2002a, 2003b, 2004), Uzun,
Blaisdell, and Lyrintzis (2002, 2003), and Uzun (2003) into a quiescent environment is
shown in Figure 6.7 along with the defining coordinate system. In this setup, two distinct
regions are used: the first is the LES domain, where the filtered equations of motion
are solved using a subgrid-scale flux closure. It is in this portion of the domain that the
inflow boundary conditions are specified, the jet develops, and, ultimately, the sound is
generated. Outside this area is the wave equation domain, where, with approximation,
the sound is propagated to locations far removed from the jet plume. There have also
been several studies (Bogey, Bailly, and Juve 2000a; Bogey and Bailly 2005a, 2005b,
2006; Morris et al. 2002) that retain a portion of the far acoustic field in the near-field
solution in lieu of a separate wave-propagation zone.
6.2.2.1 Near-field discretization
Many of the previous discussions in this book on numerical discretization apply equally
well to the simulation of jet flows and to the radiated noise. One point, however, that
should be emphasized is the correspondence between the dispersion relation of the
247
13:36
P1: IBE
0521871441c06
248
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
numerical scheme (including the temporal discretization) and that of the radiated sound.
Because air is an approximately nondispersive medium for low-amplitude sound waves
at the conditions of interest in jet noise, the numerical method must reflect this by
not introducing any artificial dispersion. Often it is necessary to correctly capture the
sound propagation over a distance of many wavelengths for a range of frequencies so
that numerics-induced dispersion or dissipation will be exaggerated.
The most popular numerical methods of this sort are the dispersion-relationpreserving (DRP) scheme of Tam and Webb (1993) (see also Bogey and Bailly 2004)
and the combination of a Pade derivative scheme (Lele 1992) and a high-order time
advancement scheme, usually a RungeKutta variant. The mixing layer calculations of
Lui and Lele (2001, 2002a, 2002b, 2003) used an optimized compact scheme coupled
with the fourth-order, low-dispersion RungeKutta scheme of Hu et al. (1996) (using
the low-storage version of Stanescu and Habashi 1998) and found the method to be
highly accurate and numerically efficient.
The resolving characteristics of finite difference methods have not been completely
translated into finite element or finite volume schemes. Such schemes are quite convenient when complex geometries are involved, as is true for most engineering applications, and their efficient use for aeroacoustic problems is needed. Initial attempts
at using finite elements to include the nozzle in jet LES calculations by Al-Qadi
and Scott (2002), DeBonis and Scott (2002), and Jansen, Maeder, and Reba (2002),
for example, show promise but, to date, have had mixed success. The DRP scheme
has also been applied to jet flows with the nozzle present by Dong and Mankbadi
(1999).
6.2.2.2 Boundary conditions
For simulations of jet exhaust noise and other flow-generated noise situations the boundary conditions are crucial in establishing the environment into which the flow develops.
Aside from the need for nonreflecting boundary treatments, jet noise predictions have
been found to be sensitive to the inlet conditions associated with the quasi-laminar
annular shear layer (Morris et al. 2002; Bodony and Lele 2002a; Bogey, Bailly, and
Juve 2003).
In real jet flows, the annular shear layers exiting the engine nozzle are extremely
thin (0 /D j O(103 ), where 0 is the shear-layer momentum thickness and D j the
jet diameter) and turbulent. In an LES of the corresponding flow, the shear layers are
thicker (0 /D j O(102 )) with imposed quasi-laminar disturbances both due to the
limited available resolution. (For reference, the DNS of a low-Reynolds-number jet by
Freund (2001) had an initial momentum thickness of 0 /D j = 0.02.) The thicker initial
shear layer implies that, for the same velocity difference across the shear layer, the rate
of energy extraction from the mean to the shear-layer disturbances is lower than in the
real flow. This appears to increase the sensitivity of the jet-radiated sound to the inlet
conditions. In particular, the degree of azimuthal correlation of the inlet disturbances is
found to substantially influence the radiated noise levels with more strongly correlated
disturbances leading to increased radiated sound.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
13:36
249
90
60
Figure 6.8. OASPL directivity at a distance of 30D j from

the unheated, Mach 0.9 jet exit. Nearly axisymmetric disturbances, ; azimuthally decorrelated disturbances, ;
data of Stromberg McLaughlin, and Troutt (1980), ; MolloChristensen, Kolpin, and Martucelli (1964), ; Freund (2001),
; Bogey, Bailly, and Juve (2000a), .
30
0
90
100
110
120
This relationship between the azimuthal correlation of inlet disturbances and the
radiated sound was established experimentally by Zaman (1985), who found that the
noise from axisymmetric disturbances was approximately 5 dB higher than from lesscorrelated disturbances. Similar changes in sound levels have been found in the available
LES studies, and Figure 6.8 exhibits the change found by Bodony and Lele (2002a). Note
that the inlet disturbances used by Bodony and Lele (2002a) in the nearly axisymmetric
disturbances (solid line) case of Figure 6.8 were identical in form to those used in the
DNS of Freund (2001). Bogey, Bailly, and Juve (2003) further studied the effect of
the inlet conditions on the radiated noise by varying the initial shear-layer thickness,
the azimuthal mode composition, and overall forcing amplitude of the quasi-laminar
disturbances. They also found that more strongly correlated inlet disturbances increase
the jet noise. In addition they noted that a thinner initial shear-layer thickness reduces
the downstream-radiated noise while increasing the noise levels at observers situated
near 90 from the jet axis.
The introduction of disturbances into the annular shear layer must be such that, in
addition to inducing natural turbulence development, no spurious noise is generated in
the vicinity of the inlet plane. That boundary noise generation can be a problem is easily
seen in the simulations of Constantinescu and Lele (2001). There the inlet conditions
radiated sound at levels similar to the jet itself (see their Figure 2) and have marked
peaks in the sound spectrum.
For the avoidance of such difficulties, two approaches have been used successfully
to date. The first method attempts to construct disturbance profiles that are divergence
free, or nearly so. Bogey, Bailly, and Juve (2003) and Uzun (2003), for example, use
disturbances of the form
u x = f 1 ( )g1 (t)h 1 (x, r ),
(6.3)
u r = f 2 ( )g1 (t)h 2 (x, r ),
(6.4)
u = 0,
(6.5)
130
P1: IBE
0521871441c06
250
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where the functions h 1 and h 2 are chosen such that u = 0 (see also Glaze and
Frankel 2003). This technique works well for cold low- and moderate-Mach-number jets
where the density is nearly uniform. One could generalize this somewhat by specifying
that, for example,
u x = f 1 ( )g1 (t)h 1 (x)i 1 (r )
(6.6)
with u r , u used similarly and requiring that the functions { f k , h k , i k }3k=1 be such that u
is divergence free. This yields a set of differential equations relating the f, h, i that may
be solved with additional constraints. Similarly, one may also vary the inlet momentum
thickness 0 as a function of time and azimuthal location; that is,
u x
Uj
0 (, t) =
1
4B
= (1 tanh [0 (, t) (r/ro ro /r )]) ,

St U
+ 1m,n Amn cos Dmnj t j + mn cos (m + mn ) ,
(6.7)
(6.8)
where B is the time-averaged momentum thickness and the functions , , and St

specify the fluctuation randomness. Used originally by Freund (2001), this method has
been used subsequently by Zhao et al. (2001a) and Bodony and Lele (2002a, 2003b,
2004).
The second approach uses the nonradiating eigenfunction solutions to the linearized
spatial instability problem associated with the inlet mean profile as the disturbance profiles. Assuming a modal decomposition of a disturbance flow variable q (x, r, , t) as
) exp{i(t kx n)}, for integer n, we find that the pressure disturbance ampliq(r
tude p for an axisymmetric mean flow is governed by the cylindrical Rayleigh equation

1 d
2k
( ku)2
n2
1
du d
d2
2
+
p
+
k
p = 0
p
+
dr 2
r
dr
( ku) dr dr
r2
a2
(6.9)
subject to the boundary conditions
<
| p|
p 0
as r 0,
(6.10)
as r .
(6.11)
The spatial stability problem is characterized by fixing to be real and solving for
the complex-valued eigenvalue k. The remaining flow variables are given as linear
combinations of p and its radial derivative.
This method is guaranteed to yield nonradiating disturbances provided the phase
speed /kr , with k = kr + iki , of each (, n, k) triplet is subsonic relative to the
ambient fluid. This condition is satisfied in unheated jets with M j < 1.5 and is a function
of the temperature ratio T j /T for heated jets. When the disturbance phase speed is
supersonic, the instability waves radiate directly to the far field (Tam and Morris 1980).
Application of this method to rectangular jets has been discussed by Rembold and
Kleiser (2003) and by Zhao et al. (2001a) and Bodony and Lele (2003b, 2004) for
circular jets.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.2.2.3 Far-field sound extrapolation
In experimental investigations of jet noise, sound measurements are generally taken in

the acoustic far field, usually at distances of 30D j and beyond from the nozzle exit.
This region, where the nearly incompressible pressure fluctuations directly associated
with the turbulence of the jet are negligible, is generally uniform; thus, the propagation
of sound is greatly simplified relative to the environment nearer to the jet. It is therefore not necessary to include the acoustic far field in the LES but instead to restrict the
computational domain laterally to a few jet diameters away from the jet centerline. Continuation of the sound field to the far field is done either analytically or with a simplified
numerical method. One such technique, using a Kirchhoff surface, is described below;
alternatives may be found in the literature. For the Ffowcs WilliamsHawking surface,
see Brentner and Farassat (1998) and the references therein as well as Uzun (2003); for
a time-marching numerical solution of the wave equation, see Freund (2000).
The Kirchhoff surface method is based on an integral solution of the wave equation

2 2 2
(6.12)
a /t 2 / x j x j ( p p) = 0
for an inviscid, quiescent medium with constant sound speed a and has been reviewed
by Lyrintzis (1994). The presence of a uniformly moving fluid outside the jet, as there
would be for a jet in flight, can be accommodated by a change of coordinates. This
simple linear model of sound propagation is valid for noise of most high-Reynoldsnumber applications and for the frequency ranges and propagation distances of interest.
Extremely loud jets (with sound-pressure intensity levels in excess of 150 dB) or jets
with crackle (Ffowcs Williams, Simson, and Virchis 1975) may require a more complicated description of the propagation. The application to jet noise is straightforward:
pressure fluctuations about the average p = ( p p) are collected on a cylindrical
shell of radius Rs surrounding the jet. The shell is chosen to be sufficiently far away
from the hydrodynamic region of the jet so that the sound propagation is sufficiently
described by Equation (6.12) and extends from the computational inlet to the exit. The
ends of the shell, which are perpendicular to the jet axis, are generally not included
because they do not lie in linear regions of the flow; their absence may be accounted for
(Freund, Lele, and Moin 1996) but is often ignored with the restriction that the far-field
sound predictions are valid over a limited range of angles from the jet axis.
Using the pressure fluctuations specified on the shell, we can solve the wave equation (6.12) in the partially transformed form

2
n2
d2
1 d
2
+ 2 k 2 p = 0,
+r
(6.13)
dr 2
dr
a
r
which may also be obtained directly from Equation (6.9). Equation (6.13) is Bessels
; k, n, ) =
equation and is subject to the boundary conditions that, for r = Rs , p(r
s ; k, n, ), and that as r the solutions represent outgoing traveling waves.
p(R
2
k 2 > 0 radiate as sound and contribute to the
Only those waves that satisfy 2 /a
far field. Near r = Rs , however, the pressure field may be dominated by the nonradiating
components. The far-field pressure fluctuations are found by integrating Equation (6.13)
251
13:36
P1: IBE
0521871441c06
252
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
to the desired value of r and inverse transforming. The overall solution at a given field
point (x, r, , t) is then seen to be
1
( p p)(x, r, , t) =
(2 )2
+

N
/21
s ; k, n, )
p(R
n=N /2
2 k2
r 2 /a

ei(tkxn) dkd, (6.14)
(m)
2
2
2
Hn
Rs /a k
(m)
Hn
(m)
where Hn is the nth-order Hankel function with m = 1 for > 0 and m = 2 for
< 0. The radial location of the Kirchhoff surface is important to ensure that the
nonlinearity, nonuniformities, or both of the jet near field are sufficiently low to justify
the assumptions made in the methods derivation. If the surface is placed too close to
the near field, the sound levels can be overpredicted by 57 dB (Brentner and Farassat
1998; Uzun 2003).
6.2.3 Results for jet simulations
When examined in detail, the structure of a turbulent jet is complex with a wide range
of spatial and temporal scales involved in the dynamics. The far-field noise is generally
broadband with spectral characteristics that depend on the angle from the jet axis and
jet operating conditions. Experimental investigations have shown that higher-frequency
sources are typically located closer to the nozzle exit while lower-frequency sources
can be found farther downstream (NASA Glenn Research Center 2001). Moreover, the
turbulent structure and corresponding sound sources are functions of the jet exit Mach
number M j (Freund, Lele, and Moin 2000) and jet temperature ratio T j /T . To date,
most of the LES studies of far-field noise have been of moderate-to-high subsonic jets
owing, in part, to the existence of quality experimental data and the desire to avoid
cases with shocks present. Notable exceptions to this are the incompressible jet studies
of Boersma (2002, 2003) and the perfectly expanded Mach 2.1 unheated jet of Morris
et al. (2002). Because the simulation of moderate-to-high subsonic jets has received
the most attention, it will be the main topic in the remainder of this discussion. A brief
survey of the relevant studies will precede a more detailed review of quantitative results.
6.2.3.1 Brief literature survey
Initial attempts at using LES to study compressible jets began in the late 1990s. Estivalezes and Gamet (1996) and Gamet and Estivalezes (1998) used MacCormacks
scheme to investigate the near-jet region of a hot, Mach 2 jet in two and three dimensions with a Kirchhoff surface to extract the far-field sound. Choi et al. (1999) compared
MGBK predictions with LES predictions of an off-design supersonic nozzle using a
Kirchhoff surface. This was followed by the simulation of Dong and Mankbadi (1999)
of nozzle-ejector mixer operated at a jet Mach number of 1.5 with Mach 0.4 coflow.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Boersma and Lele (1999) performed an exploratory study investigating the suitability
of LES for the jet noise problem, focusing on the subgrid-scale modeling. They found
that the compressible subgrid-scale model of Moin et al. (1991) performed reasonably
well in capturing the mean and rms fields but noted some dependence on the boundary
conditions. Constantinescu and Lele (2001) furthered the work of Boersma and Lele
(1999) by applying a highly accurate centerline treatment (Constantinescu and Lele
2002) and focusing on the boundary conditions for a Mach 0.9 jet. The resulting nearfield data compared favorably with the experimental data, and they found that using
non-Favre-weighted variables improved the robustness of the solver. It is difficult to extract far-field acoustic data from their investigation because their inlet condition appears
to be strongly radiating, corrupting the sound field but apparently having little effect
on the hydrodynamic field. Bodony and Lele (2002a, 2003b, 2004) refined the work of
Boersma and Lele (1999) and Constantinescu and Lele (2001, 2002) by implementing
improved boundary conditions, allowing the necessary entrainment through the lateral
boundaries and a physically realizable inlet disturbance specification.
The Mach 0.9 jet of Bogey et al. (2000a), Bogey, Bailly, and Juve (2003), and
Bogey and Bailly (2005a, 2005b, 2006), using a rectangular grid solver and various
subgrid-scale closures, compared favorably with the available turbulence experimental data in the self-similar region of the jet plume. In addition, their computational
domain included a portion of the far field in which they collected pressure fluctuation
data. The spectra of pressure fluctuations had anomalous behavior at lower frequencies,
which they attributed to the boundary conditions. When the low-frequency portion of
the acoustic spectra was neglected, Bogey et al. (2000a) found overall sound-pressure
levels (OASPL) and directivity patterns to be in reasonable agreement with the experimental data. Zhao et al. (2001a) found similar acoustic field characteristics using a
dynamic subgrid-scale model and Kirchhoff surface. More recently, Bogey, Bailly, and
Juve (2003) refined their initial results by, through boundary condition improvement,
removing the excess low-frequency fluctuations in their acoustic spectra and performing
a parameter study on the inlet conditions. Uzun, Blaisdell, and Lyrintzis (2002, 2003),
also using a rectangular grid solver, computed the sound of a series of isothermal jets at
varying Reynolds numbers with particular attention to the location of the Kirchhoff
and Ffowcs WilliamsHawkings surfaces. Bodony and Lele (2004) have examined the
ability of LES to capture heating effects in moderate-to-high subsonic jets.
A unique study by Andersson (2003) of a Mach 0.75 jet (heated and unheated)
was the computational counterpart to a series of measurements at the same conditions
using the nozzle geometry of the experiment. When a finite volume method is used the
unheated and heated jets slightly underpredict the streamwise rms fluctuations along
the centerline with a peak location farther upstream than measured experimentally.
OASPL directivity comparisons show 3 dB agreement over a wide range of angles
for the unheated and hot jets.
In working toward the development of a general jet noise prediction methodology
Shur, Spalart, and Strelets (2003) investigated the use of a multiblock solver with an
implicit LES model, in which the subgrid-scale dissipation is accomplished through
253
13:36
P1: IBE
0521871441c06
254
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
0.2
0.15
0.1
0.05
0
0
10
20
30
40
50
Figure 6.9. Centerline distribution of streamwise root-mean-square fluctuations. Legend: , Bodony

and Lele (2004); , Bogey, Bailly, and Juve (2003); , Lau, Morris, and Fisher (1979); , Zaman (1986);
, Arakeri et al. (2003).
numerical upwinding, on an impressive array of jet conditions and effective nozzle

geometries, including nozzles with chevrons. The nozzles are not directly included,
however, and the influence of the nozzle geometries is modeled through a specified
inflow velocity profile. A Ffowcs WilliamsHawking surface provides extrapolation
of the sound to the far field. Without explicitly forcing unsteadiness into the incoming
fluid they found, similar to Andersson (2003), OASPL agreement of 3 dB over a range
of conditions.
6.2.3.2 Discussion of LES acoustic results
There are believed to be two main types of sound generation in subsonic turbulent
jet plumes: the turbulence-generated sound concentrated in the vigorous mixing region near the end of the potential core and the lower-frequency, large-scale instability
wave-generated noise. (We exclude in this discussion other important noise sources
especially the high-frequency noise generated in the thin, near-nozzle shear layers. See
Section 6.2.4.2.) Each noise component is argued to have a unique acoustic spectrum
(Tam, Golebiowski, and Seiner 1996; Goldstein 2003), and the overall sound produced
by a jet contains significant contributions from both. It is thus necessary for an LES
noise prediction to capture both phenomena.
For the turbulence-generated sound associated with the end of the potential core,
simple reasoning suggests that a necessary, but by no means sufficient, statistical quantity to be captured accurately is the centerline distribution of the turbulence fluctuation
levels. Several data are shown in Figure 6.9 of the streamwise rms velocity fluctuations. (Note: In this and other figures, the data have been shifted along the x-axis
to yield a common potential core collapse location.) The rapid rise of the fluctuation
levels is due to the merging of the annular shear layers and the subsequent induced
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
255
0.5
0.4
0.3
0.2
0.1
0
0
10
20
30
40
50
Figure 6.10. Centerline distribution of density root-mean-square fluctuations normalized by the difference ( j ). Legend: , LES data of an unheated, Mach 0.9 jet by Bodony and Lele (2004); ,
Panda and Seasholtz (2002).
mixing. In the data shown, the fluctuations peak near 1215% of the jet exit velocity.
The LES predictions appear to be consistently higher than the experimental measurements, which is probably a result of the weakly turbulent shear layers of the simulations
relative to the fully turbulent experimental shear layers. In all cases the width of the peak
is roughly the same with a uniform rate-of-decrease fluctuation level with increasing
axial position.
In those studies with excessive azimuthal correlation of inlet disturbances (Bodony
and Lele 2002a; Morris et al. 2002), the increased axisymmetric coherence manifests
as a large overshoot of the centerline u x,rms levels. A 20% increase in the peak rms
value was typical. Downstream of the peak, however, there was a rapid decay of the
fluctuation intensities, and thus beyond x/ro = 25, roughly, there was agreement with
experimental data in the self-similar region. It should be emphasized, in light of these
data, that validation of jet LES data in the self-similar range is not sufficient and may
yield inaccurate conclusions.
The role of the density fluctuations in sound generation has not yet been well established owing, in part, to the difficulty is measuring within the jet. Lighthill (1952)
argued that the velocity fluctuations are primarily responsible for the quadrupole stress
and thus that the density may be taken as uniform with a value equal to that of the
density in the ambient fluid. Crow (1970) put the neglect of density in the sound
source on firm ground for incompressible jets. For higher-speed or hot jets, or both,
this is not expected to be the case, but there are not yet sufficient data from which
to draw a conclusion. The density fluctuations in unheated jets have recently been
measured by Panda and Seasholtz (2002) using the Rayleigh scattering technique.
Their results further support the neglect of the density fluctuation considerations in
sound generation. Figure 6.10 shows the density rms levels along the jet centerline for
13:36
P1: IBE
0521871441c06
256
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.11. Far-field OASPL taken at a distance of 30D j

from the nozzle exit. Legend: , Bodony and Lele (2004); ,
Bogey, Bailly, and Juve (2003); , Freund (2001); , Stromberg
et al. (1980); , Mollo-Christensen et al. (1964).
an unheated, Mach 0.9 jet normalized by the difference j . The LES-predicted

density fluctuations are roughly twice those measured by Panda and Seasholtz (2002).
That the LES-predicted sound fields are in agreement with those of laboratory-measured
jets (see Figure 6.11) further suggests the insensitivity of the sound to density fluctuations for unheated, subsonic jets. The role of the density fluctuations for hot jets is just
beginning to be established (Panda et al. 2004; Bodony and Lele 2004).
Beyond the fluctuation levels, it is the spacetime structure of the turbulence that
plays a central role in the statistical properties of the sound field (Goldstein 2003). More
precisely it is the spacetime character of the sound source that is important, but, through
approximation, sound field may be related to the velocity correlation u i (x, t)u j (x +
, t + ). Two parameters that characterize the correlation are the integral length
scale and the moving-axis, Lagrangian integral time scale T . Measurements of these
quantities is difficult, and limited data are thus available. Davies, Fisher, and Barratt
(1963) found that, for a low-Mach-number jet (M j near 0.45), the Lagrangian time
scale could be correlated to the local inverse shear,

4.5
,

T =
U /r
(6.15)
for locations near the lip line at the end of the potential core. This argument implicitly
states that the turbulence in the shear layer and potential core region is essentially
equivalent to homogeneously sheared turbulence (Blaisdell, Mansour, and Reynolds
1993). That such a relation exists in the potential core region of a turbulent jet is
unclear. However, Figure 6.12 shows that, for the limited data available, the correlation
is approximately retained in the LES for those points near the lip line (r/ro = 1). The
filled-in circles of Figure 6.12 correspond to measurements taken just downstream of
the potential core, beginning with r/ro = 0.5 and increasing in increments of 0.1. The
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
257
4
3.5
3
Figure 6.12. Integral Lagrangian time scale
of streamwise fluctuations near the end of
the potential core. Legend: , LES data of
Bodony and Lele (2004) with r /r o > 1; , LES
data of Bodony and Lele (2004) with r /r o < 1;
, correlation of Davies et al. (1963).
2.5
2
1.5
1
5
10
open circles are for r/ro > 1 with similar increments. For those locations near the lip
line, the correlation of T with the local inverse shear holds in the manner of Davies et al.
(1963). There is a rapid departure for locations r/ro > 1, which is believed due to the
quasi-laminar nature of the incoming, annular shear layers. These results suggest that,
at least for r/ro < 1, the jet spacetime structure of the turbulence is relatively well
captured in the LES. Similar agreement was found in the Mach 0.75 jets of Andersson
(2003).
With the possible exception of the density fluctuations, the agreement between the
LES data and experiment would suggest that the radiated noise for the turbulencegenerated noise should be properly predicted. This is borne out in Figure 6.11 of
the OASPL directivity where, to a large extent, the LES-predicted levels are within
the scatter of the experimental data. Near the maximum intensity angle of 30 , the
LES data have little scatter relative to the other angles and are consistently below
the DNS data of Freund (2001), which may be due to low-Reynolds-number effects
in the DNS. Over the downstream angles, the LESexperimental data agreement is
encouraging.
The spectral content of the sound as a function of polar angle is given in Figure 6.13, where the simulation data of Bodony and Lele (2004) have been scaled to
the physical dimensions of apparatus used by Lush (1971) for his 195-m/s jet. Note
that the narrow-band spectra of the LES have been synthesized into 1/3-octave band
spectra for the comparison. At lowest polar angle, both the peak frequency and amplitude are well captured by the simulation. For the intermediate angle of 45 the
simulation again captures the peak frequency and amplitude but with little energy in
frequency components much beyond the peak frequency. By 90 , the peak frequency
is missed and the simulation spectra are evidently a low-pass filtered version of the
experimental data. The energy at higher frequencies that is missing in the simulation
but present in the measurements is generated in two regions: (i) near the nozzle exit
15
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
258
0 521 87144 1
printer: sheridan
November 20, 2006
60
40
20
10
10
10
10
60
40
20
10
10
10
10
60
40
20
10
10
10
10
Figure 6.13. One-thirdoctave spectral comparisons of an unheated, Mach 0.5 jet of Bodony and Lele
(2004) with the 195-m/s data of Lush (1971). Legend: , simulation; , Lush (1971).
and (ii) within the jet plume very near the potential core with the first location being
the more dominant. The effect of missing the near-nozzle shear layers is quite evident,
and yet the predictions are quite good for those frequencies that are present in the
simulations.
Bogey and Bailly (2006) and Bodony and Lele (2004) have found (see Figure 6.14)
that the azimuthal correlation of the sound in the far field is in relatively good agreement
with the experimental measurements of Maestrello (1976). At small angles to the jet
axis, where the large-scale instability waves are dominant, the sound field shows a large
degree of azimuthal correlation around the jet, indicating the presence of strong, axisymmetric disturbances. Michalke (1984) notes that the jet is a highly selective amplifier
with a strong preference for axisymmetric motions. Thus, the creation of axisymmetric
1
0.75
Figure 6.14. Azimuthal correlation of the far-field sound
field of LES data from Bogey, Bailly, and Juve (2003)
( ) and Bodony and Lele (2004) () compared with
experimental data by Maestrello (1976) () at = 30 and
= 90 .
0.5
0.25
0
0.25
0
30
60
90
120
150
180
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
modes through convection is manifested in the sound field at these low angles. For
larger angles to the jet axis, where the large-scale instability waves are less important,
a more rapid decorrelation with increasing separation angle is found.
The question of the robustness of these results to changes in the subgrid model is
an important one that has only recently been investigated. By surveying the literature
cited throughout this article, one can see that, for those studies that use high-fidelity
numerics, the variety of methods taken to compute the jet near field, the acoustic
far field, or both do not have a significant effect on the turbulence intensities or on the
radiated sound field. Bogey and Bailly (2005a) examined the sound-field changes due
to different subgrid-scale models (dynamic coefficient Smagorinsky with and without a
turbulence kinetic energy correction) and to a simulation using selective filtering alone
to remove high-frequency energy. The kinetic energy correction appears to have little
effect. In changing from the usual LES procedure of including a subgrid-scale model to
the use of selective filtering, Bogey, Bailly, and Juve (2003) reported a downstream shift
in the jet development and small, but statistically meaningful, changes in the turbulence
field. The acoustic fields do differ in this case; in particular, the spectra show a shift
toward lower frequencies for a fixed point caused by the lateral shift of the jet as a
whole when the selective filtering procedure is used. However, the two models have
roughly the same rate of energy decrease at larger Strouhal numbers, suggesting that the
high-frequency noise is not significantly affected. Likewise, Zhao et al. (2001a) found
that a dynamic mixed model produced a jet with slightly higher turbulence level and a
louder sound field relative to the dynamic Smagorinsky alone but without significant
overall change.
6.2.4 Future directions of jet acoustics
On the basis of the results presented in the previous section, it appears that the use
of LES to predict the noise radiated by high subsonic jets is justified. There is evidence that the predicted noise-generation-relevant turbulence and radiated acoustic
field characteristics are in agreement with existing experimental and numerical data.
Before LES as a noise prediction tool becomes widespread, however, several issues
need to be addressed. What follows is a brief discussion of the LES-specific issues that
are deemed important; it is by no means exhaustive.
6.2.4.1 Resolution effects on the near and far fields
As discussed in Section 6.2.3.2, the peaked spectrum at 30 is believed to be related

to the large-scale jet motions with length scales on the order of the jet diameter. It is
of interest to know the degree to which a coarse-grid LES, which is not able to predict
the near-field turbulence levels, accurately can reproduce the low-frequency noise. One
That is, those studies with an implementation that did not introduce numerical diffusion, dispersion, or
both.
259
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
260
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.15. Far-field OASPL taken at a distance of 30D j from

the nozzle exit. Legend: , 100,000-point simulation; ,
1,000,000-point simulation of Bodony and Lele (2004); , Freund (2001); , Stromberg et al. (1980); , Mollo-Christensen
et al. (1964).
such comparison for the OASPL is given in Figure 6.15 and for the spectra in Figure 6.16
(dashed curves in (a) and (b)). The coarse LES, with one-tenth the number of grid
points but with the same inflow and lateral boundary conditions, is able to reproduce
the OASPL to within 2 dB and retains the spectral peak location and amplitude near
St = 0.2. The higher-frequency noise is clearly missed. That the poorly resolved LES,
in terms of the turbulence, is capable of estimating the low-frequency noise at shallow
angles may be of interest because the MGBK and TamAuriault formulations (Tam
and Auriault 1998, 1999) are unable to do so in their current form (cf. Choi et al.
1999). Further work is needed to characterize the resolution dependence of the radiated
sound field to determine the usefulness of a hybrid approach using analogy-based
methods for the high-frequency noise and a coarse-grid LES for the low-frequency
noise.
SPL
100
80
(a)
60
40
-2
10
10
-1
St
10
SPL
100
80
(b)
60
40
-2
10
101
St
10
Figure 6.16. Far-field acoustic spectra taken at a distance of

30D j from the nozzle exit. Legend: , 1,000,000-point LES
simulation of Bodony and Lele (2004); , 100,000-point
LES simulation; , similarity spectra of Tam et al. (1996). In
the upper (lower) figure, the low-frequency (high-frequency)
spectra of Tam et al. (1996) are used.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.2.4.2 Subgrid-scale noise model
As a corollary to Section 6.2.4.1, one issue of the truncation of scales in an LES is

the effect on the sound generated by the missing scales of motion. Because only the
largest scales, as determined by the grid resolution, are represented in an LES, the
influence of the subgrid scales on the resolved scales is modeled in the closures used
in the mass, momentum, and energy equations. The role of the subgrid-scale motions
in the sound-generation process is, as yet, not understood. At the Reynolds numbers
of realistic flows, the subgrid-scale noise may well account for a significant fraction of
the relevant frequencies in the far-field spectrum. Bodony and Lele (2003a) recently
estimated that, for an LES of a Mach 0.9 jet to capture the St > 0.5 far-field noise at 90
from the jet axis to within 1 dB, a grid spacing of 0.0008D j 0.0016D j would be necessary for axial locations of 1 x/D j 2. Current simulations have a minimum grid
spacing of 0.01D j ; recovering the missing noise with better-resolution simulations,
including in the near-nozzle regions with thin, turbulent, annular shear layers, is a long
way off.
To illustrate the missing noise, consider the spectra of Figure 6.16. Here the difference in the two simulations, as discussed previously, is apparent: the finer-resolution
computation has a larger frequency content. Compared with the empirical data parameterized by Tam et al. (1996), both simulations fall far below for St > 1. Because the
experimental data were measured under realistic conditions, the elevated spectral levels
are due primarily to the lack of the thin, turbulent, annular shear layers present near the
nozzle exit. It is these shear layers, where the inherent time scales are much shorter than
those found in the plume, in which the high-frequency sources lie. Using the estimate
of the previous paragraph with that of Freund and Lele (2004) shows that the number
of points needed to include both the near-nozzle shear layers and the jet plume is on the
order of 110 billion. This number of points is beyond current computational capacity
and thus an alternative approach is needed.
One such approach is the statistical subgrid noise model recently proposed by
Bodony and Lele (2002b, 2003a) using Goldsteins (2003) generalized acoustic
analogy. In their formulation, a field variable, for instance the velocity u i , is decom(L)
(M)
posed into a large-scale component u i and a missing-scale component u i . The
large-scale variable would be that computed in an LES. On joining this decomposition with Goldsteins (2003) source term Bodony and Lele (2002b, 2003a) proposed
(L) (M)
a subgrid-scale noise model composed of terms of the form u i u j that represents
the interaction of the large- and missing-scale components. Further work is needed to
establish the validity of this formulation.
Another approach is to develop numerical techniques that allow one to compute
the near-nozzle region along with the turbulent plume at reasonable cost. The work
of Jansen et al. (2002), for example, illustrates the difficulty in including the nozzle
directly but clearly defines the gains that may be had in considering small sections of
the nozzle when the nozzle geometry warrants such a simplification. The introduction
of the discontinuous Galerkin (Hu et al. 1999) and compact finite volume (Piller and
Stalio 2004), with their better resolution properties, will be necessary.
261
13:36
P1: IBE
0521871441c06
262
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.2.5 Conclusions for far-field jet acoustics
The suitability of LES to the prediction of jet noise has been well established
for unheated, high-subsonic jets. The introduction of high-bandwidth, high-accuracy
numerical techniques has allowed for the computation of both the unsteady, near-field
turbulence found in the jet exhaust plume and in its radiated sound. Numerous research
groups have reported agreement between their LES results and experimental data for
both the near-field fluctuations and the far-field sound characteristics for high-subsonic
jets at moderate Reynolds numbers. The most recent investigations have exhibited the
correct acoustic spectral characteristics over a range of frequencies. Ongoing investigations are exploring the influnce of the subgrid-scale model and the grid resolution
on the radiated sound. Future work is focused on extending the LES results to higher
frequencies through the development of subgrid-scale noise models.
6.2.6 Acknowledgments
The authors gratefully acknowledge support from the Aeroacoustics Research Consortium, a government and industry consortium managed by the Ohio Aerospace Institute,
Cleveland, Ohio, U.S.A.
6.3 Cavity noise

Xavier Gloerfelt, Christophe Bogey, and Christophe Bailly
6.3.1 Introduction to cavity noise
Impinging shear layers give rise to intense, self-sustained oscillations as well as noise
radiation. Flow past a cavity belongs to this class of flows and has been extensively
studied for the last 50 years because of its practical interest, its geometrical simplicity,
and the diversity of the theoretical questions that it raises. The generic features are
assessed in detail in the reviews of Rockwell and Naudascher (1978); Tam and Block
(1978); Rockwell (1983); Komerath, Ahuja, and Chambers (1987); Howe (1997); or
Colonius (2001).
The self-sustained oscillations arise from a feedback loop consisting of the following
chain of events. The growth and convection of instability waves in the shear layer
induce large-amplitude pressure disturbances as the vortical perturbations impinge on
the downstream corner of the cavity. The upstream influence of the generated pressure
fluctuations provides further excitation of the instabilities in the shear layer especially
in its most receptive region near the upstream edge. A stable phase criterion is then
installed between the downstream and the upstream edges of the cavity. This complex
phenomenon is often greatly simplified to build lumped models such as the Rossiter
formula (Rossiter 1964) in which the free-shear layer is viewed as two-dimensional and
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.3 CAVITY NOISE
the recirculating flow is neglected. At high Mach numbers, this simple formula succeeds
in predicting the admissible Strouhal numbers, although it provides no information on
the amplitude of the self-sustained oscillations and fails to indicate which of the multiple
modes will be predominant. At low Mach numbers, for frequencies close to those of the
acoustic resonances, the acoustic forcing by the cavity resonance can overwhelm that
provided by the feedback loop. Besides, this kind of semiempirical model is independent
of the Reynolds number and cannot describe the changes between an initially laminar
or turbulent incoming boundary layer.
The ways to study cavity-flow phenomena are thus complete experiments or numerical simulations. An overview of previous simulations and recent advances is provided
in Section 6.3.2. The 2D limitations are explained in view of previous works. Recent
achievements by LES are described in Section 6.3.3. We attempt to identify important
issues such as the influence of the turbulence or the coexistence of multiple tones, which
will require further studies.
6.3.2 Overview of cavity-flow simulations
6.3.2.1 CFD simulations of cavity flow
Use of URANS approaches
The first CFD computations of unsteady cavity flows used the two-dimensional unsteady
Reynolds-averaged NavierStokes (URANS) equations with a turbulence model. The
early work of Hankey and Shang (1980), using 78 52 grid points, was promising,
showing fair agreement with frequencies of oscillations and with distributions of the
pressure coefficient along the cavity walls from experiments. With the development
of CFD codes in the 1990s, numerous URANS simulations were performed from the
subsonic to hypersonic regimes. Various turbulence models were coupled: one-equation
models based on the eddy viscosity concept of Boussinesq, such as the CebeciSmith
(see Hankey and Shang 1980) or the BaldwinLomax models (see Baysal and Stallings
1987), or two-equation models such as the k (see Suhs 1987; Shih, Hamed, and Yenan
1994) or the k models (see Zhang 1995; Henderson, Badcock, and Richards 2000).
The effectiveness of such models for separated flows remains an open question. For
instance Slimon, Davis, and Wagner (1998) found that simulations display a strong
sensitivity to the choice of the turbulence model. Tam, Orkwis, and Disimile (1995)
also showed that the results are affected by high values taken by the turbulent viscosity.
They even noticed better results for the estimation of the time-averaged surface pressure
field with a zero-equation turbulence model, which led Rona and Dieudonne (2000) to
prefer to study laminar flow motion. Orkwis et al. (1998) also showed how the results
can be affected by the resolution algorithm.
The first 3D applications were carried out in the late 1980s by Suhs (1987), Rizzetta
(1988), and Srinivasan and Baysal (1991) and captured additional features of the 3D
shear layer and recirculation inside the cavity but yielded limited improvements owing
to the relatively coarse-mesh grid used. The development of industrial RANS codes
allowed the study of complex geometrical configurations, including passive or active
devices for the control of cavity oscillations (Kim and Chokani 1990; Suhs 1993;
263
13:36
P1: IBE
0521871441c06
264
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Baysal, Ten, and Fouladi 1994; Ota et al. 1994; Lamp and Chokani 1997; Zhang et al.
1998; Soemarwoto and Kok 2001; Bortz et al. 2002).
Use of hybrid RANSLES approaches
To enhance the description of the unsteady features of the separated shear layer at high
Reynolds numbers, Shieh and Morris (2000) or Paliath and Morris (2003) tested in 2D
the use of a hybrid RANSLES approach based on the detached-eddy simulation (DES)
concept. Extensions to 3D flows were carried out recently by Sinha, Arunajatesan, and
Ukeiley (2000), Shieh and Morris (2001), and Hamed, Basu, and Das (2003). This
method aims at exploiting the best features of both approaches by solving RANS
equations in the turbulent boundary layer and performing LES in the separated shear
layer.
6.3.2.2 CAA simulations of cavity noise
Two-step methods
The two-step approaches separate the flow calculation and the noise propagation problem in order to apply the most appropriate method at each step. One of the most famous
is the acoustic analogy of Lighthill (1952), extended by Curle (1955) and Ffowcs
Williams and Hawkings (1969) (FWH) to take into account the effects of solid boundaries. Zhang, Rona, and Lilley (1995) used URANS simulations and Curles spatial
formulation to obtain far-field spectra of cavity noise, but no validation was proposed.
In Gloerfelt, Bailly, and Juve (2003), the results of a direct computation of the noise by
DNS were successfully compared with three integral methods: the FWH analogy, the
Kirchhoff method, and a wave extrapolation method based on the FWH equation. A
wave extrapolation method based on the FWH equation was also applied by Ashcroft
and Zhang (2001) to extend a compressible URANS solution to the acoustic far field.
Note that these extrapolation methods require a compressible simulation below the extrapolation surface, whereas the source terms for the acoustic analogy can be deduced
from an incompressible simulation.
Hardin and Pope proposed another two-step method: the viscous flow is obtained
from an incompressible simulation, and a correction to the constant density is defined. The acoustic radiation is then obtained from the numerical solution of perturbed, compressible equations. This technique was applied to the computation of flowinduced cavity noise by Hardin and Pope (1995), Slimon et al. (1998), and Moon et al.
(1999).
These methods suffer, however, from two difficulties: the modeling of the source
terms from aerodynamic fluctuations and the ability of the wave operator to include
complex acoustic flow interactions.
Direct noise computations
Because flow and acoustic fluctuations are solutions of the compressible NavierStokes
equations, it is possible to obtain both fields in the same calculation. However, owing
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.3 CAVITY NOISE
Figure 6.17. Transition toward a wake mode for large L/ ratio (2D DNS of the flow over an L/D = 4
cavity, at M = 0.5, and ReD = 4800). Snapshots of the vorticity field (levels in s1 ) for two successive
times T/2 and T during one oscillation period T . On the left, a shear-layer mode is observed for
L/1 63, whereas, on the right, a wake mode is visible for L/ 2 280 (Gloerfelt, Bailly, and Juve
2000).
to the great disparities between these two quantities, in classical computational fluid
dynamics (CFD), acoustical phenomena are not resolved accurately because of the numerical schemes used and inadequate grid-cell size or time step. Moreover, reflections
due to the boundary conditions can shade the physical acoustic wave field. High-order
algorithms minimizing dispersion and dissipation and nonreflecting boundary conditions have therefore been developed that permit DNC.
In the case of cavity noise, initial attempts of DNC were made for supersonic flows
by Zhang (1995) and Rona and Dieudonne (2000) for which the amplitude of the shock
waves is very strong and the acoustic waves can not travel upstream. The first computations of the noise induced by a cavity with a subsonic grazing flow were carried out
by Colonius et al. (1999), Rowley, Colonius, and Baasu (2002), and Shieh and Morris
(1999b) using 2D DNS. These simulations showed a transition toward a new flow regime
when the ratio L/ of the cavity length over the momentum thickness became large, as
already observed by Najm and Ghoniem (1991) by using a vortex method. This regime
is characterized by the shedding of a single vortex that occupies the entire cavity and
overshadows the role of the smaller-scale vortices of the separated shear layer. The periodic ejection of this structure is associated with an increase of the cavity drag. The same
numerical bifurcation was also noted by Gloerfelt et al. (2000), as illustrated in Figure
6.17. Rowley et al. (2002) called the new regime a wake mode because of similarities
with the transition observed in the experiments of Gharib and Roshko (1987) in a water
tunnel. However, the presence of a wake mode is not seen in experiments of compressible cavity flows. Did it result from the very low Reynolds numbers imposed by DNS
or from the 2D approach? To investigate higher Reynolds numbers, Shieh and Morris
(2000) applied computational aeroacoustics (CAA) tools to solve hybrid URANSLES.
The transition to a wake mode was still observed, indicating that it could be related to
the 2D behavior rather than to the Reynolds number. A subsequent 3D study by Shieh
and Morris (2001) did not show a wake mode transition for the same configurations,
265
13:36
P1: IBE
0521871441c06
266
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
confirming the 2D nature of the bifurcation. In a 3D vortex, stretching significantly

modifies the turbulent mixing between counterrotating vortices in the recirculation
region and prevents untimely transition to the wake mode.
Another 2D DNC using DNS of the cavity flow was performed in the framework
of the third CAA workshop by Kurbatskii and Tam (1999), Shieh and Morris (1999a),
and Heo and Lee (2001) for a door cavity. Heo and Lee (2001) studied the effects of
cover plates obstructing the cavity opening. Koh and Moon (2003) simulated 2D compressible turbulent flows and compared a URANS approach using a second-moment
turbulence closure and a limiter for the energy balance with the DES model. Bidimensional behavior of the recirculation zone was visible as well in the computations of Heo,
Kim, and Lee (2003). The latter used 2D DNS with high-order numerical schemes and
probed the feedback loop through correlations.
6.3.3 Recent achievements using LES
6.3.3.1 A challenging test case
The previous simulations indicate that a 3D approach is required to describe the turbulent mixing inside the cavity and, if necessary, the turbulent character of the incoming
flow. A DNS solving all the scales (down to the Kolmogorov scale) would be impractical. To achieve reasonable computational cost, and given the dominance of coherent
structures in the shear layer, cavity flows seem a good candidate for LES. This configuration is, however, a challenging test case insofar as the flow can be viewed as a
synthesis of a simple shear flow in the mixing layer above the cavity, a complex shear
flow inside the cavity, and multiscale wall boundary layers.
Basic difficulties in cavity-flow LES are thus the modeling of the unresolved subgrid
scales (SGS), the description of the near-wall region, and the intensive computational
resources required. The first point is still subject to discussion. The most popular solution is the use of an eddy viscosity model linearly related to the resolved stress tensor
such as Smagorinsky or dynamic Smagorinsky models, which are referred to as SM and
DSM, respectively. These models are often overdissipative in high-Reynolds-number
configurations (Fureby and Grinstein 2002), and it can be inferred that the eddy viscosity
may have a nonnegligible effect on the shear-layer dynamics and that a less dissipative model would be helpful to preserve high-Reynolds-number features. Alternative
approaches have been proposed. Zang et al. (1993) used a mixed model that combined the DSM and the similarity model of Bardina for turbulent recirculating flows in
driven cavities, reducing the magnitude of the dynamically computed coefficient. Boris
et al. (1992) have suggested damping the turbulent energy by the numerical procedure.
In their monotone integrated LES (MILES), the numerical dissipation is provided by
the use of low-order upwind schemes. The use of high-order explicit filtering appears
more appropriate in order to control the numerical dissipation and the cutoff between
resolved and unresolved scales. For instance, Visbal and Rizzetta (2002) obtained better
results using compact filtering alone than with the DSM for isotropic turbulence. In the
same manner, Bogey and Bailly (2005c) compared the DSM and the use of selective
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.3 CAVITY NOISE
filtering alone for a high-Reynolds-number high-subsonic jet. Their results support the
fact that the eddy viscosity artificially decreases the Reynolds number. The second
difficulty for wall-bounded flows concerns the description of the near-wall structures.
The resolution of the inner layer exceeds present computational capabilities already at
moderate Reynolds numbers. Methods to bypass the wall layer are required to perform
high-Reynolds-number LES. One alternative is proposed by the hybrid RANSLES
approaches, which switch to RANS solutions in the boundary layers. Another possibility is the use of wall models for the inner layer (Piomelli and Balaras 2002) even if
their extension to complex configurations is not simple (see Shih, Povinelli, and Liu
2003). The application of the SGS model to wall-bounded flows is also questionable.
The simplest approach is the use of damping functions to account for the scale reduction near the walls such as the van Driest function. Several modifications reviewed
by Piomelli et al. (1989) or Sagaut, Montreuil, and Labbe (1999) have been proposed
to model the near-wall effects. This kind of model is also hardly extendable to more
complex configurations. A dynamic procedure, such as that of DSM, may overcome
this difficulty by using the information from the smaller resolved scales. Note that the
latter problem is automatically avoided in alternative approaches based on implicit SGS
contributions such as MILES or the compact or selective filtering approaches.
Owing to the preceding considerations and the computational ressources required,
very few large-eddy simulations have been carried out until now. The first applications
concerned transonic or supersonic flows over shallow cavities, representing weapon
bays of military aircrafts. Large-eddy simulations by Dubief and Delcayre (2000),
Smith (2001), or Lillberg and Fureby (2000) reproduced high-Reynolds-number configurations with fairly coarse resolutions. Rizzetta and Visbal (2002) carried out LES
of weapon bays with L/D = 5, M = 1.19, and Re L = 2 105 . The computations involved 20.6 million grid points and used high-order compact schemes and DSM to
describe the fine-scale structures. Two computations were performed with or without suppression devices based on high-frequency mass injection. Another interesting
weapons bay LES without the acoustic field was performed by Larcheveque, Sagaut,
and Le (2003). The existence of strong transverse acoustic modes in the cavity was
highlighted. Preliminary results with the DSM by Oh and Colonius (2002) can also be
quoted. Three recent aeroacoustic contributions by Larcheveque, Sagaut, Mary et al.
(2003), Gloerfelt, Bogey, Bailly, and Juve (2002), and Gloerfelt et al. (2003a) are presented below.
6.3.3.2 LES of a deep cavity in a channel
Larcheveque, Sagaut, Mary et al. (2003) performed LES of the flow over a deep cavity
with a length-to-depth ratio of L/D = 0.42, a freestream Mach number of 0.8, and a
Reynolds number Re L = 8.6 105 , reproducing exactly the experiment of Forestier,
Jacquin, and Geffroy (2003). Two LES strategies were tested: a traditional LES with
an eddy viscosity model, where t was determined by the selective mixed model,
using a second-order centered scheme, and a MILES approach in which the intrinsic
dissipation of an upwind scheme mimicked the dissipative behavior of the unresolved
267
13:36
P1: IBE
0521871441c06
268
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.18. LES of a deep cavity.
scales. In both simulations, a third-order compact RungeKutta algorithm was used

together with a wall model based on the logarithmic law. The turbulence level inside
the boundary layer was provided by the superimposition of random fluctuations in the
inlet plane.
The two LES strategies gave a very good agreement with measured spectral peak
levels. Reynolds and phase averages were also remarkably well reproduced, even if
the turbulent levels were slightly higher with MILES. This is probably related to the
more dissipative discretization scheme in the MILES approach. The acoustic field was
obtained thanks to a 2D3D domain coupling. Figure 6.18 shows a view of the computational domain. Structures are visualized using isosurface of the Q-criterion equal to
10(U /L)2 combined with a Schlieren-like picture in the background (Larcheveque,
Sagaut, Mary et al. 2003). The Q-criterion defines as vortex tubes the regions in which
the second invariant of velocity gradient tensor Q = 1/2 (i2j si2j ) is positive. This
view reveals the existence of a characteristic lattice of multiply reflected waves in the
channel similar to that observed experimentally by Forestier et al. (2003). The presence
of upper wall reflections can trigger the shedding of new vortices.
6.3.3.3 Influence of the incoming-flow turbulence
Three-dimensional large-eddy simulations were performed by the authors to describe

the influence of the small scales and of the intermittency of the turbulence. In his
pioneering work on cavity oscillations, Karamcheti (1955) noticed a different behavior whether a laminar or a turbulent boundary layer interacted with the cavity. The
frequency of the oscillations for a turbulent inflow was slightly lower than the one
measured with a laminar inflow. Moreover, a reduction in the amplitude of the pressure
fluctuations was observed together with the emergence of a low-frequency component
in the spectra for a turbulent inflow. The origin of this low-frequency component, with
roughly half the frequency of the fundamental, remains unexplained. One of Karamchetis configurations was reproduced by Gloerfelt, Bogey, Bailly, and Juve (2002)
using LES with two different incoming flows: a laminar Blasius boundary layer with
no forcing, and a turbulent boundary layer generated by superimposing random Fourier
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.3 CAVITY NOISE
modes on a turbulent mean profile. The length-to-depth ratio was L/D = 3, with
D = 2.54 mm as in the experiment, and the freestream Mach number was M = 0.8,
giving a Reynolds number Re L = 1.46 105 . The governing equations were integrated in time using an explicit low-storage, six-step RungeKutta scheme optimized
in the frequency space. The derivatives were calculated by using optimized finite differences with an eleven-point stencil for the convective fluxes, and fourth-order finite
differences for the viscous and heat fluxes. As part of the algorithm, a selective filtering using an eleven-point stencil was incorporated in each direction to eliminate
grid-to-grid oscillations. The coefficients of the RungeKutta, of the finite differences,
and of the filtering are given in Bogey and Bailly (2002). The computations lasted
40 hours for 5.4 million grid points (CPU time of 0.6 s per grid point and per iteration). The effects of the unresolved SGS were modeled via the Smagorinsky eddy
viscosity with a self-adaptive van Driest damping function near the solid boundaries
for the turbulent inflow simulation. No model was applied for the initially laminar
case because SM cannot describe the turbulent transition. For examining the role of
the turbulence model, incorporated solely in the initially turbulent simulation, the turbulent case was also performed with the use of selective filtering alone (i.e., without an explicitely added SGS model) in a subsequent study by Gloerfelt, Bogey, and
Bailly (2002).
The results are displayed in Figure 6.19 for the three large-eddy simulations.
On the left, snapshots of the instantaneous vorticity modulus (top view at
x2 = 0.06 D + sideview at x3 = 0) are shown. The levels are between 0 and 2 106 s1
for the three cases. The vorticity snapshots indicate that the coherent structures are actually made up of several small-scale vortices more concentrated with the laminar
inflow than with the turbulent inflow. The vorticity concentration is also enhanced
more with SM than with the explicit filtering. The features of the shear layer, with finer
scales without SM, support the fact that the effective Reynolds number is reduced by
the use of an added eddy viscosity. In the right column of Figure 6.19, the pressure
field in the computational domain and OASPL is shown at the point (D, 4.6D, 0)
in decibels (ref 2 105 Pa). The pressure grayscale is the same for the three cases,
with levels between 3000 and 3000 Pa (Gloerfelt, Bogey, and Bailly 2002; Gloerfelt,
Bogey, Bailly, and Juve 2002). The structure of the radiated field showed a strong downstream directivity owing to the convection of acoustic wave fronts by the mean flow
combined with the reflections on the cavity walls. The acoustic emission was directly
linked to the impingement of the coherent structures, and the pressure levels seemed
correlated with the vorticity concentration of the structure. The highest amplitude is
found for the laminar configuration in agreement with Karamchetis results (Karamcheti 1955). The comparison of acoustic spectra for the turbulent configurations with or
without SM indicated that the levels are lower without the model. A part of the energy
is dumped into the more numerous small scales, which induces a less coherent interaction with the downstream edge, and, consequently, reduces the acoustic emission. The
results also showed the emergence of a low-frequency component for the turbulent case
without SGS model, as in Karamchetis measurements.
269
13:36
P1: IBE
0521871441c06
270
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.19. Influence of the boundary layer turbulence on cavity noise. LES of the L/D = 3, M = 0.8
configuration of Karamchetis experiments with laminar inflow (top); turbulent inflow with SM (CS =
0.18, van Driest damping) (middle); and turbulent inflow with selective filtering alone (bottom).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.3 CAVITY NOISE
Figure 6.20. Mode switching. Three isocontours of instantaneous modulus of vorticity (1, 2, and 3 106 s1 ) are superimposed on the instantaneous vorticity averaged over the spanwise direction (grayscale
between 1.2 and 1.2 106 s1 ) for two instants during one cycle. A mode I, with one
big roll in the shear layer, is visible on the top,
and a mode II, with two coherent rolls, is identifiable on the bottom (Gloerfelt et al. 2003a).
6.3.3.4 Switching between competitive modes
The previous configuration (L/D = 3, M = 0.8) was simulated with periodic boundary conditions in the spanwise direction by Gloerfelt et al. (2003a), allowing a finer
resolution in this direction. The same algorithms were used, and the transition toward
a realistic turbulent boundary layer ahead of the cavity was successfully validated. The
SGS contribution used SM with a Smagorinsky constant CS = 0.1 and an adaptive van
Driest function near the walls. Attention was primarily focused on the intermittence
of the turbulence, leading to the coexistence of two peaks in the spectra for turbulent
conditions (Karamcheti 1955). The switching between modes I and II of the cavity oscillations was demonstrated (see Figure 6.20). This switching was not random in time
but followed a cyclical pattern with a successive alternation between the two modes
even if a certain level of intermittency was observed in this succession. In turbulent
conditions, the coherent structures display a shape of clusters of small scales, and the
alternation of different sizes of the dominant structures can proceed by a reorganization of these clusters. The examination of the early shear-layer growth revealed that
the formation of large-scale structures was possible through a collective interaction
after the upstream corner, corresponding to the fusion of several smaller vortices shedded at the most unstable frequency of the shear layer. This phenomenon appears to
be of fundamental interest in the selection process of the main oscillation frequencies.
Another interesting finding is the strong unsteadiness of the recirculating flow within
the cavity, which seems severely coupled to the shear-layer modulations. However, the
role of the recirculating flow on the switching phenomenon is not clear because the
coexistence of different modes has been noticed in the self-sustained oscillations for
configurations without adjacent recirculating flows. The acoustic field clearly indicates
the predominance of the lower component, whereas the Schlieren visualizations of
271
13:36
P1: IBE
0521871441c06
272
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Karamcheti indicate the predominance of the high component. The development of

the lower mode in the LES was enhanced by the numerics especially by the eddy
viscosity and thus a more violent impingement on the downstream corner of the
cavity was induced and dominated the acoustic radiation.
6.3.4 Concluding remarks for cavity noise
For high-Reynolds-number flows, the dynamics of the small scales and the intermittence of the turbulence may lead to the coexistence of different modes with the flow
switching between them. This phenomenon can hardly be described by RANS methods and is better suited to LES. The first complete aeroacoustic computations are very
recent. The large-eddy simulations described here reveal the dependence of the results
on the resolution and on the SGS model. The radiated field induced by the cavity flow
is greatly sensitive to small flow modifications such as the concentration of the coherent structures. The computation of the correct acoustic levels in the far field is thus
a challenging problem. It requires subsequent investigations on the role of the SGS
models and on the effects of the numerical resolution. Significant progress relies on
the direct comparisons with experimental databases. Such databases with aerodynamic
and acoustic information have been lacking and are just beginning to emerge (Forestier
et al. 2003).
The simultaneous obtainment of the aerodynamic and acoustic fields from the compressible NavierStokes equations is made possible by the use of very accurate algorithms and of quiet boundary conditions. This can provide further insights into the
noise-generation mechanisms. In particular, the investigation of the nonlinear interactions between modes by Gloerfelt et al. (2003b), the characterization of the phase
relationship between the two corners of the cavity especially the identification of the
acoustic feedback path or the study of the growth of the shear-layer instabilities may
help toward a better understanding of the selection process in order to design reduced
models and efficient control strategies.
6.4 Aeroelastic noise

Sandrine Vergne, Jean-Marc Auger, Fred Peri
e,
Andre Jacques,
and Dimitri Nicolopoulos
6.4.1 Introduction to aeroelastic noise
Aerodynamic noise is one of the major contributors to internal vehicle noise; it becomes
dominant above 400 Hz at driving speeds exceeding 100 km/h when compared with
structure-borne, power train, and tire noise for which substantial noise reduction has
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.4 AEROELASTIC NOISE
been achieved. The interaction of the flow with the geometrical singularities of the
vehicle body produces transient turbulent flows, often detached, resulting in an increased
aerodynamic noise generation. Two main phenomena are involved:
r Aeroacoustic noise: The production and decay of turbulent structures within the
flow yields the generation of acoustic sources. That is propagative pressure waves
by opposition to the mostly convective nature of the flow around the car. The pressure
waves produce a loading of the vehicle body and side windows which do transmit
part of the noise inside the vehicle.
r Aeroelastic noise: The convected pressure waves fluctuations (pseudo noise), although not propagative, do also load the vehicle structures and therefore induce
vibrations that will radiate noise inside the passenger compartment.
This section focuses mainly on this second kind of noise, which is also named
aerovibroacoustics.
Flow-induced vibration phenomena are a ubiquitous feature in numerous engineering
applications from dynamic stability (e.g., airplane stability, deformations of building
structures, and bridges under wind loads) to structural radiation (e.g., the trailing-edge
noise described in Howe (2000) or the flow-induced sound inside a plane or a car
due to the pressure fluctuations of the boundary layer on the exterior of the vehicle).
In the first case, the vibrations are undesirable, causing material fatigue at best and
catastrophic failure at worst. The interactions between flow and structure are strong. In
the second case, flow-induced forces on bodies and structures induce vibration. This
vibration, in turn, induces additional noise and may even modify the unsteady fluid
flow. In many cases, the structural surfaces do vibrate, but these vibrations cannot alter
the flow motion. Turbulent structures remain unaffected by the acoustic effects. This
assumption corresponds to the implicit hypothesis, made in acoustic analogies, that
the acoustic pressures are small in comparison with the turbulence pressures. These
flowstructure interactions may be considered weak.
Several research tasks are dedicated to the improvement of numerical and experimental methods of building hybrid structuralacoustic models based on a combination of experimental and numerical data. The global purpose is to combine equivalent
sources or forces derived from experiments and vibroacoustic, transfer path models in
order to calculate the pressure at the passengers ears within the vehicle compartment
(Mein, Dupuy, and Bohineust 2002; Bohineust, Bardot, and Dupuy 1996; Bohineust
et al. 1999). These hybrid models are now available for several stages of the design and
development process for power train and tire noise but are still under development for
aerodynamic noise calculation.
Such hybrid models have been applied to test cases like a flat plate, either coupled
or not coupled to a cavity, and submitted to a turbulent boundary layer flow. The structure problem could be solved using the normal-mode analysis technique: the structure
is assumed to respond in its normal modes of vibration, which describe a condition
of resonance. Therefore, given a known description of the fluid characteristics on the
273
13:36
P1: IBE
0521871441c06
274
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
surface, the vibroacoustics response of the system could be evaluated. For a turbulent
boundary layer, the excitation is described by a semiempirical wave-number frequency
spectrum model such as the well-known Corcos model (Corcos 1964), the Smolyakov
and Tkachenko model (Smolyakov and Tkachenko 1991), theoretical ones such as the
Chase model (Chase 1987), or the Ffowcs Williams model (Ffowcs Williams 1982). In
most cases (Graham 1997; Leclercq 1999), computed acoustic responses radiated by the
plate strongly overestimate the experimental results above a threshold frequency, and
the accuracy of the predicted response levels is very sensitive to the spacetime characteristics of the turbulent flow (i.e., the coherence and phase velocity of the wall-pressure
fluctuations).
Moreover, research efforts are aimed at introducing numerical approaches to calculate aerodynamic noise. Many studies (Perot et al. 2004; Lokhande, Sovani, and
Xu 2003; Reichl et al. 2003) deal with the estimation of the aerodynamic noise by
using LES computation, but the acoustics receivers are put outside the vehicle. Theses
numerical approaches can help toward an understanding of the physical mechanisms responsible for the noise generation. To better understand the coupling between turbulence
and vibroacoustics, we consider a numerical method based on unsteady, fully coupled
resolution of 3D compressible NavierStokes equations and structural conservation
laws. This method has been applied to several types of problems such as aeroacoustic
simulation of a rear side-mirror vehicle (Obrist, Nicolopoulose, and Jacques 2002),
flow fan noise (Roy and Cho 1999), and automotive exhaust to resolve coupled problems featuring noise transmission through a vibrating structural part (Nicolopoulos,
Perie, and Jacques 2004; Sakurai, Endo, and Perie 2002). In this section, we will focus on the ability of the Radioss CFDCAA code to calculate a typical case of many
industrial applications: the vibroacoustic response of an elastic platecavity system
excited by a turbulent flow. The behavior of the model is studied in terms of its three
components: hydrodynamic (flow field and wall pressure fluctuations), structural (plate
accelerations), and acoustic (cavity response). These different components are compared with an existing experimental database obtained by Peugeot Societe Anonyme
(PSA) (Leclercq 1999).
6.4.2 Fluidstructure interaction
We consider the plate to be driven by the wall-pressure fluctuations induced by the

turbulent flow and by the acoustics fluctuations due to plate motions (Figure 6.21). This
model is given by the equations of motion,
ms
2w
+ Ds w = ( pt + p2 p1 ),
+
2
t
t
t
(6.16)
and acoustics (Helmholtz equation),

p0
1 2 pi
= 0,
c02 t 2
(6.17)
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
275
Figure 6.21. The plate model.
where w(x1 , x2 , ) is the plate normal velocity, Ds is the plate bending stiffness, m s is
the plate mass per unit area, and is a mechanical damping coefficient of the panel.
The plate bending stiffness Ds is related to the plate Youngs modulus E, thickness
h, and Poissons ratio , by Ds = Eh 3 /12(1 2 ). The plate vibrations are excited by
turbulent wall-pressure fluctuations pt (xi , ) on the upper surface of the plate. The
terms p1 (xi , ) and p2 (xi , ) are the internal and external acoustic pressure generated
by the motion of the panel, respectively.
This system of equations is completed by the boundary conditions of the plate
(simply supported, clamped) and by the condition linking pi and w at the plate surface:
w
pi
= 0
.
x3
t
(6.18)
The external flow is governed by the NavierStokes equations

(u i )
= 0,
+
t
xi
( pi j i j )
(u i ) (u i u j )
=
,
+
t
x j
x j
(6.19)
(6.20)
where u i are the flow velocity components, the volumic mass, c0 the sound velocity,
and i j the viscosity shear stress.
An unsteady slipping boundary is applied on the moving panel as
u n(t) = w n(t),
(6.21)
where u is the flow velocity, w the panel velocity, and n(t) the normal to the surface
that is dependent on time t.
In most literature (Blake 1970a, 1970b; Kraichnan 1956) for boundary layer excitations, authors have assumed that whatever surface motion occurs does not influence the
fluid dynamics (i.e., u n(t) = 0). This weak-coupling assumption is physically valid
for most fluidstructure interaction. The weak-coupling assumption is not used in the
considerations presented in this section.
13:36
P1: IBE
0521871441c06
276
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.4.3 Numerical simulation
The approach taken in this section for computational aeroacoustics attempts to take
several physics into account in a single simulation:
r Unsteadiness of the flow involving macroscopic scales;
r Dissipation due to microscopic turbulent scales;
r Acoustic contribution of macroscopic turbulence;
r Acoustic wave propagation, including impedance of boundaries;
r Noise transmission to structures; and
r Structural response.
The CAA methodology is grounded on the following numerical choices:
r The transient solution of the full, compressible NavierStokes equations that al-
low for the propagation of pressure waves as well as fluid flow. Its double precision
implementation guarantees enough accuracy for both average values (flow) and fluctuations of the calculated variables (macroscale turbulence and acoustic waves). In
these classes of aeroacoustic problems, the dynamic range of the pressure fluctuation
can be from 2 105 Pa up to several thousand Pa.
r An explicit time integrator (first-order) that provides enough time resolution to
capture bifurcation phenomena. This method yields a fairly cheap computational
cost per element per time step (at least one magnitude order cheaper than an implicit
solution) at the expense of a high number of time steps.
The simulations are performed with the explicit module of Radioss CFDCAA
developed by Mcube and Mecalog. The implementation of NavierStokes equations is
based on the arbitrary Lagrangian Eulerian (ALE) formulation, as presented in Donea
(1982): each point in space has the usual material velocity

u as well as a grid velocity
v describing the a priori arbitrary movement of nodal points. Conservation laws may
be written as
+ [(
u
v ) ]
u = 0,
t
(6.22)
+ [(
u
v ) ] +
u = 0,
t
(6.23)
+ [(
u
v ) ]e + (e + p)
u = 0.
t
(6.24)
Note that, for

u =
v , Equations (6.22)(6.24) reduce to the Lagrangian descrip
tion, whereas for v = 0, they describe the Eulerian case. As a consequence, a single
formulation is able to describe the evolution of physical variables in the laboratory
reference frame and in a grid with any arbitrary movement (e.g., owing to the grid
deformation required to accommodate deformation and movement of structures).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
277
The spatial integration is achieved with a second-order finite element formulation

with a streamline upwind PetrovGalerkin (SUPG) integration scheme (Brooks and
Hughes 1982). The time integration is performed with an explicit time-centered integration scheme whose stability is governed by Courants condition,

h
,
(6.25)
t = min
c +
u
where h is the minimum element dimension and c is the speed of sound. These result in
very small time steps that, together with SUPG, provide minimal numerical diffusion
and enough accuracy to capture and to convey the resolved vortices.
These schemes allow turbulent structures greater than 10 elements to be resolved
by the grid. Smaller scales are taken into account by the Smagorinsky SGS model with
Cs = 0.1. The proposed method does not pretend to resolve turbulence near walls.
A special treatment (Benyahia 2003) assumes a logarithmic profile and computes the
turbulent viscosity at the boundary layer as if all turbulent structures were unresolved
there. Special care is also taken for sound absorption due to unresolved turbulence by
adding a viscous bulk behavior:
Pvisc = sgs V /V,
(6.26)
where Pvisc is the corrective pressure dependent of the density , the subgrid viscosity sgs for the LES, and the flow velocity fluctuation rate V /V . This correction has
enhanced the sound-pressure wave reflections on structures through boundary layers.
Note that the compressible nature of Equations (6.22)(6.24) takes into account
acoustic wave propagation within the computational domain. This allows direct capture
and propagation of the aeroacoustic noise sources, avoiding the tedious data storage
and propagation calculations needed by the volume terms in Lighthills approach.
However, it requires modeling the radiation at the computational domain boundaries.
The linearized Euler equation of Bayliss and Turkel (1982b) provides an efficient
treatment of this problem:
u n
c
p
= c
+
(P p),
t
t
2lc
(6.27)
where u n is the velocity normal at the boundary, and P is the stagnation pressure. The
first term on the right-hand side is the Sommerfeld impedance, and the last one is a relaxation term toward the average value either imposed or derived by assuming continuity:
p
= 0.
(6.28)
This equation is equivalent to a high-pass filter. Moreover, a Fourier analysis of Equation (6.27) shows that these two terms match the radiation impedance of a monopole
at distance 2lc . This means that the computational domain boundaries should be constructed so that the simulated acoustic sources are far enough apart to be seen as
monopoles.
13:36
P1: IBE
0521871441c06
278
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.22. Sketch of the experimental setup: small ruler of size L = 0.010 m and H = 0.009 m and
elastic plate.
The finite element implementation allows a simple nodal approach to connect different parts either by simply sharing nodal connectivity or through the use of interfaces.
Several nonconforming interfaces are available for CAA applications:
r Fluidfluid,
r Fluid wallstructure,
r Tied structures, and
r Periodic.
By its nature the ALE formulation allows a two-way coupling of the fluid and the
Lagrangian domains. This means that noise can be transmitted to structures and vice
versa. It also means that structures can be modeled through the same formulation using
a wide variety of finite elements and kinematics conditions, namely,
r thin or thick shell, quads, or triangles;
r beams;
r springs and dampers;
r rigid bodies; and so on.
6.4.4 Application
6.4.4.1 Test case description
An existing test case is used to evaluate model results as defined and fully investigated by
PSA (Leclercq 1999). It was built in order to analyze and to improve the comprehension
of the mechanisms of energy transfers between a turbulent flow and the vibroacoustics.
In the experimental setup (Figure 6.22), a flat plate is excited by a turbulent flow and
coupled to a cavity. This system should represent a very simplified vehicle structure
for which the flat plate and the cavity represent, respectively, a vehicle panel and the
cabin; the flow velocity is of the order of 140 km/h, for which the aerodynamic noise is
dominant. As a way of approaching the complex flows observed around the vehicles, the
flow is perturbed by a small ruler, which could represent a seal or a windshield wiper.
Upstream, a turbulent boundary layer fully develops on a thick, rectangular, flat
table, 2.4 m long and 0.9 m wide. At a distance of 0.05 m before the fence, the turbulent
boundary layer thickness is estimated at 0.02 m. The 40 m/s flow and its turbulent
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
279
0.1
Figure 6.23. Wall-pressure transducer positions on the

elastic steel plate.
x (m)
P1: IBE
0521871441c06
35 31
40
0
39
27
25
0
x1 (m)
boundary layer are then perturbed by a small ruler, 0.009 m high and 0.01 m wide,
mounted on the floor (Figure 6.23). Downstream, the resulting turbulence excites a
1-mm-thickness elastic steel plate (0.366 m 0.306 m) coupled to a cavity (0.370 m
long, 0.310 m wide, and 0.280 m high).
6.4.4.2 Experimental setup
An experimental study has been conducted to characterize this test case accurately
(Leclercq 1999). Measurements involve flow spatiotemporal characteristics and vibroacoustics response. Velocity profiles (particle image velocimetry, hot-wire anemometry) are measured to characterize the flow field. Because the flexible plate cannot be
equipped, wall-pressure measurements are made over a rigid plate in a first step. The
area is equipped with 42 wall-pressure transducers with their positions as shown in Figure 6.23. The transducers are developed by PSA to characterize the wall pressure field,
accurately (Leclercq and Bohineust 2002). Note that this approach implicitly assumes
a weak coupling (i.e., that pressure fluctuations are dominated by the turbulence and
not modified by plate vibrations).
In a second step, the vibroacoustics response of the flexible plate coupled to the
acoustic cavity is then studied by fitting the experimental setup with transducers to
measure the plate vibratory field (six accelerometers) and the cavity acoustic field (five
microphones) (Figure 6.24).
6.4.5 Simulation model
6.4.5.1 Mesh description
Mesh sizes and general dimensions
The computational domain begins 50 mm upstream of the obstacle and is 960 mm long,
as shown in Figure 6.25.
As conditions for accurately performing aeroacoustic simulation, it has been established by Perie (2002) that
r The Strouhal number associated with the element dimension must be less than 1/12
in the acoustic sources region. This gives in our case a smallest mesh size length of
3 mm (the frequency range of interest being 403000 Hz).
r To capture the structures generated by the obstacle (whose dimensions are
10 mm 9 mm) a mesh requires smaller size. Twenty elements per side of the
0.1
13:36
P1: IBE
0521871441c06
280
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
x3 (m)
24
26
32
Figure 6.24. Vibroacoustics measurements: (+) accelerometers and () microphone positions.
30
0.1
x2 (m)
0.1
x1 (m)
obstacle give a size of 0.5 mm. As the minimal size of the whole mesh, this 0.5 mm
induces the critical time step (Courants condition), 1.3 ms, which should be compared with the 0.6 s of physical duration necessary for simulation.
r Twelve elements per smallest wavelength are required in the propagation zones, in
order to have minimal attenuation and dispersion when acoustic waves travel across
the mesh. This gives 25 mm in our case.
r Mesh consists of hexahedral elements. The mesh size in the transverse direction (y)
has the constant value of 5.2 mm.
The plate and the cavity
The plate is meshed with shell elements so that its flexion is represented by at least six
elements of size h per smallest wavelength. Considering that the celerity of the flexion
waves is given by
C f = (B/Ms )1/4 1/2 ,
B=
Figure 6.25. General mesh dimensions.
Eh 3
,
12(1 2 )
(6.29)
(6.30)
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
281
Figure 6.26. General sizes of the mesh.
where B is the bulk modulus, Ms the surface mass, the pulsation, E the Young
modulus, the Poisson coefficient, and h the plate thickness, we find that this gives a
size of 8 mm for the plate elements size.
Because only acoustic propagation is simulated in the cavity, the following values
were used: 6.2 mm in x , 10.6 mm in y , and 15 mm in z . They all widely satisfy the
propagation criterion of 25 mm.
Mesh description
First, an automatic triangle mesh of the xz plane is generated. The triangular elements
are then recombined to create quadrangular elements. This allows a greater flexibility in
the mesh size repartition. Then, the 2D mesh is extruded along the spanwise direction.
This mesh is particularly well refined around the obstacle, where the cells keep a
structured topology. The mesh density areas can be seen in Figures 6.26 and 6.27.
The average 0.5-mm-density region follows the high shear layer whose location was
identified in the first model. Near the plate, the mesh dimension is 1.0 5.2 1.0 mm3 .
The total size of this mesh is 1,098,310 hexahedral elements, 1,710 shell elements, and
1,141,433 nodes.
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0.05
0.1
Figure 6.27. Detailed mesh around the obstacle.
0.15
0.2
13:36
P1: IBE
0521871441c06
282
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.4.5.2 Numerical model
Air is modeled using a linearized compressibility assumption equivalent to an isothermal equation of state.
Connection between the fluid main part, the cavity, and the plate
The fluid, the cavity, and the plate meshes are connected through a nonconforming
LagrangeLagrange interface. The nodes of the fluid and of the cavity in contact with
the plate are set Lagrangian.
Initial and boundary conditions
r For initializing the fluid flow, a velocity of 40 m/s, oriented in the x-direction, is
used as an initial condition for the first time step.
r Continuous fluid motion is forced through an imposed velocity condition at the inlet
of the system; the velocity profile Vx (z) close to the wall describes the upstream
boundary layer as estimated in the experiment. The y and z components of the
velocity are set to zero.
r A constant pressure of 0 Pa is applied at the outlet of the computational domain,
whereas the upper boundary has a zero pressure gradient.
r A silent boundary treatment is applied at both outlet and at the upper boundary: the
frequency components of the pressure waves greater than the cutting frequency of
273 Hz are absorbed. The acoustic impedance at boundaries is in fact equivalent to
the radiated impedance of a monopole located at 100 mm away.
r On the cavity sides and on the floor (including the obstacle), the velocity is set to
zero (no slip).
r Symmetry conditions are imposed on the lateral planes (Vy = Vnormal = 0).
r For modeling the sealing of the plate, a series of springs are used and their stiffnesses are adjusted to approach the experimental resonance frequencies. These
resonance frequencies are the damping coefficient are obtained by classical vibration measurements (a simple freedom degrees extraction method is used to posttreat
the frequencies response of a punctual excitation). The translation stiffness is set
to 1.5 107 N/m, and the rotation stiffness to 15 N/m. The damping coefficients
have been input as 50% of the critical damping, which gives in translation 85 kg/s
and in rotation 0.278 103 kg/s. Those values were calibrated earlier during the
evaluation of the hybrid method described in Section 6.4.3.
r The plate is set Lagrangian.
r The two-way fluidstructure coupling is achieved by the ALE formulation applied
to the fluid regions, whose meshes actually deform near the surface of the plate.
Analysis
Two phases are required to perform the aeroacoustic simulation:

r During a highly transient phase, the flow and the plate vibrations converge toward
a stabilized regime until an energy and acoustic convergence is reached. This phase
lasts 0.1 s of physical time.
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
283
Figure 6.28. Contours of the streamwise

mean velocity U1 /U0 near the ruler in the
median plane (levels between 0.4 and 1.2;
white line: isocontour of 0).
x3 /h
10
5
0
10
x1/ h
20
30
r Once convergence is obtained, the simulation is continued in order to record pressure
signals at selected locations. The sampling frequency is 10 kHz. A duration of 0.5 s

is necessary to have 10 consecutive blocks of 512 records ready to perform averaged
spectra.
The total computing time is 70 elapsed hours and 550 CPU hours on a CRAY SX6 with
eight processors.
6.4.6 Numerical results
6.4.6.1 Flow-field analysis
Mean flow
To validate the simulation, one must first evaluate the CAA codes ability to simulate
the flow accurately. Figure 6.28 presents the modulus of the mean velocity scaled
on U0 near the ruler in the vertical median plane. The figure clearly shows that the
flow separates on the ruler, the single recirculating region encompassing both the top
and the rear faces. The flow reattaches on the horizontal wall at approximately 20 h.
This flow separation generates turbulence, as shown by the nondimensional streamwise
fluctuations u rms /U0 in Figure 6.29. The more significant turbulent kinetic energy is
observed near the reattachment point (about 1520 h).
The simulation leads to an overprediction of the reattachment length from those
observed in the literature. With a block of length 2 h, Moss and Baker (1980) measured
the reattaching point at a distance of 12 h downstream from the front face. Mohsen
(1967) measured similar separation length downstream from a fence (around 1112 h).
The simulated profiles for the mean streamwise velocity U1 and the root-meansquare (rms) velocities u 1 and u 3 , respectively, in the streamwise direction and in
the direction normal to the wall are compared with experimental data at about 42 h
downstream from the ruler (Figure 6.30). The experimental mean velocity profile is
always strongly perturbed by the fence, which induced a mean velocity drop and a
strong turbulent layer increase. The numerical mean velocity profile is less disturbed
10
Figure 6.29. Streamwise velocity fluctuations urms /U0 near the ruler in the median
plane (levels between 0 and 0.3; black line:
isocontour of 0.25).
x3/ h
P1: IBE
0521871441c06
5
0
10
20
x1/h
30
13:36
P1: IBE
0521871441c06
284
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.30. Velocity profiles at the origin of the reference:

(-) measurement; (+ + +) simulation. (a) Mean streamwise
velocity U1 , (b) rms streamwise velocity u1 , (c) rms normal
velocity u3 .
by the fence. The nondimensional streamwise mean velocity U1 /U0 and rms velocity
u 1 /U0 are overestimated by the computation in the turbulent boundary layer. Thus,
the boundary layer thickness calculated is underestimated: the momentum thickness is
around 0.26 h for 0.65 h measured.
Mean pressure and rms levels
The mean pressure distribution downstream from the ruler is similar to the observation of Mohsen (1967) and downstream fences in terms of absolute levels and trends
(Figure 6.31). The minimum level appears in the separation bubble region, and the
maximum just downstream from the reattachment point. The extent of the region with
rapid pressure rise is longer for this computation (about 10 h) than in Mohsens (1967)
observation (8 fence heights).
At about x1 = 18 h, the maximum rms pressure level is observed about 0.13 times
the dynamic pressure q (Figure 6.32). This value is greater than the observation of
Mohsen (1967), who found a factor of 0.06 for a 51.8 m/s velocity downstream from
a fence. The comparisons with our experimental data show a good agreement for the
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
285
0.5
Cp
0.25
0
Figure 6.31. Pressure coefficient.
20
40
60
x /h
80
100
points situated in the middle of the flat plate (Figure 6.33). For the two upstream points,
the level is overestimated, and this is consistent with the overpredicted reattachment
length already observed.
6.4.6.2 Wall-pressure fluctuations
In the classical random vibration theory, the turbulent loading function of structures
subject to low-speed flow is described by the spacetime cross correlation between two
points separated by = (1 , 2 ) or its time and spatial Fourier transform, the wavevector frequency spectrum. Rather than study directly the cross spectrum, we can use
the following expression:
S pp (x, , f) = pp (x, f) (x, , f)e(i2f/Up (x,,f)) ,
(6.31)
where
r pp (x, f) is the autospectrum of the fluctuating wall pressures,
r U p (x, , f) is the turbulent eddy convection velocity in the nominal flow direction,
and
r 2 (x, , f) is the coherence function (Bendat and Piersol 1980), which is sometimes
called the coherency-squared function.

The shape and properties of these three functions of space and frequency are important for the computation of the dynamic response of a panel subjected to a similar
0.15
0.125
Figure 6.32. Wall-pressure-fluctuation rms levels calculated at various positions along the median plane.
prms / q
P1: IBE
0521871441c06
0.1
0.075
0.05
0.025
0
20
40 60
x1 / h
80
100
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
286
0 521 87144 1
printer: sheridan
November 20, 2006
0.1
prms / q
35 31
27
25
Figure 6.33. Comparison of calculated ( ) and measured (+ + +) wall-pressure fluctuation rms levels on the
plate.
0.05
0
20
39
30
40
x /h
50
60
excitation. In this study, we use the same formalism to validate our computations: the
autospectrum is calculated at different positions in the flow, the coherence functions are
evaluated for streamwise and spanwise point pairs, and the phase velocity is computed
along the main flow direction.
Power spectral density (PSD)
As shown in Figures 6.34(a) and 6.34(b), the wall-pressure fluctuation PSD simulated on the flat plate are in good agreement with experimental results obtained for
points located downstream far enough from the ruler (points 25 and 40 in Figure 6.23).
The discrepancies appear for frequencies higher than 500 Hz; levels become underpredicted (about 15 dB at 1000 Hz), corresponding to a cutoff effect of the
model. Some oscillations on computated spectra are mainly caused by the limited
number of available blocks (only 21 blocks of 512 points) used to perform spectrum averaging. For points closer to the ruler (points 31 and 35 in Figure 6.23), results show important discrepancies as long as the location goes up (Figures 6.34(c)
and (d)): the power spectral density (PSD) intensities are overvalued for low frequencies ( f < 500 Hz). In this region close to the reattachment point, the rms
pressure level is still decreasing quickly, as shown in Figure 6.32. Therefore, the result
comparison is quite sensitive to the prediction of the recirculation length.
In conclusion, far enough from the reattachment zone, the simulation correctly
the predicts the low-frequency structures, and thus the wall-pressure power spectral
densities, up to a frequency around 500700 Hz.
Cross-spectral densities (CSD)
The coherence 2 is compared at different positions in the flow for streamwise and
spanwise transducer pairs with separation. The phase velocity Up is only measured
along the main flow direction. For all experimental and numerical comparisons, the
computed coherence in the streamwise direction is overrated for high frequencies (Figure 6.35) for which as already observed in the section on PSD, the limits of computation
are reached: in fact, the wall-pressure PSD showed that there is no excitation at high
frequencies (predicted pressure levels are 20 dB below for f > 700 Hz). For lower frequencies, the computed coherence is also overvalued for the smaller separation distance
(Figure 6.35(a): = 0.011 m) but underestimated when the distance increases.
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
287
r ef
(f) / P2 ]
Channel 25
100
80
70
60
200
400
600
f (Hz)
Channel 40
800
1000
200
400
800
1000
200
400
800
1000
200
400
800
1000
100
ref
(f) / P2 ]
10 log10 [
pp
90
80
70
60
600
f (Hz)
Channel 31
ref
100
80
70
600
f (Hz)
Channel 35
110
ref
(f) / P2 ]
10 log10 [
pp
90
100
pp
Figure 6.34. PSD of wall-pressure fluctuations measured on the flat plate (point locations in Figure 6.23): (-) measurement;
(- - -) simulation (Pref = 2 105 Pa).
(f) / P2 ]
10 log
10
pp
90
10 log10 [
P1: IBE
0521871441c06
90
80
70
600
f (Hz)
13:36
P1: IBE
0521871441c06
288
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.35. Coherence 2 of wall-pressure fluctuations measured on the flat plate for streamwise
points using the same notation as in Figure 6.34. (a)
1 = 0.011 m, (b) 1 = 0.040 m, (c) 1 = 0.109 m.
As mentioned in Section 6.4.4.2, wall-pressure fluctuations are measured on a rigid

test plate. Noise due to plate vibration is not taken into account; an experimental analysis
done by comparing a rigid and an elastic plate case at downstream locations shows more
important pressure fluctuations and a coherence drop for frequencies below 200 Hz
(Leclercq 1999) for the elastic case as against the rigid one. This is in good agreement
with the numerical observations, which undervalued the experimental coherence for
the greatest separation distance (Figure 6.35(c)).
The comparisons of spanwise coherence are similar with those in the streamwise
direction (Figure 6.36). Furthermore, coherence variation versus separation distance is
correctly reproduced by the model.
The computed phase velocity (established from cross-spectrum phase of Bendat
and Piersol 1980) is in good agreement with the experimental results (Figure 6.37); the
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.36. Coherence 2 of wall-pressure fluctuations measured on the flat plate for spanwise points using the same notation as in Figure 6.34. (a) 2 = 0.010 m,
(b) 2 = 0.030 m, (c) 2 = 0.078 m.
error is lower than 10%. The convection velocity is calculated correctly by the explicit integration scheme together with the nondiffusive SUPG spatial numerical
scheme.
6.4.6.3 Vibroacoustics response
The plate response to aeroelastic loading is owed, on the one hand, to the intensity
and to the spatial distribution of the stress, and, on the other hand, to the intrinsic
characteristics of the structure that is, as represented by its modal base. For the lower
frequencies ( f < 1000 Hz), the modal density is weak and the spatial coupling between
modal shapes and the wall-pressure loading (coherence, convection speed) determines
the plate response. The comparisons of numerical and experimental acceleration PSD
aa show a good agreement (Figure 6.38). The frequency shift is estimated to be due
to the parametric adjustment of the plate eigenfrequencies, which are controlled by
289
13:36
P1: IBE
0521871441c06
290
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.37. Phase velocity Up of wall-pressure fluctuations measured on the flat plate for streamwise
points using the same notation as in Figure 6.34.
(a) 1 = 0.011 m, (b) 1 = 0.011 m, (c) 1 = 0.040 m.
structural boundary conditions (i.e., spring parameters). For frequencies greater than
700 Hz, the computed PSD level is underestimated (about 10 dB); again the limits of
the simulation are reached.
For acoustic pressure p in the cavity, results are quite similar to those found for the
plate acceleration a (Figure 6.39). In conclusion, vibroacoustic quantities are correctly
modeled in the frequency range below 700 Hz. For frequencies above 700 Hz, the
underprediction of numerical results increases with increasing frequency.
6.4.7 Mesh influence
Unfortunately, present computer performance limitations do not allow use of a refined

mesh everywhere in the computational domain. Despite ever-increasing computational
performance, it is unlikely that this situation will significantly change in the near future.
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
291
Channel 26
2
10log10[aa(f) / aref]
140
120
Figure 6.38. Acceleration PSD aa channel 26

(aref = 1. 106 ms2 ): () measurement;
(- - - -) simulation.
100
80
60
250
500
f (Hz)
750
1000
Models able to compute acoustic frequencies up to 5000 Hz around a side mirror

and window with simulations long enough to record 20 samples will feature over
10 million nodes and need to be computed for about 1 million time steps. Conducting
such a simulation in a reasonable engineering turnaround time requires access to highly
powerful supercomputers featuring hundreds of processors.
For the case studied here, the numerical model initially ran for 19 days on Fujitsu
VPP5000 (one processor). The time was reduced to 55 hours by using a CRAY SX6
(eight processors), and more recently to 37 hours on thirty-two Intel IA64 processors.
The CPU time is getting reasonable, but still requires large computing servers. Therefore, a careful mesh design phase must be carried out to make sure that the CPU time
is kept under control; meanwhile the quality of the results is satisfactory. The general
idea is to alternate as fast as possible between a noise generation fine mesh criteria
to a noise propagation coarser one. This is achieved by making sure that the highly
turbulent zones are meshed with the fine criteria. Several techniques can be used to
do this, among which is the examination of experimental or existing numerical RANS
visualizations of the flow. In this case, an initial mesh has been built and a careful
examination of the flow results led to the development of a second mesh, with the same
number of elements but a better spatial repartition:
r The first one was achieved by using a block-structured technique, which allows
refinement close to the obstacle (where the mesh size is 0.5 5.2 0.5 mm3 )
Channel 30
100
Figure 6.39. Acoustic pressure PSD pp

channel 30 (Pref = 2. 105 Pa) using the same
notation as in Figure 6.38.
2
10log10[pp(f) / pref]
P1: IBE
0521871441c06
75
50
25
0
250
500
f (Hz)
750
1000
13:36
P1: IBE
0521871441c06
292
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
0.1
0.09
0.08
0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
0.05
0.1
0.15
0.2
Figure 6.40. First mesh: detailed mesh around the obstacle.
and coarsening in the far field to reach the value of 10.0 5.2 6.6 mm3 (Figure 6.40). The total number of elements for this mesh is 974,752 hexahedral elements, 1710 quad shells, and 1,017,424 nodes. This simulation was designed to be
accurate up to a frequency of about 750 Hz.
r The second one (see Section 6.4.5.1) was, in particular, refined near the flat plate
in order to better capture the smaller structures generated by the obstacle, which
propagate above the plate. This mesh was supposed to solve turbulent structures up
to 3200 Hz in the streamwise direction but only up to around 615 Hz in the spanwise
direction.
10 log 0 [ (f) / P2 ]
1
pp
ref
The comparison of wall-pressure fluctuation spectra shows some gain above 500 Hz
(Figures 6.41, 6.42) when the mesh is refined in the streamwise direction (second
model), but the improvement is not as good as expected: the spanwise mesh size
probably also needs to be refined to really improve the upper frequency limit of the
model.
Channel 25
100
90
80
s2
70
60
50
s1
200
400
600
f (Hz)
800
1000
Figure 6.41. Comparison of the two simulations wall-pressure fluctuations DSP on

the plate: m = measurement (), s2 ( ),
s1 (. . ).
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
293
6.5 TRAILING-EDGE NOISE
Channel 30
100
10log10[pp(f) / pref]
P1: IBE
0521871441c06
Figure 6.42. Comparison of the two simulations

acoustic pressure in the cavity using the same notation as in Figure 6.41.
75
50
25
0
250
500
f (Hz)
750
6.4.8 Conclusions for aeroelastic noise prediction
In this section, we have demonstrated the capability of the Radioss CFDCAA code
to predict aeroelastic and vibroacoustic phenomena due to turbulence in the wake of
a ruler. This approach exhibits accurate predictions below approximately 700 Hz. For
higher frequencies, the levels are underestimated. A refined mesh in the streamwise
flow provides somewhat better results, but the poor spanwise resolution of the model
is believed to limit the development of structures smaller than 0.05 m (corresponding
to frequencies higher than 500 Hz). Another research track is to evaluate the dissipation of wall functions. The fully coupled numerical approach presented here provides
better results than those obtained by a hybrid method combining modal analysis for the
plate and the cavity with a semiempirical model of the wall-pressure loading, which
overestimated the acceleration and acoustic radiation of the plate above 700 Hz.
The main advantage of a compressible ALECFD code like Radioss CFDCAA is
its ability to solve flow field; noise, source generation; propagation, including model
behaviors; and radiation and transmission to structures in a single simulation.
6.4.9 Acknowledgment
The authors would like to acknowledge Dominik Obrist of CRAY Computer Inc. and
Ian Godfrey of FUJITSU Systems for their active support throughout the study.
6.5 Trailing-edge noise

Roland Ewert and Eric Manoha
6.5.1 Introduction to trailing-edge noise simulation using LES
The applications described in Sections 6.5.2 and 6.5.3 address the general context of
the numerical prediction of the aerodynamic noise generated by aircraft in the vicinity
1000
13:36
P1: IBE
0521871441c06
294
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
of urban airports. It is known that, for the largest existing and future aircraft, the noise
radiated in the approach phase is evenly generated by the engines (working at reduced
thrust) and the airframe. Airframe noise sources are known to be mainly located in
landing gears and in high-lift devices, slats, and flaps, which are deployed during the
approach.
Among all the complex noise-source mechanisms associated with high-lift devices,
trailing-edge noise has probably received the most extensive attention (mostly theoretical and experimental) and is a perfect candidate for the evaluation of the emerging
CAA techniques of numerical simulation based on LES. The trailing-edge noise mechanism is associated with the acoustic scattering of the airfoil turbulent boundary layer
convected at the trailing edge. In this process, the quadrupolar (acoustically inefficient)
nature of the convected eddies is transformed into dipolar sources, which are much
more acoustically efficient.
The numerical simulation of the complete problem, including the local noise-source
simulation and the far-field noise radiation, faces the difficulty of handling fluctuations
covering a very extended range of length scales and amplitudes: turbulent eddies that
generate noise have small lengh scales but high energy, whereas the acoustic waves
radiated away have comparatively very long wavelengths but small amplitude. Thus,
trailing-edge noise simulation is still beyond the capabilities of complete DNS, and
hybrid methods are used in most practical cases. Figure 6.43 sketches the possible
numerical strategies, showing how the near-field turbulent flow and the far-field noise
are computed separately. The idea is to divide the physical space into several domains
in which specific physical mechanisms are simulated by using the most adequate set of
equations with the most economic discretization strategy.
Computational fluid dynamic techniques are used to simulate the near-field
flow, which contains the aerodynamic noise sources. Available techniques include
steady RANS computations, in conjunction with stochastic models of the wavenumberfrequency spectrum of the turbulence (Bechara et al. 1994; Bailly, Lafon,
and Candel 1995; Kalitzin, Kalitzin, and Wilde 2000; Billson, Eriksson, and
Davidson 2003), unsteady RANS methods (Singer et al. 1999; Khorrami, Singer,
and Berkman 2002), LES (Seror et al. 1999; Wang and Moin 2000; Piomelli 2001;
Boudet, Casalino, Jacob, and Ferrand 2003; Boudet, Grosjean, and Jacob 2003; Manoha,
Delahay, Sagaut et al. 2001; Manoha et al. 2002; Ewert, Meinke, and Schroder 2002;
Ewert, Schroder, Meinke, and El-Askary 2002; Ewert et al. 2003; Sorguven, Magagnato, and Gabi 2003), and nonlinear disturbance equations (Chyczewski, Morris, and
Long 2000a; Terracol et al. 2003). This local flow solution has to be coupled to an
acoustic numerical technique for the prediction of far-field noise. The most practical
formulations are the integral methods such as the Ffowcs WilliamsHawkings equation (Ffowcs Williams and Hawkings 1969), the boundary element method (BEM)
(Manoha et al. 1999), and the Kirchhoff integral (Prieur and Rahier 1998; Rahier and
Prieur 1997). Integral methods assume that, beyond a given distance from the noise
sources and body surfaces, the sound propagates in a medium at rest or moving with
uniform velocity. This assumption becomes a strong limitation when the radiated noise
results from a surface integration or a control interface located near solid walls, where
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
295
Local turbulence prediction:

Unsteady CFD (DNS, LES, unsteady RANS)
Steady CFD (RANS) + stochastic model
Flow
Airfoil
Turbulent flow
Low-turbulence
Nonuniform
mean flow
Outer region
with uniform
flow
Integral methods:
- Acoustic analogy
- Kirchhoff integral
- BEM
Acoustic propagation through

inhomogeneous flow using any
modified formulation of
linearized Euler equations
Far-field
observation
point
Figure 6.43. Numerical simulation of airfoil aerodynamic noise: possible hybrid strategies.
velocity gradients are significant. In that case, only the (linearized or nonlinear) Euler equations in perturbation form (and derivatives of these equations) governing the
acoustic propagation may account for the propagation in nonhomogeneous flows. This
is obtained at the price of a significant computational effort because the propagation
domain must be meshed separately with an adequate resolution with respect to the
smallest acoustic wavelength and also because finite difference, high-order schemes
are needed to ensure numerical accuracy and low dispersion of acoustic wave propagation (Tam and Webb 1993). Moreover, applications to realistic geometries, including
airfoils, need three-dimensional curvilinear grids (Grogger et al. 2000) on which the
use of high-order schemes is not straightforward. However, the domain in which Euler
equations is used can be strictly limited to regions in which velocity gradients are significant: in most practical airframe noise problems, an external boundary can be found
beyond which the flow can be assumed uniform, and thus integral methods can be used
for the noise prediction at very long distances from the airfoil.
The critical point is the coupling technique between the LES and the Euler domain.
Two different techniques are described in this section, and both are applied to the simulation of trailing-edge noise. The first coupling technique, developed in Section 6.5.2, is
based on a volume right-hand-side source term for linear disturbance equations determined from the unsteady flow data provided by the LES. In this case, the acoustic grid
13:36
P1: IBE
0521871441c06
296
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
used for the acoustic simulation can be completely dissociated from the aerodynamic
grid used for the LES, the results of which are used as a local-volume source term
for acoustic propagation. This method has been followed by several authors (Bechara
et al. 1994; Bailly et al. 1995; Billson, Eriksson, and Davidson 2002; Ewert, Meinke,
and Schroder 2000, 2001a), often using different formulations of the source term. Section 6.5.3 decribes a more straightforward technique based on direct injection, without
additional filter or source-term reconstruction, of LES perturbation data (obtained by
subtracting the time-averaged mean flow) at the border of the Euler domain. This process relies on the compressibility of LES and its ability to simulate the local acoustic
field correctly (Manoha, Delahay, Sagaut et al. 2001; Manoha et al. 2002).
6.5.2 Trailing-edge noise simulation using LES and APE
6.5.2.1 Hybrid simulation approach
This section concerns a hybrid acoustic simulation approach to airframe noise utilizing
LES to resolve the unsteady compressible flow problem in the hydrodynamic near field
and subsequently applying an acoustic analogy based on linear acoustic perturbation
equations (APE) to determine the related sound radiation toward the far field.
Computational results of the considered hybrid simulation concept are presented
for trailing-edge noise generated at a flat plate at Mach number M = 0.15, Reynolds
number Re = 7 105 , and an airfoil at M = 0.088 and Re = 8 105 . Figure 6.44
sketches the approach for the plate problem. Separating the analyses of the flow field
and the acoustic field in general makes it possible to take advantage of the disparity of the turbulent and acoustic length scales at low Mach numbers, where the latter
scale with M 1 compared with the former. For the trailing-edge noise problem, the
LES is limited to a subdomain close to the trailing edge. Hence, 2 106 mesh points
are sufficient for the resolution of the Re = 7 105 problem. The acoustic domain
has a substantially larger extension compared with the LES domain owing to the increased grid spacing allowed for the acoustic simulation and comprises, apart from
the LES region, the remaining geometry not resolved in the unsteady flow simulation.
The acoustic simulation based on APE considers mean-flow convection and refraction
effects of a nonuniform mean flow such that the computational domain of the LES has
to comprise only the significant acoustic source region in the immediate vicinity of the
trailing edge. The nonuniform mean flow in the acoustic domain in general is provided
by a steady-state RANS solution.
The linearized acoustic perturbation equations can be considered modifications of
the linearized Euler equations (LEE), which completely prevent the unbounded growth
of hydrodynamic instabilities in critical mean flows. Because the LEE describe the
propagation of acoustic, vorticity, and entropy waves, unstable solutions of the LEE
in a globally unstable mean flow occur. If the LEE are excited by additional sources,
the instabilities will cause divergent solutions that are limited neither by nonlinear
saturation nor by viscous damping (Ewert et al. 2001a).
On the basis of APE that encode a convected wave equation for nonuniform
mean flows, an acoustic analogy is derived from the full governing-fluid equations
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.44. Sketch of the computational domains to determine trailing-edge noise with the hybrid
approach. The LES domain encompasses the vicinity of the trailing edge, whereas the computational
aeroacoustics (CAA) domain includes the whole airfoil owing to the less stringent demands concerning
the grid resolution (xacoustic xLES /M, M = O(101 )). Thus, scattering at the leading edge can be
captured and directivities can be predicted.
that permits the coupling of the flow and acoustic simulation via appropriate source
terms. The source reconstruction from the compressible flow simulation involves only
simple numerical operations that allow a straightforward computation.
For the discretization of the acoustic perturbation equations, numerical methods
developed in the framework of CAA are applied. Section 6.5.2.2 discusses the methods used in the acoustic simulation step in their application to the trailing-edge noise
problems. Section 6.5.2.3 presents the simulation results for the flat plate at M = 0.15
and the transonic airfoil at M = 0.088.
6.5.2.2 Acoustic simulation step
Acoustic perturbation equations (APE)
The acoustic perturbation equations (APE), which have been proposed in Ewert and
Schroder (2003) see also Section 5.4 constitute an equation system for the pressure
and velocity perturbations ( p , u )T :

p
p
2

(6.32)
+ c0 0 u + 2 u 0 = c02 qc ,
t
c0

p
(6.33)
= q m .
u t + u 0 u +
0
The subscript 0 denotes mean-flow quantities (i.e., 0 , p0 , and u 0 are the density,
pressure, and the velocity, respectively, of the time-averaged flow). Furthermore,
297
13:36
P1: IBE
0521871441c06
298
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
c0 = p0 /0 is the sonic speed with the isentropic exponent. The right-hand-side

sources qc and q m will be defined in the discussion of the APE-based acoustic analogy.
The homogeneous equation system with the sources qc = 0, q m = 0 can be shown to
be equivalent to the convective wave equation for irrotational mean flows. To obtain the
wave operator the perturbation velocity u is split into an irrotational plus a remaining
part that contains all the vorticity:
u = + u r .
(6.34)
Because u r is not solenoidal, the decomposition becomes uniquely defined after the
additional condition is imposed that the unsteady pressure be expressed only in terms
of the unsteady acoustic potential by
p = 0
D0
,
Dt
(6.35)
with the substantial time derivative D0 /Dt = /t + u 0 . Introduction of Equations

(6.34) and (6.35) into Equations (6.32) and (6.33) yields the equivalent system for the
variables (, u r ):

1
(0 u r )
D0 1 D0
(
=
,
(6.36)
L =
0
Dt c02 Dt
0
0

u r t + (u 0 u r ) = 0.
(6.37)
The left-hand side of Equation (6.36) is the convective wave operator for an acoustic
potential . It is Pierces approximate wave equation (Pierce 1990), which was also
derived by Goldstein (1978) and recently used in Golubev and Atassi (1998) and Cooper
and Peake (2001). The wave operator is equivalent to the linearized wave operator of
Mohrings acoustic analogy (Mohring 1979, 1999), which reads

qtot
D0 1 D0 B
1

,
(6.38)
0 B =
LB =
2
Dt c0 Dt
0
0
where B denotes the perturbations of the total enthalpy, which is linked, with the pressure, to linear order via D0 B /Dt = 1/0 p /t with thermal conductivity and viscous
effects neglected. As discussed by Howe (1998), for homentropic, high-Reynoldsnumber flows, Howes acoustic analogy (Howe 1975) agrees with Equation (6.38).
Mohring (1999) showed that the wave operator L is stable for arbitrary mean flows.
Unlike the LEE, the entropy mode is a priori completely removed in the APE system
and the behavior of the vortical perturbations is governed by Equation (6.37). Taking
the curl of Equation (6.37) yields the vorticity equation
r t = 0;
(6.39)
that is, the vorticity convection property of the LEE is also removed. Because vortical
Equation (6.37) does not support growing instabilities and Equation (6.36) is stable,
the equivalent APE system does not possess instabilities. This is an important feature
if linear perturbations are forced with sources, because growing instabilities will cause
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
299
divergent solutions that are limited by neither nonlinear saturation nor viscous effects.
Instabilities appear in the LEE for unstable mean flows.
Although the homogeneous APE system describes wave propagation properly in
irrotational mean flows, even in mean flows with mild levels of mean vorticity the
homogeneous perturbation equation can resolve convection effects with good accuracy,
as was demonstrated in Ewert and Schroder (2003), who, to quantify these convection
errors, considered the propagation of acoustic waves of frequency f through a shear
layer of thickness 2 and velocity jump u 0 . Hence, in the shear layer a mean vorticity
of magnitude 0 u 0 / is apparent. Then the convection error is a function of a
Strouhal number defined by the mean vorticity 0 and acoustic frequency f (i.e.,
St = f /0 ). The irrotational case is defined by the high-frequency limit of the Strouhal
number tending to infinity (i.e., St = f /0 ). However, it was demonstrated that
the wave operator still resolves refraction effects for Strouhal numbers of order O(1)
properly (see also the discussion in Section 6.5.2.3 and Figure 6.60).
APE-based acoustic analogy
An acoustic analogy with dependent variables p and u can be found that recasts the
NavierStokes equations in nonlinear disturbance formulation such that the left-hand
side equals the APE system. The remaining terms lumped together on the right-hand side
constitute the sources (Ewert and Schroder 2003, 2004). The right-hand side sources
become

0 D 0 s
,
qc = u +
c p Dt
+ T s0 s T0
q m = ( u)

(u )2

+
,
2
(6.40)
(6.41)
where (. . .) := (. . .) (. . .) denotes perturbations of terms obtained by subtracting

from an actual term its time average, indicated by an overline. A major vortex source
term is the perturbed Lamb vector L = u , that is,

= 0 u u 0 u .
q m = L = ( u)
(6.42)
A similar vortex source term appears in the acoustic analogies of Powell (1964), Howe
(1975), and Mohring (1979). It can also be easily computed from a compressible unsteady flow simulation. The extension of the acoustic analogy concept from a scalar
wave equation to a set of linearized inhomogeneous equations was recently also discussed by Goldstein (2003).
13:36
P1: IBE
0521871441c06
300
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Ehrenfried, Meyer, and Dillmann (2003) integrated Mohrings (1999) acoustic analogy equation (6.38), numerically using the system
B t + u 0 B +
c02
c2
w
= 0 qs ,
0
0
(6.43)
w

+ 0 B = ( L)
+ u 0 w
t
(6.44)
with w
= ( u ) . For more detailed information about the right-hand-side sources, refer
to Ehrenfried et al. (2003). Therefore, Equations (6.32) and (6.33) and Equations (6.43)
and (6.44) constitute two formally different sets of perturbation equations for the same
wave operator equation (6.38). As was discussed by Mohring (1999), the wave operator
L of Equation (6.38) is self-adjoint. However, the APE can be shown not to have the
self-adjoint property. Hence, the adjoint equations differ formally from the APE but
must also encode the self-adjoint wave operator L of Equation (6.38). In what follows
to the
it is demonstrated that, by appropriately renaming the variables from B and w
adjoint variables of the APE, for instance ( pa , u a )T , the formal system equations (6.43)
and (6.44) give the adjoint equations to the APE, and vice versa. Furthermore, as was
discussed in Ewert, Schroder et al. (2002), the system equations (6.32) and (6.33) can
also be used to integrate Mohrings (1999) acoustic analogy equation (6.38). However,
unlike Equations (6.43) and (6.44), the system equations (6.32) and (6.33) are based
on the usual primitive variables ( p , u )T and hence immediately give exact solution to
the most important acoustic variable in the complete acoustic domain p .
Computational methods
The computational problem to solve for perturbation variables U = ( , u , p )T over

a mean flow field (0 , u 0 , p0 )T looks formally like
U
U

+ Ai
+ H U = S.
t
xi
(6.45)
In Equation (6.45), Einsteins summation condition holds (i.e., products of terms with
equal indices are summed over all spatial directions), and a source vector S appears
on the right-hand side. For the acoustic perturbation equations, Equations (6.32) and
(6.33), the matrices and source terms become, in 2D,
u 0i
qc
0 1i
0 2i
0
0
u 01 1i 1i /0
u 01 1i
, S = qm1 ,
Ai =
(6.46)
0
u 01 2i
u 02 2i 2i /0
qm2
0
p0 1i
u
0i
xi
H =
0
p0 2i
u 0i
0
x1
u 01
x1
0
x2
u 02
x1
u 01
x2
u 02
x2
0
c02
x1
0
c02
x2
c02 qc
12
x1
0
,
0
12
2
0

c02 xi
u 0i
c02
(6.47)
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.45. Coordinate system and nomenclature used to determine corrections for a 2D acoustic
simulation; R denotes a 2D polar radius vector in the x1 x2 plane, R = x3 the spanwise coordinate of
the 3D radius vector r = |
r |, and r = (x1 , x2 , x3 ) = (R cos , R sin , R)T .
where u 0i denotes the ith component of the mean flow u 0 and i j is the Kronecker
symbol with i j = 0 for i = j and i j = 1 otherwise. The components of the source
vector S are defined by Equations (6.40) and (6.41), where qmi denotes the ith component of q m . Note that the governing equation for the density is decoupled from the
remaining part such that the first equation could be dropped. The spatial discretization is accomplished with the fourth-order DRP scheme of Tam and Webb (1993).
For use on curvilinear multiblock grids, Equation (6.45) is extended with a metric, where the metric terms are also computed consistently with the DRP scheme.
At far-field boundaries just the asymptotic radiation boundary condition of Tam and
Webb (1993) is used because, even at outflow boundaries, no nonacoustic convection modes are apparent. The solid-wall boundary condition is implemented based
on a ghost point beneath the surface that yields the proper surface-wall normal pressure gradient (Tam and Dong 1994). For the suppression of high-frequency spurious waves, artificial selective damping (according to Tam and Dong 1993) has been
used. The time integration is carried out with the fourth-order, alternating, two-step,
low-dissipation, and low-dispersion RungeKutta scheme (LDDRK56) proposed by
Hu et al. (1996).
Determination of volume sources
The volume sources are computed from Equation (6.42) expressing the velocity fluctuations and the fluctuating vorticity through the LES solution. The acoustic simulations
are carried out in 2D using only the x1 and x2 components of the source term. As was
discussed in Manoha et al. (1999), Oberai, Roknaldin, and Hughes (2002), Casper and
Farassat (2003), and Ewert et al. (2003), for cylindrical geometries of constant 2D cross
sections and if one considers observer positions just in the spanwise, symmetric x1 x2
plane (see Figure 6.45), the 3D problem can be reduced to an equivalent 2D problem.
According to the findings of Manoha et al. (1999), Casper and Farassat (2003), and
301
13:36
P1: IBE
0521871441c06
302
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Oberai et al. (2002), only the zeroth spanwise wave number of the spanwise, Fouriertransformed source needs to be considered. Oberai et al. (2002) and Ewert et al. (2003)
used a reduced 2D source in a pure 2D simulation to correct the sound-pressure levels
later from 2D (decay R 1/2 ) to 3D (decay R 1 ). Casper and Farassat (2003) utilized only the zeroth spanwise wave-number component of the source for a 3D acoustic
simulation. The latter approach allows the spanwise acoustic grid resolution to be determined by the acoustic rather than the turbulent length scale; hence, the full wing span
L is resolvable in the acoustic simulation step. This avoids, furthermore, the necessity
of an additional 2D-to-3D sound-pressure-level correction.
However, note that, in general, for the flow and acoustic problem, respectively, different demands concerning the spanwise grid resolution and extension exist that are
naturally taken into account with a hybrid prediction method. In particular, consider the
trailing-edge problem resolved with a direct method that combines both the LES and the
acoustic simulations. In the LES the spanwise extension of the computational domain
is typically only on the order of the turbulent length scale, and periodic boundary conditions are applied to reduce the computational cost to an acceptable level. Because in
low-Mach-number flows the turbulent length scale is significantly smaller than the
acoustical length scale, applying periodic boundary conditions for the compressible
problem means to simulate an unphysically correlated acoustic source in the spanwise direction, which yields radiated acoustic waves that decay R 1/2 . Physically,
the waves generated by one LES slice of spanwise width decay R 1 . The spanwise, unphysically correlated acoustic signal was observed by Manoha, Delahay, Sagaut
et al. (2001) by analyzing the acoustic signal contained in the LES of a trailing-edge
flow. The correlated source causes an overprediction of the sound-pressure level (SPL).
Kato et al. (1993) stated the SPL to be about 14 dB too high. This issue was also observed by Wang and Moin (2000). To recover the proper physical decay, one could use
an acoustic computation with absorbing boundaries in the spanwise direction. However, a computational acoustic domain of small spanwise width yields acoustic waves
that leave the sidewise boundary surfaces at very shallow angles. Most nonreflecting
boundary conditions will not work properly under these conditions (i.e., the reflection
causes a false decay law of the sound waves). On the other hand, if the full wingspan
L is considered, the spanwise resolution limited by the flow length scale restricts the
computational problem because of high memory requirements.
Because an integral length scale of the acoustic source in the spanwise direction
is small compared with the acoustic wavelength for small Mach numbers, the source
is compact in the spanwise direction; hence, an acoustic simulation with a spanwise
extension and periodic boundary conditions is equivalent to a 2D acoustic simulation
with a spanwise-averaged acoustic source

1 /2

q d x3 .
(6.48)
q =
/2
Here, q indicates a component of the spanwise-averaged source, and
q is a component
of the source from the full 3D LES simulation. If the source equation (6.48) is used
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
303
in a 2D acoustic simulation, the sound-pressure correction from 2D to 3D becomes,

according to Ewert et al. (2003),

1 i k2

.
(6.49)
p (0, R, , ) p(R,
, )
2
R
In Equation (6.49),
p ( R, R, , ) denotes the 3D, and p(R,
, ) the 2D,
frequency-related Fourier transform of the sound pressure. Furthermore, = 2 f
is the angular frequency, k = /c0 the wave number, and i = 1. Hence, the correction is frequency dependent and only applies in the x1 x2 plane, where R = 0 (see
Figure 6.45, which also defines all other appearing quantities). The result agrees with
the findings of Oberai et al. (2002) that were derived for a vanishing mean flow. In
addition, the derivation of Equation (6.49) also incorporates the convection effect of a
uniform low-Mach-number mean flow. The correction does affect the final spectral SPL
distribution; however, it has no impact on the -dependent directivity, obtained from a
2D simulation. Furthermore, the mean-flow amplification remains the same. In other
words, the 2D simulation delivers directivities and Doppler mean-flow amplifications
that are also valid in the x1 x2 plane for the 3D case. From Equation (6.49) the SPL
correction can be deduced to be

f 2
S P L 3D, = S P L 2D + 10 log
.
(6.50)
Rc0
Apparently the 2D sound-pressure spectrum is shifted 3 dB/octave toward the higher
frequencies by the correction. The 3D-corrected pressure S P L 3D, of Equation (6.50)
is the sound radiated by one slice of width . For a finite spanwise extension L (Figure
6.45), an additional correction

L
(6.51)
S P L 3D,L = S P L 3D, + 10 log

has to be used. This correction was given by Kato et al. (1993) and is based on the
assumption that all slices of width along the wingspan are uncorrelated, whereas
the spanwise extension L is small compared with R. The correction also only affects
the final spectral SPL distribution but leaves the directivity unchanged.
On the basis of Taylors hypothesis of convecting vorticity, the nondimensionalizing
of Equation (6.49) yields, with k = /c0 , u 0 /v , where v is the length scale of
a vortical disturbance, the scaling

p M 1/2 p.
In other words, the correction Equation (6.49) serves also as the correction of the powerscaling law (i.e., the sound intensities in 3D scale with the power of the Mach number
increased by one compared with the 2D case).
In general the computational domain of the LES is a small subdomain inside the
computational domain of the acoustic simulation (see Figure 6.44). Because the vanishing and generation of vorticity are always related to a strong noise-source mechanism,
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
304
0 521 87144 1
printer: sheridan
November 20, 2006
x2
u0
W 1(x1)
0.75
LES subdomain
0.5
l1
x1c
l2
0.25
0
x1
l1
Plate
x1
l2
Figure 6.46. LES subdomain at the trailing edge (left) and horizontal weighting function (right).
the simple truncation of the vorticity-source-term Equation (6.42) at the boundaries

of the LES domain in the mean-flow direction will cause spurious noise sources in
the acoustic simulation. The spurious sound generated owing to the termination of the
source of Lighthills (1952) acoustic analogy was discussed by Wang, Lele, and Moin
(1996) for a vortex-shedding airfoil. The problem was also addressed by Crighton
(1991, 1993). Furthermore, the same difficulty was observed by Mankbadi et al. (1994)
and Mitchell et al. (1999) in jet noise computations. Mitchell et al. (1999) used slowly
decaying Lighthill source terms at the outflow boundaries to damp the spurious sound
source.
For the APE source-term equation (6.42) the spurious noise-source mechanism was
discussed by Ewert, Meinke, and Schroder (2002). It was shown that the spurious noise
sources can be avoided by smoothly switching the sources on and off inside the resolved
source region in the mean-flow convection direction. Because the vortical structures are
floating for the flat-plate problem mainly in the x1 -direction parallel to the plate, this is
accomplished applying a spatial weighting function W1 (x1 ) in that direction according
to Figure 6.46.
Ewert, Meinke, and Schroder (2002) demonstrated that a low-pass behavior for
spurious sound is obtained, which is a function of the nondimensional wave number
d = 2 d/v , where, according to Figure 6.46, d = (l2 l1 )/2 is the width of the
on- and offset zone and v is the wavelength of a vortical disturbance (related to an
acoustic wavelength of same frequency via v = M ). Figure 6.47 depicts the lowpass behavior for a Hann window-weighting function. As a rule of thumb, d should be
chosen larger than the largest vortical structure v that is related to the lowest resolved
frequency via f = u c /v with the convection velocity u c 0.66u 0 .
Altogether, the nonzero components of the vortex source vector S = (0, qm1 ,
qm2 , 0)T of Equation (6.45) are determined for the 2D acoustic simulation by the first
two components of
1
q m (x1 , x2 ) =
/2
/2
[(
x ) u (
x )] W1 (x1 )d x3 .
(6.52)
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
305
Figure 6.47. Damping function |F (d)| over vortical

wave number in x1 -direction = 2/v scaled with
the damping zone on- and offset width d = (l 2 l 1 )/2.
Damping |F(d)|
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
10
15
6.5.2.3 Simulation of trailing-edge noise

Large-eddy simulation I: Flat plate
An LES is performed for the compressible flow over the trailing edge of the flat
plate. The discretization of the convective fluxes of the NavierStokes equations is
based on a second-order accurate advective upstream-splitting method, described in
Liou and Steffen (1993), with a central approximation for the pressure derivative. The
viscous stresses are discretized using central differences of second-order accuracy.
Furthermore, an explicit five-step RungeKutta time-stepping scheme of second-order
accuracy is used for the temporal integration. The coefficients are chosen to maximize
the stability of a central scheme. The scheme is described in more detail in Meinke
et al. (2002). The Reynolds number based on the length of the plate is Re = 7 105 ,
and the freestream Mach number is M = 0.15. The flat plate possesses a thickness
d = 0.175%. To reduce the computational effort while capturing the essential physics,
numerical simulations are conducted in a domain that contains only 25% of the total plate length. Figure 6.48 shows a plane of the LES grid. Seventeen points in the
-5
-10
10
Figure 6.48. LES grid with partially resolved plate and every second grid point shown; dimensions are
related to the inflow boundary layer 0 , which scales with 1/52.228 related to the plate length l (i.e.,
the complete plate extends from x = 52.228 to 0.0).
20
13:36
306
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
2
1.5
1
0.5
P1: IBE
0521871441c06
-0.5
-1
-1.5
-2
-2
-1
Figure 6.49. Acoustic grid scaled with the plate length l (i.e., plate extends from x = 1.0 to 0.0).
Every fourth grid point is shown.
spanwise direction have been used to yield a grid with 2.22 106 points. The spanwise
extension is 0.640 . The boundary layer thickness at the inlet is 0 /l = 1/52.228. The
instantaneous inflow data are generated via an auxiliary LES of an adiabatic flat-plate
boundary layer using a compressible rescaling method. More details can be found in
El-Askary et al. (2001).
Large-eddy simulation II: Airfoil
An LES for the flow over a complete 3D airfoil is performed using a block-structured
mesh with a C-type inner block and an O-type outer block. The free-flow Mach number
is M = 0.088. The Reynolds number based on the chord length is Re = 8 105 ;
therefore, a spanwise extension of the mesh of 0.3 percent chord suffices. Experience
has shown that an LES for the flow over an airfoil requires a resolution of the nearwall cells in inner law scaling in the range of x + 100, y + 2, and z + 20,
where x, y, z denote the streamwise, normal, and spanwise coordinates, respectively.
This requirement is approximately satisfied over the whole airfoil. The complete mesh
contains 7,323,651 grid points, 1686 of which are distributed on the airfoil 373 in
the wake, 197 in the normal, and 17 in the spanwise direction. Capturing the laminar
turbulent transition, which is known from experiments to occur just downstream of the
nose region, was accomplished by clustering the mesh in the area 0 x 0.2 (Figure 6.50). Furthermore, a high resolution was generated near the trailing edge to ensure
that the vortical structures in the boundary layers on the upper and lower surface and
in the wake were well captured because these structures determine the acoustic source
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
307
Figure 6.50. Enlargement of the leading- (left) and trailing-edge region (right).
terms of the acoustic perturbation equations and as such the overall acoustical field.
On the right of Figure 6.51 the vortical structures in the aft region of the airfoil and in
the wake are visualized. The streaky structures in the near-wall region are elongated
near the trailing edge, and spanwise vortices form and grow in the wake region. The
left of Figure 6.51 shows the vortical structures in the suction-side boundary layer. The
occurrence of multiple streaky structures in the spanwise direction indicates the LES
to have a sufficiently large spatial extension in the spanwise direction. More details can
be found in Ewert et al. (2003).
Acoustic simulation I: Flat plate
The acoustic grid is shown in Figure 6.49. The flat plate extends from 1 to 0, and
the finite thickness is resolved. The grid is clustered in the wake of the plate to resolve
the source term properly. Computing the acoustic source term equation (6.52) entails
sampling 503 time levels of the LES with a time increment of t = 4 103 (reference
time tref = l/c , plate length l). Using 10 points per period is sufficient to resolve
X
Z
Figure 6.51. Visualization of the instantaneous flow field: Coherent structures on the airfoil surface
(left) and vortex structures in the wake (right).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
308
0 521 87144 1
printer: sheridan
November 20, 2006
-1.0E-03 -6.0E-04 -2.0E-04 2.0E-04
6.0E-04
1.0E-03
0.03
0.02
0.01
0.01
0.02
0.03
0.04
0.02
-2.0E-02 -1.2E-02 -4.0E-03 4.0E-03
0.02
0.04
Figure 6.52. APE source terms L = [

u ] =
(L x , L y )T , L x (top), L y (bottom), and CAA grid.
1.2E-02
2.0E-02
0.03
0.02
0.01
0
0.01
0.02
0.03
0.04
0.02
0.02
frequencies up to 42 kHz, whereas the lowest resolution is, according to the number of
time levels, on the order of 800 Hz. Owing to the fine grid resolution of the LES close
to the plate surface, the temporal resolution of the LES is much finer such that only
every 160th time step is used.
Figure 6.52 shows snapshots of the source terms that follow from Equation (6.52).
The quantitative comparison evidences that the main contribution to the source term
comes from the y-component. The visible vortical structures mainly convect in the
downstream direction, where the temporal resolution of the acoustic simulation is sufficient to resolve the source term properly. Note that the source based on Lambs vector
is small close to the wall owing to the no-slip boundary condition where the LES has
its finest resolution. The upper-half planes in Figure 6.52 show the acoustic grid in the
vicinity of the plate. The acoustic grid possesses 2 105 points, 17,585 of which are
located in the LES domain. The source term is computed on the LES grid and then
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
309
p
-1E-06
1E-06
2
1.5
1
0.5
P1: IBE
0521871441c06
-0.5
-1
-1.5
-2
-2.5
-2
-1.5
-1
-0.5
0.5
1.5
Figure 6.53. Pressure contours of the trailing-edge problem M = 0.15, Re = 7 105 , at time level
T = 3.0 using APE solution equations (6.32, 6.33).
bilinearly interpolated onto the acoustic grid. Entropy fluctuations are neglected in the
source computation.
The source term is ramped up and down at the interior LES boundaries by applying
a weighting function to avoid spurious sound sources. According to the discussion of
Section 6.5.2.2 and Figure 6.47, the spatial extension of the ramping function must be
on the order of the wavelength of the vortical disturbances. Because the LES domain is
enclosed by a C-shaped sponge layer (Figure 6.48), a natural damping zone is obtained
at the outflow boundary where the source will decay to zero. At the inflow boundary
the computed source is weighted over a width d = 0.15.
Figure 6.53 shows a snapshot of the pressure field that evidences acoustic waves
generated at the trailing edge. No spurious sound sources (e.g., from the LES inflow
boundary located at x = 0.25) are visible. Figure 6.54 shows pressure contours for
the same time level computed on a finer grid with a grid refinement factor of 1.5 for each
coordinate direction. The contours of the coarse and the fine grid simulation evidence
a grid-independent solution. Figure 6.55 shows the directivity for four different frequencies and a circle centered at the trailing edge with radius r = 1.5. The directivities
agree qualitatively very well with the shapes of the approximate harmonic, noncompact
edge-noise Greens function according to Howe (2001). In the compact limit, the
product of wave number and chord length tends to zero (i.e., kl 0). In this limit, the
directivity recovers a dipole eight-characteristic sin (). For the opposite case with
13:36
310
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
p
-1E-06
1E-06
1.5
0.5
P1: IBE
0521871441c06
-0.5
-1
-1.5
-2
-1.5
-1
-0.5
0.5
Figure 6.54. Pressure contours of the trailing-edge problem M = 0.15, Re = 7 105 at time level
T = 3.0 with APE solution equations (6.32, 6.33) on acoustic grid refined by a factor of 1.5 in each
coordinate direction.
kl , the ratio of chord length to wave length tends to infinity (i.e., the cardioid directivity of the semi-infinite flat plate sin(/2)) follows. With increasing wave numbers,
a multilobe pattern appears whose envelopes converge against the cardioid shape. Figure 6.56 compares the directivities from the APE system equations (6.32) and (6.33)
210
1.510
110
510
kl=3.072
kl=9.215
kl=15.358
kl=21.502
210
110

2 2 that is,
Figure 6.55. Directivity 1/2 (, r, f ) with =
nondimensional PSD of p 2 (, r )/ c

2
2
2
0 (, r, f )d f = p (, r )/ c ; APE simulation for r = 1.5, origin at the trailing edge, for various
nondimensional wave numbers.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
311
210
1.510
110
510
p, kl=9.215
p, kl=15.358
B, kl=9.215
B, kl=15.358
210
110
Figure 6.56. Comparison of the trailing-edge noise directivities 1/2 (, r, f ) for r = 1.5 applying Equations (6.32) and (6.33) ( =
PSD of p 2 ) and Mohrings
(1999) acoustic analogy ( =

PSD of B 2 ),
Equations (6.43) and (6.44).
with those based on the perturbation system equations (6.43) and (6.44), Section 6.5.2.2,
for the acoustic analogy of Mohring (1999). Note that the acoustic variable of the acoustic analogy of Mohring (1999) that is, the total perturbation enthalpy B is related
in the far field to the perturbation pressure via p 0 B . The directivities of both
methods agree qualitatively and quantitatively well.
Figure 6.57 depicts the PSD in three receiving points with a distance r 1.5 to the
trailing edge. The three receiving points are related to the polar angles = 45 , 90 ,
and 135 , respectively, where all angles are defined similarly to those of Figure 6.45.
The SPL and dimensional frequencies are computed based on the chord length of 0.2 m
70
o
60
= 90
= 135
S P L [dB]
50
= 45
40
30
20
10
10
f [kHz]
15
20
Figure 6.57. Sound-pressure level (SPL) versus frequency for a receiving point in r = 1.5 above the
trailing edge for various directions (see Figure 6.45).
13:36
P1: IBE
0521871441c06
312
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
and ambient thermodynamic properties. The dimensional pressure levels are related to
a reference pressure of pref = 2 105 Pa such that the SPL follows from

peff
(6.53)
, p0dB = 2 105 Pa,
S P L[dB] = 20 log
p0dB
with peff := ( f )1/2 , where (, r, f ) is the dimensional PSD of p 2 , and f =
1 Hz the dimensional unit frequency. The definition of the PSD used here is based on
the dimensional single-sided PSD

2 3
2

p (, r ) =
(, r, f ) d f =
cl (, r, f ) d f .
(6.54)
0
In Equation (6.54), denotes the nondimensional, single-sided PSD that follows

2
and f ref = c /l as reference pressure
from an acoustic simulation using pref = c
and frequency, respectively, for nondimensionalizing with the subscripts indicating
undisturbed flow properties. Hence,
2
( f )1/2
peff = c
(6.55)
in Equation (6.53) with f = f / f ref . The spectral distributions evidence the decline
of the SPL over the frequency band.
Acoustic simulation II: Airfoil
The LES data sample for the airfoil in a subsonic flow contains 350 discrete time levels
that cover a nondimensional time interval T = 2.975, which is nondimensionalized with
the sonic speed and the chord length as references. With a temporal resolution of 20
points per time period, the data sample is sufficient to resolve physically frequencies up
to 5 kHz based on a chord length of 0.4 m, which was chosen as the high-frequency limit.
Furthermore, the time interval yields a frequency increment of T 1 (i.e., physically
f = 350 Hz). However, because it needs about three time units to propagate the initial
disturbances out of the CAA domain plus at least one data sample period to sample
results (e.g., for directivities), the source data sample has to be fed periodically into the
acoustic simulation. To avoid the strong spurious sound sources that appear because
the sample data ends do not match smoothly, the samples need to be manipulated to
prevent spurious sound sources without losing essential physical information. Because
spurious sound sources can be avoided by smoothly ramping up and down the timedependent sources, the source data sample has been weighted with a window function.
The -window depicted in Figure 6.58 evidences the least-steep ramp function to
appear if the whole half-data period is used to switch on the source. However, for the
)/2 is used. In addition
airfoil source computation a Hann window N (t) = (1 cos 2t
T
to the temporal data, windowing is added to one weighted data sample with the same
weighted sample time shifted by T /2 to yield finally the weighted source (e.g., for the
first half-period),

T
T
N t+
, 0 t T /2.
q(t) = q(t) N (t) + q t +
2
2
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
313
0.4
0.3
N(t)
0
( u)y
0.2
0.1
0
-
0.1
0.2
0.3
0.4
0
0.5
1.5
2.5
3.5
4.5
5.5
t, T=2.97518
0.4
0.3
t
T/2
T/2
( u)y
0.2
f(t)
0.1
0
-
0.1
0.2
0.3
-0.4
0
t, T=2.97518
Figure 6.58. Generation of a periodical source term via window weighting.
Figure 6.58 depicts on the right a comparison between unweighted and weighted sources
with the Hann window applied (the left-hand-side sketch shows, for illustration, a
Bartlett window). One can easily show that all source components that are 2n/T periodical remain unchanged by the filtering based on the Hann window. Figure 6.58
also sketches the 2/T periodical source obtained from the filtering; hence, the filtering
doubles the lowest resolved frequency as well as the frequency increment. With the
periodic source, after a transient period in any receiving point a correspondingly periodical signal is obtained that can be exploited to compute Fourier transforms (i.e., no
additional windowing is necessary to compute sound pressure spectra).
Spurious noise at interior boundaries is prevented by switching the source term of
the airfoil simulation on and off over a region of 0.4 chord length in accordance with
the lowest resolved frequencies.
Figure 6.59 shows details of the curvilinear multiblock mesh used for the acoustic
simulation of the transonic airfoil case. The computational domain is resolved with a
Figure 6.59. CAA grid.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
314
0 521 87144 1
printer: sheridan
November 20, 2006
2
1.5
p
0.0010
0.0009
0.0008
0.0006
0.0005
0.0004
0.0003
0.0002
0.0001
-0.0001
-0.0002
-0.0003
-0.0004
-0.0005
-0.0006
-0.0008
-0.0009
-0.0010
0.5
0
-0.5
-1
-1.5
-2
-1.5
-1
-0.5
0.5
1.5
2.5
0.00035
0.0003
Figure 6.60. A harmonic test source over the

trailing edge (top) and directivities obtained
for M = 0, 0.088 with LEE and APE (bottom)
applied.
LEE, M=0
LEE, M=0.088
APE, M=0.088
0.00025
0.0002
0.00015
1E-04
5E-05
0
-5E-05
-0.0001
-0.0002
-0.0001
1E-04
0.0002
10-block mesh that contains 86,902 grid points. Because, for acoustic simulations, an
equidistant Cartesian grid is optimal, an H-topology in the far field has been chosen in
combination with a C-grid topology around the wing section. The topology yields two
singular points where the continuation of the grid lines is not unique. It was found that
those points only cause strong spurious signals in combination with convecting vorticity
passing through them. However, vorticity waves are a priori avoided when applying the
APEs. The grid was chosen to have a resolution of 10 points per wavelength for the
highest resolved frequency of 5 kHz based on a chord length of 0.4 m.
The acoustic simulation was carried out with the time-averaged mean flow from the
LES simulation that in this case covers the whole acoustic domain. Figure 6.60 shows the
sound field generated by a pointlike harmonic test source with an equivalent frequency
of 5 kHz that was placed above the trailing edge. The simulations were carried out
with and without viscous mean flow applying the LEE and the APE. The directivities
depicted on the top of Figure 6.60 evidence that the APE with mean flow delivers
the same directivities as the LEE simulation for nonuniform mean flows. Interestingly,
the simulations with a mean flow with a Mach number of just M = 0.088 show clear
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
315
uy)
0.2
-0.2000 -0.0667 0.0667
0.15
0.2000
0.1
0.05
0.05
0.1
0.9
Figure 6.61. APE source term (

u ) y (top) and sound
radiation from the trailing edge (bottom).
1.1
1.2
2
1.5
1
0.5
P1: IBE
0521871441c06
0.5
1
1.5
2
differences compared with the case without mean flow. That is, even at very small Mach
numbers convection effects seem to be important that are not described by methods
based on the simple wave equation. Figure 6.61 (top) presents snapshots of the APE
source term close to the trailing edge. The sound radiated from the trailing edge is
depicted on the bottom panel of Figure 6.61. Analogous to the flat plate case, no waves
are radiated into the up- and downstream direction. Owing to the nonuniform mean
flow, the acoustic lobes on the airfoil pressure side are more pronounced than that on
the suction side.
6.5.3 Trailing-edge noise simulation using LES, Euler equations,

and the Kirchhoff integral
6.5.3.1 Objective
The objective is to simulate the noise radiated by an NACA-0012 airfoil with a blunted
trailing edge. The proposed strategy combines three different techniques.
13:36
P1: IBE
0521871441c06
316
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
A near-field compressible LES accurately simulates the local aerodynamic noise

sources and is able to provide the acoustic field generated at a very short distance,
including refraction effects through the inhomogeneous unsteady flow and scattering
effects on solid surfaces.
Then this LES local solution is coupled to a Euler solver in a midfield domain. LEScomputed perturbations are directly injected into the Euler domain, in which they propagate outward through the nonuniform mean flow, naturally accounting for refraction
effects in regions where mean velocity gradients are significant. The radial extent of
this Euler domain is set such that the mean flow is rather uniform beyond this limit.
Finally, the acoustic field radiated at the external boundary of the Euler domain
becomes the entry data of a Kirchhoff integration, which provides the noise radiated in
the far field.
6.5.3.2 Large-eddy simulation
This section is devoted to the unsteady flow simulation via LES. First, the main features of the LES are recalled. Then the airfoil geometry and the computational grid
are presented as well as the numerical procedure and storage method. Several basic
aerodynamic results are given at the end of this section.
Code implementation
The LES method described below has been implemented in the FLU3M solver, an
industrial CFD software that has been developed at the Office National dEtudes et
de Recherches Aerospatiales (ONERA) for several years. This solver is based on the
discretization of the compressible NavierStokes equations on multiblock, structured
meshes by a finite volume technique. A second-order accurate, implicit temporal integration is achieved thanks to an approximate Newton method.
Filtered NavierStokes equations
These equations are detailed in Chapter 5. A dimensionless form of the NavierStokes
equations is retained, which means that the local instantaneous velocity

u , tempera
ture T , density , abscissa x , dynamic viscosity , and thermal conductivity are

normalized, respectively, by U0 , T0 , 0 , C, 0 , and 0 . The three-dimensional, unsteady, filtered NavierStokes equations are used in conservative form for a viscous,
compressible Newtonian fluid. Any flow variable can be written as = + ,
where represents the large-scale part of the variable and its small-scale part. The
Favre filtering operator = /, classicaly defined as a convolution product on the
computational domain (Sagaut 2001), is assumed to commute with time and spatial
derivatives. The eddy viscosity must be expressed by an SGS model. The equations are
supplemented by a filtered equation of state.
Selective mixed-scale model
The selective mixed-scale model, developed by Sagaut and Lenormand, has been
retained, because it realizes a good compromise between accuracy, stability, and
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
computational cost (Lenormand, Sagaut, and Ta Phuoc 2000; Lenormand, Sagaut, Ta

Phuoc, and Comte 2000). Moreover, the use of a selective function allows this model
to handle transitional flows, which is one of the key points of the present application.
This selective mixed-scale model is detailed in Chapter 5.
Numerical method
The NavierStokes equations are discretized using a cell-centered finite volume technique and structured, multiblock meshes. The viscous fluxes are discretized by a secondorder accurate centered scheme. For efficiency reasons, an implicit time integration is
employed to deal with the very small grid size encountered near the wall. Indeed the
large disparity between the acoustic wave speed and the advection velocity at low Mach
numbers renders explicit time integration inefficient. This occurs because numerical
stability considerations impose small time steps on the acoustic waves, whereas the
physics is mainly driven by the solenoidal part of the flow, whose time scale, being associated with advection, is larger. Then a three-level backward differentiation formula
is used to approximate the temporal derivative of Q c , leading to second-order accuracy.
An approximate Newton method is employed to solve the nonlinear problem. At each
iteration of this inner process, the inversion of the linear system relies on the lower-upper
symmetric GaussSeidel (LU-SGS) implicit method originally proposed in Yoon and
Jameson (1987). More details about these numerical points are available in Mary and
Sagaut (2001) and Weber and Ducros (2000). Usually LES requires high-order centered
schemes for the Euler fluxes discretization (with spectral-like resolution) to minimize
dispersive and dissipative numerical errors. However, such schemes cannot be applied
easily in complex geometry. Indeed, most of the aerodynamic codes able to deal with
such a geometry are based on a finite volume technique in order to handle degenerated
cells. Thus, getting a high-order method becomes very time consuming owing to the to
high-order quadrature needed to compute the fluxes along the cell boundaries. Because
several works, for instance Wu et al. (1999), have shown that LES can be carried out
with a low-order centered scheme in case of sufficient mesh resolution, only a secondorder accurate scheme is employed in this study. However, a special effort has been
carried out to minimize the intrinsic dissipation of the scheme. Thus, the second-order
accurate hybrid upwind-centered discretization developed in Mary and Sagaut (2001)
has been used in this study to achieve a good compromise between robustness and
accuracy. The key point of the scheme is the use of a sensor that allows some numerical dissipation to be introduced locally when a numerical wiggle is detected on one
of the primitive variables. Therefore, the effect of the SGS model is not influenced
by the numerical dissipation as long as oddeven decoupling is not detected in the
flow solution.
Geometry, computational grid, and numerical procedure
A two-dimensional (constant section in the spanwise direction) NACA-0012 airfoil

with a C = 0.6096 m chord and a blunted trailing edge (TE) of thickness H = 2.5 mm
317
13:36
P1: IBE
0521871441c06
318
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
(0.4% of the chord) has been retained for this study with reference to the airfoil noise
experiment conducted at NASA in 1980 by Brooks and Hodgson (1981). In this experiment, the airfoil had a span S = 0.46 m (or 75% of the chord), whereas in the
present simulation the LES computational domain has a spanwise extent representing
only 3.3% of the chord. The upstream flow velocity is 69.45 m/s (maximal velocity
in Brooks and Hodgson 1981), the Mach number is 0.205, and the Reynolds number
based on upstream velocity and chord is 2.86 million. The airfoil angle of attack with
respect to the flow direction is 5 . The 3D curvilinear computational grid is obtained by
replication in the spanwise direction ( y) of a 2D curvilinear structured grid made of two
domains. Domain 1 is located upstream from the TE (C-shape), with 309 points along
the airfoil body and 97 points in the radial direction. Domain 2 is located downstream
from the TE with 227 points in the z direction (including 35 points on the TE bluntness)
and 103 points in the x direction. The influence of boundary conditions is minimized by
using large computational domain. As shown in Figure 6.62, this 2D grid extends nearly
10 chords above and below the profile as well as upstream and downstream. Regions in
the vicinity of solid walls are highly refined. The whole mesh is made of 1.76 million
points. The horizontal plane z = 0 is a plane of symmetry. In any (x, z) plane, the
smallest cells are located at the TE corners with dimensions (in wall units) z + = 1.5
in the direction normal to the wall and x + = 15 along the chord. From the corners,
the grid is streched in the streamwise direction (x) with a stretching coefficient equal to
1.06 (upstream) and 1.09 (downstream). In the z direction, the stretching coefficients
are equal to 1.23 toward the z = 0 plane and 1.12 toward the grid periphery.
This 2D grid is replicated 32 times in the spanwise direction (y) with a constant
step y = 103 C. A no-slip condition is applied at the airfoil surface and a periodic
conditon is imposed in the spanwise direction. Nonreflecting characteristic boundary
conditions are applied for the far field. Moreover, a steady RANS computation using
BaldwinLomax models provides an initial flow solution. The time accuracy of the
results is ensured by taking the physical time step equal to 0.5 s, meaning a sampling
frequency of 2 MHz. An initial phase of 100,000 time steps was achieved; then the useful
computation was performed over 130,000 time steps with a total duration of 65 ms,
representing the convection of the flow over 7.5 airfoil chords. The requirements of
acoustic computations led to a storage of one time sample every 100th LES time step,
meaning a storage sampling of 20 kHz, with a useful frequency band of 10 kHz. The
total computation cost was 360 CPU hours on an NEC-SX5 computer.
Aerodynamic results
In this section, the aerodynamic field computed via LES near the airfoil is analyzed in
detail. This field is dominated by turbulent structures in the airfoils boundary layers
and wake. At small distance of the airfoil, the pressure field is mostly acoustic and
is generated by the acoustic scattering at the airfoil TE of (i) the turbulent boundary
layers convected on both sides (broadband noise) and (ii) the alternated vortex shedding
generated by the TE bluntness (narrow-band component). Owing to the stretching of
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.62. Computational grid.
the LES computational grid, which acts as an acoustic low-pass frequency filter, this
acoustic field cannot radiate farther than a half-chord from the body.
Figure 6.63 displays contours of instantaneous Mach-number isovalues showing
the development of the turbulent boundary layers and the wake. The transition between laminar and turbulent flows occurs at x/C = 0.15 on the suction side and at
x/C = 0.9 on the pressure side. Mean velocity profiles near the TE were obtained by
time-averaging instantaneous data. The integration of these velocity profiles provided
turbulent boundary layer (TBL) momentum thicknesses of 0.8 mm on the pressure
side and 1.5 mm on the suction side. Brooks and Hodgsons (1981) measurement on a
similar airfoil was slightly higher (3.9 mm), which could be explained by the roughness
strips of random carbon they used to trip the boundary layer artificially at x/C = 0.15.
319
13:36
P1: IBE
0521871441c06
320
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.63. Contours of instantaneous Mach-number isovalues.
It is thought that the excessive slenderness of the computed TBL may also be
explained by an insufficient grid refinement in the chord direction x near the walls in
the transition region. However, it will be shown in the next section that the computed
flow presents general qualitative and quantitative features that make it suitable for
aeroacoustic applications. Figure 6.64 shows details of the instantaneous (above) and
time-averaged (below) flow streamlines around the blunted trailing edge. Time-averaged
results have been favorably compared with the results of the mean flow computed
separately using a RANS code.
6.5.4 Unsteady pressure-field analysis

6.5.4.1 Surface-pressure fluctuations
Time histories
Figure 6.65 (top) shows time histories of the wall-pressure fluctuations near the TE
at symmetrical points located on the pressure and suction sides. This plot exhibits a
time period of 2 ms, meaning a dominant mechanism at a frequency near 5 kHz, and a
clear phase opposition between both sides. This opposition is confirmed by the crossspectrum phase between both signals shown in Figure 6.65 (bottom). These features
are typical of a mechanism of alternated vortex shedding at the TE.
Power spectral densities
Figure 6.66 shows the evolution along the chord (from the leading edge up to the TE)
of the wall-pressure PSD (010 kHz) on the suction side (above) and the pressure side
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.64. Flow streamlines at the trailing edge: instantaneous (above) and time-averaged
(below).
(below). Spectra are characterized, on both airfoil sides, by a narrow-band component

emerging out of a wideband continuum. Located near 4.7 kHz, the narrowband bump
is due to the vortex-shedding mechanism. It is perceived on the whole chord length
with identical levels on both airfoil sides. It will be shown later that this component
corresponds to acoustic waves generated at the TE and propagating in the upstream
direction. The wideband continuum is generated by the convected turbulent boundary
layers. Because the airfoil has a positive 5 angle of attack, the laminar-to-turbulent
transition occurs at 15% of the chord on the suction side and at 90% of the chord
on the pressure side. From this transition point toward the TE, the boundary layer
thickness grows, enriching spectra with higher-frequency components. Notice that, on
both airfoil sides, the vortex-shedding component emerges from the continuum near the
TE (x/C > 0.98), then is masked by the near-field boundary layer, and finally emerges
again at x/C < 0.85.
Comparison with experimental data
In their airfoil noise experiment (performed with the same velocity, tripped TBL, but
zero incidence), Brooks and Hodgson (1981) measured surface pressure with arrays
321
13:36
P1: IBE
0521871441c06
322
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.65. Wall-pressure fluctuations on pressure and suction sides near the TE. Above: time histories; below: cross-spectrum phase.
of pinhole sensors of 0.34 mm diameter. The closest distance between sensors and the
TE was at 2.54 mm. Figure 6.67 compares the spectrum measured by this sensor with
computed data at the same position (x/C = 0.9958). The comparison is qualitatively
satisfying: the same narrow-band component emerges by 7 dB from a wide-band
continuum. However, the quantitative comparison shows that predicted levels are overestimated by nearly 3 dB and that the predicted vortex-shedding frequency (4.7 kHz)
is also overestimated (3.5 kHz in the experiment). Because this frequency is governed by a Strouhal number based on the wake thickness (sum of the TE thickness
and TBL displacement thicknesses), this overestimation may be partly explained by
the excessive slenderness of the computed TBL with respect to the experimental one.
However, this cannot be explained solely by the absence of transition triggering in the
simulation.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.66. Evolution of the wall-pressure PSD along the chord.
Wave-numberfrequency spectra
The spectral densities presented earlier in this section are single-point data that do not
provide any information on the propagative features of the wall-pressure field. This
information can be revealed by processing a discretized, space-time, Fourier transform of any function p(x, t), providing the so-called wave-numberfrequency spectrum (k, ). For example, a monodimensional wave with frequency 0 and velocity
V0 propagating in the x-direction, for instance p(x, t) = ei0 (tx/V0 ) , will appear in the
(k, ) domain as a Delta function centered at k = 0 /V0 and = 0 . With the same
process, a white noise propagating at celerity c0 in the x-direction will appear as a
ridge along the line k = /c0 or along the line k = /c0 if the sound travels in the
opposite (x) direction.
Figure 6.68 presents the results of this process applied to the wall-pressure field computed via LES on the airfoil pressure side near the TE at x/C = 0.9. The wave-number
323
13:36
P1: IBE
0521871441c06
324
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.67. Evolution of the wall-pressure PSD along the chord.
domain extent [/x, /x] depends on the local space sampling x of the pressure field or the size of the grid cells. A wide-band convective ridge can be observed
along the k = /Uc line generated by the convection in the flow direction, of eddies
inside the TBL at an average velocity Uc slightly inferior to the local mean flow velocity
Figure 6.68. Wave-numberfrequency spectrum of wall-presssure fluctuations on the airfoil suction

side at x/C = 0.9.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.69. Wave-numberfrequency spectrum of wall-pressure fluctuations on the airfoil suction

side at x/C = 0.5.
Ue . Another wide-band component can be observed along the k = /(c0 Ue ) corresponding to an acoustic propagation in the direction opposite to the flow with a peak at
4.7 kHz, the vortex-shedding frequency. Figure 6.69 presents the same results obtained
at midchord (x/C = 0.5). The convective ridge and the upstream acoustic component
are again present. One interesting result is a new component k = /(c0 + Ue ), which
corresponds to acoustic waves (also with a peak at 4.7 kHz) propagating in the flow
direction (downstream), which may be generated by the diffraction, at the airfoil leading
edge, of the acoustic waves propagating upstream.
6.5.4.2 Flow-pressure fluctuations

Instantaneous maps of pressure fluctuations
The existence of a sound field generated at the TE, strongly suggested by the surfacepressure results above, is confirmed in Figure 6.70, which shows instantaneous isovalues
of the pressure fluctuations inside the flow. At every point of the LES grid, pressure fluctuations are computed by subtracting the time-averaged pressure to the instantaneous
pressure. The grayscale table is adapted to the low values of the acoustic fluctuations
(6 Pa), although there are much larger pressure fluctuations (up to 500 Pa) in the TBL
and in the wake. Concentric waves are clearly observed near the TE, with a wavelength
corresponding to the vortex-shedding frequency 4.7 kHz which, as shown earlier in this
section, dominates the source spectrum by a few decibels. An animation was generated
from 300 identical maps plotted for successive time steps (t = 5.105 s), allowing the
acoustic radiation in the fluid and the turbulence convection along the airfoil chord to
be visualized simultaneously with their respective celerity.
325
13:36
P1: IBE
0521871441c06
326
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.70. Instantaneous isovalues of

pressure fluctuations obtained from LES
data.
Low-pass filtering by the CFD grid
It is interesting to note that the wave pattern corresponding to the vortex-shedding noise
vanishes at a half-chord from the airfoil, while larger wavelengths are observed much
farther away. This is explained by the radial stretching of the LES grid, which acts on
the noise field as a low-pass filter. It is known that the propagation of an acoustic wave
will not be correctly simulated if its wavelength is discretized with less than four or
six cells. To illustrate this, Figure 6.71 shows the evolution of the wall-pressure spectra
Figure 6.71. Evolution of the wall-pressure

spectra along the vertical grid line x = C
(starting from the TE upper corner) with respect to the vertical distance z.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
327
Figure 6.72. Spanwise evolution of the coherence

of the surface-pressure field on the suction side
near the TE (x/C = 0.9958). The frequency bandwidths are integrated.
along the vertical grid line x = C (starting from the TE upper corner) with respect to the
distance from the TE (in logarithmic scale). This plot clearly shows that low-frequency
waves propagate much farther than high-frequency waves.
6.5.4.3 Spanwise coherences
Inside the TBL and the wake, turbulent structures or eddies are convected at low velocities, which means that typical spanwise length scales may be smaller than the spanwise
extent of the LES domain and correctly simulated. To illustrate this, Figure 6.72 shows
the spanwise evolution of the coherence of the surface-pressure field near the TE in
several frequency bandwidths. It is defined as
2 (y, f ) =
|S pp (0, y, f )|2

,
|S p (0, f )||Sp(y, f )|
(6.56)
where S p (y, f ) denotes the surface-pressure spectrum at the spanwise position y, and
S pp (y1 , y2 , f ) denotes the surface-pressure cross spectrum between spanwise positions
y1 and y2 . The coherence rapidly vanishes at any frequency bandwidth, meaning that
integral length scales are smaller than the spanwise extent of the LES domain. It can be
seen that the largest length scales correspond to the vortex-shedding frequency, which
is a mechanism significantly correlated in the spanwise direction.
Figure 6.73 shows spanwise coherence of the pressure field above the TE on the grid
line j0 = 54 located at distance z 0 = 37.9 mm from the TE, where pressure fluctuations
are expected to be mostly acoustic. As explained in the next paragraph, this particular
grid line will be the location of the LESEuler coupling. Figure 6.74 shows that typical
length scales seem to be much longer than the spanwise extent of the LES grid
especially at low frequency and the frequency of vortex shedding. In other words, the
very-near-field turbulent flow is strongly three-dimensional, whereas the acoustic field
simulated at some distance from the airfoil is mostly two-dimensional, which confirms
that a two-dimensional LESEuler coupling is reliable.
13:36
P1: IBE
0521871441c06
328
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.73. Spanwise evolution of the coherence

of the pressure field at distance z0 = 37.9 mm from
the TE at x/C = 1. The integrated frequency bandwidths are the same those used in as Figure 6.72.
6.5.4.4 LESEuler coupling process
This section describes how the acoustic field generated at a very short distance by LES
propagates through the nonuniform mean flow by resolution of the Euler equations.
The main steps of this LESEuler coupling are (i) the derivation of an acoustic grid
and (ii) the splitting of LES data into a mean flow and a perturbation field.
Propagation code
A code named E3P (propagation via Euler equations under a small perturbation hypothesis) has been developed at ONERA by Redonnet, Manoha, and Sagaut (2001)
relying on the discretized Euler equations in a perturbation formulation based on
the splitting of the total field in a mean flow (subscript o) and a perturbation field
(subscript p):

[t uo + Fo ] + t up + Fp = qo + qp + qs .
(6.57)
Here, the vectorial quantities u, F, and q, respectively, denote the unknowns, fluxes,
and source terms. These equations are discretized in conservative form and can either
Figure 6.74. Final problem-adapted acoustic grid.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
be linearized (only first-order terms are kept) or nonlinearized (all linear and nonlinear
terms are conserved). Cartesian or curvilinear 2D3D (monodomain) structured grids
can be handled. In the space domain, the discretization uses high-order, finite difference
explicit or implicit schemes. In the time domain, multistep schemes (AdamsBashforth
or RungeKutta) are implemented. Specific boundary conditions are used for solid surfaces, nonreflexive borders, periodicity, or perturbation injections. Nonuniform mean
flows are taken into account. The characteristics of the code, as well as several test cases
of acoustic scattering on solid objects within nonuniform 2D and 3D flows, have been
described in detail in Redonnet et al. (2001). The numerical tools developed in E3P
are now implemented in a multipurpose CFDCAA solver named sAbrinA (solver
for aeroacoustics broadband interactions from aerodynamics) built from the general
multiblock solver FLU3M, which already houses the LESDNS solvers. This new
solver is another step toward an integrated tool for simultaneously coupling CFD and
propagation calculations.
Acoustic grid derivation
The simulation of viscous turbulent flows via RANS or LES methods requires a high
level of grid refinement near the solid walls and otherwise uses strong grid stretching
in the far field to benefit from the almost uniform flow conditions. Consequently, CFD
grids are notoriously unadapted to the simulation of acoustic propagation via the Euler
equations, which must be discretized on homogeneous grids with cells of rather constant
size all over the domain. In the present case, the coupling of LES and Euler equations
is required to create a new grid for the acoustic domain.
An acoustic grid was derived from the LES grid following specific constraints:
(i) the homogeneity of grid refinement; (ii) the average cell size (with respect to the
smallest wavelengths); (iii) intrinsic limitations of the E3P code, which, for example,
cannot handle multiblock structured grid; and (iv) the coupling interface does not
intercept the wake. This acoustic grid is shown in Figure 6.74. The interior border, on
which LES data will be injected in the Euler domain, follows the airfoils surface at an
average distance of 1% of the chord length. The outer border of the acoustic domain is
approximately one chord away from the airfoil because it was confirmed that the mean
flow is quasi-uniform beyond this distance. All cells have a quasi-constant size close
to min /6, for the smallest considered wavelength min (corresponding to the highest
frequency of interest 10 kHz, for instance). The grid involves 183 C-lines and 866
radial lines, or 158,478 points. Figure 6.75 presents a closer view of this grid near
the airfoil.
Data injection process
The E3P code is linked with unsteady CFD computations via an interface: a 2D (respectively, 1D) interface is required for a 3D (respectively, 2D) Euler calculation. The
principle of the data injection process is to reinitialize the solution vector u p at each
time step on a ghost-point distribution outside the Euler domain with the injected perturbation values. The number of ghost points depends on the maximum order of the
329
13:36
P1: IBE
0521871441c06
330
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.75. Final problem-adapted acoustic grid (closer view).
used spatial scheme and spatial filter. In the present application, a sixth-order scheme
(seven-point stencil) is used along with a tenth-order filter (eleven-point stencil), and
thus the filter stencil requires five ghost points outside the domain, the data being finally
injected on six points (five ghost points plus the interface point).
Mean flowperturbation splitting
The E3P code handles perturbations that propagate on a steady inhomogeneous flow. In
our case, the mean flow was obtained at any point of the LES domain by time-averaging
the unsteady data. This method assumes that the LES computation was long enough
to reach a stabilized mean flow, which was actually untrue: its average magnitude is
subject to a slow drift. This problem was minimized by limiting time averaging to the
last quarter (300 time steps) of the whole stored LES available duration (1300 time
steps) to ensure that the mean flow would be as stabilized as possible, but this problem
may have induced a bias in the coupling process.
LESEuler coupling result
Figure 6.76 shows isovalue contours of an instantaneous pressure fluctuation field

computed from (i) LES inside the injection interface and (ii) E3P (from LES data
injection) outside the injection interface. Figure 6.77 shows a closer view of the contours
in Figure 6.76 centered on the airfoil. This view shows that there is no discontinuity at
the injection interface between the LES wave fronts and the E3P wave fronts.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.76. Isovalue contours of instantaneous pressure fluctuation field (range 2 Pa, black and
white) computed from (i) LES inside the injection interface and (ii) E3P (from LES data injection) outside
the injection interface.
6.5.4.5 Full simulation, including LES, Euler, and Kirchhoff integration
Finally, this section describes the full three-step CAA process: the acoustic field radiated
at the external boundary of the Euler domain (where the mean flow can be considered
as uniform) becomes the entry data of a Kirchhoff integration, which provides the noise
radiated in the far field.
Kirchhoff formulation
The KIM (Kirchhoff integration method) code provides the noise radiated by any threedimensional surface in a flow with uniform velocity in the direction x given the density
(or pressure with isentropy assumed or p = c02 ) and its normal gradient along this
surface. The formulation is (Prieur and Rahier 1998)

1
DXY
1

d Sd,
(6.58)
f K ( X , Y , ) t +
( X , t) =
4
d
c0
Figure 6.77. Isovalue contours of instantaneous

pressure fluctuation field (range 2 Pa, black and
white) computed from (i) LES inside the injection interface and (ii) E3P (from LES data injection) outside
the injection interface (closer view).
331
13:36
P1: IBE
0521871441c06
332
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
where D X Y is the distance, corrected from convection effects, between the source point,
located on the integration surface, and the observer point (both defined in a reference
framework in which the surface S is at rest):

DXY =
d=
dM0 (X 1 Y1 )
,
2
(1 2 )i j + 2 (X i Yi )2 .
(6.59)
Here, M0 is the mean-flow Mach number and is a PrandtlGlauert factor (1 M 2 ).
The expression f K ( X , Y , ) is given by

n ( X Y ) + M02 n 1
f K ( X , Y , ) = 2 (1 M02 )
d
Y1
n
"
!
n ( X Y )
1
+
.
M0 n 1 +
c0
d
(6.60)
This formulation assumes that the pressure field on S satisfies the convected wave
equation:

D2
2
a0 = 0
D 2
with
,
=
+ U0
D
X1
(6.61)
which governs the propagation of acoustic waves in a medium with uniform velocity U0 in the direction X 1 . The spacetime discretization of the Kirchhoff integration
on a given closed control surface assumes that temporal fluctuations of fluid pressure
are time sampled (with time step t and duration N t) at any cell (with maximal
dimension ) of a two-layer (with maximal normal separation n between both
layers) surface grid. If a and Ta = a /c0 denote the acoustic wavelength and period,
the crucial parameters of the process are /a , n/a , and t/Ta . The parameter
/a must be as small as possible to ensure that the acoustic field will be correctly
discretized on the control surface especially if the code is not implemented with a
noncompact cell treatment. The parameters n/a and t/Ta must also be minimized to ensure that the normal gradients and time derivative of the pressure fluctuations
will be accurately computed. Finally, a long duration N t is necessary for (i) statistical accuracy of the spectral processes applied to random signals and (ii) limiting the
effects of truncation due to the delay between maximum and minimum propagation
times D X Y /c0 .
Final result
The whole three-step CAA process is finally applied to the NACA-0012 case, including
a far-field noise prediction performed by use of a Kirchhoff integration based on the
pressure field and its normal derivative on the external boundary of the Euler domain.
Figure 6.78 shows the superimposition of a pressure fluctuation field obtained from
LES (below the LESEuler interface), a pressure fluctuation field obtained from E3P
by injection of LES data on the new acoustic grid (between the two interfaces), and a
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.6 BLUNT BODIES (CYLINDER, CARS)
Figure 6.78. Isovalue contours (range 3 Pa black

and white) of instantaneous pressure fluctuation
field computed from (i) LES data inside the injection interface, (ii) Euler data (from LES data injection) between the injection interface and the Kirchhoff control surface, and (iii) Kirchhoff integration
data beyond the Kirchhoff control surface.
pressure fluctuation field obtained from Kirchhoff by integration of Euler data on the
surface indicated by the highest black line (above the E3PKirchhoff interface). The
wave fronts are continuous from one domain to another. An interesting point is that
acoustic waves emitted by the airfoils leading edge are hardly visible on this figure.
This phenomenon has been already observed on wall-pressure wave-numberfrequency
spectra computed at midchord on the suction side (Manoha, Delahay, Sagaut et al. 2001)
and confirms that the leading edge acts as a geometrical singularity on which incident
acoustic waves coming from the trailing edge are scattered.
6.6 Blunt bodies (cylinder, cars)

Franco Magagnato
6.6.1 Overview of blunt-body simulations
The flow-induced noise around blunt bodies such as cylinders and cars is described in
this section. The numerical simulations of this flow field have been under investigation
for many years. Owing to the high computational requirements (especially for the
flow around a car), the first calculations were made with the help of RANS (e.g., Fritz,
Magagnato, and Rieger 1991, 1992). Beaudan and Moin (1994) performed the first LES
of the flow past a circular cylinder at a relatively low Reynolds number of Re = 3900.
333
13:36
P1: IBE
0521871441c06
334
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Later, others performed LES calculations of the flow past circular cylinders but at
higher Reynolds numbers (Breuer 1997; Frohlich et al. 1998; Magagnato and Gabi
2000). All these calculations were done without aeroacoustical predictions. It appears
that only a few investigations of the aeroacoustical aspects of the flow past a cylinder
have been performed up to now. Batten et al. (2002a) have computed the flow past
a square cylinder using their new method called limited numerical scales (LNS) and
a conventional unsteady RANS approach based on a cubic eddy-viscosity turbulence
model. They found that the predicted SPL of the LNS approach was some 20 dB higher
compared with predictions made through URANS. They attributed this to the reduced
levels of numerical dissipation in the LNS calculations.
Because the computational requirements of an LES in conjunction with the LEE
of the far-field noise are too demanding, most calculations on cylinders and cars have
been done so far with the acoustic analogy (mainly with Ffowcs WilliamsHawkings).
Experience in computational aeroacoustics shows that, in the far field, the accuracy of
acoustic signals is determined by the accuracy of the flow-field prediction rather than
by acoustic analogy. Although it is well known that URANS equations cannot explore
unsteady flow accurately, this approach is still a very popular numerical procedure
in computational aeroacoustics because of the relatively low CPU time and memory
requirements. URANS equations resolve only a narrow band of frequencies present
in the flow and average the rest of the frequency domain. Directional numerical solution has to be carried out to resolve all of the fluctuations. However, because DNS
requires very fine grid resolution and therefore very high CPU time and memory, today DNS is only affordable for very low Reynolds numbers and simple geometries.
A good compromise is LES, which resolves a fairly large range of frequencies and
models only a small part of the flow. It is commonly assumed that the unresolved part
of the flow consists of very small eddies that exhibit a homogeneous and isotropic
structure. As long as this assumption is valid, even an algebraic model, like the SGS
model from Smagorinsky, provides reasonable results. To keep this assumption valid
requires that all important scales be resolved, which is a manageable task for small
Reynolds numbers and simple geometries. However, in industrially relevant flows, for
example in turbomachines, very complicated turbulence structures prevail. Resolving
all relevant turbulence scales is challenging work even with supercomputers using high
processing power and the memory access rate available today. Unless the computational
mesh is fine enough to capture all relevant scales of turbulence, unresolved fluctuations
cannot be considered isotropic or homogenous. Better turbulence models than algebraic ones have to be employed to model the unresolved part fairly. For this purpose
a two-equation turbulence model has been proposed by Magagnato and Gabi (2000).
This so-called adaptive k model can be used for all cell Reynolds numbers in the
unsteady case. It has the property of reducing to a DNS if the temporal and spatial
resolution of the flow field is on the order of the Kolmogorov microscale, but if the
fluctuations are not resolved the model reduces to a standard two-equation model. Another feature of the model is that the backward scattering of energy transfer is taken into
account.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.79. Numerical mesh (2D plane).
6.6.2 Circular cylinder

6.6.2.1 Etkin test case
A comparison between 2D and 3D URANS, LES, and the experiment from Etkin,
Korbacher, and Keefe (1957) of the flow past a circular cylinder has been made by
Pantle (2002) and Magagnato, Pantle, and Gabi (2002). Etkin et al. (1957) performed
experiments in the turbulent flow region of the flow past a circular cylinder. The acoustical frequency spectrum of the computation was expected to show some noise owing
to the turbulent sources entering the acoustical computation. For the comparison, an
observer at a distance of 0.6 m perpendicular to the main flow direction above the
cylinder was considered. The main flow velocity was U = 68.6 m/s, giving a Mach
number of M = 0.2; the cylinder diameter was d = 0.0125 m. The flow medium was
air at ambient temperature, and the Reynolds number based on the cylinder diameter was Re = 60,000. The unsteady computation was performed using the k model
of Speziale, Abid, and Anderson (1990). The turbulence level was set according to
the results of Etkin et al. (1957) to 0.3%. For the turbulent computation some information about turbulent length scale or eddy viscosity was also required that could
not be found in the experimental description. Thus, as a starting value, a reasonable
eddy viscosity ratio was chosen. Owing to previous computing experiences it was set
to t /l = 1.0.
The computational schemes were of second-order accuracy both in space and time
(Magagnato 1998). The numerical mesh consisted of about 50,000 cells (Figure 6.79)
for the 2D case, whereas about 3,200,000 cells were used for the 3D case. The meshes
showed a distance between the cylinder and the mesh boundary of about 17 diameters in
the upstream direction, 20 diameters in the directions perpendicular to the flow, and 43
diameters downstream. At the boundaries upstream and parallel to the flow, a far-field
condition was applied; downstream, the static pressure was set. In the third direction
(along the cylinder) there was a symmetry condition for the 2D case, whereas a periodic
335
13:36
P1: IBE
0521871441c06
336
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
condition was used in 3D. For the LES and the 3D URANS calculation the spanwise extension of the cylinder was chosen three times the diameter of the cylinder. This is more
than necessary to capture the correlations of the turbulent length scales (approximately
one diameter is sufficient) but may be insufficient for the acoustical length scales of
this low-Mach-number flow, which are much larger compared with the turbulent ones
(see Manoha, Delahay, Sagaut et al. 2001). For the LES, the SmagorinskyLilley SGS
model was used. The numerical scheme was a central differencing scheme in combination with a classical artificial dissipation scheme based on Jameson et al. (1981) but
with a strongly reduced numerical dissipation compared with a URANS calculation. In
many previous calculations it was found that the numerical dissipation of a compressible
solver is absolutely necessary for the stability of the calculation but must be reduced in
the range of 3 to 5% compared with URANS calculations to avoid overdamping of the
turbulent small-scale fluctuations. The time step used for the acoustical computation
was t = 2 105 s.
Pantle (2002) found that the 2D URANS calculation gave an SPL that was too
high owing to the two-dimensionality of the flow field and consequently larger lift
and drag fluctuations compared with the LES results. The well-known, strong, threedimensional flow field observed in experiments obviously could not be predicted by the
2D simulation, and consequently an overprediction of the SPL of about 8 dB was found.
The same conclusions have been drawn by Boudet et al. (2002). They investigated the
flow past a cylinder measured by Michard, Jacob, and Grosjean (2002). Their prediction
overestimated the SPL at the vortex-shedding frequency by 25%. The question was now,
could a 3D URANS calculation resolve the three-dimensional structure of this flow,
resulting in a reduction of the SPL? Boudet, Casalino et al. (2003) as well as Pantle
(2002) found that a 3D URANS was not able to predict a 3D flow field owing to
the high level of eddy viscosity distribution in the wake of the cylinder. Therefore the
3D URANS gave essentially a 2D flow-field solution with the same SPL. Only in the
LES, a three-dimensional flow field was predicted and the lift and drag oscillations
were considerably reduced as well because the SPL was predicted about 15 dB lower
compared with the URANS calculations. Pantle (2002) found in a calculation that, in
comparison with the experiment of Etkin et al. (1957), the main peak of the SPL was
predicted 5.5 dB lower than the experimental value of SPL = 117 dB. The fundamental
frequency predicted by the 2D, 3D URANS calculations as well as LES was at 1260 Hz,
whereas in the experiment it was predicted at 1000 Hz. Because this was compared in
the one-third-octave band spectrum, an uncertainty remains about the real discrepancy
between calculations and experiment. In Figures 6.80 and 6.81 the SPL obtained by
the computation and the comparison with the noise-reduced experimental findings is
shown. Because these measurements were taken 45 years ago it is not certain how
reliable they are.
6.6.2.2 ECL test case
A more recent experiment was carried out in the test facilities of the Ecole Centrale
de Lyon by Michard et al. (2002) and Jacob et al. (2005). Again a circular cylinder
and an airfoil in the wake of a cylinder were investigated. Because this experiment is
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
337
Exp. (Etkin)
LES
130
SPL [dB]
120
110
100
90
80
103
104
f [Hz]
Figure 6.80. Sound-pressure level of LES.
considered a benchmark validation case for CFD codes, both aerodynamic and acoustic
data were measured. Another reason besides precise and purposive measurements is
that, although all solid bodies are stationary, the flow in the experiment imitates the flow
at rotating blades. The experimental setup is shown in Figure 6.82. The airfoil chord is
aligned with the center of the cylinder cross section, and it is placed one chord length
after the cylinder to avoid significant feedback of the airfoil onto the shedding. The
experiments were carried out with different cylinder diameters and inlet velocities. In
the numerical simulations of Magagnato, Sorguven, and Gabi (2003), one configuration
is chosen. In this reference configuration the diameter of the cylinder is 0.01 m and 0.3
m long in the spanwise direction. An anechoic wind tunnel provides air with 70 m/s
inlet velocity and 1% turbulent intensity. The Reynolds number based on the cylinder
diameter was Re = 4.8 104 .
Exp. (Etkin)
URANS
130
SPL [dB]
120
110
100
90
80
102
103
104
f[Hz]
Figure 6.81. Sound-pressure level of 3D URANS.
13:36
P1: IBE
0521871441c06
338
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.82. Experimental setup.
The observer point for the acoustic measurements was placed at a distance of R =
1.38 m from the airfoil midpoint at an observation angle of = 90 in the midspan plane.
The resolution of measurements is given as 2 Hz. Because the detailed experimental
data are not yet freely available, the experimental data presented here are taken from the
published papers and are therefore subject to rounding errors. The computational grid
for the cylinder-only configuration was similar to the one used for the Etkin et al. (1957)
case. The grid for the cylinder extended 20 times the cylinder diameter upstream and in
the directions parallel to the main flow and 40 times the downstream. It consisted of 72
blocks and 3.2 106 control volumes in the finest mesh. All simulations were carried
out with up-to-the-wall integration, which means that y + is kept less than or equal to
1 in the finest mesh levels. As determined by the space correlation considerations, the
spanwise length was set to two times the cylinder diameter in both grids and resolved
with 65 control volumes in the finest grid levels. This was considered necessary to
resolve the flow field better in this direction.
The total number of control volumes was limited by increasing the expansion ratio
beyond the vertical region. The effects of this are discussed in the next section by
comparison of two configurations. The boundary conditions used for the acoustical
calculations are extraordinarily important because they can affect the acoustic data
easily. In the simulations far field, static pressure and periodic boundary conditions are
used. The far-field boundary condition can be considered nonreflecting or at least weakly
reflecting because Riemanns invariants have been used. At the outlet, the static pressure
is kept constant. These boundary conditions are indeed reflective, but two factors allow
us to neglect the eventual reflections. First, the outlet boundary is set far away from the
cylinder. Hence, the variations of the pressure near the boundary are small. Secondly, the
smallest wavelength the control volumes near the boundary can resolve is much higher
than for the wavelengths relevant to the acoustic analysis. The eventually reflected
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
339
Figure 6.83. Instantaneous streamlines in 2D URANS simulation.
waves would therefore not affect the acoustic computation. The fluid medium was air
at room temperature, and the thermodynamic data were set accordingly. Velocity in
the far field was 70 m/s and had a block profile. The outlet static pressure was set to
p = 104,326.23 Pa. Turbulence intensity was set to 1% in the far field as it was measured
in the experiment. Regarding that, the eddy viscosity ratio was set equal to 5.
The computed acoustic signal needs a correlation owing to the differences in numerical and experimental configurations. Because the spanwise length of the cylinder and
the airfoil is much smaller in the computations than in the experiments, the numerical
acoustic signal has to be correlated. For this purpose the following empirical formula
according to Etkin et al. (1957) was used:
I =
0 2 2 6
L c L St V ,
32a03
(6.62)
where 0 is the reference density, a0 is the reference speed of sound, St the Strouhal
number, V the inlet velocity, and cL the fluctuating force coefficient. By applying
Equation (6.62) to two different lengths L exp and L sim and rearranging, we get

L exp
S P L = 20
.
(6.63)
L sim
Because the length in the spanwise direction is 0.3 m in the experiment and 0.02 m in
the simulation, Equation (6.63) yields S P L of 23.52 dB. All computational acoustic
results were corrected with this value. It should be mentioned that other correlations have
been proposed. Kato et al. (1994) proposed a correlation that gives a smaller correction
than Equation (6.63). As an overview of the flow around the cylinder, the streamlines
and the contours of the velocity component parallel to the main flow v are shown
in Figures 6.83 and 6.84. These are instantaneous snapshots of the 2D URANS and
LES (adaptive k ) calculations, respectively. Both are results at the finest grid levels.
In 2D simulations only a few vortices are resolved; they represent a pure periodic
structure. After about four times the cylinder diameter in the wake, the streamlines
become completely parallel to the main flow again. The results of LES simulations
exhibit a much larger von Karman vortex street. The frequency of the vortex structure
is still dominated by the same shedding frequency, but a spectral broadening around
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
340
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.84. Instantaneous streamlines in LES with adaptive model.
this frequency is observed. The vortices have lost their uniformity and exhibit a more
chaotic structure.
Figures 6.85 through 6.87 show the lift and drag coefficients history of the cylinder
in different grid levels of the LES. Here the lift fluctuates around cL = 0 with an
amplitude around cL = 1. The perfect periodicity of lift and drag coefficients in the
third-finest grid level shows that the flow is still two-dimensional, although the grid
is 3D. This signifies that the grid is not fine enough to resolve the three-dimensional
structures in the flow. In a finer grid level (Figure 6.86), three-dimensionality is captured.
Fluctuations with different frequencies are generated, but the amplitudes are of the
same extent. In the finest grid level, the shedding frequencies cover a larger range, and
the amplitudes of the lift and drag coefficients sink dramatically. This shows that the
turbulent energy of the flow is distributed in all three dimensions and is totally threedimensional. It has to be mentioned here that the velocity component in the spanwise
direction is 10 m/s in the second-finest grid, whereas it is 50 m/s in the finest one.
The 2D simulations result in a pure, periodic acoustic signal at the observation point, as seen in Figure 6.88. The acoustic signal fluctuates between
3 106 kg/m3 . Correspondingly, fast Fourier transform analysis shows a clear
peak at the dominant frequency f = 1612 Hz, S P L = 106 dB and its harmonics at
2
CL, CD
Figure 6.85. Lift and drag coefficients of the cylinder in the

third-finest grid.
0.0375
t [G]
0.0725
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
341

second-finest grid.
CL, CD
0.1025
t [G]
0.1075

finest grid.
CL, CD
P1: IBE
0521871441c06
Figure 6.88. Acoustic density fluctuations of 2D

URANS simulation in the finest grid.
0.1325
t [G]
0.1375
13:36
P1: IBE
0521871441c06
342
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.89. Sound-pressure level of 2D URANS

simulation in finest grid.
f = 3274 Hz, S P L = 71 dB, and f = 4887 Hz, S P L = 71 dB (Figure 6.89). Although the tendency of the sound-pressure spectrum is predicted correctly, both frequencies and amplitudes of the peaks are miscalculated.
The second-finest mesh in 3D simulations results in a more chaotic acoustic signal
(Figure 6.90). The corresponding sound-pressure spectra indicate that fluctuations with
different frequencies besides the dominating one exist (Figure 6.91). The amplitudes
are on average about 3 106 kg/m3 . Because the amplitudes in acoustic signals are
comparable with the 2D case, the SPL is also on the same level. Therefore frequency is
predicted better ( f = 1466 Hz), but the amplitude is similar to the 2D case (S P L = 109
dB).
Figures 6.92 and 6.93 give the acoustic signal and the sound-pressure spectra of
the finest grid in 3D simulations. Here, the amplitudes of the acoustic signal sink
dramatically in comparison with the coarser grids. This is a clear result of the threedimensionality of the calculated flow field. Figure 6.93 shows that the SPL is damped
because of these phenomena. Also, the frequency of the dominating fluctuations is
predicted better. Here, the main peak is at 1466 Hz and 84 dB, which corresponds
Figure 6.90. Acoustic density fluctuations of LES

with adaptive model in second-finest grid.
13:36
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
343
Num
Exp
100
Figure 6.91. Sound-pressure level of LES with

adaptive model in the second-finest grid.
SPL [dB]
80
60
40
2000
f [Hz]
4000
6000
Figure 6.92. Acoustic density fluctuations of

3D LES simulation in finest grid.
Num.
Exp.
80
Figure 6.93. Sound-pressure level of LES with

adaptive model in finest grid.
SPL [dB]
P1: IBE
0521871441c06
60
40
2000
f [Hz]
4000
6000
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
344
0 521 87144 1
printer: sheridan
November 20, 2006
Num.
Exp.
SPL [dB]
80
Figure 6.94. Sound-pressure level of LES in finest

grid with the Smagorinsky and Lilley model.
60
40
2000
f [Hz]
4000
6000
to about 7% and 0.5% error in frequency and SPL, respectively. One cannot see the
positions of harmonics clearly because of the many fluctuations in the computational
SPL curve. It was assumed that these would vanish with more acoustic data.
Figures 6.93 and 6.94 show the comparison between two different LES models, the
Smagorinsky and Lilley model and the adaptive k model. Both provide similar results
with respect to the position of the peak, whereas the adaptive k model is slightly better.
At lower and higher frequencies, both models achieve higher SPL curves than in the
experiment, but the adaptive k model is about 50% better than the Smagorinsky and
Lilley model in these regions.
Montavon et al. (2002) have made calculations as well as measurements on a cylinder in cross flow at two different Reynolds numbers. The low Reynolds number of
Re = 3900 was calculated as an LES using CFX-5 commercial software. The aerodynamic results were in very good agreement compared with their own experiments as
well as with published work in the literature (Lourenco and Shih 1993). The higher
Reynolds number of Re = 140,000 gave less satisfactory aerodynamics results compared with the experiments by Cantwell and Coles (1983). Although the distribution of
the fluctuating pressure coefficient agrees fairly well with the experiment of Norberg
(1992), the back pressure coefficient was predicted to be about 1, which underestimates the measurements by about 20%. This could be due to a low resolution of the
flow field, for only 225,000 elements were used for both Reynolds numbers. The acoustical results were obtained with SYSNOISE (LMS SYSNOISE 2002), which solves
the acoustic analogy equation of Ffowcs WilliamsHawkings with the direct boundary
element method. At the higher Reynolds number, the prediction of the overall SPL was
found to be about 5 dB higher than in the experiment, whereas at the lower Reynolds
number the overestimation was on the order of 10 dB. The reason for this error was not
clearly understood, but Montavon et al. (2002) assumed that it had to be connected with
the overprediction of the drag coefficient by 30%, which could mean that the dipole
sources were generally overestimated.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.95. The Ahmed body from Ahmed et al. (1984) (isosurface of zero streamwise velocc
ity from Kapadia et al., 2003). Reprinted with permission from SAE Paper # 840300 1984
SAE
International.
6.6.3 Car
The prediction of the flow field around a car is without any doubt of great practical importance. In the early days of the application of CFD in car aerodynamics, the
drag and lift generated by a car were the main driving forces. Nowadays car designers are interested additionally in the pressure distribution around a specific shape in
order to get correct boundary conditions for internal flow-field predictions (e.g., the
engine cooling device or the air conditioning inlet and outlet position). Recently the
noise generated by the flow around a car became increasingly important for the car
industry mainly because of comfort factors but also in response to more stringent noise
regulations.
It appears that only very few URANS, DES, or LES results have been reported
in the open literature so far. Perzon and Davidson (2000) have calculated the flow
around the Association for Structural and Multidisciplinary Organization (ASMO)
model with URANS. The geometry of the ASMO model is mainly used for testing
CFD codes. Perzon and Davidson (2000) have found the pressure distribution around
the car in the symmetry plane to be in good agreement with the measurement whereas
the total drag deviates significantly. Hinterberger, Garcia-Villalba, and Rodi (2003) and
Howard and Pourquie (2002) have made large-eddy simulations around the Ahmed
body (Ahmed, Ramm, and Faltin 1984). The Ahmed body is a generic car shape
with a relatively simple geometry and extensive measurement data are available (see
Figure 6.95).
Kapadia, Roy, and Wurtzler (2003) have performed a DES about the same geometry.
The resolution of these calculations ranged from 1,700,000 points (Kapadia et al. 2003)
up to 18,500,000 (Hinterberger et al. 2003). The comparison of the predicted flow
structure with the experiment was well captured, but some discrepancy, especially in
the lower part of the slant back, exists for all calculations. None of them have predicted
the flow-induced noise.
345
13:36
P1: IBE
0521871441c06
346
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.96. Mesh for the CFD model from Volkswagen.
Another experiment around a generic car shape has been made by the Volkswagen
AG (see Wustenberg and Hupertz 1995). This geometry (see Figure 6.96) is a bit closer
to realistic cars and is as well very detailed.
An LES of this CFD model has been made by Magagnato, Pritz, and Gabi (2005)
using a 3D grid provided by Volkswagen AG. The full mesh has been obtained by
mirroring the original grid at the symmetry plane. A total of 5,200,000 points covers
the flow field, which extends four lengths in the upstream direction, three lengths in
the lateral direction, five lengths in the downstream direction, and three lengths of the
car perpendicular to the car. The freestream velocity was u = 63 m/s with a turbulence
level of T u = 1%; the density was = 0.62, and the temperature was T = 292 K. This
results in a Reynolds number of about Re = 7,350,000 based on the length of the car.
It is clear that the number of grid points above mentioned can not adequately resolve
the flow field in such a way that a large part of it lies well inside the inertial subrange
of the energy spectrum. It is probably more a very large eddy simulation.
The simulation has been made by using the Smagorinsky and Lilley SGS model, a
second-order central differencing scheme in space, and a second-order accurate implicit
scheme in time (dual time stepping). The calculations were done on an IBM SP2 parallel
computer using 32 processors. The global computational time was about 80 h. In Figures
6.97 and 6.98 the streamlines and the vorticity magnitude at an instantaneous time are
shown. The strong three-dimensional character of the flow field especially in the
wake is clearly visible.
The comparison of the pressure distribution in the symmetry plane of the CFD
model between experiment and calculation shows a very good agreement over the roof
and in the major part of the car underbody (see Figure 6.99). Except close to the front
of the underbody, a considerable deviation can be observed. This is very surprising
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.97. Streamlines in the wake of the CFD model.
because the flow in this region is mainly a potentially dominated flow that is very easily
predictable by CFD. One possible explanation could be the different application of the
wall-boundary condition in front of the car. In the experiment, a boundary layer sucking
in front of the car is used, whereas in the calculation a symmetry condition is applied in
the front part ending with a wall boundary condition just below the front part of the car.
Unfortunately, no acoustical data have been measured in the experiment. Thus, the
calculated SPL with the acoustic analogy at an observer point of x = 10 m, y = 10 m,
and z = 1 m is only for demonstration purposes.
The SPL collected at 512 samples is shown in Figure 6.100. The main peak in
the SPL is about 57.9 dB at a frequency of 41 Hz. Over a range of about 100 Hz the
Figure 6.98. Vorticity in the wake of the CFD model.
347
13:36
348
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.99. Pressure coefficient in the symmetry plane of the CFD model.
amplitude of the SPL remains on the order of 50 dB, whereas at higher frequencies it
drops very quickly.
Another approach of predicting aeroacoustics noise generated by a car is used by
Grunewald and Basel (see Isensee and Frenzel 2003). They have been using the BEM
in conjunction with unsteady calculation methods for the prediction of noise in a recent
study of the German Ministry of Education and Research (Leiser Verkehr).
There are also some calculations published with commercial codes (Correa, Massa,
and Zajas 2003; Mendonca 2002) calculating the flow around a generic mirror with
LES and acoustic analogy. The corresponding experiment has been set up to be used
60
50
40
SPL [dB]
P1: IBE
0521871441c06
30
20
10
100
200
300
400
500
f [Hz]
Figure 6.100. Sound-pressure level at an observer point of x = 10 m, y = 10 m, and z = 1 m.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.7 INTERNAL FLOWS
as a simplified mirror producing noise in the passenger compartment. Many car manufacturers are concerned about passively reducing the noise emitted to the driver and
the passengers.
6.7 Internal flows

Philippe Lafon, Fabien Crouzet, and Jean Paul Devos
6.7.1 Introduction to internal flows
Among the many aeroacoustics phenomena that may occur in internal flows, the experience of the industrial context leads us to consider the following first:
r Acoustic fluctuations due to turbulence-generated noise at low Mach number,
r Self-sustained oscillations and flow acoustic coupling at low Mach number, and
r Aeroacoustic instabilities at high Mach number.
These phenomena are similar to the ones observed in free flows, but in the context
of confined flows, many features of these phenomena are enhanced; for example, flow
acoustic interactions are stronger.
The numerical approaches based on LES or very large eddy simulation (VLES) we
present in this chapter were developed in an industrial context. Thus, direct acoustic
simulations by compressible DNS are out of our scope.
Section 6.7.2 presents a hybrid approach based on the computation of incompressible LES and LEE. This approach is able to take into account the generation and
the propagation of acoustics disturbances due to the presence of control flow devices (diaphragms, valves, etc.) in industrial ducts. In Section 6.7.3, a direct approach based on the computation of nonlinear Euler equations is able to model the
flowacoustics interactions at low Mach numbers when there are feedback phenomena and acoustic resonances. In Section 6.7.4, the same set of equations is solved
in order to model transient aeroacoustic phenomena in high-Mach-number flows.
All these cases were defined and computed to obtain information about real problems
occurring in industrial configurations.
6.7.2 Computation of acoustic fluctuations due to turbulence-generated
noise at low Mach number
6.7.2.1 Source modeling
The modeling approach used here is a hybrid one. Aeroacoustic methods that separately
consider aerodynamic and acoustic modeling are called hybrid approaches.
This work was carried out within the framework of the PREDIT research program supported by the
French government.
349
13:36
P1: IBE
0521871441c06
350
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Approaches based on Lighthills analogy are of this type, but because they rely on
an integral solution they are not well suited for confined problems. That is why Euler
equations have been chosen since the first developments (Bechara et al. 1994; Bailly,
Lafon, and Candel 1996; Bogey et al. 2002).
A source term has been defined on the momentum equation of the Euler system.
This source was built to match Lilleys equation for a one-dimensional sheared flow.
This source term is expressed by

#
u it
u it
u jt
.
(6.64)
Si = u jt
x j
x j
The subscript t in the source term expression denotes turbulent fluctuating quantities. This source term may be calculated by several methods. An approximate approach
is possible by using stochastic modeling for turbulent fields (Bechara et al. 1994; Lafon
1997; Longatte, Lafon, and Candel 1998; Bailly and Juve 1999). Here we use LES data
to calculate the source term.
The source term can also be formulated in terms of vorticity (Howe 1975). In this
case, it is more appropriate to use a vortex method for computing the flow (Hofmans
1998).
Note that in supersonic cases the knowledge of a broadband turbulent source term
is not necessary because, in such situations, the noise radiation is mainly due to the excitation of instability waves. Thus, a compressible LES with suitable inflow excitations
can be used (Dong and Mankbadi 1996).
6.7.2.2 Flow and acoustic computations
LES computation
LES computations are carried out with the ESTET code developed by the R&D Division
of Electricite de France. It solves LES equations for incompressible flow using a conservative finite volume method on a Cartesian grid. The classical Smagorinski model
is used, and periodic boundary conditions are imposed in the transverse direction.
Acoustic computation
The source terms need to be calculated from the transient LES results. It was shown by
Crouzet et al. (2002) that these results have to be very clean because any nonphysical
disturbances in the LES results lead to large errors in the source terms.
Acoustic computations are carried out with the EOLE3D code developed by the
R&D Division of Electricite de France. It solves the LEE on a structured Cartesian grid
using a DRP scheme proposed by Tam and Webb (1993).
6.7.2.3 Results in the case of a diaphragm in a duct
The configuration of a 3D duct obstructed by a 2D diaphragm is chosen because some

experimental data are available. This simple case is considered to behave, at least
qualitatively, like many industrial cases when turbulent acoustic sources are generated
downstream of flow control devices (e.g., diaphragm, valves, etc.).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
351
6.7 INTERNAL FLOWS
80 mm
35 mm
150 mm
3 mm
350 mm
40 mm
Figure 6.101. Aerodynamic computational domain. Reprinted by permission of the American Institute
of Aeronautics and Astronautics, Inc.
The computational domain for the LES computations is presented in Figure 6.101.
The computational domain for the acoustic that is presented in Figure 6.102 is much
longer, and thus the acoustic far field can be obtained in the duct. The LES grid is
refined close to the diaphragm, whereas the acoustic one is uniform.
Figure 6.103 displays a snapshot of the instantaneous longitudinal velocity for
U = 14 ms1 .
Figure 6.104 displays the acoustic results. The upper graph shows the evolution of the
acoustic power radiated by the diaphragm sources with respect to the average velocity
in the duct. The results are compared with the classical U 4 law, and the agreement is
good. The lower graph compares the computed and the experimental spectra of the
acoustic power.
6.7.3 Computation of flow acoustic coupling in low-Mach-number
ducted flows
Fluidacoustic coupling
In confined flow, self-sustained oscillations due to feedback phenomena or acoustic resonances are very common. The basic physics and analysis may be found in Rockwell
and Naudascher (1978). More recently, several computational studies analyzed these
kinds of flows in free or confined configurations (e.g., Dequand, Hulshoff et al. 2003;
Gloerfelt, Bailly, and Juve 2003; Lafon et al. 2003). In such flows, large coherent eddies drive the whole dynamics. So, computing LES or VLES in these cases seems to
be relevant. Furthermore, such flow dynamics and associated acoustics are nonviscous
phenomena. Thus, one might expect to calculate these configurations by means of nonlinear Euler equations, provided some suitable wall treatment is introduced if boundary
layers have a great influence on flow development.
80 mm
35 mm
400 mm
3 mm
1200 mm
40 mm
Figure 6.102. Acoustic computational domain. Reprinted by permission of the American Institute of
Aeronautics and Astronautics, Inc.
13:36
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.103. LES velocity field; longitudinal component (U = 14 ms1 , t = 6.6 102 s). Reprinted
by permission of the American Institute of Aeronautics and Astronautics, Inc.
120
Experiments
Computation
4
U law
115
Acoustic Power (dB ref = 10 12 W )
110
105
100
95
90
85
80
75
0
10
10
Mean velocity (m/s)
10
80
EOLE3D computation
Experiments
70
60
12
352
CUFX063/Wagner
Acoustic power (dB) ref 10
P1: IBE
0521871441c06
50
40
30
20
10
100
200
300
400
Frequency (Hz)
500
600
Figure 6.104. Acoustic results: acoustic power radiated by the diaphragm with respect to the mean
velocity (top) and acoustic power spectrum for U = 14 ms1 (bottom). Reprinted by permission of the
American Institute of Aeronautics and Astronautics, Inc.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.7 INTERNAL FLOWS
Figure 6.105. Geometry of the cavity duct system.
Fluidacoustic computation using nonlinear Euler equations
The Euler equations are solved on a 2D curvilinear grid using a conservative finitevolume method (Lafon and Devos 1993). The total variation diminishing algorithm
developed by Harten (1983) is applied. For inlet and outlet boundary conditions, characteristic variables, and compatibility equations are used. As is proposed in Poinsot and
Lele (1992), additional terms are introduced in the equations in order, first, to drive the
solutions toward a target state at infinity and second to ensure absorbing conditions for
acoustic waves.
Results in the case of a ducted cavity
The geometry of the confined cavity we have studied is shown in Figure 6.105. This
configuration is a 2D case modeling a real industrial problem that has been also treated
with 3D numerical methods and 2D and 3D experiments on a scale model.
The values of the parameters defining the case are as follows:
r d = 0.05 m
r e = 0.008 m
r h = 0.02 m
r H = 0.137 m
r L = 0.073 m;
the sound speed is 343 m s1 . The U0 mean velocity in the duct is 62.8 m s1 . This
value of the velocity is the one that produced the strongest fluctuations measured in the
experimental models.
Figure 6.106 shows snapshots of the time evolution during an oscillation period of
the pressure in the whole duct and the vorticity in the cavity. Note that the evolution
of the flow in the cavity and of the acoustics in the duct are coupled: it is shown that,
when the eddies interact with the downstream corner of the cavity (t0 , t0 + T ), the
acoustic response in the duct is clearly organized in accordance with the first transverse
mode pattern. Experimental results and more detailed numerical analysis showed that
353
13:36
P1: IBE
0521871441c06
354
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.106. Snapshots of the pressure in the duct (left) and the vorticity in the cavity (right) during a
period of the oscillation. From top to botton: t0 , t0 + T/4, t0 + T/2, t0 + 3T/4, t0 + T .
the fluctuations in the duct are at a maximum when this coupling occurs (Devos and
Lafon 2003; Lafon et al. 2003).
6.7.4 Computation of aeroacoustic instabilities in high-Mach-number

ducted flow
6.7.4.1 Presentation of the test case
Industrial applications involving high-pressure steam flows are controlled by systems

like pressure-reducing valves or diffusers. Downstream of these sudden expansions,
the flow may become unstable by strong aeroacoustic coupling. To determine the fundamental mechanisms of these instabilities, researchers have carried out experimental
studies. Numerous results are given and analyzed in the paper of Meier et al. (1978).
Several different cases are presented, but the case of a rectangular duct with a sudden
enlargment is very typical of many industrial configurations. The geometry of the case
is presented in Figure 6.107. The parameter that controls the evolution of the flow is the
ratio of the upstream pressure on the downstream pressure, = Pupstream /Pdownstream . At
high values of the flow is entirely supersonic. When is decreased, the flow becomes
unstable and a cyclic oscillation appears. Complex flow patterns can be characterized
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
355
6.7 INTERNAL FLOWS
H2
H1
H1
= 0.3
H2
L = 7.23
H2
L = 240 mm
Figure 6.107. Geometry of the sudden enlargment. Reprinted by permission of the American Institute
of Aeronautics and Astronautics, Inc.
involving normal and oblique shock waves, formation of supersonic jets, reentry of
fluid at the downstream boundary, and so on.
6.7.4.2 Results in the case of a sudden enlargment in a 2D duct
Results for two pressure ratios are presented in Figures 6.108 and 6.109: 1 = 5.5
and 2 = 2.65. For the highest pressure ratio the flow is not fully detached from the
walls and the flow oscillates between two states: one involving a strong normal shock
associated with two supersonic jets on the sides and the other one in which the normal
shock disappears. For the lowest pressure ratios, the flow is detached from one wall and
oscillates from a state in which shock patterns are still present to another one in which
the structures are fully destroyed.
6.7.5 Conclusions for internal flow prediction
We have presented methods able to deal with aeroacoustic phenomena in internal flows.
These methods, based on transient flow computations (LES and VLES), give data
for defining source terms in hybrid approaches for subsonic flow. When applied to
Figure 6.108. Snapshots of the Mach number for 1 = 5.5. Reprinted by permission of the American
Institute of Aeronautics and Astronautics, Inc.
13:36
P1: IBE
0521871441c06
356
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.109. Snapshots of the Mach number for 2 = 2.65. Reprinted by permission of the American
Institute of Aeronautics and Astronautics, Inc.
compressible computations, these methods are able to catch flow and acoustic feedback
and aeroacoustic coupling directly. In comparison with free-space situations, internal
configurations have the following characteristics:
r There is no need to have a very large grid in order to deal with the acoustic far field.
r The influence of solid boundaries is of course greater both for flow computation
(boundary layers) and for acoustics computation (resonances).
6.8 Industrial aeroacoustics analyses

Fred Mendonca
6.8.1 Introduction to industrial aeroacoustics analyses
It is true of aeroacoustics, as with most simulation-based engineering, that the engineer

is pulled between two extremes: whether to undertake fully rigorous analyses with
all the implied resource demands or, instead, to perform analyses of limited scope
with quantified risk because the time scales in the industrial design cycles require as
much. When the analyses involve large-eddy simulation (LES), this dilemma is brought
sharply into focus. In this section we attempt to qualify important questions relating to
the aeroacoustics simulation process to help the analyst make value judgements before,
during, and after LES calculations, concentrating on data analysis and interpretation,
because the processing of information ultimately defines the true worth of the outcome.
This section should be read with an industry perspective. It aims to extract maximum
value from the modeling process by gathering useful information from computational
fluid dynamics (CFD) analyses to maximize knowledge of the acoustics properties
of simulated systems. In industry, these systems are diverse, covering many sectors
most notably transportation. Some current topics of interest are listed in the following
paragraphs.
The commercial aerospace operators main noise-associated concerns are well
documented: high-lift devices (Agarwal and Morris 2004), landing gears (Souliez
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
6.8 INDUSTRIAL AEROACOUSTICS ANALYSES
et al. 2002), jet nozzles (Morris and Farassat 2002), and cabin and cockpit climate
control devices (Mendonca et al. 2005). Military operations focus on the structural
and aerodynamic instabilities of weapons and weapons bays (Henshaw 2000; Allen,
Mendonca, and Kirkham 2005).
Automobile manufacturers and component suppliers are keen to demonstrate expertise in applied technologies for noise minimization in several areas. External components such as sideview mirrors (Siegert, Schwarz, and Reichenberger 1999), A-pillars,
and wiperblades directly excite the vehicles panels, which transmit noise to the drivers
ears. Under the hood, noise from turbomachines such as the cooling fan (Algermissen,
Siegert, and Spindler 2001) and turbocharger can be heard above the idling engine.
Ducting and climate control system components such as the blower fan (Barone
et al. 2003; Read et al. 2004; Dubief et al. 2005) and flaps introduce noise directly
into the passenger compartment. Open apertures such as sunroofs and windows suffer
buffeting caused by oscillation of the separated shear layer at the apertures leading edge.
In what follows, we exploit a variety of CFD and postprocessing approaches, covering steady-state and transient calculations pertinent to all of the preceding topics of
interest. The exploitation leads toward an expert-system analysis template for industrial aeroacoustics. Postprocessing of steady-state CFD to resynthesization noise or to
gain an approximate measure of the grid frequency cutoff are useful precursors to fully
time-accurate transient calculations. Later, we turn our attention to noise propagation.
Some examples of the hybrid approach coupling transient CFD to a separate noise
propagation analysis are presented, but we start with a reminder of some basic building
blocks in the process.
6.8.2 Preliminary considerations
6.8.2.1 Hybrid nature of flow-noise generation and propagation mechanisms
We begin with the premise that general purpose, commercially available CFD codes,
widely used in industry and capable of producing quality flow solutions for complexgeometry mainstream applications using second-order accurate discretization methods
in space and time, are sufficiently accurate to predict flow-induced noise sources
through direct capture of the dynamic flow structures but are insufficient to propagate
the associated noise to the far field (see also Section 6.6). This limitation arises because
the spatial discretization is only of second-order accuracy, but this is not necessarily a
major impediment. In fact, the two mechanisms, noise-source generation and propagation, are conveniently separated. Acoustic pressures are typically orders of magnitude
smaller than the flow-fluctuating pressures (pseudo-sound). The flow generates noise
and influences its propagation via, for example, convection and diffusion. Conversely,
the sound field generally does not generate flow. This useful dissection allows for the
flow analysis and the propagation analysis to be performed separately and consecutively
in that order and is referred to as the hybrid method for aeroacoustic simulation.
One notable exception is aeroacoustic resonance, in which feedback occurs between
the fluctuating flow and propagating pressure waves. This circumstance only occurs
when the geometry contains duct or cavity length scales that are equivalent to the
357
13:36
P1: IBE
0521871441c06
358
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
acoustic wavelengths excited by the flow. Examples are resonators in automotive intake
and exhaust systems (in which these devices are purposely introduced), passenger cabins
with open windows or sunroofs (in which the resonance is unintended and distinctly
unpleasant), and aerospace cavities such as those housing the landing gear and weapons
stores. Aeroacoustics resonance has been successfully simulated by Inagaki et al. (2002)
and is further illustrated here in Section 6.8.4.1 to stress that compressibility is the
mechanism by which acoustic-flow feedback is facilitated.
6.8.2.2 Domain size, meshing strategy, and temporal resolution
The engineer should include in the CFD analysis sufficient volume around the subject
geometry to ensure
r all the flow noise mechanisms of interest are contained within,
r the boundary values are known, and
r that the numerical treatment at boundaries allows for wave transmission, or
r that the boundaries are placed far enough away for artificial reflections to be dissi-
pated before they influence the noise-generation mechanisms.

Meshes should as far as possible be orthogonal and isotropic (unit aspect ratio) the
former to preserve the spatial accuracy, and the latter to ensure that the shape of the noisegenerating flow structures captured by the simulation is not distorted. Some readers will
take the assertion of cell isotropy to be excessive, imposing an unreasonable burden on
resolution especially in the boundary layer: instead, the choice of cell structure should
be consistent with resolution of target-flow features. For example, vortex streaks are
stretched features confined to the boundary layer; therefore, streamwise cell stretching
is a natural and consistent choice. Isotropic meshes, aimed at resolving authentically
three-dimensional structures in the bulk flow, do not preclude the use of high-aspectratio cells in the boundary layer aimed at resolving streaks.
Time-accurate (transient) simulations will be able to resolve frequencies to a maximum of the inverse of twice the time-step increment, corresponding to the Nyquist
requirement of two points in a wave form. The author prefers to recommend a more
conservative ten points per wave form to give a better chance for the amplitude, and
hence the noise intensity, to be resolved.
6.8.3 A two-step CFD modeling process (steady-state and transient)
We define two steps in the modeling process commencing with CFD in steady state,
after which useful postprocessing is performed as a precursor to calculations in the
transient mode.
6.8.3.1 Steady-state calculations
Synthetic reconstruction of the turbulent fluctuations from steady-state calculations can

lead to useful approximations of relative noise-source magnitudes and locations. One
such method is described in Section 6.8.3.2. Similar synthetic reconstructions may be
used to account for the broadband contributions lost in the SGS (see Section 4.2.7),
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
359
or to introduce perturbations at the domain inlet with time-varying coherent structures

representative of turbulence statistics (Benhamadouche, Jarrin, and Laurence 2003).
Through postprocessing of steady-state RANS solutions, a mesh frequency cutoff
(MFC) measure is calculable. This simple method of Mendonca et al. (2005) gives a
reasonable approximation of the limiting frequency the mesh is locally able to resolve.
The validity of the method is verified through direct comparison with the predicted
spectra from a transient solution and measurements.
6.8.3.2 Steady-state postprocessing
Approximation of noise sources through synthetic reconstruction
Standard turbulence closures include transport equations for the mean kinetic energy
k of the mean turbulent structures and their dissipation rate , or other turbulence
quantities from which k and may be derived. Together with the mean flow, this
information is used to resynthesize the fluctuating velocity (Debatin 1999; Kraichnan
1970) vi (x, t) based on a Fourier decomposition of the von Karman energy spectrum
E(, t) across wave numbers as follows:

3 2
E(, t)d = vi = k,
(6.65)
2
0

2 E(, t)d = .
(6.66)
2
0
Reconstruction of the fluctuations, sometimes referred to as stochastic noise generation and radiation (SNGR), arises from the wave-number decomposition
vi (x, t) =
N

2
vn cos [n (x V t) + n + n t] en .
(6.67)
n=0
This equation describes the turbulent fluctuation as the sum of N harmonic waves n
with a specific orientation en , amplitude n , turbulent frequency n , convective speed
V , and phase shift n that have been selected from a set of appropriate probability
density functions.
The noise characteristics associated with the fluctuations can be inferred through
the Lilley (1969) acoustic analogy. If we neglect viscous and entropy effects, the Lilleyequation source term (left-hand side of Equation (6.68) below depicting gradients of
the instantaneous velocity vi, j,k ) can be separated into various contributions from the
turbulent fluctuations (B), the mean flow distortion (A), and mixed meanturbulent
gradients (C). From dimensional analysis, the largest contribution comes from (B):
B
A
%&
'
%&
'$
vk v j vi
vk v j vi
= 2
2
2
xi xk x j
xi xk x j
xi xk x j
vk
v j
vi
%&
'
vk v j vi
vk v j vi
6
.
6
xi xk x j
xi xk x j
(6.68)
13:36
P1: IBE
0521871441c06
360
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.110. Opel 2004 Astra (courtesy of Opel AG).
Figure 6.110 illustrates an isosurface of the power of the synthetic source represented
by term (B), which henceforth will be referred to herein as the Lilley source, arising
from the steady solution flow over a full automobile.
There are two obvious benefits to using this methodology as a postprocessing
tool in steady-state calculations. First, the relative magnitudes of sources are identifiable; this is useful either to compare one design against another or to rate the relative source magnitudes from different components on the same vehicle. We see in
Figure 6.110 that the largest shear-noise sources originate from the sideview mirror,
wiperblade cavity, A-pillar, wheelwell, and wake. Secondly, the method provides information about the location and expanse of the sources. Thus, in the preparation of a
transient analysis on more detailed or localized analyses, it is useful in deciding where
to refine the mesh, where to place domain boundaries, and where to locate monitoring
points.
The often-cited idealized wing-mirror study of Siegert et al. (1999) demonstrates
the suitability of this method well. From Figure 6.111, the analyst is able to identify clearly those regions in which the Lilley source indicator dominates and is therefore able to plan a concentration of mesh density there. The inset in Figure 6.111
shows four levels of successive refinement, the finest level adjacent to the mirror surface, with refinements extending into the separated wake region The analyst should
also be aware of the limitations of this technique. As a steady-state method, it is incapable of representing noise sources that arise from large-scale transient features
(large-scale here means larger than the grid scale) such as vortex shedding or moving
boundaries.
Mesh frequency cutoff (MFC) measure of Mendonca
The main value in performing an MFC analysis as postprocessed from a steady-state

solution is to be able to decipher in advance of a transient calculation whether the chosen
grid has sufficient resolution to capture the turbulence flow structures in the frequency
range of interest.
Given a cell dimension and local turbulent kinetic energy k, the smallest length
scale of a turbulent eddy structure captured by the mesh is 2; its associated isotropic
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.111. Lilley turbulence shear-source distribution illustrated by isosurfaces for the idealized
wing-mirror example of Siegert et al. (1999).
(
2
fluctuation velocity is
k . Therefore, the maximum frequency f MC reasonably
3
resolved by the local grid spacing is
(
2
k
3
.
(6.69)
f MC =
2
The choice of k, the mean turbulent kinetic energy contained by the mean turbulent
structures, to represent the turbulence velocity fluctuation is easily justified. Most widely
used RANS turbulence models solve transport equations for it directly; for those that
do not, it is easily derivable. The transport equation for k is derived directly from the
unsteady NavierStokes equations and therefore contains contributions from all the
important mechanisms: convection, diffusion, production, and dissipation. So long as
the mean flow features are well captured in the modeling, the k solution also tends to
be reasonably grid independent.
This measure is used for aeroacoustics in preference to one derived from the turbulence time scale (i.e., the ratio of kinetic energy to its dissipation rate k/, which is
widely used to estimate the time scales for fully resolved LES). This is more representative of the dissipative scales, which, in frequency and energy content, are, it is hoped,
orders removed from the human hearing range or peak sensitivity.
Because this measure is derived from a steady-state solution, some limitations are
inherent. The frequencies associated with time-varying, large-scale motions such as
vortex shedding, which convect through the mesh, will not be accounted for. Instead,
its usefulness is to approximate the frequencies of the turbulence scales modeled in
RANS that become resolved in LES. In other words, this measure is more valid for the
broadband and less so for narrow-band excitations (Mendonca et al. 2005).
361
13:36
P1: IBE
0521871441c06
362
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.112. Mesh frequency cutoff (MFC) estimate; idealized wing-mirror example of Siegert
et al. (1999).
The maximum frequency f MC represents an important step in the analysis process. For this value to be correctly applied and interpreted, certain conditions should
be met. First, the mesh should be isotropic (unit aspect ratio) as recommended earlier; in this case, the cube root of the volume becomes an appropriate measure of
the hexahedral cell dimension (appropriate modifications are required for perfect
tetrahedra, polyhedra, etc.). Secondly, the boundary conditions applied to k, especially at the inflow, should be realistic. Thirdly, the mesh must be fine enough to
capture the important mean-flow features. Most of the cases chosen in this section
to support the MFC measure, although quite typical of industrial applications, are lowMach-number flows. Further validations are necessary for transonic and higher-Machnumber flows.
With respect to the idealized wing mirror of Siegert et al. (1999), Figure 6.112
shows the local mesh frequency cutoff distribution f MC . As expected, we observe
discontinuities in the measure across the mesh refinement interface. Close to the mirror
surface, where the grid is most densely packed, the measure suggests that frequencies
well in excess of 1 kHz will be resolved. In the separated shear layer, the cutoff frequency
is lower (somewhere in the region 800900 Hz). Practical justification of this measure
through comparison with a fully time-accurate transient calculation is given in Section
6.8.3.3 (see Figure 6.113).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.113. Predicted versus measured pressure spectra; idealized wing-mirror example of Siegert
et al. (1999).
6.8.3.3 Transient calculations
Transient calculations permit the direct computation of the noise sources (within the limits of the modeling used), and are useful in determining the spectra in the noise-source
region directly (Figure 6.113 is typical) or accumulating flow data for compilation of
equivalent noise sources for noise propagation to the far field.
Figure 6.113 shows the computed versus measured pressure spectra at a point in
the wake of the idealized wing mirror. We observe a good qualitative and quantitative
(within 2 dB) prediction over a wide frequency range up to about 900 Hz, beyond
which the prediction trails off from the measured spectrum. This is indicative of a
lack of spatial resolution and is usually manifested by an underprediction of soundpressure level. The energy contained by (unresolved) eddy structures smaller than the
grid dimensions is lost in the subgrid scale. Section 4.2.7 demonstrates that there are
possibilities for reconstructing the spectral contribution from the subgrid scales using
SNGR, for example, but note that such methods also have inherent limitations.
It is useful to note that, with the ability to identify the cutoff value from a steadystate calculation, the analyst is afforded a useful additional step in the process. He or
she is free to refine the mesh before the transient simulation by direct extrapolation
of the MFC measure. For example, if the measure indicates 500 Hz when 1000 Hz is
targeted, the user is obliged to increase the spatial resolution locally by a factor of 2 in
all directions.
With respect to the choice of time-step size, though the Nyquist criterion suggests
the minimum frequency resolvable in a time-based simulation is half the inverse of
363
13:36
P1: IBE
0521871441c06
364
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
the time step, a more conservative recommendation is made here: that is to use onetenth of the inverse of the time step to allow for a better capture of the waveform and
amplitude of any oscillating signal. The recommendation is supported by Mendonca
et al. (2005), who have shown that a cutoff proportional to the time-step size is observed
in the predicted spectra. In contrast to the underpredictions that result from inadequate
spatial resolution, the temporal cutoff is manifested in the simulated spectrum by a leveling off in the sound-pressure level. This can be an indication that the solution is unable
to distinguish between, or separate out, the energy content of the resolved structures
across wave numbers that are higher than the applied temporal resolution allows.
The following sections provide an overview of some additional key issues: turbulence modeling in unsteady flow, inlet perturbations, discretization practices, and
global-to-local domain boundary mapping. We discuss the practicability of turbulence
modeling for industrial applications.
Turbulence modeling
We now offer some limited examples of turbulence modeling for transient flows. Despite
the title of this book, we start with RANS models applied to unsteady flows, so-called
URANS, to provide a suitable comparison with LES-based models. We see that the
former have inherent limitations in their ability to predict broadband flow excitations.
RANSLES combinations are capable of overcoming these limitations by invoking
LES where it is necessary while retaining the economy of RANS where LES would
otherwise be unaffordable. Necessary, for aeroacoustics, is taken to mean the bulk
flow particularly separated flow regions rather than boundary layers.
The studies of Mendonca et al. (2003), Allen and Mendonca (2004), and Allen
et al. (2005) into the M219 high-speed cavity flows note that URANS performs well
where flow excitations result in a dominant narrow-band tone but potentially fails
where multiple modes of roughly the same order of magnitude exist, or when the flow
excitations are broadband or both. In this example, the performance of one particular
hybrid approach, DES (Spalart et al. 1997), is compared with a URANS model for a
well-documented high-speed cavity flow (Henshaw 2000). A range of mesh sizes has
been assessed ranging from 1 to 2.8 million cells, containing 0.8 to 2.5 million cells in
the cavity and shear layer, respectively, and showing minimal mesh dependency in the
results.
Figure 6.114 illustrates the instantaneous flow fields from DES and URANS simulations in a section of the three-dimensional, 4-in.-deep cavity configured without bay
doors (L/D = 5, W/D = 1) at M = 0.85. Both calculations were performed on the
same mesh, which was designed for URANS with wall functions, using the same timestep size, 2 105 s. The computational times for the DES and URANS calculations
are virtually identical. Many more eddy structures are prevalent in the DES solution,
implying greater fidelity in capturing broadband noise.
This is confirmed by the measures of rms pressure along the cavity ceiling, illustrated
in Figure 6.115, and spectra at the midpoint along the ceiling shown in Figure 6.116.
URANS fails to pick up the second and fourth (Rossiter) acoustic modes observed at
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.114. Mach 0.85 cavity: symmetry-plane snapshot at t = 0.3 s; DES/k (top), and
URANS/k (bottom).
approximately 380 Hz and 800 Hz, respectively, demonstrating a major failing of the
model. DES, by contrast, predicts the background levels and modal behavior extremely
well.
Inlet perturbations
At the inlet of the computational domain, mean flow parameters are usually well known;
these are a necessary requirement for RANS simulations, and we note happily that
experimentalists are well attuned to this need and usually provide high-quality data
needed for simulation. In LES, by contrast, correct modeling can depend strongly
on the perturbations about the mean flow. For example, the primary breakup of the
liquid core discharging from a circular duct into the atmosphere, as is typical in a
spray nozzle, requires instabilities from the duct flow to initiate breakup of the free
surface (Buonfiglioli and Mendonca 2005), as shown in Figure 6.117. Instabilities may
be accounted for by a full, transient LES precursor calculation in the duct, but this
procedure is inordinately expensive and, for the moment, impractical in any industrial
application.
Cheaper solutions are becoming better understood and more widely applied. These
are based on synthetic excitations, which, unlike white noise, are correlated in
space and time. One such method, described by Benhamadouche et al. (2003), uses
knowledge of the local mean-flow turbulence, k and , to impose synthetic vortical
structures in the plane of inlet boundary and streamwise perturbations normal to it. The
synthetic structures are shown to be self-sustaining and also to generate fully developed
turbulence within a few duct dimensions downstream of the perturbed inlet.
365
13:36
P1: IBE
0521871441c06
366
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.115. Overall (a) and band-limited (b) Prms along cavity ceiling centerline.
Boundary mapping global to local domain interfacing
Irrespective of the speed of computing, it will always be desirable with LES to focus
on localized domains. Localized domain analysis depends crucially on the provision of
well-defined boundary conditions primarily at the inlet with respect to mean profiles.
In the following example (Read and Mendonca 2004), mapping of boundary conditions from a steady-state calculation in a global domain to a transient simulation in a
localized domain is described. The focus of this case is the aeroacoustic behavior of the
Audi A2s wing mirror. A steady-state calculation was first made around the full vehicle. A localized domain mesh was then generated (seen in dark color in Figure 6.118)
containing approximately 3.8 million cells; many more would be required for LES on
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.116. PSD (kPa2 /Hz) at location x/L = 0.45.
the full vehicle. The steady-field values were mapped onto the boundaries of the local
domain that is, velocities at the upstream and four adjacent side planes and pressure
at the localized domain outlet; all were held at the steady values for the duration of the
transient calculation.
Figure 6.119 shows a snapshot of the resulting transient DES calculation, which
is dominated by large-eddy structures in the wake of the wing mirror. The pressure
spectrum at a point on the mirror face is shown in Figure 6.120 and, when compared
with the measurement, gives immediate credence to this methodology. Additionally,
mapping of the turbulence levels together with synthetic perturbations allows for even
nonsteady inlet conditions.
Figure 6.117. Diesel injector primary liquid spray breakup; liquid-free surface with synthetic inlet perturbation (top) and without inlet perturbation (bottom).
367
13:36
P1: IBE
0521871441c06
368
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.118. Audi A2 full-vehicle geometry with localized domain shown in dark (courtesy of Audi AG).
6.8.3.4 Transient postprocessing
For all transient flow calculation aimed at aeroacoustic sources prediction, we recommend that some basic sanity checks be performed. By exemplification on industrial
cases, these sanity checks also justify the recommended processes defined earlier in
this section. We address data sampling in an attempt to highlight pitfalls and necessary
minimum requirements.
Transient data sampling and spectral postprocessing
To illustrate some key features and pitfalls in the processing of transient data, we return
to the high-speed cavity case. The unsteady pressures (measured and simulated) are
Figure 6.119. Instantaneous velocity magnitude field (courtesy of Audi AG).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
369
Figure 6.120. SPL against frequency at Microphone 4 (courtesy of Audi AG).
inherently noisy, as is apparent from Figure 6.121 corresponding to the samples at the
downstream end of the cavity at x/L = 0.95.
The experimental data (Henshaw 2000) are compiled using 34 windows (rectangular
boxcar with zero overlap), each of 0.1 s duration and corresponding to approximately
50 freestream passes over the length of the cavity. Overlapping of the data, using for instance Hanning windowing with 50% overlap, has negligible effect. From the rms pressure curves of the experimental data in Figure 6.122, it is evident that the minimum sample of 1.0 s must be processed to obtain a result comparable to the original 3.4-s sample.
From these assessments, we can begin to make the following recommendations for
this case. LES simulations should provide data equivalent to an elapsed time of 0.5 s,
allowing the initial 0.1 s to be discarded and the remaining 0.4 s to be processed. This
translates into the first 50 freestream passes being discarded and a minimum of 200
further passes required for statistically steady data to be produced. Shorter samples
should be interpreted with care.
Figure 6.123 shows the Prms curves from the simulation (DES) over a range of
sample periods averaged in the same manner as the experimental data but with the first
0.1 s of the sample discarded to eliminate startup effects. The curves begin to converge
from the 0.2-s sample onward, and a sample with a total duration of 0.4 s is required to
obtain a negligible level of deviation between windows.
Pressure (Pa)
Pressure (Pa)
28000
28000
24000
24000
20000
20000
16000
16000
12000
12000
8000
8000
4000
4000
4000
4000
8000
12000
0.0
8000
0.05
Experiment
0.1
0.15
0.2
0.25
12000
0.3
0.0
Time (s)
DES
0.05
0.1
0.15
0.2
0.25
0.3
Time (s)
Figure 6.121. Experimental (left) and simulated (right) pressure trace to 0.3 s at x/L = 0.95.
13:36
P1: IBE
0521871441c06
370
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.122. Sampling effects on overall Prms (kPa) along the cavity ceiling for M219 experimental
data.
6.8.4 Postprocessing through acoustic coupling
As stated in Section 6.8.2.1, the assumption of the hybrid nature of aeroacoustics allows
the separation of these two physical mechanisms (source generation and propagation)
during simulation. Two classes of simulation method are enumerated.
The first involves integrals (volume, surface, line) using information that comes
directly from transient or even steady-state CFD calculations that are enumerated here
Figure 6.123. Sampling effects on overall Prms (kPa) along the cavity ceiling for CFD data.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
mainly through reference for brevity. Some limitations are inherent; that is, they do
not account for multiple surface reflections; they do not allow for surface or volume
attenuation (damping) and therefore are not generally advisable for internal flows; and
they do not generally allow for refraction effects owing to interaction of the acoustic waves with nonuniform mean flow (particularly shear layers). Morris and Farassat
(2002) described the use of acoustic analogies and alternative methods for jet noise
predictions, all of which are based on steady-state (RANS) k-based prediction, and
concentrate on directionality in the prediction of the propagated jet noise spectrum.
Also based on RANS calculations, Agarwal and Morris (2004) demonstrate propagation of noise generated in the slat-cove region of a high-lift device through the mean
flow into the far field using only three empirical constants. Greschner et al. (2004)
have made a study of different RANS-based DES calculations of the noise generated
and propagated in a complex transient interaction between a cylinder wake and downstream aerofoil, illustrating the use of permeable Ffowcs WilliamsHawkings surface
integrals.
The second class involves frequency-domain-boundary or finite element computations formulated around acoustic analogies, that compute the propagation, including
reflections and surface-volume damping, of aeroacoustic sources that have been derived from transient CFD calculations. Recently, efforts have been made to incorporate
interaction of the acoustics with mean-flow variations, but as yet these have not been
validated fully. The next section presents some key applications using this class of
source-propagation coupling.
6.8.4.1 Frequency-domain methods
Frequency-domain methods offer an efficient methodology for noise propagation simulation using commercially available software. They require input of equivalent sources
such as monopoles, surface or rotating dipoles, and quadrupoles. It is typical for the
latter two types to be compiled from synthesized sources or at best from transient LES
simulations that contain modeled contributions from both narrow-band and broadband
sources.
Surface and volume sources
In the following example describing one-way coupling between a precursor CFD calculation and frequency-domain acoustics postcalculation, important features are demonstrated with the primary objective of simulating resonance in a cavity.
r First, the CFD calculation has the following characteristics:
It captures broadband and narrow-band excitations.

The frequency of the narrow-band oscillation (due to the shear-layer fluctuations
at the neck of the cavity) is linearly proportional to the cavity bypass velocity.
The amplitude of the narrow-band oscillations maximizes at the cavity resonance
frequency.
371
13:36
P1: IBE
0521871441c06
372
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.124. Resonator geometry: application challenge from BEHR GmbH in the DESTINY-AAC
(Detached Eddy Simulation for the Transportation Industry Aero Acoustics) project (courtesy of BEHR
GmbH).
The underlying second-order numerics are insufficient to amplify any excitation

at the cavity resonance frequency. Virtual microphones in the cavity do not detect
any amplification of power at the cavity resonance frequency except when the
shear-layer oscillation corresponds with it.
r Second, the characteristics of the acoustics calculation are as follows:
Volume sources from the CFD containing both broadband and narrow-band power
are exported onto the acoustics mesh.
Subsequent solution of the propagation of the sources reproduces an amplification
of the acoustic power at the cavity resonance frequency even though the shear
layer oscillates at a different frequency.
Figure 6.124 shows a schematic of the rig and typical flow conditions. On the
basis of the connecting rod length and diameter, box volume, and orifice correction
L = r/2, a resonance frequency of 358 Hz found analytically using
)
r2
a
(6.70)
f0 =
2
V (l + l)
agrees closely with the measured resonance.
In the CFD simulation, transient effects are dominated by the oscillation of the shear
layer at the neck of the resonator and subsequent downstream shedding (see velocity
contours in Figure 6.125). The system responds to three upstream bulk velocities (4,
8, and 12 m/s) with a near-linear increase in the frequency of the shear-layer oscillation (183, 305, and 538 Hz, respectively), but nonlinear influence on magnitude of the
fluctuations. The maximum of the three amplitudes occurs at the intermediate velocity,
wherein the frequency is closest to the resonators resonant frequency, clearly demonstrating the effects of compressibility in the interplay between velocity and pressure in
the CFD solution.
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.125. Velocity contours (top), pressuretime traces (bottom left), and spectral magnitude (bottom right) at three bulk velocities (4, 8, and 12 m/s) taken at the neck of the resonator (courtesy of
BEHR GmbH).
Coupling between a CFD solution (here obtained by STAR-CD) and a commercial

acoustics code (here ACTRAN) requires interpolated values of the Lighthill tensor at the
nodes of a volume acoustics mesh (Caro et al. 2005). The subsequent frequency-domain
calculation for the 8 m/s inlet bulk-velocity case, computed here with a resolution of
25 Hz, shows, in Figure 6.126, a magnification of the acoustic pressure at the resonant
frequency. This result requires that excitation from the CFD contain broadband power
spectral distribution with nonzero power at the resonant frequency. The only mechanism
through which this is possible is LES turbulence modeling, which is here achieved
using DES.
Fan sources
In this final example, the noise source and subsequent radiation from a subsonic compressor fan is modeled.
The noise from the blower fan, containing 47 blades rotating at 3770 rpm in a typical
automotive heating, ventilation, and air-conditioning system was assessed in isolation
(see Figure 6.127). Experimental measurements were taken in the semianechoic chamber at Denso Thermal Systems at the nine far-field points illustrated in Figure 6.127
around the rig comprising the blower housing (scroll), electric motor, cylindrical inlet
duct, and a rectangular outlet duct with a filter at the outlet.
373
13:36
P1: IBE
0521871441c06
374
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.126. Acoustic response for 8 m/s case at the microphone (courtesy of BEHR GmbH and Free
Field Technology, Caro et al. 2005).
A mesh consisting of approximately 1.5 million cells was initially run steady-state
using the multiple-rotating frames of reference method and then restarted transiently
using the k variant of DES with the mesh around the fan blades rotating relative to
the static volumes across a sliding interface. The flow field evolved to a pseudosteady
condition after approximately eight full rotations of the fan. Data were then output
for propagation in the commercial acoustic code SYSNOISE to the far-field observer
locations. A generalized Ffowcs WilliamsHawkings method was used. It requires the
storing of forces on an individual blade of the fan (taken to be representative) over
one complete rotation of the fan. Fuller details can be found in Barone et al. (2003).
Sources from the housing surface and the volume attenuation effects of the filter were
Figure 6.127. Experimental prototype with inlet cylinder and outlet filter removed and locations of
far-field monitors (courtesy of Denso Thermal Systems).
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
375
Figure 6.128. Comparison of steady-state RANS and snapshots from the DES calculation (courtesy
of Denso Thermal Systems; Barone et al. 2003).
neglected. Figure 6.128 shows a velocity magnitude in a section through the blower
from both the RANS and DES calculations at different times during one rotation of the
fan. The capture and subsequent convection of large eddies can clearly be seen in the
DES snapshots.
Figure 6.129 shows the predicted sound fields from the acoustics code on the exterior
of the blower housing and noise propagated to far-field locations. Figure 6.130 gives
a comparison of the computed result and experiment for the blade-passing frequency
(BPF). Agreement within 5 dB is observed. More importantly, the directional variation
of noise magnitude is correctly modeled.
(a)
(b)
Figure 6.129. Surface acoustic pressure (Pa) on the exterior model (a) and acoustic pressure in the
far field (b) (courtesy of Denso Thermal Systems and LMS International; Barone et al. 2003).
13:36
P1: IBE
0521871441c06
376
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
Figure 6.130. Computed and measured dB(A) levels at the nine microphone locations at the bladepassing frequency (BPF) (courtesy of Denso Thermal Systems and LMS International; Barone
et al. 2003).
6.8.5 Conclusions for industrial aeroacoustics analyses
Various techniques for adding value to the aeroacoustics calculations through CFD and
acoustics postprocessing have been assessed. These include steady-state analyses of
noise source locations, their relative magnitudes, and synthesized velocity fluctuations.
For transient calculations, the importance of using appropriate turbulence modeling for
tonal and broadband flow excitations has been stressed as has the need for thoughtful
meshing and calibration of discretization practices on the subgrid model. Justifications
have been offered for the use of the approximate steady-state methods through equivalent transient validations and, in particular, the use of a steady-state MFC analysis in
the assessment of mesh suitability for capturing acoustic-generation mechanisms in the
desired frequency range.
Even though the theories and methodologies applied to LES and aeroacoustics are
constantly evolving, the processes described herein (especially Section 6.8.3) offer a
framework in which to exploit CFD calculations on industrial applications in a controlled and thoughtful manner, each step adding a little more knowledge toward the
understanding of the aeroacoustics signature of complete systems or their components.
The usefulness of coupling CFD to acoustic propagation has been demonstrated, showing that one-way interaction between the flow and acoustics for a variety of industrial
applications can be sufficient.
6.8.6 Acknowledgments
The author acknowledges the involvement of Airbus, BAE SYSTEMS, QinetiQ, Bombardier Transportation, Audi, BMW, DaimlerChrysler, Opel AG, Air International,
BEHR, Denso, and Valeo, all of which have contributed significantly in the form of
experimental facilities, measurement data, people time, engineering insight, and project
13:36
P1: IBE
0521871441c06
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 20, 2006
funding. Willingness on the part of CFD code vendors to pursue detailed studies bridging research methods and commercial utilization on industrial cases and investing in
increased accuracy and robust validated modeling practices has been essential. The
benefits are embodied in CFD results from the commercial code STAR-CD (2004)
reported extensively here.
In particular, the author is grateful to colleagues A. Read, R. Allen, M. Buonfiglioli,
and collaborators T. Rung, K. Debatin, V. Joshi, F. Brotz, M. Schrumpf, M. Islam,
F. Klimetzek, F. Barone, P. Durello, F. Dubief, F. Werner, and A. Senf.
377
13:36
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Conclusions
Claus Wagner, Pierre Sagaut, and Thomas Huttl
Noise is becoming generally accepted as an environmental and even health hazard to the
population. Many noise sources are man-made especially transportation noise from
road traffic, aircraft, and trains. Other technical noise sources can also be annoying
such as wind turbines or cooling and climate systems. Governmental reactions to noise
problems and loss of public acceptance are laws, regulations, decrees, and guidelines
for the certification of noise-emitting vehicles and machines as well as temporal or
spatial limitations of their use. Aircraft and jet engine manufacturers in particular face
increasingly stringent noise requirements for near-airport operations worldwide.
Aerodynamic noise is one of the major contributors to external vehicle noise
emission. It also contributes to internal vehicle noise owing to the transmission of
the externally generated noise through structure and window surfaces into the cabin.
Aerodynamic noise becomes dominant at driving speeds exceeding 100 km/h when
compared with structure-borne, power train, and tire noise for which substantial noise
reduction has been achieved. The interaction of the flow with the geometrical singularities of the vehicle body produces unsteady turbulent flows, often detached, resulting
in an increased aerodynamic noise radiation.
To achieve these noise reductions, the European Commission, for example, has laid
out a series of research objectives. In order to meet the challenging goals proclaimed,
the design process needs to be supported by computer-based noise prediction tools.
7.1 Governing equations and acoustic analogies

In principle, the way to predict aerodynamic noise generation and propagation is
straightforward. The governing equations are those of mass, momentum, and energy
that involve more unknowns than equations. The additional information needed to obtain a complete set of equations is provided by empirical information in the form of
constitutive equations. An excellent approximation can be obtained by assuming the
fluid to be locally in thermodynamic equilibrium. This implies, for a homogeneous fluid,
that two intrinsic-state variables fully determine the state of the fluid. For acoustics it is
378
3:46
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
7.1 GOVERNING EQUATIONS AND ACOUSTIC ANALOGIES
convenient to choose the mass density and the specific entropy s as variables. When
there is no source of entropy, the sound generation is dominated by the fluctuations of
the Reynolds stress.
With respect to local thermodynamic equilibrium, it is reasonable to assume that
transport processes are determined by linear functions of the gradients of the flow-state
variables. This leads to a Newtonian fluid behavior and to the fully compressible and
unsteady NavierStokes equations, which can be numerically solved by means of direct
numerical simulation (DNS) without any additional physical assumptions or models.
All the dynamically active scales of motion must be represented in the simulation to
ensure reliable results. This means that the grid spacing x and the time step t must
be fine enough to capture the dynamics of the smallest scales of the flow down to the
Kolmogorov scale and that the computational domain must be large enough to represent
the largest scales. These criteria lead to a high computational cost, which is responsible
for the fact that DNS is nowadays almost only used for theoretical analysis and accurate
understanding of flow dynamics and is not a brute force engineering tool.
A classical technique for reducing the complexity of the simulation (and then lowering the computational effort) is to apply an averagingfiltering procedure to the Navier
Stokes equations, yielding new equations for a variable that is smoother than the original solution of the NavierStokes equations because the averagingfiltering procedure
removes the small scales and high frequencies of the solution. Although the high frequencies are no longer captured by the computation, their action on the resolved scales
can be taken into account via the use of a statistical model. The most popular averaging and filtering operations lead to the Reynolds-averaged NavierStokes (RANS)
equation, which relies on a statistical average leading to steady computations in the
general case. Unsteady RANS (URANS) can also be obtained by applying the statistical average of a conditional or phase-averaging procedure, or both. Note that the RANS
approach does not allow an explicit control of the complexity of the simulation because
the cutoff frequency can not be specified during the averaging procedure. The second
popular filtering approach leads to the large-eddy simulation (LES) technique, which
is based on a filtering operator that results in unsteady 3D computations. The filtering
procedure can be explicitly associated with the application of a convolution filter to the
DNS solution or implicitly imposed by numerical errors, the computational mesh, or
modeling errors, or even by the blending of these two possibilities.
The requirement of controlling the numerical error appears more stringent for LES
than for DNS because the LES cutoff is supposed to occur within scales that are much
more energetic than for DNS, leading to a much higher level of numerical error. This
may become very problematic if the numerical scheme introduces some artificial dissipation (artificial viscosity, upwind scheme, filter, etc.) because the amount of numerical
nonphysical dissipation may happen to overwhelm the physical drain of resolved kinetic
energy associated with the energy cascade. Practical experience reveals that the best
results are obtained using centered, nondissipative schemes, which make it possible to
capture a much broader resolved band of scales than dissipative methods; however, because nondissipative schemes do not provide stabilization, they require much finer grids
379
3:46
P1: JZZ
0521871441con
380
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
CONCLUSIONS
to ensure numerical stability. Centered schemes introduce dispersive errors instead of

dissipative errors.
Although statistical modeling is in general not suitable for aeroacoustic prediction
owing to the removal of the unsteadiness, LES provides the unsteady flow with the use of
a subgrid-scale model. Several subgrid-scale models are available that can be classified
according to the assumption made during their derivation. Functional modeling uses the
dynamics of the isotropic turbulence as a basis. Both theoretical and numerical studies
show that the net effect of the subgrid scales is a drain of kinetic energy from the resolved
scales. A simple way to account for that net drain of energy is to parameterize it as an
additional dissipation. This is done by defining an eddy viscosity in the same spirit as in
RANS modeling. Note that subgrid viscosity models for incompressible flows are relevant models for the deviatoric part of the subgrid tensor only. As a consequence, the subgrid kinetic energy, which is tied to the trace of the subgrid tensor, is not a direct output of
these models and must be modeled separately if it is required for some physical analysis.
The second class of subgrid-scale models use structural modeling. Many ways of
reconstructing or approximating subgrid scales have been proposed that rely on either
approximate deconvolution, scale similarity, or deterministic vortex models. Deconvolution models are not able to reconstruct scales smaller than the grid spacing x and so
are not able to account for nonlocal energy transfer across the cutoff. As a consequence,
they must be supplemented by another model specially designed for this purpose. A
simple way to do this is to use functional models of the eddy-viscosity type, which are
based on the description of the kinetic energy transfers associated with nonlocal interactions. Thus, a full deconvolution model is obtained by operating a linear combination
of a deconvolution-type structural model with a functional model of the eddy-viscosity
type. All the eddy-viscosity models exhibit a constant that was set when the isotropic
turbulence case was considered. An idea to minimize modeling errors is to adjust that
constant at each point and at each time step to obtain the best possible adaptation of the
selected subgrid model to the local state of the resolved field. This can be done using
the dynamic procedure, which relies on the Germano identity.
For more complex applications and industrial problems, coupling of acoustic and
aerodynamic prediction has been attempted to several hybrid RANSLES approaches
that show promise for computational aeroacoustics problems. Various new ones have
been described, ranging from global ones that solve a single set of equations throughout
the entire domain to zonal approaches that explicitly impose pure RANS or pure LES
in individual zones according to some initial domain decomposition.
In comparison with traditional LES, hybrid RANSLES methods are intended to
allow larger mesh spacings and hence larger time steps, and thus they can result in a
smaller portion of the total frequency spectrum being directly resolved. This is not a
weakness of the hybrid models themselves but simply an expected result of exploiting
the coarser spatial and temporal resolutions with which hybrid methods are able to operate. The issue of unresolvable noise sources has been discussed, and two possible approaches to help extend the range of predicted frequencies have been described. Both approaches involve synthetic reconstruction of an unsteady velocity field that reproduces
3:46
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
key properties of the underlying (statistically represented) turbulence. One approach

considered was a nonlinear acoustics solver using a stochastic subgrid model; the other
involves augmenting the frequency spectra of traditional LES or hybrid RANSLES to
separately model the generation and transmission of the noise from unresolvable scales.
Finally, we note that the choice of both numerical method and subgrid-scale closure
is expected to continue to play a crucial role in LES, hybrid RANSLES, and acoustics.
In most existing LES or hybrid RANSLES codes, the subgrid-scale model and numerical flux treatments are independent of one another, typically resulting in a mixing
that is too strong.
Aerodynamic noise occurs because of two basically different phenomena. The first
one is impulsive noise a result of moving surfaces or surfaces in nonuniform flow
conditions. The displacement effect of an immersed body in motion as well as the
nonstationary aerodynamic loads on the body surface generate pressure fluctuations
that are radiated as sound. This kind of noise is deterministic and relatively easy to
extract from aerodynamic simulations because the required resolution in space and time
to predict the acoustics is similar to the demands from the aerodynamic computation.
This noise appears primarily with rotating systems such as helicopter rotors, wind
turbines, turbine engine fans, and ventilators. In particular, if the surfaces move at
speeds comparable to the speed of sound or there is an interaction between a rotor and
a stator wake, these tonal noise components can be dominant.
The other noise mechanism is the result of turbulence and therefore arises more
or less powerfully in nearly every engineering application because turbulence is by
its very nature stochastic and therefore has a broad frequency spectrum. Interestingly
enough, turbulent energy is converted to acoustic energy most efficiently in the vicinity
of sharp edges as, for example, at the trailing edge of an aircraft wing. In this case the
uncorrelated turbulent eddies flowing over the upper and lower sides of the edge have
to relax with each other, generating very locally strong equalizing flows that result in
highly nonstationary pressure spikes. Another major source of turbulence sound is jet
flows, in which the shear layer in the mixing zone again radiates into the far field.
Given the nonlinearity of the governing equations, it is very difficult to predict the
sound production of fluid flows. This sound production occurs typically at high-speed
flows for which nonlinear inertial terms in the equation of motion are much larger
than the viscous terms (high Reynolds numbers). Because sound production represents
only a very minute fraction of the energy in the flow, the direct prediction of sound
generation is very difficult. This difficulty is particularly dramatic in free space and
at low subsonic speeds. Solving the complete, fully coupled, compressible Navier
Stokes equations requires tremendous computational resources because especially on
small-Mach-number flows flow and acoustics represent a multiscale problem with
its inherent difficulties. The problem is that the small acoustic perturbations are not
drowned out by numerical errors of the much larger aerodynamic forces. Space and time
resolution for the aerodynamic data combined with the large regions up to an observer
in the far field give rise to ridiculously high numbers of cells and time steps; however,
even if the computer power were available, the discretization schemes well known
381
3:46
P1: JZZ
0521871441con
382
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
CONCLUSIONS
from numerical fluid dynamics do not work very well in computational aeroacoustic
applications because they usually have higher dispersion. Typically a plane wave is
severely distorted and sometimes dampened after being transported along a distance of
just a few wavelengths. Such a distance is clearly too short for the common case of an
observer in the far field.
A more promising approach for technically relevant aeroacoustic problems is to
apply hybrid methods. With this approach the near-field aerodynamics are computed
to obtain velocity and pressure fluctuations, which form the acoustic source terms for
a separate computation of the far-field acoustics. The reason for splitting off aerodynamics from aeroacoustics is the great disparity of levels and length scales between the
flow and aeroacoustic fields.
In acoustics, one considers small perturbations of a flow. This allows linearizing
the conservation laws and constitutive equations. To derive a wave equation for the
prediction of noise propagation, one not only linearizes the basic equations but friction
and heat transfer are also neglected.
Using Greens theorem we can obtain an integral equation that includes the effects
of the sources, the boundary conditions, and the initial conditions on the acoustic field.
Greens function is defined as the response of the flow to an impulsive point source.
To determine the source from any measured acoustical field outside the source region,
we need a physical model of the source. This is typical of any inverse problem in
which the solution is not unique. When using microphone arrays to determine the
sound sources responsible for aircraft noise, one usually assumes that the sound field
is built up of so-called monopole sound sources; however, sound sources are more
accurately described in terms of dipoles or quadrupoles. Under such circumstances it
is hazardous to extrapolate such a monopole model to angles outside the measuring
range of the microphone array or to the range of flow Mach numbers other than used
in the experiments.
It is often stated that Lighthill has demonstrated that the sound produced by a
free, turbulent, isentropic flow has the character of a quadrupole. The reason is that,
because in many flows there is no net volume injection owing to entropy production nor
any external force field, the sound field can at most be a quadrupole field. Therefore,
Lighthills statement is actually that we should ignore any monopole or dipole emerging
from a poor description of the flow. Sound, for example, produced by turbulence in a
free jet has the character of a quadrupole distribution field. Because the vortices that
produce the sound are convected with the main flow, there will be a significant Doppler
effect at high Mach numbers. This results in a radiation field mainly directed about
the flow direction. Owing to convective effects on the wave propagation, the sound is,
however, deflected in the shear layers of the jet. This explains that, along the jet axis,
there is a so-called cone of silence.
We can use the Greens function formalism to determine the effect of the movement
of a source on the radiated sound field. The problem of a source, observer, and scattering
objects moving together steadily in a uniform stagnant medium is the same as the
problem of a fixed source, observer, and objects in a uniform mean flow.
3:46
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
Lighthill proposed an approach to the case of an arbitrary source region surrounded

by a quiescent fluid. The key idea of Lighthill was to derive from the exact equations of
mass conservation and momentum conservation a nonhomogeneous wave equation that
reduces to the homogeneous wave equation in a region surrounding the listener. When
the entropy term and the external forces are negligible, the flow will only produce sound
at high velocities corresponding to high Reynolds numbers. He therefore assumed that
viscous effects are negligible and reduce the sound source to the nonlinear convective
effects. Hence, we can calculate the source term from a numerical simulation that
ignores any acoustic wave propagation and subsequently predicts the sound production
outside the flow. In extreme cases of low-Mach-number flow, a locally incompressible
flow simulation of the source region can be used to predict the sound field. One can state
that such an integral formulation combined with Lighthills analogy allows us to obtain
a maximum of information concerning the sound production for a given fact flow field.
A spectacular example of this is Lighthills prediction that the power radiated to free
space by a free, turbulent, isothermal jet scales as the eighth power U08 of the jet velocity.
The first step in making Lighthills analogy useful is to identify a listener around
which the flow behaves like linear acoustic perturbations and is described by the homogenous wave equation. This is an assumption valid in many applications. When we
listen, under normal circumstances, to a flute player we have conditions that are quite
reasonably close to these assumptions. At this stage, the most important contribution
of Lighthills analogy is that it generalizes the equations for the fluctuations and p
to the entire space even in a highly nonlinear source region. Then, the next step is to
introduce approximations to estimate the source terms.
The integral formulation of Lighthills analogy can be generalized for flows in the
presence of walls. To do so, Curle used the free-space Greens function and, instead of the
pressure p as aeroacoustical variable, the density . Using his theory, we understand
easily that a rotor blade moving in a nonuniform flow field will generate sound owing
to the unsteady hydrodynamic forces on the blade. At low Mach number this will easily
dominate the Doppler effect because of the rotation. Wind rotors placed downwind of
the supporting mast are cheap because they are hydrodynamically stable. There is no
need for a feedback system to keep them in the wind. However, the interaction of the
wake of the mast with the rotor blades causes dramatic noise problems.
Although the formulation of Curle just discussed assumes a fixed control surface S,
the formulation of Ffowcs Williams and Hawkings allows the use of a moving control
surface S(t). The key idea is to include the effect of the surface in the differential
equation. For the formulation of Ffowcs Williams and Hawkings describing the effect
of moving boundaries, p is used again as the aeroacoustical variable. In acoustics
this would have been immaterial because the two variables are related by the equation
of state. Actually, in aeroacoustics there is a subtle difference that appears when we
compare the source terms of the two wave equations. In that case, we see that when p is
used as the aeroacoustical variable, the effect of entropy fluctuations has the character
of a monopole sound source. On the other hand, when is used, the apparently same
effects produce a quadrupole distribution that is qualitatively different. Of course, there
383
3:46
P1: JZZ
0521871441con
384
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
CONCLUSIONS
is no difference if we consider the exact equations, but if we do not introduce any

approximation the analogy is just a reformulation of basic equations without much
use. Clearly, we have to be careful in selecting the aeroacoustic variable. Other choices
of aeroacoustic variables lead to different analogies; however, such analogies become
quite obscure. They do not, moreover, provide much intuitive insight and can only be
used numerically.
In many cases such analogies tend to avoid the problem induced by the inability of
the Lighthill analogy to distinguish between propagation and production of sound waves
in a strongly nonuniform flow, which induces refraction. This becomes very important
in supersonic flow. In such cases the source is not compact. One of the problems is that
the sound source deduced from Lighthills analogy is spatially rather extended, leading
to slowly converging integrals. For low-Mach-number isothermal flow the aerodynamic
sound production is entirely due to mean-flow velocity fluctuations, which may be described directly in terms of the underlying vortex dynamics. This is more convenient
because vorticity is in general limited to a much smaller region in space than the corresponding velocity field (Lighthills sound sources). This leads to the idea of using an irrotational flow as reference flow. The result is called vortex sound theory, which is not only
numerically efficient but also allows us to translate the very efficient vortex-dynamical
description of elementary flows directly into sound production properties of these flows.
There are techniques that employ an integrated form of the relevant acoustic propagation equation that is, either Kirchhoffs surface integral or the Ffowcs Williams
Hawkings (FWH) equation. In this case the sound pressure at an observer at a specific
point in time is computed by an integration-of-source term along a surface either a
physical one or surrounding the aerodynamic area and possibly additional volume
integrals outside the surface in the case of the FWH equation. Owing to the finite
speed of sound and the deterministic relationship between emission and observer time
of a signal, there has to be some kind of interpolation of the data at least on one side.
In the case of parts of the integration surface or volume moving at transonic speeds,
the integrals become highly singular because of the Doppler effect, which leads to
difficulties regarding the numerical stability of the procedure.
When the flow is confined, the acoustical energy can accumulate in resonant modes.
Because the acoustical particle displacement velocity can reach the same order of
magnitude as the main flow velocity, the feedback from the acoustical field to the
sound sources can be very significant. This leads to the occurrence of self-sustained
oscillations we call whistling. In spite of the backreaction, the ideas of the analogy
appear to remain useful.
7.2 Numerical errors

The main problem of computational aeroacoustics (CAA) is the disparity of energy,
length, and time scales between the aerodynamics and the aeroacoustics especially
at smaller Mach numbers. In an aerodynamic simulation, we of course introduce
3:46
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
7.3 INITIAL AND BOUNDARY CONDITIONS
numerical errors. To be at the same level as the acoustics we are interested in, these
errors have to be five to nine orders of magnitude smaller than the intended physical
values. To obtain acceptable signal-to-noise ratios for sound levels, we have to add at
least one more order of magnitude. In this sense, basically every flow simulation is
very loud, counting just the numerical errors that introduce numerical noise. Common
computational fluid dynamics (CFD) schemes are adapted to stationary simulations
and therefore just suppress acoustic waves; thus, for the aerodynamic community the
problem seems to be solved.
This means that, for CAA, diffusion and dispersion errors have to be reduced to the
lowest possible level. This is of major importance because the amplitudes of acoustic
waves are several orders of magnitude smaller than the average aerodynamic field amplitudes. In addition, their length scales, typically the principal acoustic wavelengths,
are some orders of magnitude larger than the dimensions of the sound-generating perturbations (vortices and turbulent eddies). Moreover, sound generated by turbulence is
broadband noise with often three orders of magnitude difference between the largest
and the smallest acoustic wavelengths. Finally, acoustic waves propagate at the speed
of sound (which is not necessarily comparable to the mean flow velocity) over large distances in all spatial directions, whereas aerodynamic perturbations are only convected
by the mean flow. Furthermore, one is usually interested in the noise level at the far
field, implying that the waves have to be traced accurately over long distances.
Similar arguments apply to the temporal discretization. The acoustic waves have
to be tracked accurately in time. Again, more important than the temporal order of
accuracy of the time integration method are its dispersive and dissipative behavior. For
the solution to be accurate, these errors have to be minimized as much as possible.
7.3 Initial and boundary conditions

At the very heart of CAA and CFD lies the proper handling of boundary conditions.
Because acoustics is a radiation problem, basically all the sound energy will sooner or
later try to leave the computational domain. On solid walls, reflections obviously arise,
which we can handle in a straightforward manner just as in CFD. However, on artificial
far-field boundaries the physics dictates a straight pass-through without any spurious
reflections. Although this requirement seems to be obvious, it is indeed very hard
to fulfill sufficiently. Several concepts have been developed for this specific problem
optimized for one application or another, but the downside is that a generally accepted
solution does not yet exist.
As far as accuracy is concerned, the perfectly matched layer (PML) technique
appears to be the most accurate among all the methods reviewed above. However, the
PML equations are not yet available for every type of nonreflecting boundaries likely
to be encountered in practical situations. So far, they have only been developed for a
parallel mean flow in a direction aligned with one of the coordinates. The technique
is still very much under active development. In this regard, the alternatives are the
385
3:46
P1: JZZ
0521871441con
386
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
CONCLUSIONS
absorbing-zone techniques that can be applied to a wider class of problems and are often
coupled with the characteristic boundary conditions. For problems with a centralized
and compact noise source, the radiation condition offers an effective alternative.
With a few exceptions such as predictions of the transition process or fundamental
turbulence investigations, in most cases the initial conditions play a subsidiary role
because the statistically steady-state flow status, which should be reached independently
of the initial conditions, is of major concern. An example of an exception is the decay of a
homogeneous isotropic flow often used for basic investigations in turbulence research.
For this purpose a cubical integration domain with periodic boundary conditions in
all directions is chosen requiring the initialization of the flow field. On the basis of
the scalar energy spectrum a homogeneous isotropic and divergence-free vector field
is generated in spectral space and then transformed back to the physical space by
fast Fourier transformation. This procedure leads to reasonable initial conditions with
prescribed values for the turbulent kinetic energy, the dissipation rate, and the smallest
and largest wave numbers of the flow field.
7.4 Examples
From the presented examples we conclude that direct noise computation is an outstanding approach for the sound generated by free-shear flows because it permits turbulence
events to be correlated with the sound far field. In this context, LES seems the better tool
to clarify Reynolds-number effects especially for subsonic jet noise. The dependence
on Reynolds number is particularly expected for the noise-generation mechanisms observed in mixing layers whose initial state may be either transitional or turbulent.
The suitability of LES to the prediction of jet noise has been well established for
cold, high-subsonic jets. The introduction of high-bandwidth, high-accuracy numerical
techniques has allowed for the computation of both the unsteady, near-field turbulence found in a jet exhaust plume and its radiated sound. Numerous research groups
have reported agreement between their LES results and experimental data for both the
near-field fluctuations and the far-field sound characteristics for high-subsonic jets at
moderate Reynolds numbers. The most recent investigations have exhibited the correct
acoustic spectral characteristics over a range of frequencies. Ongoing investigations
are exploring the role the subgrid-scale (SGS) model and the grid resolution has on
the radiated sound. Future work is focused on extending the LES results to higher
frequencies through the development of SGS noise models.
For high-Reynolds-number flows, the dynamics of the small scales and the intermittence of the turbulence may lead to the coexistence of different modes, the flow
switching between them. This phenomenon can hardly be described by RANS methods and is better suited to LES. The first complete aeroacoustic computations are very
recent. The large-eddy simulations described here reveal the dependence of the results
on the resolution and on the SGS model. The radiated field induced by the cavity
3:46
P1: JZZ
0521871441con
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
7.4 EXAMPLES
flow is greatly sensitive to small flow modifications such as the concentration of the
coherent structures. The computation of the correct acoustic levels in the far field is
thus a challenging problem. It requires subsequent investigations of the role of the SGS
models and on the effects of the numerical resolution. Significant progress has relied
on direct comparisons with experimental databases. Such databases with aerodynamic
and acoustic information have been lacking and are just beginning to emerge.
As an example, it was also demonstrated that the Radioss CFDCAA code is capable
of predicting aeroelastic and vibroacoustic phenomena due to turbulence in the wake
of a ruler. The presented approach exhibits accurate predictions below about 700 Hz.
For higher frequencies, the levels are underestimated. A refined mesh in the streamwise
flow provides somewhat better results, but the poor spanwise resolution of the model is
believed to limit the development of structures smaller than 0.05 m (corresponding to
frequencies higher than 500 Hz). Another research track is to evaluate the dissipation of
wall functions. The fully coupled numerical approach presented here provides results
opposite to those obtained by a hybrid method combining modal analysis for the plate
and the cavity with a semiempirical model of the wall-pressure loading, which overestimated the acceleration and acoustic radiation of the plate above 700 Hz. The main advantage of a compressible arbitrary LagrangianEulerianCFD code like Radioss CFD
CAA is its ability to solve the following in a single simulation: flow field; noise; source
generation; propagation, including model behaviors; and radiation and transmission to
structures.
The whole three-step CAA process was finally applied to the NACA-0012 case,
including a far-field noise prediction performed by use of a Kirchhoff integration based
on the pressure field and its normal derivative on the external boundary of the Euler
domain. It was shown that the wave fronts are continuous from one domain to another.
An interesting point is that acoustic waves emitted by the airfoils leading edge are
hardly visible. This confirmed that the leading edge acts as a geometrical singularity
on which incident acoustic waves coming from the trailing edge are scattered.
Because the computational requirements of an LES in conjunction with the linearized Euler equations of the far-field noise are too high, most calculations on cylinders and cars have been done so far with the acoustic analogy (mainly with FWH).
Experience in computational aeroacoustics shows that, in the far field, the accuracy of
acoustic signal is determined by the accuracy of the flow-field prediction rather than
by acoustic analogy. Although it is well known that URANS equations cannot explore
unsteady flow accurately, they are still very popular in computational aeroacoustics
because of the relatively low CPU time and memory requirements. URANS equations
resolve only a narrow band of present frequencies in the flow and average the rest of the
frequency domain. Direct numerical simulation has to be carried out to resolve all of
the fluctuations. However, because DNS requires very fine grid resolution and therefore
very high CPU time and memory, today it is only affordable for very low Reynolds
numbers and simple geometries. A good compromise is LES, which resolves a fairly
large range of frequencies and models only a small part of the flow. It is commonly
387
3:46
P1: JZZ
0521871441con
388
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
CONCLUSIONS
assumed that the unresolved part of the flow consists of very small eddies that exhibit
a homogeneous and isotropic structure. As long as this assumption is valid, even an
algebraic model, like the SGS model from Smagorinsky, provides reasonable results. To
keep this assumption valid, one must resolve all important scales, which is an affordable
task for small Reynolds numbers and simple geometries.
3:46
P1: JZZ
nomenclature
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
14:39
APPENDIX A
Nomenclature
A.1 Symbols
A1 , A2
B1 , . . . , B7
c
C
Cf
E
G
J 1
k
kc = /
L
M
Msgs
Mt
p
qT
R
Re
T
T
u
Uk
u
subgrid stress tensors in the compressible momentum

equation
subgrid stress tensors in the compressible energy equation
speed of sound
cross-stress tensor
skin friction coefficient
total energy
convolution filter kernel
Jacobian of the curvilinear grid transform
wave number
LES cutoff wave number
Leonard tensor
Mach number
subgrid-scale Mach number
turbulent Mach number
pressure
heat flux
subgrid Reynolds stress tensor
Reynolds number
temperature
Kolmogorov time scale
velocity vector
kth component of the contravariant flux
friction velocity
389
P1: JZZ
nomenclature
CUFX063/Wagner
390
0 521 87144 1
printer: sheridan
NOMENCLATURE
A.2 Greek symbols

LES cutoff length
test filter cutoff length

x, y, z mesh size in direction x, y, z
t
time step of the computation
kinetic energy dissipation rate

numerical dissipation rate
num
subgrid dissipation rate
sgs
Kolmogorov length scale
molecular diffusivity
dynamic viscosity
kinematic viscosity
frequency
density
subgrid stress tensor

viscous stress tensor
v
LES cutoff time
A.3 Mathematical operators

ui

+

E
v
qT
ith component of the velocity vector

filtered part of
subgrid part of
mass-weighted filtered part of
Fourier transform of
expressed in boundary layer wall units
synthetic filtered total energy
synthetic filtered viscous stress tensor
synthetic filtered heat flux
tensorial product
convolution product
October 23, 2006
14:39
P1: JZZ
abbreviations
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
14:40
APPENDIX B
Abbreviations
AIAA
ALE
ALESSIA
APE
ASD
ASME
ASMO
AUSM
b.c.
BC
BEM
BL
BMBF
CAA
CEAS
CFD
CFL
CNRS
CPU
CSD
CUHD
CULD
DES
DESTINY
DFG
DG
DGLR
American Institute of Aeronautics and Astronautics

arbitrary LagrangianEulerian formulation
application of large-eddy simulation to the solution
of industrial problems
acoustic perturbation equations
artificial selective damping
American Society of Mechanical Engineers
Association for Structural and Multidisciplinary
Optimization
advective upstream-splitting method
boundary conditions
boundary conditions
boundary element method
logarithmic buffer layer
Bundesministerium fur Bildung und Forschung
computational aeroacoustics
Confederation of European Aerospace Societies
computational fluid dynamics
CourantFriedrichsLewy
Centre National de la Recherche Scientifique
central processing unit
cross-spectral densities
compact upwind with high dissipation
compact upwind with low dissipation
detached-eddy simulation
Detached Eddy Simulation for the Transportation Industry
Deutsche Forschungsgemeinschaft
discontinuous Galerkin
Deutsche Gesellschaft fur Luft- und Raumfahrt
- Lilienthal - Oberth e.V.
391
P1: JZZ
abbreviations
CUFX063/Wagner
392
0 521 87144 1
printer: sheridan
October 23, 2006
ABBREVIATIONS
DLR
DNC
DNS
DRP
DSM
ECL
ENO
ENSAM
ERCOFTAC
EU
FD
FDS
FFT
FSM
FV
FWH
IBM
IL
KFVS
KIM
LCRU
LDDRK
LDDRK4
LDDRK5
LDDRK5-6
LDDRK6
LEE
LES
LHS
LMFA
LNS
LSTM
LURC
LURU
LU-SGS
MGB
MGBK
Deutsches Zentrum fur Luft- und Raumfahrt

direct noise computation
direct numerical simulation
dispersion-relation-preserving
dynamic Smagorinsky model
essentially nonoscillatory
Ecole Nationale Superieure dArts et Metiers
European Research Community on Flow, Turbulence
and Combustion
European Union
finite difference
flux-difference splitting
fast Fourier transformation
flow simulation methodology
finite volume
Ffowcs WilliamsHawkings
International Business Machines
intensity level
kinetic flux vector splitting
Kirchhoff integration method
L-operator central and R-operator upwind-biased
low-dissipation and low-dispersion RungeKutta
scheme
four-stage optimized low-dissipation and low-dispersion
RungeKutta scheme
five-stage optimized low-dissipation and low-dispersion
RungeKutta scheme
alternating two-step low-dissipation and
low-dispersion RungeKutta scheme
six-stage optimized low-dissipation and low-dispersion
RungeKutta scheme
linearized Euler equations
large-eddy simulation
left-hand side
Laboratoire de Mecanique des Fluides et Acoustique
limited numerical scales
Lehrstuhl fur Stromungsmechanik
L-operator upwind-biased and R-operator central
L-operator and R-operator upwind-biased
lower-upper symmetric GaussSeidel
ManiGliebeBalsa
ManiGliebeBalsaKhavaran
14:40
P1: JZZ
abbreviations
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 23, 2006
393
ABBREVIATIONS
MILES
MPI
MUSCL
NACA
NASA
NLAS
NLD
NLDE
NRBC
OASPL
OL
ONERA
PC
PIV
PML
POD
PPW
PREDIT
PSA
PSD
PWL
R&D
RANS
RHS
RK4
RKDG
rms
RNP
SA
sAbrinA
SATIN
sgs; SGS
SM
SNGR
SOCF
SPL
SUPG
svp
SWING
14:40
monotonically integrated large-eddy simulation

message-passing interface
monotone upstream-centered schemes for conservation
laws
National Advisory Committee for Aeronautics
National Aeronautics and Space Administration
nonlinear acoustics solver
nonlinear disturbance
nonlinear disturbance equations
nonreflecting boundary conditions
overall sound-pressure level
logarithmic outer layer
Office National dEtudes et de Recherches Aerospatiales
personal computer
particle image velocimetry
perfectly matched layer
proper orthogonal decomposition
points per wavelength
Programme de Recherche et dInnovation dans les
Transports Terrestres
Peugeot Societe Anonyme
power spectral density
sound power level
research and development
Reynolds-averaged NavierStokes
right-hand side
fourth-order RungeKutta
RungeKutta discontinuous Galerkin
root mean square
Reynolds-number-preserving
SpalartAllmaras (model)
solver for aeroacoustics broadband interactions from
aerodynamics
statistical approach to turbulence-induced noise
subgrid scale
Smagorinsky model
stochastic noise generation and radiation
second-order commuting filters
sound-pressure level
streamline-upwind PetrovGalerkin
(integration scheme)
spurious vortical perturbations
simulation of wing-flow noise generation
P1: JZZ
abbreviations
CUFX063/Wagner
394
0 521 87144 1
printer: sheridan
ABBREVIATIONS
TBL
TBLE
TE
TEM
TurboNoiseCFD
turbulent boundary layer

thin boundary layer equations
trailing edge
transverse electromagnetic
turbomachinery noise-source CFD models for
low-noise aircraft designs
TVD
total variation diminishing
URANS, (U)RANS unsteady Reynolds-averaged NavierStokes
VLES
very large eddy simulation
VS
viscous sublayer
WALE
wall-adapting local eddy viscosity
WENO
weighted essentially nonoscillatory
WKB
WentzelKramersBrillouin
October 23, 2006
14:40
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
12:29
References
Abarbanel, S., Gottlieb, D. and Hesthaven, J.S. (1999): Well-posed perfectly matched layers for
advective acoustics, Journal of Computational Physics, 154:266283.
Abel, M., Stojkovic, D. and Breuer, M. (2003): Development of wall models for LES based on
nonlinear stochastic estimation, Fifth Workshop on DNS and LES: DLES-5, Munich University of
Technology, Germany, Aug. 2729.
Abel, M., Stojkovic, D. and Breuer, M. (2004): Nonlinear stochastic estimation of wall models for
LES, International Journal of Heat and Fluid Flow, 27(2):267278.
Adams, N. (2002): A subgrid-scale deconvolution approach for shock capturing, Journal of Computational Physics, 178:391426.
Adams, N.A. and Shariff, K. (1996): A high-resolution hybrid compact-ENO scheme for shockturbulence interaction problems, Journal of Computational Physics, 127:2751.
Adams, N.A. and Stolz, S. (2001): Deconvolution methods for subgrid-scale approximation in LES,
Modern Simulation Strategies for Turbulent Flow, B.J. Geurts (ed.), Edwards, Flourtown, PA,
pp. 2144.
Agarwal, A. and Morris, P.J. (2000): Direct Simulation of Acoustic Scattering by a Rotorcraft Fuselage,
AIAA Paper 2000-2030.
Agarwal, A. and Morris, P.J. (2004): Broadband Noise from the Unsteady Flow in a Slat Cove, AIAA
Paper 2004-854.
Ahmed, S.R., Ramm, G. and Faltin, G. (1984): Some Salient Features of the Time-Averaged Ground
Vehicle Wake, SAE Technical Paper Series 840 300.
oruk, Y. and Long, L.N. (2000): Computational Simulations of Fore and Aft Radiation
Ahuja, V., Ozy
from Ducted Fans, AIAA Paper 2000-1943.
Algermissen, G., Siegert, R. and Spindler, T. (2001): Numerical Simulation of Aeroacoustic Sound
Generated by Fans under Installation Conditions, AIAA Paper 2001-2174.
Allen, R. and Mendonca, F. (2004): DES Validations of Cavity Acoustics over the Subsonic to Supersonic Range, AIAA Paper 2004-2862.
Allen, R., Mendonca, F. and Kirkham, D. (2005): RANS and DES turbulence model predictions of
noise on the M219 cavity at M = 0.85, International Journal of Aeroacoustics, 4(1,2):135152.
Al-Qadi, I.M.A. and Scott, J.N. (2002): High Reynolds Number Simulation of Under-expanded Supersonic Rectangular Jet, AIAA Paper 2002-2474.
American Academy of Pediatrics (1997): Policy Statement: Noise: A hazard for the fetus and newborn
(RE9728), Pediatrics, 100(4):724727.
Andersson, N. (2003): A Study of Mach 0.75 Jets and Their Radiated Sound Using Large-Eddy
Simulation. Ph.D. thesis, Chalmers University of Technology, Goteborg, Sweden.
Ando, Y. (1969): On the sound radiation from semi-infinite circular pipe of certain wall thickness,
Acustica, 22:219225.
395
P1: JZZ
references
CUFX063/Wagner
396
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Anthoine, J., Buchlin, J.-M. and Hirschberg, A. (2002): Effect of nozzle on resonance in large srm:
theoretical modeling, Journal of Propulsion and Power, 18:304311.
Arakeri, V.H., Krothapalli, A., Siddavaram, V., Alkislar, M. and Lourenco, L. (2003): On the use
of microjets to suppress turbulence in a Mach 0.9 axisymmetric jet. Journal of Fluid Mechanics,
490:7598.
Arunajatesan, S. and Sinha, N. (2001): Unified Unsteady RANS-LES Simulations of Cavity Flowfields, AIAA Paper 2001-0516.
Ashcroft, G. and Zhang, X. (2001): A Computational Investigation of the Noise Radiated by FlowInduced Cavity Oscillations, AIAA Paper 2001-0512.
Atkins, H. and Casper, J. (1994): Nonreflective boundary conditions for high-order methods, AIAA
Journal, 32:512518.
Atkins, H.L. and Lockard, D.P. (1999): A High-Order Method Using Unstructured Grids for the
Aeroacoustic Analysis of Realistic Aircraft Configurations, AIAA Paper 99-1945.
Atkins, H.L. and Shu, C.-W. (1996): Quadrature-Free Implementation of Discontinuous Galerkin
Method for Hyperbolic Equations, ICASE Report No. 96-51.
Atkins, H.L. and Shu, C.-W. (1998): Quadrature-free implementation of the discontinuous Galerkin
method for hyperbolic equations, AIAA Journal, 36:775782.
Avital, E.J., Sandham, N.D. and Luo, K.H. (1996): Sound Generation Using Data from Direct Numerical Simulations of Mixing Layers, AIAA Paper 96-1778.
Baggag, A., Atkins, H., Ozturan,

C. and Keyes, D. (1999): Parallelization of an Object-Oriented
Unstructured Aeroacoustics Solver, ICASE Report 99-11.
Bagwell, T.G., Adrian, R.J., Moser, R.D. and Kim, J. (1993): Improved approximation of wall shear
stress boundary conditions for large-eddy simulation, Near-Wall Turbulent Flows, R.M.C. So, C.G.
Speziale and B.E. Launder (eds.), Elsevier Science Publ., New York, pp. 265275.
Bailly, C., Bogey, C. and Juve, D. (2000): Computation of Flow Noise Using Source Terms in
Linearized Eulers Equations, AIAA Paper 2000-2047.
Bailly, C., Candel, S. and Lafon, P. (1996): Computation of jet mixing noise from statistical source
models and a compressible turbulence closure, Journal of Sound and Vibration, 194(2):219242.
Bailly, C. and Juve, D. (1999): A Stochastic Approach to Compute Subsonic Noise Using Linearized
Eulers Equations, AIAA Paper 99-1872.
Bailly, C., Lafon, P. and Candel, S. (1995): A Stochastic Approach to Compute Noise Generation and
Radiation of Free Turbulent Flows, AIAA Paper 95-092.
Bailly, C., Lafon, P. and Candel, S. (1996): Computation of Noise Generation and Propagation for
Free and Confined Turbulent Flows, AIAA Paper 96-1732.
Balaras, E. and Benocci, C. (1994): Subgrid scale models in finite-difference simulations of complex
wall bounded flows, AGARD CP 551, Dec. 1994, pp. 2.12.6, 74th Fluid Dynamics Symp., Chania,
Crete, Greece, April 2004.
Balaras, E., Benocci, C. and Piomelli, U. (1996): Two-layer approximate boundary conditions for
large-eddy simulations, AIAA Journal, 34(6):11111119.
Balsa, T.F. and Gliebe, P.R. (1997): Aerodynamics and noise of coaxial jets, AIAA Journal,
15(11):15501558.
Balsa, T.F., Gliebe, P.R., Kantola, R.A., Mani, R., Strings, E.J. and Wong II, J.C.F. (1978): High
Velocity Jet Noise Source Location and Reduction, FAA Technical Report No. FAA-RD-76-79.
Bardina, J., Ferziger, J.H. and Reynolds, W.C. (1983): Improved turbulence models based on large
eddy simulation of homogeneous, incompressible, turbulent flows, Report TF-19, Thermosciences
Division, Department of Mechanical Engineering, Stanford University.
Barone, F., Durello, P., Read, A., Mendonca, F., Tournour, M. and El Hachemi, Z. (2003): Investigation
of the tonal noise radiated by subsonic fans using the aero-acoustic analogy, Fan Noise 2003, 2nd
International Symposium, Senlis, France, Sept. 2325, 2003.
Batchelor, G. (1976): An Introduction to Fluid Dynamics, Cambridge University Press, Cambridge,
UK.
Batten, P., Goldberg, U. and Chakravarthy, S. (2000): Sub-grid Turbulence Modeling for Unsteady
Flow with Acoustic Resonance, AIAA Paper 00-0473.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Batten, P., Goldberg, U. and Chakravarthy, S. (2002a): Acoustics Predictions Using Less than LES,
Proceedings of the LES for Acoustics Workshop, Gottingen.
Batten, P., Goldberg, U. and Chakravarthy, S. (2002b): LNS An Approach Towards Embedded LES,
Batten, P., Goldberg, U. and Chakravarthy, S. (2002c): Reconstructed Sub-grid Methods for Acoustics
Predictions at all Reynolds Numbers, AIAA Paper 2002-2516.
Batten, P., Goldberg, U. and Chakravarthy, S. (2004): Interfacing statistical turbulence closures with
large eddy simulation, AIAA Journal, 42(3):485492.
Batten, P., Ribaldone, E., Casella, M. and Chakravarthy, S. (2004): Towards a Generalized Non-linear
Acoustics Solver, AIAA Paper 2004-3001.
Bayliss, A. (1985): A Fourth-Order Scheme for the Unsteady Compressible NavierStokes Equations,
ICASE Report 85-44.
Bayliss, A. and Turkel, E. (1982a): Far-field boundary condition for compressible flows, Journal of
Computational Physics, 48:182199.
Bayliss, A. and Turkel, E. (1982b): Outflow boundary conditions for fluid dynamics, SIAM Journal
on Scientific and Statistical Computing, 3:250253.
Baysal, O. and Stallings, R.L. (1987): Computational and experimental investigation of cavity flowfields, AIAA Journal, 26(1):67.
Baysal, O., Ten, G.-W. and Fouladi, K. (1994): NavierStokes computations of cavity aeroacoustics
with suppression devices, ASME Journal of Vibration and Acoustics, 116:105112.
Beam, R.M. and Warming, R.F. (1976): An implicit finite-difference algorithm for hyperbolic systems
in conservation-law form, Journal of Computational Physics, 22:87110.
Beaudan, P. and Moin, P. (1994): Numerical Experiments on the Flow Past a Circular Cylinder at
Sub-critical Reynolds Number. Technical Report TF-62, Stanford University.
Bechara, W., Bailly, C., Lafon, P. and Candel, S. (1994): Stochastic approach to noise modeling for
free turbulent flows, AIAA Journal, 32(3):455463.
Bechert, D.W. (1980): Sound absorption caused by vorticity shedding, demonstrated with a jet flow.
Journal of Sound and Vibration, 70:389405.
Bechert, D.W., Michel, U. and Pfizenmaier, E. (1978): Experiments on the transmission on sound
through jets. AIAA Journal, 16:873874; also AIAA Paper 77-1278.
Bechert, D.W. and Pfizenmaier, E. (1975): On the amplification of broadband jet noise by a pure tone
excitation. Journal of Sound Vibration, 43:581587.
Bendat, J.S. and Piersol, A.G. (1980): Engineering Applications of Correlation and Spectral Analysis,
Wiley-Interscience Publication.
Benhamadouche, S., Jarrin, N. and Laurence, D. (2003): Synthetic turbulent inflow conditions for
large eddy simulation, Journal of Turbulence Heat and Mass Transfer, 4:467474.
Benyahia, A. (2003): Contribution a` la modelisation numerique du comportement de la couche limite
instationnaire: validation par comparaison aux donnees experimentales sur un profil oscillant. Ph.D.
thesis, Universite de la Mediterrane, France.
Berenger, J.-P. (1994): A perfectly matched layer for the absorption of electromagnetic waves, Journal
of Computational Physics, 114:185200.
Berglund, B., Preis, A. and Rankin, K. (1990): Relationship between loudness and annoyance for ten
community sounds. Environment-International, 16:523531.
Biedron, R.T., Rumsey, C.L., Podboy, G.G. and Dunn, M.H. (2001): Predicting the Rotor-Stator
Interaction Acoustics of a Ducted Fan Engine, AIAA Paper 2001-0664.
Billson, M., Eriksson, L.E. and Davidson, L. (2002): Acoustic Source Terms for the Linear Euler
Equations in Conservative Form, AIAA Paper 2002-2582.
Billson, M., Eriksson, L.E. and Davidson, L. (2003): Jet Noise Prediction Using Stochastic Turbulence
Modeling, AIAA Paper 2003-3282.
Blaisdell, G.A., Mansour, N.N. and Reynolds, W.C. (1993): Compressibility effects on the growth
and structure of homogeneous turbulent shear flow, Journal of Fluid Mechanics, 256:443485.
Blaisdell, G.A., Spyropoulos, E.T. and Qin J.H. (1996): The effect of the formulation of nonlinear
terms on aliasing errors in spectral methods, Applied Numerical Mathematics, 21:207219.
12:29
397
P1: JZZ
references
CUFX063/Wagner
398
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Blake, W.K. (1970a): Mechanics of Flow-Induced Sound and Vibration. Volume I. General Concepts
and Elementary Sources, Academic Press.
Blake, W.K. (1970b): Mechanics of Flow-Induced Sound and Vibration. Volume II. Complex FlowStructure Interaction, Academic Press.
Blake, W.K. (1986): Mechanics of Flow-induced Sound and Vibration, Volume I. Academic Press,
Orlando.
Blake, W.K. and Powell, A. (1986): The development of contemporary views of flow-tone generation, Recent Advances in Aeroacoustics, A. Krothapalli, C.A. Smith, Springer, Berlin, New York,
pp. 247345.
Blevins, R.D. (1990): Flow-Induced Vibration, 2nd ed., Van Nostrand Reinhold, New York.
Bodony, D.J. and Lele, S.K. (2002a): Large-eddy simulation of turbulent jets and progress towards
a subgrid scale noise model, Proceedings of the International Workshop on LES for Acoustics, October 78, German Aerospace Center, DLR, Gottingen, Germany, DGLR-Report 200203.
Bodony, D.J. and Lele, S.K. (2002b): Spatial Scale Decomposition of Shear Layer Turbulence and
the Sound Sources Associated with the Missing Scales in a Large-Eddy Simulation, AIAA Paper
2002-2454.
Bodony, D.J. and Lele, S.K. (2003a): A Stochastic Subgrid Scale Noise Model for Noise Predictions
of Subsonic Jets, AIAA Paper 2003-3252.
Bodony, D.J. and Lele, S.K. (2003b): Large-Eddy Simulation of Jet Engine Exhaust Noise, ERCOFTAC
Bulletin 58, Theme Issue: Aeroacoustics: 2124.
Bodony, D.J. and Lele, S.K. (2004): Jet Noise Prediction of Cold and Hot Subsonic Jets Using LargeEddy Simulation, AIAA Paper 2004-3022.
Boersma, B.J. (2002): Large-eddy simulation of the sound field of a round turbulent jet, Proceedings
of the International Workshop on LES for Acoustics, October 78, German Aerospace Center,
DLR, Gottingen, Germany, DGLR-Report 2002-03.
Boersma, B.J. (2003): Large-Eddy Simulation of the Sound Field of a Round Turbulent Jet,
ERCOFTAC Bulletin 58, Theme Issue: Aeroacoustics: 5154.
Boersma, B.J. and Lele, S.K. (1999): Large Eddy Simulation of a Mach 0.9 Turbulent Jet, AIAA
Paper 1999-1874.
Bogey, C. and Bailly, C. (2002): A Family of Low Dispersive and Low Dissipative Explicit Schemes
for Computing Aerodynamic Noise, AIAA Paper 2002-2509.
Bogey, C. and Bailly, C. (2004): A family of low dispersive and low dissipative explicit schemes for
flow and noise computations, Journal of Computational Physics, 194(1):194214.
Bogey, C. and Bailly, C. (2006): Computation of a high Reynolds number jet and its radiated noise
using LES based on explicit filtering, Computers and Fluids, 35(10):13441358. See also: Direct
computation of the sound radiated by a high-Reynolds number, subsonic round jet, Presented at
the CEAS Workshop From CFD to CAA, November 78, 2002, NTUA, Athens, Greece.
Bogey, C. and Bailly, C. (2005a): Decrease of the effective Reynolds number with eddy-viscosity
subgrid-scale modeling, AIAA Journal, 43(2):437439. See also LES of a High Reynolds,
High Subsonic Jet: Effects of the Subgrid Modellings on Flow and Noise, AIAA Paper 20033557.
Bogey, C. and Bailly, C. (2005b): Effects of inflow conditions and forcing on subsonic jet flows and
noise, AIAA Journal, 43(5):10001007. See also LES of a High Reynolds, High Subsonic Jet:
Effects of the Inflow Conditions on Flow and Noise, AIAA Paper 2003-3170.
Bogey, C., Bailly, C. and Juve, D. (1999): Computation of Mixing Layer Noise Using Large-Eddy
Simulation, AIAA Paper 99-1871.
Bogey, C., Bailly, C. and Juve, D. (2000a): Computation of the Sound Radiated by a 3-D Jet Using
Large-Eddy Simulation, AIAA Paper 2000-2009.
Bogey, C., Bailly, C. and Juve, D. (2000b): Numerical simulation of the sound generated by vortex
pairing in a mixing layer, AIAA Journal, 38(12):22102218.
Bogey, C., Bailly, C. and Juve, D. (2002): Computation of flow noise using source terms in linearized
Eulers equations, AIAA Journal, 40(2):235243.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Bogey, C., Bailly, C. and Juve, D. (2003): Noise investigation of a high subsonic, moderate Reynolds
number jet using a compressible LES, Theoretical Computational and Fluid Dynamics, 16(4):273
297.
Bogey, C., Gloerfelt, X. and Bailly, C. (2003): An illustration of the inclusion of soundflow interactions in Lighthills equation, AIAA Journal, 41(8):16041606.
Bohineust, X., Bardot, A. and Dupuy, F. (1996): Vehicle noise design modifications analysis using
truncated transfer path techniques, ISMA 21, Leuven, Belgium, Sept. 1820, 1996.
Bohineust, X., Floch, C., Sol, A. and Giardi, H. (1999): Hybrid structure acoustic vehicle models,
Proc. 3 Stuttgarter Symoposium, Krastsahrwesen und Verbrennungsmotoren, Feb. 2325, 1999.
Boris, J.P., Grinstein, F.F., Oran, E.S. and Kolbe, R.L. (1992): New insights into large-eddy simulation,
Fluid Dynamics Research, 10:199228.
Bortz, D.M., Rubio, A.D., Banks, H.T., Cain, A.B. and Smith, R.C. (2002): Control of open bay
acoustics by harmonic mass injection, International Journal of Aeroacoustics, 1(1):6581.
Boudet, J., Casalino, D., Jacob, M.C. and Ferrand, P. (2002): Unsteady RANS computations of the
flow past an airfoil in the wake of a rod, ASME Fluid Engineering Division Summer Meeting,
Montreal, Canada, FEDSM2002-31343.
Boudet, J., Casalino, D., Jacob, M.C. and Ferrand, P. (2003): Prediction of Sound Radiated by a Rod
using Large Eddy Simulation. In Proceedings 8th AIAA/CEAS Aeroacoustics Conference, 2003.
Boudet, J., Grosjean, N. and Jacob, M. (2003): Wakeairfoil interaction as broadband noise source,
Computational Aeroacoustics: From Acoustic Sources Modeling to Far-Field Radiated Noise
Prediction, Colloquium EUROMECH 449, December 912, Chamonix, France.
Bradshaw, P. (1977): Compressible turbulent shear layers, Annual Review of Fluid Mechanics, 9:35
54.
Brentner, K.S. and Farassat, F. (1998): An analytical comparison of the acoustic analogy and Kirchhoff
formulation for moving surfaces, AIAA Journal, 36(8):13791386.
Breuer, M. (1997): Numerical and modelling influences on large-eddy simulations for the flow past a
circular cylinder, Proceedings of the 11th Turbulent Shear Flow Conference, Sept. 1997, Grenoble,
France.
Breuer, M. (1998): Large-eddy simulation of the sub-critical flow past a circular cylinder: Numerical
and modeling aspects, International Journal for Numerical Methods in Fluids, 28:12811302.
Breuer, M. (2000): A challenging test case for large-eddy simulation: High Reynolds number circular
cylinder flow, International Journal of Heat and Fluid Flow, 21(5):648654.
Breuer, M. (2002): Direkte numerische Simulation und LargeEddy Simulation turbulenter
Stromungen auf Hochleistungsrechnern, Habilitationsschrift, Universitat ErlangenNurnberg,
Berichte aus der Stromungstechnik, Shaker Verlag, Aachen, Germany.
Breuer, M. and Jovicic, N. (2001): An LES investigation of the separated flow past an airfoil at
high angle of attack, 4th Workshop on DNS & LES, Enschede, The Netherlands, July 1820,
2001, ERCOFTAC Series, 8, pp. 165172, Direct and LargeEddy Simulation IV, B.J. Geurts, R.
Friedrich, and O. Metais (eds.), Kluwer Academic Publishers, Dordrecht.
Bridges, J.E. and Hussain, A.K.M.F. (1987): Roles of initial condition and vortex pairing in jet noise,
Journal of Sound Vibration, 117(2):289311.
Broeckhoven, T., Smirnov, S., Ramboer, J. and Lacor, C. (2003): Finite volume formulation of compact
upwind and central schemes with artificial selective damping, CD-ROM Proc. Second M.I.T. Conference on Computational Fluid and Solid Mechanics., Boston, June 1720, 2003, SIAM Journal
on Scientific and Statistical Computing 21(3):441467.
Brooks, A.N. and Hughes, T.J.R. (1982): Streamline upwind/PetrovGalerkin formulations for convection dominated flows with particular emphasis on the incompressible NavierStokes equations,
Computer Methods in Applied Mechanics and Engineering, 32:199259.
Brooks, T.F. and Hodgson, T.H. (1981): Prediction and comparison of trailing edge noise using
measured surface pressures, Journal of Sound and Vibration, 78(1):69117.
Brouwer, H.H. (1992): On the use of the method of matched asymptotic expansions in propeller
aerodynamics and acoustics, Journal of Fluid Mechanics, 242:117143.
12:29
399
P1: JZZ
references
CUFX063/Wagner
400
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Brown, G.L. and Roshko, A. (1974): On density effects and large structure in turbulent mixing layers,
Journal of Fluid Mechanics, 64(4):775816.
Bruggeman, J.C. (1987): The propagation of low-frequency sound in a two-dimensional duct system
with T joints and right angle bends: T joints theory and experiment, Journal of the Acoustical
Society of America, 82:10451051.
Bruggeman, J.C., Hirschberg, A., van Dongen, M.E.H., Wijnands, A.P.J. and Gorter, J. (1991): Selfsustained aero-acoustic pulsations in gas transport systems: Experimental study of the influence of
closed side branches, Journal of Sound and Vibration, 150:371393.
Bui, T.T. (1999): A Parallel, Finite-Volume Algorithm for Large-Eddy Simulation of Turbulent Flows.
NASA TM-206570.
Buonfiglioli, M. and Mendonca, F. (2005): LES-VOF simulation of primary diesel break-up with
synthetic inlet perturbations, ILASS Americas, 18th Annual Conference on Liquid Atomisation
and Spray Systems, Irvine, CA, May 2005.
Butterworth-Hayes, P. (2004): Europe reaches for aerospace dominance, Aerospace America, August
2004, American Institute of Aeronautics and Astronautics, pp. 3640.
Cabelli, A. (1980): The acoustic characteristics of duct bends, Journal of Sound and Vibration,
68:369388.
Cabot, W.H. and Moin, P. (1999): Approximate wall boundary conditions in the large-eddy simulation
of high Reynolds number flow, Journal of Flow, Turbulence and Combustion, 63(1,4):269291.
Cantwell, B. and Coles, D. (1983): An experiment study of entrainement and transport in the turbulent
near wake of a circular cylinder, Journal of Fluid Mechanics, 136:321374.
Carati, D., Ghosal, S. and Moin, P. (1995): On the representation of backscatter in dynamic localization
models, Physics of Fluids, 7(3):606616.
Carati, D., Winckelmans, G.S. and Jeanmart, H. (1999): Exact expansions for filtered-scales modelling
with a wide class of LES filters, Direct and Large Eddy Simulation III, Voke, Sandham and Kleiser
(eds.), Kluwer Academic Publishers, Dordrecht, pp. 213224.
Cargill, A.M. (1982): Low-frequency sound radiation and generation due to the interaction of unsteady
flow with a jet pipe, Journal of Fluid Mechanics, 121:59105.
Caro, S., Ploumhans, P., Gallez, X., Brotz, F., Schrumpf, M., Read, A. and Mendonca, F. (2005):
Aeroacoustic Simulation of the Noise Radiated by a Helmholz Resonator in a Duct, AIAA Paper
2005-3067.
Casper, J. and Farassat, F. (2003): Trailing Edge Noise Prediction Based on a New Acoustic Formulation, AIAA Paper 2002-2477.
Casper, J. and Meadows, K.R. (1995): Using High-Order Accurate Essentially Non-oscillatory
Schemes for Aeroacoustic Applications, AIAA Paper 95-0163.
Castillo, L. and Walker, J. (2002): Effect of upstream conditions on the outer flow of turbulent bounday
layers, AIAA Journal, 40:12921299.
Chase, D.M. (1987): The character of the turbulent wall pressure spectrum at subconvective wavenumbers and a suggested comprehensive model, Journal of Sound and Vibration, 112:125147.
Choi, D., Barber, T.J., Chiappetta, L.M. and Nishimura, M. (1999): LES Simulation of High Reynolds
Number Jet Flows, AIAA Paper 99-0230.
Chollet, J.P. and Lesieur, M. (1981): Parametrization of small scales of three-dimensional isotropic
turbulence utilizing spectral closures, Journal of Atmospheric Sciences, 38:27472757.
Chung, Y.M. and Sung, H.J. (1997): Comparative study of inflow conditions for spatially evolving
simulation, AIAA Journal, 35(2):269274.
Chyczewski, T.S. and Long, L.N. (1996): A Direct Numerical Study of a Supersonic Rectangular Jet
Using the Full NavierStokes Equations, AIAA Paper 96-1730.
Chyczewski, T., Morris, P. and Long, L. (2000a): Trailing Edge Noise Prediction: Large-Eddy Simulation of Wall Bounded Shear Flow Using the Nonlinear Disturbance Equations, Applied Research
Laboratory, The Pennsylvania State University, State College, PA, Final Report, Contract No.
N-00014-99-0290, Sept. 26, 2000.
Chyczewski, T., Morris, P. and Long, L. (2000b): Large-Eddy Simulation of Wall Bounded Shear
Flow Using the Nonlinear Disturbance Equations, AIAA Paper 2000-2007.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Clark, R.A., Ferziger, J.H. and Reynolds, W.C. (1979): Evaluation of subgrid-scale models using an
accurately simulated turbulent flow, Journal of Fluid Mechanics, 91(1):116.
Cockburn, B. (1999): Discontinuous Galerkin methods for convection dominated problems. In
H. Deconinck and T.J. Barth (eds.), Lecture Notes in Computational Science and Engineering,
Springer-Verlag, 9:69224.
Cockburn, B., Hou, S. and Shu, C.-W. (1990): TVB RungeKutta local projection discontinuous
Galerkin finite element method for conservation laws IV: The multidimensional case, Mathematics
of Computation, 54(190):545581.
Cockburn, B., Karniadakis, G.E. and Shu, C.-W. (2000): Discontinuous Galerkin methods. Theory, computation and applications. In Lecture Notes in Computational Science and Engineering,
Springer-Verlag, 11:350.
Cockburn, B. and Shu, C.-W. (1994): Nonlinearly stable compact schemes for shock calculations,
SIAM Journal on Numerical Analysis, 31:607627.
Collatz, L. (1966): The Numerical Treatment of Differential Equations, Springer-Verlag.
Collatz, L. (1972): Hermitian Methods for Initial-Value Problems in Partial Differential Equations,
Proceedings of the Royal Irish Academy Conference on Numerical Analysis, Topics in Numerical
Analysis, J.J.H. Miller (ed.), pp. 4161.
Collino, F. and Monk, P. (1998): The perfectly matched layer in curvilinear coordinates, SIAM Journal
on Scientific and Statistical Computing, 19(6):2016.
Colonius, T. (2001): An Overview of Simulation, Modeling and Active Control of Flow/Acoustic
Resonance in Open Cavities, AIAA Paper 2001-0076.
Colonius, T., Basu, A.J. and Rowley, C.W. (1999): Numerical Investigation of the Flow Past a Cavity,
AIAA Paper 99-1912.
Colonius, T., Lele, S.K. and Moin, P. (1993): Boundary conditions for direct computation of aerodynamic sound generation, AIAA Journal, 31(9):15741582.
Colonius, T., Lele, S.K. and Moin, P. (1997): Sound generation in a mixing layer, Journal of Fluid
Mechanics, 330:375409.
Comte, P., Silvestrini, J.H. and Begou, P. (1998): Streamwise vortices in large-eddy simulations of
mixing layers, European Journal of Mechanics B Fluids, 17(4):615637.
Constantinescu, G.S. and Lele, S.K. (2001): Large-Eddy Simulation of a Nearly Sonic Turbulent Jet
and Its Radiated Noise, AIAA Paper 2001-0376.
Constantinescu, G.S. and Lele, S.K. (2002): A highly accurate technique for the treatment of flow
equations at the polar axis in cylindrical coordinates using series expansions, Journal of Computational Physics, 183:165186.
Cooper, A.J. and Peake, N. (2001): Propagation of unsteady disturbances in a slowly varying duct
with mean swirling flow, Journal of Fuid Mechanics, 445:207234.
Corcos, G.M. (1964): The structure of the turbulent pressure field in boundary layer flows, Journal
of Fluid Mechanics, 18(3):353378.
Correa, J., Massa, B. and Zajas, S. (2003): Fluent Announces Coupling with LMS SYSNOISE for
Acoustics Modeling, press release PR61, Fluent Inc., August 6, 2003, Lebanon, NH, USA and
Leuven, Belgium, URL: http//www.fluent.com/about/news/pr/pr61.htm.
Cremer, L. and Heckl, M. (1988): Structure-Borne Sound, 2nd ed., translated and revised by E.E.
Ungar, Springer-Verlag, Berlin.
Crighton, D.G. (1985): The Kutta condition in unsteady flow, Annual Review of Fluid Mechanics,
17:411445.
Crighton, D.G. (1991): Airframe noise. In Aeroacoustics of Flight Vehicles, H.H. Hubbard (ed.),
NASA TR 90-3052.
Crighton, D.G. (1992): private communication.
Crighton, D.G. (1993): Computational aeroacoustics for low Mach number flows. Computational
Aeroacoustics, J.C. Hardin and M.Y. Hussaini (eds.), Springer-Verlag, New York, pp. 50
68.
Crighton, D.G., Dowling, A.P., Ffowcs Williams, J.E., Heckl, M. and Leppington, F.G. (1992): Modern
Methods in Analytical Acoustics, Lecture Notes, Springer-Verlag, London.
12:29
401
P1: JZZ
references
CUFX063/Wagner
402
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Crouzet, F., Lafon, P., Buchal, T. and Laurence, D. (2002): Aerodynamic Noise Prediction in Internal
Flows Using LES and Linear Euler Equations, AIAA Paper 2002-2568.
Crow, S.C. (1970): Aerodynamic sound emission as a singular perturbation problem, Studies in Applied
Mathematics, 49(1):2144.
Curle, N. (1955): The influence of solid boundaries upon aerodynamic sound, Procedings of the Royal
Society of London, Series A, 231:505514.
Dakhoul, Y.M. and Bedford, K.W. (1986a): Improved averaging method for turbulent flow simulation. Part 1: Theoretical development and application to Burgers transport equation, International
Journal for Numerical Methods in Fluids, 6:4964.
Dakhoul, Y.M. and Bedford, K.W. (1986b): Improved averaging method for turbulent flow simulation.
Part 2: Calculations and verification. International Journal for Numerical Methods in Fluids, 6:
6582.
Dalmont, J.P., Nederveen, C.J., Dubos, V., Ollivier, S., Meserette, V. and te Sligte, E. (2002): Experimental determination of the equivalent circuit of an open side hole: Linear and non-linear behaviour,
EAA, 46:567575.
Dalmont, J.P., Nederveen, C.J. and Joly, N. (2001): Radiation impedance of tubes with different flanges:
Numerical and experimental investigations. Journal of Sound and Vibration, 244:505534.
Davies, P.O.A.L., Fisher, M.J. and Barratt, M.J. (1963): The characteristics of the turbulence in the
mixing region of a round jet, Journal of Fluid Mechanics, 15:337367.
Deardorff, J.W. (1970): A numerical study of three-dimensional turbulent channel flow at large
Reynolds numbers, Journal of Fluid Mechanics, 41:453465.
Debatin, K. (1999): Lokalisation aeroakustischer Quellen in Scherschichtdominanten Stromungen,
Universitat Karlsruhe.
DeBonis, J.R. and Scott, J.N. (2002): Large-eddy simulation of a turbulent compressible round jet,
AIAA Journal, 40(7):13461354.
Deng, X. and Maekawa, H. (1997): Compact high-order accurate nonlinear schemes, Journal of
Deng, X. and Mao, M. (1997): Weighted Compact High-Order Nonlinear Schemes for the Euler
Equations, AIAA Paper 97-1941.
Deng, X. and Mao, M. (2001): High-Order Dissipative Weighted Compact Nonlinear Schemes for
Euler and NavierStokes Equations, AIAA Paper 2001-2626.
Dequand, S., Hulshoff, S., Van Kuijk, H., Willems, J. and Hirschberg, A. (2003): Helmholtz-like
resonator self-sustained oscillations, Part 2: Detailed flow measurements and numerical simulations,
AIAA Journal, 41(3):416423.
Dequand, S., Luo, X., Willems, J. and Hirschberg, A. (2003): Helmholtz-like resonator self-sustained
oscillations, Part 1: Acoustical measurements and analytical models, AIAA Journal, 41:408415.
Dequand, S., Willems, J.F.H., Leroux, M., Vullings, R., van Weert, M., Thieulot, C. and Hirschberg, A.
(2003): Simplified models of flue instruments: Influence of mouth geometry on the sound source,
Journal of the Acoustical Society of America, 113:17241735.
Devos, J.P. and Lafon, P. (2003): Computation of the Noise Generated by a Shallow Cavity in a Duct
and Test of a Solution for Lowering the Noise, AIAA Paper 03-3105.
Disselhorst, J.H.M. and van Wijngaarden, L. (1980): Flow in the exit of open pipes during acoustic
resonance, Journal of Fluid Mechanics, 99:293319.
Doak, P.E. (1989): Momentum potential theory of energy flux carried by momentum fluctuations,
Doak, P.E. (1995): Fluctuating total enthalpy as a generalized field, Acoustical Physics, 44:677685.
Domaradzki, J.A. and Liu, W. (1995): Approximation of subgrid-scale energy transfer based on the
dynamics of the resolved scales of turbulence, Physics of Fluids, 7(8):20252035.
Domaradzki, J.A., Liu, W. and Brachet, M.E. (1993): An analysis of subgrid-scale interactions in
numerically simulated isotropic turbulence, Physics of Fluids, 5(7):17471759.
Domaradzki, J.A. and Rogallo, R.S. (1990): Local energy transfer and nonlocal interactions in homogeneous, isotropic turbulence, Physics of Fluids A, 2:412426.
Domaradzki, J.A. and Yee, P.P. (2000): The subgrid-scale estimation model for high Reynolds number
turbulence, Physics of Fluids, 12(1):193196.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Donea, J. (1982): An arbitrary LagrangianEulerian finite element method for transient dynamic
fluid-structure interaction, Computer Methods in Applied Mechanical Engineering, 33:689723.
Dong, T.Z. and Mankbadi, R.R. (1996): Large-Eddy Simulations of Sound due to Turbulence Mixing
inside an Ejector, AIAA Paper 96-1731.
Dong, T.Z. and Mankbadi, R.R. (1999): Simulation of unsteady flow in nozzle-ejector mixer, Journal
of Propulsion and Power, 15(4):539543.
Doris, L., Tenaud, C. and Ta Phuoc, L. (2000): LES of spatially developing 3D compressible mixing
layer, C. R. Acad. Sci. Paris, t. 328, Serie IIb:567573.
Dowling, A.P. and Ffowcs Williams, J.E. (1983): Sound and Sources of Sound, Ellis Horwood
Publishers, Chichester, UK.
Dubief, F., Naji, S., Lewis, M. and Mendonca, F. (2005): Prediction of Aeroacoustic Noise in Automotive Centrifugal Fans: Validation of CFD and Acoustic Simulation Techniques, AIAA Paper
2005-2921.
Dubief, Y. and Delcayre, F. (2000): On coherent-vortex identification in turbulence, Journal of
Turbulence, 1:011.
Dubos, V., Keefe, D., Kergomard, J., Dalmont, J.P. and Nederveen, C.J. (1999): Theory of sound
propagation in a duct with a branched tube using modal decomposition, Acustica-Acta Acustica,
85:153169.
Ducros, F., Comte, P. and Lesieur, M. (1996): Large-eddy simulation of transition to turbulence in a
boundary layer developing spatially over a flat plate, Journal of Fluid Mechanics, 326:136.
Ducros, F., Ferrand, V., Nicoud, F., Weber, C., Darracq, D., Gacherieu, C. and Poinsot, T. (1999):
Large-eddy simulation of shock/homogeneous turbulence interaction, Journal of Computational
Physics, 152:517549.
Ducros, F., Laporte, F., Soulères, T., Guinot, V., Moinat, P. and Caruelle, B. (2000): High-order fluxes
for conservative skew-symmetric-like schemes in structured meshes: Application to compressible
flows, Journal of Computational Physics, 161:114139.
Durand, W.F. (1963): Aerodynamic Theory, Dover, New York.
Durrieu, P.P.J.M., Hofmans, G.C.J., Ajello, G., Boot, R.J.J., Auregan, Y., Hirschberg, A. and Peters,
M.C.A.M. (2001): Quasi-steady aero-acoustic response of orifices, Journal of the Acoustical Society
of America, 110:18591872.
Ehrenfried, K., Meyer, C. and Dillmann, A. (2003): Simulation of Sound Propagation in Boundary
Layers Based on Mohrings Acoustic Analogy, AIAA Paper 2003-3272.
Ekaterinaris, J.A. (1999a): Implicit, High-Order Accurate in Space Algorithms for NavierStokes
Equations, AIAA Paper 99-3257.
Ekaterinaris, J.A. (1999b): Implicit, high-resolution, compact schemes for gas dynamics and aeroacoustics, Journal of Computational Physics, 156:272299.
Ekaterinaris, J.A. (1999c): New formulation of HardinPope equations for aeroacoustics, AIAA Journal, 37:10331039.
El-Askary, W.A., Ewert, R. and Schroder, W. (2002): On large eddy simulation as a tool to predict
acoustical fields, Proc. Workshop LES for Acoustics, Gottingen.
El-Askary, W.A., Meinke, M. and Schroder, W. (2001): Towards the numerical analysis of trailing-edge
noise, Proceedings of the German Aerospace Conference, Hamburg, DGLRJT, 2001176.
El-Askary, W.A., Schroder, W. and Meinke, M. (2003): LES of Compressible Wall-Bounded Flows,
Engquist, B. and Majda, A. (1977): Absorbing boundary conditions for the numerical solution of
waves, Mathematics of Computation, 31:629651.
Erlerbacher, G., Hussaini, M.Y., Kreiss, M.O. and Sarkar, S. (1990): The analysis and simulation of
compressible turbulence, Theoretical and Computational Fluid Dynamics, 2:7395.
Erlerbacher, G., Hussaini, M.Y., Speziale, C.G. and Zang, T.A. (1992): Toward the large-eddy simulation of compressible turbulent flows, Journal of Fluid Mechanics, 238:155185.
Estivalezes, J.L. and Gamet, L. (1996): From jet flow computations to far-field noise prediction.
FED-Vol. 238, 1996 Fluids Engineering Division Conference, Volume 3.
Etkin, B., Korbacher, G.K. and Keefe, R.T. (1957): Acoustic radiation from a stationary cylinder in a
fluid stream (aeolian tones), Journal of the Acoustical Society of America, 29:3036.
12:29
403
P1: JZZ
references
CUFX063/Wagner
404
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Evans, G.W. and Maxwell, L. (1997): Chronic noise exposure and reading deficits: The mediating
effects of language acquisition, Environment and Behaviour, 29:638656.
Eversman, W. (1993): Theoretical models for duct acoustic propagation and radiation, Aeroacoustics
of Flight Vehicles, Theory and Practice Volume 2: Noise Control, H.H. Hubbard (ed.), The
Acoustical Society of America.
Ewert, R., Meinke, M. and Schroder, W. (2000): Aeroacoustic Source Terms for the Linearized EulerEquations, AIAA Paper 2000-2046.
Ewert, R., Meinke, M. and Schroder, W. (2001a): Comparison of Source Term Formulations for a
Hybrid CFD/CAA method, AIAA Paper 2001-2200, 7th AIAA/CEAS Aeroacoustic Conference.
Ewert, R., Meinke, M. and Schroder, W. (2001b): Computation of aeroacoustic sound via hybrid
CFD/CAA methods, NATO RTO/AVT Symposium on Development in Computational Aero- and
Hydro-Acoustics, Manchester, UK, Oct. 811.
Ewert, R., Meinke, M. and Schroder, W. (2002): Computation of Trailing Edge Noise via LES and
Acoustic Perturbation Equations, AIAA Paper 2002-2467.
Ewert, R. and Schroder, W. (2003): Acoustic perturbation equations based on flow decomposition via
source filtering, Journal of Computational Physics, 188(2):365398.
Ewert, R. and Schroder, W. (2004): On the simulation of trailing edge noise with a hybrid LES/APE
method, Journal of Sound and Vibration, 270:509524.
Ewert, R., Schroder, W., Meinke, M. and El-Askary, W.A. (2002): LES as a Basis to Determine Sound
Emission, AIAA Paper 2002-0568.
Ewert, R., Zhang, Q., Schroder, W. and Delfs, J. (2003): Computation of Trailing Edge Noise of a 3D
Lifting Airfoil in Turbulent Subsonic Flow, AIAA Paper 2003-3114.
Falkner, V.M. (1943): The resistance of a smooth flat plate with turbulent boundary layer, Aircraft
Engineering, 15:6558.
Fan, T.C., Xiao, X., Edwards, J.R., Hassan, H.A. and Baurle, R.A. (2002): Hybrid LES/RANS Simulation of a Shock Wave/Boundary Layer Interaction, AIAA Paper 2002-0431.
Fan, T.C., Xiao, X., Edwards, J.R., Hassan, H.A. and Baurle, R.A. (2003): Hybrid LES/RANS Simulation of a Mach 3 Shock Wave/Boundary Layer Interaction, AIAA Paper 2003-0080.
Farassat, F. (1981): Linear acoustic formulas for calculation of rotating blade noise, AIAA Journal,
19(9):11221130.
Fedorchenko, A.T. (2000): On some fundamental flaws in present aeroacoustic theory, Journal of
Sound Vibration, 232(4):719782.
Fernholz, H.H. (1971): Ein halbempirisches Gesetz fur die Wandreibung in kompressiblen turbulenten
Grenzschichten bei isothermer und adiabater Wand, ZAMM, 51:T146T147.
Ffowcs Williams, J.E. (1969): Hydrodynamic noise, Annual Review of Fluid Mechanics, 1:197222.
Ffowcs Williams, J.E. (1982): Boundary layer pressures and the Corcos model: A development to
incorporate low wavenumber constraints, Journal of Fluid Mechanics, 125:925.
Ffowcs Williams, J.E. and Hawkings, D.L. (1969): Sound generated by turbulence and surfaces in arbitrary motion, Philosophical Transactions of the Royal Society of London Series A, 264(1151):321
342.
Ffowcs Williams, J.E., Simson, J. and Virchis, V.J. (1975): Crackle: An annoying component of jet
noise, Journal of Fluid Mechanics, 71:251271.
Fischer, P. and Iliescu, T. (2001): A 3D channel flow simulation at Re = 180 using a rational LES
model. DNS/LES Progress and Challenges, Liu , Sakell and Beutner (eds.), Greyden Press, Columbus, OH, pp. 283290.
Fleig, O., Iida, M. and Arakawa, C. (2002): Aeroacoustics simulation around a wind turbine blade using
compressible LES and linearized Euler equations, Proceedings of the International Workshop on
LES for Acoustics, 78 October, German Aerospace Center, DLR, Gottingen, Germany, DGLRReport 2002-03.
Fletcher, N.H. and Rossing, T.D. (1998): The Physics of Musical Instruments, Springer-Verlag, New
York, 2nd ed.
Forestier, N., Jacquin, L. and Geffroy, P. (2003): The mixing layer over a deep cavity at high-subsonic
speed, Journal of Fluid Mechanics, 475:101145.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Fortune, V., Lamballais, E. and Gervais, Y. (2001): Study of Temperature Effects on Radiated Noise
from Mixing Layers Using DNS, AIAA Paper 2001-2257.
Freund, J.B. (1997): Proposed inflow/outflow boundary condition for direct computation of aerodynamic sound, AIAA Journal, 35(4):740743.
Freund, J.B. (1999): Acoustic Sources in a Turbulent Jet: A Direct Numerical Simulation Study, AIAA
Paper 99-1858.
Freund, J.B. (2000): A simple method for computing far-field sound in aeroacoustic computations.
Journal of Computational Physics, 157:796800.
Freund, J.B. (2001): Noise sources in a low-Reynolds-number turbulent jet at Mach 0.9, Journal of
Fluid Mechanics, 438:277305.
Freund, J.B. (2003): Noise-source turbulence statistics and the noise from a Mach 0.9 jet, Physics of
Fluids 15(6):17881799.
Freund, J.B. and Lele, S.K. (2004): Computer simulation and prediction of jet noise, High Speed Jet
Flows, G. Raman, D. Mclaughlin, and P. Morris (eds.), Taylor & Francis, London.
Freund, J.B., Lele, S.K. and Moin, P. (1996): Calculation of the radiated sound field using an open
Kirchhoff surface, AIAA Journal, 34(5):909915.
Freund, J.B., Lele, S.K. and Moin, P. (1998): Direct Simulation of a Mach 1.92 Jet and Its Sound
Field, AIAA Paper 98-2291.
Freund, J.B., Lele, S.K. and Moin, P. (2000): Compressibility effects in a turbulent annular mixing
layer. Part 1. Turbulence and growth rate, Journal of Fluid Mechanics, 421:229267.
Fritz, W., Magagnato, F. and Rieger, H. (1991): Computational Methods for Prediction of Turbulent
Flows around Automobiles, VDI-Kongress, Fellbach.
Fritz, W., Magagnato, F. and Rieger, H. (1992): Computational methods for prediction of turbulent
flows around automobiles, Proceedings of the 3rd International Conference of Innovation and
Reliability in Automotive Design and Testing, Florence.
Frohlich, J., Rodi, W., Kessler, P., Parpais, S., Bertoglio, J.P. and Laurence, D. (1998): Large-eddy
simulations of flow around circular cylinders on structured and unstructured grids, in E.H. Hirschel
(ed.), Notes on Numerical Fluid Mechanics, 66:319338.
Fu, D. and Ma, Y. (1997): A high order accurate difference scheme for complex flow fields, Journal
Fureby, C. and Grinstein, F.F. (1999): Monotonically large-eddy simulation of free shear flows, AIAA
Journal, 37(5):544556.
Fureby, C. and Grinstein, F.F. (2002): Large-eddy simulation of high-Reynolds number free and
wall-bounded flows, Journal of Computational Physics, 181:6897.
Fureby, C. and Tabor, G. (1997): Mathematical and physical constraints on large-eddy simulations,
Theoretical and Computational Fluid Dynamics, 9:85102.
Gaitonde, D. and Shang, J.S. (1997): Optimized compact-difference-based finite-volume schemes for
linear wave phenomena, Journal of Computational Physics, 138:617643.
Gaitonde, D. and Visbal, M. (1999): Further Development of a NavierStokes Solution Procedure
Based on Higher-Order Formulas, AIAA Paper 99-0557.
Gaitonde, D. and Visbal, M. (2000): Pade-Type Higher-Order Boundary Filters for the NavierStokes
Equations, AIAA Journal, 38:2103.
Gamet, L., Ducros, F., Nicoud, F. and Poinsot, T. (1999): Compact finite difference schemes on nonuniform meshes. Application to direct numerical simulations of compressible flows, International
Journal for Numerical Methods in Fluids, 29:159191.
Gamet, L. and Estivalezes, J.L. (1998): Application of large-eddy simulations and Kirchhoff method
to jet noise prediction, AIAA Journal, 36(12):21702178.
Garnier, E., Mossi, M., Sagaut, P., Deville, M. and Comte, P. (1999): On the use of shock-capturing
schemes for large-eddy simulation, Journal of Computational Physics, 153:273311.
Garnier, E., Sagaut, P. and Deville, M. (2001): A class of explicit ENO filters with application to
unsteady flows, Journal of Computational Physics, 170:184204.
Garnier, E., Sagaut, P. and Deville, M. (2002): Large-eddy simulation of shock/homogeneous turbulence interaction, Computers and Fluids, 31(2):245268.
12:29
405
P1: JZZ
references
CUFX063/Wagner
406
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Gedney, S.D. (1996): An anisotropic perfectly matched layer-absorbing medium for the truncation of
FDTD lattices, IEEE Transactions on Antennas Propagation, 44:1630.
Georgiadis, N.J., Alexander, J.I.D. and Reshotko, E. (2003): Hybrid Reynolds-averaged
NavierStokes/large-eddy simulations of supersonic turbulent mixing, AIAA Journal, 41(2):
218229.
Germano, M. (1986): A proposal for a redefinition of the turbulent stresses in the filtered Navier
Stokes equations, Physics of Fluids, 29(7):23232324.
Germano, M. (1987): On the Non-Reynolds Averages in Turbulence, AIAA Paper 87-1297.
Germano, M. (1992): Turbulence: The filtering approach, Journal of Fluid Mechanics, 238:325336.
Germano, M. (1999): From RANS to DNS: Towards a bridging model, Direct and Large-Eddy
Simulation III, Kluwer Academic Publishers, Dordrecht, pp. 225236.
Germano, M., Piomelli, U., Moin, P. and Cabot, W.H. (1991): A dynamic subgrid-scale eddy viscosity
model, Physics of Fluids A, 3(7):17601765.
Geurts, B.J. and Holm, D.D. (2003): Regularization modeling for large-eddy simulation. Physics of
Fluids, 15(1):L13L16.
Gharib, M. and Roshko, A. (1987): The effect of flow oscillations on cavity drag, Journal of Fluid
Ghosal, S. (1996): An analysis of numerical errors in large-eddy simulations of turbulence, Journal
Ghosal, S. (1999): Mathematical and physical constraints on large-eddy simulation of turbulence,
Ghosal, S. and Moin, P. (1995): The basic equations for the large-eddy simulation of turbulent flows
in complex geometry, Journal of Computational Physics, 118:2437.
Giles, M.B. (1990): Non-reflecting boundary conditions for Euler equation calculations, AIAA Journal,
28:20502058.
Glaze, D.J. and Frankel, S.H. (2003): Stochastic inlet conditions for large-eddy simulation of a fully
turbulent jet, AIAA Journal, 41(6):10641073.
Gloerfelt, X., Bailly, C. and Juve, D. (2000): Calcul direct du rayonnement acoustique dun e coulement
affleurant une cavite, C. R. Acad. Sci., t. 328, Serie IIb:625631.
Gloerfelt, X., Bailly, C. and Juve, D. (2001): Direct calculation of cavity noise and validation of
acoustic analogies, Proc. of NATO/RTO Applied Vehicle Technology Panel, Symposium on Aging
Mechanisms and Control, Development in Computational Aero- and Hydro-Acoustics, Manchester,
UK, 811 October.
Gloerfelt, X., Bailly, C. and Juve, D. (2003): Direct computation of the noise radiated by a subsonic
cavity flow and application of integral methods, Journal of Sound and Vibration, 226(1):119146.
Gloerfelt, X., Bogey, C. and Bailly, C. (2002): LES of the noise radiated by a flow over a rectangular
cavity, ERCOFTAC Workshop on LES for Acoustics, 78 October, DLR Gottingen, Germany.
Gloerfelt, X., Bogey, C. and Bailly, C. (2003a): Numerical evidence of mode switching in the flowinduced oscillations by a cavity, International Journal of Aeroacoustics, 2(2):99124.
Gloerfelt, X., Bogey, C. and Bailly, C. (2003b): Numerical Investigation of the Coexistence of Multiple
Tones in Flow-Induced Cavity Noise, AIAA Paper 2003-3234.
Gloerfelt, X., Bogey, C., Bailly, C. and Juve, D. (2002): Aerodynamic Noise Induced by Laminar and
Turbulent Boundary Layers over Rectangular Cavities, AIAA Paper 2002-2476.
Goldstein, M.E. (1976): Aeroacoustics, McGraw-Hill Book Company, New York.
Goldstein, M.E. (1978): Unsteady vortical and entropic disturbances of potential flows round arbitrary
obstacles, Journal of Fluid Mechanics, 891:433468.
Goldstein, M.E. (2003): A generalized acoustic analogy, Journal of Fluid Mechanics, 488:315
333.
Golubev, V.V. and Atassi, H.M. (1998): Acoustic-vorticity waves in swirling flows, Journal of Sound
and Vibration, 209(2):203222.
Goodrich, J.W. (1999): A Comparison of Numerical Methods for Computational Aeroacoustics, AIAA
Paper 99-1943.
Gottlieb, D. and Turkel, E. (1976): Dissipative two-four methods for time dependent problems, Mathematics of Computation, 30:703723.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Graham, W.R. (1997): A comparison of models for the wavenumber-frequency spectrum of turbulent
boundary layer pressures, Journal of Sound and Vibration, 206(4):541565.
Greschner, B., Thiele, F., Casalino, D. and Jacob, M.C. (2004): Influence of Turbulence Modelling on
the Broadband Noise Simulation for Complex Flows, AIAA Paper 2004-2926.
Grogger, H.A., Delfs, J.W., Lauke, T.G., Lummer, M. and Yin, J. (2000): Simulation of leading-edge
noise of airfoils using CAA based on body-fitted grids, 7th International Congress on Sound and
Vibration, Garmisch-Partenkirchen, Germany, 2000.
Groll-Knapp, E. and Stidl, H.-G. (1999): Literaturzusammenstellung zum Thema Auswirkungen von
Fluglarm auf den Menschen, Institut fur Umwelthygiene der Universitat Wien, Vienna, Dezember.
Grotzbach, G. (1981): Numerical simulation of turbulent temperature fluctuations in liquid metals,
International Journal of Heat and Mass Transfer, 24:475490.
Grotzbach, G. (1987): Direct numerical and large-eddy simulation of turbulent channel flows, Encyclopedia of Fluid Mechanics, Cheremisinoff, N.P. (ed.), 6:13371391, Gulf Publ., Houston.
Grotzbach, G. and Worner, M. (1999): Direct numerical and large-eddy simulations in nuclear applications, International Journal of Heat and Mass Transfer, 20:222240.
Gutin, L. (1948): On the Sound Field of a Rotating Propeller, NACA, No. TM 1195 (Original in
Russian: Zhurnal Tekhnicheskoi Fiziki 12:7683, 1936).
Hagstrom, T. and Nazarov, I. (2003): Perfectly Matched Layers and Radiation Boundary Conditions
for Shear Flow Calculations, AIAA Paper 2003-3298.
Hamed, A., Basu, D. and Das, K. (2003): Detached Eddy Simulation of Supersonic Flow over a Cavity,
Hankey, W.L. and Shang, J.S. (1980): Analyses of pressure oscillations in an open cavity, AIAA
Journal, 18(8):892898.
Hardin, J.C. and Pope, D.S. (1994): An acoustic/viscous splitting technique for computational aeroacoustics, Theoretical and Computational Fluid Dynamics, 6:323340.
Hardin, J.C. and Pope, D.S. (1995): Sound generation by flow over a two-dimensional cavity, AIAA
Journal, 33(3):407412.
Harten, A. (1983): High resolution schemes for hyperbolic conservation laws, Journal of Computational Physics, 49:357393.
Harten, A. (1994): Multiresolution Representation and Numerical Algorithms: A Brief Review, ICASE
Report No 94-59.
Harten, A. (1996): Multiresolution representation of data: A general framework, SIAM Journal of
Numerical Analysis, 33(3):12051256.
Hein, S., Hohage, T. and Koch, W. (2004): On resonances in open systems, Journal of Fluid Mechanics,
506:255284.
Heller, H.H., Holmes, D.G. and Covert, E.E. (1971): Flow induced pressure oscillations in shallow
cavities, Journal of Sound and Vibration, 18(4):545555.
Henderson, J., Badcock, K. and Richards, B.E. (2000): Subsonic and Transonic Transitional Cavity
Flows, AIAA Paper 2000-1966.
Henshaw, M.J. de C. (2000): M219 Cavity Case in Verification and Validation Data for Computational
Unsteady Aerodynamics, RTO-TR-26, AC/323(AVT)TP/19.
Heo, D.N., Kim, J.W. and Lee, D.J. (2003): Study on Noise Characteristics of an Open Cavity with
Cross-Correlation Analysis, AIAA Paper 2003-3104.
Heo, D.N. and Lee, D.J. (2001): Numerical Investigation of the Cover-Plates Effects on the Rectangular
Open Cavity, AIAA Paper 2001-2127.
Higdon, R.L. (1987): Numerical absorbing boundary conditions for the wave equation, Mathematical
Computation, 49:6590.
Hileman, J. and Samimy, M. (2001): Turbulence structures and the acoustic far field of a Mach 1.3
jet, AIAA Journal, 39(9):17161727.
Hill, W.G. and Greene, P.R. (1977): Increased turbulent jet mixing rates by self-excited acoustic
oscillations, Transactions of the ASME, Journal of Fluids Engineering, 99:520525.
Hinterberger, C., Garcia-Villalba, M. and Rodi, W. (2003): LES of flow around the Ahmed body:
Lecture notes in applied mechanics, Proceedings of the UEF Conference on The Aerodynamics
of Heavy Vehicles: Trucks, Busses and Trains, Monterey, CA, 2002, Springer-Verlag.
12:29
407
P1: JZZ
references
CUFX063/Wagner
408
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Hinze, J.O. (1975): Turbulence, McGraw-Hill, New York.

Hirsch, C. (1988): Numerical Computation of Internal and External Flows, Volume 1, John Wiley &
Sons, New York.
Hirsch, C. (1990): Numerical Computation of Internal and External Flows, Volume 2: Computational
Methods for Inviscid and Viscous Flows, Wiley Series in Numerical Methods in Engineering, John
Wiley & Sons, New York.
Hirschberg, A. (1995): Aero-acoustics of wind instruments. Mechanics of Musical Instruments, A.
Hirschberg, J. Kergomard, and G. Weinreich (eds.), pp. 291369, Vienna New York. International
Centre for Mechanical Sciences, Springer-Verlag, Udine, Italy.
Hirschberg, A., Bruggeman, J.C., Wijnands, A.P.J. and Smits, N. (1989): The whistler-nozzle and
horn as aero-acoustic sound sources in pipe systems, Acustica, 68:157160.
Hirschberg, A., Gilbert, J., Msallam, R. and Wijnands, A.P.J. (1996): Shock waves in trombones.
Journal of the Acoustical Society of America, 99(3):17541758.
Hirschberg, A. and Schram, C. (1995): A primitive approach to aeroacoustics, Waves in Fluids, Y.
Auregan, A. Maurel, and V. Pagneux (eds.), Springer-Verlag, New York, pp. 291369.
Hixon, R. (1998): A New Class of Compact Schemes, AIAA Paper 99-0367.
Hixon, R. (1999): Prefactored Compact Filters for Computational Aeroacoustics, 37th AIAA
Aerospace Sciences Meeting, AIAA Paper 99-0358.
Hixon, R. (2000): Prefactored small-stencil compact schemes, Journal of Computational Physics,
165:522541.
Ho, C.-M. and Huerre, P. (1984): Perturbed free shear layers, Annual Review of Fluid Mechanics,
16:365424.
Hoffmann, G. and Benocci, C. (1995): Approximate wall boundary conditions for large-eddy simulations, Advances in Turbulence V, R. Benzi (ed.), Kluwer Academic Publishers, pp. 222228.
Hofmans, G.C.J. (1998): Vortex Sound in Confined Flows, Ph.D. thesis, Technical University of
Eindhoven.
Hofmans, G.C.J., Boot, R.J.J., Durrieu, P.P.J., Auregan, Y. and Hirschberg, A. (2000): Aeroacoustic
response of a slit-shaped diaphragm in a pipe at low Helmholtz number, Part 1: Quasi-steady results,
Holger, D.K., Wilson, T.A. and Beavers, G.S. (1977): Fluid mechanics of the edge tone, Journal of
the Acoustical Society of America, 62:11161128.
Howard, R.J.A. and Pourquie, M. (2002): Large-eddy simulation of the Ahmed reference model,
Journal of Turbulence, 42(3):118.
Howe, M.S. (1975): Contributions to the theory of aerodynamic sound, with application to excess jet
noise and the theory of the flute, Journal of Fluid Mechanics, 71:625673.
Howe, M.S. (1979): Attenuation of sound in a low Mach number nozzle flow, Journal of Fluid
Howe, M.S. (1980): The dissipation of sound at an edge, Journal of Sound and Vibration, 70:407411.
Howe, M.S. (1997): Edge, cavity and aperture tones at very low Mach numbers, Journal of Fluid
Howe, M.S. (1998): Acoustics of Fluid-Structure Interactions, Cambridge University Press,
Cambridge, UK.
Howe, M.S. (2000): Analytical Representations of Fluid-Structure Interaction Noise for Large Eddy
Simulations, Report AM-00-003, Boston University, College of Engineering.
Howe, M.S. (2001): Edge-source acoustic Greens function for an airfoil of arbitrary chord, with
application to trailing edge noise. Quarterly Journal of Mechanics and Applied Mathematics,
54(1):139155.
Howe, M.S. (2002): Theory of Vortex Sound, Cambridge University Press, Cambridge, UK.
Hu, F.Q. (1996a): On absorbing boundary conditions of linearized Euler equations by a perfectly
matched layer, Journal of Computational Physics, 129:201219.
Hu, F.Q. (1996b): On Perfectly Matched Layer as an Absorbing Boundary Condition, AIAA Paper
96-1664.
Hu, F.Q. (2001): A stable, perfectly matched layer for linearized Euler equations in unsplit physical
variables, Journal of Computational Physics, 173:455480.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Hu, F.Q. (2002): On Constructing Stable Perfectly Matched Layers as an Absorbing Boundary Condition for Euler Equations, AIAA Paper 2002-0227.
Hu, F.Q. (2004): On Using Perfectly Matched Layer for the Euler Equations with a Non-uniform Mean
Flow, AIAA Paper 2004-2966.
Hu, F.Q. (2005): A perfectly matched layer absorbing boundary condition for linearized Euler equations with a non-uniform mean flow, Journal of Computational Physics, 208:469492.
Hu, F.Q. and Atkins, H.L. (2003): A Discrete Analysis of Non-reflecting Boundary Conditions for
Discontinuous Galerkin Method, AIAA Paper 2003-3301.
Hu, F.Q., Hussaini, M.Y. and Manthey, J.L. (1996): Low-dissipation and low-dispersion RungeKutta
schemes for computational acoustics, Journal of Computational Physics, 124:177191.
Hu, F.Q., Hussaini, M.Y. and Rasetarinera, P. (1999): An analysis of the discontinuous galerkin method
for wave propagation problems, Journal of Computational Physics, 151:921946.
Hubbard, H.H. (1995): Aeroacoustics of Flight Vehicles: Theory and Practice. Volume 1 Noise Sources;
Volume 2 Noise Control (NASA Reference Publication 1258), Acoustical Society of America.
Hubbard, H.H. and Shepherd, K.P. (1991): Aeroacoustics of large wind turbines, Journal of the
Acoustical Society of America, 89:24952508.
Hughes, T.J.R., Feijoo, G.R., Mazzei, L. and Quincy, J.B. (1998): The variational multiscale method
A paradigm for computational mechanics, Computational Methods in Applied Mechanical Engineering, 166:224.
Hughes, T.J.R., Mazzei, L. and Jansen, K.E. (2000): Large-eddy simulation and the variational multiscale method, Computers in Visual Sciences, 3:4759.
Hughes, T.J.R., Mazzei, L., Oberai, A.A. and Wray, A.A. (2001): The multiscale formulation of largeeddy simulation: Decay of homogeneous isotropic turbulence, Physics of Fluids, 13(2):505512.
Hughes, T.J.R., Oberai, A.A. and Mazzei, L. (2001): Large-eddy simulation of turbulent channel flows
by the variational multiscale method, Physics of Fluids, 13(6):17841799.
Hygge, S (1993): A comparison between the impact of noise from aircraft, road traffic and trains on
long-term recall and recognition of a text in childred aged 1214 years. Schriftenreihe des Vereins
f u r Wasser-, Boden- und Lufthygiene 88:416427.
Iaccarino, G. and Durbin, P. (2000): Unsteady 3D RANS Simulations Using the v2-f Model, Center
for Turbulence Research, Annual Research Briefs.
Ihme, M. and Breuer, M. (2002): Sound prediction for a turbulent co-flowing jet based on LES, Proceedings of the International Workshop on LES for Acoustics, 78 October, German Aerospace
Center, DLR, Gottingen, Germany, DGLR-Report 2002-03.
Inagaki, M., Murata, O., Kondoh, T. and Abe, K. (2002): Numerical prediction of fluid-resonant
oscillation at low mach number, AIAA Journal, 40(9):18231908.
Ingard, K.U. (1959): Influence of fluid motion past a plane boundary on sound reflection, absorption,
and transmission, Journal of the Acoustical Society of America, 31(7):10351036.
Ingard, K.U. and Labate, S. (1950): Acoustic circulation effects and the non-linear impedance of
orifices, Journal of the Acoustical Society of America, 22:211218.
Ingard, U. and Ising, H. (1967): Acoustic non-linearity of an orifice, Journal of the Acoustical Society
of America, 42:617.
int Panhuis, P. (2003): Calculations of the Sound Pressure Reflection Coefficient and the Acoustic
End Correction of a Semi-Infinite Circular Pipe Issuing a Subsonic Cold or Hot Jet with Co-Flow,
Internal report, MWL, Aeronautical and Vehicle Engineering, Project EU-ARTEMIS GRD1-200025507, KTH, Stockholm. Supervisors H. Boden, S.W. Rienstra.
Isensee, S. and Frenzel, J. (2003): Leiser Verkehr, Larmforschung im Forschungsprogramm Mobilitat
und Verkehr, BMBF Publik, Bundesministerium fur Bildung und Forschung (BMBF) (ed.), Bonn,
June 2003, URL: http://www.bmbf.de/pub/leiser verkehr.pdf.
Israeli, M. and Orszag, S.A. (1981): Approximation of radiation boundary conditions, Journal of
Jackson, J.D. (1999): Classical Electrodynamics, John Wiley & Sons, New York, 3rd edition.
Jacob, M.C., Boudet, J., Casalino, D. and Michard, M. (2005): A rod-airfoil experiment as benchmark
for broadband noise modelling, Theoretical and Computational Fluid Dynamics, 19(3):171196.
See also Proc. of the 3rd SWING Aeroacoustics Workshop, Stuttgart, 2002.
12:29
409
P1: JZZ
references
CUFX063/Wagner
410
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Jameson, A., Schmidt, W. and Turkel, E. (1981): Numerical Solution of the Euler Equations by Finite
Volume Methods Using RungeKutta Time-Stepping Schemes, AIAA Paper 81-1259.
Jansen, K., Maeder, T. and Reba, R. (2002): Finite-Element Based Large-Eddy Simulation of the
Near-Nozzle Region of a Compressible Round Jet, AIAA Paper 2002-2358.
Jayatunga, C., Kroeff, G., Carotte, J.F., McGuick, J.J. and Petersson, A.T. (2001): Computational
aero-acoustic studies of an exhaust diffuser, Proc. of NATO/RTO Applied Vehicle Technology Panel,
Symposium on Aging Mechanisms and Control, Development in Computational Aero- and HydroAcoustics, Manchester, UK, 811 October.
Jenvey, P.L. (1989): The sound power from turbulence: A theory of the exchange of energy between
the acoustic and non-acoustic fields, Journal of Sound and Vibration, 131:3766.
Johnson, C. and Pitkarata, J. (1986): An analysis of the discontinuous Galerkin method for a scalar
hyperbolic equation, Mathematics of Computation, 46(176):126.
Jones, D.S. (1964): The Theory of Electromagnetism, Pergamon Press, Oxford, UK.
Jones, D.S. (1972): Aerodynamic sound due to a source near a half-plane, Journal of the Institute for
Mathematics and Applications, 9:114122.
Jordan, S.A. (1999): A large-eddy simulation methodology in generalized curvilinear coordinates,
Junger, M.C. and Feit, D. (1986): Sound, Structures and their Interaction, MIT Press, Cambridge,
Massachusetts, 2nd edition.
Kalitzin, G., Kalitzin, N. and Wilde, A. (2000): A Factorization Scheme for RANS Turbulence Models
and SNGR Predictions of Trailing Edge Noise, AIAA Paper 2000-1982, 6th AIAA-CEAS Aeroacoustics Conference, Lahaina (Hawaii), 1214 June, 2000.
Kalitzin, N. and Wilde, A. (2000): Stochastic modeling of turbulence in CAA simulations of airframe noise, Second Aeroacoustic Workshop, H. Korner and J.W. Delfs (eds.), TU Dresden Press,
Braunschweig.
Kanwal, R.P. (1998): Generalized Functions: Theory and Technique, Birkhauser, Boston, 2nd edition.
Kapadia, S., Roy, S. and Wurtzler, J. (2003): Detached Eddy Simulation over a Reference Ahmed
Car Model, AIAA Paper 2003-0857, 41st Aerospace Sciences Meeting and Exhibit, 69 January,
Reno, NV.
Karamcheti, K. (1955): Acoustic Radiation from Two-Dimensional Rectangular Cutouts in Aerodynamic Surfaces, Technical Note 3487, NACA.
Karniadakis, G.E. and Sherwin, S.J. (1999): Spectral/hp Element Methods for CFD, Numerical Mathematics and Scientific Computing, Oxford University Press, New York.
Kato, C., Iida, A., Hattori, M. and Inadama, S. (2000): Numerical Simulation of Aerodynamic Sound
Source in the Wake of a Complex Object, 6th AIAA/CEAS Aeroacoustics Conference, Lahaina,
Hawaii, AIAA Paper 2000-1942.
Kato, C., Iida, A., Takano, Y., Fujita, H. and Ikegawa, M. (1993): Numerical Prediction of Aerodynamic
Noise Radiated from Low Mach Number Turbulent Wake, AIAA Paper 93-0145.
Kato, C., Shishido, S., Miyazawa, M., Yoshiki, H., Ito, H. and Tsubota, H. (2002): Numerical prediction
of aerodynamic noise radiated from a propeller fan, The Fifth JSME-KSME Fluids Engineering
Conference, Nagoya, Japan.
Kato, C., Takano, Y., Iida, A., Fujita, H. and Ikakawa, M. (1994): Numerical prediction of aerodynamic sound by large-eddy simulation, Transaction of the Japan Society of Mechanical Engineers,
Part B, 60(569):126132.
Kato, C., Takano, Y., Iida, A. and Ikegawa, M. (1991): Numerical computation of aerodynamic noise
radiation by the large-eddy simulation (Japanese), 5th Numerical Fluid Dynamics Symposium,
Tokyo, Japan, 195198.
Keating, A., Piomelli, U., Balaras, E. and Kaltenbach, H.-J. (2004): Inflow conditions for large-eddy
simulation, Submitted to Physics of Fluids.
Kennedy, C.A., Carpenter, M.H. and Lewis, R.M. (1999): Low-Storage Explicit RungeKutta
Schemes for the Compressible NavierStokes Equations, ICASE Report No. 99-22.
Kergomard, J., Garcia, A. and Taguy, G. (1989): Analysis of higher order mode effects in an expansion
chamber using modal theory and equivalent electrical circuits, Journal of Sound and Vibration,
129:457475.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Khavaran, A. and Bridges, J. (2004): Modelling of Turbulence Generated Noise in Jets, AIAA Paper
2004-2983.
Khavaran, A., Bridges, J. and Freund, J.B. (2002): A Parametric Study of Fine-Scale Turbulence
Mixing Noise, AIAA Paper 2002-2419.
Khorrami, M.R., Singer, B.A. and Berkman, M.E. (2002): Time-accurate simulations and acoustic
analysis of slat free shear layer, AIAA Journal, 40(7):12841291.
Khorrami, M., Singer, B. and David, P.L. (2002): Time-Accurate Simulations and Acoustic Analysis
of Slat Free Shear Layers: Part II, AIAA Paper 2002-2579.
Kim, I. and Chokani, N. (1990): NavierStokes Simulation of Unsteady Supersonic Cavity Flowfield
with Passive Control, AIAA Paper 90-3101-CP.
Kim, J.W. and Lee, D.J. (1996): Optimized compact finite difference schemes with maximum resolution, AIAA Journal, 34(5):887893.
Kim, J.W. and Lee, D.J. (2000): Adaptive Nonlinear Artificial Dissipation Model for Computational
Aeroacoustics, AIAA Paper 2000-1978.
Kirchhoff, G. (1877): Vorlesungen u ber Mathematische Physik. Bd. 1: Mechanik. 2. Aufl. Teubner,
Leipzig.
Klein, M., Sadiki, A. and Janicka, J. (2003): A digital filter based generation of inflow data for
spatially developing direct numerical or large-eddy simulations, Journal of Computational Physics,
186(2):652665.
Knaepen, B., Debliquy, O. and Carati, D. (2002): Subgrid-scale energy and pseudo pressure in largeeddy simulation, Physics of Fluids, 14(12):42354241.
Kobayashi, H. and Shimomura, Y. (2003): Inapplicability of the dynamic Clark model to the large
eddy simulation of incompressible turbulent channel flows, Physics of Fluids, 15(3):L29L32.
Kobayashi, M.H. (1999): On a class of pade finite volume methods, Journal of Computational Physics,
156:137180.
Kobayashi, T. and Satake, M. (1991): Numerical prediction of aerodynamic sound using large-eddy
simulation (Japanese), 5th Numerical Fluid Dynamics Symposium, Tokyo, Japan, pp. 191194.
Koh, S.R. and Moon, Y.J. (2003): Aeroacoustic Computations of Turbulent Flows over an Open Cavity,
Kolbe, R., Kailasanath, K., Young, T., Boris, J. and Landsberg, A. (1996): Numerical simulations of
flow modification of supersonic rectangular jets, AIAA Journal, 34:902908.
Kolmogorov, A.N. (1941): Dissipation of energy in the locally isotropic turbulence, Comptes Rendus
(Doklady) de lAcademie des Sciences de lURSS, XXXII(1):1618.
Komerath, N.M., Ahuja, K.K. and Chambers, F.W. (1987): Prediction and Measurement of Flows over
Cavities A Survey, AIAA Paper 87-0166.
Kopriva, D.A., Woodruff, S.L. and Hussaini, M.Y. (2000): Discontinuous spectral element approximation of Maxwells equations. Discontinuous Galerkin Methods: Theory, Computation and Applications, B. Cockburn, G.E. Karniadakis, and C.W. Shu (eds.), Lecture Notes in Computational
Science and Engineering, 11:355361. Springer Verlag, New York.
Kosloff, R. and Kosloff, D. (1986): Absorbing boundary conditions for wave propagation problems,
Koutsavdis, E.K., Blaisdell, G.A. and Lyrintzis, A.S. (1999): On the Use of Compact Schemes with
Spatial Siltering in Computational Aeroacoustics, AIAA Paper 99-0360.
Koutsavdis, E.K., Blaisdell, G.A. and Lyrintzis, A.S. (2000): Compact schemes with spatial filtering
in computational aeroacoustics, AIAA Journal, 38:713.
Kraichnan, R.H. (1956): Pressure fluctuation in turbulent flow over a flat plate, Journal of the Acoustical Society of America, 38(3):378380.
Kraichnan, R.H. (1969): Diffusion by a random velocity field, Physics of Fluids, 13(1):2231.
Kraichnan, R.H. (1970): Diffusion by a random velocity field, Journal Computational Physics,
13(1):2231.
Kraichnan, R.H. (1971): Inertial-range transfer in two- and three-dimensional turbulence, Journal of
Fluid Mechanics, 47(3):525535.
Kraichnan, R.H. (1976): Eddy viscosity in two and three dimensions, Journal of Atmospheric Sciences,
33:15211536.
12:29
411
P1: JZZ
references
CUFX063/Wagner
412
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Kravchenko, A.V. and Moin, P. (1997): On the effect of numerical errors in large-eddy simulations of
turbulent flows, Journal Computational Physics, 131:310322.
Kundu, P.K. (1990): Fluid Mechanics, Academic Press, San Diego.
Kurbatskii, K.K. and Tam, C.K.W. (1999): Direct numerical simulation of automobile cavity tones,
3rd Computational Aeroacoustics Workshop, pp. 371383, Ohio Aerospace Institute, Cleveland,
OH, Nov. 810.
Labourasse, E. (2002): Reconstruction des fluctuations turbulentes par une approche hybride
RANS/LES. Ph. D. thesis (11 December 2002), Onera, Universite Pierre et Marie Curie Paris VI.
Labourasse, E. and Sagaut, P. (2002): Reconstruction of turbulent fluctuations using a hybrid
RANS/LES approach, Journal of Computational Physics, 182(1):301336.
Lacor, C., Smirnov, S. and Baelmans, M. (2000): A finite volume formulation for compact schemes on
arbitrary meshes with applications to LES, CFD 2000, N. Satofuka (ed.), pp. 479484. SpringerVerlag 1st Int. Conf. on CFD, Kyoto, 1114 July 2000.
Lacor, C., Smirnov, S. and Baelmans, M. (2004): A finite volume formulation of compact central
schemes on arbitrary, structured grids, Journal of Computational Physics, 198:535566.
Lafon, P. (1997): Noise of confined flows, Von Karman Institute for Fluid Dynamics, Lecture Series
1997-07, Sept. 1518, 1997.
Lafon, P., Caillaud, S., Devos, J.P. and Lambert, C. (2003): Shallow cavity noise aeroacoustical
coupling and its structural effects on a piping steam line, Journal of Fluids and Structures, 18:695
713.
Lafon, P. and Devos, J.P. (1993): Numerical Prediction of Instabilities in Transonic Internal Flows
Using an Euler TVD Code, AIAA Paper 93-0072.
Lamp, A.M. and Chokani, N. (1997): Computation of cavity flows with suppression using jet blowing,
Journal of Aircraft, 34(4):545551.
Landau, L.D. and Lifshitz, E.M. (1987): Fluid Mechanics, Pergamon Press, Oxford, UK, 2nd edition.
Larcheveque, L., Sagaut, P. and Le, T.-H. (2003): Large-Eddy Simulations of Flows in Weapon Bays,
Larcheveque, L., Sagaut, P., Mary, I., Labbe, O. and Comte, P. (2003): Large-eddy simulation of a
compressible flow past a deep cavity, Physics of Fluids, 15(1):193200.
Larsson, J. (2002): Computational Aero Acoustics for Vehicle Applications, Licentiate, Chalmers
University of Technology, Dept. of Thermo and Fluid Dynamics.
Lau, J.C., Morris, P.J. and Fisher, M.J. (1979): Measurements in subsonic and supersonic free jets
using a laser velocimeter, Journal of Fluid Mechanics, 93(1):127.
Leclercq, D.J.J. (1999): Modelisation de la reponse vibro-acoustique dune structure couplee a` une
cavite en presence dun e coulement turbulent, Ph.D. thesis, Departement Genie Mecanique, Universite de Technologie de Compiègne, France.
Leclercq, D.J.J. and Bohineust X. (2002): Investigation and modelling of the wall pressure field
beneath a turbulent boundary layer at low and medium frequencies, Journal of Sound and Vibration,
257(3):477501.
Lee, S., Lele, S.K. and Moin, P. (1992): Simulation of spatially evolving turbulence and the applicability of Taylors hypothesis in compressible flow, Physics of Fluids A, 4(7):15211530.
Lee, S., Lele, S. and Moin, P. (1993): Direct numerical simulation of isotropic turbulence interacting
with a weak shock wave, Journal of Fluid Mechanics, 251:533562. Corrigendum in 264:373
374.
Lee, S., Lele, S. and Moin, P. (1997): Interaction of isotropic turbulence with shock waves: Effect of
shock strength, Journal of Fluid Mechanics, 340:225247.
Leith, C.E. (1990): Stochastic backscatter in a subgrid-scale model: Plane shear mixing layer, Physics
of Fluids A, 2(3):297299.
Lele, S.K. (1992): Compact finite difference schemes with spectral-like resolution, Journal of Computational Physics, 103:1642.
Lemaire, S., Marquez, B. and Jansen, K.E. (1999): A Comparison between Lighthill Analogy and
the Method of Kirchhoff Surfaces for Aeroacoustic Predictions in Finite Elements, AIAA Paper
99-0233, 37th AIAA Aerospace Sciences Meeting.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Lenormand, E., Sagaut, P. and Ta Phuoc, L. (2000): Large-eddy simulations of subsonic and supersonic
channel flow at moderate Reynolds number, International Journal for Numerical Methods in Fluids,
32:369406.
Lenormand, E., Sagaut, P., Ta Phuoc, L. and Comte, P. (2000): Subgrid-scale models for large-eddy
simulations of compressible wall bounded flows, AIAA Journal, 38:13401350.
Leonard, A. (1974): Energy cascade in large-eddy simulations of turbulent fluid flows, Advances in
Geophysics A, 18:237248.
Lesieur, M. (1997): Turbulence in Fluids (3rd edition), Kluwer, Dordrecht.
Leslie, D.C. and Quarini, G.L. (1979): The application of turbulence theory to the formulation of
subgrid modelling procedures, Journal of Fluid Mechanics, 91(1):6591.
Lesser, M.B. and Crighton, D.G. (1975): Physical acoustics and the method of matched asymptotic
expansions, Physical Acoustics, Volume XI, W.P. Mason and R.N. Thurston (eds.), Academic Press,
New York.
Lesser, M.B. and Lewis, J.A. (1972): Applications of matched asymptotic expansions methods to
acoustics I. The WebsterHorn equation and the stepped duct, Journal of the Acoustical Society of
America, 51:16641669.
Levine, H. and Schwinger, J. (1948): On the radiation of sound from an unflanged circular pipe,
Physical Review, 73:383406.
Li, X., Ma, Y. and Fu, D. (2000): DNS of incompressible turbulent channel flow with upwind compact
scheme on non-uniform meshes, CFD Journal, 8(4):536543.
Lighthill, M.J. (1952): On sound generated aerodynamically. Part I: General theory, Proceedings of
the Royal Society of London, A211:564587.
Lighthill, M.J. (1954): On sound generated aerodynamically. Part II: Turbulence as a source of sound,
Proceedings of the Royal Society of London, A222:132.
Lillberg, E. and Fureby, C. (2000): Large-Eddy Simulations of Supersonic Cavity Flows, AIAA Paper
2000-2411.
Lilley, G.M. (1969): Generation of Sound in a Mining Region, Lockheed Aircraft Company, F-3361571-C-1663.
Lilley, G.M. (1974): On the Noise from Jets, Technical Report AGARD CP-131.
Lilley, M.G. (2004): The source of aerodynamic noise, International Journal of Aeroacoustics, 2:241
253.
Lilly, D.K. (1992): A proposed modification of the Germano subgrid-scale closure method, Physics
of Fluids A, 4(3):633635.
Lin, S.Y., Yu, F. and Shih, S.-C. (1997): Numerical Study of MUSCL Schemes for Computational
Liou, M.S. and Steffen, C.J. (1993): A new flux splitting scheme, Journal of Computational Physics,
107:2339.
Liu, C. and Liu, Z. (1993): High order finite difference and multigrid methods for spatially evolving
instability in a planar channel, Journal of Computational Physics, 106:92100.
LMS SYSNOISE. (2002): LMS SYSNOISE Rev 5.6, Computational Vibro-Acoustics, On-Line Users
manual, LMS International NV, Interleuvenlaan 68, Leuven, Belgium.
Lockard, D.P., Brentner, K.S. and Atkins, H.L. (1994): High-Accuracy Algorithms for Computational
Lokhande, B., Sovani, S. and Xu, J. (2003): Computational Aeroacoustics Analysis of a Generic Side
View Mirror, SAE Paper 2003-01-1698.
Longatte, E., Lafon, P. and Candel, S. (1998): Computation of Noise Generation by Turbulence in
Internal Flows, AIAA Paper 98-2332.
Lord, W. and Feng, J. (2000): Physics of jet noise suppression, Proceedings of the Jet Noise Workshop,
NASA/CP2001-211152, pp. 617630, Cleveland, OH, NASA Glenn Research Center.
Lourenco, L. and Shih, C. (1993): Characteristics of the Plane Turbulent Near Wake of a Circular
Cylinder, A Particle Image Velocimetry Study. Adrian, Springer-Verlag.
Lui, C. and Lele, S.K. (2001): Direct Numerical Simulation of Spatially Developing, Compressible,
Turbulent Mixing Layers, AIAA Paper 2001-0291.
12:29
413
P1: JZZ
references
CUFX063/Wagner
414
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Lui, C. and Lele, S.K. (2002a): A Numerical Investigation of Broad-Band Shock Noise, AIAA Paper
2002-0074.
Lui, C. and Lele, S.K. (2002b): A Numerical Study of Shock-Associated Noise, AIAA Paper 20022530.
Lui, C. and Lele, S.K. (2003): Sound Generation Mechanism of Shock-Associated Noise, AIAA
Paper 2003-3315.
Lund, T.S., Wu, X. and Squires, K.D. (1998): Generation of turbulent inflow data for spatiallydeveloping boundary layer simulations, Journal of Computational Physics, 140:233258.
Lush, P.A. (1971): Measurements of subsonic jet noise and comparison with theory, Journal of Fluid
Lyrintzis, A.S. (1994): Review: The use of Kirchhoffs method in computational aeroacoustics, ASME
Journal of Fluids Engineering, 116:665676.
Maestrello, L. (1976): Two-Point Correlations of Sound Pressure in the Far Field of a Jet: Experiment.
NASA-TM-X-72835. Technical Report, NASA Langley Research Center, Hampton, VA.
Magagnato, F. (1998): KappaKarlsruhe parallel program for aerodynamics, Tasc Quarterly,
1(4):215270. Gdansk.
Magagnato, F. and Gabi, M. (2000): A new adaptive turbulence model for unsteady flow fields in
rotating machinery, Proceedings of the 8th International Symposium on Transport Phenomena and
Dynamics of Rotating Machinery, Volume I, Honolulu, Hawaii.
Magagnato, F., Pantle, I. and Gabi, M. (2002): Prediction of aero-acoustic noise around a circular
cylinder by large-eddy simulation, Proceedings of the International Workshop on LES for Acoustics, German Aerospace Center, DLR, Gottingen, Germany, DGLR-Report 2002-03.
Magagnato, F., Pritz, B. and Gabi, M. (2005): LES on Internal and External Flow Fields Using a
Compressible Solver. To be published at Proceedings of the ERCOFTAC International Symposium
on Engineering Turbulence Modelling and Measurements -ETMM6-, Sardinia, Italy.
Magagnato, F., Sorguven, E. and Gabi, M. (2003): Far Field Noise Prediction by Large Eddy Simulation
and Ffowcs WilliamsHawkings Analogy, AIAA Paper 2003-3206.
Magnient, J.C., Sagaut, P. and Deville, M. (2001): A study of built-in filter for some, eddy viscosity
models in large-eddy simulation, Physics of Fluids, 13(5):14401449.
Mahesh, K., Lee, S., Lele, S. and Moin, P. (1995): The interaction of an isotropic field, of acoustic
waves with a shock wave, Journal of Fluid Mechanics, 300:383407.
Mahesh, K., Lele, S. and Moin, P. (1997): The influence of entropy fluctuations on the interaction of
turbulence with a shock wave, Journal of Fluid Mechanics, 334:353379.
Mandal, J.C. and Deshpande, S.M. (1994): Kinetic flux vector splitting for Euler equations, Computer,
and Fluids, 23:447.
Manhart, M. (2001): Analyzing near-wall behavior in a separating turbulent boundary layer by DNS,
4th Workshop on DNS & LES, Enschede, The Netherlands, July 1820, 2001, ERCOFTAC Series,
vol. 8, Direct and LargeEddy Simulation IV, B.J. Geurts, R. Friedrich, and O. Metais (eds.), Kluwer
Academic Publishers. Dordrecht, pp. 8188.
Mani, M. and Paynter, G.C. (2002): Hybrid Turbulence Models for Unsteady Simulation of Jet Flows,
Mankbadi, R.R., Hayder, M.E. and Povinelli, L.A. (1994): Structure of supersonic jet flow and its
radiated sound, AIAA Journal, 32(5):897906.
Mankbadi, R.R., Hixon, R. and Povinelli, L.A. (2000): Very Large-Eddy Simulations of Jet Noise,
Mankbadi, R.R., Hixon, R., Shih, S.-H. and Povinelli, L.A. (1995): On the Use of Linearized Euler
Equations in the Prediction of Jet Noise, AIAA Paper 95-0505, 33rd Aerospace Sciences Meeting.
Mankbadi, R.R., Hixon, R., Shih, S.-H. and Povinelli, L.A. (1998): Use of linearized Euler equations
for supersonic jet noise prediction, AIAA Journal, 36(2):140147.
Mankbadi, R.R., Shih, S.-H., Hixon, R. and Povinelli, L.A. (1995): Direct Computation of Sound Radiation by Jet Flow Using Large-Scale Equations, AIAA Paper 95-0680, 33rd Aerospace Sciences
Meeting.
Mankbadi, R.R., Shih, S.H., Hixon, R. and Povinelli, L.A. (2000): Direct computation of jet noise
produced by large-scale axisymmetric structures, Journal of Propulsion and Power, 16(2):207215.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Manoha, E., Delahay, C., Redonnet, S., Ben Khelil, S., Guillen, P., Sagaut, P. and Mary, I. (2001):
Numerical prediction of the unsteady flow and radiated noise from a 3D lifting airfoil, Proc. of
NATO/RTO Applied Vehicle Technology Panel, Symposium on Aging Mechanisms and Control,
Development in Computational Aero- and Hydro-Acoustics, Manchester, UK, 8-11 October.
Manoha, E., Delahay, C., Sagaut, P., Mary, I., Khelil, S.B. and Guillen, P. (2001): Numerical Prediction
of the Unsteady Flow and Radiated Noise from a 3D Lifting Airfoil, AIAA Paper 2001-2133.
Manoha, E., Elias, G., Troff, B. and Sagaut, P. (1999): Towards the Use of Boundary Element Method
in Computational Aeroacoustics, AIAA Paper 99-1980.
Manoha, E., Herrero, C., Sagaut, P. and Redonnet, S. (2002): Numerical Prediction of Airfoil Aerodynamic Noise, AIAA Paper 2002-2573, 8th CEAS/AIAA Aeroacoustics Conference, Breckenridge
(CO), USA, 1719 June, 2002.
Manoha, E., Troff, B. and Sagaut, P. (2000): Trailing-edge noise prediction using large-eddy simulation
and acoustic analogy, AIAA Journal, 38:575583.
Marble, F.E. and Candel, S.M. (1977): Acoustic disturbance from gas non-uniformities convected
through a nozzle, Journal of Sound and Vibration, 55(2):225243.
Mary, I. and Sagaut, P. (2001): Large-Eddy Simulation of Flow around an Airfoil near Stall, AIAA
Paper 2001-2559.
Mary, I. and Sagaut, P. (2002): Large-eddy simulation of the flow around an airfoil near stall, AIAA
Journal, 40(2):11391145.
Mason, P.J. and Brown, A.R. (1994): The sensitivity of large-eddy simulations of turbulent shear flow
to subgrid models, Boundary Layer Meteorology, 70:133150.
Mason, P.J. and Callen, N.S. (1986): On the magnitude of the subgrid-scale eddy coefficient in largeeddy simulations of turbulent channel flow, Journal of Fluid Mechanics, 162:439462.
McCormack, R.W. (1969): The Effect of Viscosity in Hypervelocity Impact Cratering, AIAA Paper
69-354.
McDonough, J.M. and Bywater, M. (1986): Large-scale effects on local small-scale chaotic solutions
to Burgers equation, AIAA Journal, 24(12):19241930.
Meier, G.E.A., Grabitz, G., Jungowski, W.M., Witczak, K.J. and Anderson, J.S. (1978): Oscillations
of the supersonic flow downstream of an abrupt increase in duct cross-section, Mitteilungen aus
dem Max Planck-Institut fur Stromungsforschung und der aerodynamischen Versuchsanstalt, 65:1
172.
Mein, M., Dupuy, F. and Bohineust, X. (2002): Experimental Characterization of Aeroacoustic
Sources: Subsonic Flow over a Forward-Backward Step, AIAA Paper 2002-2507.
Meinke, M., El-Askary, W.A., Zhang, Q. and Schroder, W. (2004): Large eddy simulation of airfoil
flow, Fourth Aeroacoustics Workshop, W. Schroder (ed.), TU Dresden Press, Aachen, Germany.
Meinke, M., Schroder, W., Krause, E. and Rister, T. (2001): A comparison of second- and sixth-order
methods for large-eddy simulation, Computers and Fluids, 31:695718.
Meinke, M., Schroder, W., Krause, E. and Rister, Th. (2002): A comparison of second- and sixth-order
methods for large-eddy simulations, Computers and Fluids, 31(4):695718.
Mendonca, F. (2002): Aeroacoustics DESTINY-AAC industrial project, In: STAR-CD Dynamics,
p. 5, winter 2002/2003, URL: http://www.cd-adapco.com/news/19/destiny.htm.
Mendonca, F., Allen, R., de Charentenay, J. and Kirkenham, D. (2003): CFD Prediction of Narrowband
and Broadband Cavity Acoustics at M = 0.85, AIAA Paper 2003-3303.
Mendonca, F., Read, A., Caro, S., Debatin, K. and Caruelle, B. (2005): Aeroacoustic Simulation of
Double Diaphragm Orifices in an Aircraft Climate Cooling System, AIAA Paper 2005-2976.
Meneveau, C. and Katz, J. (2000): Scale-invariance and turbulence models for large-eddy simulation,
Annual Review of Fluid Mechanics, 32:132.
Meneveau, C., Lund, T.S. and Cabot, W.H. (1996): A Lagrangian dynamic subgrid-scale model of
turbulence, Journal of Fluid Mechanics, 319:353385.
Menter, F.R. (1997): Eddy viscosity transport models and their relation to the k- model, ASME
Journal of Fluids Engineering, 119:876884.
Menter, F.R., Kuntz, M. and Bender, R. (2003): A Scale-Adaptive Simulation Model for Turbulent
Flow Predictions, AIAA Paper 2003-0767.
Michalke, A. (1984): Survey on jet instability theory, Progress in Aerospace Sciences, 21:159199.
12:29
415
P1: JZZ
references
CUFX063/Wagner
416
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Michalke, A. (1989): On the propagation of sound generated in a pipe of circular cross-section with
uniform mean flow, Journal of Sound and Vibration, 134:203234.
Michard, M., Jacob, M.C. and Grosjean, N. (2002): An experimental characterization of the flow past
an airfoil in the wake of a circular rod, Proceedings of ASME FEDSM02, FEDSM2002-31344.
Miles, J.W. (1947): The diffraction of sound due to right-angle joints in rectangular tubes, Journal of
the Acoustical Society of America, 19:572579.
Mitchell, B.E., Lele, S.K. and Moin, P. (1995a): Direct computation of the sound from a compressible
co-rotating vortex pair, Journal of Fluid Mechanics, 285:181202.
Mitchell, B.E., Lele, S.K. and Moin, P. (1995b): Direct Computation of the Sound Generated by Vortex
Pairing in an Axisymmetric Jet, AIAA Paper 95-0504.
Mitchell, B.E., Lele, S.K. and Moin, P. (1997): Direct computation of Mach wave radiation in an
axisymmetric supersonic jet, AIAA Journal, 35(10):15741580.
Mitchell, B.E., Lele, S.K. and Moin, P. (1999): Direct computation of the sound generated by vortex
pairing in an axisymmetric jet, Journal of Fluid Mechanics, 383:113142.
Miyake, Y., Bando, K. and Hori, J. (1993): Analysis of sound source distribution by means of largeeddy simulation (Japanese), Transactions of the Japan Society of Mechanical Engineers, Part B,
59(567):34753481.
Mohring, W. (1978): On vortex sound at low Mach number, Journal of Fluid Mechanics, 85:685691.
Mohring, W. (1979): Modelling low Mach number noise, Mechanics of Sound Generation in Flows,
E.-A. Muller (ed.), Springer-Verlag.
Mohring, W. (1999): A well posed acoustic analogy based on a moving acoustic medium, SWING,
Aeroacoustic Workshop, P. Koltzsch and N. Kalitzin (eds.), Dresden, Germany. DFG-BMBF: University of Stuttgart, RWTH Aachen, Dresden University of Technology, DLR Braunschweig.
Mohring, W. (2001): Energy conservation, time-reversal invariance and reciprocity in ducts with flow,
Journal of Fluid Mechanics, 431:223237.
Mohsen, A.M. (1967): Experimental Investigation of the Wall Pressure Fluctuations in Subsonic
Flows, Boeing Commercial Airplane Division Report, D6-17094.
Moin, P. and Mahesh, K. (1998): Direct numerical simulation: A tool in turbulence research, Annual
Review of Fluid Mechanics, 30:507538.
Moin, P., Squires, K., Cabot, W. and Lee, S. (1991): A dynamic subgrid-scale model for compressible
turbulence and scalar transport, Physics of Fluids A, 3:27462757.
Mollo-Christensen, E., Kolpin, M.A. and Martucelli, J.R. (1964): Experiments on jet flows and jet
noise far-field spectra and directivity patterns, Journal of Fluid Mechanics, 18:285301.
Mongeau, L., Kook, H., Brown, D.V. and Zorea, S.I. (1997): Analysis of interior pressure oscillations
induced by flow over vehicle openings, Noise Control Engineering Journal, 45:223234.
Montavon, C. (2002): A big bang in aero-acoustics, CFX Update, 22:1011, Autumn 2002.
Montavon, C., Jones, I., Szepessy, D., Henriksson, R., el-Hachemi, Z., Dequand, S., Piccirillo, M.,
Tournour, M. and Tremblay, F. (2002): Noise propagation from a cylinder in a cross flow: Comparison of SPL from measurements and from a CAA method based on a generalised acoustic analogy.
IMA Conference on Computational Aeroacoustics, University of Greenwich, 911, April 2002.
Moon, Y.J., Sung, R.K., Cho, Y. and Chung, J.M. (1999): Aeroacoustic computations of the unsteady
flows over a rectangular cavity with a lip, 3rd Computational Aeroacoustics Workshop, pp. 347353,
Ohio Aerospace Institute, Cleveland, Nov. 810.
Moore, C.J. (1977): The role of shear layer instabilities in jet exhaust noise, Journal of Fluid Mechanics,
80:321367.
Morfey, C.L. (1973): Amplification of aerodynamic noise by convected flow inhomogeneities, Journal
of Sound and Vibration, 31:391397.
Morfey, C.L. (1976): Sound radiation due to unsteady dissipation in turbulent flows, Journal of Sound
and Vibration, 48:95111.
Morfey, C.L. (2003): The role of viscosity in aerodynamic sound generation, International Journal
of Aeroacoustics, 2:230240.
Morinishi, Y. and Vasilyev, O.V. (2002): Vector level identity for dynamic subgrid scale modeling in
large eddy simulation, Physics of Fluids, 14(10):36163623.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Morley, A.W. (1939): Estimation of aeroplane noise level: Some empirical laws with an account of the present experiments on which they are based, Aircraft Engineering, 11(123):187
189.
Morris, P. J. and Farassat, F. (2002): Acoustic analogy and alternative theories for jet noise prediction,
Morris, P.J., Long, L.N., Bangalore, A. and Wang, Q. (1997): A parallel three-dimensional computational aeroacoustics method using non-linear disturbance equations, Journal of Computational
Physics, 133:5674.
Morris, P.J., Long, L.N. and Scheidegger, T. (1999): Parallel Computations of High Speed Jet Noise,
AIAA Paper 99-1873.
Morris, P.J., Long, L.N., Scheidegger, T.E. and Boluriaan, S. (2002): Simulations of supersonic jet
noise, International Journal of Aeroacoustics, 1(1):1741.
Morris, P.J., Scheidegger, T. and Long, L.N. (2000): Jet Noise Simulations for Circular Nozzles, AIAA
Paper 2000-2080.
Morris, P.J., Wang, Q., Long, L.N. and Lockhard, D.P. (1997): Numerical Predictions of High-Speed
Jet Noise, AIAA Paper 97-1598.
Morse, P.M. and Feshbach, H. (1953): Methods of Theoretical Physics Volume I, McGraw-Hill Book
Company, New York.
Morse, P.M. and Ingard, K.U. (1968): Theoretical Acoustics, McGraw-Hill Book Company, New York.
Moss, W.D. and Baker, S. (1980): Re-circulating flows associated with two-dimensional steps, Aeronautical Quarterly, August: 151172.
Muchinsky, A. (1996): A similarity theory of locally homogeneous anisotropic turbulence generated
by a Smagorinsky-type LES, Journal of Fluid Mechanics, 325:239260.
Muller, E.A. and Obermeier, F. (1988): Vortex sound, Fluid Dynamics Research, 3:4351.
Munt, R.M. (1977): The interaction of sound with a subsonic jet issuing from a semi-infinite cylindrical
pipe, Journal of Fluid Mechanics, 83(4):609640.
Munt, R.M. (1990): Acoustic radiation properties of a jet pipe with subsonic jet flow: I. The cold jet
reflection coefficient, Journal of Sound and Vibration, 142(3):413436.
Murlis, J., Tsai, H.M. and Bradshaw, P. (1982): The structure of turbulent boundary layers at low
Reynolds numbers, Journal of Fluid Mechanics, 122:1356.
Musafir, R.E. (1997): A discussion on the structure of aeroacoustic wave equations, Proceedings of
the 4th Congress on Acoustics, Marseille, France, pp. 923926, Toulouse, France, April 1997.
Teknea edition.
Myers, M.K. (1980): On the acoustic boundary condition in the presence of flow, Journal of Sound
Myers, M.K. (1986a): An exact energy corollary for homentropic flow, Journal of Sound and Vibration,
109:277284.
Myers, M.K. (1986b): Generalization and Extension of the Law of Acoustic Energy Conservation in
a Nonuniform Flow, AIAA Paper 86-0471.
Myers, M.K. (1991): Transport of energy by disturbances in arbitrary flows, Journal of Fluid Mechanics, 226:383400.
Najjar, F.M. and Tafti, D.K. (1996): Study of discrete test filters and finite difference approximations
for the dynamic subgrid-scale stress model, Physics of Fluids, 8(4):10761088.
Najm, H.N. and Ghoniem, A.F. (1991): Numerical simulation of the convective instability in a dump
combustor, AIAA Journal, 29(6):911919.
Nance, D.V., Viswanathan, K. and Sankar, L.N. (1997): Low-dispersion finite volume scheme for
aeroacoustic applications, AIAA Journal, 35:255262.
NASA Glenn Research Center. (2001): Proceedings of the Jet Noise Workshop, NASA/CP2001211152, Cleveland, OH, 11070.
Nederveen, C.J. (1998): Acoustical Aspects of Woodwind Instruments, Northern Illinois University
Press, DeKalb, 2nd edition.
Nelson, P.A., Halliwell, N.A. and Doak, P.E. (1983): Fluid dynamics of flow excited resonance,
Part 2: Flow acoustic interaction, Journal of Sound and Vibration, 91:375402.
12:29
417
P1: JZZ
references
CUFX063/Wagner
418
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Nelson, P.M. (1987): Transportation Noise Reference Book, Butterworths & Co. (Publishers) Ltd.,
London, UK.
Nicolopoulos, D., Perie, F. and Jacques, A. (2004): Direct Numerical Simulation of Aero-Acoustic
Phenomena, Whitepaper, M 3 , France.
Nicoud, F. and Ducros, F. (1993): Subgrid-scale stress modelling based on the square of the velocity
gradient tensor, Flow, Turbulence and Combustion, 62(3):183200.
Norberg, C. (1992): Pressure forces on a circular cylinder in a cross flow, IUTAM Symposium on Bluff
Body Wakes, Dynamics and Instabilities, Gottingen, pp. 275278.
Norton, M.P. (1989): Fundamentals of Noise and Vibration Analysis for Engineers, Cambridge University Press, Cambridge, UK.
Oberai, A.A., Roknaldin, F. and Hughes, T.J.R. (2002): Trailing-Edge Noise Due to Turbulent Flows,
Report No. 02-002, Boston University.
Oberlack, M. (1997): Invariant modeling in large-eddy simulation of turbulence, Annual Research
Briefs Center for Turbulence Research, 322.
Obermeier, F. (1985): Aerodynamic sound generation caused by viscous processes, Journal of Sound
and Vibration, 99:111120.
Obrist, D., Nicolopoulos, D. and Jacques, A. (2002): Aero acoustic simulation of a side mirror compared with experimental results, Internoise, August 2002.
Oh, K.J. and Colonius, T. (2002): Large Eddy Simulation of the Compressible Flow over an Open
Cavity, Proceedings of ASME FEDSM02, Montreal, Canada, July 1418. ASME Paper FEDSM
2002-31352.
Orkwis, P.D., Sekar, B., Chakravarthy, S. and Peroomian, O. (1998): Comparison of three Navier
Stokes solvers for supersonic open cavity simulations, AIAA Journal, 36(5):865867.
Orlanski, I. (1976): A simple boundary condition for unbounded flows, Journal of Computational
Physics, 21:251269.
Ota, D.K., Chakravarthy, S.R., Becker, T. and Sturznegger, T. (1994): Computational study of resonance suppression of open sunroofs, ASME Journal of Fluids Engineering, 116:877882.
Ovenden, N.C. (2002): Near Cut-On/Cut-Off Transitions in Lined Ducts with Flow, AIAA Paper
2002-2445.
Ovenden, N.C. (2005): A uniformly valid multiple scales solution for sound in flow ducts with cut-on
cut-off transition, Journal of Sound and Vibration, 286:403416.
Ovenden, N.C., Eversman, W. and Rienstra, S.W. (2004): Cut-On Cut-Off Transition in Flow Ducts:
Comparing Multiple-Scales and Finite-Element Solutions, AIAA Paper 2004-2945.
oruk, Y. and Long, L.N. (1996a): A new efficient algorithm for computational aeroacoustics on
Ozy
parallel processors, Journal of Computational Physics, 125:135149.
oruk, Y. and Long, L.N. (1996b): Computation of sound radiating from engine inlets, AIAA
Ozy
Journal, 34:894901.
oruk, Y. and Long, L.N. (1997): Multigrid acceleration of a high-resolution computational aeroaOzy
coustics scheme, AIAA Journal, 35:428433.
Page, G.J., McGuirk, J.J., Behrouzi, P. and Hossain, M. (2001): A CFD coupled acoustics approach
for the prediction of coaxial jet noise, Proceedings of the NATO RTO-AVT Symposium on Ageing
Mechanisms and Control, Part A Developments in Aero and Hydro-Acoustics, Manchester, UK,
811 Oct. 2001.
Pagneux, V. and Froelich, B. (2001): Influence of low Mach number shearflow on acoustic propagation
in ducts, Journal of Sound and Vibration, 246:137155.
Paliath, U. and Morris, P.J. (2003): Numerical Simulation of Two-Dimensional Laminar and Turbulent
Cavity Flows, AIAA Paper 2003-3233.
Panda, J. and Seasholtz, R.G. (2002): Experimental investigation of density fluctuations in high-speed
jets and correlation with generated noise, Journal of Fluid Mechanics, 450:97130.
Panda, J., Seasholtz, R.G., Elam, K.A., Mielke, A.F. and Eck, D.G. (2004): Effect of Heating on
Turbulent Density Fluctuations and Noise Generation from High Speed Jets, AIAA Paper 20043016.
Pantano, C. and Sarkar, S. (2001): A subgrid model for nonlinear functions of a scalar, Physics of
Fluids, 13(12):38033819.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Pantle, I. (2002): Stromungsakustik auf der Basis akustischer Analogie mit LES und URANS, Ph.D.
thesis, Karlsruhe University.
Paterson, A.R. (1983): A First Course in Fluid Dynamics, Cambridge University Press, Cambridge,
UK.
Pereira, J.M.C., Kobayashi, M.H. and Pereira, J.C.F. (2001): A fourth-order-accurate finite volume
compact method for the incompressible NavierStokes solutions, Journal Computational Physics,
167:217243.
Pereira, M.J., Kobayashi, M.H. and Pereira, C.J. (2000): High order compact schemes in finite volume
context, ECCOMAS 2000, Barcelona, B. Suarez E. Onate, G. Bugeda (eds.), CIMNE.
Perie, F. (2002): A direct approach to CAA, Internoise 2002, Dearborn, MI, USA, August 1921,
2002.
Peroomian, O., Chakravarthy, S., Palaniswamy, S. and Goldberg, U.C. (1998): Convergence Acceleration for Unified-Grid Formulation Using Preconditioned Implicit Relaxation, AIAA Paper
98-0116.
Perot, F., Auger, J.M., Giardi, H., Bailly, C. and Juve, D. (2004): Computation of the noise generated
by low Mach number flows around a cylinder and a wall-mounted half-cylinder, 10th AIAA/CEAS
Aeroacoustic Conference, Manchester, UK, AIAA Paper 2004-2859.
Perzon, S. and Davidson, L. (2000): On transient modelling of the flow around vehicles using the
Reynolds equation, ACDF 2000, Beijing, China, pp. 720727.
Peters, M.C.A.M. and Hirschberg, A. (1993): Acoustically induced periodic vortex shedding at sharp
edged open channel ends: Simple vortex models, Journal of Sound and Vibration, 161:281299.
Peters, M.C.A.M., Hirschberg, A., Reijnen, A.J. and Wijnands, A.P.J. (1993): Damping and reflection
coefficient measurements for an open pipe at low Mach and low Helmholtz numbers, Journal of
Petropoulos, P. (2000): Reflectionless sponge layers as absorbing boundary conditions for the numerical solution of Maxwells equations in rectangular, cylindrical and spherical coordinates, SIAM
Journal of Applied Mathematics, 60:1037.
Phillips, O.M. (1960): On the generation of sound by supersonic turbulent shear layers, Journal of
Fluid Mechanics, 9(1):128.
Pierce, A.D. (1981): Acoustics: An Introduction to Its Physical Principles and Applications, McGrawHill Book Company, New York. Also available from the Acoustical Society of America.
Pierce, A.D. (1990): Wave equation for sound in fluids with unsteady inhomogeneous flow, Journal
of the Acoustical Society of America, 87(6):22922299.
Piller, M. and Stalio, E. (2004): Finite-volume compact schemes on staggered grids, Journal of
Computational Physics, 197(1):299340.
Piomelli, U. (2001): Large-Eddy Simulation of the Sound Emission from Unsteady Flow Separation,
Final Report on Grant NAG-1 1880, Department of Mechanical Engineering, College Park, MD.
Piomelli, U. and Balaras, E. (2002): Wall-layer models for large-eddy simulation, Annual Review of
Piomelli, U., Ferziger, J.H., Moin, P. and Kim, J. (1989): New approximate boundary conditions for
large-eddy simulations of wall-bounded flows, Physics of Fluids A, 1(6):10611068.
Piomelli, U., Streett, G.L. and Sarkar, S. (1997): On the computation of sound by large-eddy simulation,
Journal of Engineering Mathematics, 32(5):217236.
Poinsot, T.J. and Lele, S.K. (1992): Boundary conditions for direct simulations of compressible viscous
flows, Journal of Computational Physics, 101:104129.
Povitsky, A. (2001): High-Order Compact Simulation of Wave Propagation in a Non-uniform Flow,
Povitsky, A. and Morris, P.J. (2000): A higher-order compact method in space and time based on
parallel implementation of the Thomas algorithm, Journal of Computational Physics, 161:182
203.
Powell, A. (1964): Theory of vortex sound, Journal of the Acoustical Society of America, 36(1):177
195.
Powell, A. (1990): Some aspects of aeroacoustics: From Rayleigh until today, Journal of Vibration
and Acoustics, 112:145159.
12:29
419
P1: JZZ
references
CUFX063/Wagner
420
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Prandtl, L. and Tietjens, O.G. (1957): Fundamentals of Hydro- and Aeromechanics, Dover, New York.
Prieur, J. and Rahier, G. (1998): Comparison of Ffowcs WilliamsHawkings and Kirchhoff Rotor
Noise Calculations, AIAA Paper 98-2376.
Pruett, C.D. (2000): Eulerian time-domain filtering for spatial large-eddy simulation, AIAA Journal,
38(9):16341642.
Pruett, C.D., Zang, T., Chang, C.-L. and Carpenter, M.H. (1995): Spatial direct numerical simulation
of high-speed boundary layer flows. Part I. Algorithmic considerations and validation, Theoretical
Computational Fluid Dynamics, 7:4976.
Pulliam, T.H. and Chaussee, D.S. (1981): A diagonal form of an implicit approximate-factorization
algorithm, Journal of Computational Physics, 39:347363.
Quemere, P. and Sagaut, P. (2002): Zonal multi-domain RANS/LES simulation of turbulent flows,
International Journal for Numerical Methods in Fluids, 40:903925.
Quinn, M.C. and Howe, M.S. (1984): On the production and absorption of sound by lossless liners in
the presence of mean flow, Journal of Sound and Vibration, 97(1):19.
Rahier, G. and Prieur, J. (1997): An efficient Kirchhoff integration method for rotor noise prediction
starting indifferently from subsonically or supersonically rotating meshes, 53rd Annual Forum of
the American Helicopter Society, Virginia Beach, VA, AprilMay 1997.
Rai, M.M. and Moin, P. (1991): Direct Numerical Simulation of Transition and Turbulence in a
Spatially Evolving Boundary Layer, AIAA Paper 91-1607.
Ramboer, J. and Lacor, C. (2002): A finite volume formulation for compact upwind schemes , Proc.
2nd Int. Conf. on CFD, Sydney, July 2002.
Ramboer, J., Smirnov, S. and Lacor, C. (2002): Use of compact schemes in CAA and LES applications,
Proc. of Workshop LES for Acoustics, DLR Goettingen, 78 October 2002.
Ramboer, J., Smirnov, S. and Lacor, C. (2003): Finite Volume Formulation of Upwind Compact
Schemes for CAA and LES Applications, AIAA Paper 2003-3967.
Ravichandran, K.S. (1997): Higher order KFVS algorithms using compact upwind difference operators, Journal of Computational Physics, 130:161173.
Read, A. and Mendonca, F. (2004): Numerical simulation methods for aeroacoustics with an overview
of industrial test cases, 5th MIRA International Vehicle Aerodynamics Conference, UK, October
2004.
Read, A., Mendonca, F., Barone, F., Durello, P., Carena, F., Gallez, X., Ploumhans, P. and Caro, S.
(2004): Comparison between Measured and Predicted Tonal Noise from a Subsonic Fan Using
a Coupled Computational Fluid Dynamics (CFD) and Computational Acoustics (CA) Approach,
Redmore, T.L. and Mulholland, K.A. (1982): The application of mode coupling theory to the transmission of sound in the sidebranch of a rectangular duct system, Journal of Sound and Vibration,
85:323331.
Redonnet, S., Manoha, E. and Sagaut, P. (2001): Numerical Simulation of Propagation of Small
Perturbations Interacting with Flows and Solid Bodies, AIAA Paper 2001-2223.
Reichl, C., Grogger, H.A., Krenn, Ch., Kaltenhauser, A., Holl, M., Mann, M. and Lang, H. (2003):
Aeroacoustics characterisation of an exterior mirror applying numerical and experimental techniques, Computational Aeroacoustics, Euromech Colloquium No. 449, Chamonix, France.
Rembold, B. and Kleiser, L. (2003): Noise Prediction from Rectangular Jet Flow Using LES, AIAA
Paper 2003-3281.
Ricot, D., Maillard, V., Bailly, C. and Juve, D. (2001): Numerical Simulation of the Unsteady Flow
past a Cavity and Application to Sun Roof-Buffeting, AIAA Paper 2001-2112.
Rienstra, S.W. (1981a): Sound diffraction at a trailing edge, Journal of Fluid Mechanics, 108:443460.
Rienstra, S.W. (1981b): On the acoustical implications of vortex shedding from an exhaust pipe,
Transactions of the ASME, Journal of Engineering for Industry, 103(4):378384.
Rienstra, S.W. (1983): A small Strouhal number analysis for acoustic wavejet flowpipe interaction,
Journal of Sound and Vibration, 86(4):539556.
Rienstra, S.W. (1984): Acoustic radiation from a semi-infinite annular duct in a uniform subsonic
mean flow, Journal of Sound and Vibration, 94(2):267288.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Rienstra, S.W. (1985): Contributions to the theory of sound propagation in ducts with bulk-reacting
lining, Journal of the Acoustical Society of America, 77(5):16811685.
Rienstra, S.W. (1986): Hydrodynamic instabilities and surface waves in a flow over an impedance
wall, G. Comte-Bellot and J.E. Ffowcs Williams (eds.), Proceedings IUTAM Symposium Aeroand Hydro-Acoustics 1985 Lyon, Springer-Verlag, Heidelberg, pp. 483490.
Rienstra, S.W. (1988): 1-D reflection at an impedance wall, Journal of Sound and Vibration, 125:43
51.
Rienstra, S.W. (1999): Sound transmission in slowly varying circular and annular ducts with flow,
Rienstra, S.W. (2002): A classification of duct modes based on surface waves, Wave Motion, 37(2):119
135.
Rienstra, S.W. (2003): Sound propagation in slowly varying lined flow ducts of arbitrary cross section,
Rienstra, S.W. and Hirschberg, A. (2001): An Introduction to Acoustics, Report IWDE 0103 May
2001, Eindhoven University of Technology, Eindhoven, the Netherlands.
Rizzetta, D.P. (1988): Numerical simulation of supersonic flow over a three dimensional cavity, AIAA
Journal, 26(7):799807.
Rizzetta, D.P. and Visbal, M.R. (2002): Large-Eddy Simulation of Supersonic Cavity Flowfields
Including Flow Control, AIAA Paper 2002-2853.
Rockwell, D. (1983): Oscillations of impinging shear layers, AIAA Journal, 21(5):645664.
Rockwell, D. and Naudascher, E. (1978): Review Self-sustaining oscillations of flow past cavities,
ASME Journal of Fluids Engineering, 100:152165.
Roe, P.L. (1981): Approximate Riemann solvers, parameter vectors, and difference schemes, Journal
Rogallo, R.S. (1981): Numerical Experiments in Homogeneous Turbulence, NASA Technical Memorandum, TM81315.
Rogallo, R.S. and Moin, P. (1984): Numerical simulation of turbulent flows, Annual Review of Fluid
Roger, M. (2004): Noise from moving surfaces. Advances in Aeroacoustics and Applications, J.
Anthoine (ed.), Brussels, Von Karman Institute, Lecture Notes Series 20045.
Rona, A. and Dieudonne, W. (2000): A Flow-Resonant Model of Transonic Laminar Open Cavity
Instability, AIAA Paper 2000-1967.
Ronchi, C., Ypma, M. and Canuto, V.M. (1992): On the application of the Germano identity to
subgrid-scale modeling, Physics of Fluids A, 4(12):29272929.
Roozen, N.B., Bockholts, M., van Eck, P. and Hirschberg, A. (1998): Vortex sound in bass-reflex
ports of loudspeakers: Parts i & ii, Journal of the Acoustical Society of America, 104:1914
1924.
Rossiter, J.E. (1964): Wind Tunnel Experiments on the Flow over Rectangular Cavities at Subsonic
and Transonic Speeds, Technical Report RAE-ARC-R&M 3438, Aeronautical Research Council
Reports and Memoranda.
Rowley, C.W., Colonius, T. and Basu, A.J. (2002): On self-sustained oscillations in two-dimensional
compressible flow over rectangular cavities, Journal of Fluid Mechanics, 455:315346.
Roy, S. and Cho, P. (1999): Designing Axial Flow Fan for Flow and Noise, SAE, September 1999.
Saffman, P.G. (1992): Vortex Dynamics, Cambridge Monograph on Mechanics and Applied Mathematics. Cambridge University Press, Cambridge, UK.
Sagaut, P. (2001): Large-Eddy Simulations of Incompressible Flows. An Introduction, Springer-Verlag,
Berlin.
Sagaut, P. (2002): Large-Eddy Simulation for Incompressible Flows. An Introduction, Springer-Verlag,
Berlin, 2nd edition.
Sagaut, P. (2004): Private communication with M. Breuer.
Sagaut, P., Garnier, E., Tromeur, E., Larchevèque, L. and Labourasse, E. (2003): Turbulent Inflow Conditions for LES of Supersonic and Subsonic Wall-Bounded Flows, AIAA Paper 20030068.
12:29
421
P1: JZZ
references
CUFX063/Wagner
422
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Sagaut, P. and Grohens, R. (1999): Discrete filters for large-eddy simulation, International Journal
for Numerical Methods in Fluids, 31:11951220.
Sagaut, P., Montreuil, E. and Labbe, O. (1999): Assessment of some self-adaptive SGS models for
wall-bounded flows, Aerospace Science and Technology, 6:335344.
Sakurai, M., Endo, M. and Perie, F. (2002): Development of the exhaust systems radiation noise simulation technology, ASME 2003 Fluids Engineering Division Summer Meeting, Honolulu (USA),
July 611, 2003.
Salvetti, M.V. and Banerjee, S. (1994): A priori tests of a new dynamic subgrid-scale model for
finite-difference large-eddy simulations, Physics of Fluids, 7(11):28312847.
Sandham, N.D., Li, Q. and Yee, H.C. (2002): Entropy splitting for high-order numerical simulation
of compressible turbulence, Journal of Computational Physics, 178:307322.
Scheichl, S. (2004): On the calculation of the transmission line parameters for long tubes using the
method of multiple scales, Journal of the Acoustical Society of America, 115:534555.
Schlichting, H. (1968): Boundary-Layer Theory, McGraw-Hill Book Company, New York, 6th edition.
Schram, C. and Hirschberg, A. (2003): Application of vortex sound theory to vortex-pairing noise:
Sensitivity to errors in flow data, Journal of Sound and Vibration, 266:10791098.
Schreiber, L. (1995): Straenverkehrslarm. Taschenbuch der Technischen Akustik, M. Heckl and H.A.
Muller (eds.), Springer-Verlag, Berlin, 2nd edition, pp. 348354.
Schroder, W., Meinke, M., Ewert, R. and El-Askary, W. (2001): LES of the turbulent flow around a
sharp trailing edge, Direct and Large-Eddy Simulation IV, Kluwer Academic Publishers, Dordrecht,
pp. 353363.
Schumann, U. (1975): Subgrid scale model for finite difference simulations of turbulent flows in plane
channels and annuli, Journal of Computational Physics, 18:376404.
Scotti, A. and Meneveau, C. (1997): Fractal model for coarse-grained nonlinear partial differential
equations, Physical Review Letters, 78(5):867870.
Seror, C. and Sagaut, P. (2000): Subgrid-scale contribution to noise production in decaying isotropic
turbulence, AIAA Journal, 38(10).
Seror, C. and Sagaut, P. (2002): On the noise prediction computed by hybrid methods, Proc. Workshop
LES for Acoustics, Gottingen.
Seror, C., Sagaut, P., Bailly, C. and Juve, D. (1999): Subgrid Scale Contribution to Noise Production in
Decaying Isotropic Turbulence, AIAA Paper 99-1979, 5th CEAS/AIAA Aeroacoustics Conference,
Seattle, USA, 1012 May 1999.
Seror, C., Saguat, P., Bailley, C. and Juve, D. (2000): Subgrid-scale contribution to noise production
in decaying isotropic turbulence, AIAA Journal, 38(10):17951803.
Seror, C., Sagaut, P., Bailley, C. and Juve, D. (2001): On the radiated noise computed in large-eddy
simulation, Physics of Fluids, 13(2):476487.
Shao, L., Sarkar, S. and Pantano, C. (1999): On the relationship between the mean flow and subgrid stresses in large-eddy simulation of turbulent shear flows, Physics of Fluids, 11(5):1229
1248.
Shapiro, A.H. (1953): The Dynamics and Thermodynamics of Compressible Flow, Volume I, The
Ronald Press Co., New York.
Shen, W.Z. and Srensen, J. (2001): Acoustic modeling of turbulent airfoil flows, AIAA Journal,
39(6):10571064.
Shieh, C.M. and Morris, P.J. (1999a): A parallel numerical simulation of automobile noise involving
feedback, 3rd Computational Aeroacoustics Workshop, pp. 363370, Ohio Aerospace Institute,
Cleveland, Nov. 810.
Shieh, C.M. and Morris, P.J. (1999b): Parallel Numerical Simulation of Subsonic Cavity Noise,
AIAA Paper 99-1891.
Shieh, C.M. and Morris, P.J. (2000): Parallel Computational Aeroacoustic Simulation of Turbulent
Subsonic Cavity Flow, AIAA Paper 2000-1914.
Shieh, C.M. and Morris, P.J. (2001): Comparison of Two- and Three-Dimensional Turbulent Cavity
Flows, AIAA Paper 2001-0511.
Shih, S.H., Hamed, A. and Yeuan, J.J. (1994): Unsteady supersonic cavity flow simulations using
coupled k and NavierStokes equations, AIAA Journal, 32(10):20152021.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Shih, T.-H., Povinelli, L.A. and Liu, N.-S. (2003): Application of generalized wall function for complex
turbulent flows, Journal of Turbulence, 4:015.
Shim, I.B., Kim, J.W. and Lee, D.J. (1999): Numerical Study on Radiation of Multiple Pure Tone
Noise from an Aircraft Engine Inlet, AIAA Paper 99-1831.
Shur, M.L., Spalart, P.R. and Strelets, M.K. (2003): Noise prediction for increasingly complex jets,
Colloquium EUROMECH 449, Dec. 912, 2003, Chamonix, France.
Siegert, R., Schwarz, V. and Reichenberger, J. (1999): Numerical Simulation of Aeroacoustic Sound
Generated by Generic Bodies Placed on a Plate: Part II Prediction of Radiated Sound Pressure,
AIAA Paper 99-1895.
Sijtsma, P., Oerlemans, S. and Holthusen, H. (2001): Location of Rotating Sources by Phased Array
Measurements, AIAA Paper 2001-2169.
Singer, B.A., Brentner, K.S., Lockard, D.P. and Lilley, G.M. (1999): Sound Prediction of an Airfoil
in the Wake of a Cylinder, AIAA Paper 99-0231.
Sinha, N., Arunajatesan, S. and Ukeiley, L.S. (2000): High-Fidelity Simulation of Weapons Bay
Aeroacoustics and Active Flow Control, AIAA Paper 2000-1968.
Slimon, S.A., Davis, D.W. and Wagner, C.A. (1998): Far-Field Aeroacoustic Computation of Unsteady
Cavity Flow, AIAA Paper 98-0285.
Smagorinsky, J. (1963): General circulation experiments with the primitive equations. I: The basic
experiment, Monthly Weather Review, 91(3):99165.
Smirnov, S., Lacor, C. and Baelmans, M. (2001): A Finite Volume Formulation for Compact Schemes
with Applications to LES, AIAA Paper 2001-2546.
Smirnov, A., Shi, S. and Celik, I. (2001): Random flow generation technique for large-eddy simulations
and particle-dynamics modeling, ASME Journal of Fluids Engineering, 123:359371.
Smith, B.R. (2001): Large-eddy simulation of supersonic cavity flow with an unstructured grid flow
solver, 3rd AFOSR International Conference on Direct Numerical Simulation and Large-Eddy
Simulation (TAICDL), Arlington, TX, Aug. 59.
Smith, B.R., Jordan, J.K., Bender, E.E. and Rizk, S.N. (2000): Computational Simulation of Active
Control of Cavity Acoustics, AIAA Paper 2000-1927.
Smith, M.J.T. (1989): Aircraft Noise, Cambridge University Press, Cambridge, UK.
Smolyakov, A.V. and Tkachenko, V.M. (1991): Model of a field of pseudosonic turbulent wall pressures and experimental data, Soviet Physics Acoustics, 37(6):289292.
Snyder, R.D. and Scott, J.N. (1999): Comparison of Numerical Schemes for the Analysis of Aeroacoustics, AIAA Paper 99-0354.
Soemarwoto, B.I. and Kok, J.C. (2001): Computations of three-dimensional unsteady cavity flow to
study the effect of different downstream geometries, AGARD RTO-AVT Symposium on Development
in Computational Aero- and Hydroacoustics, Manchester, UK.
Sorguven, E., Magagnato, F. and Gabi, M. (2003): Simulation of acoustic scattering from a trailing
edge, Computational Aeroacoustics: From Acoustic Sources Modeling to Far-Field Radiated Noise
Prediction, Colloquium EUROMECH 449, December 912, 2003, Chamonix, France.
Souliez, F.J., Long, L.N., Morris, P.J. and Sharma, A. (2002): Landing gear aerodynamics noise
prediction using unstructured grids, International Journal of Aeroacoustics, 1(2):115135.
Spalart, P.R. (1988): Direct simulation of a turbulent boundary layer up to Re = 1410, Journal of
Fluid Mechanics,187:6198.
Spalart, P.R. (2000): Strategies for turbulence modeling and simulations, International Journal of
Heat and Fluid Flow, 21:252263.
Spalart, P.R., Jou, W.H., Strelets, M. and Allmaras, S.R. (1997): Comments on the feasibility of LES
for wings and on a hybrid RANS/LES approach, 1st AFOSR Int. Conf. on DNS/LES, August 48,
1997, Ruston, LA. In Advances in DNS/LES, C. Liu and Z. Liu (eds.), Greyden Press, Columbus,
OH.
Spalart, P.R. and Leonard, A. (1987): Direct numerical simulation of equilibrium turbulent boundary
layers, Proc. 5th Sym. on Turbulent Shear Flows, Ithaca NY, Springer-Verlag, pp. 235252.
Spalart, P.R. and Squires, K.D. (2003): The status of detached-eddy simulation for bluff bodies, UEF
Conf. Aerodynamics of Heavy Vehicles: Trucks, Buses and Trains, Asilomar, December 26, 2002,
Springer-Verlag.
12:29
423
P1: JZZ
references
CUFX063/Wagner
424
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Speziale, C.G. (1996): Computing non-equilibrium flows with time-dependent RANS and VLES,
15th International Conference on Numerical Methods in Fluid Dynamics, Monterey.
Speziale, C.G. (1998): Turbulence modeling for time-dependent RANS and VLES: A review, AIAA
Journal, 36(2):173184.
Speziale, C.G., Abid, R. and Anderson, E.C. (1990): A Critical Evaluation of Two-Equation Models
for Near Wall Turbulence, ICASE Report, No. 9046.
Spille-Kohoff, A. and Kaltenbach, H.-J. (2001): Generation of turbulent inflow data using a prescribed
shear-stress profile, DNS/LES Progress and Challenges, C. Liu, L. Sakell, and T. Beutner (eds.),
Greyden, Columbus, OH, pp. 319326.
Sreedhar, M. and Ragab, S. (1994): Large-eddy simulation of longitudinal stationary vortices, Physics
of Fluids, 6(7):25012514.
Srinivasan, S. and Baysal, O. (1991): NavierStokes calculations of transonic flows past cavities,
ASME Journal of Fluids Engineering, 113:368376.
Stanescu, D. and Habashi, W.G. (1998): 2N -storage low dissipation and dispersion Runge-Kutta
schemes for computational acoustics, Journal of Computational Physics, 143:674681.
Stanescu, D., Hussaini, M.Y. and Farassat, F. (2002): Aircraft Engine Noise Scattering A Discontinuous Galerkin Spectral Element Approach, AIAA Paper 2002-0800.
Stanley, S. and Sarkar, S. (2000): Influence of nozzle conditions and discrete forcing on turbulent
planar jets, AIAA Journal, 38:16151623.
STAR-CD and pro-STAR User Guide (2004): STAR-CD and pro-STAR User Guide, CD-adapco,
URL: http://www.cd-adapco.com.
Stolz, S. and Adams, N.A. (1999): An approximate deconvolution procedure for large-eddy simulation,
Physics of Fluids, 11(7):16991701.
Stolz, S. and Adams, N.A. (2001): LES of supersonic boundary layers using the approximate deconvolution model, Direct and Large-Eddy Simulation IV, Kluwer Academic Publishers, Dordrecht,
pp. 269276.
Stolz, S., Adams, N.A. and Kleiser, L. (2001a): An approximate deconvolution model for large-eddy
simulations with application to incompressible wall-bounded flows, Physics of Fluids, 13(4):997
1015.
Stolz, S., Adams, N.A. and Kleiser, L. (2001b): The approximate deconvolution model for large-eddy
simulations of compressible flows and its application to shock-turbulent-boundary-layer interaction,
Streett, C.L. and Macaraeg, M.G. (1989): Spectral multi-domain for large-scale fluid dynamic simulations, Applied Numerical Mathematics, 6:123139.
Strelets, M. (2001): Detached Eddy Simulation of Massively Separated Flows, AIAA Paper 20010879.
Stromberg, J.L., McLaughlin, D.K. and Troutt, T.R. (1980): Flow field and acoustic properties of a
Mach number 0.9 jet at a low Reynolds number, Journal of Sound Vibration, 72:159176.
Suhs, N.E. (1987): Computations of Three-Dimensional Cavity Flow at Subsonic and Supersonic
Speeds, AIAA Paper 87-1208.
Suhs, N.E. (1993): Unsteady Flow Computations for a Three Dimensional Cavity with or Without an
Acoustic Suppression Device, AIAA Paper 93-3402.
Taasan, S. and Nark, D.M. (1995): An Absorbing Buffer Zone Technique for Acoustic Wave Propagation, AIAA Paper 95-0164.
Tam, C. and Dong, Z. (1993): A study of the short wave components in computational acoustics,
Journal of Computational Acoustics, 1(1):130.
Tam, C. and Dong, Z. (1994): Wall boundary condition for high-order finite-difference schemes in
computational aeroacoustics, Theoretical Computational Fluid Dynamics, 6(6):303322.
Tam, C.-J., Orkwis, P.D. and Disimile, P.J. (1995): Comparison of BaldwinLomax turbulence models
for two-dimensional open cavity computations, AIAA Journal, 34(3):629631.
Tam, C.K.W. (1995): Computational Aeroacoustics: Issues and Methods, AIAA Paper 950677.
Tam, C.K.W. (1998a): Advances in numerical boundary conditions for computational aeroacoustics,
Journal of Computational Acoustics, 6:377402.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Tam, C.K.W. (1998b): LES for Aeroacoustcs, AIAA Paper 98-2805.

Tam, C.K.W. (2001): Noise from high-speed jets, J. Anthoine and C. Schram (eds.), Advances in
Aeroacoustics, VKI Lecture Series 20012, Brussels, Von Karman Institute for Fluid Dynamics,
pp. 134.
Tam, C.K.W. and Auriault, L. (1998): Mean flow refraction effects on sound radiated from localized
sources in a jet, Journal of Fluid Mechanics, 370:149174.
Tam, C.K.W. and Auriault, L. (1999): Jet mixing noise from fine-scale turbulence, AIAA Journal,
37(2):145153.
Tam, C.K.W. and Block, P.J.W. (1978): On the tones and pressure oscillations induced by flow over
rectangular cavities, Journal of Fluid Mechanics, 89(2):373399.
Tam, C.K.W., Golebiowski, M. and Seiner, J.M. (1996): On the Two Components of Turbulent Mixing
Noise from Supersonic Jets, AIAA Paper 1996-1716.
Tam, C.K.W. and Morris, P.J. (1980): The radiation of sound by the instability waves of a compressible
plane turbulent shear layer, Journal of Fluid Mechanics, 98(2):349381.
Tam, C.K.W. and Shen, H. (1993): Direct Computation of Nonlinear Acoustic Pulses Using HighOrder Finite Difference Schemes, AIAA Paper 93-4325.
Tam, C.K.W. and Webb, J.C. (1993): Dispersion-relation-preserving finite difference schemes for
computational acoustics, Journal of Computational Physics, 107:262281.
Tam, C.K.W., Webb, J.C. and Dong, Z. (1993): A study of the short wave components in computational
acoustics, Journal of Computational Acoustics, 1:130.
Tang, L. and Baeder, J.D. (1997): Uniformly Accurate Compact Difference Schemes, AIAA Paper
97-2093.
Temkin, S. (2001): Elements of Acoustics, Acoustical Society of America, New York. Reprint.
Tennekes, H. and Lumley, J.L. (1972): A First Course in Turbulence, MIT Press, Cambridge,
MA.
Terracol, M., Labourasse, E., Manoha, E. and Sagaut, P. (2003): Simulation of the 3D Unsteady Flow
in a Slat Cove for Noise Prediction, AIAA Paper 2003-3110.
Terracol, M., Manoha, E., Herrero, C., Labourasse, E., Redonnet, S. and Sagaut, P. (2005): Hybrid
methods for airframe noise numerical prediction, TCFD, 19(3):197227.
Terracol, M., Sagaut, P. and Basdevant, C. (2001): A multilevel algorithm for large-eddy simulation
of turbulent compressible flows, Journal of Computational Physics, 167(2):439474.
Thompson, K.W. (1990): Time-dependent boundary conditions for hyperbolic systems II, Journal of
Thompson, M.C., Hourigan, K. and Welsh, M.C. (1992): Acoustic sources in a tripped flow past a
resonator tube, AIAA Journal, 30:14841491.
Thomson, K.W. (1987): Time dependent boundary conditions for hyperbolic systems, Journal of
Tolstykh, A.I. and Shirobokov, D.A. (1995): Fifth-order compact upwind method and its applications
to three-dimensional compressible NavierStokes equations, Proc. 1st Asian CFD Conf.
Toro, E.F. (1992): Viscous flux limiters, Notes on Numerical Fluid Mechanics, Vos, Rizzi, and
Rhyming (eds.), Vieweg Verlag Wiesbaden, 35:592600.
Travin, A., Shur, M., Strelets, M. and Spalart, P.R. (2000): Detached-eddy simulations past a circular
cylinder, International Journal of Flow Turbulence and Combustion, 63:293313.
Troff, B., Manoha, E. and Sagaut, P. (1997): LES of Trailing Edge Flow with Application to Radiated
Noise, AIAA Paper 1997-120.
Tyler, J.M. and Sofrin, T.G. (1962): Axial flow compressor noise studies, Transactions of the Society
of Automotive Engineers, 70:309332.
Urbin, G. and Knight, D. (2001): Large-eddy simulation of a supersonic boundary layer using an
unstructured grid, AIAA Journal, 39(7):12881295.
Uzun, A. (2003): 3-D Large-Eddy Simulation for Jet Aeroacoustics. Ph.D. thesis, Purdue University,
West Lafayette, IN.
Uzun, A., Blaisdell, G.A. and Lyrintzis, A.S. (2002): Recent Progress Towards a Large-Eddy Simulation Code for Jet Aeroacoustics, AIAA Paper 2002-2598.
12:29
425
P1: JZZ
references
CUFX063/Wagner
426
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Uzun, A., Blaisdell, G.A. and Lyrintzis, A.S. (2003): 3-D Large-Eddy Simulation for Jet Aeroacoustics, AIAA Paper 2003-3322.
van Kuiken, G.D.C. (1995): Thermodynamics of Irreversible Processes: Applications to Diffusion and
Rheology, John Wiley & Sons, Chichester, UK.
Van Leer, B. (1979): Towards the ultimate conservative difference scheme.V. A second-order sequel
to Godunovs method, Journal of Computational Physics, 32:101136.
van Lier, L., Dequand, S., Hirschberg, A. and Gorter, J. (2001): Aeroacoustics of diffusers: An
experimental study of typical industrial diffusers at Reynolds numbers of O(105 ), Journal of the
Acoustical Society of America, 109:108117.
van Wijngaarden, L. (1968): On the oscillations near and at resonance in open pipes, Journal of
Engineering Mathematics, 11:225240.
van Wijngaarden, L. (1972): One-dimensional flow of liquids containing small gas bubbles, Annual
Review of Fluid Mechanics, 4:369396.
Vasilyev, O., Lund, T.S. and Moin, P. (1998): A general class of commutative filters for LES in complex
geometries, Journal of Computational Physics, 146:82104.
Verge, M.P., Fabre, B., Hirschberg, A. and Wijnands, A.P.J. (1997a): Sound production in recorderlike
instruments, I: Dimensionless amplitude of the internal acoustic field, Journal of the Acoustical
Society of America, 101:29142924.
Verge, M.P., Hirschberg, A. and Causse, R.R. (1997b): Sound production in recorderlike instruments,
II: A simulation model, Journal of the Acoustical Society of America, 101:29252939.
Vergne, S., Auger, J.-M., GStyr, N. and Perie, F. (2002): Simulation of cavity aero-elastic noise
induced by an external turbulent flow perturbed by a small ruler, Proceedings of the International
Workshop on LES for Acoustics, 78 October, German Aerospace Center, DLR, Gottingen,
Germany, DGLR-Report 2002-03.
Verzicco, R., Iafrati, A., Riccardi, G. and Fatica, M. (1997): Analysis of the sound generated by the
pairing of two axisymmetric co-rotating vortex rings, Journal of Sound and Vibration, 200:347358.
Vichnevetsky, R. (1981): Propagation through numerical mesh refinement for hyperbolic equations,
Mathematics and Computers in Simulation, 23:344.
Visbal, R.M. and Gaitonde, V.D. (1998): Higher-Order-Accurate Methods for Complex Unsteady
Subsonic Flows, AIAA Paper 98-0131.
Visbal, R.M. and Gaitonde, V.D. (1999): Higher-order-accurate methods for complex unsteady flows,
AIAA Journal, 37:12311239.
Visbal, M.R. and Gaitonde, D.V. (2001): Very high-order spatially implicit schemes for computational
acoustics on curvilinear meshes, Journal of Computational Acoustics, 9(4): 12591286.
Visbal, R.M. and Gaitonde, V.D. (2002): On the use of higher-order finite-difference schemes on
curvilinear and deforming meshes, Journal of Computational Physics, 181:155185.
Visbal, M.R. and Rizzetta, D.P. (2002): Large-eddy simulation on curvilinear grids using compact
differencing and filtering schemes, ASME Journal of Fluids Engineering, 124:836847.
von Terzi, D.A. and Fasel, H.F. (2002): A New Flow Simulation Methodology Applied to the Turbulent
Backward-Facing Step, AIAA Paper 2002-0429.
Vreman, B. (1995): Direct and Large-Eddy Simulation of the Compressible Turbulent Mixing Layer.
Ph.D. Thesis, Twente University.
Vreman, B., Geurts, B. and Kuerten, H. (1997): Large-eddy simulation of the turbulent mixing layer,
Wagner, C. (2001): An eddy viscosity scaled similarity model, Z. Angew. Math. Mech., 81(Suppl. 3):
491492.
Wagner, C. (2003): A non-dissipative dynamic subgrid scale similarity model, Proc. of the Third
International Symposium on Turbulence and Shear Flow Phenomena, Sendai, Japan, 2527 June
2003.
Wagner, C. and Huttl, T. (2002): LES for acoustics, Proceedings of the International Workshop on
LES for Acoustics, 78 October, German Aerospace Center, DLR, Gottingen, Germany, DGLRReport 2002-03.
Wagner, S., Bareiss, R. and Guidati, G. (1996): Wind Turbine Noise, Springer-Verlag, Berlin.
12:29
P1: JZZ
references
CUFX063/Wagner
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Walz, A. (1969): Boundary Layers of Flow and Temperature, MIT Press, Cambridge, MA.
Wang, M., Lele, S.K. and Moin, P. (1996): Computation of Quadrupole Noise from Airfoil VortexShedding, Stanford University CTR Manuscript 158.
Wang, M. and Moin, P. (2000): Computation of trailing-edge flow and noise using large-eddy simulation, AIAA Journal, 38(12):22012209.
Weber, C. and Ducros, F. (2000): Large-eddy and Reynolds-averaged NavierStokes simulation of
turbulent flow over an airfoil, International Journal of Computational Fluid Dynamics, 13:327
355.
Welsh, M.C. and Stokes, A.N. (1984): Flow-resonant sound interaction in a duct containing a plate,
Part I: Semi-circular leading edge, Journal of Sound and Vibration, 95:305323.
Werner, H. and Wengle, H. (1993): Large-eddy simulation of turbulent flow over and around a cube
in a plate channel, 8th Symp. on Turb. Shear Flows, F. Durst, R. Friedrich, B. Launder, F. Schmidt,
U. Schumann, and J. Whitelaw (eds.), Springer Verlag, Berlin, pp. 155168.
Westley, R. and Lilley, G.M. (1952): An Investigation of the Noise Field from a Small Jet and Methods
for Its Reduction, Technical Report 53, College of Aeronautics, Cranfield, UK.
Wilson, R.V., Demuren, A.O. and Carpenter, M. (1998): Higher-Order Compact Schemes for Numerical Simulation of Incompressible Flows, ICASE Report 98-13.
Wilson, T.A., Beavers, G.S., De Coster, M.A., Holger, D.K. and Regenfuss, D. (1971): Experiments
on the fluid mechanics of whistling, Journal of the Acoustical Society of America, 50:366372.
Winant, C.D. and Browand, F.K. (1974): Vortex pairing: The mechanism of turbulent mixing layer
growth at moderate Reynolds number, Journal of Fluid Mechanics, 63(2):237255.
Winckelmans, G.S., Wray, A.A., Vasilyev, O.V. and Jeanmart, H. (2001): Explicit-filtering largeeddy simulation using the tensor-diffusivity model supplemented by a dynamic Smagorinsky term.
Wu, X., Jacobs, R., Hunt, J. and Durbin, P. (1999): Simulation of boundary layer transition induced
by periodically passing wake, Journal of Fluid Mechanics, 398:109153.
Wustenberg, H. and Hupertz, B. (1995): Moglichkeiten und Grenzen der Aerodynamik-Berechnung
mit Zwei-Gleichungs-Turbulenzmodellen und Feinauflosung der Grenzschicht. Seminar Aerodynamik des Kraftfahrzeugs, Haus der Technik, Essen.
Xu, J., Stanescu, D., Hussaini, M.Y. and Farassat, F. (2003): Computation of Engine Noise Propagation
and Scattering off an Aircraft, AIAA Paper 2003-0542.
Yoon, S. and Jameson, A. (1987): An LU-SSOR Scheme for the Euler and NavierStokes Equation,
AIAA-Paper 87-600.
Yoshizawa, A. (1979): A statistical investigation upon the eddy viscosity in incompressible turbulence,
Journal of the Physical Society of Japan, 47(5):16651669.
Yoshizawa, A. (1984): Statistical analysis of the deviation of the Reynolds stress from its eddy-viscosity
representation, Physics of Fluids, 27(6):13771387.
Yu, S.-T., Hsieh, K.-C. and Tsai, Y.-L.P. (1995): Simulating waves in flows by RungeKutta and
compact difference schemes, AIAA Journal, 33:1421429.
Zaman, K.B.M.Q. (1985): Effect of initial condition on subsonic jet noise, AIAA Journal, 23(9):1370
1373.
Zaman, K.B.M.Q. (1986): Flow field and near and far sound field of a subsonic jet, Journal of Sound
Zang, Y., Street, R.L. and Koseff, J.R. (1993): A dynamic mixed subgrid-scale model and its application
to turbulent recirulating flows, Physics of Fluids A, 5(12):31863196.
Zhang, H.L., Bachman, C.R. and Fasel, H.F. (2000a): Application of a New Methodology for Simulations of Complex Turbulent Flows, AIAA Paper 2000-2535.
Zhang, H.L., Bachman, C.R. and Fasel, H.F. (2000b): Reynolds-Averaged NavierStokes Calculations
of Unsteady Turbulent Flow, AIAA Paper 2000-0143.
Zhang, X. (1995): Compressible cavity flow oscillation due to shear layer instabilities and pressure
feedback, AIAA Journal, 33(8):14041411.
Zhang, X., Chen, X.X., Rona, A. and Edwards, J.A. (1998): Attenuation of cavity flow oscillation
through leading edge flow control, Journal of Sound and Vibration, 221(1):2347.
12:29
427
P1: JZZ
references
CUFX063/Wagner
428
0 521 87144 1
printer: sheridan
November 17, 2006
REFERENCES
Zhang, X., Rona, A. and Lilley, G.M. (1995): Far-Field Noise Radiation from an Unsteady Supersonic
Cavity, AIAA Paper 95-040.
Zhao, H. and Voke, P.R. (1996): A dynamic subgrid-scale model for low-Reynolds-number channel
flow, International Journal for Numerical Methods in Fluids, 23:1927.
Zhao, W., Frankel, S.H. and Mongeau, L. (2001b): Numerical simulations of sound from confined
puslating axisymmetric jets, AIAA Journal, 39(10):18681874.
Zhao, L. and Cangellaris, A.C. (1996): GT-PML: Generalized theory of perfectly match layers and
its application to the reflectionless truncation of finite-difference time-domain grids, IEEE Transactions on Microwave Theory and Techniques, 44:25552563.
Zhao, W., Frankel, S.H. and Mongeau, L. (2000): Large-Eddy Simulation of Sound Radiation from a
Subsonic Turbulent Jet, AIAA Paper 00-2078.
Zhao, W., Frankel, S.H. and Mongeau, L. (2001a): Large-eddy simulations of sound radiation from
subsonic turbulent jets, AIAA Journal, 39(8):14691477.
Zhong, X. (1998): High-order finite difference schemes for numerical simulation of hypersonic
boundary-layer transition, Journal Computational Physics, 144:662709.
Zhou, Y. (1993): Interacting scales and energy transfer in isotropic turbulence, Physics of Fluids A,
5(10):25112524.
Ziada, S. and Buhlmann, E.T. (1992): Self-excited resonances of two side-branches in close proximity,
Journal of Fluids and Structures, 6:583601.
12:29
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
18:15
Index
90 bend, 207
a posteriori test, 18
a priori evaluation, 125
a priori test, 18
absorbing cell, 154
absorbing conditions, 353
absorbing zone, 217, 218, 219, 220, 222, 386
acoustic analogy, 16, 17, 21, 55, 128, 273, 296,
299, 300, 311, 334
acoustic feedback, 272, 356
acoustic flow feedback, 358
acoustic flow interaction, 264
acoustic perturbation equation (APE), 147, 224,
296, 297
acoustic resonances, 263, 349, 351
acoustic scattering, 174, 196, 294, 318, 329
acoustic wave propagation, 143, 295
acoustical admittance, 46
acoustical displacement, 34
acoustical flow separation, 34
acoustics receivers, 274
Adams-Bashforth scheme, 201, 329
advective upstream splitting method (AUSM), 173,
305
adverse pressure gradient, 228
aeroacoustic coupling, 354, 356
aeroacoustic instability, 354
aeroacoustic resonance, 357, 358
aerodynamic instabilities, 357
aerodynamic noise, 273
aeroelastic loading, 289
aeroelastic noise, 272, 273, 293
aeroelastic phenomena, 387
aeroengine, 58
aerospace cavity, 358
aerovibroacoustics, 273
Ahmed body, 345
air-conditioning, 345, 373
aircraft, 378
aircraft engine, 189, 197

aircraft industry, 222
aircraft manufacturer, 4, 378
aircraft noise, 2, 3, 39, 382
aircraft propeller, 52
aircraft turbine, 71
aircraft wing, 58, 381
airfoil, 209, 225, 294, 295, 318, 336, 371
airfoil leading edge, 325, 333, 387
airfoil noise, 318
airfoil trailing edge, 226
airframe noise, 222, 232, 294, 295, 296
airplane stability, 273
ALESSIA, 6
algebraic turbulence model, 172, 212
aliasing error, 109, 110
analogy of Howe, 67
analytic acoustic propagation model, 145
analytical transport technique, 11
anechoic pipe termination, 81
anechoic wind tunnel, 337
angle of attack, 318
annoyance, 2, 3
annular shear layer, 248, 254, 257, 261
annuli, 210
antidissipative effect, 124
A-pillar, 357, 360
approach, 294
approximation error, 9
arbitrary Lagrangian Eulerian formulation (ALE),
276
Archimedes, 63
artifical noise, 13
artificial boundary, 234
artificial damping, 219
artificial dispersion, 248
artificial dissipation, 110, 171, 175, 187, 189, 219,
336, 379
artificial inflow boundaries, 203, 204
429
P1: PJU
0521871441ind
430
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
artificial inflow data, 207

artificial noise, 235
artificial reflection, 358
artificial selective damping (ASD), 173, 175, 301
artificial velocity fluctuations, 143
artificial viscosity, 110
ASMO model, 345
asymmetric DRP scheme, 175
asymptotic radiation boundary condition, 301
attached boundary layer, 209
attached eddy, 213
attached flow, 212
audio range, 33
autocorrelations, 205, 208
automobile, 360
automobile manufacturer, 357
automotive exhaust, 274, 358
automotive heating, 373
automotive intake, 358
autospectrum, 285, 286
auxiliary filter, 117
auxiliary simulation, 206, 207, 226
axisymmetric jet, 238
axisymmetric mixing layer, 239, 241
axisymmetric nozzle, 246
axisymmetric shear layer, 241, 243, 244
azimuthal correlation, 258
backscattering, 14, 113, 114
backward facing step, 213
backward scattering, 334
Baldwin-Lomax model, 263, 318
bell-shaped linear pulse, 189
benchmark problem, 181, 195, 197
benchmark validation case, 337
Bessel equation, 251
Bessel function, 69, 70
bidiagonal interpolation, 193
bidiagonal third-order accurate scheme, 193
bidimensional mixing layer, 239
bifurcation, 266, 276
bilinearly interpolation, 309
Biot-Savart law, 234, 235, 237
blade-passing frequency (BPF), 375
Blasius boundary layer, 268
blended wing-body combination, 196
blower fan, 373
blunt body, 333
blunted trailing edge, 320
bluntness, 318
boundary condition, 248
boundary element method (BEM), 294, 344, 348
boundary layer, 318, 356
Boussinesq closure, 135
Boussinesq eddy viscosity concept, 263
Boussinesq model, 134
Boussinesq-like model, 126

box filter, 94
breathing mode, 161
broadband, 350, 358, 361, 364, 371, 372, 373
broadband amplification phenomenon, 232
broadband flow excitations, 376
broadband noise, 6, 8, 15, 17, 128, 148, 156, 167,
252, 318, 364, 385
buffer zone, 218, 220
buffeting, 357
buoyancy, 63, 91, 214
bypass, 58
bypass ratio, 18
bypass stream, 18
bypass the wall layer, 267
cabin climate control device, 357
car, 154, 333, 334, 345, 387
car aerodynamics, 345
car manufacturer, 349
car roof, 346
car underbody, 346
Cartesian grid, 181
car-wing-mirror, 154
catastrophic failure, 273
cavity, 262, 263, 265, 266, 357, 364, 386, 387
cavity acoustics, 172
cavity ceiling, 364
cavity drag, 265
cavity flow, 15, 263, 266, 272
cavity noise, 262, 264, 265
cavity opening, 266
cavity oscillations, 263, 268, 271
cavity resonance, 263
cavity resonance frequency, 371, 372
cavity wall, 269
cell isotropy, 358
cell-centered finite volume, 317
cell-centered version, 192
centered scheme, 111
central difference scheme, 195, 305, 346
central scheme, 171, 194
centrifugal forces, 214
C-grid topology, 314
channel flow, 187, 206, 213, 226
chaotic acoustic signal, 342
chaotic structure, 340
characteristic decomposition, 217
characteristic nonreflecting boundary condition,
217
characteristic splitting, 217
chevron, 246, 254
Chimera technique, 14
Cholesky decomposition, 144
circular cylinder, 20, 21, 132, 186, 333, 334, 335,
336, 387
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
431
INDEX
circular duct, 365

circular jet, 175, 239, 241, 250
closed pipe, 75
cockpit climate control device, 357
coexistence of different modes, 271, 272
coherence function, 285
coherency-squared function, 285
coherent structure, 87, 129, 209, 238, 242, 269,
272, 387
combustion, 36, 42, 63, 91
combustion noise, 3, 8
commercial software, 148, 164, 316, 344, 345,
348, 357, 374
compact central scheme, 176, 191
compact filtering, 266
compact finite difference, 176
compact finite element projection method, 195
compact finite volume, 261
compact region, 42
compact scheme, 175, 178, 182
compact source, 48
compact upwind scheme, 176, 190, 191
competitive mode, 271
compressible subgrid model, 125
compressor fan, 373
computational aeroacoustics (CAA), 6, 7, 223,
297
computational cost, 11
computational transport method, 11
condensation, 36
cone of silence, 52, 382
conservative finite volume method, 353
control device, 350
convecting point source, 236
convecting vortex problem, 237
convective boundary conditions, 204
convective ridge, 324, 325
convolution, 235
convolution filter, 90, 92, 93, 95, 96, 101, 379
convolution product, 316
cooling and climate system, 378
cooling fan, 357
coordinate transformation, 185, 206
Coriolis acceleration, 31
Coriolis force, 66
corotating vortex pair, 173, 241
correlation length and time, 151
counterclockwise rotating vortex, 236
counterrotating vortices, 236, 266
coupling, 149
coupling mechanisms, 10
coupling region, 10
coupling technique, 295, 330
Courant condition, 140, 277, 280
Courant number, 112
cover plate, 266
crackle, 251
Crank-Nicolson scheme, 201
cross correlation, 16, 205, 208, 285
cross spectrum, 285, 320
cross-spectral density (CSD), 286
cross-stress tensor, 97
cruise flight condition, 52, 209
C-type mesh, 306
Curles formulation, 20, 59, 60, 62, 142, 143, 264,
383
curved surface, 214
curvilinear coordinates, 185
curvilinear grid, 185, 301, 313, 353
cutoff, 266, 286, 379
cutoff frequency, 69, 72
cutoff length, 89, 90, 92, 93, 100, 101, 104, 111,
112, 115, 116, 117
cutoff mode, 69
cutoff time, 93
cutoff wave number, 17, 92
cuton mode, 69, 71
cyclical pattern, 271
cylinder wake, 371
cylindrical Rayleigh equation, 250
damping, 151, 219
damping coefficient, 154
damping function, 267
damping zone, 234, 235, 237
data injection process, 329
dealiasing, 171
decay of homogeneous isotropic flow, 202
decaying turbulence, 202
decomposition, 207
deconvolution model, 119, 120, 124, 380
deep cavity, 267
detached eddy simulation (DES), 131, 213, 264,
345, 364
deterministic vortex model, 380
diaphragm, 71, 75, 76, 81, 349, 350, 351
differentiation error, 109, 110
diffraction, 54, 325
diffusion, 113
diffusion error, 9, 10, 11, 13, 385
diffusor, 85, 354
digital filtering of random data, 208
dilatation field, 160, 163, 240, 241
dipole, 39, 42, 43, 44, 47, 48, 58, 59, 60, 294, 309,
344, 382
direct injection, 296
direct method, 9
direct noise computation (DNC), 239, 244, 264,
265, 386
direct numerical simulation (DNS), 7, 15, 18, 19,
58, 89, 129, 198, 239, 246, 249, 294, 334,
349, 379, 387
18:15
P1: PJU
0521871441ind
432
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
directivity, 242
directivity function, 79
directivity pattern, 253
Dirichlet boundary conditions, 202, 204
discontinuity, 193
discontinuous boundary, 149
discontinuous Galerkin method, 176, 195
discrete tone, 148, 156
discretization error, 124
discretization grid, 111, 151
discretization scheme, 9, 11, 13, 167, 170, 179, 180
discretization stencil, 168, 172
dispersion, 113, 199, 248, 265
dispersion error, 9, 10, 11, 13, 111, 168, 169, 170,
173, 174, 177, 180, 189, 193, 196, 199, 200,
317, 380, 385
dispersion relation, 247
dispersion relation preserving (DRP), 173, 248,
301, 350
dispersive behavior, 178, 181, 192, 197, 201, 385
dissipation, 31, 113, 153, 195, 199, 248, 265, 267,
276
dissipation error, 111, 168, 169, 170, 175, 189,
193, 196, 199, 200, 317, 380
dissipation rate, 115, 132, 138, 144
dissipative behavior, 197, 201, 385
dissipative filter, 188
dissipative model, 266
dissipative weighted compact scheme, 193
divergence free, 249, 250
divergent solutions, 296
domain decomposition, 148, 163
door cavity, 266
Doppler amplification, 50, 303
Doppler effect, 11, 49, 52, 60, 382, 383, 384
Doppler factor, 50, 51, 52
double spiral structure, 241
downstream corner, 262, 272, 353
downstream edge, 269
downstream shedding, 372
downwind interpolation, 194
downwind-biased formulation, 194
driven cavity, 266
dual time stepping, 346
duct, 357
duct flow, 206, 350, 354
duct mode, 69
ducted cavity, 353
ducted turbofan, 171
dynamic model, 123
dynamic procedure, 121
dynamic Smagorinsky model, 243, 259, 266
dynamic subgrid scale model, 135, 253
eardrum, 34
eddy, 109
eddy viscosity, 263

eddy viscosity model, 18, 114, 123, 266, 267
edge-tone, 85
effective angular frequency, 201
effective filter, 92, 112
eigenfunction, 68, 69, 70
eigenfunction solutions, 250
eigenvalue, 69, 70, 71, 72
eighth-order filter, 188
Einsteins convention, 26, 300
ejection boundary condition, 211
ejections, 211
elastic plate-cavity system, 274
eleven-point stencil, 269, 330
elliptic equations, 201
end correction, 78, 80, 83, 84, 85
energy aliasing, 189
energy cascade, 110, 111, 113
energy spectra, 208
engine cooling device, 345
engine noise, 294
engine nozzle, 248
entropy splitting, 111
environmental pollution, 1
equilibrium turbulent-boundary layer, 228, 229
error control reliability, 199
error reduction, 167
essentially nonoscillatory scheme (ENO), 173,
193
Etkin test case, 335
Euler flux, 317
Eulerian time derivative, 26
European Union, 5
exact amplification factor, 199, 200
exhaust plume, 246, 386
exit zone, 219, 220
explicit filter, 187
explicit scheme, 177
explicit time integration, 317
external mixer, 245
external vehicle noise, 378
extrapolation, 150, 203, 214
extremely loud jet, 251
factorization error, 183
fan blade, 374
fan noise, 62
far-field radiation, 149, 175
Favre filtering operator, 316
feedback, 6, 14, 24, 57, 60, 150, 337, 349, 351,
357, 383, 384
feedback loop, 262, 263, 266
Ffowcs Williams-Hawkings equation, 7, 11, 23,
216, 294, 334, 344, 384, 387
Ffowcs Williams-Hawkings method, 60, 62, 133,
175, 264, 374, 383
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
433
INDEX
Ffowcs Williams-Hawkings surface, 133, 251, 253,

254, 371
Fifth Research Framework Program, 6
fifth-order accurate scheme, 190
filter kernel, 93, 100, 102, 118, 119
finite difference, 19, 105, 214, 269, 295
finite difference method, 92, 104, 248
finite difference scheme, 172, 173, 176, 181, 182,
186, 189, 192, 194
finite element, 173, 277, 278, 371
finite element method, 20, 92, 95, 105, 195
finite element scheme, 248
finite volume, 19, 95, 316, 317
finite volume method, 92, 104, 105, 195, 196, 253
finite volume scheme, 176, 181, 193, 194, 248
five stage algorithm, 201
five stage Runge-Kutta scheme, 200, 305
flap, 158, 294, 357
flat plate, 187, 231
flat plate boundary layer, 213
flow acoustic coupling, 349, 351
flow acoustic interaction, 349
flow simulation methodology (FSM), 213
flow with discontinuities, 191
flow-induced noise, 357
flow-induced vibration, 273
fluid acoustic computation, 353
fluid-structure interaction, 273, 274
flute, 85, 383
flux difference splitting, 172
flux splitting, 191, 192
flux vector splitting, 192, 193
Fourier analysis, 37, 41, 109, 277
Fourier decomposition, 359
Fourier difference scheme, 179, 180
Fourier footprint, 169
Fourier method, 95
Fourier mode, 167
Fourier mode analysis, 167
Fourier model, 144
Fourier reconstruction, 144
Fourier series, 188
Fourier space, 94, 235
Fourier spectral method, 109
Fourier synthesis, 49
Fourier transformation, 37, 47, 71, 73, 94, 156,
173, 202, 235, 285, 302, 303, 313, 323, 340,
386
Fourier wave, 168, 175, 187, 196
four-level time integration, 201
fourth-order accurate scheme, 171, 179, 182, 183,
186, 187, 199
free flight experiment, 205
free jet, 58, 59
free shear layer, 205, 262
free vortex, 67
free-stream condition, 212

frequency domain, 180, 197
frequency domain method, 371
frequency space, 269
frequency spectrum, 8
frozen spatial turbulence field, 207
functional modeling, 112, 113
fundamental equation of thermodynamics, 28
fuselage, 174
fuselage-nacelle configuration, 197
fuselage-wing-nacelle configuration, 197
Galerkin projection, 195
Galerkin-type procedures, 94
Galilean invariance, 93, 97
Gaussian filter, 94
Gaussian wave, 181
general curvilinear grids, 100
generalized acoustic analogy, 261
generalized coordinates, 102
generic car shape, 345, 346
generic mirror, 348
geometrical singularity, 333
Germano identity, 121, 122, 380
Germanos consistent decomposition, 97
ghost cell, 150
global mass conservation, 204
Godunov theorem, 172
Godunov-type scheme, 172
gray area, 148
gray zone, 135
grazing flow, 265
Greens function, 24, 36, 37, 38, 47, 48, 49, 59, 62,
70, 71, 72, 74, 75, 76, 309, 382, 383
Greens theorem, 196, 382
grid fequency cutoff, 357
grid filter, 92
grid stretching, 203, 219, 220
gridpoints per wavelength (PPW), 178
grid-to-grid oscillation, 269
Gutins principle, 60
half-plane, 54
Hankel function, 252
Hardin-Pope two-step approach, 173
health, 1, 3
hearing loss, 1, 2
heated jet, 186, 250, 253
helicopter rotor, 8, 381
Helmholtz equation, 274
Helmholtz number, 83, 84, 223
Helmholtz resonator, 85
hemisperical half-cylinder, 154
Hermitian interpolation, 179, 181
hexahedral cell, 362
hexahedral elements, 154, 280, 292
18:15
P1: PJU
0521871441ind
434
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
high resolution LES, 209

high-bypass turbofan aeroengine, 58
higher-order moments, 208
high-lift airfoil, 152, 158
high-lift device, 294, 356, 371
high-lift wing, 158
high-lift wing device, 154
high-order accuracy, 187
high-order explicit filtering, 266
high-pass filter, 277
high-pressure steam, 354
high-speed cavity, 368
high-speed train, 5, 15, 21, 22
homogeneous isotropic flow, 386
hot jet, 253, 255
hot wire anemometry, 20, 279
H-topology, 314
h-type method, 176
human hearing range, 361
hybrid approach, 355
hybrid CFD CAA method, 357
hybrid LES-Kirchhoff method, 173
hybrid LES-LEE approach, 349
hybrid LES-Lighthill method, 173
hybrid RANS-LES approach, 213, 216, 264, 267,
380, 381
hybrid RANS-LES method, 128, 130, 133, 138,
139, 140, 142, 143, 145, 147, 148, 156, 163,
164, 165
hybrid upwind-centered discretization scheme, 317
hybrid URANS-LES, 265
hyperbolic system, 196, 215
ideal gas law, 28
impedance, 40, 44, 46, 69, 71, 77, 78, 79, 81, 83
impermeability condition, 209, 210
impinging flow, 164
impinging shear layer, 262
implicit filter, 187
implicit hypothesis, 273
Implicit method, 112
implicit temporal integration, 316, 317
implicit tridiagonal interpolation, 193
impulsive noise, 8, 381
inclined wing, 203, 204
incoming boundary layer, 263
industrial RANS code, 263
inertial range, 107, 115
inflow boundary conditions, 203, 204
inflow section, 153
inherent damping, 195
inhomogeneous wave equation, 142
in-house code, 5
initial conditions, 201, 202
injection interface, 330
inlet conditions, 238, 248, 249
inlet disturbances, 249

instability, 141, 262, 298, 354
instability theories, 238
instability wave, 262, 350
insulator, 21, 22
integral error, 201
integral scale, 105
intensity level, 33
intermittency, 271, 272
internal flow, 349, 355
internal mixer, 245
internal vehicle noise, 272, 378
interpolation, 14, 144
interpolation coefficient, 182
interpolation function, 179
intrinsic dissipation, 317
invariants of the transformation, 184
inverse energy cascade, 113
inverse Fourier transform, 208
inverse Jacobian, 184
inverse problem, 38, 39, 382
isothermal jet, 253
isotropic decaying turbulence, 139
isotropic filter, 100, 101
isotropic mesh, 358, 362
isotropic turbulence, 16, 17, 105, 106, 113, 121,
266, 380
Jacobian derivatives, 183
Jacobian matrix, 215, 217
Jacobian transformation, 172, 181, 183, 185, 186
Jameson dissipation, 189
Jameson scheme, 336
Jameson-Schmidt-Turkel scheme, 175
jet aircraft, 197
jet centerline, 251
jet core, 241
jet engine, 4
jet engine manufacturer, 4, 378
jet flow, 198
jet noise, 58, 232
jet nozzle, 357
jet plume, 247, 253, 254, 258, 261
Kelvin-Helmholtz instability, 161
kinetic energy spectrum, 108, 115
Kirchhof integral, 11, 175, 264, 294, 315, 316,
331, 332, 384
Kirchhoff approach, 133
Kirchhoff integration, 387
Kirchhoff interface, 333
Kirchhoff surface, 133, 251, 252, 253
Kirchhoffs method, 7, 19
Kolmogorov hypothesis, 108
Kolmogorov length scale, 131, 139, 222
Kolmogorov microscale, 334, 379
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
435
INDEX
Kolmogorov scale, 89, 105, 107, 108, 266

Kolmogorov spectrum, 115, 117
Kolmogorov time scale, 106
Kutta condition, 32, 55, 82, 83, 87
Kutta-Zhukhovski theorem, 236
Lagrangian integral time scale, 256
Lagrangian interpolation, 179
Lagrangian time derivative, 26
laminarization, 138
laminar-turbulent transition, 306
landing, 225
landing gear, 294, 356, 358
Laplace eigenvalue problem, 68
Laplace equation, 41
Laplace transform, 201
large-scale instability wave, 258
latency factor, 134
latency parameter, 137
lateral interface, 152
Lattice Boltzmann method, 105
law of the wall, 210
Lax-Friedrichs splitting, 191
leading edge, 55, 159, 160, 333, 357, 387
length scale, 153, 167, 208
Leonard stress tensor, 96
Leonards decomposition, 96
Lerays regularization, 119
LES-CAA coupling, 225
LES-Euler-coupling, 327, 328, 330
Lienard-Wiechert potential, 50
Lighthill source term, 246
Lighthill stress tensor, 133
Lighthill tensor, 21, 59, 373
Lighthill-Curles equation, 20, 22
Lighthills analogy, 7, 16, 17, 18, 25, 55, 56, 57, 58,
59, 62, 64, 65, 223, 241, 246, 264, 304, 350,
383, 384
Lighthills equation, 20, 142
Lilley model, 344, 346
Lilley source, 360
Lilleys equation, 246, 350
limited numerical scales (LNS), 133, 334
limiter, 172, 191, 196, 266
line integral, 186
linear combination model, 119
linear disturbance equations, 295
linear instability theory, 239
linear stochastic correlation tensor, 211
linearization error, 183
linearized acoustic perturbation equations, 296
linearized Euler approach, 7
linearized Euler equation (LEE), 11, 20, 21, 147,
170, 173, 174, 196, 198, 199, 216, 220, 221,
223, 241, 277, 295, 296, 298, 299, 334, 349,
387
linearized spatial instability problem, 250

lip line, 256, 257
listener, 37, 43, 47, 50, 51, 55, 56, 60, 63, 64, 65,
72, 73, 74, 383
local adaptive mesh refinement, 138
local isotropy hypothesis, 90
local polynomial expansion, 176
localized mesh clustering, 138
logarithmic buffer layer, 210
logarithmic law, 268
logarithmic outer layer, 210
Lorentz transformation, 52, 221
low dispersion, 295
low-dispersion scheme, 199, 200, 301
lower-upper symmetric Gauss-Seidel method, 317
low-frequency noise, 259, 260
low-frequency structures, 286
low-order scheme, 317
low-order trigonometric interpolation, 180
low-order upwind scheme, 266
low-pass filter, 187, 220, 319, 326
low-storage form, 198
low-storage scheme, 269
mass injection, 267
mass-conservation law, 27
material fatigue, 273
matrix inversion, 195
McCormack scheme, 201
mean flow-perturbation splitting, 330
medium resolution LES, 209
mesh smoothness, 197
mesh-frequency cutoff, 359, 360, 362, 376
message passing interface (MPI), 207
method of images, 73
military aircraft, 267
mixed model, 266
mixed scale model, 116
mixing layer, 16, 161, 162, 175, 186, 198, 238,
239, 241, 248, 266, 386
mixing zone, 381
mixing-length model, 213
modal analysis, 293, 387
modal decomposition, 250
mode, 68, 69, 70, 71, 72, 79, 82, 88, 144, 259
modified wave number, 109
molecular Prandtl number, 214
monoatomic gas, 99
monopole, 39, 42, 43, 44, 47, 48, 49, 51, 58, 59,
60, 63, 77, 277, 282, 371, 382, 383
monotone integrated LES (MILES), 110, 111, 126,
266, 267
monotone upstreamcentred scheme, 172
Morkovins hypothesis, 125, 229
moving body, 36, 51
moving surface, 8, 214
18:15
P1: PJU
0521871441ind
436
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
multiblock solver, 253

multidimensional stencil, 186
multilayer absorbing zone, 219
multilevel approach, 121
multilevel closure, 121
multilevel simulations, 112
multiple modes, 364
multiple streaky structures, 307
multiple tones, 263
multiple-rotating frames of reference, 374
multipole, 48, 49
multipole expansion, 47, 48, 49
Munt model, 83
Munts solution, 83, 84
NACA0012 airfoil, 19, 332, 387
nacelle, 197
narrow-band, 371
narrowband component, 321
narrowband excitations, 361, 371
narrow-band power, 372
narrowband spectra, 257
narrowband tone, 364
natural damping zone, 309
near-field noise, 149, 163
nearly axisymmetric disturbances, 249
near-wall damping, 213
near-wall region, 147, 212
near-wall structure, 210, 267
Neumann boundary conditions, 202, 203, 214, 215
Newton method, 317
Newtonian fluid, 29, 89, 91
NLDE decomposition, 151
noise carpet, 11
noise footprint, 5
noise reduction, 15, 18
noise regulations, 345
noise requirements, 378
non-Cartesian grid, 172, 181, 182, 186
noncompact upwind scheme, 190
nondissipative central difference stencil, 195
nondissipative numerical method, 209
nondissipative scheme, 379
non-Favre weighted variables, 253
nonisothermal flow, 214
nonlinear acoustics solvers, 130, 145, 154, 165
nonlinear discontinuities, 189
nonlinear disturbance equations, 146, 148, 149,
150, 165, 172, 175, 294, 299
nonlinear Euler equations, 295, 351, 353
nonlinear perturbation equations, 165
nonlinear saturation, 296
nonlinear stochastic estimation, 212
nonreflecting boundary conditions, 152, 204, 215,
217, 219, 248, 265, 302, 318, 338, 385
nonreflexive border, 329
nonresolveld wave number, 190

nonuniform grid, 172
nonuniqueness of source, 38
nonzonal approach, 212, 213
normal shock, 355
no-slip boundary condition, 208, 212, 308, 318
nozzle, 57, 77, 84, 204, 246, 248, 251, 252, 254,
257, 258, 261
nozzle shear layer, 261
nozzle-ejector mixer, 252
numerical amplification factor, 199
numerical bifurcation, 265
numerical dissipation, 108, 110, 111, 120, 126,
219, 266, 317, 334, 336
numerical error, 13, 92, 105, 110, 111, 124, 126,
151
numerical filter, 92, 112
numerical instability, 113, 114, 187
numerical noise, 13, 232
numerical reflection, 219
numerical stability, 110, 123
numerical stabilization, 111
Nyquist criterion, 363
Nyquist frequency, 113, 140, 153
Nyquist grid wavelength, 140
Nyquist length, 112
Nyquist requirement, 358
Nyquist theorem, 92
oblique shock, 355
obstacle, 207
open apertures, 357
open pipe end, 75
open pipe termination, 76, 77, 78, 79, 80, 81
open window, 358
optimization of dissipation and dispersion, 199
optimized Runge-Kutta scheme, 200
oscillating signal, 364
oscillation, 155, 161, 220, 263, 268, 286, 336, 353,
354, 355, 357, 372
oscillation frequency, 271
O-type mesh, 306
outer ear, 34
outer region, 212
outflow boundary conditions, 203, 204
outflow interface, 152
overall sound pressure level (OASPL), 253, 269
overdissipative model, 266
overlapping region, 150
overshoot, 174, 179
Pade derivative scheme, 248
Pade-type scheme, 176
pairing, 239
pairing frequency, 240
pantograph, 21
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
437
INDEX
parabolic equations, 201

paradox of dAlembert, 66
parallel computer, 207
particle image velocimetry (PIV), 279
passenger cabin, 358
passenger compartement, 349
perceived aircraft noise, 5
perfectly matched layer (PML), 217, 220, 222,
385
periodic boundary conditions, 202, 206, 226, 329,
350
periodic ejection, 265
periodic recycling, 153
perturbation injections, 329
phase shift, 78
Phillips analogy, 16
pipe flow, 206
piston engine, 5
plane channel, 150, 152, 209, 210, 212
plane jet, 238
plane mixing layer, 239
plane wave, 9, 33, 34, 35, 39, 41, 46, 53, 54, 68,
69, 72, 77, 83
plane wave mode, 72
POD-Galerkin method, 95
point source, 236
points per wavelength (PPW), 109
Poisson equation, 212
policemans whistle, 34
polyhedra, 362
poor-resolution LES, 209
postprocessing, 357, 358, 359, 360, 370
potential core, 254, 256, 258
power law, 211
power spectral density (PSD), 286, 311, 320
power train noise, 272, 273, 378
power-scaling law, 303
Prandtl number, 30, 35, 91, 126
Prandtl-Glauert factor, 332
Prandtl-Glauert transformation, 52, 221
predictor simulation, 150
predictor-corrector scheme, 195, 201
prefiltering technique, 111
prescribed energy spectrum, 208
pressure boundary conditions, 214
pressure pulse, 194
pressure reducing valve, 354
pressure spike, 381
prismatic, 134
projector operators, 94
propeller, 23, 36
propeller noise, 62
proper orthogonal disposition (POD), 95
pseudo noise, 273
pseudo sound, 357
pseudospectral methods, 95
Q-criterion, 268
quadrangular elements, 281
quadrilateral grid, 196
quadrupole, 39, 42, 44, 47, 48, 52, 63, 245, 255,
371, 382, 383
quasi-laminar annular shear layer, 248
quasi-laminar disturbances, 248, 249
quasi-laminar flow, 153
quasi-two-dimensional, 158
quiet boundary conditions, 272
q-wave, 189
radiation boundary condition, 218
radiator, 58
random error, 57, 67
random Fourier mode, 268
random velocity disturbance, 241
random-number generator, 208
Rayleigh scattering technique, 255
rear side-mirror vehicle, 274
reattaching flow, 283
reattachment, 13
reattachment length, 283, 285
reattachment point, 283, 284, 286
reattachment region, 140, 148
reattachment zone, 286
recirculating flow, 263, 266, 271
recirculation, 211
recirculation bubble, 161, 162
recirculation inside the cavity, 263
recirculation length, 286
recirculation region, 266, 283
recirculation zone, 266
reconstruction, 150
reconstruction step, 191
rectangular boxcar, 369
rectangular duct, 69, 70, 354
rectangular jet, 18, 173, 175, 250
rectangular nozzle, 246
recycled inlet plane, 143
recycling approach, 138
recycling techniques, 141
recycling zone, 139
reentry of fluid, 355
reflecting boundary conditions, 203
reflection, 204, 265, 269, 371, 385
reflections at the interfaces, 151
refraction, 82, 296, 316, 371
relaminarization, 13
rescaled boundary conditions, 206
rescaling, 138, 144, 150, 153, 206, 207, 226, 227,
228, 229, 231, 306
resolution error, 174
resolved scales, 266
resonance, 34, 148, 273, 356, 358
resonance frequency, 282, 372
18:15
P1: PJU
0521871441ind
438
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
resonant modes, 384

resonator, 88, 358, 372
resynthesizing noise, 357
reverse flow, 203
reverse-flow reciprocity principle, 74
Reynolds approach, 208
Reynolds averaging, 210
Reynolds stress, 30, 133
Reynolds stress tensor, 136
Riemann problem, 172
Riemann solver, 195, 196
Riemanns invariants, 338
road traffic, 378
road-contact noise, 5
robustness, 259
robustness of the solver, 253
rocket engine, 85
Roes scheme, 172, 193, 194
Rossiter acoustic modes, 364
Rossiter formula, 262
rotating blade, 337
rotating channel flow, 212, 213
rotating dipole, 371
rotating mode, 70
rotating quadrupole, 241
rotating surface, 214
rotor blade, 383
rotor-stator interaction, 8, 60, 71, 381
ruler, 278, 286, 387
Runge-Kutta discontinuous Galerkin method, 196
Runge-Kutta method, 196
Runge-Kutta scheme, 170, 197, 198, 199, 200,
248, 268, 269, 301, 329
Runge-Kutta stage, 170
saturation, 88
scalar energy spectrum, 202
scalar transport equation, 214
scale reduction, 267
scale similarity model, 18, 118, 380
scaled boundary condition, 226
scaling law, 8, 58, 106, 107, 109, 207, 226, 227
scaling rule, 58, 59
scattering, 52, 54, 79, 316, 333, 382, 387
Schlieren visualization, 271
Schlieren-like view, 160, 268
second law of Newton, 26
second moments, 144, 153, 165
second-moment turbulence closure, 266
second-order accuracy, 167
second-order accurate scheme, 178, 179, 187, 317,
346, 357
second-order centered scheme, 267
second-order central scheme, 169, 171
second-order dissipation, 189
second-order filter, 188
second-order statistic, 208

selective mixed-scale model, 267, 316, 317
self-similar region, 253
self-sustained oscillations, 24, 84, 85, 88, 262, 263,
271, 349, 351, 365, 384
self-sustaining physical disturbances, 153
semianalytic model, 129
semidiscretized equation, 168, 169, 170
semiempirical model, 263
semiempirical sources, 10
semiperiodic boundary condition, 226
sensitivity, 151
separated flow, 263
separated flow regions (detached eddies), 213
separated shear layer, 264, 265, 362
separation, 66, 85, 87, 111, 132, 148, 164, 211,
212, 213
separation bubble, 284
separation zone, 159, 160
seven-point stencil, 174, 330
shallow cavity, 267
sharp cutoff filter, 94, 115, 117
sharp edge, 32, 58, 67, 71, 76, 85, 86, 87, 88, 381
sharp leading edge, 173
shear flow, 266
shear layer, 52, 80, 87, 88, 106, 238, 239, 241, 242,
244, 248, 249, 254, 255, 256, 261, 262, 266,
269, 381
shear layer instability, 272
shear layer oscillation, 372
shear region, 161
shedding, 337
shedding frequency, 339, 340
shedding of vorticity, 87
shifted boundary conditions, 211
shock, 3, 252
shock capturing, 108, 110, 111, 126
shock corrugation, 108, 109
shock deformation, 108
shock pattern, 355
shock sensor, 111
shock wave, 27, 34, 108, 126, 173, 265
shock-turbulence interaction, 108, 109
side mirror, 291
sideview mirror, 357, 360
sign convention, 37, 68
silencer, 81
silent artificial boundary, 234
similarity model of Bardina, 266
simplified mirror, 349
single-point statistics, 153
six stage algorithm, 201
six stage Runge-Kutta scheme, 200
Sixth Research framework Program, 6
sixth-order accurate scheme, 182, 183
sixth-order compact scheme, 186
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
439
INDEX
sixth-order Fourier difference scheme, 180

sixth-order Taylor scheme, 180
skin friction, 153, 209
skin friction coefficient, 106
slat, 158, 196
slat cove, 152, 154, 158, 159, 371
slats, 294
slats trailing edge, 163
sleep, 1, 2
sliding interface, 374
Smagorinsky constant, 134, 140, 271
Smagorinsky model, 18, 19, 20, 22, 115, 116, 132,
134, 135, 239, 241, 259, 266, 269, 277, 334,
344, 346, 350, 388
Smagorinsky-Lilley model, 336
smooth flow, 191
solid surface, 329
solid wall boundary condition, 203, 208
solution of dAlembert, 40, 41
Sommerfeld impedance, 277
Sommerfeld solution, 54
sound intensity, 33
sound pressure level, 1, 33, 156, 157, 302, 312,
336, 347, 363
sound pressure level correction, 302
Spalart-Allmaras model, 132, 158
spatial correlation, 153, 208, 338
spatial discretization scheme, 167, 168, 169, 170
spatial stability problem, 250
spatially developing mixing layer, 238
spational discretization scheme, 199
spational DRP scheme, 200
spectral broadening, 339
spectral element implementation, 197
spectral element method, 105, 197
spectral gap, 114, 115
spectral method, 105, 110, 176
spectral/hp element method, 176
spectral-like resolution, 317
spherical dipole field, 181
spherically symmetric wave, 39, 41, 42
spherical-shaped eddies, 144
spiraling mode, 69
splitter plate, 85
splitting, 116, 328, 382
sponge layer, 218, 219, 220, 309
spray nozzle, 365
spurious effects, 113
spurious noise, 249, 304, 313
spurious perturbation, 233, 234
spurious reflections, 14, 385
spurious signal, 314
spurious sound, 77, 232, 234, 235, 237, 304, 309,
312
spurious velocity field, 233
spurious vortical perturbations, 233
spurious wave, 9, 174, 301

square cylinder, 20, 334
stability, 199
stabilization, 110, 379
stagnation flow, 186
stagnation region, 140
standard central scheme, 177
statistical subgrid noise model, 261
stencil width, 171
stochastic modeling, 350
stochastic noise generation and radiation (SNGR),
7, 359
stochastik fluctuations, 208
Stokes assumption, 91
Stokes hypothesis, 29
Stokes no-slip boundary condition, 208
strain rate, 213
stratification, 91, 92
streak, 209, 358
streaky structure, 106, 307
streamline, 203, 339, 346
streamline curvature, 161
streamline upwind Petrov-Galerkin (SUPG)
integration, 277, 289
streamwise vortices, 238
Strouhal number, 65, 79, 80, 81, 83, 84, 88, 223,
259, 263, 279, 299, 322, 339
structural modeling, 112, 113
structural response, 276
structure function model, 17, 116
structure-borne noise, 272, 378
subfilter scales, 112
subgrid dissipation, 110
subgrid kinetic energy, 117
subgrid mode, 115
subgrid model, 92, 96, 103, 104, 106, 110, 111,
113, 117, 239
subgrid model filter, 92
subgrid Reynolds stress tensor, 97
subgrid scale, 94, 112, 113, 115, 119
subgrid scale model, 16, 17, 19, 90, 110, 112, 129,
145, 166, 186, 253, 259, 262, 266, 316
subgrid scale noise model, 261, 262
subgrid tensor, 21, 97, 102, 112, 117, 118
subgrid terms, 111
successive grid refinement, 202
suction side, 161
sudden enlargement, 355
sudden expansion, 354
sunroof, 357, 358
super-bang phenomenon, 50
superposition, 153
superposition of perturbations, 202
supersonic jet, 15, 16, 18, 355
suppression device, 267
surface integral, 133
18:15
P1: PJU
0521871441ind
440
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
INDEX
surface-pressure cross spectrum, 327

surrounding node, 168
Sutherland law, 91
sweeps, 211
switching, 271
symmetric DRP scheme, 173
symmetry breaking, 202
synthesize turbulent fluctuations, 150
synthesized disturbances, 154
synthesized fluctuations, 154
synthesizing turbulent data, 153
synthetic model, 144, 153
synthetic noise, 143
synthetic perturbations, 367
synthetic reconstruction, 138, 141, 143, 147, 153,
154, 156, 164, 165, 358, 359, 380
synthetic turbulence, 142, 145, 153, 154
synthetic velocity fluctuations, 143
synthetic vortical structure, 365
synthetical reconstruction, 165
T106 turbine blade, 152
tabs, 246
takeoff, 52, 225
Taylor difference scheme, 179, 180
Taylor expansion, 167, 171, 173, 177, 182, 185
Taylor microscale, 108
Taylor series, 188
Taylor series expansions, 47, 119
Taylors hypothesis, 207, 303
temperature boundary conditions, 214
temporal correlation, 153, 208
temporal discretization scheme, 167, 168, 170,
197, 248
temporally evolving mixing layer, 238, 239
tenth-order accurate scheme, 178
tenth-order filter, 330
test filter, 104, 116
tetrahedral mesh, 134, 197, 362
theoretical filter, 92, 112
thin boundary layer equations (TBLE), 212
third-order accurate scheme, 176, 190
third-order upwind scheme, 189
three-dimensional curvilinear grid, 295
three-dimsnsional mixing layer, 239
threshold of hearing, 34
threshold of pain, 34
thunder, 74
time advancement scheme, 248
time and length scales, 165
time and space correlation, 144
time discretization, 201
time integration scheme, 199
time scale, 153, 208
time series, 208
tire noise, 272, 273, 378
T-joint, 71
tonal flow excitations, 376
tonal noise, 3, 8, 15, 129, 189, 381
total variation diminishing (TVD), 353
trailing edge, 19, 20, 55, 58, 82, 152, 159, 161,
163, 225, 273, 294, 296, 302, 305, 306, 307,
309, 311, 315, 333, 381, 387
trailing edge noise, 201, 223, 231, 233, 293, 294,
295, 296, 297, 305
train, 3, 378
transient period, 153
transition, 132, 265, 266
transition process, 202
transition region, 320
transition to turbulence, 106, 205
transition triggering, 322
transitional flow, 238
transitional shear layer, 238, 239
transverse acoustic mode, 267
transverse electromagnetic wave, 181
transverse mode pattern, 353
transverse unstable mode, 158
triangular elements, 281
triangular mesh, 196
triple decomposition, 149
truncation error, 177
turbine blade, 206
turbine engine, 8
turbine engine fan, 381
turbocharger, 357
turbofan, 58
turbofan engine, 245
turbofan revolution, 58
turbomachine, 334, 357
TurboNoiseCFD, 6
turbulence intensity, 205
turbulence investigations, 202
turbulence level, 346
turbulence model, 9, 16, 19, 263
turbulence-generated noise, 349
turbulence-related noise, 148
turbulent boundary layer, 206, 264, 268, 273, 294
turbulent bubble, 163
turbulent eddy, 10, 167, 294, 327
turbulent energy cascade, 106
turbulent fluctuation, 358
turbulent structures, 153, 238, 327
two point correlation, 153
two-engine jet aircraft, 197
two-equation turbulence model, 334
two-layer model, 212, 213
two-point correlation, 105, 206, 211, 246
two-point statistic, 208
two-step method, 264
two-step Runge-Kutta scheme, 301
two-way coupling, 150
18:15
P1: PJU
0521871441ind
CUFX063/Wagner
0 521 87144 1
printer: sheridan
October 26, 2006
441
INDEX
unheated jet, 250, 252, 253, 255, 256, 262

uniform Euler flow, 194
uniform grid, 176
uniqueness of source, 38
United States, 5
unphysical streak, 209
unresolved scales, 154, 164, 243, 266
unstaggered formula, 192
unsteady Reynolds averaged Navier-Stokes
(URANS), 150, 263, 294, 334
unwanted oscillations, 187
upstream edge, 262
upstream traveling perturbations, 203, 204
upwind biased, 172, 194
upwind compact scheme, 189
upwind discretization, 183
upwind scheme, 171, 190
valve, 349, 350
van Cittert iterative procedure, 118, 126
van Driest damping, 134, 269
van Driest function, 267, 271
van Driest transformation, 227
vehicle noise, 3
ventilation, 373, 381
ventilator, 8
very large eddy simulation (VLES), 130, 186, 200,
346, 349
vibration, 3, 273
vibroacoustics, 273, 274, 278, 387
vibroacoustics response, 279, 289
viscous boundary layer, 31
viscous damping, 296
viscous dissipation, 58
viscous limiter function, 166
viscous sublayer, 128, 210, 211
von Karman constant, 213
von Karman energy spectrum, 359
von Karman length scale, 139, 141
von Karman vortex street, 339
vortex, 194, 236, 266
vortex merging, 238
vortex method, 350
vortex pairing, 57, 161, 163, 239, 240, 242
vortex ring, 87, 88
vortex shedding, 35, 55, 67, 82, 83, 84, 132, 161,
162, 163, 265, 268, 318, 320, 321, 360, 361
vortex shedding airfoil, 304
vortex shedding frequency, 20, 322, 325, 327, 336
vortex shedding noise, 326
vortex sheet, 55, 82
vortex sound, 65
vortex sound theory, 25, 32, 80, 84, 86, 384
vortex streaks, 358
vortex structure, 339
vortex-blade interaction, 3
vortical structures, 245
wake, 20, 22, 60, 154, 203, 318, 336, 360, 367
wake mode, 265, 266
wake mode transition, 265
wake thickness, 322
wall boundary layer, 266
wall function, 141, 154, 210, 214
wall model, 107, 131, 210, 211, 212, 267, 268
wall resolving LES, 107
wall shear stress, 211, 212
wall stress model, 107
wall units, 209
wall-adapting local eddy viscosity model (WALE),
116
wall-damping function, 123
wall-modeling strategy, 214
wall-normal distance, 210
water tunnel, 265
water-air mixture, 64
wave extrapolation method, 264
wave-number-frequency spectrum, 323
wave-propagation zone, 247
wave-vector frequency spectrum, 285
weak-coupling assumption, 275, 279
weakly reflecting boundary conditions, 338
weapon bay, 267, 357, 358
weapons, 357
weighted essentially nonoscillatory scheme
(WENO), 173
weighting function, 178
whistle, 84
whistler-nozzle, 84
whistling, 24, 77, 84, 88, 384
white noise, 153, 208, 323, 365
wideband continuum, 321, 322
Wiener-Hopf method, 79, 82
wiggle, 317
wiggle detection, 111
wind rotor, 383
wind tunnel, 203, 204, 205
wind turbine, 3, 8, 174, 378, 381
window, 291, 357
windshield wiper, 278
wing, 203
wing mirror, 154, 360, 362, 363, 366, 367
wing span, 302
wiperblade, 357, 360
zero-equation turbulence model, 263
zero-order truncation error, 180
zero-pressure gradient boundary layer, 229
zonal approach, 212
zonal hybrid method, 148
18:15

LES For Acoustics by Wei Shyy and Michael J. Rycroft

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

LES For Acoustics by Wei Shyy and Michael J. Rycroft

Загружено:

Авторское право:

Доступные форматы

P1: JZZ

This page intentionally left blank

November 22, 2006

November 22, 2006

LARGE-EDDY SIMULATION FOR ACOUSTICS

also part-time Professor at Ecole

November 22, 2006

November 22, 2006

Cambridge Aerospace Series

J. M. Rolfe and K. J. Staples (eds.): Flight Simulation

November 22, 2006

Large-Eddy Simulation for

November 22, 2006

CAMBRIDGE UNIVERSITY PRESS

Cambridge, New York, Melbourne, Madrid, Cape Town, Singapore, So Paulo

November 22, 2006

List of Figures and Tables

1.2 Introduction to computational aeroacoustics

1.3 State of the art: LES for acoustics

1.3.1 Broadband noise prediction in general

Theoretical Background: Aeroacoustics . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

November 22, 2006

Approximations and alternative forms of the basic

Theoretical Background: Large-Eddy Simulation . . . . . . . . . . . . . . . . . . . . 89

November 22, 2006

3.5 Extension of subgrid models for the compressible case

Use of Hybrid RANSLES for Acoustic Source Predictions . . . . . . . . . . . . 128

Numerical Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

Introduction to discretization schemes

5.2 Boundary conditions for LES

Outflow boundary conditions

5.3 Boundary conditions: Acoustics

November 22, 2006

5.3.4 Perfectly matched layers

5.4 Some concepts of LESCAA coupling

and Roland Ewert

5.4.1 LES inflow boundary

Applications and Results of Large-Eddy Simulations for Acoustics . . . . . 238

Introduction to jet acoustics

6.3 Cavity noise

Introduction to cavity noise

6.4 Aeroelastic noise

Introduction to aeroelastic noise

6.5 Trailing-edge noise

November 22, 2006

6.6 Blunt bodies (cylinder, cars)

Introduction to industrial aeroacoustics analyses

7.1 Governing equations and acoustic analogies

7.2 Numerical errors

7.3 Initial and boundary conditions

November 22, 2006

November 22, 2006

List of Figures and Tables

Magnetic levitation hover train project Transrapid

November 22, 2006

LIST OF FIGURES AND TABLES

Initial startup transient at probe 111 (cylinder rear face) predicted by

November 22, 2006

LIST OF FIGURES AND TABLES

LES of a turbulent boundary layer for Re0 = 1400 and M = 0.4.

November 22, 2006

LIST OF FIGURES AND TABLES

Sketch of the experimental setup

APE source terms L = [