A comprehensive study of detached-eddy

simulation

vorgelegt von

Charles Mockett
M.Eng

Von der Fakultät V – Verkehrs- und Maschinensysteme

der Technischen Universität Berlin
zur Erlangung des akademischen Grades
Doktor der Ingenieurwissenschaften (Dr.-Ing.)
genehmigte Dissertation

Promotionsausschuss:
Vorsitzender: Prof. Dr.-Ing. Wolfgang Nitsche
Berichter: Prof. Dr.-Ing. Frank Thiele
Prof. Mikhael Strelets

Tag der wissenschaftliche Aussprache: 17.04.2009

Berlin 2009
D 83
ISBN (online) 978 3 7983 2159 5
ISBN (print) 978 3 7983 2160 1

∞ Gedruckt auf säurefreiem alterungsbeständigem Papier

Druck/ docupoint GmbH Magdeburg

Printing: Maxim-Gorki-Straße 10, 39108 Magdeburg

Vertrieb/ Universitätsverlag der TU Berlin

Publisher: Universitätsbibliothek
Fasanenstr. 88 ( im VOLKSWAGEN-Haus), D-10623 Berlin
Tel.: (030)314-76131; Fax.: (030)314-76133
E-Mail: [email protected]
http://www.univerlag.tu-berlin.de/
To Silke and Frederick

i
ii
 Abstract
Detached-eddy simulation (DES) is a prominent example of a new family of meth-
ods that seek to bridge the gap existing in terms of computational cost and predic-
tive accuracy between Reynolds-averaged Navier–Stokes (RANS) and large-eddy
simulation (LES) approaches for turbulent flows. As such, DES targets computa-
tional resources expected to become widespread in the near future, and is expected
to be particularly valuable for scenarios in which RANS is known to be unreliable
(e.g. strongly-separated flow) or in which resolution of unsteady turbulent motion is
required (e.g. computational aeroacoustics).

In this work a comprehensive study of DES has been carried out. In the literature
review, a detailed introduction of the method and its various derivative versions is
offered together with a summary of the relevant turbulent flow physics. The presen-
tation and discussion of results draws upon a wide selection of test cases, ranging
between canonical and complex configurations and with a comprehensive variety
of turbulent flow phenomena. The topics covered include elementary implementa-
tion issues and the verification and calibration of model features, the fundamental
interaction of the model with the numerical scheme, the detailed comparison of DES
results with wind tunnel experiments, the dependency of DES on the underlying
RANS model in different scenarios, the sensitivity of DES to the choice of numerical
time step, the importance of long simulation times for reliable statistics, the per-
spectives for a dynamic determination of the model parameter, the performance of a
recent extension of the DES formulation when applied as a wall-modelled LES and a
proposed extension to this that relaxes the stringent requirements on the wall-normal
grid spacing. A concise summary of the principle findings is provided in the conclu-
sion, which includes links to the relevant sub-sections for easy reference.

iii
iv
 Zusammenfassung
Die Detached-Eddy-Simulation (DES) ist ein typischer Vertreter einer neuen Familie
von Methoden, welche in Hinblick auf die Vorhersagegenauigkeit und den numeri-
schen Aufwand eine Brücke zwischen Lösungsverfahren auf Basis Reynolds-gemit-
telter Navier–Stokes-Gleichungen (RANS) und Grobstruktursimulationen (LES) für
die Berechung turbulenter Strömungen bilden. Die DES ist daher so ausgelegt, dass
sich ihr Ressourcenbedarf durch die typischerweise in naher Zukunft weitverbreitet
zur Verfügung stehenen Rechenkapazitäten decken lässt. Besonders geeignet ist ihre
Anwendung dabei in Szenarien, in denen RANS-Modelle bekannterweise unzuver-
lässig sind, wie z.B. stark abgelösten Strömungen, oder wo die Auflösung instatio-
närer turbulenter Strukturen erforderlich ist, etwa bei aeroakustischen Anwendun-
gen.

Im Rahmen dieser Arbeit wurden umfassende Studien zur DES durchgeführt. So

wird im Überblick zum aktuellen Stand der Forschung zunächst die Methode in ih-
ren verschiedenen Versionen detailliert vorgestellt und eine Zusammenfassung der
relevanten strömungsphysikalischen Phänomenen gegeben. Die anschließende Dis-
kussion der Ergebnisse erstreckt sich über eine breite Auswahl von Testfällen, von
vereinfachten bis hin zu komplexen Konfigurationen, um so unsterschiedlichste Tur-
bulenzphänomenen zu erfassen. Dabei umfassen die Untersuchungen verschiede-
ne Themenbereiche, wie elementare implementierungstechnische Fragen, die Veri-
fikation und Kalibrierung von Modellkomponenten, die grundlegende Wechselwir-
kung zwischen Modell und numerischem Schema, detaillierte Vergleiche von DES-
Ergebnissen mit Windkanaldaten, die Modellabhängigkeit der DES in verschiedenen
Situationen, der Einfluss des Zeitschrittes auf die DES, die Wichtigkeit von langen Si-
mulationszeiten, um verläßliche statistische Größen zu gewinnen, die Perspektiven
einer dynamische Bestimmung des Modellparameters, die Bewertung einer kürz-
lich entwickelten Erweiterung der DES–Formulierung als Wandmodellierung für ei-
ne LES, und eine vorgeschlagene Erweiterung, um die strikten Anforderungen an
die wandnormale Gitterauflösung zu lockern. Am Abschluss wird eine kompakte
Zusammenfassung der prinzipiellen Erkentnisse gegeben, wobei Verweise auf die
relevanten Abschnitte ein Nachschlagen erleichtern sollen.

v
vi
Thanks and acknowledgements

I would like to express my profound graditude to the following people, without

whose help this work would not have been possible, or at least not nearly so enjoy-
able:

First and foremost to Prof. Frank Thiele, who has provided a fantastic working envi-
ronment and research facilities, the freedom to follow my own initiative and contin-
ual support and guidance.

To Prof. Mikhael Strelets for his valuable and detailed review of the manuscript as
examiner for the PhD defence. Likewise, to Prof. Wolfgang Nitsche for his role as
chairman of the examining committee.

To the following, who have directly assisted in many of the investigations and in
some cases reviewed sections of the manuscript: Dr. Ulf Bunge, Angelo Carnarius,
Thilo Knacke, Dr. Rodolphe Perrin, Thorsten Reimann and Dr. Jianping Yan.

To Prof. Mikhael Shur and Prof. Mikhael Strelets, and more recently Dr. Philippe
Spalart for the invigorating discussions and open exchange of ideas, which I have
valued tremendously over the years.

To my current and former colleagues at the Institute of Fluid Mechanics and Engi-
neering Acoustics (ISTA), for their companionship, continual assistance and many
fruitful discussions, as well as for the continual stream of interesting reading matter.

To Martin Franke, Lars Oergel and Norbert Schönwald for their invaluable IT sup-
port (and, indeed, patience).

For helping with the “necessary evils” of the administrative world: Gabrielle Grün-
dig-Hänszke, Lilly Lindemann and Edelgard Schröder.

For the generous provision of data and pictures, and the permission to reproduce
them: Cambridge University Press, Dr. Sergio Hoyas, Prof. Tee Tai Lim, Prof. Chris-
toffer Norberg, Dr. Philippe Spalart, Dr. Katrina Swalwell and Dr. Jianping Yan.

To the following members of the CFD turbulence modelling community for their
very important input in the form of many memorable and lively discussions and
informal coaching: Jonas Ask, Dr. Bertrand Aupoix, Dr. Sofiane Benhamadouche,
Dr. Marianna Braza, Prof. Lars Davidson, Dr. Sébastien Deck, Prof. Tom Gatski,
Dr. Werner Haase, Prof. Charles Hirsch, Dr. Yannick Hoarau, Dr. Suad Jakirlic, Dr. To-
bias Knopp, Dr. Johan Kok, Dr. Robin Langtry, Prof. Dominique Laurence, Prof.

vii
Michael Leschziner, Dr. Ivan Mary, Dr. Florian Menter, Prof. Christoffer Norberg,
Prof. Shia-Hui Peng, Dr. Alistair Revell, Prof. Thomas Rung, Dr. Dieter Schwamborn,
Dr. Lionel Temmerman, Dr. Dominic von Terzi, Dr. Fabrizio Tessicini, Dr. Harmen
van der Ven, Dr. Stefan Wallin, Dr. Keith Weinman, Dr. Don Wu.

Finally, on the home front I am deeply grateful to my family and friends, both in
England and in Berlin, for the tireless support and patience, and for politely putting
up with increasingly abstract responses to the question “So how is work going?”.

* * *

This work was undertaken in the framework of the European FLOMANIA1 and
DESider2 projects, both of which were excellently coordinated by Dr. Werner Haase.
Many of the more time consuming computations were conducted on the IBM pSeries
690 at the Zuse-Institut Berlin (ZIB) and the Norddeutschen Verbund für Hoch- und Höch-
stleistungsrechnen (HLRN)3 , with valuable technical support provided by Dr. Wolf-
gang Baumann. This substantial support is gratefully acknowledged.

1
FLOMANIA (Flow Physics Modelling – An Integrated Approach) was funded by the European Commu-
nity represented by the CEC, Research Directorate General, in the 5th Framework Program, under Con-
tract No. G4RD-CT2001-00613. See http://cfd.mace.manchester.ac.uk/flomania/index.html and
Haase et al., 2006 [55] for more details.
2
DESider (Detached-Eddy Simulation for Industrial Aerodynamics) was funded by the European Commu-
nity represented by the CEC, Research Directorate General, in the 6th Framework Program, under Con-
tract No. AST3-CT-2003-502842. See http://cfd.mace.manchester.ac.uk/desider/index.html and
Haase et al., 2009 [56] for more details.
3
http://www.hlrn.de

viii
Contents

I State of the art and literature review 1

1 Introduction 3
1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Outline of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2 Physical background 7
2.1 Laminar to turbulent transition . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 The turbulent energy cascade . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2.1 Kolmogorov’s hypotheses . . . . . . . . . . . . . . . . . . . . . . 9
2.3 The turbulence energy spectrum . . . . . . . . . . . . . . . . . . . . . . 11
2.3.1 Important length scales . . . . . . . . . . . . . . . . . . . . . . . 13
2.4 Wall-bounded turbulence . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4.1 Universality of boundary-layer profiles . . . . . . . . . . . . . . 16
2.4.2 Empirical correlations for boundary layer flows . . . . . . . . . 18
2.5 Bluff body flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.1 Vortex shedding . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.5.2 Circular cylinder wake flow regimes . . . . . . . . . . . . . . . . 23

3 Simulation and modelling of fluid turbulence 29

3.1 Numerical simulation of fluid flow . . . . . . . . . . . . . . . . . . . . . 29
3.2 Overview of computational approaches to turbulent flow . . . . . . . . 31
3.3 Direct numerical simulation (DNS) . . . . . . . . . . . . . . . . . . . . . 33
3.4 Statistical modelling of turbulence (RANS) . . . . . . . . . . . . . . . . 34
3.4.1 Reynolds-averaged Navier–Stokes equations . . . . . . . . . . . 35
3.4.2 The Boussinesq eddy viscosity hypothesis and Reynolds stress
models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.4.3 Modelling hierarchy . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.4.4 Wall boundary conditions . . . . . . . . . . . . . . . . . . . . . . 39
3.4.5 Unsteady RANS . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.5 Partial resolution of turbulence (LES) . . . . . . . . . . . . . . . . . . . . 40
3.5.1 Modelling of the unresolved turbulence . . . . . . . . . . . . . . 42
3.5.2 Requirements on the numerical scheme and resolution . . . . . 45
3.6 Hybrid RANS-LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
3.6.1 Nomenclature for filtering and averaging . . . . . . . . . . . . . 50

ix
Contents

3.7 Detached-eddy simulation . . . . . . . . . . . . . . . . . . . . . . . . . . 50

3.7.1 Philosophy of the method: “The ‘D’ in DES” . . . . . . . . . . . 50
3.7.2 The original DES formulation: DES97 . . . . . . . . . . . . . . . 53
3.7.3 Problems with the original formulation . . . . . . . . . . . . . . 55
3.7.4 Enhanced versions of DES . . . . . . . . . . . . . . . . . . . . . . 58
3.7.5 Incorporation of WMLES functionality . . . . . . . . . . . . . . 60
3.7.6 Other modifications to the DES formulation . . . . . . . . . . . 64

II Methodology 69

4 Numerical flow solver 71

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2.1 Diffusion term . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.2 Convection term . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.2.3 Time discretisation . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.3 Pressure–velocity coupling . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.4 Boundary conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.5 Solution of the linear equation system . . . . . . . . . . . . . . . . . . . 77

5 Implementation of detached-eddy simulation 79

5.1 Preceding work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.2 RANS models applied . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.3 Overview of implemented versions . . . . . . . . . . . . . . . . . . . . . 79
5.4 Hybrid numerical convection scheme . . . . . . . . . . . . . . . . . . . 81
5.5 Model-specific implementation details . . . . . . . . . . . . . . . . . . . 82
5.5.1 Calibration of the DES model parameter CDES . . . . . . . . . . 82
5.5.2 Ψ functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.5.3 Alternative DES length scale substitutions . . . . . . . . . . . . 83
5.5.4 Specific features of the CEASM-DES implementation . . . . . . 83
5.5.5 IDDES implementation . . . . . . . . . . . . . . . . . . . . . . . 85

6 Description of test cases 87

6.1 Decay of isotropic turbulence . . . . . . . . . . . . . . . . . . . . . . . . 87
6.2 Fully developed turbulent channel flow . . . . . . . . . . . . . . . . . . 90
6.3 NACA0012 airfoil in deep stall . . . . . . . . . . . . . . . . . . . . . . . 95
6.4 NACA0021 airfoil in deep stall . . . . . . . . . . . . . . . . . . . . . . . 97
6.5 Circular cylinder in a square channel . . . . . . . . . . . . . . . . . . . . 98
6.6 Separating-reattaching flow over a bump in a rectangular duct . . . . . 104

III Validation, demonstration and analysis 111

x
 Contents

7 Validation of the implementation 113

7.1 Influence of the numerical convection scheme . . . . . . . . . . . . . . . 113
7.2 Functionality of the hybrid convection scheme . . . . . . . . . . . . . . 115
7.3 Calibration of the CDES parameter . . . . . . . . . . . . . . . . . . . . . 116
7.4 Validation of the Ψ functions . . . . . . . . . . . . . . . . . . . . . . . . 118
7.5 Validation of boundary layer shield function for CEASM-DES . . . . . 121

8 Demonstration and Analysis 125

8.1 RANS and LES activity in DES . . . . . . . . . . . . . . . . . . . . . . . 125
8.1.1 Instantaneous snapshot of the flow field . . . . . . . . . . . . . . 125
8.1.2 Distribution of the RANS and LES zones . . . . . . . . . . . . . 126
8.2 Variation of CDES between models and the scope for a “dynamic DES” 130
8.3 Alternative length scale substitutions in DES . . . . . . . . . . . . . . . 135
8.4 Sensitivity of DES to the RANS model used . . . . . . . . . . . . . . . . 140
8.4.1 Massively-separated external flow: NACA0021 at α = 60◦ . . . 141
8.4.2 Separating-reattaching internal flow: Bump in a rectangular duct147
8.5 Comparison of DES with experiment for the flow around a circular
cylinder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.5.1 Transition behaviour . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.5.2 Instantaneous flow . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.5.3 Time-averaged flow . . . . . . . . . . . . . . . . . . . . . . . . . 151
8.5.4 Spectral content . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155
8.5.5 Phase-averaged flow field . . . . . . . . . . . . . . . . . . . . . . 160
8.6 Time step sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
8.7 Comparison of DES with experiment for the bump test case . . . . . . 166
8.8 Investigation of wall-modelled LES of channel flow . . . . . . . . . . . 168
8.8.1 Comparison of DES97 and IDDES . . . . . . . . . . . . . . . . . 168
8.8.2 Examination of IDDES functionality . . . . . . . . . . . . . . . . 172
8.8.3 Numerical cost of WMLES and grid considerations . . . . . . . 180
8.8.4 Combination with the hybrid-adaptive boundary condition . . 181
8.9 IDDES application to complex flow . . . . . . . . . . . . . . . . . . . . . 186

9 Conclusion 191
9.1 Summary of findings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
9.2 Closing comments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194

IV Appendices 195

A The governing equations of fluid motion 197

A.1 Conservation of mass . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
A.2 Conservation of momentum . . . . . . . . . . . . . . . . . . . . . . . . . 197

B RANS models 199

xi
Contents

B.1 One-equation models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199

B.1.1 Spalart–Allmaras . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
B.1.2 Edwards modification . . . . . . . . . . . . . . . . . . . . . . . . 200
B.1.3 SALSA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
B.2 Two-equation models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
B.2.1 Wilcox k − ω . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
B.2.2 LLR k − ω . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
B.2.3 LL k − ε . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
B.3 CEASM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204

C Hybrid-adaptive wall boundary condition 205

D Derivation of model-specific Ψ functions 209

D.1 Spalart–Allmaras model without “trip terms” . . . . . . . . . . . . . . . 209
D.2 Spalart–Allmaras model with Edwards modification . . . . . . . . . . . 210
D.3 SALSA model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
D.4 Wilcox k − ω model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
D.5 LLR k − ω model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
D.6 LL k − ε model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
D.7 CEASM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215

E Obtaining resolved and modelled turbulent statistics 217

E.1 Method for RANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
E.2 Method for LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
E.3 Method for DES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

F Phase averaging of quasi-periodic turbulent flows 219

G Figures for the bump test case 221

xii
xiii
List of symbols and abbreviations

Upper-case Roman
A coefficient of Smagorinsky-form of DES implementation
B intercept constant of log law of the wall
C constant from Pope’s model spectrum
CD coefficient of drag, integrated over entire span
CDES DES model constant
Cd coefficient of drag, integrated over single spanwise section
CL coefficient of lift, integrated over entire span
Cl coefficient of lift, integrated over single spanwise section
Cl IDDES model-specific constant for f l
C 0 large scale constant from Pope’s model spectrum
Cp pressure coefficient
C pb base pressure coefficient on the rear surface of a bluff body
CS Smagorinsky constant
Ct IDDES model-specific constant for f t
Cw near-wall constant for IDDES Δ function
Cη dissipative scale constant from Pope’s model spectrum
Cμ turbulent viscosity constant in k − ε model
D cylinder diameter
D diffusion term of transport model
D wake width of bluff body
E log law parameter E = eκB
E (κ ) energy spectrum function
LDDES DDES length scale
L̃DDES modified DDES length scale from IDDES
LDES DES length scale
LIDDES IDDES length scale (combining DDES and WMLES functions)
LLES LES length scale
LRANS RANS length scale
LWMLES length scale of hybrid WMLES model
M Mach number
Re Reynolds number
Recrit critical Reynolds number
P production term of transport equation
Sij strain rate tensor
Sφ source term of φ
S∗ rate of strain invariant
St Strouhal number
St∗ Roshko’s universal Strouhal number for bluff bodies
Tu free stream turbulence intensity
Ui time or ensemble averaged velocity vector

xiv
 +
U normalised mean velocity (magnitude) in wall units
Uδ mean velocity at channel centreline
Uτ friction velocity
Ui  phase-averaged velocity vector
i
U filtered (resolved) velocity vector
Ui filtered or averaged velocity vector (in hybrid RANS-LES)

Lower-case Roman
c airfoil chord length
dw wall-normal distance
f frequency
f amp solution-dependent magnitude function for f restore
fd DDES boundary layer shield function
f˜d DDES RANS-LES blending function for IDDES
f dt DDES modelled log-law region shield function for IDDES
f hill grid-dependent shape function for f restore
fl blending function for viscous sublayer from f amp
f 0 large scale function from Pope’s model spectrum
f restore IDDES RANS length scale boosting function
f step IDDES hybrid weighting function of RANS-LES blending
ft blending function for modelled log-law region from f amp
fη dissipative scale function from Pope’s model spectrum
h grid spacing
hmax maximum grid cell edge length
hwn grid cell edge length in wall-normal direction
k turbulent kinetic energy
k base pressure parameter
 length scale
0 length scale of the largest eddies
n normal component of xi
p̃ pressure
p modified pressure
p0 large scale constant from Pope’s model spectrum
rd modelled turbulent boundary layer sensor from DDES
rdl IDDES viscous sublayer sensor from f l
rdt IDDES modelled log-law region sensor from f t
t time
u x-component of velocity
ui velocity vector
ũ i periodic fluctuation of velocity vector
u i fluctuation of velocity vector
u i incoherent fluctuation of velocity vector
u i u j Reynolds stress tensor

xv
 ũ i ũ j coherent part of Reynolds stress tensor
u i u j incoherent part of Reynolds stress tensor
u i u j  phase-averaged turbulent stresses
u  u
i j sub-grid stress tensor
u mean convective velocity averaged over outflow plane
|us | velocity magnitude of free streamline at separation
u∞ free stream x-component of velocity
|u∞ | free stream magnitude of velocity
uη Kolmogorov velocity scale
v y-component of velocity
w z-component of velocity
xi position
y+ normalised wall distance in wall units
y+hyb y+ from hybrid-adaptive wall boundary condition

Upper-case Greek
Γφ diffusivity coefficient of φ
Δ LES filter width
Δt time step size
Δx grid cell spacing in x-direction
Δf fixed filter width
Ψ function to prevent wall-term activation in LES-mode
Ψ body forces
Ωij rate of rotation tensor
Ω∗ rate of rotation invariant

Lower-case Greek
α angle of attack
α IDDES near-wall grid function from f step
β dissipative scale constant from Pope’s model spectrum
δ channel half-height
δij Kronecker delta
 dissipation term in a transport equation
ε rate of dissipation of k
η Kolmogorov length scale
θ effective viscosity to wall-normal distance parameter of hybrid-adaptive
wall boundary condition
κ von Kármán constant
κ wavenumber
λ2 vortex core criterion
μ dynamic viscosity

xvi
 ν kinematic viscosity
νt turbulent/eddy (kinematic) viscosity
ν̃ SA model eddy viscosity
ρ density
σ hybrid blending function for convection scheme
τ period of a harmonic oscillation
τij stress tensor
τw wall shear stress
τη Kolmogorov time scale
φ generic conserved intensive flow quantity
φij pressure – rate of strain redistribution tensor
ϕ phase angle
ϕ blending function for hybrid-adaptive wall boundary condition
ω turbulent frequency

Superscripts

+ quantity normalised in frictional wall units (w.r.t. Uτ )

Subscripts

c quantity based on airfoil chord length

re f reference value of a quantity
sgs sub-grid scale quantity
w value at a solid boundary (wall)
0 largest turbulent scales
∞ free stream value at infinite distance from configuration

Symbols

x ensemble averaged x
x phase averaged x
x filtered x
|x| magnitude of a vector xi
x fluctuation of x
x̃ coherent, quasi-periodic fluctuation of x
x  incoherent, random fluctuation of x

xvii
Abbreviations
2C two-component
3C three-component
CEASM compact explicit algebraic Reynolds stress model
CFD computational fluid dynamics
DDES delayed DES
DES detached-eddy simulation
DES97 the original DES formulation
DNS direct numerical simulation
ELAN the CFD solver: “Elliptic Analysis of the Navier–Stokes equations”
FD finite difference
FE finite element
FV finite volume
GIS grid-induced separation
IDDES improved DDES
L laminar vortex shedding regime
LES large-eddy simulation
LL Lien–Leschziner k − ε model
LLM log-layer mismatch
LLR linear local realisable k − ω model
MSD modelled stress depletion
PIV particle image velocimetry
RANS Reynolds-averaged Navier–Stokes
RMS root mean square
SA Spalart–Allmaras model
SAE Spalart–Allmaras model with Edwards correction
SALSA strain-adaptive linear Spalart–Allmaras model
SAS scale-adaptive simulation
SGS sub-grid scale model
SST shear stress transport model
TrBL transition-in-boundary-layer vortex shedding regime
TRPIV time-resolved PIV
TrSL transition-in-shear-layer vortex shedding regime
TrW transition-in-wake vortex shedding regime
URANS unsteady RANS
VLES very-large-eddy simulation
WCX Wilcox k − ω model
WMLES wall modelled LES
X-LES extra-large-eddy simulation

xviii
 Part I

State of the art and literature review

1
1 Introduction

1.1 Background
Fluid flow is an important phenomenon in many aspects of the natural sciences and
engineering. From aerodynamics to meteorology, chemical processing to medical ap-
plications, fluid dynamics plays a central role. Following the pioneering theoretical
and experimental work of the last centuries, the application of computer simulation
has in recent times established itself as a tool to further understanding and to aid
the solution of practical problems. With the rapid increase in computing power and
the improvement of simulation algorithms, the use of computational fluid dynamics
(CFD) for practical applications is clearly set for continued growth. Despite this opti-
mistic outlook, significant barriers to accuracy and efficiency remain to be overcome
before CFD can be considered a truly everyday tool.

Turbulent flows have long presented a considerable obstacle to the accuracy, appli-
cability and popularity of CFD in industrial application. Although direct numerical
simulation (DNS) allows the resolution of the complete range of turbulent scales, the
computational cost renders this approach unfeasible for many decades to come. At
the other end of the spectrum, the Reynolds-averaged Navier–Stokes (RANS) equa-
tions significantly reduce the computational cost whilst introducing uncertainties
due to modelling approximations. Large-eddy simulation (LES) lies conceptually be-
tween the extremes of pure turbulence simulation (DNS) and pure turbulence mod-
elling (RANS), whereby only the largest energy-containing turbulent motions are
resolved, with the influence of the smallest unresolved scales provided by a model.
Although offering a reduced computational cost compared to DNS, LES still proves
too expensive for most practical applications. This expense is particularly high in at-
tached turbulent boundary layers, as the turbulence scales are very small compared
to the geometry in question. It is therefore widely acknowledged that a gap exists be-
tween RANS and LES, and recent activity in the development of hybrid RANS-LES
methods is aimed at addressing this. One such method, which perhaps shows the
greatest promise for widespread practical application, is detached-eddy simulation
(DES).

1.2 Objectives
The global objective to which this work contributes has been outlined in the previous
section: The problem presented by turbulent flow prediction, which strongly limits
the reliability and applicability of CFD in engineering. The study focuses on DES,
which seeks to blend the level of direct turbulence resolution and modelling empiri-
cism to an optimum level targeted at the computing capability expected in the near

3
1 Introduction

future. As such, DES and hybrid RANS-LES methods in general attempt to provide
a bridge between the current industrial workhorse of RANS models and pure LES
methods that will not become affordable until the mid-term to distant future. The
work forms part of wider efforts to enhance the reliability and applicability of DES.

Specific objectives of the thesis work are formulated as follows:

• To review available DES methods and to establish their relationship to alterna-
tive hybrid RANS-LES approaches.

• To implement the selected methods in the in-house CFD solver “ELAN” and
to validate these implementations using carefully selected test cases and pro-
cedures.

• To document the experience gained in this implementation and validation pro-

cess, which should provide a useful reference to readers wishing to implement
DES methods themselves.

• To assess the level of predictive accuracy that can be achieved using DES in a
range of situations representative of diverse target applications.

• To identify any problems encountered and to investigate potential remedies.

• To document remaining open issues and suggest corresponding future work

based on these findings.
Acknowledging that the suitability of various turbulence prediction methods de-
pends on the flow phenomena concerned, considerable value is placed on the consid-
eration of a wide range of test cases, which exhibit contrasting and complimentary
flow physics. It is with this aim in mind that the adjective “comprehensive” is used
in the title. Such an approach is intended to provide the greatest possible degree
of generality of the conclusions drawn. The thesis also targets completeness in the
sense that all phenomenological arguments are founded on concepts described in the
literature review, rather than relying too strongly on external references.

Finally, the ultimate goal of the methods examined is considered to be their applica-
tion to engineering problems. This principle implicitly defines the criteria for success
or failure in the interpretation of results. This furthermore serves to eliminate any
modifications from consideration that would require excessive levels of user input
or that are fundamentally limited to simple geometries.

1.3 Outline of the thesis

An overview of the structure of the thesis will be given to assist orientation. The
thesis is organised into three parts plus a collection of appendices. Part I contains the
introduction and literature review. The goal of the literature review is to summarise

4
 1.4 Conventions

all physical and numerical background information to be drawn upon in the later
discussions. As such, the literature review is organised into two chapters. Chapter 2
provides physical descriptions of the relevant turbulent flow phenomena, whereas
chapter 3 describes existing numerical methods for the simulation and modelling of
turbulent flow, with a particular emphasis on the DES approach.

Part II describes the methodology of the investigations, beginning in chapter 4 with

a concise description of the ELAN CFD solver applied. In chapter 5, details concern-
ing the DES implementation are given and in chapter 6 the selection of test cases is
presented. Each case is described in detail, from the characteristics of the underlying
flow phenomena to the numerical details and post processing techniques.

The results are presented and discussed concurrently in Part III, which consists of a
multi-faceted collection of investigations. These are roughly grouped into two chap-
ters: In chapter 7, investigations targeted at the validation of key implementation
features are collated, which provide a foundation for the demonstration and analysis
of chapter 8. Each of the investigations is presented in a self-contained manner, with
conclusions drawn for each. These conclusions are collated concisely in chapter 9,
with explicit references to the more detailed discussions in preceding chapters. A
particularly important outcome of this work is the collection of directions identified
for further study, which are also collated in chapter 9.

To enhance the readability, all derivations above a certain level of complexity are
placed in the appendices. These also contain a collection of topics that can be consid-
ered non-essential background information, which some readers may however find
useful.

1.4 Conventions
Certain nomenclature conventions have been adopted throughout the document,
which are summarised here. Where feasible, tensor notation with the Einstein sum-
mation convention will be used in equations. When used, vectors are identified with
bold type. A capital C with a subscript will be reserved for quantities that are truly
constant, e.g. Cμ , whereas a lower-case f with subscript will be applied for varying
parameters or functions, e.g. f Cμ .

5
1 Introduction

6
2 Physical background

The phenomenon of fluid turbulence introduces one of the greatest challenges facing
CFD. A range of methods are available to incorporate turbulence effects in fluid flow
simulations, each with their respective strengths and weaknesses. Before these are
reviewed in Chapter 3, a summary of the characteristics of turbulent flow relevant to
these methods will be given, which naturally does not constitute an exhaustive treat-
ment. For this, the interested reader is referred to the extensive available literature
(e.g. [169], [61], [116]). Furthermore, an attempt will be made to convey an under-
standing of turbulent phenomena on a physical level, restricting the presentation of
mathematical relationships to those necessary for discussion and analysis in the fol-
lowing chapters.

Bradshaw’s definition of turbulence [14] makes an ideal introduction to this section,

giving a preview of many key aspects:
“Turbulence is a three dimensional time dependent motion in which vor-
tex stretching causes velocity fluctuations to spread to all wavelengths
between a minimum determined by viscous forces and a maximum de-
termined by the boundary conditions. It is the usual state of fluid motion
except at low Reynolds numbers.”
The important role that turbulence plays in science, engineering and nature (out-
lined in Chapter 1) is highlighted by the recognition that turbulent flow is the rule
and laminar flow the exception. The quotation also identifies turbulent motion to
be three-dimensional and unsteady in nature. The fluctuations in the velocity field
are apparent in the form of overlapping eddies over a wide range of scales, a state of
affairs brought about by the phenomenon of vortex stretching. Analysis of the vortic-
ity equation (obtained by applying the curl operator to the Navier–Stokes equations)
reveals that the intensity of a line vortex increases when it is stretched. This vortex
stretching term is inactive in the two-dimensional form of the equations, hence the
statement that turbulence is three-dimensional. The sheer and obvious complexity of
turbulence has long fascinated and baffled scientists, not just through its nature, but
also the fact that such a random state can arise from the deterministic Navier–Stokes
equations. The apparent futility of the struggle to describe, explain and understand
turbulence motivated the following lament by Horace Lamb [79]:
“I am an old man now, and when I die and go to heaven there are two
matters on which I hope for enlightenment. One is quantum electrody-
namics, and the other is the turbulent motion of fluids. And about the
former I am rather optimistic.”

7
2 Physical background

2.1 Laminar to turbulent transition

Nonetheless, much progress has been made in finding “laws” or expressions of recur-
ring patterns of behaviour. The pioneering work on analysing and predicting the oc-
currence of a laminar or turbulent flow state was carried out by Osborne Reynolds in
the late 19th century [119]. Reynolds observed the flow of fluids through glass pipes,
using injected dye for visualisation. Initially the thread of dye flowed smoothly,
however after a certain distance the thread was seen to develop regular waves fol-
lowed by a transition to fully-turbulent flow, which rapidly mixed the dye with the
surrounding fluid. An everyday example of this transition from laminar to turbu-
lent flow is the behaviour of a buoyant plume of cigarette smoke. Figure 2.1 shows
a cigarette placed in a low-turbulence wind tunnel that provides a smooth cross
flow [113]. The initially laminar smoke plume develops regular vortex rings, which
then degrade into a disordered turbulent state.

Figure 2.1: Laminar to turbulent transition illustrated with an everyday example [113] (repro-
duced with permission).

A central parameter in the study of turbulent flows bears Reynolds’ name, originat-
ing from the pipe flow study outlined above [119]. The Reynolds number is the ratio
of inertial forces to viscous forces acting on the fluid and can be expressed as
ure f re f
Re = , (2.1)
ν
where ure f and re f are reference values of velocity and length particular to the flow
in question and ν = μ/ρ is the kinematic viscosity. Laminar flow occurs at low val-
ues and turbulent flow at high values of Re.

The transition to turbulence can occur via a variety of different mechanisms, and is
highly sensitive to disturbances such as free stream fluctuations or geometric imper-
fections. Because of this at most a range of typical values for the critical Reynolds
number can be given. In practical applications, predicting the precise location of
boundary layer transition is both highly important and exceedingly challenging,

8
 2.2 The turbulent energy cascade

with many different methods currently under development. Transition prediction

is generally considered a separate discipline to the modelling of turbulence itself,
and is therefore not considered in this work.

2.2 The turbulent energy cascade

A principal concept in the understanding of fluid turbulence is the energy cascade.
As identified in the introductory paragraphs of this chapter, turbulence consists of
overlapping three-dimensional eddies of different scales. The majority of the kinetic
energy is contained in the largest eddies, the length scale 0 of which is of the order of
magnitude of the geometry or flow feature in question (e.g. bluff body wake, bound-
ary layer thickness). These large eddies draw their energy from the velocity gradients
in the mean flow, and are anisotropic and strongly affected by the mean flow topol-
ogy. A cascade process then ensues driven by vortex stretching, in which energy
is continually passed to successively smaller scales. The energy of the smallest ed-
dies is then dissipated to heat by the action of molecular viscosity. This fundamental
concept has been lyrically expressed by Lewis Richardson [123] as follows:
“Big whorls have little whorls,
Which feed on their velocity;
And little whorls have lesser whorls,
And so on to viscosity.”

2.2.1 Kolmogorov’s hypotheses

More specific details of the turbulent cascade emerge from Kolmogorov’s theory [77],
which is expressed as three hypotheses. These provide important information such
as the size of the smallest scales and the characteristics of the turbulence across the
length scales. The first of the hypotheses describes the nature of the small-scale mo-
tions.

Kolmogorov’s hypothesis of local isotropy

At sufficiently high Reynolds number, the small-scale turbulent motions
(  0 ) are statistically isotropic.

The directional information of the anisotropic large scales is therefore lost along the
turbulent cascade, resulting in locally-isotropic small scales. An important outcome
of this is that the small motions have a universal character, which implies that these
may be modelled more simply than the large scales.

Kolmogorov’s first similarity hypothesis:

In every turbulent flow at sufficiently high Reynolds number, the statis-
tics of the small-scale motions have a universal form that is uniquely de-
termined by ν and ε.

9
2 Physical background

The variables upon which this universal form depends arise from considerations of
the dominant physical processes. The small-scale motions are considered to oper-
ate like a conveyer belt, transferring the energy received from the larger scales to be
dissipated by the viscosity. This rate of dissipation, ε, is therefore one of the deter-
mining parameters together with ν. The range of scales at which this process occurs
is referred to as the universal equilibrium range.

0
10

-1
η / l0 , uη / u0 , τη / τ0

10

-2
10

η / l0
-3
10 uη / u0
τη / τ0

-4
10 0 1 2 3 4 5
10 10 10 10 10 10
Re

Figure 2.2: Ratios of smallest to largest turbulent scales as functions of Re.

The very smallest eddies are characterised by the Kolmogorov length, velocity and
time scales given by

η ≡ (ν3 /ε)1/4 , (2.2)

u η ≡ (ε ν)1/4 , (2.3)

τη ≡ (ν/ε)1/2 , (2.4)
respectively. An important feature that justifies the premise that the Kolmogorov
scales correspond to the smallest turbulent motions is the Reynolds number based
on these scales, u η η/ν, which has a value of unity.

It can furthermore be observed that the smallest scales become smaller (compared to
the largest scales, 0 ) as the Reynolds number of the flow increases. The ratios of the
Kolmogorov scales to the largest scales as a function of Reynolds number are given
in Eqs.( 2.5 – 2.7) and depicted graphically in Fig. 2.2.

10
 2.3 The turbulence energy spectrum

η/0 ∝ Re−3/4 , (2.5)

u η /u0 ∝ Re−1/4 , (2.6)

τη /τ0 ∝ Re−1/2 . (2.7)

There exists therefore at high Reynolds number an intermediate range of scales that
are large enough not to be affected by viscosity ( η) yet small compared to the
energy-containing large scales and not affected by the boundary conditions of the
flow (  0 ). As a consequence of this and of the first similarity hypothesis:

Kolomogorov’s second similarity hypothesis:

In every turbulent flow at sufficiently high Reynolds number, the statis-

tics of the motions of scale  in the range 0  η have a universal
form that is uniquely determined by ε, independent of ν.

The universal equilibrium range is therefore subdivided into two subranges, the dis-
sipation range, where  approaches η and viscous effects dominate, and the larger iner-
tial subrange where viscous effects are negligible. The scale range of the largest eddies
is referred to as the energy-containing range. Pope [116] suggests that the boundary be-
tween the dissipation and inertial subranges is located at 60η. The energy-containing
range is suggested to occupy 61 0 <  < 60 , such that the boundary between the
energy-containing and universal equilibrium ranges is located at 16 0 . A schematic
of the energy flow and regions of the turbulence cascade is given in Fig. 2.3

2.3 The turbulence energy spectrum

To examine the distribution of energy over different scales, the energy spectrum func-
tion of wavenumber E (κ ) is useful (the wavenumber of a particular length scale is
given by κ = 2π/). As a consequence of the second Kolmogorov hypothesis, it
is found that E (κ ) ∝ κ −5/3 in the inertial subrange [116]. For isotropic turbulence,
the energy spectrum function has a universal form, depending only on the Reynolds
number of the flow.

Pope [116] proposed a model spectrum, which returns E (κ ) for a given length scale of
the large eddies, 0 = k3/2 /ε and value of ε. The expression for the model spectrum
is

11
2 Physical background

Production from Dissipation to heat

macroscopic velocity via molecular viscosity
gradients

Transfer of energy to smaller

scales

Universal equilibrium range

Energy−containing Inertial subrange Dissipation range
range
Anisotropic
Determined by boundary
conditions Determined by ε only Determined by ε and ν
Non−universal character Universal character Universal character

6l 0 l0 1 6
/ l0 60η η
Large Small

Figure 2.3: Schematic diagram of the energy flow and regions of the turbulent energy cascade.

E (κ ) = C ε2/3 κ −5/3 f 0 (κ 0 ) f η (κη ) ,

⎛ ⎞5/3+ p0
κ 0
f  0 ( κ 0 ) = ⎝  ⎠ f η (κη ) = e− β{[(κη ) +Cη ] −Cη } ,
4 4 1/4
, (2.8)
( κ  0 ) 2 + C 0

where the constant values are C = 1.5, C0 = 6.78, p0 = 2, β = 5.2 and Cη = 0.4.
The Reynolds number based on the large scales is defined as Re0 = k1/2 0 /ν, and
examples of the model spectrum for various values of Re0 and 0 = ε = 1 are
shown in Fig. 2.4. The increasing separation of large and small scales with increas-
ing Reynolds number can be clearly seen, as can the slope of exponent −5/3 for the
inertial subrange. The physical interpretation of the low wavenumber region (scales
larger than the energy-containing range) is of a merging of large vortices, in contrast
to the stretching and break up of smaller vortices in the energy cascade at higher
wavenumber [169].

The various ranges of the turbulent cascade are annotated on the model spectra of
Fig. 2.5 for two Reynolds numbers. The values of the Kolmogorov length scale η
and the borders between the different regions are taken from the definitions in Sec-
tion 2.2.1. The inertial subrange is seen to exist only at the higher Reynolds number,
being eliminated by an overlap of the energy-containing range and dissipation range
at lower Reynolds number.

12
 2.3 The turbulence energy spectrum
0
10

-1
10

-2
10
E ∝ κ −5/3
-3
E(κ) 10

-4
10
Rel =1000
0
-5
10
Rel =10000
0
-6
10 Rel =100000
0

-7
10 0 1 2 3 4
10 10 10 10 10
κ

Figure 2.4: Model energy spectra E (κ ) [116] for varying turbulence Reynolds number Re0 . Val-
ues of 0 = 1 and ε = 1 are assumed.

Universal equilibrium range

Dissipation range
E(κ)

E(κ)

Energy−containing
range

Energy−containing
Dissipation range range

Inertial
subrange

κ 2π/η κ 2π/η

(a) Re 0 = 1000 (b) Re 0 = 100000

Figure 2.5: Model energy spectra E (κ ) [116] annotated with the regions of the turbulence en-
ergy cascade. Values of 0 = 1 and ε = 1 are assumed.

2.3.1 Important length scales

The length scale of the largest eddies, 0 , and the Kolmogorov length scale, η, have
been introduced in the above sections. Some further length scales are of importance
to later material, which shall be introduced briefly. For a more detailed account of
the derivations and physical interpretations of these, the reader is referred to the text-
book of Pope [116].

13
2 Physical background

Two important length scales are derived from the streamwise and lateral autocorre-
lation functions f (r, t) and g(r, t), the former of which in isotropic turbulence com-
pletely determines the two-point correlation Rij (r, t). The first of these is the longi-
tudinal integral scale

∞
L11 (t) ≡ f (r, t)dr , (2.9)
0
which is simply the area under the curve of f (r, t) and is characteristic of the larger
eddies.

The longitudinal and lateral Taylor microscales are somewhat more difficult to define
intuitively. They are geometrically obtained from the curvature of the longitudinal
and transverse autocorrelation functions at r = 0, describing the intersection of an
osculating parabola with the axis f = 0 or g = 0. The longitudinal and lateral Taylor
microscales are defined as
  −1/2

 2u2rms 1 

λf =   and λ g = − g (0, t) . (2.10)
∂u 2 2
1
∂x1
√
It can be shown that in isotropic turbulence λ g = λ f / 2.

The length scale characterising the largest eddies can be defined in terms of k and ε
as

k3/2
L≡ , (2.11)
ε
and some relationships between the various length scales can be determined analyt-
ically as

λg √
= 10Re− L
1/2
, (2.12)
L
√
λ g = 10η 2/3 L1/3 . (2.13)
At high Reynolds number therefore, λ g lies between η and L. The Taylor microscale
is often used to characterise the Reynolds number of isotropic turbulence:
 1/2
urms λ g 20
Reλ ≡ and Reλ = ReL . (2.14)
ν 3

14
 2.4 Wall-bounded turbulence

2.4 Wall-bounded turbulence

The cascade process and properties of the range of scales in turbulent flow intro-
duced above have profound consequences as far as the treatment of turbulence in
CFD is concerned. However, analysis has so far been limited to the simplest con-
ceivable occurrence of turbulence, that of statistically isotropic (and hence homoge-
neous) turbulence. Although describing important aspects (isotropy is the state to
which turbulent flows return in the absence of external forcing), the effects of solid
boundaries must be included in a complete discussion. It has been mentioned in
Section 2.2 that the larger, energy-containing scales are strongly affected by the flow
topology and geometry. For the majority of engineering applications, the proxim-
ity of solid boundaries (referred to as walls) is the most important mechanism by
which anisotropy occurs in turbulence. The properties of wall-bounded turbulent
flows (relevant to both internal flows such as pipes and external flows such as airfoil
boundary layers) will be summarised here.

The principal effect of wall proximity on turbulence is a damping of the fluid fluctu-
ations in the direction normal to the wall. The energy of these fluctuations is redis-
tributed to the directions tangential to the wall, which are hence amplified. As such,
turbulence far away from the wall is relatively isotropic, and becomes increasingly
anisotropic nearer to the wall. Indeed an alternative description is that the turbu-
lence is strongly three-dimensional away from the wall, and becomes increasingly
two-dimensional nearer to the wall. The fluid in direct contact with the wall exhibits
zero relative velocity to the boundary (the “no-slip” condition, u w = vw = ww = 0),
and the damping effect on the wall-normal turbulent fluctuations naturally increases
with proximity to the wall. As a consequence, the turbulent motions nearest to the
wall are eliminated entirely, and molecular viscosity becomes the dominant energy
dissipation mechanism. Correspondingly, turbulent boundary layers are charac-
terised by distinct zones arranged tangentially to the wall within which different
physical phenomena predominate.

There are three key canonical wall-bounded flows, namely that along an un-inclined
flat plate and the flow through planar channels and circular pipes. The flat plate
boundary layer is characterised by zero pressure gradient and a continually thick-
ening boundary layer in the streamwise (x) direction. The domain in the spanwise
(z) direction is considered to be homogeneous, so that the mean streamwise velocity
profile in the wall-normal (y) direction U (y) is a function of the streamwise location
only. The early boundary layer is laminar, whereas at some downstream position a
transition region to turbulence occurs. As outlined in Sect. 2.1, the transition position
is characterised by the Reynolds number. The local Reynolds number for a flat plate
is defined as

U∞ x
Rex = , (2.15)
ν

15
2 Physical background

where U ∞ is the mean uniform on-flow velocity and x is the distance downstream
of the beginning of the plate. The critical Reynolds number is quoted by Massey [87]
as lying in the range 3 × 105 < Rex < 5 × 105 , whereas Pope [116] gives a value of
Rex ≈ 106 . The precise value varies strongly from case to case, depending for exam-
ple on such parameters as free-stream turbulence or surface roughness. More detail
on boundary layer transition is given by Schlichting [143].

By contrast, the internal wall-bounded flows (pipes and channels) require a negative
pressure gradient in the streamwise direction to overcome the wall shear stresses
caused by the boundary layer. As is the case for the flat plate, the boundary layer
begins in a laminar state followed (at a sufficiently high Reynolds number) by transi-
tion to turbulence. The boundary layer at each wall grows continually until meeting
the boundary layer edge from the opposing wall at the central line of a pipe or plane
of a channel. Following this, the velocity profile remains fixed, and is hence referred
to as fully-developed. In the fully-developed state, the channel or pipe velocity profiles
U (y) are hence independent of the streamwise direction in addition to the spanwise
or radial directions.

2.4.1 Universality of boundary-layer profiles

The outermost regions of the turbulent boundary layers of a flat plate and fully-
developed channel can be expected to differ, as the former approaches a laminar
region of uniform velocity whereas the latter approaches the edge of an opposing
turbulent boundary layer. Apart from this distinction however, it emerges that the
boundary layer profiles of zero-pressure gradient flat plates and fully-developed
channels exhibit a universal behaviour described by the “law of the wall”1

The basis of the law of the wall is the the renormalisation of flow and geometric
quantities in terms of the wall shear stress, τw . Such scaled quantities are referred
to as “wall friction units” and denoted by the superscript + . The mean wall friction
velocity is defined as

τw
Uτ = , (2.16)
ρ
and is used to normalise the mean streamwise velocity and wall-normal distance as
follows:
U τ dw + U
y+ =, U = . (2.17)
ν Uτ
A wall friction Reynolds number, Reτ , can be formulated as
1
The law of the wall is a much-debated concept, and many researchers question the existence of a true
universal character (see e.g. Bradshaw & Huang [16] for a review of this issue). Nonetheless, the law
of the wall can be said to apply at least approximately, and forms the cornerstone of many modelling
approaches.

16
 2.4 Wall-bounded turbulence

Uτ δ
Reτ =
, (2.18)
ν
where δ represents the boundary layer thickness (corresponding to the channel half
height in fully-developed flow).

As mentioned in the introductory passages, a series of regions exist across the bound-
ary layer, each characterised by different physical phenomena. A description of these
zones and their approximate extents in terms of y+ and y/δ is given by Pope [116],
which is summarised below.

Inner layer (y/δ < 0.1): The mean velocity profile is determined by U τ and y+ ,
independent of U δ and δ.

Viscous wall region (y+ < 50): The effect of molecular viscosity is significant.

Viscous sublayer (y+ < 5): The Reynolds shear stress is negligible in comparison
to the viscous shear stress. The velocity profile is linear, given by
+
U = y+ . (2.19)

Outer layer (y+ > 50): The effect of molecular viscosity on the mean velocity profile
is negligible.

Overlap region (y+ > 50, y/δ < 0.1): The region where the inner and outer layers
overlap, which only occurs at high Reynolds number.

Log-law region (y+ > 30, y/δ < 0.3): The “log-law” applies:
1 +
ln y+ + B .
U = (2.20)
κ
The constant κ, known as the von Kármán constant, describes the exponent of the
profile and B its intercept. The values of κ and B are obtained through curve-fitting
to experimental or DNS data, so some scatter is apparent. Widely-accepted values
are B = 5.2 and κ = 0.41.

Buffer layer (5 < y+ < 30): The transitional region between the viscous sublayer
and the log-law region.

To illustrate the relationship of these zones to the mean velocity profile, Fig. 2.6
+
shows the profile U (y+ ) obtained at Reτ = 2000 from direct numerical simulation
by Hoyas and Jiménez [63, 67]2 , upon which the zones are depicted.
2
Data used with permission.

17
2 Physical background

30

inner layer
25
overlap
viscous wall region region
20

buffer
layer
+

15
U

viscous sublayer
outer layer
10
log−law region

5 DNS Hoyas et al., Reτ = 2000

+ -1 +
U = κ ln y + B
0
0 1 2 3 4
10 10 10 10 10
+
y

Figure 2.6: Mean velocity profile for a fully-developed turbulent channel flow at Re τ = 2003
computed using DNS by Hoyas and Jiménez [63, 67] compared to the log-law pro-
file of Eq. 2.20. The physical zones are plotted as given by Pope [116].

2.4.2 Empirical correlations for boundary layer flows

For reasons that will become apparent in Chapter 3, direct numerical simulation is
considered the most reliable source of benchmark data for turbulent channel flow,
but is limited to low Reynolds numbers due to computational expense. To provide
engineers with information on the higher Reynolds numbers of practical relevance,
experimental measurements are therefore indispensable. A number of mathematical
expressions have been proposed that correlate principal characteristics of the flow
with known input parameters. These have been calibrated to experimental mea-
surements and are hence only as accurate as the data upon which they are based.
Such correlations are employed in this study to give an indication of the accuracy of
simulations above the DNS Reynolds numbers and hence represent distillations of
experimental data.

Reichardt [118] proposed a correlation describing the inner layer velocity profile in
wall units. Reichardt’s correlation is given by
 
+ 1   + y+ −y+ /3
U = ln 1 + κy+ + 7.8 1 − e−y /11 − e , (2.21)
κ 11
which is plotted in comparison to the DNS profile of Hoyas and Jiménez in Fig. 2.7. It
can be seen that the Reichardt correlation, although describing the viscous sublayer
and buffer layer well, predicts a log-law region with an excessive intercept value B.

18
 2.4 Wall-bounded turbulence

Reichardt also formulated a wake term for the outer layer, which is not employed
here.

30
DNS Hoyas et al., Reτ = 2000 24
25 Reichardt
+ + 22
U =y
+ -1 +
20 U = κ ln y + B
20
+

+
15
U

U
18

10 16

5 14

0 12
0 1 2 3 4 2 3
10 10 10 10 10 10 10
+ +
y y

Figure 2.7: Comparison of the normalised mean velocity profiles between the DNS of Hoyas
and Jiménez [63, 67] and the Reichardt correlation of Eq. 2.21. The right hand figure
shows a zoom of the log law region.

Although Reτ is the dependent parameter for the self-similarity, a more accessible
quantity is a Reynolds number based on the mean bulk velocity U b , where

2δ
1
Ub = U dy . (2.22)
2δ
0
The bulk Reynolds number is hence given by
U b 2δ
. Reb = (2.23)
ν
A relationship between Reτ and the bulk Reynolds number, Reb has been suggested
for channel flows by Pope [116] as

Reτ ≈ 0.09 Re0.88

b . (2.24)
An approximate relationship between the mean centreline velocity (at y = δ) and U τ
is also given by Pope:
Uδ
≈ 5 log10 Reb . (2.25)
Uτ
Further useful correlations based on the bulk Reynolds number have been formu-
lated for channel flow by Dean [34]. These are based on an extensive review of the
experimental data available at the time (1978) and claim validity within the range
6 × 103 ≤ Reb ≤ 6 × 105 . The skin friction coefficient is correlated to Reb by

19
2 Physical background

C f ≈ 0.073 Re−
b
1/4
, (2.26)
where C f is defined as follows:
τw
Cf = 2
. (2.27)
1 ρ Ub
2
The ratio between the centreline and bulk velocities is furthermore proposed by Dean
to follow
Uδ
≈ 1.28 Re−b
0.0116
. (2.28)
Ub
The agreement of Dean’s correlation for C f with more recent DNS data is demon-
strated in Fig. 2.8 to be largely satisfactory. The C f values returned by Dean’s corre-
lation consistently exceed those from the DNS however.

0.01
Dean
DNS: Moser et al.
DNS: Hoyas et al.
0.008
Cf

0.006

0.004

0.002 4 4 4 4
0 2×10 4×10 6×10 8×10 1×10
5

Reb

Figure 2.8: Comparison of Dean’s correlation for C f (Reb ) [34], Eq. 2.26, with more recent DNS
data of Moser et al. [100] (at Re τ = 180, 395, 590) and Hoyas et al. [63] (at Reτ =
180, 550, 950, 2000).

It is essential to close the description of the law of the wall and the empirical correla-
tions with a remark as to their validity. The law of the wall is only valid for boundary
layers in which a weak streamwise pressure gradient is present, allowing the neglec-
tion of history effects and the reduction to a simple one-dimensional system. In prac-
tice however, strong distortions of the boundary layer profile can be caused by phe-
nomena such as adverse pressure gradients (particularly arising across shock waves),
surface curvature, separation and re-attachment and three-dimensional skewing ef-

20
 2.5 Bluff body flows

fects (occurring for example on swept wings). The law of the wall and by extension
any modelling approaches reliant upon it would prove to be highly inaccurate in
such situations.

2.5 Bluff body flows

An important class of flows, particularly relevant to many of the test cases studied
in this work and to detached-eddy simulation in general, are bluff body flows. The
term “bluff” is essentially intuitive in nature, referring more to flow characteristics
than it does to geometric features. As a result, it is hard to define precisely what is
meant by a bluff body, although it is perhaps helpful to consider that a streamlined
body constitutes its opposite.

For flows at all but very low Reynolds numbers, bluff bodies are characterised by
strong flow separation and generally unsteady wakes. Furthermore, bluff body flows
exhibit a negligible contribution of the skin friction drag in comparison to the pres-
sure drag. For flows around streamlined bodies however, the skin friction drag plays
an important role. Examples of bluff body flows in industrial applications include
buildings, undersea oil pipelines, heavy goods vehicles and wings or rotor blades in
deep stall.

Due to the opportunities that these offer for simplified two-dimensional analysis,
prismatic bluff bodies have long been the focus of much study – none more so than
circular cylinders. As remarked by Roshko [129], the wake structure of different bluff
bodies is similar in nature. The circular cylinder is therefore an important canonical
flow, many of the characteristics of which can be translated to other bluff bodies.

A brief overview of the characteristics of bluff body flows relevant to this work
will be presented. The emphasis is therefore on the behaviour near to the critical
Reynolds number (introduced in Section 2.5.2) as well as on a selection of influenc-
ing parameters such as blockage, aspect ratio and free stream turbulence.

2.5.1 Vortex shedding

Over a wide range of Reynolds numbers (detailed in the case of a cylinder in Sec-
tion 2.5.2) the wake behind many bluff bodies is characterised by the periodic and
asymmetric shedding of large scale vortices. This phenomenon is more prevalent for
high aspect ratio or prismatic bluff bodies, but by no means limited to these. An ex-
perimental visualisation of the vortex shedding behind a circular cylinder is shown
in Fig. 2.9. It can be seen that both the wake width and the size of the shed vortices
are of the order of the cylinder diameter. This phenomenon was first analysed theo-
retically by von Kármán in 1912 [176], and is as a result known as the von Kármán
vortex street.

21
2 Physical background

Figure 2.9: Smoke-wire visualisation of the vortex shedding behind a circular cylinder at Re =
8000 [103] (reproduced with permission).

The alternating shedding of vortices from the upper and lower surfaces of the bluff
body gives rise to a strongly fluctuating lift force. Before von Kármán’s analysis, the
frequency of this fluctuation had been studied by Strouhal [165] and was found to
depend on the free stream velocity and cylinder diameter3 . The non-dimensional
frequency, or Strouhal number, is defined as
fD
St =
, (2.29)
|u∞ |
where f is the dominant frequency, D is the cylinder diameter and | u ∞ | is the free
stream velocity magnitude. For bluff bodies other than the circular cylinder, D is to
be substituted by the frontal height of the bluff body. In particular, for an inclined
flat plate or airfoil, c sin α is to be used (with c the chord length and α the angle of
incidence).

A further step forward in the analysis of bluff body wake flow was achieved by
Roshko [129], who demonstrated that the centreline pressure coefficient on the rear
surface of a bluff body (known as the “base pressure”, C pb ) arises due to the vortex
formation. When the region of this formation is shifted further downstream (for
example through the use of a splitter plate), an increase of the base pressure and
corresponding decrease of the drag is observed. This is furthermore accompanied
by a reduction in the shedding frequency. One outcome of Roshko’s analysis was
the derivation of a “universal Strouhal number” St∗ for bluff bodies, based upon the
velocity magnitude of the free streamline at boundary layer separation, | u s | and the
wake width D  :
f D
St∗ =. (2.30)
|us |
As | u s | is a difficult quantity to establish, a formulation employing the base pres-
3
In fact, it emerged that the value of St is not constant but dependent on parameters such as the Reynolds
number. The Strouhal number remains however a useful non-dimensionalised frequency.

22
 2.5 Bluff body flows

sure parameter k = 1 − C pb is more convenient. This relates to the conventional
Strouhal number as follows:
St D 
St∗ =
. (2.31)
k D
Roshko established a value of St∗ ≈ 0.16, which was demonstrated to hold for a
range of bluff body geometries and over a wide range of Re.

This analysis was further extended by Abernathy [1] for sharp-edged flat plates over
a wide range of angles of incidence. It was observed experimentally that the wake
width could be expressed as D  = 1.41c sin α, giving
f 1.41 c sin α
St∗ = . (2.32)
k |u∞ |
The value St∗ ≈ 0.15 was found to hold for 0.35◦ ≤ α ≤ 90◦ .

Another outcome of Roshko’s work was some more insight into the concept of “bluff-
ness”, namely:

• For two bodies with the same frontal area, the bluffer diverges the flow more
strongly creating a larger wake and experiencing a higher drag.

• Bluffer cylinders have a lower shedding frequency, which is inversely propor-

tional to the wake width4 .

2.5.2 Circular cylinder wake flow regimes

The wake flow behind a circular cylinder shows a high sensitivity to the Reynolds
number, passing through a number of distinct regimes as the Reynolds number is
varied. An excellent introduction and review of this behaviour is given by Zdrav-
kovich [189], a summary of which is presented here. At very low Reynolds numbers
the flow remains attached around the entire cylinder (“creeping” flow), separating to
a steady attached recirculation region at slightly higher Re. As Re is increased further
the von Kármán instability gives rise to the alternating vortex shedding described in
the previous section. The formation of the vortex street is a laminar phenomenon up
to Re ≈ 200. The entire laminar flow regime is denoted with the abbreviation L by
Zdravkovich.

As Re is increased further, transition to turbulence takes place – at first in the far wake
(denoted TrW), then moving forward to the free shear layers (TrSL) and finally reach-
ing the cylinder boundary layers (TrBL). The behaviour of the Strouhal number and
mean drag coefficient over a wide range of Re was reported by Schewe [142] and is re-
produced in Fig. 2.10. The rapid drop in CD and increase in St at Re ≈ 3 × 105 marks
4
With reference to the work of Fage & Johansen [41].

23
2 Physical background

Figure 2.10: Variation of Strouhal number (above) and mean drag coefficient, CD , (below) of
a circular cylinder with Reynolds number. From Schewe [142] (reproduced with
permission).

the onset of turbulence in the attached boundary layers. This causes a dramatic rear-
ward shift of the separation point and a corresponding narrowing of the wake. This
phenomenon is referred to as the “drag crisis” and the Reynolds number at which
it occurs the critical Reynolds number, Recrit . The TrSL regimes at Re < Recrit are
correspondingly referred to as subcritical.

The TrSL regime begins with the formation of instability waves in the shear layer,
which were identified and analysed by Bloor [11]. At higher Re, these form small-
scale eddies with a size of the order of the shear layer thickness. At still higher
Re, occasional bursts of incoherent turbulence are observed in the shear layer. The
transition waves and turbulent bursts are considered analogous to the Tollmien–
Schlichting waves and turbulent spots [143] observed in boundary layer transition.

In contrast to the Strouhal number increase, the drop in drag coefficient begins grad-

24
 2.5 Bluff body flows

ually from Re ≈ 1 × 105 as the TrBL regime is entered. This gradual decrease is
caused by sporadic occurrences of boundary layer turbulence, varying from one vor-
tex shedding cycle to the next. Humphreys [64] furthermore observed the formation
of a cellular pattern of the separation position in the spanwise direction, which was
stable in time. This was shown to be caused by alternating laminar and turbulent
separation along the span, with a cell size of roughly 1.4D to 1.7D. The steep region
of the drag crisis is characterised by two discontinuous jumps, labelled A and B in
Fig. 2.10. The first of these represents the establishment of a single laminar separa-
tion bubble on one side of the cylinder, giving rise to an asymmetric flow condition
discovered by Bearman [7]. The second jump is caused by the formation of a bubble
on the opposite side of the cylinder and the re-establishment of symmetry.

Abbreviation: Description: Re range:

L1 Creeping flow 0 < Re < (4-5)
L2 Steady separation regime (4-5) < Re < (30-48)
L3 Periodic laminar regime (30-48) < Re < (180-200)
TrW Transition-in-wake regime (180-200) < Re < (350-400)
TrSL1 Transition waves in shear layers (350-400) < Re < (1k-2k)
TrSL2 Transition eddies in shear layers (1k-2k) < Re < (20k-40k)
TrSL3 Burst to turbulence (20k-40k) < Re < (100k-200k )
TrBL0 Precritical regime (100k-200k ) < Re < (300k-340k )
TrBL1 one-bubble regime (300k-340k ) < Re < (380k-400k )
TrBL2 two-bubble regime (380k-400k ) < Re < (500k-1M)
TrBL3 supercritical regime (500k-1M) < Re < (3.4M-6M)
TrBL4 post-critical regime Re > (3.5M-6M)

Table 2.1: Overview of vortex shedding regimes for circular cylinders [189].

The vortex shedding regimes can be broken down into sub-regimes describing the
development of the vortex shedding behaviour. A summary of the shedding regimes
is given together with their abbreviations and Reynolds number range according to
Zdravkovich [189] in Tab. 2.1. Much of the complexity of this behaviour arises due to
the wide range of positions for the flow separation on the smooth cylinder surface,
and is hence not necessarily typical for all bluff bodies. A contrasting example is that
of a normal flat plate, where a constant CD was reported by Abernathy [1]5 across
the range 6000 < Re < 600000.

Particularly around the critical Reynolds number, the flow exhibits a very strong
5
With reference to experiments by Flachsbart [45]

25
2 Physical background

sensitivity to a number of influencing parameters. A selection of these relevant to

this work will be summarised in the following.

Influence of free stream turbulence

The flow around circular cylinders has been shown to exhibit a strong sensitivity to
the turbulence properties of the oncoming free stream flow. This is a central issue
concerning the comparison of wind tunnel experimental data both with industrial
applications (usually involving high levels of free stream turbulence) and with sim-
ulations (typically with a steady laminar on-flow). A review and discussion of this
issue for bluff bodies in general has been published by Bearman & Morel [8], and a
later set of detailed experiments for cylinder flow was conducted by Norberg [102].

The interactions of free stream turbulence with the flow around bluff bodies involve
three principal mechanisms: enhanced mixing and entrainment by the wake, distor-
tion of the oncoming turbulence itself by the bluff body and accelerated transition
to turbulence. Of these, the latter is of the greatest relevance to this work. For sub-
critical cylinder flows, the position of transition in the free shear layer can be shifted
upstream by the presence of free stream turbulence. To fully characterise the free
stream turbulence (assumed to be isotropic far upstream of the bluff body), informa-
tion on both the turbulence intensity and the integral length scale is required. The
integral length scale must be comparable with the characteristic scales of the flow,
such as the cylinder diameter and shear layer thickness, in order to have an effect.

Figure 2.11: Variation of the mean base pressure coefficient, C pb , with Reynolds number for
various values of free stream turbulence intensity, Tu [102] (reproduced with per-
mission).

26
 2.5 Bluff body flows

Figure 2.11 shows the effect of different levels of free stream turbulence intensity Tu 6
(obtained using upstream turbulence-generating grids) on the base pressure coeffi-
cient over the range 10 < Re < 106 [102]. A clear reduction of the critical Reynolds
number by more than 100000 is seen as Tu is increased from Tu = 0.1% to Tu = 1.4%.
A mild increase in the drag coefficient (i.e. reduction of C pb ) is furthermore seen in
the subcritical regime.

Influence of blockage coefficient

Wind tunnel blockage is caused by the constraining effects of the test section walls on
the object tested and can be an important influence parameter for bluff body flows.
The blockage coefficient is defined as the ratio of the bluff body frontal height (e.g.
cylinder diameter) to the test section height. The effect of this parameter on the flow
around a circular cylinder was studied by Richter et al. [124], whose results for the
mean drag coefficient variation with Reynolds number are reproduced in Fig. 2.12.

Increased blockage coefficient is seen to lead to a reduction of the critical Reynolds

number as well as an increase in the drag coefficient at subcritical Re, qualitatively
the same effect as that of free stream turbulence discussed above. The increase of
subcritical drag coefficient is however much stronger than is the case for free stream
turbulence. The effect on the critical Reynolds number can be explained by the ac-
celeration of the shear layer flow leading to earlier transition to turbulence.

A further effect of wind tunnel blockage is an increase in the shedding frequency. An

interesting observation of Richter et al. [124] is that the universal Strouhal number of
Roshko, Eq. (2.31), remains valid despite the effects of confinement.

Influence of aspect ratio

The effect of aspect ratio was studied experimentally by Szepessy & Bearman [168]
using end plates with varying spanwise separation for aspect ratio values between
0.25 and 12 and across the Reynolds number range 8000 to 140000. The greatest sen-
sitivity to aspect ratio was seen for the root-mean-square of the lift fluctuation, which
increased strongly for decreasing aspect ratio. This was explained by the observed
accompanying increase in spanwise correlation of the vortex shedding. Additionally,
a reduction of Recrit was measured for low aspect ratio cases. A convergence of the
measured quantities was seen for aspect ratio values in excess of around five for high
subcritical Reynolds numbers.

This study also has important implications for the three-dimensionality of the flow.
Low frequency modulation of the force fluctuations is commonly observed for bluff
bodies around the high subcritical Reynolds number range (e.g. [64, 167]). This be-
haviour is characterised by a temporary reduction of the vortex shedding strength,

6
Tu = 2
3k / |U |

27
2 Physical background

Figure 2.12: Effect of blockage coefficient on the mean drag coefficient of a circular cylinder
around the critical Reynolds number range. From Richter et al. [124] (reproduced
with permission).

which occurs fairly randomly approximately every 10 to 20 vortex shedding cy-

cles. The occurrence of these “weak shedding cycles” was observed by Szepessy
& Bearman [168] to be strongly reduced as the end plates were moved closer to-
gether, giving rise to the interpretation that these are associated with sudden three-
dimensionality of the wake.

28
3 Simulation and modelling of fluid turbulence

3.1 Numerical simulation of fluid flow

The Navier–Stokes equations provide the fundamental description of fluid motion,

and are based solely on the assumption that the fluid can be considered as a contin-
uum of differentiable variables. The derivation and characteristics of the governing
equations of fluid motion are summarised in Appendix A. Since all of the presented
applications concern incompressible flows of Newtonian fluids, the incompressible
Navier–Stokes equations are considered, resulting in considerable simplification of
their formulation and implementation. In conservation form, the continuity and mo-
mentum equations are
∂u i
= 0 and (3.1)
∂xi
Du i 1 ∂p ∂
=− + (2νSij ) , (3.2)
Dt ρ ∂xi ∂x j
respectively. u i is the velocity vector, ρ is the density, p is the modified pressure (see
Section A.2), ν is the kinematic viscosity and Sij is the strain rate tensor.

The analytical solution of the Navier–Stokes equations is only possible for a very
confined number of flows with very simple boundary conditions, or with the inclu-
sion of further assumptions and approximations. When analytical solution is not
possible, the most appropriate method is to obtain approximate solutions using nu-
merical procedures. In order to do this, a discretisation method is required, which ap-
proximates the differential equations as a system of algebraic equations that can be
solved computationally. CFD therefore works on the basis of the discretised Navier–
Stokes equations, coupled with a range of algorithms to achieve iterative solutions.
Three principal discretisation methods are commonly used: finite differences, finite
volumes and finite elements. Of these, the most widespread in industrial CFD is the
finite volume method, which also forms the basis of the ELAN CFD solver [184] used
in these investigations.

The starting point of the finite volume discretisation method is the integral form of
the governing equations. The fluid domain is divided completely into a finite num-
ber of tessellating control volumes (i.e. the numerical grid described in the following
paragraphs), with the conservation equations applied to each of these. Once approx-
imations are applied for the integral of the flow variables over the control volume,
the interpolated values at the control volume faces and the integration over these,
an algebraic equation for each control volume is obtained. These equations are ex-

29
3 Simulation and modelling of fluid turbulence

pressed in terms of the values of the fluid variables at the control volume centre and
those of the neighbouring control volumes. The reasons for the popularity of the fi-
nite volume method are probably the inherent conservation of the scheme together
with the intuitive nature of the quantities to be approximated. The particulars of the
finite volume scheme employed in the ELAN solver used in this work are outlined in
Chapter 4. For a detailed review of finite volume discretisation methods, the reader
is referred to the textbook of Ferziger and Perić [44].

The discretised system of equations must be solved iteratively on a spatial domain

consisting of discrete points corresponding to the fluid region of the problem un-
der investigation. The design of this computational grid is very important, having
a strong impact on the accuracy and stability of the solution. Regions with strong
gradients in the flow variables such as boundary layers or shear layers must be re-
solved using a sufficient number of points, where the number required depends on
the order of accuracy of the method (higher-order methods requiring less points).
Furthermore, certain constraints may exist on the shape of grid cells, with sharp in-
ternal angles (sheared cells) or high aspect ratios causing instabilities or errors in the
numerical solution. As a result of these factors, grid generation for complex geome-
tries can be a very time consuming process and is considered a specialism in itself.

The main classification of meshes in CFD concerns the manner in which individual
points are related to their neighbours to form grid volumes. In a structured grid,
every non-boundary point has a constant number of neighbours, which are found
by the code using a system of index directions. As such, structured grids consist
of rectangles in two dimensions or hexahedra in three dimensions. To be able to
capture complex geometries, multiple regions or blocks can be combined, each of
which must contain a constant number of points in each index direction. Structured
codes benefit from simplified implementation and higher numerical efficiency, as the
neighbours of each point are implicitly known and the solution matrix has a regular
diagonal structure. Furthermore, in regions of parallel flow (e.g. boundary layers)
structured grids can be aligned with the flow direction resulting in higher accuracy
of the solution. Hexahedral cells can also be generated with very high aspect ra-
tios, efficiently allowing fine resolution of the strong wall-normal gradients found
in boundary layers. However, the constraints on the grid design can lead to great
difficulty in meshing complex geometries. Furthermore, necessary local refinement
in one flow region often causes refinement in other less important regions, which is
wasteful of computational resources. Very high aspect ratio cells can also give rise to
numerical convergence problems. An example of a structured grid around an airfoil
is shown in Fig. 3.1(a).

The principal alternative methodology is unstructured grid generation. Instead of in-

dices to allocate the neighbours of each grid cell, the neighbours are explicitly listed
in connectivity tables. This has the advantage of allowing arbitrary cell shapes, such
as tetrahedra, prisms and other polyhedra, which in turn allow strong local refine-

30
 3.2 Overview of computational approaches to turbulent flow

(a) structured (b) unstructured

Figure 3.1: Examples of structured and unstructured grids around a NACA0012 airfoil (n.b. the
structured grid is intended for viscous computation with resolution of the boundary
layers, whereas the unstructured grid is intended for inviscid computation).

ment where necessary and coarsening of the grid where acceptable. Unstructured
grids for complex geometries can be rapidly generated using automatic algorithms,
vastly reducing the work load in grid generation, which is the key reason why un-
structured codes are favoured in industrial CFD. However, the storage of connectiv-
ity data and a sparse solution matrix brings an associated computational overhead,
and irregular cell shapes cause a reduction in accuracy for parallel flows such as
boundary layers. As such, a common trade off is to use a hybrid methodology of
hexahedral or prismatic cells in boundary layer regions, and tetrahedra outside of
this. An example of an unstructured grid is shown in Fig. 3.1(b) 1 .

Naturally, the necessary computing resources (RAM and CPU time) are directly link-
ed to the number of points in the numerical grid. For large problems, a solution can
be obtained in parallel by a number of CPUs in a supercomputer or cluster. The prin-
cipal method of parallel computing used in CFD is domain decomposition, whereby
the grid is divided up and distributed among the computing units. An exchange of
data between the boundary cells on neighbouring CPUs is necessary as the solution
progresses, which causes a parallelisation overhead and a demand on the network-
ing speed between CPUs. Nonetheless, a lot can be achieved on relatively low cost
clusters of networked PCs, which has contributed to an increased accessibility of
CFD.

3.2 Overview of computational approaches to turbulent flow

A wide variety of methods exist to represent turbulence and its effects in a simula-
tion, each with characteristic advantages and drawbacks in terms of computational
cost and simulation quality. In principle, it is possible to simulate turbulent flows
1
Courtesy of the German Aerospace Center (Deutsches Zentrum für Luft- und Raumfahrt).

31
3 Simulation and modelling of fluid turbulence

with perfect accuracy, however for practical applications this would require comput-
ing power not expected to exist for many decades to come. Reducing the compu-
tational burden to manageable levels necessitates the introduction of simplifications
and modelling assumptions concerning the turbulence behaviour. These represent
sources of uncertainty and strong differences can therefore arise between alternative
models. As such, the selection of the appropriate method and model presents a con-
siderable demand on the knowledge and experience of the user. Before these issues
are clarified in more detail in the remaining sections of this chapter, an overview of
the principal simulation and modelling methodologies will be given.

RANS LES DNS

100% modelled ~90% resolved 100% resolved
turbulence turbulence turbulence
Ready: 1990 Ready: 2070 Ready: 2080

Increasing computational cost and method accuracy

Figure 3.2: Diagram of the relative computational cost, method accuracy and forecast “readi-
ness” date [153] for the principal turbulence treatment methods.

The principal classes of turbulence method can be arranged, somewhat simplistically,

along a conceptual axis of increasing computational cost and predictive accuracy.
Figure 3.2 provides such a qualitative representation, together with a description of
the underlying methodology. The axis also represents one of decreasing empiricism,
from Reynolds-averaged Navier–Stokes (RANS) methods (Sect. 3.4) in which the en-
tire range of turbulent motions are described by mathematical models through to
direct numerical simulation (DNS, Sect. 3.3), in which all turbulent motions are di-
rectly resolved in space and time with no empirical modelling. Recalling from Chap-
ter. 2 that turbulence is inherently three-dimensional and unsteady in nature, it is
easily recognised that DNS is correspondingly a four-dimensional simulation prob-
lem. By contrast, for statistically steady turbulent flows RANS allows a steady-state
description of the mean flow field. Furthermore, for geometrically two-dimensional
problems (such as the flow around an airfoil profile), a two-dimensional spatial do-
main can be employed in a RANS computation resulting in a dramatic reduction of
computational expense compared to DNS.

Large-eddy simulation (LES, Sect. 3.5) consists of simulating the energy-containing

scales of turbulent motion, whilst the relatively isotropic and universal scales are
modelled. Although therefore lying conceptually between RANS and DNS, its place-
ment very close to DNS in Fig. 3.2 is not accidental; certainly in the vicinity of
solid boundaries the “large eddies” are indeed very small as described in Sect. 2.4.

32
 3.3 Direct numerical simulation (DNS)

Furthermore, the method naturally inherits the requirement of three space dimen-
sions and one in time from DNS. Hence, the numerical cost and scaling of this with
Reynolds number verges upon that of DNS [153].

To relate this overview of computational expense to the realities of engineering appli-

cation, the “readiness date” of each method as estimated by Spalart [153] are quoted2 .
The message is clear: there is no escaping the introduction of modelling empiricism
for complex high Reynolds number applications for the foreseeable future. This fact
has motivated a tremendous effort in the development and validation of more ac-
curate and computationally efficient RANS models (e.g. the European FLOMANIA
project [55]). An observation of particular relevance to this work is that of a con-
siderable gap in the spectrum of turbulence methods for the “near future” between
RANS and LES. This is the motivation for an increasing amount of work on the de-
velopment of hybrid RANS-LES methods for many conceivable target applications,
as outlined in Sect. 3.6.

3.3 Direct numerical simulation (DNS)

In DNS, the entire range of turbulent scales is resolved directly in space and in
time. Consideration of some of the properties of turbulent flows outlined in Chap-
ter 2 helps to explain why DNS is an inherently expensive technique. Firstly, turbu-
lence is three-dimensional in space and unsteady in time, which makes DNS a four-
dimensional problem. Secondly, the shrinking of the Kolmogorov scales relative to
the largest turbulent scales with Reynolds number (i.e. increasing scale separation
described in Section 2.2.1) means that the numerical cost of DNS grows strongly for
higher Reynolds numbers. For isotropic turbulence, a minimum domain size of eight
times the integral length scale and a maximum spatial resolution of Δx ≈ 2.1η are
quoted by Pope [116]3 . As the motion of the smallest eddies must also be temporally
resolved, the numerical time step size must be balanced to the spatial resolution. An
appropriate guideline is that a fluid particle may only traverse a fraction of a grid
cell per time step, as embodied in the Courant–Friedrichs–Lewy (CFL) criterion [31].

These considerations allow a derivation of the numerical cost of DNS for isotropic
turbulence. For this, the required number of grid points and time steps is the domi-
nating factor, which has been derived by Pope [116] to be proportional to 160Re30 or
0.55Re6λ . For non-isotropic flows, this scaling is seen to be even more severe. Pope
furthermore demonstrates that typically well over 99% of the modes lie within the
dissipative range, a fact that emphasises the attractiveness of the large-eddy simu-
lation concept to be described in Section 3.5. As was given in Fig. 3.2, Spalart esti-
2
This forecast is based on the first “grand challenge” (i.e. non-routine) application of each method to a com-
plete transport aircraft or car, and assumes a fivefold increase in computational power every five years.
The description LES in Fig. 3.2 corresponds to Spalart’s designation “QDNS” (quasi-DNS), Sect. 3.5.2.
3
A multiple of the Kolmogorov scale is quoted because this in fact underestimates the dissipative scale size,
Fig. 2.5.

33
3 Simulation and modelling of fluid turbulence

mates the date of the first DNS “grand challenge” application to a complete aircraft
at 2080 [153].

Particular attention is paid to numerical schemes for DNS, with accurate, low-dissi-
pative methods considered essential. In regard to the numerical cost, efficient sch-
emes with massive parallelisation capabilities are furthermore of great importance.
Spectral or pseudo-spectral methods are hence favoured for applications in which
homogeneous directions exist. As a result, DNS is usually applied to flows involv-
ing simple geometries.

Provided that sufficient resolution, an adequate numerical scheme and appropriate

boundary conditions are used, DNS contains no empiricism and can be considered
accurate. As such, despite the numerical cost, DNS proves immensely useful for fun-
damental turbulence research, providing access to statistical quantities all but impos-
sible to obtain experimentally. DNS is also highly important as a source of numerical
databases for the validation and calibration of more economical methods.

Figure 3.3: Instantaneous vorticity contours at a slice normal to the spanwise direction through
the Reτ = 2000 channel DNS computation of Hoyas & Jiménez [63, 67] (reproduced
with permission). Flow is from left to right and only half of the streamwise domain
is shown.

A recent “grand challenge” DNS study is the fully-developed turbulent channel flow
computation at Reτ = 2000 of Hoyas & Jiménez [63, 67]. The computation of this
unprecedentedly high Reynolds number was carried out using a grid of 18 billion
(×109 ) cells and 2048 CPU cores of the Mare Nostrum at Barcelona Supercomputing
Centre. The resolution in the wall-normal direction was adjusted to Δy ≈ 1.5η. The
instantaneous snapshot from this simulation shown in Fig. 3.3 gives a remarkable
impression of both the fine resolution of the DNS as well as the complexity of turbu-
lent flow in general.

3.4 Statistical modelling of turbulence (RANS)

The opposite end of the conceptual scale to DNS involves the statistical description
of turbulence by mathematical models. Based on the Reynolds-averaged Navier–
Stokes (RANS) equations (Sect. 3.4.1), these models assume a steady mean flow field

34
 3.4 Statistical modelling of turbulence (RANS)

with no turbulent motion being resolved directly. Correspondingly, steady-state so-

lutions can be sought and highly anisotropic grids can be used with large tangential
spacing near solid boundaries. Furthermore, there is no requirement to compute
in three spatial dimensions for two-dimensional applications such as airfoil cross-
sections. As a result, the computational costs are several orders of magnitude lower
than DNS, and RANS methods have already found widespread industrial applica-
tion. However, the approximations and modelling assumptions give rise to funda-
mental conceptual flaws in the simpler model categories and high levels of uncer-
tainty in all cases. As such, the simpler models must be carefully calibrated, result-
ing in applicability for only a limited range of physical phenomena. The promise
of a universal model offered by the most complex class of approaches has yet to be
demonstrated, and these are beset with the additional disadvantages of implementa-
tion complexity, computational cost and lack of numerical robustness. These issues
will be discussed at greater detail in the following.

3.4.1 Reynolds-averaged Navier–Stokes equations

The basis of all statistical turbulence models is given by the Reynolds-averaged Na-
vier–Stokes (RANS) equations. These are derived via the technique of Reynolds de-
composition [120], whereby ensemble averaging is applied to the turbulent flow field
quantities, φ( xi , t). For an ergodic process where the ensemble averaged signal at
each spatial point does not vary with time, ensemble averaging is equivalent to time
averaging. For spatially homogeneous fields, ensemble averaging is equivalent to
spatial averaging at each moment in time. For the purposes of this introduction en-
semble averaging will be considered equivalent to time averaging. Reynolds decom-
position hence splits the unsteady turbulent field φ( xi , t) into a mean and fluctuating
part, φ( xi ) and φ ( xi , t) respectively:

φ ( x i , t ) = φ ( x i ) + φ  ( x i , t ). (3.3)
In this sense, Reynolds decomposition can therefore be interpreted as an assumption
that a turbulent flow has a statistically steady state, and that all fluctuations away
from this are due to turbulence4 . This concept is depicted in Fig. 3.4 for an example
velocity component signal at a single location in space.

Applying the Reynolds decomposition to the governing flow equations gives rise to
the Reynolds-averaged Navier–Stokes (RANS) equations. The Reynolds-averaged
continuity and momentum equations for incompressible Newtonian flow are:

4
This is not always a safe assumption, as many flows also exhibit unsteadiness that is non-turbulent in
origin. This issue is discussed in Appendix F, together with a triple decomposition method applicable to
some such situations.

35
3 Simulation and modelling of fluid turbulence

u(t) U u’(t)
u’u’

= + 0
t

0 0
t t

Figure 3.4: Concept of Reynolds averaging illustrated for a velocity component time signal at
a single spatial location, u ( t).

∂U i
=0 (3.4)
∂xi

DU i 1 ∂p 1 ∂τ ij ∂u i u j
=− + − , (3.5)
Dt ρ ∂xi ρ ∂x j ∂x j
where the time-averaged viscous stress tensor is
 
∂U i ∂U j
τ ij = μ + = 2μSij . (3.6)
∂x j ∂xi
These are identical in form to the original equations, except that the time-averaged
quantities φ have replaced the instantaneous quantities φ, and the additional tensor
ρu i u j has emerged. This is the Reynolds stress tensor, which describes the influence of
the turbulent fluctuations on the mean flow field. Unfortunately this represents an
additional unknown in the equation system, leading to a closure problem. Closure
intrinsically requires the introduction of approximations – it is impossible to derive
a closed set of exact equations. RANS models are hence tasked with providing pre-
scriptions for the Reynolds stress tensor in terms of known quantities such as the
mean flow field or geometric parameters.

3.4.2 The Boussinesq eddy viscosity hypothesis and Reynolds stress

models
An important method of approximating the Reynolds stress tensor is based on the
hypothesis that the effects of turbulence are analogous to an increased viscosity. This
is intuitively justifiable when effects such as energy dissipation and increased mass
transport normal to mean flow streamlines are considered. The Boussinesq relation-
ship [13] between the deviatoric Reynolds stresses5 and the mean flow strain embod-
5
The deviatoric Reynolds stress tensor is obtained by subtracting the isotropic stress tensor, i.e. − ρu i uj +
3 ρkδij .
2

36
 3.4 Statistical modelling of turbulence (RANS)

ies this hypothesis and is formulated as

 
2 ∂U i ∂U j
− ρu i u j + ρδij k = μ t + = 2μ t Sij . (3.7)
3 ∂x j ∂xi
Here, μ t = ρνt is known as the eddy viscosity or turbulent viscosity. Note that a math-
ematical analogy is established between the Boussinesq relation and the Newtonian
stress–rate-of-strain relation (A.5). This is appropriate as mean flow strain is the pri-
mary mechanism by which turbulence is produced and sustained.

The class of models that provide a scalar value of the eddy viscosity to close the
RANS equations are hence referred to as linear eddy viscosity models (LEVM). An al-
ternative to this is the direct prescription of each component of the Reynolds stress
tensor, a technique known as Reynolds stress modelling (RSM) or second moment clo-
sure. The Reynolds stress transport equations can be derived by combining the
Navier–Stokes and RANS equations, although this doesn’t circumvent the closure
problem; on the contrary, an additional six transport equations must be solved for the
Reynolds stresses, which introduce a further twenty-two unknowns into the equa-
tion system [133].

Analysis of the Reynolds stress transport equations for simplified flow conditions
such as simple shear and homogeneous compression can be used to assess the va-
lidity of the Boussinesq hypothesis and the LEVM family. A list of fundamental
inadequacies of the LEVM framework is given by Leschziner & Drikakis [80]. For
example, no resolution of normal stress anisotropy is possible; the rigid linkage of
stresses with strain does not account for stress transport by convection or diffusion;
stresses are strongly over-estimated at high strain rates; the response to curvature
strain, normal straining and rotation is incorrect.

3.4.3 Modelling hierarchy

A wide range of RANS modelling categories therefore exist, with different levels
of physical approximation and mathematical complexity, between full differential
Reynolds stress models (DRSM) and LEVM. In an attempt to summarise this model
landscape as effectively as possible, a modelling hierarchy is drawn up in Fig. 3.5. A
series of physical approximations and mathematical model reductions lead to a num-
ber of steps on the path from DRSM to LEVM. All of these model types are capable
of resolving anisotropy of the Reynolds stresses, and the explicit algebraic Reynolds
stress model (EARSM) is a particularly popular formulation. Models of this variety
represent a compromise between the generality, complexity and expense of DRSM
on the one hand, and the physical inadequacy, simplicity and robustness of LEVM
on the other hand. The EARSM framework was first introduced by Pope [115], and
a prominent example is the model of Wallin & Johansson [178].

37
3 Simulation and modelling of fluid turbulence

Differential Reynolds stress models (DRSM)

Anisotropy−resolving models steady turbulence; structural equilibrium

Implicit algebraic Reynolds stress models (ARSM)

linear ϕij; invariants & representation theory; projection method

Explicit algebraic Reynolds stress models (EARSM)

reduction

Compact (cubic) explicit algebraic Reynolds stress models (CEASM)

reduction

Quadratic explicit algebraic Reynolds stress models

linearisation 2 equation models

1 equation models
Linear eddy viscosity models (LEVM)
"half equation" models
algebraic models

Figure 3.5: Hierarchy of RANS models in order of decreasing modelling complexity (top to bot-
tom). Adapted from an unpublished figure by U. Bunge and from Rung, 2000 [132].

At least theoretically, the model hierarchy represents a trend of increasing generality,

physical accuracy, numerical cost and complexity, as well as one of decreasing nu-
merical robustness in the direction LEVM → DRSM. In practice, however, this is not
so simple – it is often reported that a well-formulated LEVM can provide more accu-
rate predictions of certain key turbulent flow phenomena than full DRSM [55, 80].

Within the LEVM family itself, many different methods are used to determine the
eddy viscosity. Dimensional analysis reveals that two characteristic scales are re-
quired to formulate the eddy viscosity, a fact leading to the common two-equation
modelling framework. Transport equations for two characteristic scales are solved,
such as k and ε√(Jones & Launder [71]), k and ω (Wilcox [181, 182]), k and  (Rotta
[130]) or k and k (Menter et al. [91]). Applying the assumption of local equilibrium
(a balance of turbulent production and dissipation) to two-equation models allows
the derivation of one-equation models [131, 42, 89], although not all such models
have been derived in this way. The more recent one-equation models involve a trans-
port equation for the eddy viscosity and the model of Spalart & Allmaras [157] is a
well-known example. This model, like many one-equation models, requires the pre-
scription of a length scale in the form of the wall-normal distance. This effectively
restricts the model’s validity to attached boundary layer flows. Other one-equation
models, such as those of Menter [89] and Fares & Schröder[42] do not suffer from
this limitation, which is achieved by applying the Bradshaw hypothesis [15] that the
turbulent shear stress scales with the turbulent kinetic energy.

38
 3.4 Statistical modelling of turbulence (RANS)

The simplest form of LEVM are known as algebraic or zero-equation eddy viscosity
models, because they specify the eddy viscosity purely from local information with-
out consideration of transport and history effects. Algebraic models are based on the
mixing length theory of Prandtl [117] and von Kármán [175]
 
 dU 
νt = m   , m = κdw , (3.8)
dy 
which is applicable to the log-law region (Section 2.4). To describe the viscous wall
region and boundary layer edge, the van Driest damping function [173] and Kle-
banoff intermittency function [73] are often used, respectively. Algebraic models
such as those of Cebeci & Smith [25] or Baldwin & Lomax [6] have previously found
extensive application in thin boundary layers typical in external aerodynamic appli-
cations. Owing to increasing computing power however, these are seldom still in use.

One interesting intermediate between algebraic and one-equation models are the so-
called half-equation models, of which that of Johnson & King [70] is exemplary. This
model uses an algebraic expression for the wall-normal distribution of eddy viscos-
ity (based on the Cebeci–Smith model), which is scaled using an ordinary differential
equation to describe the streamwise transport of the maximum turbulence energy.
The Johnson–King model has been reported to give excellent results for the predic-
tion of adverse pressure gradient flows and separating boundary layers, as well as
the effect of separation on the mean field [80, 88], in some cases even out-performing
k − ω models.

3.4.4 Wall boundary conditions

Conventionally, two types of wall boundary condition are available for the solution
of the RANS model equations. These exhibit very different requirements on the wall
normal distance of the first grid point and any violation of these leads to a drastic
degeneration in the solution quality6 . This places a very high level of importance on
the design of the numerical grid, and contributes to the excessive human resources
typically spent on this task. Furthermore, these criteria depend strongly on the local
flow quantities, which means that prior knowledge of the solution is required for
correct grid design. In practice this often means that an iterative process is required,
causing further grid generation expenditure.

In order to address these problems, a more general unified boundary condition that
can deliver robust and reliable solutions on arbitrary grids is highly desirable. Such a
generic boundary condition has been proposed by Rung et al. [135], the formulation
of which is presented in detail in Appendix C. Similar and functionally equivalent
boundary conditions have also been proposed by Esch et al. [40] and Knopp [74].

6
For example, an under-prediction of the skin friction coefficient of up to a factor of five has been reported
for airfoil flows when inappropriate wall treatment is applied [145].

39
3 Simulation and modelling of fluid turbulence

3.4.5 Unsteady RANS

Many important flows exist for which strong unsteadiness of a non-turbulent origin
is apparent, which can be broadly classed in two categories:

• Externally-enforced unsteadiness, such as the imposition of a moving geome-

try or time-variant boundary condition.

• Unsteadiness of a non-turbulent nature inherent to the flow and arising from

hydrodynamic instabilities, e.g. the vortex shedding behind a bluff body, shock
oscillation.

In such cases the application of time-averaging for the Reynolds decomposition (and
subsequent steady-state solution of the mean flow) is inappropriate and can lead to a
significant source of error. Instead, an ensemble averaging over a suitable finite time
period is required, such that the non-turbulent unsteadiness is resolved in the mean
flow and the turbulent fluctuations are described by the RANS model. Such a sim-
ulation is most often referred to as unsteady RANS (URANS) although the acronym
TRANS for “transient RANS” is sometimes used. For bluff body flows involving
unsteadiness of the second type identified above, URANS results of an impressive
degree of accuracy have been reported together with a severe degradation of results
for steady-state simulations, e.g. by Durbin [38] and Shur et al. [147].

URANS is however not without its difficulties, which can be significant. The cen-
tral concept of resolving motions larger than the simulation time step and modelling
those smaller than this requires a significant spectral gap between the non-turbulent
unsteadiness and the largest turbulent time scales. Such a large scale separation is
seldom present, in which case the conceptual foundation of URANS becomes murky
and ill-defined. A strong and undesirable sensitivity to the choice of RANS model
is unfortunately a well-established feature of URANS (see e.g. test cases reported in
the FLOMANIA project [55]). A troublesome sensitivity to the spanwise domain ex-
tent has furthermore been reported for two-dimensional bluff bodies resolved three-
dimensionally by Shur et al. [146]. URANS simulations are furthermore saddled with
a higher computational expense than steady-state RANS, with Durbin for example
claiming a factor of two orders of magnitude in his study of separated flows [38].
The precise increase in numerical expense depends of course strongly on the flow in
question and is naturally less than that of large-eddy simulation.

3.5 Partial resolution of turbulence (LES)

The final family of turbulence methods from the overview of Fig. 3.2 is large-eddy
simulation (LES). This is dealt with last, because of the intermediate conceptual space
that this occupies between RANS and DNS, as has been summarised in Section 3.2.
As was outlined, LES inherits the requirement of three-dimensional spatial resolu-
tion and unsteady simulation from DNS. The reduction in numerical expense com-

40
 3.5 Partial resolution of turbulence (LES)

pared to DNS is achieved by the introduction of a measure of empiricism.

The conceptual basis of large-eddy simulation is the spectral separation of larger

turbulent motions from the smaller scales, whereby the former are resolved directly
and the latter approximated by some manner of model. Recalling the description
of the turbulence energy spectrum arising from the Kolmogorov hypotheses from
Section 2.2.1, the smaller scales can be considered statistically isotropic. It therefore
follows that a tenable representation of the small scale motions can be achieved with
a relatively simple model in comparison to a full RANS closure. The scale separation
is formalised by a filtering operation, which can be considered similar to a localised
averaging over a region within the filter width, Δ. Application of such a filter opera-
tor to the flow field variables results in the following decomposition:

( xi , t) + φ ( xi , t) ,
φ( xi , t) = φ (3.9)
where φ is the large-scale filtered or resolved field at each position and time and φ

is the small-scale fluctuation away from the filtered value, which is not resolved. A
schematic illustration of this process is given in Fig. 3.6.

20 u’u’(t)
u(t) U

= +
0 0 0
t t t

Figure 3.6: Conceptual sketch of filtering applied to a fully-resolved velocity signal u.

The filter operator can be explicitly applied, whereby operators such as the Gaus-
sian, top-hat or cut-off filters7 are used with a specified filter width Δ. The filter
width must be at least as large as the local grid spacing, h, and is either uniform
throughout the domain or applied as a constant multiple of h. Alternatively, it is
recognised that the discretisation scheme in itself acts as a filter function, since scales
smaller than h cannot be resolved. This is known as grid filtering or implicit filtering.
Explicit filtering has the advantage that a known filter function is specified, allowing
a controllable separation of the resolved and unresolved scales and the derivation of
models for the unresolved scales in a mathematically rigorous manner. However, the
usual specification of Δ > h can be considered wasteful in terms of the grid resolu-
tion, since turbulence is not resolved right down to the smallest structures possible.
Some explicit filter types can furthermore give rise to implementation difficulties
and additional computational cost. Implicit filtering is by contrast considered more
7
For example, a detailed study of explicit filtering is presented in the PhD thesis of S. Benhamadouche [9].

41
3 Simulation and modelling of fluid turbulence

pragmatic in the sense that it involves no additional implementation and the grid
resolution is fully-exploited. However, the exact filter behaviour is not known and
varies depending on the numerical discretisation scheme. It is furthermore the case
that the smallest resolved eddies are represented with decreasing accuracy as their
size approaches the grid spacing (dependent on the order of accuracy of the numer-
ical scheme).

When the filtering decomposition of (3.9) is applied to the Navier–Stokes equations,

the resulting filtered equations for the large-scale motion are of the same form as the
Reynolds-averaged Navier–Stokes equations (3.4, 3.5), despite the fundamental dif-
ference in the averaging applied. Due to the prevalence of implicit filtering in early
LES work, the unknown term describing the action of the unresolved scales on the
resolved scales is widely referred to as the sub-grid stress tensor and is denoted − ρu
 u .
i j

3.5.1 Modelling of the unresolved turbulence

Some means of approximating the sub-grid stress tensor is therefore required, for
which two principal classes of approach exist. In the class known as implicit LES
(ILES), a numerical scheme with an appropriate quality and quantity of numerical
dissipation is considered to provide the dissipative action of the unresolved scales.
ILES has emerged predominantly from the atmospheric and meteorological research
community, where the effects of solid boundaries are not of great importance. ILES
typically encounters problems when applied to wall-bounded flows, although some
considerable progress has been achieved recently (e.g. Hickel & Adams [60]). An-
other potential problem with ILES is that the magnitude of the sub-grid stresses re-
mains unknown, which can present a disadvantage in applications where knowledge
of the total Reynolds stresses is of importance. Implicit LES is not considered further
in this work, and reference is made to the book of Grinstein et al. [53] for further
details.

The contrasting approach to implicit LES, sometimes referred to as classical LES in-
volves the explicit introduction of a mathematical model to describe the sub-grid
turbulence. This family of methods has largely been developed within the frame-
work of aeronautics research, for which the correct representation of near-wall flows
is of central importance. A brief description of some important sub-grid scale models
will now be given.

The Smagorinsky model

The first, simplest and still perhaps the most widely-applied sub-grid scale model is
that of Smagorinsky [150]. The first application of the Smagorinsky model (together
with a simple wall function) to the three-dimensional simulation of turbulence was
published by Deardorff in 1970 [35] for the case of fully-developed channel flow.
The model is an eddy viscosity model, employing the Boussinesq hypothesis (Sec-

42
 3.5 Partial resolution of turbulence (LES)

tion 3.4.2) to relate the sub-grid scale stresses to the local flow strain. The model is a
simple explicit algebraic expression, which was derived heuristically by Smagorin-
sky8 as

νsgs = (CS Δ )2 S ∗ , (3.10)

where

S∗ = 2Sij Sij . (3.11)
Under the assumption that the cut-off wave number lies within the inertial subrange,
Lilly [82] derived a value of CS = 0.1825 for isotropic turbulence9 . Deardorff [35]
found a much lower value to be appropriate in his pioneering channel flow investi-
gation, which has been widely corroborated by other researchers quoting values of
0.08 < CS < 0.11 for shear-driven flows.

A number of deficiencies are inherent to the Smagorinsky model, which can be sum-
marised as follows:
• The Smagorinsky parameter CS is clearly not a constant, as shown by the strong
difference in values appropriate for homogeneous and sheared flows. The
specification of a global value is inappropriate for realistic flows with mixed
turbulence phenomena.

• The positive and uniform value of CS causes the generation of eddy viscosity in
sheared laminar flow, and the model is hence incapable of predicting transition
to turbulence.

• For the same reason, the model contribution does not vanish in the limit of a
fully-resolved DNS.

• The effect of wall proximity in the viscous sublayer (Section 2.4) is not correctly
reproduced by the model, requiring the addition of damping terms such as the
van Driest formulation [173].

• The model is purely dissipative, i.e. the sub-grid scales always extract energy
from the resolved scales. Although this cascade principle (Section 2.2) is true
in the Reynolds-averaged sense, the reverse process of backscatter from small
to large scales can occur locally within the resolved field. The Smagorinsky
model cannot predict backscatter.

Dynamic sub-grid scale models

These fundamental problems with the Smagorinsky closure listed above form the
motivation for the dynamic sub-grid scale procedure of Germano et al. [51]. In this
8
It is however possible to derive the Smagorinsky model by many other means [144].
9
This value cited by B. Aupoix in [3].

43
3 Simulation and modelling of fluid turbulence

method, the appropriate value of CS is computed dynamically and locally in space

and time using a test filter larger than the LES filter width Δ (a test filter width of
2Δ was found by Germano et al. to be optimal). In this manner, the stress-strain
relationship is evaluated within the “test window” of the smallest resolved scales of
motion and CS is computed at each time step as a function of position from the in-
formation already contained in the resolved velocity field. This procedure is explicit
and algebraic in nature, hence a minimal additional computational cost is involved
per iteration.

The original Germano formulation demonstrated much improved performance and

generality, addressing all of the problems experienced with the Smagorinsky model:
• The dynamically-determined value of CS was shown to be capable of produc-
ing high-quality results for both homogeneous and shear-driven flows.

• CS = 0 is returned in laminar regions allowing the simulation of transition to

turbulence.

• CS = 0 is likewise returned in the limit of full DNS resolution.

• The correct wall-normal distribution of eddy viscosity is returned in the vis-

cous sublayer without resorting to damping functions.
• Locally negative values of CS are produced, allowing a reproduction of back-
scatter effects.
However, when applied to each grid point individually, excessive regions of negative
CS can be predicted, which lead to computational instabilities. For this reason, Ger-
mano et al. suggested averaging of the CS values in homogeneous flow directions,
thereby rendering the dynamic model inapplicable to complex geometries. Lilly [83]
proposed a modified technique for determining CS based on a least-squares method,
which although reducing the problem, still requires some degree of averaging for
stability.

Firm theoretical foundations for the dynamic procedure were provided by the work
of Ghosal et al. [52], from which a dynamic method free from the constraints of av-
eraging was introduced. A k-equation based dynamic model was proposed, which
is however associated with a high numerical cost due to the need to solve a complex
integral equation for the localised parameters. A dynamic one-equation subgrid-
scale (DOESGS) model was later proposed by Davidson [32], based on a transport
equation for the subgrid-scale kinetic energy, k sgs . By including the dynamically-
determined parameters in the k-equation source terms rather than directly into the
impulse equations, a greater degree of numerical stability is achieved: the oscillations
of the dynamic parameters are smoothed naturally by the convection and diffusion
terms of the transport equation. Despite the higher numerical effort per iteration
required by the solution of the transport equation, the DOESGS model in fact ex-
hibits a lower numerical cost than the dynamic Smagorinsky model [151, 144]. This

44
 3.5 Partial resolution of turbulence (LES)

is because less iterations are required per time step owing to the greater numerical
stability.

Yoshizawa k equation model

As mentioned in the preceding section, a family of transport equation based subgrid-
scale models exist. A prominent example is that of Yoshizawa [188]. The motivation
of the model was a desire to eliminate the assumption, inherent to the Smagorin-
sky model, that a balance between subgrid energy production and destruction ex-
ists. A transport equation for the subgrid-scale turbulent kinetic energy, k sgs was
formulated, which reduces to the Smagorinsky model under the local-equilibrium
assumption:

  
Dk sgs ∂ νsgs ∂k sgs k3/2
sgs
= ν+ + Pk sgs − Cε (3.12)
Dt ∂x j σsgs ∂x j Δ

νsgs = Ck Δ k1/2
sgs , (3.13)
with

Pk sgs = νsgs S ∗2 . (3.14)

Only approximate values for the model constants were provided by Yoshizawa [188],
hence those published by Fureby [48] are given:

σsgs = 1.0 Cε = 1.05 , Ck = 0.07 . (3.15)

Such models include transport and history effects of the subgrid quantities, how-
ever there is no dynamic character: the model constants are fixed. This model has
been described due to its relevance in some DES formulations and to discussions in
Sects. 3.7.6 & 8.3.

3.5.2 Requirements on the numerical scheme and resolution

First and foremost, it is worth repeating the statement made in Section 3.2 that LES
requires a three-dimensional spatial domain as well as temporal resolution, a prop-
erty inherited from DNS and owing to the fundamental dimensionality of turbulence
(Chapter 2).

Whereas in ILES use is made of the numerical scheme’s inherent dissipative char-
acter to act as a subgrid-scale model, in the derivation of classical LES models the
assumption is made that no dissipation is presented by the numerical scheme. The
requirement of negligible numerical dissipation therefore constitutes one of the prin-
cipal demands placed by classical LES on the numerics. The widely propagated doc-
trine of ensuring the kinetic energy conservation of the numerical scheme [10, 43, 86]
is indeed centred upon this issue. Although such requirements are also placed on

45
3 Simulation and modelling of fluid turbulence

DNS numerics, in contrast to both RANS and DNS the solution is not smooth at the
grid level in LES.

In the general situation that the local structure of the turbulence is not known in ad-
vance, it should be ensured that an isotropic grid resolution is employed. Assuming
isotropic large eddies, the resolution capability will be dictated by the coarsest grid
direction, meaning that strong grid anisotropy can be wasteful in LES. Only in very
specific cases can anisotropic grids be recommended, for example in the near wall
region when the predominant flow direction is known in advance. Because of the
streak-like structures of the large-eddies here (see e.g. [49]), a coarser grid resolution
in the streamwise direction relative to the spanwise direction is routinely applied in
LES of canonical wall-bounded flows [128]. The distribution of the grid resolution
should furthermore be as smooth as possible. Sudden jumps in the grid resolution
can be very detrimental, particularly when oriented normal to the mean streamwise
direction (an issue investigated in detail by Vanella et al. [174]). When judiciously
applied tangentially to the flow direction, discontinuities in the grid resolution can
however prove both benign and beneficial [9, 144].

Concerning the necessary grid resolution for LES, it can be generally said that a min-
imum resolution up to a cut-off wave number in the inertial subrange is required.
This arises from Kolmogorov’s hypotheses (Sect. 2.2.1) and the general philosophy of
resolving the large, geometry-specific eddies and modelling only the more isotropic
and universal small eddies. Indeed, the derivation of many SGS models is founded
upon this assumption. The scales corresponding to the onset of the inertial subrange
depend however on the length scale of the largest eddies (recall Pope’s estimate [116]
of 61 0 , cited in Sect. 2.2) and hence on the local nature of the turbulence. Indeed, the
extent of the inertial subrange depends on the Reynolds number, vanishing entirely
at lower Re (Fig. 2.5). This presents LES of low Reynolds number flows with con-
ceptual difficulties10 . Even at higher Re, the dependence on the local length scale
renders a-priori grid design difficult11 , and leads to the reliance on rules of thumb
for particular situations. These can be principally divided into guidelines for bound-
ary layer flows and “detached” flows.

LES of boundary layer flows is dominated by the need to resolve the streak structures
existing around the buffer layer. The grid dimensioning is scaled in wall friction
units (Sect. 2.4.1), although quite some scatter is observed in published resolution
recommendations (a review of which is given by Sagaut [138]). The appropriate res-
olution will undoubtedly vary depending on the accuracy of the numerical scheme
and subgrid-scale model [128]. Commonly-cited values are however Δxmax + ≈ 40
10
A further situation where the inertial subrange vanishes is the near-wall damping effect on turbulence. In
contrast to the Reynolds number reduction effect however, this is because of a shrinking of the large scales
0 instead of a growth of the Kolmogorov scales η. For this reason, fully-resolved LES hence approaches
DNS levels of resolution near the wall.
11
One notable approach however involves the extraction of length scale estimates from a precursor RANS
simulation, which are employed to define the required LES grid resolution [2].

46
 3.5 Partial resolution of turbulence (LES)

and Δz+max ≈ 20 in the streamwise and spanwise directions respectively. The wall-
normal grid spacing should be condensed towards the wall giving rise to Δy+ max ≈ 1
at the wall and Δy+ max ≈ max ( Δx ; Δz ) near the boundary layer edge. For this a
geometric stretching is often used, with a stretching ratio k ≤ 1.15 often cited as a
guideline (e.g. [148]).

The domain size for LES of boundary layer flows is determined by the need to cap-
ture the important large scales of the energy-containing range (estimated at 60 by
Pope, Sect. 2.2). For channel flow, Moser et al. [100] report that a tangential domain
size of L x = 2πδ and L z = πδ is sufficient to provide a near complete decay in the
two-point correlations over the half-lengths in each direction. A further study to
identify the “minimal flow unit” for DNS of channel flow by Jiménez & Moin [68]
found this to scale in wall units, with a minimum L + +
x ≈ 250 to 350 and L z ≈ 100
cited. Combining the domain size and resolution guidelines allows the grid scal-
ing of LES for wall-bounded flows to be derived. Assuming the ability to vary Δx
and Δz optimally across the boundary layer, Chapman [26] determined the required
number of grid points to scale as Nxyz ∝ Re1.8 . For structured solvers however,
such “nested” optimised grids are not possible, and Reynolds [121] gives a more
expensive Nxyz ∝ Re2.2 . This dominance of the fine near-wall structures on the com-
putational cost of LES for wall-bounded flows motivated Spalart [159, 153] to refer
to such simulations as quasi-DNS or QDNS. Because the size of the largest eddies
grows linearly with wall distance, whereas the Kolmogorov length scale grows more
slowly (proportional to d1/4
w [116]), the wall region penalty is relatively more severe
for LES than it is for DNS.

The resolution requirement for detached turbulent flows in the absence of the length
scale reducing effect of wall proximity is less severe. The requirement to resolve
the energy-containing range sufficiently gives rise to a fixed number of grid points
irrespective of Re0 . Assuming Pope’s estimate of the extent of the energy contain-
ing range, this results in a minimum box size of 363 grid points, and Spalart [154]
quotes a more pragmatic minimum of 323 cells. As mentioned, the difficulty of the
dependency on 0 leads to a lack of clear a-priori LES refinement criteria for detached
flows. As such, a range of diagnostic methods to assess the sufficiency of grid reso-
lution for an existing LES solution field exist. One approach is to compute the tur-
bulence energy spectra and inspect the existence of a κ −5/3 exponent at the highest
resolved wave numbers as an indicator of resolution up to the inertial subrange. This
is however an arduous process when applied throughout the turbulent region, and
is furthermore strictly invalid for anisotropic flows. Another set of methods is based
upon the expectation that a minimum ratio of resolved to modelled turbulent kinetic
energy should result from a sufficiently-resolved LES, with the value 80% suggested
by Pope [116]. Upon this basis, the diagnostic index LES_IQ (for “LES index of qual-
ity”) was proposed by Çelik et al. [24]. In still another approach, Šarić et al. [177]
conducted an LES grid resolution estimate based on a multiple of the Kolmogorov
length scale, obtaining its distribution from a precursor Reynolds stress model com-

47
3 Simulation and modelling of fluid turbulence

putation12 .

In addition to spatial resolution, a sufficiently fine time step is required to capture

the motion of the resolved turbulent structures properly. This issue was investigated
by Choi & Moin [27] for DNS of turbulent channel flow at Reτ = 180 using a fully-
implicit method. For very large time steps, the flow relaminarised due to the time
filtering effect on the resolved turbulent fluctuations. In contrast, convergence of
the solution was obtained for the two finest time step sizes tested. In between these
extremes, an over-prediction of streamwise Reynolds stresses and under-prediction
of the spanwise and normal components was seen, qualitatively mimicking the effect
of insufficient spatial resolution and/or excessive numerical dissipation. The concept
of a CFL criterion has been introduced in Sect. 3.3 as a means of checking the balance
of spatial with temporal resolution. Indeed, the converged fine steps reported by
Choi & Moin were for values of CFL = 0.5 and 1. The criterion of CFL ≤ 1 also
corresponds to the advice given by Spalart [154] for LES.

3.6 Hybrid RANS-LES

The level of interest in hybrid RANS-LES approaches has increased dramatically in
recent years. This is believed to be motivated by the following factors:
• Dissatisfaction with the RANS modelling paradigm (as described in Sections
3.2 and 3.4): Despite many years of intense effort, the accuracy of RANS meth-
ods leaves much to be desired. Sources of uncertainty are significant and a
“universal model” remains elusive. This is particularly true for flows featuring
separation.

• Excessive expense of fully-resolved pure LES methods (Section 3.5): Although

more accurate than RANS models, the computational cost of pure LES remains
prohibitive at high Reynolds numbers, particularly for wall-bounded flows.

• The significant gap between the computing demands of pure RANS and LES
(Fig. 3.2): Commonly-available computing power is expected to occupy this
region in the near future.

• The growth of interest in unsteady flow prediction: For example in aeroa-

coustic and flight mechanical applications, and for greater accuracy in aero-
dynamic design near the limits of performance envelopes where separation
and unsteadiness are characteristic. URANS approaches reveal a number of
key weaknesses in this respect (Section 3.4.5).
As a result, there has been an explosion of new methods reported in the literature.
The common and distinguishing features between these are often difficult to discern,
12
In this last case however, it would appear as if the factor of η applied to determine the LES resolution
should depend on the Reynolds number. Such approaches to identify the inertial subrange based on the
Kolmogorov scales hence appear less suitable than equivalents based on the large eddy length scales.

48
 3.6 Hybrid RANS-LES

as are hypothetical or proven advantages or suitability to different applications. A

valiant attempt to combat this problem by summarising and grouping the methods
into a taxonomy has been published by Fröhlich & von Terzi [47]. Their approach
has been to define and class methods based on key aspects of their formulation,
and the resulting taxonomy is fairly complex. Nonetheless the work serves as an
excellent overview of the diversity of approaches pursued, to which the reader is re-
ferred. Another highly competent and comprehensive review can be found in Sagaut
et al. [139], which furthermore offers a formulation-oriented nomenclature system of
its own. A detailed review of hybrid RANS-LES approaches other than DES is con-
sidered outside of the scope of this work, however an original perspective on taxon-
omy will be described and employed.

In the proposed taxonomy, it is acknowledged that several different perspectives ex-

ist as to the goals of hybrid RANS-LES. This aspect has not been considered in as
much depth as the formulation issues by Fröhlich & von Terzi [47], however it offers
a much simpler means to categorise hybrid RANS-LES methods by ignoring formu-
lation issues entirely. Four such “hybrid RANS-LES ambitions” are hence identified
as:
1. Very large-eddy simulation (VLES): Allowing coarser grids for LES within the
LES region by enabling a filter width at scales larger than the inertial subrange.
2. Embedded LES: An arbitrary, user-defined LES region of interest within an
encompassing RANS domain.
3. Wall-modelled LES (WMLES): RANS applied to the near-wall region of an LES
computation to reduce tangential grid resolution requirements (with the RAN-
S/LES interface inside the boundary layer).
4. Regional RANS/LES activity dependent on flow physics: Division of the RANS
and LES domains on the basis of local flow characteristics, e.g. attachment or
separation, respectively.
Because there isn’t a single method that lays claim to all of these ambitions or target
applications, their consideration is particularly important. Unfortunately these are
seldom explictly declared by the method authors, which appears to contribute signif-
icantly to widespread misunderstanding. Attempts to apply various hybrid methods
to fundamentally inappropriate target applications is an all too common occurrence
in the literature. The terminology of the hybrid RANS-LES ambitions is hence pro-
posed to assist clarity of discussion within this work and, it is hoped, within the
research community. It must of course be acknowledged that nomenclature is arbi-
trary and subjective in nature, and different meanings for these terms are widespread
(e.g. some interpretations of URANS as VLES, or limitation of the definition hybrid
RANS-LES to a much more specific class of approaches than the umbrella term used
here). This taxonomy approach is furthermore proposed as a good framework for the
formulation of best practice guidance for industrial application of hybrid RANS-LES
methods.

49
3 Simulation and modelling of fluid turbulence

3.6.1 Nomenclature for filtering and averaging

In Section 3.4 the nomenclature U i was introduced to refer to the averaged velocity
vector in pure RANS approaches, whereas in pure LES U i is the filtered velocity
vector (Sect. 3.5). However, for many hybrid RANS-LES methods, these concepts are
used interchangeably in the same solution field (which doesn’t “know” whether it
is averaged or filtered). The upper case symbol Ui will hence be used to refer to the
velocity vector in hybrid RANS-LES, which can be considered to be either averaged
or filtered depending on the activity of the RANS or LES mode respectively. In cases
where a distinction is necessary, the subscript sgs will be applied to refer to sub-grid
scale quantities. Concerning DES, as will become clear, the same variable is used for
the eddy viscosity in both RANS and LES operation. The notation νt will hence be
used universally.

3.7 Detached-eddy simulation

Detached-eddy simulation is the particular hybrid RANS-LES method forming the
subject of this work, which has since its introduction in 1997 [159] enjoyed wide-
spread acceptance in the industrial CFD community. In this section a detailed ac-
count of the development history of DES will be given, beginning with a discussion
of the central philosophy of the method in Sect. 3.7.1. The original formulation will
then be introduced in Sect. 3.7.2, before a summary of its principal limitations is
given in Sect. 3.7.3. The extensive and largely successful work undertaken to rectify
these problems will then be described in Sects. 3.7.4, 3.7.5 & 3.7.6.

The driving forces behind the development of DES are the research groups of P.
Spalart (Boeing Aircraft Company, USA) and M. Shur, M. Strelets and A. Travin
(St. Petersburg Technical University and New Technologies and Services, Russia),
and the publications of these authors are considered the authoritative sources of in-
formation on DES. In addition, considerable testing and validation work has been
carried out in collaborative European research programs, notably the EU-funded
FLOMANIA [55] and DESider [56] projects from which the majority of this work
has originated. The interested reader is also referred to the recent review article of
Spalart [156] for an overview of prominent application examples and a discussion of
the DES state of the art.

3.7.1 Philosophy of the method: “The ‘D’ in DES”

To begin this discussion on the conceptual foundation of DES, the definition pro-
vided by Travin et al. [171] will be cited:
“A Detached-Eddy Simulation is a three-dimensional unsteady numeri-
cal solution using a single turbulence model, which functions as a sub-
grid-scale model in regions where the grid density is fine enough for
a large-eddy simulation, and as a Reynolds-averaged model in regions
where it is not.”

50
 3.7 Detached-eddy simulation

The first specification of a three-dimensional and unsteady solution relate to the

properties inherited from LES, described in Sect. 3.5. As a result, any two-dimensio-
nal or steady-state solution of the DES equations is a misnomer as well as a concep-
tual error. The following item, that DES involves a single turbulence model, relates
to what is often referred to as the “non-zonal” nature of DES (together with the sin-
gle solution field aspect). However, as commented in the review of Fröhlich & von
Terzi [47], the terms zonal and non-zonal are of limited use in categorising hybrid
RANS-LES methods, not least because of inconsistent use in the literature. Finally,
a fundamental coupling of the RANS and LES activity to the local grid resolution is
expressed in the definition. As will be discussed in Sections 3.7.2 & 3.7.3, this aspect
is central to the formulation of DES as well as to some of its early shortcomings.

The above definition precisely outlines key attributes of the DES formulation whilst
avoiding unnecessary rigidity and formalism13 . However, it doesn’t address the am-
bitions or uses of the method directly. The four hybrid RANS-LES ambitions in-
troduced in Section 3.6 will now be revisited in an effort to elucidate the targeted
application range of DES as well as to demarcate it from other hybrid RANS-LES
strategies. The terminology of “natural” and “extended” uses of DES outlined by
Spalart [155] will be drawn upon.

Natural uses, or “the ‘D’ in DES”: Natural uses of DES involve the RANS treat-
ment of attached turbulent boundary layers in their entirety, whilst LES is applied to
regions of massively separated flow (only detached eddies are simulated). Natural
uses of DES hence correspond to the fourth hybrid RANS-LES ambition, whereby the
flow-physical criterion dictating the RANS or LES operation is the attached or sepa-
rated status of the local flow. As a consequence, any incursion of the RANS–LES in-
terface inside turbulent boundary layers is a violation of this requirement. Expressed
differently, any DES method must be capable of modelling an entire boundary layer
with RANS14 [155].

Extended uses: Extended uses of DES correspond to the third hybrid RANS-LES
ambition, namely wall-modelled LES (WMLES). The adjective extended is applied
for two reasons: Firstly, WMLES was not the ambition envisaged in the original pub-
lication of the method [159], which carefully expresses the the motivation of the for-
mulation in terms of natural DES usage. Secondly, the first tests of DES-based WM-
LES for channel flow [101] revealed a fundamental inaccuracy of such simulations
(to be described in more detail in Sect. 3.7.3).

The rationale motivating the intended natural use of DES can be explained when
some fundamental properties of pure RANS and LES approaches are considered
13
Note, for example, that the precise formulation of the RANS–LES switching mechanism is left open and
that the method is not limited to a particular RANS model.
14
This requirement hence eliminates DES formulations based on simple algebraic mixing length models
unless they include additional terms for the viscous sublayer and outer layer. It furthermore differentiates
DES from many WMLES methods proposed.

51
3 Simulation and modelling of fluid turbulence

with respect to attached and separated flow. As described in Sect. 3.4, RANS mod-
els can be well-tuned to handle simple boundary layer flows at minimal compu-
tational cost, however exhibit significant deficiencies in massively-separated flows
(even when computed in unsteady mode, i.e. URANS, Sect. 3.4.5). LES by con-
trast inherently delivers more accurate predictions due to the limitation of mod-
elling empiricism to the smallest turbulent scales (Sect. 3.5). However, as described
in Sect. 3.5.2, the numerical cost for wall-bounded flows is enormous and increases
strongly with Reynolds number. For detached flows far from the influence of walls,
however, the cost of LES is manageable and Reynolds number independent15 . These
motivational factors underlying DES are summarised diagrammatically in Fig. 3.7.

Attached boundary layers Massively−separated flow

D
RANS models well−tuned and inexpensive (U)RANS highly inaccurate
LES unfeasibly expensive, numerical
E LES expense manageable &
expense highly Re−dependent S Re−independent

Figure 3.7: Diagrammatical summary of the motivation of DES.

Having been of use in describing what DES is, it is informative to draw upon the
hybrid RANS-LES ambitions of Sect. 3.6 to discuss what DES is not. Firstly, DES is
not a method for VLES: There is no ingredient in DES targeted at allowing a coarser
grid in the LES-mode region than would be acceptable for the Smagorinsky model.
As a result, the usual requirements for LES spatial and temporal resolution outlined
in Sect. 3.5.2 must be adhered to in the region where LES-mode operation is desired
in DES.

Secondly, DES is not an embedded LES: Although it is in principle possible to con-

trol the local RANS or LES functionality via the grid design, an arbitrary placement
of these regions will not give the desired behaviour. Such arbitrary zone placement
requires considerably more elaborate methods than DES, particularly concerning the
formulation of interface conditions. In particular, where a RANS domain is con-
vected into an LES domain, physically-viable resolved turbulent fluctuations must
be explicitly injected into the solution. This is closely related to the grey area prob-
lem to be described in Sect. 3.7.3.

The original DES concept sketch from the presentation of the 1997 paper is shown
in Fig. 3.8, which illustrates well the envisioned application area of DES. The exten-
sive surface area covered by thin boundary layers is typical of external aerodynamic
applications. These can be reliably treated using RANS, and would incur unman-
15
In the statement “Reynolds number independent”, the Reynolds number based on the largest eddy length
scale 0 is considered. Of course, situations could be envisaged where 0 itself depends on the configura-
tion Reynolds number (e.g. decrease of the wake width behind a cylinder at supercritical Re, Sect. 2.5.2.

52
 3.7 Detached-eddy simulation

Figure 3.8: Original concept sketch of DES from the presentation of the 1997 paper [159]. Cour-
tesy P. Spalart.

ageable numerical expense with LES or even WMLES. The separated region of large-
scale turbulence behind the spoiler contrastingly represents an ideal application re-
gion for LES for the reasons already outlined. The onset of this region is furthermore
clearly-defined by sharp geometric features, which raises the question of the predic-
tive quality of DES in cases with more sensitive flow separation. In such cases, the
separation point will be dictated by the RANS model, which as identified in Sect. 3.4
is a distinct disadvantage. It is clear therefore that a hybrid RANS-LES method aim-
ing to improve separation point prediction relative to RANS must involve some LES
element in the upstream boundary layer and hence the higher associated costs. Nat-
ural DES in contrast only aims to improve the flow prediction following separation.

The shallow area of separation and re-attachment labelled with a question mark in
Fig. 3.8 reflects some of the conceptual problems with DES already identified at the
outset. These will be outlined in Sect. 3.7.3, following the description of the original
DES formulation in the next section.

3.7.2 The original DES formulation: DES97

The motivation and intended application area of DES (described in the previous sec-
tion) was well identified and discussed in the original proposal of 1997 [159], to-
gether with the first method formulation. Accompanying the publication of the first
major revision [158] (DDES, Sect. 3.7.4), the authors proposed the title of DES97 for
the original formulation, which is hence adopted in this work. Because of the em-
phasis on dividing the RANS and LES operation between attached and separated
flow, respectively, RANS models that incorporate the wall-normal distance, dw , as a
length scale were first considered a natural choice for the DES97 formulation. The
Spalart–Allmaras (SA) model [157], the equations of which are listed in Appendix B,
was hence chosen. The DES length scale was introduced to replace dw in all terms of
the SA model equations:

53
3 Simulation and modelling of fluid turbulence

 
LDES97 = min (dw ; LLES ) , LLES = CDES Δ , Δ = max Δ i ; Δ j ; Δ k . (3.16)

The wall-normal distance occurs in the destruction term of the ν̃ transport equation,
which is proportional to 1/d2w . Near to the wall, where dw < LLES , the DES length
scale is equal to the RANS model length scale, and the model formulation is identical
to the SA RANS model. Far from the wall, where LLES < dw , the length scale LLES
is employed, which consists of a measure of the local grid cell size, Δ, multiplied
by a model constant CDES . Δ corresponds to the LES filter width, and is formulated
as the maximum cell length in each index direction16 . This formulation is proposed
assuming that no knowledge of the local turbulent structure is available in advance,
such that the smallest resolvable isotropic eddies would scale with the coarsest grid
cell dimension.

It was shown that under assumed conditions of local turbulent equilibrium, i.e.
equality of the production and destruction terms in the model transport equations,
a proportionality of eddy viscosity to the strain rate and square of the filter width
analogous to the Smagorinsky model, Eq. (3.10), is returned (see e.g. Appendix D).
This provides evidence that the subgrid-scale modelling in the LES-mode operation
of DES is viable, and furthermore that CDES can be considered an analogue to the
Smagorinsky constant. The value CDES = 0.65 was first calibrated for SA-based DES
on the basis of decaying, isotropic turbulence (DIT) by Shur et al., 1999 [148] along-
side the publication of the first three-dimensional DES results.

The formulation hence successfully relates near wall regions to RANS modelling and
finer grid regions far from the wall to LES as desired. However, no consideration of
the boundary layer thickness is included, such that a danger of LES-mode activity
inside attached boundary layers is present. The original formulation however relied
on the expectation that RANS grids for external aerodynamic applications typically
involve grid spacings tangential to the wall much larger than the boundary layer
thickness.

Generalisation to other RANS models

The connection between the DES methodology and models with a length scale based
on the wall-normal distance was originally believed to be fundamental [159]. How-
ever, later publications demonstrated the applicability of the DES methodology to
RANS models in general. The model’s specific turbulence length scale was employed
in place of dw in the DES length scale definition as:

LDES97 = min ( LRANS ; LLES ) . (3.17)

The first demonstrations of DES with alternative RANS models were published by
16
This formulation hence assumes hexahedral structured grids and an equivalent must be formulated for
unstructured grid applications with arbitrary cell geometries.

54
 3.7 Detached-eddy simulation

Travin et al. [172] and Strelets [163], who presented a formulation based on the
Menter SST model [90]. The value of CDES was found to vary relative to the SA
model value, and specific values were calibrated against DIT for each of the k − ε
and k − ω branches of the SST model. The generalisation of the DES methodology
to other RANS models was an important step: As commented in Sect. 3.7.1, DES can
be expected to be “at the mercy” of the underlying RANS model in cases with sen-
sitive flow separation. Furthermore, the non-generality of RANS models has led to
the establishment of preferred models for different applications in the framework of
industrial best practice [23, 55].

The generalised methodology for the DES modification is in fact less clear cut than
the length scale substitution in SA model, where dw is only present in a single term
of the transport equation. For models such as SST however, the length scale is in-
corporated explicitly or implicitly in a number of terms in both transport equations.
Indeed, Strelets [163] acknowledged this freedom of choice concerning the term se-
lected for substitution of LDES . The decision taken to substitute only the length scale
in the k equation destruction term was motivated as follows: Firstly, the modifica-
tion should be as simple as possible as long as the basic requirement of a derivable
Smagorinsky-like form is met. Secondly, such a substitution provides the greatest
degree of similarity with the established SA-DES. The substitution of the DES length
scale in alternative terms was subsequently investigated by Bush & Mani [22] and
Yan et al. [186] and incorporated in the DES-like approach “X-LES” of Kok et al. [76].
These formulations will be described in Sect. 3.7.6 and this issue will be returned to
in the analysis summarised in Sect. 8.3.

Early results obtained with DES97 were highly encouraging, with considerable ad-
vantages relative to URANS methods demonstrated for flows featuring massive sep-
aration [55, 94, 148, 163, 171].

3.7.3 Problems with the original formulation

A number of shortcomings have however been identified with DES97, some of which
were anticipated from the outset [159] and some of which emerged upon further
scrutiny. Many of these have been successfully addressed in later revisions and some
remain open issues, apparently fundamental to the methodology itself. These will
be introduced in the following sub-sections before the respective enhancements are
described in Section 3.7.4.

Erroneous activation of near-wall damping terms in LES mode

Certain RANS models (including the SA model) contain additional damping terms
to ensure correct near-wall behaviour, such as in the viscous sublayer. When applied
in DES, the activity of the LES length scale can under certain circumstances con-
spire to cause a spurious damping of the subgrid eddy viscosity to near zero levels.
Such damping terms are often sensitised to quantities such as the near-wall distance

55
3 Simulation and modelling of fluid turbulence

and eddy viscosity levels, becoming active only when low levels of both are present.
In LES-mode regions with very fine grids however, the small value of Δ together
with low levels of subgrid-scale eddy viscosity activate these damping terms. This
effect also destroys the analogy to the Smagorinsky model, as the effective Smagorin-
sky parameter value (a model-specific expression built up of damping terms, model
constants and CDES ) ceases to be constant. Although discovered earlier, this error
mechanism was first published in 2006 [158]. In the same paper, a methodology to
derive an antidote function that restores the desired Smagorinsky model behaviour
was presented. This remedy is described in Sect. 3.7.4.

Incursion of LES mode inside the boundary layer

As described in Sect. 3.7.1, the DES97 formulation is based on the assumption that
the tangential grid spacing near the wall exceeds the boundary layer thickness by a
good margin. Only then will the interface where dw = CDES Δ be located outside
of the boundary layer as required. The problems to be expected when this condition
is violated were indeed anticipated in the original publication, however it was as-
sumed that grids fine enough to cause them would be unfeasibly expensive for the
foreseeable future.

However, it did not take long before an occurrence of this problem was reported (by
Menter & Kunz [92]) for the case of a highly-loaded airfoil with a small separation
region near the trailing edge. The separation point predicted by DES97 was found
to move further upstream than the location found by RANS (using the same model).
As a result, the authors named the phenomenon “grid-induced separation” (GIS),
due to the central role played by the grid resolution. The mechanism was identified
as an encroachment of the RANS-LES interface inside the boundary layer, giving rise
to reduced levels of eddy viscosity on the LES-mode side. However, whilst proving
too fine for the correct functioning of DES97, the grid spacing was at the same time
not fine enough to resolve turbulent structures above the interface in the sense of a
WMLES.

In a later analysis of the problem by Spalart et al. [158], the problem caused by such
ambiguous grids was entitled “modelled stress depletion” (MSD), reflecting the ef-
fect of reducing the RANS Reynolds stresses. MSD can of course give rise to reduced
skin friction prediction without the GIS phenomenon arising.

MSD and GIS probably represent the most serious deficiency of the DES97 method as
far as industrial applicability is concerned. Such an over-sensitivity to the grid res-
olution results in impractical grid design constraints as guaranteeing a sufficiently
coarse tangential resolution can in many cases be impossible, particularly for struc-
tured grid methodologies17. Furthermore, the result that a refinement of the grid can

17
The example of a planar channel with a backward-facing step can be considered: A streamwise refinement
approaching the step would give rise to the same streamwise refinement along the opposite flat wall.

56
 3.7 Detached-eddy simulation

in fact deteriorate the results is both paradoxical and undesirable.

A number of proposals to combat this problem have been made, most notably by
Menter & Kunz in the same paper [92] and by Spalart et al. [158]. These will be
introduced in Sect. 3.7.4.

The grey area problem

The grey area problem was identified in the original DES publication [159] and refers
to a region of undefined modelling existing between the RANS mode and fully-
developed LES zones. The severity of the grey area problem is furthermore highly
case-specific. The grey area can be described by considering a turbulent boundary
layer which at some point separates from a solid surface to become a free shear layer.
It is assumed for the sake of simplicity that the grid resolution is sufficient for a well-
resolved LES immediately following separation (which is unlikely to be the case in
practice). As established, the attached boundary layer should be handled entirely
by RANS mode with all turbulence modelled and none resolved. The problem arises
when considering the region immediately following separation where LES should be
carried out, in which the majority of the turbulent kinetic energy should be resolved.
However, this resolved turbulence is not convected from the upstream RANS so-
lution. Hence, when the DES switches to LES mode in such cases, the initial lack
of resolved turbulence results in something that is neither RANS nor LES, due to
insufficient modelled and resolved turbulence respectively. Instead, as far as the re-
solved field is concerned, a kind of pseudo laminar-turbulent transition must occur
resulting in fully-developed resolved turbulence further downstream. This problem
is fundamental to all hybrid RANS-LES methods and the focus of much study. In
the case of DES and similar hybrid methods however, the problem is exacerbated by
the transport of eddy viscosity from the RANS boundary layer into the LES-mode
region. Similarly, the grid cannot in general be expected to be sufficiently fine for
such thin shear layer LES resolution. Both effects further delay the development of
resolved structures.

It is clear that the extent and severity of the grey area problem is highly dependent
on the flow in question and the strength of the shear layer instability. Indeed for
massively-separated flows (e.g. bluff body wakes), the impact of recirculating re-
solved turbulence upon the early shear layer provides an amplification mechanism
to the development of resolved turbulence, resulting in a negligible grey area. This
feedback mechanism is weaker or non-existent for cases of more shallow flow separa-
tion, in which case a significant grey area problem can be expected. This is the reason
why DES is considered particularly suitable for massively separated flows. However,
taking the example of wing or airfoil aerodynamics, it is precisely the weakest occur-
rences of separation in the early stall onset that are of greatest relevance to predicting
the off-design performance. For cases with an extensive grey area therefore, the net
deficit in turbulent shear layer mixing (resolved or modelled) can be expected to re-

57
3 Simulation and modelling of fluid turbulence

sult in delayed flow reattachment and exaggerated recirculating regions.

It should be noted that in some publications the term grey area is applied to the
RANS/LES interface in a WMLES setting and the associated problem of log-layer
mismatch to be described in the next subsection. Although this is clearly related in
some ways, the term grey area is applied solely to the RANS/LES interface occurring
between attached and separated flow in this work.

The grey area problem remains an open issue in DES, effectively limiting its rec-
ommended application range. A remedy but not a cure could emerge from consid-
erations of alternative length scale substitutions, as will be described in Sect. 8.3.
It must be acknowledged however that a full resolution of the grey area problem
would require the explicit addition of realistic turbulent fluctuations at the onset of
the LES-mode region, which is expected to be a highly complex task.

Log-layer mismatch in WMLES scenarios

Recalling that WMLES is considered an extended and not a natural use of DES, such
simulations are nonetheless a highly attractive prospect. Detailed tests of the perfor-
mance of DES97 for fully-developed turbulent channel flow at a range of Reynolds
numbers were performed by Nikitin et al. [101]. These test were considered success-
ful in certain key respects: The resolved turbulence in the LES-mode channel core
was sustained and high Reynolds numbers could be simulated with wall-tangential
grid spacing unlimited in wall units. However, a consistent pathology was identified
in the form of a kink in the velocity profiles between the RANS and LES log-layers,
which has since become known as log-layer mismatch (LLM) [170]. This gave rise to
an unacceptable under-prediction of the skin friction coefficient of the order of 15%.

LLM is in fact a very common feature of hybrid RANS-LES methods targeting WM-
LES applications, having been reported for a wide range of simpler formulations
(e.g. by Baggett [5], Hamba [57] and Piomelli et al. [114]). Opinions vary on both the
cause of the phenomenon and its remedy (see e.g. Hanjalić et al. [58]), however the
super buffer-layer that emerges between the RANS and LES log-layers is consistently
associated with oversized streak structures known as super streaks.

The LLM problem has been addressed in a recent major extension to the DES concept,
to be described in Sect. 3.7.5.

3.7.4 Enhanced versions of DES

Many of the problems outlined above have been tackled in subsequent enhanced
DES versions. The DES improvements aimed at natural DES applications (Sect. 3.7.1)
will be described in this section, whereas an improvement enabling extended WM-
LES usage is treated in the following Section 3.7.5.

58
 3.7 Detached-eddy simulation

Correction of RANS model damping term behaviour

A correction function to resolve the problem of unwanted activity of the RANS
model damping terms in LES mode was published for the SA-model by Spalart et
al. [158]. As this function is model-specific, a general methodology for analysing the
need for such a function in different RANS models and, where needed, the derivation
thereof was described. The correction function, Ψ, functions as an antidote to restore
the Smagorinsky model behaviour for all values of eddy viscosity. The correction
function is incorporated in the LES length scale

LLES = CDES ΨΔ , (3.18)

and for models where no such function is required (e.g. the 1988 Wilcox k − ω
model [181]), Ψ = 1 can be set.

The published function, including numerical limiters, for the Spalart–Allmaras mo-
del with trip terms activated is
⎧ ⎫
⎨ 1 − C Cκb12 f ∗ [ f t2 + (1 − f t2 ) f v2 ] ⎬
Ψ2SA = min 102 ; w1 w
  , (3.19)
⎩ f v1 max 10−10 ; 1 − f t2 ⎭

where f w∗ is the asymptotic value of the function f w for high eddy viscosity values.
The methodology for deriving these functions is described in more detail in Ap-
pendix D, where the derivations for the RANS models considered in this work are
also reported.

The Ψ function methodology was shown by its authors to be an effective remedy

of the damping function problem, whilst leaving the behaviour of these functions
unaltered in RANS mode.

Protection of RANS mode operation in boundary layers

To resolve the problem of MSD caused by encroachment of LES-mode activity inside
the boundary layer, a fairly significant modification of the original DES97 formula-
tion was found to be necessary. Instead of depending purely on the grid (and turbu-
lence length scale in the case of the generalised formulation), some aspect of solution
dependence must be incorporated into the DES length scale definition. Only with a
suitable sensor for the presence of a turbulent boundary layer can it be ensured that
this is modelled using pure RANS.

The first such proposed shield function, known as a GIS-shield, was proposed by
Menter & Kunz together with their diagnosis of the GIS problem [92]. This was a
model-specific fix for DES based on the SST model, which exploited the SST model’s
F2 function to sense the presence of an attached turbulent boundary layer. Inside this
region, LES-mode operation of the DES was disabled.

An analogous and more generally-applicable function was later proposed by Spalart

59
3 Simulation and modelling of fluid turbulence

et al. [158], for which the boundary layer sensor, rd , the shield function, f d and the
corresponding modification to the DES length scale are
ν +ν
rd = "t  , (3.20)
∂Ui ∂Ui
κ2 d2w max ; 10−10
∂x j ∂x j
# $
f d = 1 − tanh (8rd )3 , (3.21)

LDDES = LRANS − f d max (0 ; LRANS − LLES ) . (3.22)

The sensor function rd is based strongly on the function r from the Spalart–Allmaras
model, Eq. (B.5), however slightly modified so as to be applicable to any eddy viscos-
ity model. The hyperbolic tangent blending function in f d was tuned such that the
earliest onset of LES mode occurs just outside the boundary layer whilst avoiding an
excessively strong shield function. The function assumes the value f d = 0 inside a
turbulent boundary layer, blending smoothly to f d = 1 at the boundary layer edge.
Incorporated into the DES length scale definition as it is, the shield function delays
the switch to LES mode until outside the turbulent boundary layer. For this reason,
the method was named delayed DES (DDES).

DDES has been shown to successfully shield attached boundary layers irrespective
of the grid resolution, and hence represents a significant robustness enhancement.
The method has hence been proposed to replace DES97 entirely. The solution depen-
dency however does give rise to a new peculiarity, namely the possibility of a dual
solution. In the presence of resolved turbulence near the wall, the shield function
recedes back inside the boundary layer. This enables WMLES-type simulations to
produce sustained resolved turbulence, however the LLM problem remains [170].
For cases such as a periodic channel, the solution dependency boils down to a de-
pendency on initial conditions: For simulations begun from a smooth initial solution
field, the shield function grows from each wall and meets in the middle, returning a
valid RANS solution. However, if the solution is initialised with turbulent fluctua-
tions, the shield function activity remains near to the wall and the nominal WMLES
situation emerges. In an asymmetric test, with one side initialised with fluctuations,
the other with a smooth RANS profile, the LES side was shown to prevail, engulfing
the entire channel core [149].

3.7.5 Incorporation of WMLES functionality

The LLM imperfection identified for DES by Nikitin et al. [101] and described in
Sect. 3.7.3 has been tackled in the most recent DES development work. The moti-
vation for this extended capability is first and foremost the promise that WMLES
offers in a wide range of applications, particularly internal flows with thick bound-
ary layers. The necessity of proper WMLES prediction is however also a result of
the dual-solution nature of DDES. To illustrate this, the example of a separating-re

60
 3.7 Detached-eddy simulation

attaching flow such as a backward-facing step is considered. Even if a natural DES

is targeted, the impact of resolved turbulence on the boundary layer at the reattach-
ment point gives rise to some unavoidable WMLES content in the simulation. A
WMLES situation will arise even with boundary layer shielding such as DDES ap-
plied, as described in the above section. The fact that WMLES regions can occur in
natural DES simulations therefore underscores the importance of such capability.

The outcome of this effort is the “improved DDES” (IDDES) formulation, first pub-
lished by Travin et al. [170], which was followed by a more detailed discussion (and
nomenclature revision for enhanced clarity) by Shur et al. [149]. IDDES consists of
a novel hybrid RANS-LES formulation for pure WMLES application, which is in-
corporated into the DDES concept. As such, the WMLES model will first of all be
described in the next subsection, followed by the complete combined model.

The hybrid model for WMLES

The hybrid RANS-LES model for WMLES has been formulated to eliminate LLM,
whilst maintaining a formulation compatible with the general DES approach. As
such, the method is in principle applicable to any RANS model and non-zonal in
nature. Furthermore, unlike many existing hybrid RANS-LES methods for WMLES
(see e.g. [47, 58]), it is in no way constrained to channel flows or flows with a homo-
geneous direction and requires no specific user input (such as interface placement or
input from DNS data).

The principal ingredients of the method can be summarised as a near-wall modifica-

tion of the LES filter, Δ, and a more sudden transition between the RANS and LES
length scales than given by DES97 or DDES. The method is based on a number of
sensors of different boundary layer regions and blending functions that have been
formulated through a combination of theoretical considerations and empirical tun-
ing.

The motivation for a modification to the grid filter definition is the observation that
strongly different values of the Smagorinsky constant are required for homogeneous
and sheared turbulent flows (Sect. 3.5.1). An alternative to local variation of the
Smagorinsky constant applied in dynamic LES was proposed for IDDES, based on
the observation that the Smagorinsky constant always appears as a product with Δ
in the model equation. The model parameter CDES is therefore held constant and the
value of Δ is reduced near to the wall from the standard “maximum” formulation,
as given by Eq. (3.23).

Δ = min [max (Cw dw ; Cw hmax ; hwn ) ; hmax ] , (3.23)

where Cw is an empirical constant with the value 0.15, hmax is the maximum of the
grid spacing in all three directions (i.e. the standard DES formulation of Δ), and
hwn is the grid spacing in the wall-normal direction. This formulation gives rise to
three distinct zones: far from the wall, where Δ = hmax , very near to the wall, where

61
3 Simulation and modelling of fluid turbulence

Δ = Cw hmax and a linear blending in between. The blending is given by Δ = Cw dw

for grids with a stretching ratio less than 1 + Cw = 1.15 in the wall-normal direction,
and is steeper than this for stronger stretching ratios. Adoption of this Δ definition
was shown to enable fully-resolved LES of channel flow using the Smagorinsky con-
stant value obtained by calibration against DIT [170, 149], whereas the cubic root of
the cell volume and Δ = hmax variants give very poor results.

The blending between RANS and LES length scales is governed by the hybrid length
scale LWMLES , the basic weighting function of which is f step. This is constructed such
that in the near wall region f step = 1 and LWMLES = LRANS ; away from the wall
f step = 0 and LWMLES = LLES . This function is dependent on the grid construction
only, more precisely upon the ratio of the wall-normal distance and the maximum
cell length dw /hmax . The formulation of f step is given in Eq. (3.24).
 
f step = min 2e−9α ; 1
2
(3.24)
dw
α = 0.25 − (3.25)
hmax
It can be seen by inspection that the begin of the switch between RANS and LES
length scales occurs when α = (ln 0.5/ − 9)1/2 = ±0.2775, i.e. when dw = 0.53 hmax .
The shape of f step is shown in Fig. 3.9, from which the sharp switch at dw /hmax =
0.5275 and rapid drop to zero at around dw /hmax ≈ 1 can be seen.

1
fstep
0,8 frestore / Ψ famp

0,6

0,4

0,2

0
0 0,2 0,4 0,6 0,8 1 1,2
dw / hmax

Figure 3.9: The grid dependent RANS-LES blending function f step and the grid dependent part
of f restore (i.e. max [( f hill − 1) , 0]) as functions of the wall normal distance ratio with
maximum cell spacing, dw /hmax .

The second component function of LWMLES is intended to counteract the undesirable

reduction of the RANS modelled stresses near to the interface caused by interac-

62
 3.7 Detached-eddy simulation

tion with the LES region of reduced eddy viscosity. The RANS length scale is there-
fore multiplied by a positive “boosting” function f restore, which itself consists of both
grid dependent and solution dependent components, f hill and f amp respectively. The
complete constitution of f restore is given in Eqs. (3.26–3.32).

f restore = max [( f hill − 1) ; 0] Ψ f amp (3.26)

%
2e−11.09α if α ≥ 0,
2

f hill = − 2 (3.27)
2e 9α if α < 0.
f amp = 1 − max ( f t ; f l ) (3.28)
 3 
f t = tanh Ct2 rdt (3.29)
 10 
f l = tanh Cl rdl
2
(3.30)
νt
rdt = "  (3.31)
∂Ui ∂Ui
κ2 d2w max ; 10−10
∂x j ∂x j
ν
rdl = "  (3.32)
∂Ui ∂Ui
κ2 d2w max ; 10−10
∂x j ∂x j

The grid dependent part, f hill , is a function of the ratio dw /hmax in a similar manner
to f step, indeed identically when f step < 1. f hill provides the shape function of f restore,
whereas the (locally varying) amplitude of f restore depends on the solution field, and
is given by f amp. This component is built of sensor functions for the viscous sublayer
and for the modelled log law region, rdl and rdt respectively, which are analogous
to the sensor function for the complete modelled turbulent boundary layer rd from
DDES (Eq. 3.20). Indeed, by comparing Eqs. (3.31) and (3.32) with Eq. (3.20), it can
be seen that rd = rdt + rdl . The parameters Cl and Ct are model-specific constants,
which must be tuned such that f amp is virtually zero when either f t or f l are close to
unity. The values quoted by the authors [170, 149] are Cl = 3.55 and Ct = 1.63 for the
Spalart–Allmaras model, and Cl = 5.0 and Ct = 1.87 for the Menter SST model. The
grid-dependent part of f restore is shown in Fig. 3.9, where it can be seen that f restore is
only active in the RANS simulation region, i.e. where f step = 1.

The functions f step and f restore are combined to give the WMLES length scale, Eq.
(3.33), which is substituted into the background RANS model according to standard
DES practice (Sect. 3.7.2).
 
LWMLES = f step (1 + f restore) LRANS + 1 − f step LLES (3.33)
The behaviour of the solution-dependent parameters and blending functions is pre-

63
3 Simulation and modelling of fluid turbulence

sented together with a validation of the IDDES implementation for fully-developed

channel cases in Part III.

Combination of WMLES and DDES functionality: IDDES

The combination of the hybrid model presented above with DDES is intended to al-
low “mixed mode” computations, for which the model must be capable of selecting
the appropriate DDES or WMLES functionality for different flow regions. As such,
the combined method can be considered as an extension to DDES allowing a wider
range of flow types to be computed successfully, but with full backwards compati-
bility.

The authors of IDDES did not find a way of combining the original DDES length scale
definition, Eq. (3.22), with the WMLES length scale, Eq. (3.33), and presented the
reformulated L̃DDES given in Eq. (3.34) [149]. This was demonstrated to be equivalent
to the original.

 
L̃DDES = f˜d LRANS + 1 − f˜d LLES (3.34)
& '
f˜d = max (1 − f dt ) ; f step (3.35)
# $
f dt = 1 − tanh (8rdt )3 (3.36)

The modified DDES length scale formulation can be combined with the WMLES
length scale to give the IDDES length scale of Eq. (3.37).
 
LIDDES = f˜d (1 + f restore) LRANS + 1 − f˜d LLES (3.37)
In regions of resolved turbulence, rdt  1 and f dt ≈ 1. This is the same behaviour
shown by the DDES shield function, described in Sect. 3.7.4. Hence, f˜d = f step and
LIDDES = LWMLES . In the absence of resolved turbulence, f restore = 0 and conse-
quently LIDDES = L̃DDES . A demonstration of the behaviour of the IDDES length
scale for a mixed-mode computation is presented in Part III, where a separating-
reattaching internal flow is computed with a steady-state RANS inflow condition.

3.7.6 Other modifications to the DES formulation

A number of further modified DES formulations have been proposed by various
authors, of which two shall be briefly described.

The Breuer et al. modification [18] and “EDDES” of Riou et al. [125, 126]
A direct comparison of SA-DES97 with Smagorinsky LES was conducted for the
massively-separated flow over an inclined flat plate by Breuer et al. [18]. The DES
was seen to exhibit very high levels of eddy viscosity in the separated shear layer

64
 3.7 Detached-eddy simulation

in comparison to the Smagorinsky model, which caused a delay in the develop-

ment of resolved turbulent
 structures. This was attributed to two factors: Firstly,
the definition of Δ = 3 Δ i Δ j Δ k in the LES compared to Δ = max(Δ i ; Δ j ; Δ k ) for
DES resulted in a much larger value of Δ for DES in the highly-stretched shear layer
cells. Secondly, the f v1 , f v2 and f w damping terms were observed to deviate from the
asymptotic values expected far from the wall. A modification was hence proposed,
whereby


Δ= 3
Δi Δ j Δk

f v1 = 1

f v2 = 0

fw = 1 (3.38)
was set in the zone of LES-mode operation (i.e. when CDES Δ < dw ). Indeed, the eddy
viscosity levels and flow field in the early shear layer were seen to adopt the LES be-
haviour and the mean velocity and Reynolds stress profiles became more similar to
those of the LES.

The same modification was later proposed by Riou et al. [125, 126], apparently in-
dependently, although in conjunction with DDES and similarly applied to the LES-
mode region only. The method was baptised “extended DDES” (EDDES) and the
same improvement in early shear layer resolved turbulence was reported.

These modifications are clearly intended to tackle the grey area issue, which is ach-
ieved with a certain degree of success. The extent to which these modifications to
the damping functions can be generalised to DES based on different RANS models
and indeed the necessity of this remains to be seen however. Furthermore, it is not
entirely certain how the modification of the damping terms interacts with, overlaps
with or indeed replaces the Ψ function18 . It is also unclear why the value f w = 1 is
specified, and it indeed appears unjustified: As emerges from the analysis leading to
the derivation of the Ψ function (see [158], Sect. 3.7.4 and Appendix D), the asymp-
totic (i.e. high νt ) value f w∗ = 0.424 is derived for the LES-mode operation of SA-DES.
Setting f w = 1 in the LES mode region hence corresponds to lowering the effective
Smagorinsky constant value. Furthermore, in normal RANS operation f w = 1 oc-
curs within the log law region of a boundary layer only, with the function reducing
near the boundary layer edge finally tending to f w = 0 in free shear flow far from
the wall. In addition to the reduced effective CS therefore, the application of these
modifications exclusively to the LES-mode region hence gives rise to a discontinuity

18
If the SA-model trip terms are not active (i.e. f t1 = f t2 = 0), the modification of Eq. (3.38) indeed causes a
constant behaviour of the coefficient A of the Smagorinsky form of the model (presented in Appendix D).

65
3 Simulation and modelling of fluid turbulence

of the solution field at the interface. This could be perceived as contradictory to the
“non-zonal” philosophy of DES.

Alternative substitution of the DES length scale by Bush & Mani [22] and in the
“X-LES” method of Kok et al. [76]
Two alternative hybrid RANS-LES methods have been proposed by Bush & Mani
(2001) [22] and Kok et al. (2004) [76], which are closely related to DES in a number of
ways. The first method was not given a name by its authors, and is hence referred to
by their initials BM. The second method was called “extra-large-eddy simulation”
and the corresponding acronym X-LES will be applied. These methods are both
based on two-equation RANS models and both involve substitution of an analogy
to the DES97 length scale in the dissipation term of the k-equation, k , as follows:

k3/2
k, BM = (3.39)
LBM

LBM = min ( LRANS ; CBM Δ ) (3.40)

k3/2
k, X-LES = β k (3.41)
LX-LES
√ 
LX-LES = min k/ω ; CX-LES Δ . (3.42)

The Bush & Mani hybrid length scale is clearly identical to the DES97 length scale,
and hence CBM = CDES . The value for this constant was quoted as CBM = 0.1,
although no detailed calibration study was carried out. In the case of X-LES, the hy-
brid length scale differs only due to a formulation issue: The parameter β k = 0.09
of the TNT k − ω RANS model [75] upon which the method is based is not in-
cluded in the RANS length scale definition given by Kok et al. [76], rather as a co-
efficient of the destruction term. This is however not a fundamental difference, as
CX-LES = β k CDES . The authors calibrated CX-LES = 0.06 against DIT, which hence
corresponds to CDES = 0.667.

A common feature of both formulations, and a point where they differ from standard
DES, concerns an additional substitution of the hybrid length scale in the expression
for the eddy viscosity:

√
νt, BM = Cμ LBM k (3.43)
√
νt, X-LES = LBM k . (3.44)
It is readily shown that such a dual substitution gives rise to a subgrid-scale model
of identical form to the k equation model of Yoshizawa [188] (See Eq. (3.12) in Section

66
 3.7 Detached-eddy simulation

3.5.1). In LES mode therefore, the auxiliary equation (for ε or ω) of the background
RANS model is hence decoupled from the solution field. In both papers the possibil-
ity to obtain an identifiable SGS model in LES mode, not just under the assumption
of local equilibrium19 , is touted as a particular advantage. Although perhaps ele-
gant, it is however open to debate to what extent this represents an advantage: It
cannot be stated that the SGS version of standard DES formulations is a priori worse
than any “identifiable” SGS model20 .

This additional length scale substitution is not considered a sufficient criterion to

differentiate these methods from DES: The precise formulation is left open in the
method definition given by Travin et al. [171] (Sect. 3.7.1) and both methods appear
to target the fourth hybrid RANS-LES ambition (for X-LES this is explicitly stated,
whereas the Bush & Mani method is less clearly defined). Both methods however
each include further features that serve to differentiate them from each other, as well
as representing further differences to the standard DES formulation.

Beginning with the Bush & Mani formulation, a further additional feature concerns
the filter width Δ. This extends the standard DES filter definition of Eq. (3.9) with
additional terms based on the time step size, Δt:
 √ 
ΔBM = max Δ i ; Δ j ; Δ k ; |U | Δt ; k Δt . (3.45)
This formulation is very interesting, as it seeks to include an indicator of the temporal
filter width in the LES mode filter definition. The first Δt term is based on the local re-
solved velocity magnitude, whereas the second employs a velocity scale based on the
subgrid turbulence. As such, a kind of CFL criterion is incorporated implicitly in the
model, affecting both the filter width (and hence the level of sub-grid model activity)
and the RANS/LES switching. The method will hence seek to counter the effects of
time filtering, to be expected where an excessive time step is applied. These effects
are not taken into account by standard DES formulations, which are based solely on
the spatial filter size and assume that the user has selected an appropriate time step.
The performance and behaviour of this feature has however not been demonstrated
by the method’s authors.

The final distinguishing feature of X-LES also concerns the filter definition. The au-
thors here give preference to a fixed filter width defined by the user from case to case.
Although this is termed explicit filtering, the application of an explicitly-defined fil-
ter kernel to the resolved field (Sect. 3.5) is not meant, rather a fixed value of Δ. This
is motivated by a desire to seek grid-convergent solutions in the sense of separating
the numerical and modelling problems. By keeping the Δ value fixed and refining
the grid spacing, h, the numerical error should be diminished in such an approach.
This was pursued in the investigation of Weinman et al. [179], however concrete
19
Recall from Sect. 3.7.2 that the Smagorinsky model can be derived from the DES formulation under the
assumption of a balance between production and destruction terms.
20
As was pointed out by Prof. Strelets in his review of this manuscript.

67
3 Simulation and modelling of fluid turbulence

conclusions were hard to construe. The practical advantages of this procedure are
furthermore unclear, and it can be expected that a higher numerical cost results com-
pared to DES with Δ = h.

68
 Part II

Methodology

69
4 Numerical flow solver

4.1 Introduction
The flow solver used in this investigation is the ISTA in-house solver “ELAN” (Ell-
iptic Analysis of the Navier–Stokes equations), the foundation of which was laid by
Leiping Xue in the framework of his PhD thesis [184]. Upon this basis a number of
important extensions were implemented as reported in subsequent theses:
• The validation of a range of LES models and statistical averaging capabilities
by Schmidt (2000) [144]

• Non-linear RANS models by Lübcke (2001) [84]

• Combustion models, multi-grid acceleration methods and DES simulations for

aero-acoustic studies by Yan (2003) [185]

• Exploration of the grid deformation, turbulence modelling and DES capabili-

ties for flow control by Schatz (2003) [141]

• Fluid–structure coupling, grid deformation and the implementation and vali-

dation of DES methods by Bunge (2004) [19]
Alongside these theses specific to the ELAN solver, general descriptions and anal-
yses of related second order, finite-volume methods can be found in the literature.
As examples the textbooks of Ferziger and Perić [44] and Patankar [106], the lecture
course notes (in German) of Rung et al. [137] and the PhD thesis of Jasak [65] are
cited.

Within the scope of this work, no fundamental modifications to the numerical meth-
ods or architecture of ELAN have been undertaken. Correspondingly, a summary
level of detail will be given here with reference to precise descriptions elsewhere.
Furthermore, the purpose of this chapter is to describe the numerical schemes em-
ployed in the investigation, not to give a complete account of all ELAN features. In
particular, the incompressible mode of ELAN has been applied in all the test cases
considered. The new or modified versions of DES implemented during the course of
these studies are described in more detail in Chapter 5.

4.2 Discretisation
Discretisation transforms the underlying partial differential equations into a set of
algebraic equations. The discretisation of the solution domain has been introduced
in Section 3.1 together with an overview of the finite volume methodology. The

71
4 Numerical flow solver

principal characteristics of the ELAN solver in this respect can be summarised as

follows:
• Multi-block structured grids are used.

• Cell-centred discretisation: the variables are stored at the centre of each grid
cell, as opposed to the vertices.

• The transport equations are formulated in a curvilinear coordinate system ali-

gned with the grid, allowing body-fitted grids for complex geometries.

• Temporal discretisation is achieved with a specified time step, which is uniform

throughout the domain.
All of the transport equations to be solved, including those of the turbulence models,
can be expressed in a generic form consisting of a temporal derivative, a convection
term (these can be combined to form the substantial derivative, Eq. A.1), a diffusion
term and a source term. In Cartesian coordinates the generic transport equation is
 
∂ρφ ∂ρu j φ ∂ ∂φ
+ − Γφ = Sφ (φ ) , (4.1)
∂t
()*+ ∂x j ∂x j ∂x j ( )* +
( )* + ( )* + source term
temporal derivative
convection term diffusion term
where φ is the generic transported variable and Γ φ and Sφ are the diffusivity coef-
ficient1 and source term of the variable, respectively. Each of these terms require
specific discretisation methods, and those implemented in ELAN and applied in this
work are described in the following.

4.2.1 Diffusion term

As can be seen from Eq. 4.1 the first derivative of φ on each of the cell faces is required
to approximate the diffusive flux. These are modelled with a second order central
approximation, whereby the diffusion normal to the face is treated implicitly and the
cross diffusion explicitly.

4.2.2 Convection term

To determine the convective flux over the cell faces, the value of φ must be interpo-
lated onto the cell faces with reference to the neighbouring cell centres. A number of
interpolation schemes are available, the choice of which has a strong impact on the
quality and robustness of the solution.

Central differencing scheme (CDS)

In CDS a simple linear interpolation of φ is applied drawing on the neighbouring cell
centres. This scheme is of second order accuracy both on uniform and non-uniform
1
Here ρ is considered to be incorporated into Γφ .

72
 4.2 Discretisation

meshes [44] and exhibits very low numerical dissipation. However for convection-
dominated flows on coarse meshes where the cell Péclet number2 is large, CDS suf-
fers from spurious oscillations [106].

For LES, where negligible numerical dissipation is required together with very fine
grid resolutions to resolve the turbulent scales, CDS convection schemes are fre-
quently used. The problems of unboundedness arise typically for RANS simulations
where coarse grids are applied to modelled turbulence at high Reynolds numbers.

Upwind differencing scheme (UDS)

In such convection-dominated flows described above, it can intuitively be expected
that the flow variables at the cell face are more strongly influenced by the upstream
cell centre than by that lying downstream: this is the principle underlying upwind
schemes. In the most simple of these, φ at the cell face is set equal to the value at
the upstream cell centre. This has the advantage that boundedness of the solution is
guaranteed, so the kind of spurious oscillations that CDS exhibits are not a feature
of UDS. However, this stability comes at the expense of accuracy; UDS shows only a
first order error reduction and is characterised by high levels of numerical diffusion.
The numerical diffusion is furthermore magnified when the local flow direction is
oblique to the grid orientation, which can be guaranteed to occur in all but the most
simple of applications.

Total variation diminishing scheme (TVD)

Attempts to find a convection scheme that is both bounded and accurate have re-
sulted in a number of more complex schemes than the basic UDS and CDS. Higher-
order convection schemes are based upon interpolation using polynomials of higher
order, for which the number of local extrema increase in proportion to the polyno-
mial order. Higher-order convection schemes hence suffer from unboundedness,
and techniques of limitation have been developed as a remedy. For example, the
flux-limiting procedure of Boris & Book [12] creates differencing schemes that are
higher than first-order accurate but without spurious oscillations. A methodology
for oscillation-free flux-limited schemes is given by the notion of total variation dimi-
nution (TVD) introduced by Harten [59]. The TVD methodology can in principle be
applied to many higher-order convection schemes.

In ELAN, the higher-order convection schemes are implemented using deferred cor-
rection of the basic UDS scheme, with a user parameter that specifies the precise
scheme. All of these are however subjected to a TVD limiter, hence TVD is the
most appropriate nomenclature for the family of higher-order schemes implemented
in ELAN. These are of maximum third order accuracy, which is however reduced
when the TVD limiter is activated. The TVD scheme is reported by Schatz [141]
2
The cell Péclet number is a measure of the relative importance of convection to diffusion across a grid cell,
Pe = uΔx/Γφ (considering the x-direction).

73
4 Numerical flow solver

and Bunge [19] to give the highest quality results for RANS simulations with ELAN.
More details concerning the precise formulation are given by Xue [184] and Schatz
[140].

Flux blending
An alternative attempt to trade off between accuracy and boundedness of the scheme
has been proposed by Perić [107]. In this technique, known as flux blending, a lin-
ear blending between UDS and CDS is carried out in a ratio specified by the user.
For each specific simulation, the user can therefore decide on the required amount
of numerical diffusion to guarantee stability. Flux blending is applied uniformally
throughout the solution domain. A localised variation of the flux blending parame-
ter has been specifically developed for DES by Travin et al. [172], which is described
in Section 5.4. For the ELAN implementation of the hybrid scheme however, the
blending is between TVD and CDS rather than UDS and CDS.

4.2.3 Time discretisation

Second order backward differencing in time is applied, whereby values of φ at each
new time step (t + 1) are determined from the values at the previous two time steps
(t and t − 1) according to

∂φ  1  ( t +1) 
 = 3φ − 4φ( t) + φ( t−1) . (4.2)
∂t t +1 2 Δt
Because the temporal discretisation is implicit, a stable solution is guaranteed in-
dependent of the time step size. This is therefore adjusted such that a satisfactory
temporal resolution of transient flow phenomena is achieved.

Source term
The source term is linearised and decomposed into a constant part and a component
that is linearly dependent on φ. To preserve the diagonal dominance of the equation
system for implicit solution, the constant of proportionality of φ should always be
negative. In the case of a positive sign, the source term is handled explicitly. This pro-
cedure is particularly important for the turbulent transport equations, which exhibit
strong dominance of the source term. Further comments on source term treatment
are given by Patankar [106].

4.3 Pressure–velocity coupling

The discretised form of the incompressible Navier–Stokes equations exhibit linear
dependence of velocity on pressure and vice-versa. This inter-equation coupling re-
quires special treatment. A segregated approach is used, whereby the equations are
solved sequentially rather than simultaneously. The procedure is based on the SIM-
PLE algorithm [72] with which the pressure is iterated to convergence using a pres-

74
 4.4 Boundary conditions

sure correction. This method ensures mass conservation as the pressure equation is
derived from the continuity equation. A generalised Rhie and Chow interpolation is
used to avoid an odd-even decoupling of pressure and velocity with the collocated
storage scheme applied [105].

Only the gradient of the pressure affects an incompressible flow; the absolute value
is of no significance. In order to set the absolute value therefore, the pressure is fixed
to a user-specified value (the choice of which is immaterial) at a single point in the so-
lution field. This point should be located in a uniform, undisturbed flow region; near
to the inflow boundary is usually a good choice for external aerodynamic problems.

4.4 Boundary conditions

Because the underlying flow equations are elliptic in character, all borders of the
computational domain must be treated with appropriate boundary conditions corre-
sponding to the flow case at hand.

Treatment of pressure
For all the boundary condition types considered, the pressure is treated in the same
manner: a linear extrapolation onto the boundary condition plane is conducted and
a zero-gradient boundary condition is applied to the pressure correction equation.

In the following, the treatment of the momentum and turbulence model equations at
various boundary condition types will be described.

Inflow
At the inflow far field boundary the values of φ for the momentum and turbulence
model equations to be solved must be specified. The specification of the individ-
ual components of u i allows the variation of the global angle of attack or sideslip
in external aerodynamic calculations. The correct specification of turbulence model
parameters at the inflow boundary is for some applications far from trivial and can
have a strong effect on the results. This matter has been examined and discussed in
depth by Spalart & Rumsey [160] for the particular difficulties arising for external
aerodynamic investigations. A uniform inflow boundary can be specified in ELAN
with a single set of values, or alternatively a non-uniform profile can be read from an
input file.

Outflow
At the far field exit planes of the flow domain a convective outflow profile is ap-
plied, which allows resolved turbulent structures to be transported out of the domain
undisturbed [152]. The boundary condition is formulated as:

75
4 Numerical flow solver

∂φ ∂φ
+ u mean =0, (4.3)
∂t ∂n
where u mean is the average convective velocity over the outflow plane and n is the
component of xi normal to the outflow plane. The derivative in the normal direction
is determined to second order accuracy on the basis of the gradient in the cell centre
adjacent to the outflow plane. The time derivative is also determined to second order
accuracy as described in Section 4.2.3. In this way the value of φ can be determined
at the exit plane.

Solid walls
The no-slip condition is effectively applied to the velocity components at solid walls3 ,
i.e. u = v = w = 0. This is however not carried out directly, rather by specifying
the wall shear stress as an area force on the wall cell face [184]. For the turbulent
quantities the hybrid-adaptive boundary condition as described in Section 3.4.4 and
Appendix C, [135] and [145] is applied. This boundary condition seamlessly com-
bines the standard low-Re and high-Re formulations allowing a flexible positioning
of the first wall-normal grid vertex.

Symmetry
A symmetric boundary condition can also be applied, which treats each of the com-
ponent boundary faces as a symmetry plane. This enforces flow tangential to the
boundary face, resulting in zero flow across the plane. This boundary condition can
hence be considered as equivalent to a frictionless (Euler) wall.

Periodicity
Periodicity can also be applied to a pair of parallel boundary planes of the same
size and with the same point distribution; the flow conditions at each cell face are
set equal to those at the corresponding face on the paired boundary plane. Unlike
the symmetry plane, no restriction on through-flow or normal gradient is appar-
ent. However the enforcement of periodicity between parallel planes in the flow
can cause the amplification of harmonic modes and the attenuation of disharmonic
modes that may occur. Depending on the flow in question, a suitable separation of
the periodic planes (and sufficient resolution between these) must be ensured.

For a two-dimensional laminar, RANS or URANS simulation both symmetry and

periodic boundaries can be applied to the domain extents in the homogeneous span-
wise direction. If however a three-dimensional turbulence-resolving simulation is
conducted, the symmetric condition would artifically damp the spanwise fluctua-
tions of vortices near the boundary. The periodic boundary condition must therefore
be applied in such cases.
3
The components of velocity relative to the wall are referred to here.

76
 4.5 Solution of the linear equation system

4.5 Solution of the linear equation system

The discretisation methods described in the above sections lead to the construction of
a linear equation system to solve for the unknown flow quantities in each grid block
in an iterative manner. To this end, a three-dimensional extension by Xue [184] of the
nominally two-dimensional strongly implicit procedure (SIP) of Stone [162] is used.
The SIP solver is an incomplete lower–upper decomposition method specifically de-
signed for algebraic equations that are discretisations of partial differential equations.
The solver requires an ordered multi-diagonal solution matrix and is hence limited
to structured grids. The linearised equation systems for each transport variable are
solved sequentially to convergence using the SIP solver. The non-linear coupling
of the equations is handled by a re-computation of the coefficient matrix following
convergence of the SIP solver on all equations. The SIP solver iterations can hence
be considered as inner iterations, and the computation of the coefficients as outer
iterations.

77
4 Numerical flow solver

78
5 Implementation of detached-eddy simulation

Having summarised the numerical background of the ELAN solver in Chapter 4, a

more detailed description of the DES implementations will be given. The validation
and calibration of some of these basic features will be presented in Chapter 7.

5.1 Preceding work

As is the case for the ELAN flow solver, initial implementations of DES were inher-
ited from predecessors. The majority of the earlier implementation work was carried
out by U. Bunge as reported in his PhD thesis [19], chiefly within the framework of
the FLOMANIA project [55]. This included DES97 implementations based on the
SALSA, LLR and CEASM models (Section 5.2) and the hybrid convection scheme
of Travin et al. [172] (Section 5.4). These were tested on the stalled NACA0012 test
case as well as on a range of oscillating bluff bodies. Early applications of DES to
the study of flow control devices were reported by Schatz [141] and to aero-acoustic
source simulation by Yan [185].

5.2 RANS models applied

The RANS models to which DES has been implemented are summarised in Tab. 5.1,
together with the shorthand abbreviation with which they are referred to in this
work. The archival literature references are also given in the table. The equations,
functions and constants of each model are listed in Appendix B. The Lien–Leschziner
k − ε model is not applied for DES directly. It is included however as it serves as the
background model for the CEASM.

5.3 Overview of implemented versions

A summary of the implemented DES versions is given in Tab. 5.2, in which the sub-
stituted length scale and the available DES versions (DES97, DDES or IDDES) are
listed. The DES switch is implemented within the routine for each model, and the
specification of the DES97, DDES or IDDES length scale is currently done by manual
alteration of the source code.

Some of the results presented in Part III pre-date the implementation of the DDES
shield or the Ψ functions for low-Reynolds number term correction. It will therefore
always be stated whether the DES97, DDES or IDDES implementation has been ap-
plied. The Ψ function is here considered to be a component of DDES and IDDES, so

79
 Model name Transport Abbreviation References
variable(s)
Spalart–Allmaras model ν̃ SAE Edwards et al. (1996) [39]
with Edwards modification
Strain-adaptive linear Spalart– ν̃ SALSA Rung et al. (2003) [134]
Allmaras model
5 Implementation of detached-eddy simulation

Wilcox k − ω model k, ω WCX Wilcox (1988) [181]

Local linear realisable k, ω LLR Rung et al. (1996) [136]
k − ω model
Lien–Leschziner k − ε model k, ε LL Lien et al. (1993) [81]
Compact explicit algebraic k, ε CEASM Lübcke et al. (2002) [85]
stress model

Table 5.1: Summary of RANS models applied for DES.

80
 5.4 Hybrid numerical convection scheme

Model LRANS DES97 DDES IDDES

SAE dw • • •
SALSA dw • •
1
k 2
WCX • •
Cμ ω
1
k2
LLR • •
f βk ω
3
k2
LL
ε
3
k2
CEASM • • •
ε

Table 5.2: Overview of implemented DES versions.

these designations imply that the Ψ function is active (if required) unless explicitly
stated otherwise.

5.4 Hybrid numerical convection scheme

As discussed in Section 4.2.2, CDS is often an acceptable choice for pure LES ap-
plications due to the fine grids required to resolve the turbulent scales. For RANS
however, the unboundedness of CDS for high cell Péclet numbers results in spurious
oscillations on the typically coarser grids and more stable upwind-biased schemes
are practicable. These however prove to be too dissipative for LES.

It is clear therefore that DES (and indeed hybrid RANS-LES methods in general)
place conflicting demands on the convection scheme. To address this problem, Travin
et al. [172] proposed a form of locally-adaptive flux blending known as the hybrid
convection scheme. Similarly to standard flux blending (Section 4.2.2), a weighting
parameter 0 ≤ σ ≤ 1 between upwind-based and central-based convection is ap-
plied to determine the the cell face values φ f :

φ f = (1 − σ) φ f ,CDS + σ φ f ,TVD . (5.1)

In the ELAN implementation, blending is carried out between the TVD (σ = 1) and
CDS (σ = 0) schemes, whereas the authors of the hybrid scheme blend between
a third/fifth order upwind and fourth order central scheme. The local value of σ
depends on Δ, νt , S ∗ , Ω∗ and the convective time scale of the simulation τ = re f /ure f
and its formulation is given in Eq. (5.2)1 .
1
Some typographical errors for the formulation of l turb and C H3 were present in the original publica-

81
5 Implementation of detached-eddy simulation

   
CDES Δ
σ = σmax tanh AC H1 , A = CH2 max − 0.5 ; 0 ,
lturb g
"  
νt + ν S ∗2 + Ω∗2 0.1
lturb = , K = max ; ,
Cμ3/2 K 2 τ

CH3 Ω∗ max (S ∗ ; Ω∗ )
g = tanh B4 , B=  ∗2  ,
S + Ω ∗2
max ; 10−20
2

σmax = 1 , CH1 = 3 , CH2 = 1 , CH3 = 2 . (5.2)

This empirically-formulated blending function applies CDS to fine grid regions with
resolved turbulent content (sensed as higher vorticity and lower strain), and UDS to
coarser grid regions or regions of irrotational flow. This has been demonstrated in a
comprehensive parameter study of the scheme conducted by Bunge [19]. As such,
CDS is only applied to regions where LES is possible and desirable, with TVD used
elsewhere. The importance of applying CDS for LES is demonstrated in Section 7.1
and the correct functionality of the hybrid convection scheme is demonstrated for
DES computations in Section 7.2 (also reported by Mockett et al. [97]). In a further
confirmation of the scheme’s functionality, near-identical spectra were obtained us-
ing the hybrid scheme in comparison to CDS for calculations of decaying isotropic
turbulence (DIT) [19, 21].

5.5 Model-specific implementation details

5.5.1 Calibration of the DES model parameter CDES
The parameter CDES has been calibrated for each model using DIT. The method is
described in Sect. 6.1 and some example calibration spectra are given with an accom-
panying analysis in Section 7.3. The calibrated and implemented values of CDES are
listed in Tab. 5.3 for each model.

SAE SALSA WCX LLR CEASM

CDES 0.65 0.60 0.70 0.75 0.65

Table 5.3: Calibrated values of the CDES parameter for each model. (In the case of WCX-DES
the value for the standard length scale substitution in the destruction term of the k
equation is given.)

tion [164], and the correct formulation is presented here.

82
 5.5 Model-specific implementation details

5.5.2 Ψ functions
The problem of the unwanted activation of RANS-model low-Reynolds number te-
rms in LES mode has been described in Chapter 3 together with the remedy proposed
by Spalart et al. [158] in the form of the Ψ correction function. The necessity of such
a function depends on the specific formulation of the RANS model in question. For
those applied in this study, such an analysis has been carried out and is detailed in
Appendix D. The analysis and derivation is based upon the sub-grid scale form of
the model and its comparison to the Smagorinsky model, for which a coefficient A is
derived. If A is constant, no correction term is necessary and Ψ = 1 can be applied.
If however A is not constant, a function Ψ must be sought to correct this. The coeffi-
cients A and correction functions Ψ derived for each considered model are listed in
Tab. 5.4.

For the LLR and CEASM models, the complexity of the expression A is such that the
derivation of a corresponding Ψ remains elusive. The implications of this are anal-
ysed in Sect. 7.4, and a demonstration of the functionality of the derived corrections
is given.

5.5.3 Alternative DES length scale substitutions

Some alternative means of substituting the DES length scale, LDES , in different ex-
pressions of the background RANS model have been investigated. For this, three al-
ternative formulations of WCX-DES are implemented, which are referred to as DES1,
DES2 and DES3 depending on the terms in which LDES is substituted (described in
Section 8.3 and by Yan et al. [186]). As part of the investigation, the values of CDES
were calibrated for each variant using DIT and the implemented values are sum-
marised in Tab. 5.5.

5.5.4 Specific features of the CEASM-DES implementation

The DES implementation for the CEASM model exhibits some particular features
that warrant further description. The first of these concerns the choice of RANS
model length scales to be substituted by LDES , and the second is a model-specific
shield formulation for the prevention of grid-induced separation (GIS – Section 3.7.3).
Details of this formulation have also been published by Bunge et al. [21].

The formulation of the LL k − ε model [81], the background model used in the
CEASM, is given in Appendix B. The LL k − ε model is based on the idea of two-
layer models [127] and as such includes the wall normal distance dw in addition to
the RANS model length scale given in Tab. 5.2. The wall normal distance appears in
the damping terms f μ and C̃ε1 and is hence not substituted by LDES in order to avoid
the erroneous activation of these terms in the LES-mode region.

The availability of the wall-normal distance in the model furthermore allows the
formulation of a GIS-shield following the methodology introduced by Menter &

83
 Model A Ψ2
 
Cb1 f v1 S̃ 1 f C
SA min 102 ; − ∗ v2 2b1
Cw1 f w S∗ f v1 f f κ C
w v1 w1
⎧   −1 ⎫
Cb1 f v1 S̃ ⎨ ⎬
max( f v1 ; 10−10 )
SAE min 102 ; + f v1
2
Cw1 f w S ∗ ⎩ max(χ ; 10−10 ) ⎭
⎧   −1 ⎫
f Cb1 f v1 S̃ ⎨ max( f v1 ; 10−10 ) ⎬
SALSA min 102 ; + f v1
2
f Cw1 f w S ∗ ⎩ max(χ ; 10−10 ) ⎭
 
5 Implementation of detached-eddy simulation

Cω2 3/2
WCX 1
Cω1
⎡  ⎤2/3
f1 f Cμ S ∗
LLR ⎣ f C1 − ⎦ —
f 3/2 f βω ω
Cμ
 3/2 ⎧  3/2 ⎫
C̃ε2 f μ Cμ ⎨ Cε2 C̃ε1 ⎬
LL min 102 ;
C̃ε1 ⎩ max( f μ ; 10−10 ) C̃ε2 Cε1 ⎭
 3/2
C̃ε2 f μ f Cμ
CEASM 1
C̃ε1
Table 5.4: Overview of coefficients of the Smagorinsky form, A, and the implemented low-Reynolds number correction functions, Ψ.

84
 5.5 Model-specific implementation details

DES1 DES2 DES3

Calibrated CDES 0.70 0.95 0.90

Table 5.5: Calibrated values of CDES for the alternative length scale substitution formulations
of WCX-DES (Section 8.3, [186]).

Kunz [92]. The GIS-shield employs exclusively variables and functions already pre-
sent in the background RANS model and is formulated as follows:

0 #  $1
1 1
k = k3/2 max ; 1 − tanh A2L−ν
LRANS CDES Δ
 
k3/2 ν k
A L−ν = max 2 ; . (5.3)
dw ε d2w ε

This shield function is validated in Section 7.5 for a simple flat plate boundary layer
flow. The development of the function was shortly followed by the publication of
the generally-applicable DDES method with an equivalent functionality ([158], de-
scribed in Section 3.7.4). As a result, the CEASM-DDES formulation has since su-
perceded that of Eq. (5.3) in the ELAN implementation.

5.5.5 IDDES implementation

The IDDES method [170, 149], described in Section 3.7.5, has been implemented on
the basis of the SAE and CEASM methods as indicated in Tab. 5.2.

The modified formulation of the filter width Δ, Eq. (3.23), requires the wall-normal
grid spacing hwn and hence information at every grid cell as to which grid index di-
rection is the closest to being normal to the nearest wall. To avoid case-specific and
tedious user input a sensor has been implemented that compares the wall normal
distance of each cell vertex. In this manner, the wall-normal index direction is de-
termined as that for which the vertices exhibit the greatest difference in wall-normal
distance.

Additionally, the IDDES blending function f restore contains two model-specific pa-
rameters Cl and Ct . These must ideally be tuned for each model such that f amp is
close to zero when either f l or f t are close to one. These parameters affect the model
performance only in the limit of a pure RANS solution, where f restore = 0 would
distort the RANS model behaviour. As such, for the purposes of initial testing the
values published by Travin et al. [170, 149] for the Spalart–Allmaras and Menter SST
models have been employed for the SAE and CEASM models respectively. For the
former case, it is not expected that the published values will be too incorrect as the
models are highly similar. In the latter case however it is doubtful that the published

85
5 Implementation of detached-eddy simulation

values would be appropriate. Therefore, to consider the implementation complete

and to ensure that the model performs correctly in the full RANS limit, Cl and Cd
must be specifically tuned for the SAE and CEASM model implementations.

86
6 Description of test cases

In this final chapter of the methodology, the test cases employed in the investigations
of Part III will be described. Collating the test case descriptions in a dedicated chap-
ter is considered preferential to a grouping of descriptions together with the results,
as the investigations are structured thematically and results from some cases are em-
ployed in multiple sections.

For each test case, an overview of the flow-physical features will be given, together
with a description of the benchmark data and relevant references. The important
numerical parameters, grids and setup issues will also be reported, together with
an overview of any special post-processing techniques. Where relevant, any known
limitations on the comparability of the benchmark data with simulations will also be
reported. Finally, a list of publications where the results from each test case can be
found will be given.

6.1 Decay of isotropic turbulence

Decaying isotropic turbulence (DIT) represents mathematically the simplest conceiv-
able realisation of turbulent flow, and is therefore a fundamental test case for new
turbulence modelling or simulation techniques. In DES, a RANS model is modified
to provide LES functionality and this test case is central to the assessment of the LES
capability of the implementation. The motivation of this case can be summarised as
follows:
• DIT represents a basic and readily-verifiable test, at minimal computational
cost, of the capability of the model/solver combination to resolve turbulent
structures and to predict the energy cascade.
• The formulation of DES includes the empirical parameter CDES that must be
calibrated. DIT offers a basis for this.
• The level of numerical dissipation within a given solver can be established. In
cases where this is excessive, steps must be taken to reduce it (Sect. 3.5.2).
Demonstration of these basic LES capabilities and calibration of new model constants
are considered an essential prerequisite to the implementation of a hybrid RANS-LES
method. Although analytical methods may exist to derive the values of subgrid-scale
model parameters (e.g. the expressions for CDES as a function of CS and other RANS
model parameters derived in Appendix D and discussed in Sect. 8.2), these can in
some cases be self-contradictory (as discussed in Sect. 8.3 and [186]). Furthermore,
such derivations do not take numerical dissipation into account.

87
6 Description of test cases

Figure 6.1: Visualisation of the decaying isotropic turbulence test case (velocity vectors are
shown on the surface of the 3D, periodic grid). The data is taken from the DNS
of Jiménez et al. [69] re-sampled to 643 control volumes.

In addition to the primary goals of demonstrating LES capability and calibrating

parameters, a variety of secondary uses can be exploited. These include the demon-
stration of the functionality of model features (e.g. the derived Ψ functions, Sect. 7.4)
and a measure of the dissipation of the numerical scheme using a “no-model” test
(Sect. 7.1).

Despite, or perhaps owing to its simplicity, real isotropic turbulence is seldom found
in nature or practical applications. It is also a very difficult flow state to approximate
in a wind tunnel. Nonetheless, its usefulness as a computationally-inexpensive and
canonical turbulent flow for the purposes stated above is not diminished. Investiga-
tions based on this test case have been published in Bunge et al., 2003 [20], Bunge et
al., 2007 [21], Haase et al., 2006 [55], Haase et al., 2009 [56], Michel et al., 2007 [93],
Mockett & Thiele, 2007 [99], Mockett et al., 2008 [96] and Yan et al., 2005 [186].

Flow description
The isotropy of the flow means that statistical quantities are invariant to coordinate
system rotations, and as a consequence the statistical field is also homogeneous (in-
variant to a shift in the coordinate system). This flow is hence often referred to as
decaying isotropic homogeneous turbulence, although the adjective “homogeneous”
is redundant. The isotropy also eliminates the existence of any mean flow shear, the
mechanism by which turbulence is produced1. As a result of the lack of energy in-
put combined with the dissipative action of the turbulent cascade (Section 2.2), the
energy level decays with time.
∂Ui
1
The production term of the turbulence kinetic energy equation is Pk = − ui u j ∂xj .

88
 6.1 Decay of isotropic turbulence

Benchmark data
A range of alternative benchmark data is available for DIT, with various advantages
and disadvantages. The experiment by Comte-Bellot & Corrsin [30], referred to as
CBC in this work, was conducted at a Reynolds number too low for an inertial sub-
range to exist in the energy spectra (Sect. 2.3). Experiments also have the general dis-
advantage that the three-dimensional, unsteady velocity field cannot be measured.
As a result, the initialisation velocity field for the simulation must be generated us-
ing an inverse Fourier transform of the upstream energy spectra. Although exhibit-
ing the correct spectral distribution, such a velocity field does not contain physically
valid spatial correlations. This disadvantage is alleviated by using DNS data as a
benchmark, such as that of Wray [183, 69]. The CBC data was employed in earlier
DIT investigations with the Wray DNS data adopted for later studies. No significant
difference was observed in the basic spectral analysis that would lead to a change in
the calibrated values of CDES .

Energy spectra are available for the CBC experiment at stations corresponding to the
non-dimensional times of 0, 0.87 and 2.0. Spectra for the DNS data are available at
many different time steps, however those at the non-dimensional times of 0.19938,
1.79940 and 3.03694 were chosen for comparison, as these correspond closest to val-
ues rounded to the nearest 0.1 units.

Grids and numerical setup

The simulation of DIT is conducted in a cubic computational domain with three pairs
of periodic boundary conditions in each spatial direction, to reflect the homogeneity
of the flow. In experiments, the turbulence is generated at an upstream station by a
grid, and the measurement reference frame moves with the mean convective veloc-
ity downstream to capture the decay of the turbulence. A coordinate transformation
from space to time is conducted for the simulations, such that the computational box
is considered to be transported with the mean velocity. The experimentally spatial
decay of turbulence is therefore tracked temporally in the simulation; the initial so-
lution field is set with a suitable instance of isotropic turbulence, the decay of which
is then solved in an unsteady manner.

The grids employed were all equidistant and cubic, with external dimensions of
2π and varying grid resolutions of 163 , 323 and 643 cells. The lowest and highest
resolved wavenumbers2 corresponding respectively to the largest and smallest re-
solved scales are therefore 1 and Nx /2. The time step was 0.01 non-dimensional
units in each case, which was found to deliver CFL  1 throughout the domain and
at all time steps. For all DES simulations, pure LES mode of the model was enforced
by setting LDES = LLES and the central difference convection scheme (Sect. 4.2.2) was
applied unless stated otherwise.

2
κ = 2π/x, Sect. 2.3.

89
6 Description of test cases

Field initialisation
The velocity field for the solution initialisation was obtained from the benchmark
data, either by inverse Fourier transform of the CBC experimental spectra or by trun-
cation of the DNS field to the desired grid resolution in the case of the Wray data3 .
For the inverse Fourier transform, a tool written by St. Petersburg Technical Univer-
sity (SPTU) [148] was provided in the framework of the FLOMANIA and DESider
projects. To obtain the remaining solution quantities (pressure and turbulence model
parameters), a steady-state computation of “frozen turbulence” was carried out by
holding the velocity field fixed. In this manner, the initial subgrid eddy viscosity is
obtained by the model to be calibrated and this step must therefore be repeated for
every model variation. In earlier work [20, 19], the initial eddy viscosity was ob-
tained using the explicit Smagorinsky model. This however has the disadvantage
that the calibration result depends on the value of CS chosen, which is undesirable.
Following convergence of the frozen turbulence calculation, a restart file is produced
that can be used to initialise the unsteady decay computation.

Post processing
The complete velocity field was output at selected time steps corresponding to the
non-dimensional times of 0.87 and 2.0 in the case of the CBC-based computations,
and 0.2, 1.8 and 3.0 for the Wray DNS-based computations. Turbulence energy spec-
tra as a function of wavenumber, E (κ ), were computed from these using the SPTU
three-dimensional Fourier transform tool mentioned above.

6.2 Fully developed turbulent channel flow

The fully-developed turbulent channel flow is a very useful canonical test case for
wall-bounded flows. The test case is employed in Part III for the examination and
validation of WMLES using the IDDES method. Such results from this test case have
been published in Mockett & Thiele, 2007 [99], Mockett et al., 2008 [96] and Haase et
al., 2009 [56].

Geometry and flow conditions

The physical background of fully-developed turbulent channel flow has been sum-
marised in Section 2.4. The fully-developed flow between two parallel flat plates of
infinite extent in the streamwise (x) and spanwise (z) directions is considered. The
channel half-height in the perpendicular (y) direction is denoted δ, by which all spa-
tial coordinates are normalised. The coordinate origin is placed at the lower wall in
3
It has since been noted that this simple truncation gives rise to an aliasing effect, whereby spurious energy
is superimposed on the highest resolved wavenumbers. In very recent work, this has been countered by
employing a suitable filtering operation, however all DIT based on the Wray benchmark data presented
in this work is affected by aliasing. Thankfully, this was not seen to modify any of the conclusions and it
was not considered necessary to repeat the computations.

90
 6.2 Fully developed turbulent channel flow

the normal direction, such that the upper wall is at y = 2δ. In the streamwise direc-
tion the origin is at the entry of the numerical domain.

The flow is driven by a streamwise pressure gradient dp/dx, which is balanced by

the wall shear stress, τw . Consideration of this force balance gives the following
relationship:
dp τw
= . (6.1)
dx δ
The test case is defined such that τw = 1 and δ = 1, as a consequence of which
the pressure gradient dp/dx = 1 is applied. The density is specified as ρ = 1, as a
result of which the wall-friction velocity (Section 2.4) is U τ = 1. The desired skin
friction Reynolds number, Reτ , is hence applied by setting the molecular viscosity to
ν = 1/Reτ .

Benchmark data
For lower Reynolds number cases, DNS data is considered preferential to experimen-
tal measurements as the latter include many sources of uncertainty (e.g. the influence
of side walls and limited aspect ratio, intrusive measurement techniques and the dif-
ficulty of obtaining well-resolved data very near to the wall). The DNS database of
Moser, Kim & Mansour, 1999 [100] is employed for Reτ = 395 and Reτ = 590, and the
more recent DNS of Hoyas & Jiménez, 2006 [63] is used for Reτ = 2000. For higher
Reynolds numbers for which no DNS data are available, the empirically-tuned cor-
relations described in Section 2.4.2 are used.

Grids and numerical setup

The computational domain size specified for the mandatory grid in the DESider
project channel flow test case [56] has been used4 , which extends to L x = 6.4δ
and L z = 3.2δ, respectively. This slightly exceeds the typical values used for fully-
resolved LES quoted in Section 3.5.

Two families of grids have been applied, as depicted in Fig. 6.2; one set of wall-
refined grids with a stretched point distribution applied to the wall-normal cells,
and an equidistant, cubic grid. For the wall-refined grids the tangential point dis-
tributions of the DESider project grid have been used with 64 equidistant cells in
both directions. This results in a resolution of Δx = 0.1δ and Δz = 0.05δ, the
streamwise-elongated cells conforming to standard LES practice in cases where the
mean flow direction is known (Section 3.5). The wall-normal resolution has been ad-
justed to deliver Δy+ ≈ 1 at the wall for each Reynolds number with low stretching
ratios between 1.137 ≤ k ≤ 1.150. The cubic grid by contrast has a resolution of
Δx = Δy = Δz = 0.05δ, whereby the halving of the streamwise grid cell size with
respect to the wall-refined grids has been introduced to represent more general sit-
4
The channel flow test case within DESider was coordinated by L. Davidson of Chalmers University.

91
6 Description of test cases

Z X

(a) WR-4000
Y

Z X

(b) C-4000

Figure 6.2: Cross-section through the x/y plane of a wall-refined (WR) and cubic (C) grid.

uations in which the mean flow direction is not known at the grid generation stage.
The grid properties together with the computational case names are summarised in
Tab. 6.1.

The values of the grid cell sizes in wall units are determined from the distance be-
tween neighbouring cell vertices. Because a cell-centred solver is used, the distance
of the first cell centre to the wall is in fact half of the Δy+ | y=0 value quoted in Tab. 6.1.
It is unclear however, whether the requirement of Δy+ | y=0 ≤ 0.5 should be applied
to the nearest cell centre to the wall in such solvers. To investigate this, a refined
version of the WR-18000 grid has been generated that conforms to this requirement
and no effect on the results was observed.

Periodic boundary conditions are applied in the streamwise and spanwise directions,
with no-slip walls (as described in Section 4.4) at the channel plates. The cubic grid is
intended to investigate the applicability of the hybrid-adaptive wall boundary con-
dition ([135, 145] and Appendix C) in combination with WMLES simulations. The
driving pressure gradient is applied in the form of a corresponding pressure dif-
ference between the periodic boundaries in the streamwise direction. A constant
pressure difference is superimposed on the pressure fluctuations arising from the re-
solved turbulence of the simulation.

For the numerical convection scheme, pure CDS (Section 4.2.2) was applied in all
cases. The time step size was specified with the goal of delivering CFL < 1 through-

92
 Case: Reτ : Nx : Ny : Nz : Nxyz : Δx + : Δz+ : Δy+ | y=0 : Δy+ | y=δ : k:
WR-395 395 64 50 64 204800 39.5 19.75 2.25 50 1.138
WR-4000 4000 64 80 64 327680 400 200 2.25 524 1.150
WR-18000 18000 64 108 64 442368 1800 900 2.25 2193 1.139
C-590 590 128 40 64 327680 29.5 29.5 29.5 29.5 1.000

6.2 Fully developed turbulent channel flow

C-4000 4000 128 40 64 327680 200 200 200 200 1.000
C-18000 18000 128 40 64 327680 900 900 900 900 1.000
C-100000 100000 128 40 64 327680 5000 5000 5000 5000 1.000

Table 6.1: Summary of channel flow test cases and corresponding grid properties.
93
6 Description of test cases

out the domain5 . The time step required was estimated based on the centreline ve-
locity and the streamwise grid spacing including a safety factor. The approximate
expressions of Pope [116], Eqs. (2.24) and (2.25), were used to obtain a value for
the expected mean centreline velocity at each Reτ . Following the establishment of a
fully-developed flow state, the CFL number was inspected to ensure that the targets
had been met, and in some cases adjusted accordingly.

Field initialisation

The specification of the initial field is of central importance in LES and WMLES com-
putations of channel flow. If a smooth mean flow profile is used without any spatial
fluctuations, a quasi-laminar profile arises in the case of LES [144] or a RANS solution
results from the IDDES method [170, 149]. The superposition of simple white noise
has been demonstrated by many investigators to be inadequate (e.g. by Schmidt,
2000 [145]), as the fluctuations decay too rapidly.

The approach used to initialise the channel flow simulations in this work (in the ab-
sence of a suitable restart field from a previous simulation) involves a combination of
the mean velocity profile from a two-dimensional precursor RANS simulation with
fluctuations derived from isotropic turbulence. The isotropic turbulence data was
obtained from the 643 truncated field of the Wray DNS simulation [183], which was
used to initialise the DIT simulations (described in Section 6.1). By mapping the ho-
mogeneous turbulence from its equidistant Cartesian grid to the wall-refined chan-
nel grids, a qualitative similarity to the real flow is achieved with stretched vortices
near the wall and near-spherical vortices in the core flow. This method proved satis-
factory as a source of sustainable velocity fluctuations.

Post processing

Following the field initialisation or a simulation restart with a modified setup, the
temporal development of U τ and the bulk velocity, U b , were monitored throughout
each computation. Statistical analysis was only conducted following the establish-
ment of a fully-developed flow state with the initial transient behaviour eliminated.

The entire field was output at regular time intervals, which could be selected for
post processing following examination of the temporal evolution described above.
Reynolds averaging (Section 3.4.1) with respect to time and both homogeneous di-
rections was carried out using a purpose-written Fortran tool6 . In addition, the in-
stantaneous WMLES blending functions of the IDDES calculations were output at an
arbitrary wall-normal location to analyse the functionality of the models.

5
The CFL number is discussed in Section 8.6.
6
Based on that written by A. Carnarius, TU-Berlin.

94
 6.3 NACA0012 airfoil in deep stall

Limitations on comparability with benchmark data

No limitations on the comparability with the DNS benchmark data are anticipated,
as the benchmark DNS computations all exhibit sufficient spatial resolution and sta-
tistical convergence. The DNS computations are similarly carried out using a peri-
odic domain in the streamwise and spanwise direction meaning that no geometric
uncertainties arise. The Reichardt correlation [118], applied for the higher Reynolds
numbers and described in Section 2.4, does however include some uncertainty. The
parameters of the expression were calibrated based on what must by now be consid-
ered as outdated experimental data. The Reichardt correlation can therefore only be
interpreted as a guideline for approximate comparison. The same reservation applies
to the empirical correlations of Dean [34] and Pope [116]; as mentioned in Section 2.4,
these can only be considered as accurate as the experimental data upon which they
are based and considerable scatter is apparent.

6.3 NACA0012 airfoil in deep stall

The first three-dimensional test of the DES method was conducted on the flow ar-
ound a symmetric NACA0012 airfoil profile at very high angles of attack [148]. Due
to the apparently promising improvement in the results and relatively low numer-
ical cost, this test case subsequently became the focus of cross-validation efforts for
the implementation of DES in many CFD solvers. A good example of such use was
the European FLOMANIA project [55], in which the first DES implementations in
ELAN were also compared [19, 94]. The time-averaged lift and drag coefficients
were obtained in experiments published by Hoerner [62]. No unsteady information
is available experimentally at such high angles of attack, and important details such
as the aspect ratio of the airfoil in the experiment are unknown.

However, from the evidence of subsequent studies in the DESider project [56] (which
were based on the initial investigation of Guenot [54]), it appeared as if the absolute
agreement with experimental force values was due to a “cancellation of errors” effect
arising from an insufficient spanwise simulation domain. This effect is described in
detail in Part III.

Despite this lack of comparability with experiment, the results obtained from this test
case allow some important insight into the basic LES/RANS activity of DES solutions
(Sect. 8.1.2) and a comparison of different DES length scale substitutions (Sect. 8.3).

The symmetric NACA0012 airfoil exhibits a 12% thickness-to-chord ratio, and has
been studied here at the high angle of attack α = 60◦ . A depiction of the geome-
try and coordinate system (with origin at the profile nose) is shown in Fig. 6.3. The
incompressible flow around the profile is of Reynolds number Rec = |U | ∞ c/ν =
1 × 105 . All quantities reported are normalised by the chord length, c, and the free
stream velocity magnitude |U | ∞ .

95
6 Description of test cases

Z X

Figure 6.3: Schematic of geometry and coordinate system for NACA0012 test case.

(a) Coarse (b) Medium, fine

Figure 6.4: Near field views of numerical grids used for the NACA0012 test case (slices normal
to the spanwise, z, direction). The medium and fine grids differ only in the spanwise
resolution.

The numerical grids used were provided by M. Shur and M. Strelets of St. Petersburg
Technical University (SPTU) to the partners of the FLOMANIA project. Three refine-
ment levels were computed, with 201600, 268800 and 324800 control volumes on the
coarse, medium and fine grids respectively. All grids were of the same O-topology,
and exhibit a uniform distribution of approximately cubic cells in the near wake “fo-
cus region” [154]. Detailed information on the grids is given in Tab. 6.3, and a visual
comparison is depicted in Fig. 6.4.

The far field domain boundary is circular, and situated at a distance of Lr /2 from
the airfoil. Depending on the angle of attack, an appropriate segment of arc length
around 60◦ is allocated as an inflow boundary condition, and the remainder of the
far field boundary is designated outflow. The surface of the airfoil is treated as a no-
slip wall boundary, and the grid resolution is sufficient for full resolution to y+ = 1
everywhere. The boundary conditions in the spanwise direction are periodic. For all

96
 6.4 NACA0021 airfoil in deep stall

Grid Nr Nθ Nz Ntot Lr Lz Δz
Coarse 60 140 24 201600 30c 1c 0.0417c
Medium 80 140 24 268800 30c 1c 0.0417c
Fine 80 140 29 324800 30c 1c 0.0345c

Table 6.2: Details of the SPTU grids for the NACA0012 test case. Nr , Nθ and Nz respectively
denote the cell numbers in the radial, circumferential and spanwise directions, and
the L values are the total numerical domain size in these directions (i.e. distance from
boundary to boundary, not from geometry to boundary).

computations, the numerical time step size was set to 0.025c/|U∞ |.

6.4 NACA0021 airfoil in deep stall

Following the initial studies on the NACA0012 airfoil detailed above, a search was
conducted for a set of experimental data that included unsteady quantities. No such
data was found, however a detailed experimental campaign by Swalwell et al. [167,
166] was found for the thicker NACA0021 profile at a slightly higher Reynolds num-
ber of Rec = 2.7 × 105 . This data was collected as part of an investigation of wind
turbine blade stall, and involves long time series of force component data obtained
from arrays of surface pressure sensors scanned at a high sampling rate. It was there-
fore decided to adopt this test case instead of the NACA0012 for the extensive com-
putations conducted in the European DESider project [56]. Results for this test case
have been published in Braza et al. [56], Bunge et al., 2007 [21], Michel et al. [93],
Mockett & Thiele, 2007 [97], Mockett et al., 2008 [96], Swalwell [166] and Weinman
et al. [179] (although in the last publication a different set of grids was employed).

Apart from the profile thickness and Reynolds number, essentially the same investi-
gation was carried out as for the NACA0012 profile, so only information that differs
from the NACA0012 case will be given here. In contrast to the FLOMANIA inves-
tigation, only the deep stall attitude of α = 60◦ was studied. The geometry and
coordinate system of the NACA0021 profile is shown in Fig. 6.5.

The computational grid was again prepared by the St. Petersburg group of Profs.
Shur and Strelets, and was based on the same O-topology construction as for the
NACA0012. Results from a grid with a spanwise domain extent of 1c are reported
here. Following the findings of Guenot [54] that a spanwise domain extent of 1c was
insufficient for the NACA0012 test case at α = 45◦ , no comparison with experimental
data is presented7 . Details of the grid computed with the ELAN solver are given in
Tab. 6.4. A two-dimensional view of the grid is given in Fig. 6.6.

7
A grid with an extended spanwise domain of 3.24c was also computed within the DESider project [?].

97
6 Description of test cases

Z X

Figure 6.5: Schematic of geometry and coordinate system for NACA0021 test case.

Grid Nr Nθ Nz Ntot Lr Lz Δz
NTS 100 140 34 476000 30c 1c 0.0294c

Table 6.3: Details of the NTS grid for the NACA0021 test case. The quantities have the same
meaning as those in Tab. 6.3.

Figure 6.6: Near field views of the numerical grid used for the NACA0021 test case (slice nor-
mal to the spanwise, z, direction).

6.5 Circular cylinder in a square channel

As described in Section 2.5, circular cylinders have long served as canonical test cases
for bluff body aerodynamics. The highly unsteady, massively separated turbulent
wake that occurs at high Reynolds numbers represents a significant challenge to CFD
methods. The particular case computed in this study has been selected for two prin-
cipal reasons. Firstly, very detailed unsteady experimental data are available, which
have been obtained using state of the art non-intrusive techniques. Secondly, the rel-
atively confined experimental domain can be directly reproduced in the simulation,
thereby eliminating some of the typical sources of uncertainty in CFD–experimental

98
 6.5 Circular cylinder in a square channel

comparison. Indeed, the provision of a useful database for the validation of numer-
ical methods was one of the key aims of the experiment, and the test case has been
employed as such during the European FLOMANIA [55] and DESider [56] projects.

The test case is therefore employed for the validation of DES working at its “design
condition”, namely massively separated bluff body flows, as is presented in Sec-
tion 8.5. An investigation of time step sensitivity, Section 8.6, is also based on this
case. Published results include a validation of the numerical method by Mockett et
al. [97, 98] as well as in the framework of a joint numerical and experimental study
of the flow physics by Perrin et al. [110, 112]. Other publications employing these
results less centrally in an overview role include Michel et al., 2007 [93], Mockett
& Thiele, 2007 [97], Mockett et al., 2008 [96], as well as the recent review paper of
Spalart [156].

Geometry and flow conditions

A visual impression of the cylinder and channel geometry can be obtained from
Fig. 6.7, in which the axis system for the test case is also depicted. The cylinder is
enclosed within the square channel, which it traverses symmetrically at the channel
mid height. As a result of the confining channel, the cylinder aspect ratio (span /
diameter) is fairly low at 4.7 and the blockage coefficient (cylinder diameter / chan-
nel height) is relatively high at 0.208. This arrangement is in stark contrast with the
majority of experimental cylinder investigations, in which a simulation of “infinite
conditions” is targeted using a low blockage coefficient, a high aspect ratio and often
elaborate end-plates at the cylinder fixture to the wind tunnel side walls. The an-
ticipated effects of these geometric constraints are described in the literature review,
Section 2.5.2.

The channel is 22D long and situated in an open-jet wind tunnel. The channel inflow
is located 7D upstream of the cylinder, and measurements have confirmed that the
velocity of the open jet is uniform across the channel inlet with a turbulence intensity
of Tu = 1.5% [108]. The coordinate system has its origin at the centre of the cylinder
at the mid span position.
The Reynolds number of 140000 is realised with respect to the inlet velocity u ∞ =
15m/s, cylinder diameter of D = 0.14m and kinematic viscosity of ν = 1.5 × 10−5
m2 /s. Through experimental observation of the drag coefficient variation with Re
[108], the flow has been seen to exhibit behaviour corresponding to the precritical
regime, TrBL0, at the beginning of the drag crisis (described in Section 2.5.2). This oc-
curs at a lower Re than that quoted in the literature, which is due to the effects of the
inflow turbulence intensity and the channel blockage (as described in Section 2.5.2).

Experimental data
The principal experiments of this cylinder flow were conducted at the Institut de
Mécanique des Fluides de Toulouse (IMFT) [108, 109, 110, 111] with some accompa-

99
6 Description of test cases

Figure 6.7: Geometry of the cylinder flow domain including representation of the PIV measure-
ment planes.

nying earlier studies carried out at the Institut de Recherche de l’Ecole Navale de
Brest (IRENav) [37] on an equivalent configuration in a water tunnel. The following
data were acquired:
• Unsteady pressure measurements at a number of locations around the circum-
ference of the cylinder at the mid-span plane.

• Two-component (2C) particle image velocimetry (PIV) measurements at the

planes shown in Fig. 6.7, providing uncorrelated snapshots of the u and v ve-
locity field.

• Three-component (3C) stereoscopic PIV at the same measurement planes pro-

viding uncorrelated snapshots of u, v and w.

• 3C time-resolved PIV (TRPIV) providing the u, v and w fields with a high sam-
pling rate of 1kHz (giving a maximum resolved frequency of St = f D/2u ∞ =
4.67 observing the Nyquist condition [104]).

• Time-resolved single point laser Doppler anemometry (LDA) measurements

with a high sampling rate at discrete points within the PIV planes (from the
IRENav investigation).
An excellent agreement between the TRPIV and LDV spectra as well as between
the TRPIV and low-frequency PIV fields was demonstrated [110], which suggests
a high degree of fidelity of the measurements. For a more detailed report of the
experimental apparatus, the quoted references should be consulted.

Grid and numerical setup

The entire experimental domain, including the cylinder and all four walls of the sur-
rounding channel have been directly reproduced in the numerical domain, shown in

100
 6.5 Circular cylinder in a square channel

Fig. 6.8. No-slip boundary conditions have been applied on all walls, and the hybrid
adaptive wall boundary condition described in Section 3.4.4 has been exploited to
save grid points by using a large value of y+ on the less important channel bound-
ary layers whilst fully-resolving the cylinder boundary layer with y+
max ≤ 1.

(a) Complete domain, every second point omitted.

(b) Zoom of cylinder region, every sec-

ond point omitted.

Figure 6.8: Flow domain and numerical grid for the cylinder in a square channel.

The resulting grid consists of some 5 million volumes, with a high circumferential
resolution of 240 uniformly spaced cells applied to the cylinder and 96 in the span-
wise direction. The spanwise spacing is uniform along the central portion of the
cylinder and compressed towards each wall within a region of one diameter from
each end. The spacing of the uniform section has been matched to that of the other
grid directions in the focus region of the near wake to give roughly cubic cells here in
accordance with DES grid guidelines [154]. The grid was generated by J. Yan during
his occupation as a research assistant at ISTA, and was made available to the DESider
consortium [56] for comparative computation within the project.

The inflow plane, situated at the same location as the channel entrance in the exper-
iment, is assumed to be far enough upstream of the cylinder for unsteadiness effects
to be negligible. As such, a steady uniform velocity profile is applied here. It was
found that placing the inflow boundary condition directly at the channel entrance
caused convergence problems. The solution adopted was to add a short inflow sec-
tion of five cells upstream of the channel entrance, to the sides of which a symmetry

101
6 Description of test cases

boundary condition was applied in the y and z directions.

The time step size used for the validation test case is 3 × 10−4 s, which corresponds
to roughly 0.03D/u ∞ or a Strouhal sampling frequency of St = D/(2 Δt u ∞ ) = 15.6
(observing the Nyquist condition [104]). A variation of the time step size was also
studied with the additional value of Δt = 5 × 10−4 s computed. The majority of
the simulations were run using 32 CPUs on the IBM pSeries 690 supercomputer at
the HLRN (see Acknowledgements). Each time step was found to achieve satisfac-
tory numerical convergence of at least two orders of magnitude in the residua after
roughly 12 inner iterations. Following the establishment of a fully-developed flow
state, a minimum physical time of 4.2s or 450D/u ∞ was computed for statistical
analysis in each case.

Data capturing and post processing

Much of the post-processing necessary for the comparison with the experiments was
conducted by R. Perrin, during his post-doctoral research work at ISTA. Because he
was also responsible for the collection and analysis of the experimental data at IMFT,
it can be said that exactly the same post-processing methods were applied to both
data sets, thereby enhancing the directness of the comparison.

The circumferential distribution of mean and RMS pressure coefficient has been ob-
tained for both the simulation and the experiment. The mean experimental drag
coefficient was estimated by integration of the mean pressure on the cylinder at the
mid-span position, assuming homogeneity of the flow on a significant spanwise por-
tion and neglecting the contribution of the viscous forces. In the simulation, the drag
has been calculated both as in the experiment as well as over the entire cylinder
including pressure and friction components8 . An evaluation of the experimental as-
sumptions on the basis of the numerical simulations is therefore possible.

For the PIV data, global averaging has been carried out using about 3000 instanta-
neous fields. These can be considered uncorrelated in time due to the low sampling
rate, and uncertainties have been determined to the order of 2% [110]. Concerning
the simulation, global averaging has been performed on the entire domain during
the calculation, giving access to information about the homogeneity of the flow in
the spanwise direction and wall effects. As is reported in Sect. 8.5, the mean flow
field can be considered uniform in the spanwise direction within a region greater
than 1D from each side wall. To further improve the statistical convergence, addi-
tional averaging of the computational field was conducted in the spanwise direction
within the homogeneous region.

8
The sectional drag coefficient is denoted Cd , whereas the drag coefficient integrated over the entire span is
denoted C D .

102
 6.5 Circular cylinder in a square channel

For the analysis of unsteady flow features, the computed flow field has been stored
every time step on a 2D slice located at the mid-span position z = 0 corresponding
to the PIV and TRPIV measurements. Point velocity spectra have been computed
from the time-resolved experimental and numerical data using Welch’s averaged pe-
riodogram method [180].

The nearly periodic nature of the vortex shedding in the wake enables the use of
phase averaging to characterise the unsteady mean motion and the turbulence quan-
tities. This method allows the classical decomposition of the flow into a mean part,
a quasi-periodic fluctuation and a random fluctuation, which can be written as u i =
Ui + ũ i + u i [122]. The phase-averaged motion is then defined as Ui  = Ui + ũ i .
A more detailed description of phase averaging is given in Appendix F. An ear-
lier technique employed to obtain the phase angle ϕ required to conduct the phase
averaging involved the use of the surface pressure on the cylinder upstream of the
separation point as a trigger signal. Due to the large spatial separation between the
trigger signal point and the wake field, this however gave rise to a residual peri-
odic component in the random fluctuations (known as “phase jitter”). A much more
precise separation of the periodic and random fluctuations was achieved when the
phase angle was defined using the coefficients from a proper orthogonal decompo-
sition (POD) of the velocity field to be averaged, as described in detail by Perrin et
al. [110]. This phase averaging technique was applied to both the numerical (for the
Δt = 3 × 10− 4s case only) and experimental velocity fields, with roughly 170 fields
averaged for each phase angle in the case of the experiment, and 90 for the simu-
lation. From the simulation a two-dimensional slice through the entire domain at
z = 0 was output every time step for the purposes of phase averaging and unsteady
analysis. The simulation therefore doesn’t suffer from the limited spatial coverage of
the PIV windows used in the experiments.

Limitations on the comparability between CFD and experiment

The inlet turbulence intensity of Tu = 1.5% measured in the experiment unfortu-
nately represents a serious limitation on the comparability with simulation9 . The
cylinder flow at such precritical Reynolds number has been demonstrated to be
highly sensitive to free stream turbulence (FST), as is outlined in Sect. 2.5.2, with
strong effects seen at such levels. Furthermore, a measurement of the length scale is
not available, which is necessary for a full description of the FST.

Including FST effects, even with knowledge of the length scale, presents both RANS
and LES simulations with difficulties: In a pure RANS, the problem of a non-physical
decay between the inlet and cylinder gives rise to difficulties in the specification of
suitable inlet values of the turbulence parameters [160]. Having overcome this, the
effect of the FST on the sensitive transition processes would require a well-calibrated
and dedicated model for transition and lies outside the scope of RANS models. For
9
Attention was drawn to this important fact by C. Norberg.

103
6 Description of test cases

LES, the problem is partly one of the specification of physically-realistic resolved

inlet turbulence, and partly one of numerical expense (the inlet fluctuations must be
resolved well all the way to the cylinder). In any case, for the DES computation with
steady-state inlet conditions used here, it cannot be expected that the FST effect is
accounted for correctly.

6.6 Separating-reattaching flow over a bump in a rectangular duct

This test case was conceived within the DESider project [56] with the intention of
providing a particularly challenging combination of flow phenomena:

• Turbulent boundary layer separation from a smoothly curved surface, making

the prediction of the separation location very sensitive to modelling errors.

• A shallow separation, as opposed to the “massive separation” of the bluff body

flows, for which the grey area problem concerning the transition from mod-
elled to resolved turbulence (Sect. 3.7.3) can be expected to be particularly in-
fluential.

• Flow reattachment, the position of which should prove highly sensitive to the
level of predicted turbulent shear stress in the separated shear layer (and hence
the grey area issue).

• The reattachment and recovery region, in which resolved turbulence impinges

upon the wall boundary layer. This would require WMLES capability of the
hybrid RANS-LES model, which is not one of the target applications of natural
DES (Sect. 3.6).

• A high Reynolds number, such that pure LES computations would be unfeasi-
bly expensive.

• A low aspect ratio allowing the computation of the complete domain includ-
ing side walls, hence eliminating problems associated with the experimental
approximation of infinite spanwise conditions.

As such, the test case intentionally casts DES outside of its intended application
range, thereby demonstrating weaknesses in various areas and allowing the effect
of improvements to be assessed. Within this work, the case is employed to investi-
gate the model dependency of DES (Sect. 8.4.2) and to assess the predictive accuracy
of DDES and IDDES in comparison with experimental data (Sect. 8.7). To avoid rep-
etition, the figures concerning this test case are presented separately in Appendix G,
to which both aforementioned sections refer CHECK THIS!. Results from the bump
test case have been published in Haase et al. [56], Mockett et al. [96] and Mockett &
Thiele [99].

104
 6.6 Separating-reattaching flow over a bump in a rectangular duct

Geometry and flow conditions

The duct is rectangular, of height H = 0.28 m and width W = 0.5 m and constricted
to half its height by a bump mounted on the lower wall. The geometry of the bump
was designed by Y. Egorov and F. Menter of Ansys Inc. employing precursor RANS
and URANS computations using their CFX solver, with the goal of achieving the
desired flow phenomena outlined above. An early indication of the high sensitivity
of the flow was given by the strong model dependency of the RANS solutions in
these preliminary calculations [4]. A hyperbolic tangent geometry was decided upon
for the downstream bump curvature, given by
   1/β
Lb − x β
h = Hb tanh tan(α) , (6.2)
Hb
where Hb = 0.138 m is the bump height, L b = 0 is the position of the end of the
bump, α = 35◦ is the slope angle of bump and β = 5 is a shape contour exponent.
The upstream, contracting portion of the bump is a straight ramp with smoothed in-
tersections to the channel floor and the bump plateau. The bump geometry included
in the computational domain is shown together with the coordinate directions in
Fig. 6.9. The coordinate origin is located at the downstream end of the bump in the
spanwise symmetry plane.

Figure 6.9: Geometry of the bump flow domain including representation of the PIV measure-
ment planes.

The fluid employed was water of density ρ = 997 kg/m3 and a dynamic viscosity of
μ = 0.89 × 10−3 Pa s. The Reynolds number based on the bump height and the bulk
velocity above the bump is Reb = 1.044 × 106 .

Experimental data
The experimental data for this test case was collected during a measurement cam-
paign at ONERA in the course of the DESider project. The THALES water tunnel at
ONERA Toulouse was employed, into which the bump model was mounted. The
data acquisition consisted of:

105
6 Description of test cases

• 90 pressure tappings arranged in three parallel lines (along the x-direction) at

z = −0.170 m, z = −0.085 m and z = 0.0 m.

• Unsteady pressure tappings at four locations along the bump and downstream
wall.

• Laser Doppler velocimetry (LDV) measurements across the entire plane in-
tended to represent the entry plane in the computations, x = −0.367 m.

• LDV measurements along the symmetry plane, z = 0.

• PIV measurements at various planes normal to the spanwise direction, de-

picted in Fig. 6.9. These were located in the region of the separation and down-
stream of the reattachment point. The downstream planes were at three span-
wise locations, z = −0.125 m, z = 0.0 m and z = 0.125 m.
The PIV measurement planes at z = ±0.125 m were intended to allow an assessment
of the spanwise homogeneity and symmetry of the flow.

More details on the experimental setup, data acquisition and post processing are
given by Aupoix et al. [4].

Grid and numerical setup

The computational grid was generated by Ansys Inc. and set as a mandatory mesh
for the DESider participants. The grid consists of around 3.9 × 106 volumes, with
220, 120 and 150 cells in the streamwise, vertical and spanwise directions, respec-
tively. The boundary layers are refined appropriately for low-Re boundary condi-
tions (Sect. 3.4.4) and a relatively fine and isotropic grid spacing is evident through-
out the channel around the separated region. The streamwise grid spacing is coars-
ened gradually after around two bump heights downstream of the bump foot in
order to limit the computational expense. A collection of visualisations of the grid
are shown in Fig. 6.10.

The inlet and outlet of the computational domain were located at x = −0.367 m and
x = 2.0 m, respectively. To specify the flow at the inlet, a precursor computation
including the upstream convergent ramp portion of the bump was conducted by
Ansys Inc., using a DRSM. The goal of this was to provide common inlet profiles
for all partners as well as to match the experimental inlet data as closely as possi-
ble. Very good agreement between the Ansys inlet data and the LDV measurements
was achieved, except for the spurious vortical structures apparent in the experiment,
which are discussed later. Some visualisations of the Ansys inlet plane are given in
Fig. 6.11, showing the relatively thin boundary layers (i.e. the flow is far from fully-
developed) as well as the secondary flow predicted by the DRSM. A weak secondary
flow is apparent in the inflow channel core, whereas the corner vortex structures in-
side the boundary layers are strong by comparison. The bulk velocity in the stream-

106
 6.6 Separating-reattaching flow over a bump in a rectangular duct

Z X

(a) z slice in vicinity of bump

X Z

(b) x slice downstream of bump

-0.2 0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

X

(c) z slice through entire flow domain

Figure 6.10: Depictions of the grid at various slices through the domain.

wise plane, upon which the Reynolds number is based, is Ub = 6.751 m/s.

All computations employed the hybrid convection scheme (Sect. 5.4, [172]), and a
time step of either Δt = 8 × 10−4 s or Δt = 5 × 10−4 s was used. The computations
were conducted in parallel on 16 CPUs of the HLRN’s IBM p690 series supercom-
puter. DES computations were initialised either from precursor RANS simulations
or from existing DES solutions. An initial transient period was simulated before
commencing the statistical analysis. Quantities are not non-dimensionalised for this
case, unless explicitly stated.

Data capturing and post processing

The entire flow field was Reynolds averaged, allowing the examination of the mean
field and resolved as well as modelled Reynolds stress data throughout the do-
main. The time-averaged pressure distribution was obtained at the walls together
with pressure time traces at the locations of the experimental unsteady transducers.

107
6 Description of test cases

(a) Contours of U

(b) Cross-flow streamlines

(c) Cross-flow streamlines, zoom in corner region

Figure 6.11: Visualisations of the flow imposed at the domain inlet. Cross-flow streamlines are
integrated using the tangential V and W components only.

To compute the pressure coefficient, the same reference velocity of Ure f = 4 m/s
used in the experiment was applied. The pressure reference point was chosen at
xre f = 0.01 m at the z position currently considered.

108
 6.6 Separating-reattaching flow over a bump in a rectangular duct

Limitations on the comparability between CFD and experiment

The limitations of the comparability between CFD and experiment are based on two
principal sources of uncertainty concerning the experimental data [4]:
• The LDV and PIV measurements at the same locations show some degree of
disagreement in the separated region. This is relatively minor for the U compo-
nent, but unfortunately significant for the vertical V component. The Reynolds
stresses also show relatively strong disagreement, whereby the turbulence lev-
els of the PIV are lower than those given by LDV, particularly in regions of high
turbulence levels. The published account of the experiments places a greater
level of trust in the LDV measurements, stating that this technique will be pur-
sued in future campaigns [4].

• Spurious localised velocity defects coinciding with increased turbulence levels

are evident in the LDV inflow plane measurements. These are believed to rep-
resent streamwise vortical disturbances and occur near the side wall boundary
layer edges, outside of the upper and lower boundary layers. Their origin is not
fully understood, although it may be that they are propagated from upstream
technical installations. The vortices do not appear in the precursor RSM cal-
culation of Ansys, which apart from this otherwise give very good agreement
with the experimental inlet plane.
As a result, only the LDV profiles are plotted in the comparison with CFD, as these
are considered more reliable. It is unclear to what extent the neglected streamwise
vortices in the inflow plane will affect the computational results.

109
6 Description of test cases

110
 Part III

Validation, demonstration and analysis

111
7 Validation of the implementation

In this chapter, results will be presented and discussed that serve to demonstrate key
functionality and correct implementation of certain features of the DES components
of ELAN. The results are based primarily on simple test cases before more complex
flows are considered in Chapter 8.

7.1 Influence of the numerical convection scheme

The basic numerical convection schemes were described in Section 4.2.2, where it was
noted that the choice of scheme has a strong impact on the accuracy and robustness
of the solution. The LES-mode operation of DES corresponds to a classical or explicit
LES (see Section 3.5), for which SGS models are derived under the assumption that
the dissipative behaviour of the fine, unresolved scales is prescribed in its entirety by
the model. It is therefore essential to establish the validity of this assumption for the
underlying numerical implementation in order to demonstrate the fundamental LES
capability of the solver and to establish the most suitable numerical parameters.

One method of investigating this issue is based on the simulation of decaying, isotro-
pic turbulence (DIT). The physics of the turbulent cascade and the energy spectrum
have been introduced in Sections 2.2 and 2.3 respectively, and background informa-
tion concerning the simulation of DIT can be obtained together with implementation
details from Section 6.1. In the usual computation of DIT, the performance of sub-
grid models can be investigated for a particular Reynolds number, for which it is
assumed that the following components constitute the total level of dissipation ε tot :

ε tot = ε SGS + ε ν + ε num , (7.1)

where ε SGS , ε ν and ε num represent the SGS model, molecular and numerical contri-
butions to the dissipation, respectively. In order to isolate ε num , the SGS model is
deactivated and the molecular viscosity is set to zero. The criterion to assess whether
ε num = 0 is obtained from theoretical considerations of the energy cascade process
of inviscid turbulence. In the absence of dissipation and in the presence of a spectral
truncation at high wavenumber (i.e. a finite grid resolution), the flow can obtain a
state of absolute statistical equilibrium with an equipartition of energy over wave-
vectors. The energy spectrum then has the form E (κ ) ∝ κ2 (see e.g. Kraichnan,
1967 [78]). For flow initialised with an arbitrary spectral energy distribution, the κ2
behaviour begins at the highest wavenumbers, progressing to lower wavenumbers
and feeding on their energy content. This behaviour has been clearly demonstrated
in a DNS study by Cichowlas et al., 2005 [29], where an inertial range exhibiting
E (κ ) ∝ κ −5/3 exists at low wavenumbers with the E (κ ) ∝ κ2 at high wavenumbers

113
7 Validation of the implementation

serving as a form of virtual viscous dissipation. Such use of this theoretical result
for a test of the energy conservation of a numerical scheme was proposed by Ben-
hamadouche & Laurence, 2002 [10].

DIT has been computed on the basis of the Wray benchmark DNS data starting from
an initial velocity field that reproduces the turbulent E ∝ κ −5/3 distribution of the
inertial subrange at the highest resolved wavenumbers. The numerical grid of 643
control volumes was used, upon which the initial field was truncated from the 5123
DNS data1 . Fig. 7.1 shows the turbulent energy spectra arising at various time steps
for “no-model” inviscid simulations using the CDS, UDS and TVD schemes.

0,1 0.1 0.1

0,01 0.01 0.01

E(κ)

E(κ)

E(κ)
0,001 DNS 0.001 0.001
ELAN, t=0.2
ELAN, t=1.8
ELAN, t=3.0
2
0,0001
E = Bκ
0.0001 0.0001
1 10 1 10 1 10
κ κ κ
(a) CDS (b) UDS (c) TVD

Figure 7.1: Energy spectra for “no-model” solutions of DIT with different convection schemes,
ν = 0, 643 grid, Wray DNS benchmark data [183].

It can be seen for CDS that the high wavenumbers indeed begin to turn upwards
with time to approach the κ2 gradient, and that this behaviour spreads to succes-
sively lower wavenumbers. It therefore appears as if the numerical dissipation of the
CDS convection scheme is negligible, at least for the orthogonal and equidistant grid
used here. On the basis of this evidence, an indication is given of the numerical suit-
ability of the ELAN solver for LES when CDS is applied. It must be acknowledged
that this test alone falls short of a full demonstration of kinetic energy conservation
of the numerics, for which further tests such as temporal energy conservation for in-
viscid Taylor–Green vortices would be of value [10, 86].

The spectra returned by UDS and TVD show the opposite behaviour – the high
wavenumber motion is damped very strongly by UDS and slightly less so by TVD,
with energy levels much lower than the benchmark data returned. Even though no
SGS model was active in these simulations, it can be seen that ε num is greater than
the ε SGS that an appropriate model would provide. The UDS and TVD convection

1
The problem of aliasing in the initial velocity field, outlined in Sect. 6.1 is of no consequence here.

114
 7.2 Functionality of the hybrid convection scheme

schemes must therefore not be employed in conjunction with LES.

For an impression of the adverse impact of dissipative upwind-biased schemes on

the LES of more complex flows, reference is made to the study of channel flow by
Breuer [17]. A higher-order upwind scheme caused a strong deterioration of the solu-
tion quality in an analogous manner to a strongly excessive value of the Smagorinsky
parameter. Second order CDS in contrast produced good agreement with benchmark
DNS.

Finally, some consideration will be made of cases where weak numerical dissipa-
tion is present, such that the high wavenumber spectral exponent falls short of the
E (κ ) ∝ κ2 but still exceeds κ −5/3 . A good candidate for such behaviour would for
example be flux blending with a small percentage of UDS (Sect. 4.2.2). In such cases,
a reduced model dissipation could in principle be added to match the benchmark
spectral behaviour. Such balancing of numerical and model dissipation could in-
deed prove of use in solver architectures where a weak residual numerical dissipa-
tion cannot be eliminated. The expectation is that such a setup would perform well
in detached flow regions, however problems would be expected in the near wall re-
gion in a pure LES context: The wall damping effect (Sect. 2.4) reduces the turbulent
dissipation to zero and the spurious presence of ε num in addition to ε ν becomes criti-
cal. In the context of DES or WMLES however, the balancing of numerical and model
dissipation could indeed prove to be a viable and pragmatic approach.

7.2 Functionality of the hybrid convection scheme

The hybrid scheme of Travin et al. [172] has been implemented in order to resolve the
conflicting demands placed by the RANS and LES modes on the numerical convec-
tion scheme. The importance of this in particular for the LES mode region has been
demonstrated in Section. 7.1. The motivation of the scheme and its formulation are
described in Section 5.4. The functionality of the hybrid convection scheme for DES
of massively-separated flows is demonstrated using the cylinder and NACA0021
cases.

Figure 7.2 shows the contours of the hybrid scheme’s σ function, which blends be-
tween TVD (σ = 1) and CDS convection (σ = 0), for the cylinder and NACA0021
cases, described respectively in Sects. 6.5 and 6.42 . For both cases the σ function
tracks the unsteady wake regions of resolved turbulence very closely (visible in the
contours of vorticity magnitude on the left-hand side), ensuring a dominance of low-
dissipative CDS convection there. In the irrotational flow outside of the wake and
further from the cylinder, the more stable TVD is ensured. The hybrid scheme is
therefore seen to fulfil its intended role perfectly for this kind of flow simulation.

2
The same slice and snapshot as used in Figures 8.2 and 8.3 have been employed, allowing a direct compar-
ison with the RANS/LES functionality plotted there.

115
7 Validation of the implementation

(a) Vorticity magnitude (b) Blending function σ

(c) Vorticity magnitude (d) Blending function σ

Figure 7.2: Hybrid scheme functionality for the cylinder case (CEASM-DES, Δt = 0.03D/u ∞ ,
above) and NACA0021 case (SALSA-DES97, NTS-1c grid, below) for a slice at the
spanwise mid-section.

7.3 Calibration of the CDES parameter

Although it is possible to derive analytical relationships that provide values of CDES

in relation to the Smagorinsky constant and other parameters of the underlying
RANS model, an empirical calibration is considered essential for a number of rea-
sons. Firstly, the model expressions do not take the level of numerical dissipation of
the flow solver into account. Secondly, as has been demonstrated by Yan et al. [186],
such derivations can be conceptually dubious and even deliver contradictory values.
These issues are discussed in more detail in Sect. 6.1.

The CDES values for all implemented DES variants have been calibrated using DIT, a
test case described in Sect. 6.1. The first calibration of CDES using DIT was reported
by Shur et al. [148]. The earlier calibration work of the ELAN DES implementa-
tions was similarly based on the benchmark experimental data of Comte-Bellot and
Corrsin [30] and the general method was reported by Bunge et al. [20]. Later on,
within the framework of the DESider project, the technique was revised somewhat

116
 7.3 Calibration of the CDES parameter

and utilises the DNS benchmark data of Wray [183] for comparison. Comparative
calculations have however not demonstrated a strong enough dependency on the
method used to justify a modification of the earlier calibrated parameter values. De-
tails and discussion of the calibration of CDES can be found in [19, 20, 21, 55, 56, 163,
172, 186].

Due to the considerable previous reporting of the calibration, an exhaustive presen-

tation of the DIT spectra for all models will not be given. Instead, spectra for the
SALSA-DES implementation alone will be presented and discussed and reference is
made to Tab. 5.3, where the calibrated values of CDES for all models have been sum-
marised. Figure 7.3 compares various spectra for the SALSA-DES implementation
with the Ψ function active: For a range of CDES on the 323 and 643 grids, and a cross-
plot of spectra at the calibrated value CDES = 0.60 on the 163 , 323 and 643 grids. In
accordance with the findings of Sect. 7.1, pure central differencing has been applied
for the convection scheme.

0,1 0,1 0,1

0,01 0,01 0,01

E(κ)

E(κ)

E(κ)

DNS DNS DNS

0,001 CDES = 0.50 0,001 CDES = 0.50 0,001 3
16 grid
CDES = 0.60 CDES = 0.60 3
32 grid
CDES = 0.70 CDES = 0.70 3
64 grid
0,0001 0,0001 0,0001
1 10 1 10 1 10
κ κ κ
(a) Variation of CDES , 323 grid (b) Variation of CDES , 643 grid (c) Variation of grid resolution,
CDES = 0.60

Figure 7.3: Example energy spectra for CDES calibration computations, showing the effects of
varying CDES and grid resolution. SALSA-DES model with Ψ function active com-
pared to Wray DNS benchmark data [183] at t = 0.2, t = 1.8 and t = 3.0 (in order of
decreasing energy). All computations with pure CDS convection scheme.

The effect of varying CDES is seen to be relatively strong, with a noticeable influ-
ence on the strength of the damping at high wavenumbers. In line with expectation,
the damping is stronger for higher values of CDES . Interestingly, stronger damp-
ing of the higher wave numbers gives rise to an over-prediction of the energy in
the lower wavenumber range, with an intersection of the spectra at approximately
κmax /2 (corresponding to approximately the resolution 4Δ). Furthermore, the influ-
ence of varying CDES appears to be stronger on the coarser grid. This is in accordance
with the general properties of LES: Modelling empiricism is confined to the smallest
scales and finer resolution therefore results a reduction of the relative influence of the
model. Finally, it is worth noting that such a clear influence of the model can only

117
7 Validation of the implementation

be achieved with a low-dissipative numerical scheme. In the DESider project [56]3 ,

results submitted for solvers with strong levels of numerical dissipation exhibited a
negligible influence of CDES or equivalent model parameters.

Were the results for the 643 grid to be considered in isolation, a value of CDES = 0.50
would clearly emerge as the most appropriate choice: The agreement with the DNS
benchmark spectra is excellent at all time samples. Cross-reference to the 323 grid re-
sults however reveal a poor performance for this value, with an insufficient level of
damping apparent. The spectra for CDES = 0.60 in contrast perform better on the 323
grid, but are mildly too dissipative on the 643 grid. A significant grid dependency on
the chosen value is hence observed, with higher CDES required on coarser grids. The
final decision in favour of CDES = 0.60 was motivated by the consideration that 323
is likely to be more representative of the level of resolution affordable in more com-
plex practical applications. The comparison of the DIT spectra from three different
grid resolutions furthermore enables an assessment of the resolution requirements
for LES discussed in Sect. 3.5.2. There, minimum cube sizes of the order of 363 vol-
umes were identified based on the estimates of Pope [116], and the estimate 323 of
Spalart [154] was cited. These results appear to corroborate these estimates: The
spectra at a resolution of 323 are reasonable whereas 163 points are clearly insuffi-
cient4 .

As identified, a priority is given to the results at 323 resolution when choosing the
calibrated value of CDES for pragmatic reasons. A further ambiguity can arise be-
tween the time snapshots (e.g. on the 323 grid, CDES = 0.7 gives the best agreement
at t = 0.2 whereas this is excessively dissipative at later snapshots). It is believed
that emphasis should be placed on the later time steps when deciding the calibrated
value.

The discussion of the DIT results for the calibration of CDES will be continued in
Sect. 8.2, where the variation of CDES between models will be analysed and the scope
for a dynamic determination of CDES will be discussed.

7.4 Validation of the Ψ functions

The problem of undesired activation of the damping functions of some RANS models
in LES mode has been described in Section 3.7.4. An analysis based on the method-
ology described by Spalart et al. [158] is presented in Appendix D for the models
considered and the implemented correction functions, Ψ, are summarised in Sec-
tion 5.5.2. Here, the functionality of the Ψ functions will be demonstrated. Further-
more, for the models for which no Ψ function could be derived (LLR and CEASM),
3
In particular, see Chap. IV, Sect. 5 therein.
4
This observed grid dependency has since been found to result primarily from the spurious aliasing effect,
mentioned in Sect. 6.1, which is stronger for coarser grids. Later studies employing filtered initial velocity
fields to eliminate the aliasing effect (not shown) exhibit a much weaker (yet still apparent) sensitivity of
the calibrated CDES value upon the grid resolution.

118
 7.4 Validation of the Ψ functions

an empirical assessment is conducted.

To demonstrate the Ψ functionality, the DIT test case (described in Section 6.1) is
employed using the Wray benchmark DNS data [183]. As the damping nature of
the SGS-mode coefficient term A emerges only for low values of νt /ν (which can be
interpreted as proportional to the “sub-grid scale Reynolds number”5 ), a fine grid
resolution is required for the effect to emerge. The tests have therefore been con-
ducted using the 643 grid.

0.1 0,1

0.01 0,01
E(κ)

0.001 E(κ)
0,001

DNS DNS
no Ψ no Ψ
Ψ active Ψ active
0.0001 0,0001
1 10 1 10
κ κ
(a) SAE-DES (b) SALSA-DES

Figure 7.4: DIT spectra demonstrating the functionality of the derived Ψ functions for SAE-
DES and SALSA-DES compared to Wray [183] benchmark data at t = 0.2, t = 1.8
and t = 3.0 (in order of decreasing energy), calibrated values of CDES , 643 grid.

Figure 7.4 shows the spectra resulting from SAE-DES and SALSA-DES simulations
of DIT with and without the derived Ψ function active. In both cases it can be seen
that the high wavenumber energy is strongly over-predicted in the absence of the
Ψ function, which is due to the strong damping of sub-grid eddy viscosity. Com-
parison with the “no-model” simulations using the CDS convection scheme (Fig. 7.1
in Sect. 7.1) reveals a similar qualitative nature – in both cases a positive gradient
emerges at high wavenumbers. However, the fact that this gradient is lower than the
exponent 2 suggests that the eddy viscosity is not entirely absent. Confirming this,
Fig. 7.5 shows contour plots of the νt /ν levels active for the SAE-DES computation
with and without the Ψ function active.

A further effect to note from Fig. 7.4 is that the strength of the eddy viscosity damp-
ing appears to increase as the time step increases and the turbulence decays. This is
because, as mentioned, the damping strength of the A coefficient increases as νt /ν
decreases. As the turbulence Reynolds number decays with time, so does the sub-
grid scale Reynolds number (for a constant grid spacing) and with it the level of νt /ν.
√
2 Cμ k2 νt νt
5
The turbulence Reynolds number [116] Re L = kL
ν = kεν . As νt = ε , ReL = Cμ ν ∝ ν . The term
“sub-grid scale Reynolds number” is used as here νt = νsgs .

119
7 Validation of the implementation

(a) SAE-DES, no Ψ (b) SAE-DES, Ψ active

Figure 7.5: Plots of the eddy viscosity ratio νt /ν at slices through the DIT domain for SAE-DES
computation with and without the Ψ function, t = 3.0, CDES = 0.60, 643 grid. Note
that the contours are scaled differently.

The failure mechanism of DES based on models without the necessary Ψ function is
therefore in accordance with expectation. More importantly, the correct functionality
of the Ψ functions derived for SAE-DES and SALSA-DES has been demonstrated; in
both cases a sensible spectral energy distribution is restored. The strength of the de-
viation caused by the spurious damping is observed to be very strong, underlining
the importance of the Ψ function analysis and derivation.

0.1 0.1

0.01 0.01
E(κ)

E(κ)

0.001 0.001

DNS DNS
LLR-DES CEASM-DES
0.0001 0.0001
1 10 1 10
κ κ
(a) LLR-DES (b) CEASM-DES

Figure 7.6: DIT spectra for the DES variants for which a Ψ function could not be derived, com-
pared to Wray [183] benchmark data at t = 0.2, t = 1.8 and t = 3.0 (in order of
decreasing energy), calibrated values of CDES , 643 grid.

Attention will finally be turned to the models for which a Ψ function was seen to be

120
 7.5 Validation of boundary layer shield function for CEASM-DES

necessary but could not be derived, namely the LLR and CEASM models (details can
be found in Appendix D). The spectra returned by LLR-DES indeed demonstrate, at
least qualitatively, the typical failure mechanism observed for the SAE and SALSA
variants. The fact that the eddy viscosity damping is already present at t = 0.2 fur-
thermore suggests that the threshold level of νt /ν is higher for the LLR model. The
behaviour at t = 1.8 and t = 3 differs from that of the SAE and SALSA models how-
ever; the strength of the damping appears to be lower at t = 3 compared to t = 1.8.
An explanation for this may be the complex dependencies upon the velocity field
gradients remarked upon in Appendix D (the relevant LLR model expressions are
given in Eq. (B.20)). Until this problem is resolved, the computation of DES on the
basis of the LLR model cannot be advised. For the LLR-DES results presented in this
work (which pre-dated this analysis), an assessment of the sub-grid Reynolds num-
ber is suggested; if νt /ν is sufficiently high throughout the LES-mode region, it can
be assumed that the damping problem has not occurred.

Contrasting results have been obtained for the CEASM model, with sensible spectra
returned despite the lack of an implemented Ψ function. Precisely why this is the
case is not understood, however the implication is that the CEASM-DES formulation
is valid – at least for the range of νt /ν that emerge in the 643 DIT test. Were the
damping problem to have occurred for the CEASM model, the Ψ function derived
for the background LL k − ε model could have been tested as a potential remedy.
In light of these favourable results however, ΨLL has not been implemented for the
CEASM-DES.

7.5 Validation of boundary layer shield function for CEASM-DES

To validate the functionality of the model-specific GIS-shield function derived for
CEASM-DES [21], Eq. (5.3), the canonical test case of a boundary layer developing
along an un-inclined flat plate will be employed. Because of its simplicity and the
lack of benchmark data required, a dedicated description of the test case was not
included in Chap. 6. A flat plate of unit length has been meshed in two dimen-
sions with a uniform grid in the streamwise direction (except for some compres-
sion towards the leading edge) and stretched grid in the wall-normal direction. The
streamwise spacing over the uniform portion was set to Δx = 4 × 10−3 , targeting
the occurrence of MSD (Sect. 3.7.3) as Δx  δ. The wall-normal grid was adjusted
to y+
max < 1 at the first cell centre. A short section with a symmetry boundary con-
dition was placed upstream of the leading edge of the plate, and a block profile of
constant velocity was applied at the inflow. Outflow boundary conditions were set
at the upper and downstream domain limits. The Reynolds number was set such
that Remax = 1 × 106 .

A steady RANS computation was conducted, which serves as the benchmark: When
the shield functions as intended, an identical velocity profile should be returned by
the shielded DES. The computations with the DES97 and shielded DES model acti-

121
7 Validation of the implementation

vated were conducted steady-state and were only resolved in two dimensions. For
the purposes of this simple validation of the shield function, this is acceptable as no
resolution of LES content is sought.

30
1
25

20

tanh (AL-ν )
2
+

15
U

10 GIS shield value

+
U , RANS
+
U , DES with shield
5 +
U , DES without shield
0
0
1 10 100 1000 10000
+
y

Figure 7.7: Velocity profiles and shield function behaviour at a location near the end of the
plate, demonstrating the effectiveness of the model-specific GIS-shield formulation
for CEASM-DES.

The normalised velocity profiles (in skin friction units, Sect. 2.4.1) from a location
towards the downstream end of the plate are shown in Fig. 7.7. These  are super-
imposed with the value of the boundary layer shield function, tanh A2L−ν from
Eq. (5.3). The occurrence of MSD in the unshielded DES97 computation can be clearly
seen in the form of a strong increase in the exponent of the velocity profile above
y+ ≈ 200. The effect of this is an under-estimation of the boundary layer thickness
of 27% relative to the RANS, and an 18% under-prediction of the wall shear stress.
The magnitude of these effects of course depend on the location along the plate and
the grid resolution.

The computation including the shield function however matches the RANS profile
very well. The agreement is however not absolutely perfect, with some minor dis-
crepancy apparent in the skin friction6 . The activity of the shield function is super-
imposed over the velocity profiles in Fig. 7.7. This is seen to correctly detect the
turbulent log law region (Sect. 2.4.1) of the boundary layer, equalling unity between
y+ ≈ 6 and the boundary layer edge around y+ ≈ 2000. The fact that the shield
6
This is concluded from the level of U + outside of the boundary layer. Because of the unconfined flow
domain, the velocity magnitude outside the boundary layer should be constant irrespective of differences
in the boundary layer thickness. Changes in U + outside the boundary layer hence arise from differences
in τw .

122
 7.5 Validation of boundary layer shield function for CEASM-DES

function drops very near to the wall is due to its dependence on k and the turbulence
damping effect of the wall in the viscous sublayer. This is however not considered a
deficiency of the formulation. The remaining minor discrepancy could probably be
eliminated with some tuning of the hyperbolic tangent blending function. Since the
publication of the generalised DDES method [158] (Sect. 3.7.4) however, the motiva-
tion for this is slight.

123
7 Validation of the implementation

124
8 Demonstration and Analysis

In Chapter 7, the implementation of key DES features was validated and the proper-
ties of the numerical convection schemes were examined. This lays the groundwork
for the presentation of a comprehensive range of studies in this chapter. Some of
these seek to establish a deeper understanding of the DES method whereas others
target the derivation of best practice advice for its application in an industrial con-
text. A particular emphasis is placed on the assessment of the predictive accuracy of
DES in a comparison with experimental data at an unprecedented level of detail.

The individual sections are dedicated to a particular topic, and in some cases draw
upon multiple test cases. Likewise, results from certain cases are called upon to
support the discussion on multiple topics. Reference is hence made to Chapter 6,
where the detailed test case descriptions are collated. An attempt has been made to
order the sections in this chapter such that the latter sections build upon the former.
Due to the inherent inter-dependencies however, this has not proved possible in all
cases.

8.1 RANS and LES activity in DES

To investigate the spatial distribution of the RANS and LES modes from DES, ref-
erence will be made to three bluff body test cases, the NACA0012 and NACA0021
stalled airfoils and the circular cylinder. These are introduced in Sections 6.3, 6.4
and 6.5 respectively. The influence of grid refinement is investigated using LLR-
DES97 results from the NACA0012 coarse and fine grids. Results for SALSA-DES97
and LLR-DES97 are compared for the NACA0021 case with the NTS-1c grid, and the
CEASM-DES with the GIS-shield described in Section 5.5.4 has been applied to the
cylinder case using a time step of Δt = 0.03D/u ∞ . A qualitative impression of the
simulated flow fields will be given in Section 8.1.1 before the RANS/LES functional-
ity is analysed and discussed in Section 8.1.2.

8.1.1 Instantaneous snapshot of the flow field

A visual impression of the flow physics and the structures resolved by the DES can
be obtained from Fig. 8.1, which shows instantaneous snapshots of the vortex cores
for the cylinder and NACA0021 cases1 . The vortex cores are portrayed using a neg-
ative value of the λ2 criterion proposed by Jeong & Hussain [66]. For the cylinder,
clearly-defined horseshoe vortices are evident at the cylinder/wall junctions, as are
the breakup of the shear layer and the organisation of the chaotic turbulent eddies
1
The NACA0021 and NACA0012 cases do not differ qualitatively, hence only one of these is presented here.

125
8 Demonstration and Analysis

into a coherent vortex street pattern. The same phenomena can be observed for the
NACA0021 case, with exception of the horseshoe vortices due to the spanwise pe-
riodicity. Both cases are hence typical examples of bluff-body flows with massive
separation.

The existence of such fine-grained resolved turbulence provides evidence of the de-
sired LES-mode functionality in the separated wake region. Although the resolved
eddies are of a similar scale in the near wake region of both geometries, it is clear
that these coarsen rather rapidly downstream in the case of the NACA0021. This can
be explained when the grids are compared (Fig. 6.6 for the NACA0021 and Fig. 6.8
for the cylinder); the NACA0021 grid exhibits a much more rapid expansion of the
cell size than the cylinder grid, for which a relatively constant cell size is maintained
downstream.

(a) (b)

Figure 8.1: Instantaneous vortex core structures portrayed using isosurfaces of the λ2 crite-
rion [66] shaded by streamwise velocity. (a) cylinder case, CEASM-DES, Δt =
0.03D/u ∞ ; (b) NACA0021 case, SALSA-DES97, NTS-1c grid.

8.1.2 Distribution of the RANS and LES zones

Two-dimensional snapshots of the vorticity magnitude at a slice through the wake
region2 and the functionality of the DES are compared in Fig. 8.2 for the cylinder
case. Using the ratio of the RANS to LES length scales LRANS /LLES as an indicator,
the regions in which the alternative modes of the DES are in operation are shown. In
this representation a value less or greater than unity corresponds to the RANS and
LES modes respectively and the interface value of LRANS /LLES = 1 is highlighted by
dashed contour lines.

The model can be seen to operate in LES-mode throughout the finely-resolved tur-
bulent wake and in the outer irrotational flow region, and as RANS in the separated
shear layers and outer edges of the turbulent wake. The majority of the attached
2
This figure (and the slice shown for the SALSA-DES97 NACA0021 in Fig. 8.3) depict the same snapshot as
the examination of the σ blending function in Fig. 7.2, allowing a cross-comparison.

126
 8.1 RANS and LES activity in DES

(a) Vorticity magnitude (b) LES-mode region

Figure 8.2: Analysis of the DES functionality for the cylinder case (CEASM-DES, Δt =
0.03D/u ∞ ) for a slice at the spanwise mid-section.

(a) Vorticity magnitude, SALSA-DES97 (b) LES-mode region, SALSA-DES97

(c) Vorticity magnitude, LLR-DES97 (d) LES-mode region, LLR-DES97

Figure 8.3: Comparison of DES functionality for the one-equation SALSA-DES97 (above) and
two-equation LLR-DES97 (below) computations of the NACA0021 case at the span-
wise mid-section.

127
8 Demonstration and Analysis

(a) Coarse grid (b) Fine grid

Figure 8.4: Extent of the instantaneous LES-mode region as a function of grid resolution.
The quantity LRANS /LLES is shown for values exceeding unity such that the
shaded regions portray the LES-mode operation. LLR-DES97 computations of the
NACA0012 case.

cylinder boundary layer is also handled with LES-mode, which at first inspection
appears incorrect. This is however an outcome of the laminar boundary layer prior
to separation, which is discussed in Sect. 8.5.1. Such an unsteady interface is typi-
cal of DES based on two-equation models, where the RANS length scale is derived
locally from the model parameters (see Tab. 5.2). This behaviour can be interpreted
as a rough measure of the grid’s capability to resolve the turbulent scales present
at each point, reverting to URANS modelling when this is insufficient3 . When the
DES definition proposed by Travin et al. [171] (cited in Sect. 3.7.2) is recalled, this
switch-back to RANS mode in the separated regions with coarser grids is clearly in
accordance with the underlying concept of DES.

Examination of the values of LRANS /LLES in the region of the interface reveals a fur-
ther aspect of the switch between RANS and LES modes, namely that this is a con-
tinuous transition without any discontinuities in the solution field. There is therefore
no strong difference in the model behaviour just either side of the interface for val-
ues of LRANS /LLES ≈ 1, even though a formal transition between two contrasting
modelling frameworks has occurred.

Unlike the cylinder case, the NACA0021 airfoil has been computed using a range
of turbulence models. This case can therefore be used to demonstrate the difference
in RANS/LES switching behaviour for one-equation and two-equation models. As
demonstrated for the cylinder, the local definition of the RANS length scale causes a

3
Caution is however advised: Although denoted LRANS , the length scale constructed from the local turbu-
lence variables is not necessarily equivalent to the RANS length scale. This is because the LES activity
serves to damp k to k sgs . LRANS is hence often closer to a length scale of the subgrid-scale modelled
turbulence.

128
 8.1 RANS and LES activity in DES

switching between RANS and LES modes also in regions of separated flow for two-
equation models. This behaviour is also seen for the LLR-DES97 solutions of the
NACA0021 case in Figs. 8.3 (c) and (d). For the one-equation SALSA-DES97 solu-
tions (Figs. 8.3 (a) and (b)) however the picture is very different; RANS-mode oper-
ation is confined to a thin region encompassing the attached boundary layers. This
is because of the length scale definition LRANS = dw , and the RANS/LES switching
behaviour is hence purely grid-dependent.

Upon initial consideration it is tempting to interpret more into this difference than it
perhaps deserves. Indeed the DES switch is activated to LES mode everywhere far
from the wall, and it can be shown that under the local-equilibrium assumption this
LES mode is equivalent to the Smagorinsky model (Appendix D). However, the only
modification compared to the RANS model is that the dissipation term (Eq. B.2) be-
comes proportional to 1/Δ2 instead of 1/d2w . In cases where a large grid cell occurs
far from the wall, the RANS and LES modes therefore become essentially equiva-
lent. The length scale definition of dw used in the majority of one-equation mod-
els is in any case only valid for attached boundary layers and for this reason one-
equation models are strictly only applicable to attached flows (Sect. 3.4.3, [80, 157]).
The demonstrated behaviour of switching back to RANS mode in separated regions
of coarser grids is therefore inherent to DES regardless of the chosen background
model.

Nonetheless, two-equation models offer a minor advantage that the formal division
of RANS/LES operation can be easily inspected. It can even be considered advisable
to check to what extent the focus region [154] targeted for LES mode activity is actu-
ally computed as such. This can be used as a simple and approximate thumb-rule to
check whether the grid resolution is sufficiently fine4 , a remark that will be clarified
by the discussion to follow.

The effect of the grid resolution on the RANS/LES mode regions is portrayed in
Fig. 8.4 for the NACA0012 case. Although the difference in grid resolution is not very
strong (and indeed the “fine” grid could be considered relatively coarse by modern
standards), the impact on the RANS/LES switching can be clearly discerned. The
fine grid sustains regions of LES mode operation much further downstream than the
coarse grid, for which no LES mode is apparent further than around 1.5c downstream
of the airfoil.

4
It was however suggested by Prof. Strelets in his review of this work that visualisation of the fraction of
time spent locally in RANS or LES mode may be more useful than snapshots of the length scale ratio in
this respect.

129
8 Demonstration and Analysis

8.2 Variation of CDES between models and the scope for a “dynamic
DES”
The calibration of CDES using DIT has been demonstrated in Section 7.3 and the im-
plemented calibrated values are listed for each model in Table 5.3. These show a
certain degree of model-specific variation within the range 0.60 ≤ CDES ≤ 0.75. In
order to derive the Ψ correction function for each model, analysis of the LES-mode
behaviour using the methods described in Appendix D was applied. Here it will be
investigated whether such analysis can provide an explanation for the variation in
CDES witnessed. In so doing, the possibility of analytically deriving an appropriate
value of CDES for each model (naturally under the assumption of negligible numer-
ical dissipation, Sect. 7.1) will be assessed. Finally, a discussion will be given on the
scope for developing a “dynamic DES” method based on the derived relationships.

An expression linking CDES to the Smagorinsky model parameter CS can be derived

for each model. The first step is a comparison of the Smagorinsky model (Eq. (3.10))
with the generalised form of DES models in LES-mode (derived in Appendix D):

νt = (CS Δ )2 S ∗ = A(Ψ CDES Δ )2 S ∗ . (8.1)

By incorporating the definition of Ψ (Eq. (D.3)), the desired expression for the equiv-
alent Smagorinsky parameter value can be obtained for each model as
√
CS = A∗ CDES . (8.2)
If the values of CDES calibrated by DIT have been obtained in a consistent manner
and if the assumption of local equilibrium underlining the derivation is valid, then
similar values of CS should be returned by each model. The values of CS correspond-
ing to the calibrated CDES are given together with the asymptotic coefficient A∗ (for
high values of νt /ν) in Table 8.15 .

SA SAE SALSA WCX CEASM (LL)

   
Cb1 Cb1 f C∗b1 Cω2 3/2 Cμ Cε2 3/2
A∗
Cw1 f w∗ Cw1 f w∗ f C∗w1 f w∗ Cω1 Cε1
A∗ value 0.0986 0.0893 0.0993 0.0496 0.0416
Calibrated CDES 0.65 0.65 0.60 0.70 0.65
CS 0.2041 0.1942 0.1891 0.1559 0.1325

Table 8.1: Values of CS equivalent to the calibrated CDES parameter for each model, using
Eq. (8.2).

5
The table does not include the LLR model for which A ∗ cannot be derived and uses the expression for the
background LL k − ε model for the CEASM (which is only valid for constant Cμ = 0.09).

130
 8.2 Variation of CDES between models and the scope for a “dynamic DES”

0.1

0.01

0.01

E(κ)

E(κ)
DNS
0.001 Smagorinsky 0.001
SAE-DES
SALSA-DES
WCX-DES
CEASM-DES
0.0001
1 10 15 20 25 30 35
κ κ
(a) Spectral comparison (b) Zoom of high κ region (linear κ
scale)

Figure 8.5: Cross plot of DIT spectra for all DES variants compared to Wray [183] benchmark
data at t = 0.2, t = 1.8 and t = 3.0 (in order of decreasing energy), calibrated values
of CDES , 643 grid.

The values of CS in fact exhibit a fairly high degree of scatter, especially when recall-
ing from Section 3.5 that CS ≈ 0.2 is typical for DIT calibration whereas CS ≈ 0.1 is
used for wall-bounded flows. For the SA-based models the values are very close to
the CS ≈ 0.2, however those for the WCX and CEASM models are rather low. Cross
plotting the DIT spectra for all models reveals that the trend in equivalent CS cor-
responds precisely to the moderate differences in dissipation of the high wavenum-
ber spectra (Fig. 8.5). Spectra for the Smagorinsky model with the analytical value
CS = 0.1825 (Sect. 3.5.1, [82, 3]) are also included. Particularly at t = 0.2, the models
are ranked CEASM → WCX → Smagorinsky → SALSA → SAE correlating perfectly
with the equivalent CS trend in Tab. 8.1. This trend is maintained in the later spectra
for all models with the exception of CEASM, which appears to become increasingly
dissipative (relative to the other models) as the solution progresses. It is supposed
that this irregular behaviour is due to the lack of a derivable Ψ function for this model
and the observed variable behaviour of ALL (Section 7.4, Appendix D).

This apparent correlation between the scatter of equivalent CS values and the high
wavenumber behaviour of the spectra raises the possibility of determining CDES (CS )
for a fixed target value of CS . If the underlying assumptions in the analysis are cor-
rect, then identical spectra should be returned. The CDES values corresponding to
CS = 0.1825 are listed in Tab. 8.2, and the resulting spectra on the 323 and 643 grids
are plotted in Fig. 8.6.

At the first time instant, t = 0.2, the spectra are indeed very similar on both grids
(indeed more so than the calibrated values) and the same grid-dependent nature re-
ported in Sect. 7.3 is apparent. However, at later instants the spectra diverge strongly:
Those from the SA-model family remain similar and show a reasonable behaviour,
whereas the WCX and especially the CEASM spectra are excessively dissipative on

131
8 Demonstration and Analysis

SA SAE SALSA WCX CEASM (LL)

CDES for CS = 0.1825 0.581 0.611 0.579 0.819 0.895

Table 8.2: Values of CDES equivalent to the analytical value CS = 0.1825 for each model, using
Eq. (8.2).

0.1 0,1

0.01 0,01
E(κ)

E(κ)
DNS DNS
0.001 Smagorinsky 0,001 Smagorinsky
SAE-DES SAE-DES
SALSA-DES SALSA-DES
WCX-DES WCX-DES
CEASM-DES CEASM-DES
0.0001 0,0001
1 10 1 10
κ κ
(a) 323 (b) 643

Figure 8.6: Cross plot of DIT spectra for all DES variants compared to Wray [183] benchmark
data at t = 0.2, t = 1.8 and t = 3.0 (in order of decreasing energy), CDES values
corresponding to CS = 0.1825.

both grids. This suggests that the analytical relationship between CDES and CS can
be considered valid for the one-equation models tested, whereas effects unaccounted
for in the simple equilibrium analysis appear to play an important role in the two-
equation models. One possible explanation for this behaviour concerns the method
of length scale substitution applied in the two-equation models (Sects. 3.7.2 & 8.3).
As LDES is only substituted in the destruction term of the k equation the variable
k is limited directly, whereas the second equation variable (ε or ω) is limited only
implicitly through its dependency on k. The mixture of an LES-limited k and a “nat-
ural” value of the second equation variable in the eddy viscosity expression may be
the cause of this behaviour. Were the results for the CEASM-DES to be considered
in isolation, one might also conjecture that the Ψ problem identified with this model
(Appendix D) were to blame. However, the similar behaviour seen for the WCX-DES
contradicts this.

The implications of this are that a derivation of CDES as a function of CS can indeed
provide sensible values for some models (on the basis of this evidence, apparently
the SA family of models). However, caution is advised when considering new mod-
els for DES: Such analytically derived values should at most be used as a starting
value for an empirical calibration based on DIT and a test should always be con-
ducted.

132
 8.2 Variation of CDES between models and the scope for a “dynamic DES”

Dynamic DES?
A further implication of the demonstrated link between CDES and CS , at least for the
SAE and SALSA models, concerns the possibility to extend the SGS modelling capa-
bility of DES. In principle, enhancements developed originally for the Smagorinsky
model could be applied to the LES mode of DES. In particular, the dynamic determi-
nation of CS pioneered by Germano et al. [51] and described in Sect. 3.5.1 could offer
considerable advantages.

As outlined in that section, the Smagorinsky parameter demonstrates strong case-

dependency and a range of disadvantages of a constant CS for LES have been listed.
Dynamic determination of CS , or dynamic one-equation models incorporating anal-
ogous features have been shown to offer a wide range of advantages in LES, not least
improved generality through the elimination of the dependency on case-specific cal-
ibration. Under first consideration therefore, the motivation for investigating such a
dynamic DES is strong.

It can however also be argued that such dynamic parameter determination is of less
interest in the context of DES. For example, Smagorinsky LES solutions are strongly
sensitive to the value of CS for the near-wall flow only. For flow further from the wall
(including the log law region), the solution is only weakly dependent on CS [17, 159].
In both natural DES as well as WMLES applications, the near-wall RANS treatment
is expected to result in a low general dependency on CDES 6 . Furthermore, the ad-
vantages of dynamic LES concerning transition prediction and the deactivation of
the subgrid-scale model in the limit of full DNS resolution are of limited relevance in
the expected application scope of DES.

Nonetheless, dynamic DES is considered worth pursuing for a number of reasons,

and some initial DIT experimentation with the Germano/Lilly dynamic model [51,
83] (Sect. 3.5.1) has enforced this view. It was demonstrated in Sect. 7.3 that the cal-
ibrated value of CDES depends on the grid resolution (compare e.g. the spectra for
CDES = 0.50 on the 323 and 643 grids in Fig. 7.3). It is conceivable that an appro-
priate dynamic modification of CDES could reduce such sensitivity, and DIT com-
putations were conducted to investigate this. Figure 8.7 compares spectra from the
Smagorinsky model using CS = 0.1825 with those returned by the dynamic model
for a variation of grid resolution. The Smagorinsky model exhibits a strong grid de-
pendency, with an insufficient level of dissipation apparent on the 323 grid and very
poor spectra at 163 (which is clearly under-resolved). The dynamic determination of
CS however significantly improves the situation: The spectra at 323 overlap almost
perfectly with those at 643 , and even the under-resolved 163 spectra are improved.

6
Some evidence supporting this is given in Sect. 8.4.1. In a comparison of models with strongly differing
values of “effective CS ” for massively separated flow, the solutions were identical to within the statistical
error despite significant differences in the eddy viscosity levels in the wake. Furthermore, in an attempt
(not shown) to reduce the residual LLM in channel flow IDDES, a lower value of CDES was computed.
The influence on the mean flow was again negligible.

133
8 Demonstration and Analysis

0,1 0,1

0,01 0,01

E(κ)

E(κ)
DNS DNS
0,001 3 0,001 3
16 grid 16 grid
3 3
32 grid 32 grid
3 3
64 grid 64 grid
0,0001 0,0001
1 10 1 10
κ κ
(a) Smagorinsky model, CS = (b) Dynamic model [51, 83]
0.1825

Figure 8.7: DIT spectra for LES with the Smagorinsky model (constant CS = 0.1825) and the
dynamic model of Germano/Lilly ([51, 83]) compared to Wray [183] benchmark
data at t = 0.2, t = 1.8 and t = 3.0 (in order of decreasing energy).

This indicates therefore, that dynamic models offer the potential to improve LES on
coarser grids and to resolve the grid-dependent nature of the model parameter.

A further possible benefit of a dynamic DES concerns its usefulness in the context
of IDDES. Recall from Sect. 3.7.5 that an empirically-tuned grid-dependent near wall
damping of the filter width, Δ, was introduced to address the discrepancy between
suitable values of CDES for DIT calibration and fully-resolved LES of wall-bounded
flows. The generality of such a grid-dependent formulation remains to be demon-
strated however. It could indeed emerge, that a truly dynamic CDES reproduces the
desired behaviour with increased generality.

It appears therefore that the prospect of concrete advantages to be expected from a

dynamic DES are uncertain, however worth well worth pursuing. Future investiga-
tions in this direction must aim to demonstrate measurable advantages of the for-
mulation. If these prove to be marginal at best, then these benefits must be weighed
against the increased complexity of the formulation. Concerning the formulation, the
simplest variant would be to employ the usual dynamic determination of CS accord-
ing to Germano [51] or Lilly [83], and the simple expression of Eq. (8.2) to translate
this to the corresponding CDES . For the reasons of numerical stability and generality
outlined in Sect. 3.5.1 however, an approach along the lines of Davidson’s dynamic
one-equation SGS model would be considered preferable. This may be easily formu-
lated on the basis of a double substitution of the DES length scale (Sects. 3.7.6 & 8.3),
as such formulations give rise to a k equation SGS model.

134
 8.3 Alternative length scale substitutions in DES

8.3 Alternative length scale substitutions in DES

As described in Sect. 3.7.2, the DES methodology was generalised by Travin et al.
[171] and Strelets [163] to be applicable to RANS models in general. Whereas the
substitution of the length scale dw affected only the destruction term of the transport
equation in the original SA model based DES, for other RANS models length scales
can be identified in a range of terms. This gives rise to a degree of freedom as to
the length scale(s) to be substituted by LDES . Alternative formulations employing
an additional substitution in the eddy viscosity expression of two-equation models
have been explored by Bush & Mani [22] and Kok et al. [76], which are described
in Sect. 3.7.6. However a comparative study to assess the impact of these alterna-
tive formulations and to identify practical advantages were not presented by those
authors. A thorough investigation of this issue was published by Yan, Mockett &
Thiele [186] the principal findings of which will be summarised here. The interested
reader is referred to the original publication for more detail.

Formulations
The 1988 Wilcox k − ω model [181] chosen as the object of the study, because it is
a well-known and representative model with a simple formulation. Furthermore,
no complications arise concerning spurious behaviour of damping terms in LES
mode (as demonstrated in Appendix D). The model formulation is listed in Ap-
pendix B.2.1.

Three alternative DES formulations are derived, which entail the substitution of the
DES length scale, Eq. (3.16), in different combinations of the model terms. The dissi-
pation term of the k equation, given in Eq. (B.12), and the eddy viscosity expression,
Eq. (B.14), are reformulated as follows:

k3/2
k = Cμ kω = , (8.3)
Lk

k √
νt = = Cμ L ν k , (8.4)
ω
√
k
Lk = Lν = . (8.5)
Cμ ω
The RANS turbulence length scale is the same in both expressions and the notation
L k and L ν is employed merely to aid description in the following. The three DES for-
mulations are labelled DES1, DES2 and DES3, and the corresponding substitutions
of the DES length scale are summarised in Tab. 8.3. DES1 hence corresponds to the
standard DES practice, DES2 to the dual substitution applied by Bush & Mani [22]
and Kok et al. [76], and DES3 is the remaining formulation that completes the possi-
ble combinations of length scale substitutions.

135
8 Demonstration and Analysis

Lk Lν
DES1 LDES Lν
DES2 LDES LDES
DES3 Lk LDES

Table 8.3: Summary of the substitutions of the length scale in the k equation destruction term,
L k , and in the eddy viscosity expression, L ν , by the DES length scale LDES .

The alternative formulations in SGS mode

As described in Sect. 3.7.2, the minimum requirement upon the length scale substi-
tution to give a valid DES formulation is that a form corresponding to the Smagorin-
sky model, Eq. (3.10), can be derived in LES mode [163]. Such an analysis is there-
fore conducted to assess the alternative formulations, which also addresses possible
derivations of CDES in relation to other model parameters. Beginning with DES1, ref-
erence is made to Appendix D, where the local equilibrium assumption applied to
both model equations is shown to give rise to the following Smagorinsky-like form:
 
Cω2 3/2
νt = (CDES Δ )2 S ∗ . (8.6)
Cω1
In the spirit of the analysis of Sect. 8.2, a dependency of CDES on a given value
of CS can be derived. For the analytical value of CS = 0.1825, the corresponding
CDES = 0.82 is obtained.

For DES2, it emerges that a k equation SGS model of identical form to that of Yoshi-
zawa [188] is obtained:

  
Dk ∂ νt ∂k k3/2
= ν+ + νt S ∗2 − (8.7)
Dt ∂x j σk ∂x j CDES Δ
√
νt = Cμ CDES Δ k (8.8)
The dual substitution of LDES in fact leads to a decoupling of the ω equation from
the solution. As noted in Sect. 3.5.1, when the assumption of a balance between
the production and destruction terms is applied to the k equation SGS model, the
Smagorinsky model is indeed returned:

νt = Cμ3/2 (CDES Δ )2 S ∗ . (8.9)

Again, an expression for CDES as a function of CS can be derived, which delivers
CDES = 1.11 for CS = 0.1825. However, by comparing the Yoshizawa model with
the k equation SGS model inherent to DES2, a plethora of additional expressions for
CDES can be derived. For example:

136
 8.3 Alternative length scale substitutions in DES

DES1 DES2 DES3

CDES 0.70 0.95 0.90

Table 8.4: Calibrated values of CDES for the three alternative length scale substitutions.

Ck
CDES = = 0.78 (8.10)
Cμ

1
CDES = = 0.95 (8.11)
Cε
"
Ck
CDES = = 0.86 . (8.12)
Cμ C

Although all these values are of approximately the correct order of magnitude, the
spread is large enough to give rise to significant discrepancies in a DIT test. The
inherent contradiction arising from such analytical CDES derivations further empha-
sises the importance of empirical calibration calculations, a recurring theme through-
out this work.

Finally, for the DES3 formulation, no derivation of a Smagorinsky form of the SGS
model is possible, although both k and ω are indirectly limited through the limitation
of νt in their production terms. As such, DES3 cannot strictly count as a valid DES
formulation.

DIT calibration of CDES for the alternative formulations

The DIT test case (described in Sect. 6.1) employing the experimental benchmark
data of Comte-Bellot & Corrsin [30] was applied to ascertain appropriate values of
CDES for the DES1, DES2 and DES3 formulations. The use of DIT for such calibration
has been discussed in Sect. 7.3. The resulting values are given in Tab. 8.4.

The calibrated values exhibit a relatively strong variation, which indeed exceeds the
greatest difference between background RANS models (compare for example with
the values listed in Tab. 5.3). The DES2 and DES3 exhibit much higher values of
CDES , which appears to represent a lower “inherent dissipativity” of the formulation,
a concept to be returned to in the discussion subsection to follow.

Comparison for massively-separated flow

To compare the three DES formulations for a more complex and practically-relevant
flow, the NACA0012 test case (Sect. 6.3) was computed using the coarse grid and the

137
8 Demonstration and Analysis

DES1 DES2 DES3

CL 0.978 0.990 1.038
CD 1.635 1.664 1.740
St 0.176 0.173 0.182

Table 8.5: Time-averaged lift and drag coefficient and vortex shedding Strouhal number for the
NACA0012 case simulated using each length scale substitution.

calibrated values of CDES for each variant (Tab. 8.4). Following an elimination of ini-
tial transient behaviour, over 1000 non-dimensional time units (c/|U | ∞ ) were com-
puted for each case for statistical analysis, corresponding to over 200 vortex shedding
periods. Although this is a very long simulated time, an appreciable statistical error
can still be expected in the mean values. This arises from the highly stochastic nature
of the flow and the occurrence of strong low-frequency modulation, and is described
in detail for the similar NACA0021 case in Sect. 8.4.1.

Force spectra, plots of the mean surface pressure coefficient and statistical wake flow
quantities were compared, with a high degree of similarity observed between the for-
mulations. To give an indication of this, the time-averaged force coefficients with the
vortex shedding Strouhal number are given in Tab. 8.5. These results are not suffi-
ciently different to reveal a systematic discrepancy outside of the expected statistical
convergence error.

Results reported for a coaxial jet

Despite the negligible effect that the length scale substitution method has exhibited
for the massively separated flow around the NACA0012 profile, more recent appli-
cations of the DES1 and DES2 formulations to a coaxial jet flow have revealed a
very interesting difference. These computations, carried out by Yan et al. [187] and
funded by the European project CoJeN7 , have also been reported by Michel et al. [93],
Mockett & Thiele [99] and Mockett et al. [96]. The goal of the investigation was to
develop a simulation methodology for the development of strategies to reduce jet en-
gine noise. Such modifications include chevrons introduced to the engine nacelle lip,
and inclusion of the geometry was hence considered important. As such, DES would
appear to represent a well-suited method combining RANS of the nozzle geometry
with LES of the resolved turbulent fluctuations. However, the weakness of the shear
layer instability and lack of convective transport of turbulent disturbances through
recirculation cause exacerbation of the grey area problem discussed in Sect. 3.7.3.

The DES1 and DES2 variants were compared for a short-cowl generic coaxial jet ge-
ometry. The core stream is heated and flows at M = 0.90, whereas the bypass stream
7
Computation of Coaxial Jet Noise, Contract No. AST3-CT-2003-502790.

138
 8.3 Alternative length scale substitutions in DES

(a) DES1 (b) DES2

Figure 8.8: Density isosurfaces showing resolved structures in the jet shear layer using the
DES1 and DES2 length scale substitution variants of WCX-DES. Courtesy of J. Yan.

is at ambient temperature and M = 0.86, flowing into static surrounding air. The
configuration was computed using a grid of 3.8 × 106 cells on the Bavarian HLRB su-
percomputer. Density isosurfaces are used to visualise the resolved turbulent struc-
tures in Fig. 8.8. The two variants are seen to exhibit strong differences in the levels
of resolved turbulence in the early shear layer, with much finer structures obtained
using DES2. This represents a significant advantage for the aeroacoustic application
in question, where the resolved turbulent structures are the dominant sound source.
The accelerated development of resolved turbulence corresponds to a reduction of
the grey area extent. The mechanism believed to be responsible for this improve-
ment will explained in the following discussion.

Discussion
The initial investigations of the alternative length scale substitutions, incorporating
the DIT and NACA0012 test cases and published by Yan et al., 2005 [186], analysed
the SGS form of the variants and indicated an equivalence once the strong differences
in CDES have been balanced through calibration. That the results are indiscernible
for such a massively-separated (and LES-mode-dominated) flow is indeed reassur-
ing, however the later investigation of jet flow (Yan et al., 2007 [187]) indicates that
this is not the whole story. The DES2 form appears to significantly reduce the impact
and extent of the grey area, an important and promising result that warrants some
further discussion.

A convincing explanatory mechanism is proposed, the key to which was indeed

identified in the original paper when discussing the strongly different values of CDES
that arise. In the above subsection, these were vaguely linked to the concept of in-
herent dissipativity of the formulation, without elaboration as to the meaning of this.
The inherent dissipativity regards the strength or rapidity with which RANS levels
of eddy viscosity are damped to SGS levels once the LES length scale is active in the
model equations. That both of the variants with a direct length scale substitution

139
8 Demonstration and Analysis

in the eddy viscosity expression exhibit a lower inherent dissipativity (based on the
higher CDES values) can be analysed by considering the model terms in question. The
limitation of the eddy viscosity due to the LES length scale results in a reduction of
the production and diffusion terms in the transport equations, whereas the standard
length scale substitution only increases the destruction term. The k equation produc-
tion and destruction terms sit physically at opposite ends of the turbulent cascade,
being responsible for the generation of turbulence from mean-flow shear and its dis-
sipation into heat via molecular viscosity, respectively (Sect. 2.2). These phenomena
involve entirely different physical scales and a time lag is furthermore apparent be-
tween production and destruction. Considering the situation of a switch from RANS
to LES mode along a streamline, the response of the model in terms of damping νt
will be rapid when the production term is limited (the modelled turbulence is cut off
at its source) and more delayed when only the destruction term is increased.

Such consideration of the respective strong and weak effects of the production and
destruction terms serves to explain both the differences in the CDES values8 as well as
the shortened grey area observed for the jet flow. The encouraging results concerning
the grey area reduction should be investigated in more detail on the basis of a suitable
and more simple test case9 .

8.4 Sensitivity of DES to the RANS model used

Since the generalisation of the DES formulation from its SA-specific origin to a gene-
rally-applicable methodology ([163], Section 3.7.2), the question of the model sensi-
tivity of DES has been of high practical importance. Is this similar to the URANS
situation, where a very strong model sensitivity is observed (see e.g. [94], [55] and
[146]), or is this alleviated by the introduction of the LES ingredient? It can be ex-
pected that the model sensitivity would depend upon the flow in question, in partic-
ular upon the extent to which the RANS region influences the global flow. With the
variety of DES implementations and test cases to hand, these issues can be explored
in a comprehensive manner. Two test cases with contrasting flow physics are cho-
sen: the geometry-induced massive separation of the NACA0021 airfoil in deep stall
(Section 8.4.1) and the bump test case, for which a weaker separation from a smooth
curved surface occurs (Section 8.4.2). In both cases results are compared between
DES variants based on different RANS models, with identical grids and numerical
setup. No comparison with experiments will be made – this is done in Sections 8.5
and 8.7.

8
That the DES2 and DES3 values are so much higher than that of DES1 corroborates the interpretation that
the production term has a stronger influence than the destruction term. Furthermore, the fact that DES2
has the highest value of CDES is in line with expectation, as this exhibits the strongest limitation in LES
mode (the substitution is active in all model terms).
9
Perhaps an un-inclined flat plate of finite streamwise length, where the turbulent boundary layers coalesce
to a separated wake at the trailing edge?

140
 8.4 Sensitivity of DES to the RANS model used

8.4.1 Massively-separated external flow: NACA0021 at α = 60◦

This test case has been employed in the descriptions of the RANS and LES zone
distribution and hybrid convection scheme functionality in Sections 8.1.2 and 7.2, re-
spectively. Instantaneous snapshots of the flow field were shown, from which the
highly-unsteady massively separated wake flow could be observed. Due to the rel-
ative sharpness of the leading and trailing edges, the separation can be considered
geometry-induced and, as a result, relatively insensitive to model-specific differences
in the prediction of the upstream boundary layers. The level of similarity of the solu-
tions will be investigated by a detailed comparison of time-averaged and unsteady
flow characteristics. The test case, grid and numerical setup are described in Sec-
tion 6.4. All results in this section were computed with the NTS-1c grid, and the
SALSA-DES97, LLR-DES97 and CEASM-DES97 variants have been employed.

Statistical convergence of the mean integral force coefficients

For such highly unsteady flows, the validity of the time-averaged quantities depends
on the size of the time sample computed for averaging. To ascertain the level of statis-
tical convergence of these simulations, running averages of the integral lift and drag
coefficient have been computed and are plotted in Fig. 8.9. These curves begin at the
first time step employed in the statistical analysis (i.e. following the elimination of
initial transient behaviour) and show how the mean forces develop as the averaging
sample increases. It is seen that substantial shifts in the mean value can occur even
after a very long simulation time. This is due to the random low frequency modula-
tion of the force signals, a feature discussed in Sect. 8.5.2 for the cylinder flow case.

Despite the very long simulation times (a minimum of over 40000 time steps has been
computed in all cases, corresponding to roughly 1000 t∗ , where t∗ = c/|U∞ |, or 200
vortex shedding cycles), an appreciable statistical convergence error is still apparent.
The magnitude of this is furthermore hard to specify, due to the random drift of
the mean values (e.g. for the CEASM-DES at t∗ ≈ 1400). A statistical convergence
error of around ±0.03 units (for both CL and CD ) could be considered a conservative
estimate.

Time-averaged flow
The time-averaged integral force components are summarised in Tab. 8.6, together
with the total number of time steps over which the averaging has been conducted,
NΔt . Considering the residual statistical averaging error estimated in the preceding
section, the small differences in the values may not be interpreted as arising from
model dependency. The predicted forces are considered identical to within the esti-
mated statistical convergence error.

In addition to the lift and drag coefficients, the force coefficient components normal
and tangential to the airfoil chord have been computed using a coordinate rotation
and tabulated in Tab. 8.6. Again, these values are very similar between the variants,

141
8 Demonstration and Analysis

1.05 1.7
SALSA-DES SALSA-DES
LLR-DES LLR-DES
CEASM-DES 1.65 CEASM-DES
1

CD
CL

1.6

0.95
1.55

0.9 1.5
0 200 400 600 800 1000 1200 1400 1600 1800 0 200 400 600 800 1000 1200 1400 1600 1800
t* t*

Figure 8.9: Running averages of the lift and drag coefficients (integrated over the entire span)
over the time range used for statistical analysis.

CL CD CN CT St Wake length NΔt

SALSA-DES 0.9842 1.5916 1.8705 -0.0565 0.1855 1.725 40443
LLR-DES 0.9846 1.6204 1.8956 -0.0425 0.1878 1.697 49424
CEASM-DES 0.9701 1.6059 1.8758 -0.0372 0.1868 1.729 65457

Table 8.6: Time-averaged integral lift, drag, chord-normal and chord-tangential force coeffi-
cients, Strouhal number, wake length (defined as the saddle point position relative
to the leading edge, x = y = 0) and the number of time steps used for statistical
analysis.

-2

-1
Cp

SALSA-DES
LLR-DES
0 CEASM-DES

1
0 0.2 0.4 0.6 0.8 1
x/c

Figure 8.10: Profiles of time and spanwise-averaged pressure coefficient.

although the somewhat higher CN value for the LLR-DES deserves some further
investigation. Examination of the time and spanwise-averaged pressure coefficient
profiles in Fig. 8.10 reveals a stronger suction along the downstream airfoil surface
for the LLR-DES, corresponding to the higher CN . This suggests a difference in the
mean wake flow field, which will be described in the following.

142
 8.4 Sensitivity of DES to the RANS model used

(a) SALSA-DES97 (b) LLR-DES97

(c) CEASM-DES97

Figure 8.11: Streamlines of the time and spanwise-averaged flow field.

Streamlines for the time and spanwise-averaged flow field have been plotted in
Fig. 8.11. All models deliver the same fundamental features: the stagnation point
in a similar location on the upstream airfoil surface, separation at the leading and
trailing edges, the formation of two large counter-rotating mean vortices (the up-
permost of which is larger), a secondary separation of the re-attached flow towards
the leading edge on the rear airfoil surface and the closure of the wake at a sad-
dle point. The plots appear identical visually, however a quantitative comparison
reveals some minor differences. Interpreting the saddle point as the closure of the
time-averaged wake, a measure of the wake length can be obtained. This is defined
as the streamwise distance between the saddle point and the coordinate origin at the
profile leading edge, and the values are listed in Tab. 8.6. Whereas the values for the
SALSA and CEASM variants are very similar, the LLR-DES exhibits a shorter wake.
This physical link between the wake length, rear surface pressure level and force co-
efficient is in line with the findings of Roshko [129] (summarised in Section 2.5).

To examine the time-averaged resolved and modelled turbulence quantities returned

by each simulation, contours of the resolved and modelled turbulent kinetic energy
are shown in Fig. 8.12. These quantities are obtained using the methods described in

143
8 Demonstration and Analysis

Appendix E10 . The resolved fluctuations predicted by each model are highly similar
both in level and distribution. It can be seen that kres develops within the separated
shear layers and reaches its highest levels in the regions of the time-averaged vortex
cores and saddle point. The impinging flow on the downstream airfoil surface causes
a further amplification of the resolved fluctuations in the near wall region.

Contrastingly, a marked model dependency is observed in the time-averaged mod-

elled turbulent kinetic energy: The SALSA implementation exhibits much higher
levels of k mod throughout the wake, most markedly within the shear layers, whereas
the LLR and CEASM are similar. It could be supposed that this is an artefact of
the estimation of k from νt for the one-equation SALSA model, employing the Brad-
shaw hypothesis [15] as described in Appendix E (k is not available directly as for
the other models). To examine this possibility the time-averaged eddy viscosity is
plotted in the leading edge region in Fig. 8.13 and shows considerably higher levels
for the SALSA computation. The approximation in the post processing is therefore
not responsible for this difference. Two hypotheses are proposed:

• The equivalent Smagorinsky constant value for SALSA-DES is higher than that
of CEASM-DES (see Tab. 8.1 in Section 8.2). The equivalent CS for LLR-DES
cannot be determined analytically, however this could be expected to be similar
the value for WCX-DES upon which the LLR model is based.

• The SALSA model may inherit the behaviour identified by Breuer et al. [18] and
Riou et al. [125] for SA-DES, discussed in Section 3.7.6: The damping terms f v1 ,
f v2 and f w were observed to give rise to excessive eddy viscosity in the early
shear layer.

It is of course possible that both hypotheses are true: The first would be responsi-
ble for the general level of νt throughout the wake and the second for a localised
increase at the shear layer. The k mod levels in the wake show a similar distribution
for each model, with the increase at x/c ≈ 1.3 arising due to the increase in Δ there
(see Fig. 6.6).

Whatever the cause of this model dependent behaviour, the global solution appears
to be insensitive to different levels of eddy viscosity in the LES-mode region, with
a negligible difference observed. This is probably because the natural instability of
the shear layer is so strong for this massively-separated flow that the development
of resolved turbulent structures is not damped by the higher levels of eddy viscosity
produced by SALSA-DES. Indeed, the levels of kres in the early shear layers are very
similar for all models (Fig. 8.12). The apparent insensitivity of such natural DES so-
lutions to the general eddy viscosity level in the wake region fortifies some elements
of the discussion on dynamic DES made in Sect. 8.2.
10
It should be pointed out that much of the resolved kinetic energy is due to the coherent von-Kármán vortex
shedding and is hence not entirely turbulent in origin. Separating the coherent and incoherent, turbulent
motions would require phase averaging (Appendix F).

144
 8.4 Sensitivity of DES to the RANS model used

(a) k res, SALSA-DES97 (b) k mod , SALSA-DES97

(c) k res, LLR-DES97 (d) k mod , LLR-DES97

(e) k res, CEASM-DES97 (f) k mod , CEASM-DES97

Figure 8.12: Resolved kinetic energy (left) and time-averaged modelled kinetic energy (right)
(time and spanwise-averaged data).

Spectral content
Having established the negligible model dependency in the time-averaged forces
and wake field, the unsteady content of the solutions should be compared. The
power spectral densities of the four integral force components have been evaluated
and are compared in Fig. 8.14. The spectra are obtained from the force time traces ob-
tained at each spanwise slice, which are then averaged to provide smoothing. Each

145
8 Demonstration and Analysis

(a) SALSA-DES97 (b) LLR-DES97 (c) CEASM-DES97

Figure 8.13: Time-averaged eddy viscosity ratio, νt /ν, and streamlines compared in the leading
edge region (time and spanwise-averaged data).

Cl Cd Cn Ct
1
10
0
10
-1
10
PSD

-2
10
-3
10
-4
10
SALSA-DES
10
-5 LLR-DES
CEASM-DES
-6
100.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10

St St St St

Figure 8.14: Power spectral density of sectional force signals (lift, drag, chord-normal and
chord-tangential components).

PSD computation is conducted over a multiple of windows using the Hanning win-
dow function and then averaged11 .

The spectra are essentially identical, showing that the negligible model dependency
for this case extends to the unsteady flow prediction as well as time-averaged quan-
tities. The spectra for the different components show interesting differences, with
the tangential component exhibits a very strong peak corresponding to the coherent
vortex shedding. The corresponding Strouhal numbers (calculated as St = f c/|U∞ |)
for the vortex shedding frequency are listed in Tab. 8.6 and are very similar. The
fact that the greatest remaining differences between the spectra occur in the low fre-
quency range lends weight to the assumption that the small deviations between the
simulations are due to imperfect statistical convergence, as the low-frequency por-
tion of the spectra is the most sensitive to statistical convergence errors.

11
The spectral processing employing averaging of overlapping windows was carried out using the in-house
tool dftavg written by D. Eschricht.

146
 8.4 Sensitivity of DES to the RANS model used

8.4.2 Separating-reattaching internal flow: Bump in a rectangular duct

Having established a near independence of the results on the DES background model
for the NACA0021 flow in the preceding section, the level of model dependency will
be investigated for a strongly contrasting flow case: A separating-reattaching flow
with a very sensitive pressure-induced separation from a smoothly-curved surface
and complex three-dimensional phenomena. The flow and the expected challenges
it presents are outlined in Sect. 6.6 together with the test case setup. The CEASM-DES
with model-specific GIS shield (Sect. 5.5.4) and the SAE-DDES implementations will
be compared, which represent RANS background models occupying different levels
in the modelling hierarchy (Fig. 3.5, Sect. 3.4.3). Unfortunately, different time step
sizes were computed in each case, which strictly limits the comparability of the com-
putations. The time step sizes together with the number of time steps computed are
summarised in Tab. G.1. The statistical sample lengths, although less than those em-
ployed in the NACA0021 case, are considered adequate as no low-frequency modu-
lation occurs in this flow (e.g. shear layer flapping) [4].

The qualitative nature of the flow topology along the symmetry plane is the same for
both models (Figs. G.1): An initial separation occurs near the bump crest, followed
shortly by an early flow reattachment. The flow separates again near the bump foot
before reattaching somewhat further downstream. The location of these phenomena
is however predicted very differently by each model, with the CEASM predicting
a longer and larger recirculation region. These features are also clearly visible in
the centreline pressure coefficient plot of Fig. G.2(a): The earlier reattachment of the
SAE-DDES gives rise to an upstream shift of the pressure recovery region. It is fur-
thermore notable that the differences due to the background model are much more
significant than those between the DDES and IDDES formulations.

The positions of the separation and reattachment points are also listed in Tab. G.1,
from which it can be seen that the CEASM model predicts a slightly earlier initial
separation point than the SAE model. In fact, the difference in separation point pre-
diction amounts to only one streamwise grid cell. It is commonly observed however
that small differences in the prediction of the separation point can lead to strong dif-
ferences in the extent of the recirculation region (see e.g. the periodic hills and Obi
diffuser test cases studied in the FLOMANIA project [55]). A quantitative compari-
son of the velocity field profiles is given in Fig. G.3, from which the greater strength
and size of the recirculation region predicted by the CEASM-DES is again observed.

The modelled, resolved and total Reynolds shear stresses are compared in Fig. G.4.
Generally, the maximum levels are predicted very similarly by both models, with a
displacement of the peak locations caused by the variation in the shear layer loca-
tion. Some notable exceptions however occur, firstly in the early shear layer region
(at x = −0.079) where greater levels of modelled stress are predicted by the SAE-
DDES. This is most likely the same phenomenon observed for the NACA0021 case in
Sect. 8.4.1 (see Figs. 8.12 & 8.13), albeit for the SALSA model, which was suspected

147
8 Demonstration and Analysis

to arise from an analogous effect to that reported by Breuer et al. [18] and Riou et
al. [125] (described in Sect. 3.7.6). Secondly, the CEASM-DES profiles show greater
levels of modelled turbulence in both the upper and lower wall boundary layers
downstream of the reattachment location (around x > 0.6). The upper wall bound-
ary layer turbulence is furthermore purely modelled by the CEASM-DES, whereas
the growth of some resolved content is apparent for the SAE-DDES at the x = 0.925
station. The infiltration of resolved turbulence into the upper wall boundary layer
in the case of SAE-DDES is visible in the instantaneous vortex core plot of Fig. G.5
and the fact that this is sustained implies that the DDES shield has receded inside the
boundary layer. The GIS shield functions are plotted in Fig. G.9, which indeed re-
veals a somewhat thicker layer of shield function activity in the case of the CEASM-
DES. This is assumed to cause the absence of resolved turbulence along the upper
wall boundary layer and the greater level of modelled turbulence at the lower wall.

The issue of the model-dependency of the initial separation position will be returned
to. The analysis has so far been confined to data at the symmetry plane, however
three-dimensional effects are of great importance in this low-aspect ratio geometry.
In Fig. G.6 the three-dimensional topology of the mean flow is shown, which is qual-
itatively identical for all computations: Large reverse flow regions are predicted in
the corners downstream of the bump, which coincide with a pair of counter-rotating
streamwise vortices. These vortices induce a strong downwards motion of the core
flow near the symmetry plane (visible also in Fig. G.7), which is responsible for the
rapid reattachment observed down the bump slope. The streamwise vortices pre-
dicted by the SAE-DDES are stronger than those of the CEASM-DES, which is seen
in Fig. G.7 to give rise to a stronger W-shaped distortion of the shear layer. The shear
layer predicted by the CEASM-DES is in contrast relatively flat in the channel centre,
up to about one bump height away from each side wall. The higher degree of three-
dimensionality of the flow for the SAE-DDES is furthermore reflected in the pressure
distributions at various z positions depicted in Fig. G.2(c). The pressure profiles dif-
fer very strongly between the spanwise stations, whereas those for CEASM-DES are
highly similar (Fig. G.2(b)).

A further difference between the solutions that can be directly linked to the mod-
elling hierarchy (Sect. 3.4.3) concerns the corner vortices observed at the prescribed
inlet12 , which are depicted in Fig. 6.11. Whereas the anisotropy-resolving CEASM
model is capable of propagating these vortices downstream into the solution do-
main, these are rapidly damped by the linear SAE model: This is shown in Fig. G.8,
where instantaneous contours of the streamwise vorticity at the slice x = −0.2 are
compared. This leads to a notable difference in the corner flow topology at the sepa-
ration region predicted by each model. The extent to which this influences the down-
stream corner vortex strength or indeed the separation point prediction is however
uncertain.

12
Recall from Sect. 6.6 that the inflow data was provided by a DRSM computation by Ansys Inc.

148
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

In summary therefore, a stark contrast is seen with the results for the NACA0021
case, which were shown to be essentially model-independent. Because of the high
importance of the separation point prediction and the strong reliance of this upon
the upstream RANS modelling, a very strong level of DES model dependency is wit-
nessed for this flow.

8.5 Comparison of DES with experiment for the flow around a

circular cylinder
The test case of a circular cylinder in a square channel at Re = 140000, described
in detail in Sect. 6.5, is employed as a validation case for DES working in its in-
tended application environment of massively-separated flow. The subject of the in-
vestigation is the CEASM-DES variant with the model-specific GIS-shield function
described in Sect. 5.5.4 and by Bunge et al. [21]. The laminar cylinder boundary lay-
ers existing at this precritical Reynolds number (Sect. 2.5.2) in fact limit the validation
to the LES-mode operation of the DES13 . The highly detailed experimental data al-
lows many levels of comparison, and the result will be presented according to the
type of post-processing carried out. The comparison represents a validation of DES
to an unprecedented level of detail.

8.5.1 Transition behaviour

In a similar manner to Shur et al. [147] and Travin et al. [171], it was decided to
manually adjust the transition behaviour of the RANS model to enforce the laminar
cylinder boundary layers known to exist in the experiment. This is justifiable, as it
is widely acknowledged that the prediction of laminar to turbulent transition falls
outside the scope of RANS models. Without user intervention or the use of an ex-
ternal transition prediction method, RANS models tend to predict boundary layer
turbulence too early on in their development.

The authors cited above developed and applied a technique they named the “trip-
less” approach, whereby a laminar boundary layer is ensured up to the separation
point by exploiting the sensitivity of certain model terms to very low levels of free-
stream eddy viscosity. If applied using a uniform inflow profile however, the tripless
approach would also laminarise the side wall boundary layers. Although the tripless
approach would be applicable with a non-uniform inflow profile of the turbulence
quantities, it was decided to adopt an alternative approach. Instead, a minor adjust-
ment to the CEASM model formulation was undertaken to delay its inherent bound-
ary layer transition behaviour to a sufficient extent to ensure laminar separation on
the cylinder whilst maintaining the turbulent side walls. The laminar separation
from the cylinder is verified in the eddy viscosity ratio plot of Fig. 8.15.
13
Although the RANS mode operation is active on the channel side walls, this is not expected to play an
important role in the flow.

149
8 Demonstration and Analysis

Figure 8.15: Instantaneous velocity vectors (at every second circumferential grid line) tangen-
tial to the spanwise median plane and contours of eddy viscosity ratio.

8.5.2 Instantaneous flow

Instantaneous snapshots of the flow field have already been portrayed in Figs. 8.1
& 8.2, and a description of the resolved vortex structures was given in Sect. 8.1.1.
That discussion will now be extended to cover the temporal behaviour of the flow,
based on the time traces of the pressure lift and drag integrated over the entire cylin-
der shown in Fig. 8.16. The lift component exhibits strong oscillations around zero
with a very regular frequency and irregular amplitude. The drag component by con-
trast contains higher-frequency fluctuations of a generally weaker amplitude. The
irregularity of the time traces, with a strong degree of low-frequency modulation, is
considered analogous to the weak shedding cycle behaviour already noted for the
NACA0021 profile in Sect. 8.4.1. The strong tonal nature of the lift coefficient and
more broadband nature of the drag coefficient furthermore compares well with the
behaviour seen for the tangential and normal force components of the NACA0021,
respectively. Strong, medium and weak shedding cycles can clearly be discerned
from the lift trace, which correspond respectively to higher, intermediate and lower
levels of the drag coefficient. Because the experimental surface pressure readings
were obtained from a single tapping, no time traces of the force coefficients are avail-
able from the experiment so a direct comparison of this character is not possible.
However, the nature of the low-frequency modulation compares very well with the
time traces published by Szepessy & Bearman [168] at a similar Reynolds number
and aspect ratio, although at a much lower blockage coefficient.

Comparison of the velocity fields in the wake can however be made between the
DES and the experiment. Fig. 8.17 is adapted from the joint numerical and exper-
imental investigation of the flow physics published by Perrin et al. [112]. Velocity
field snapshots from instants identified as strong and weak shedding cycles are com-
pared, these being identified using surface pressure time traces in the case of the

150
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

3
CL
CL , CD 2 CD

-1

-2
200 300 400 500 600
t*

Figure 8.16: Time traces of lift and drag coefficient over the time range used for statistical anal-
ysis (pressure components only, integrated over the entire cylinder).

experiment and point velocity signals in the DES14 . From the snapshots, a strong
flow asymmetry and large-scale, coherent vortex shedding can be seen during the
strong shedding. By contrast, the near wake flow is symmetric and populated with
chaotic small-scale structures during weak shedding. The formation region of the
large scale vortices is furthermore displaced downstream. The comparability of the
DES and experiment in this respect is excellent. In addition to validating the ability
of the DES to predict such unsteady flow phenomena, this investigation has there-
fore also provided evidence to support the hypothesis of Szepessy & Bearman [168]
(mentioned in Section 2.5.2) that the weak shedding cycles are associated with in-
creased three-dimensionality of the wake flow. A similar observation was made for
the NACA0012 test case by Mockett et al. [94].

8.5.3 Time-averaged flow

Some time-averaged quantities are portrayed at a spanwise slice at y = 0 in Fig. 8.18.
It is seen that the mean flow is almost two-dimensional over a large portion of the
span in the central region, up to about 1D from each side wall. The regions of appre-
ciable near-wall effect then extend slowly with increasing downstream distance. It is
also apparent that the averaged quantities exhibit some statistical scatter, in particu-
lar the spanwise component of the resolved Reynolds stress. In order to enhance the
statistical convergence, additional averaging has been performed over the observed
homogeneous central region, as described in Sect. 6.5.

Running averages of the pressure lift and drag integrated over the cylinder are sh-
own in Fig. 8.19, using the technique discussed in Section 8.4.1. These plots show
14
The comparability of these quantities was demonstrated in the publication [112].

151
8 Demonstration and Analysis

0.6 0.6

0.4 0.4

0.2 0.2
y/D

y/D
0 0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2
x/D x/D

2 2

0 0
s

s
-2 -2

26.6 26.8 27 t 27.2 27.4 26.8 27 27.2 t 27.4 27.6

(a) PIV

(b) DES

Figure 8.17: Near wake behaviour at instants where the vortex shedding is strong (left) and
weak (right). The time trace depicted from the experiment (upper) is of surface
pressure, whereas the velocity at y/D = 0.5 and x/D = 1 is shown for the DES
(lower).

how the time-averaged force values develop as an increasing number of time steps
are computed for averaging, and cover the time range used for Reynolds averaging
(i.e. after the removal of initial transient character). The lift component converges
very quickly as the mean value is not changed by switches from strong to weak
shedding patterns. The fact that the lift inexplicably converges to a slightly non-zero
value is cause for some concern for this symmetric geometry. The drag force drifts
considerably however, the value still showing a fairly strong downwards trend near
the end of the statistical range (because of a protracted absence of strong shedding
behaviour). A considerable statistical convergence error must therefore be assumed
despite the respectable time sample employed for averaging (14000 time steps, cor-
responding to a physical time of 420D/U∞ or roughly 85 vortex shedding periods).

Figure 8.20 presents a comparison of the time-averaged streamlines between the

152
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

(a) U/U∞ (b) u u res /U∞

2

(c) v  v  res /U∞

2
(d) w  w  res /U∞
2

Figure 8.18: Spanwise homogeneity of the flow: various statistical quantities on the slice
y/D = 0.

0.1 1.6

0.05
1.5
CD
CL

1.4
-0.05

-0.1 1.3
200 300 400 500 600 200 300 400 500 600
t* t*

Figure 8.19: Running averages of the pressure lift and drag coefficients (integrated over the
entire cylinder) over the time range used for statistical analysis.

DES and the experiment, showing excellent qualitative agreement. The quantita-
tive agreement of the mean recirculation length (measured from the rear surface of
the cylinder at x/D = 0.5) is also very good, as summarised in Tab. 8.7. Correspond-
ingly, the mean pressure drag obtained at the spanwise median plane (z = 0) and
likewise listed in Tab. 8.7 also shows very good agreement with the experiment. The
fact that the total drag coefficient (i.e. including the effects of skin friction) integrated
over the entire span is so similar reflects the fact that pressure drag is dominant for
such bluff-body flows. This can also be interpreted as a further indicator of the high
degree of spanwise homogeneity observed above.

A more detailed quantitative comparison of the mean velocity in the wake is offered

153
8 Demonstration and Analysis

Cd,p CD,p CD Recirc. length

Experiment 1.45 — — 0.78
CEASM-DES 1.48 1.48 1.49 0.85

Table 8.7: Time-averaged drag coefficient (pressure drag on spanwise median plane, Cd,p , pres-
sure drag over whole cylinder, CD,p and total drag over whole cylinder, CD ) and
recirculation length (relative to the cylinder rear surface) for DES and experiment.

(a) PIV (b) DES

Figure 8.20: Comparison of the time-averaged streamlines between DES and experiment.

by Fig. 8.22 at vertical lines traversing the wake at a range of x/D positions. The
agreement is generally excellent, although the profiles of the vertical velocity com-
ponent show a stronger inflective nature. The small deviations could to some extent
be a result of the remaining statistical convergence error noted above.

Contour plots of the Reynolds-averaged turbulent stresses are depicted in Fig. 8.21
in comparison with 2C and 3C PIV. The qualitative and quantitative agreement of the
u  u  stress is excellent, whereas the effect of the mildly excessive recirculation length
can be seen for the v v and w w components. Furthermore, the peak levels of v v
are slightly over-predicted, whereas those of w w are slightly under-predicted.

A more detailed comparison can again be made with reference to the profiles of
Fig. 8.22. The level of u  u  at the symmetry line (y/D = 0) agrees almost perfectly
with the experiment. As the profiles move towards the shear layer however, a consis-
tent over-prediction is apparent. The vertical component, v v in contrast reproduces
the shape of the experimental plots very well, although the magnitude is under-
predicted at the more upstream stations and over-predicted further downstream.
The profiles coincide very well at x/D = 1.3. Finally, the shear stress component u  v
is in excellent agreement with the experimental data, except perhaps at x/D = 1.0,
where the DES curve is excessively inflective.

154
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

(a) u u res (b) v  v  res (c) w  w  res

Figure 8.21: Comparison of the resolved turbulent stress components from the DES below with
PIV data above.

The Reynolds stress profiles in Fig. 8.22 also compare the resolved-only and total
(i.e. resolved plus modelled) contributions (computed as described in Appendix E).
The modelled contribution is clearly negligible in comparison to the resolved com-
ponent, which could be interpreted as a sign of sufficient grid resolution (discussed
in Sect. 3.5.2). However, the resolved part includes both the coherent and incoherent
contributions to the motion, and is not purely made up of turbulent motion. To make
a statement concerning the sufficiency of the grid resolution therefore, the coherent
contribution should be subtracted using phase averaging (Appendix F).

8.5.4 Spectral content

Velocity time traces have been extracted from the TRPIV measurements for a range of
positions in the near wake shown in the lower right-hand corner of Fig. 8.23. Spec-
tra obtained for the U and V velocity components at each location are compared
between DES and the TRPIV measurements in Fig. 8.23. The agreement seen is gen-
erally excellent, with reproduction of the variations in the spectra from point to point
(e.g. the elimination of the peak due to symmetry for the U component at point 9, lo-
cated near the centre line). Some significant deviations do however occur: The peak
in the spectra represents the quasi-periodic vortex shedding, and occurs at a Strouhal
number of St = 0.21 in the experiment whereas the DES slightly over-predicts this at
St = 0.23. A fairly sharp drop off is furthermore seen at the higher frequencies, after
around St ≈ 2 (the precise value appearing to vary from point to point), which is
due to either the spatial or temporal filtering of the simulation. The simulation time
step corresponds to St ≈ 30, which when combined with the Nyquist criterion [104]
results in a temporal cut-off at St ≈ 15. It is therefore apparent that the spatial reso-
lution, rather than the time step is responsible for this behaviour.

155
8 Demonstration and Analysis

1
x / D = 0.8 x / D = 1.0 x / D = 1.3 x / D = 1.8
0.8
0.6
0.4
0.2
y/D

PIV
0 DES
-0.2
-0.4
-0.6
-0.8
-1
-0.4 0 0.4 0.8 1.2 1.6 -0.4 0 0.4 0.8 1.2 1.6 -0.4 0 0.4 0.8 1.2 1.6 -0.4 0 0.4 0.8 1.2 1.6
U / Uref U / Uref U / Uref U / Uref

1
0.8
0.6
0.4
0.2
y/D

PIV
0 DES
-0.2
-0.4
-0.6
-0.8
-1
-0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4 -0.4 -0.2 0 0.2 0.4
V / Uref V / Uref V / Uref V / Uref
1
0.8
0.6
0.4
0.2
y/D

PIV
0 DES, res
-0.2 DES, tot
-0.4
-0.6
-0.8
-1
0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5 0 0.1 0.2 0.3 0.4 0.5
2 2 2 2
u’u’ / Uref u’u’ / Uref u’u’ / Uref u’u’ / Uref
1
0.8
0.6
0.4
0.2
y/D

PIV
0 DES, res
-0.2 DES, tot
-0.4
-0.6
-0.8
-1
0 0.2 0.4 0.6 0 0.2 0.4 0.6 0 0.2 0.4 0.6 0 0.2 0.4 0.6
2 2 2 2
v’v’ / Uref v’v’ / Uref v’v’ / Uref v’v’ / Uref
1
0.8 PIV
0.6 DES, res
DES, tot
0.4
0.2
y/D

0
-0.2
-0.4
-0.6
-0.8
-1
-0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2 -0.2 -0.1 0 0.1 0.2
2 2 2 2
u’v’ / Uref u’v’ / Uref u’v’ / Uref u’v’ / Uref

Figure 8.22: Comparison of time-averaged profiles at various x/D slices between 2C-PIV and
DES (time and spanwise-averaged data).

156
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

U, point 1 V, point 1 U, point 2 V, point 2

10

PIV
1 DES
-5/3
St
0.1
PSD

0.01

0.001

0.0001
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10

St St St St
U, point 3 V, point 3 U, point 4 V, point 4
10

PIV
1 DES
-5/3
St
0.1
PSD

0.01

0.001

0.0001
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10

St St St St
U, point 5 V, point 5 U, point 6 V, point 6
10

PIV
1 DES
-5/3
St
0.1
PSD

0.01

0.001

0.0001
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10

St St St St
U, point 7 V, point 7 U, point 8 V, point 8
10

PIV
1 DES
-5/3
St
0.1
PSD

0.01

0.001

0.0001
0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10 0.01 0.1 1 10

St St St St
U, point 9 V, point 9
10

PIV
1 DES
-5/3
St
0.1
PSD

0.01

0.001

0.0001
0.01 0.1 1 10 0.01 0.1 1 10

St St

Figure 8.23: Comparison of U and V velocity spectra between DES and experiment (TRPIV) at
nine probes in the near wake region.

157
8 Demonstration and Analysis

(a)  Ωz , ϕ = 90◦ (b)  u u , ϕ = 90◦

(c)  Ωz , ϕ = 180◦ (d)  u u , ϕ = 180◦

(e)  Ωz , ϕ = 270◦ (f)  u u , ϕ = 270◦

(g)  Ωz , ϕ = 360◦ (h)  u u , ϕ = 360◦

Figure 8.24: Phase-averaged spanwise vorticity  Ωz  (left) and streamwise turbulent stress
 u u  (right) for phase angles ϕ = 90◦ , 180◦ , 270◦ and 360◦ . Contour shading
from DES. Contour lines depict values of  Ωz  = ±1 and  u  u   = 0.2 for the PIV
(——) and DES (– – –).

158
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

(a)  v  v  , ϕ = 90◦ (b)  u v  , ϕ = 90◦

(c)  v  v  , ϕ = 180◦ (d)  u v  , ϕ = 180◦

(e)  v  v  , ϕ = 270◦ (f)  u v  , ϕ = 270◦

(g)  v  v  , ϕ = 360◦ (h)  u v  , ϕ = 360◦

Figure 8.25: Phase-averaged vertical turbulent stress  v v  (left) and turbulent shear stress
 u v  (right) for phase angles ϕ = 90◦ , 180◦ , 270◦ and 360◦ . Contour shading
from DES. Contour lines depict values of  v v  = 0.2 and  u  v  = ±0.1 for the
PIV (——) and DES (– – –).

159
8 Demonstration and Analysis

8.5.5 Phase-averaged flow field

The quasi-periodic nature of the vortex shedding enables an analysis using phase-
averaging to separate the coherent vortex shedding from the incoherent turbulent
motions. Reference is made to Appendix F, where this process and the notation of
the different components are described. Phase averaging has been performed for
both the PIV and simulation data, using the surface pressure on the cylinder as a
trigger signal [108, 112], facilitating a deeper comparison of the unsteady flow pre-
diction. The left hand column of Fig. 8.24 shows the phase-averaged spanwise vor-
ticity, Ωz , for the phase angles ϕ = 90◦ , 180◦ , 270◦ and 360◦ . The contour shading
and dashed contour lines are taken from the DES, whereas the solid contour lines are
from the PIV. In both cases, the contour lines represent the values Ωz  = ±1. The
positions of the shed vortices are predicted very well by the DES, although a slight
and consistent upstream shift is observed. Considering that a downstream shift in the
wake length was seen in the time-averaged data (Fig. 8.20), this is at first sight para-
doxical. A viable mechanism to explain this can however be outlined: The snapshots
falling within the weak shedding cycles cannot easily be allocated a phase angle, due
to the breakdown of the regular shedding motion. For this reason, the weak shed-
ding events are detected using a procedure described in [110] and neglected from
the phase averaging. The discussion surrounding Fig. 8.17 indicated that the weak
shedding cycles are associated with a lengthening of the wake. The weak and strong
shedding behaviour emerges sporadically. If the DES time series employed for sta-
tistical analysis happens to involve a greater proportion of weak shedding behaviour
than the experimental sample, this would explain the phenomenon. To prove or dis-
prove this, a selective Reynolds averaging could be carried out employing the same
criteria as employed in the phase averaging to neglect the weak shedding events.

Nonetheless, the discrepancy between the phase-averaged vortex locations of the

DES and experiment is minor. The process by which the shear layers roll up into
the coherent vortex street is hence predicted very well by the DES, both qualitatively
and quantitatively.

The phase-averaged incoherent contributions to the turbulent stresses, u i u j  are
similarly plotted for the same phase angles in the right hand side of Fig. 8.24 and in
Fig. 8.25. The line contours continue to represent the same values for both the DES
and the experiment, with the thresholds u  u   = v v  = 0.2 and u  v  = ±0.1
chosen. The topology of the incoherent stresses agrees very well with the experi-
ment over all phase angles: The shear layers exhibit maxima of u  u   and u  v ,
the vortex centres maxima of u  u   and v v  and the saddle points maxima of
u  v . However, the magnitude appears in all cases to be under-predicted by the
DES. Possible explanations for this are suggested:
• The modelled turbulence quantities were not included in the phase averaging
procedure so the u i u j  data hence represents the resolved portion only. In-
cluding the modelled contribution would thus improve the agreement.

160
 8.5 Comparison of DES with experiment for the flow around a circular cylinder

• A related point concerns the grid resolution: The spectra of Fig. 8.23 demon-
strate a sharp cut-off in the resolved energy at high frequencies due to the grid
filtering. Ideally, this energy would be represented by the subgrid-scale model.
That the cut-off concerns amplitudes at least a decade lower than the strongest
resolved incoherent scales implies however that the unresolved quantities are
minimal.

• Statistical convergence: this was shown to be imperfect for the time-averaged

drag, and the situation can be expected to be much worse for the quantities
considered here – the number of samples is considerably less (around 85 per
phase angle compared to 14000 for the Reynolds-averaging) and second-order
statistics are furthermore considered. The poor statistical convergence is re-
flected in the fairly noisy appearance of the data.

• An under-prediction of the incoherent contributions to the Reynolds stresses

together with the slight over-prediction of the total Reynolds stresses (e.g. Fig.
8.21) could arise from a prediction of excessive coherent motion strength in the
simulation.
It could indeed be the case that a combination of these possibilities is in effect. It
is in any case clear that the balance of coherent and incoherent motion is strongly
dependent on the statistical sample: Predominantly coherent motion has been asso-
ciated with the strong shedding regime and an incoherent character with the weak
shedding (Fig. 8.17). The transition between these regimes is apparently random and
with a very long time scale (around an order of magnitude larger than the shedding
period). All of this serves to underline the fact that the collection of phase-averaged
data from such simulations requires extremely long time samples that well exceed
those obtainable using currently acceptable computing capacity. As described in
Section 6.5, an alternative phase averaging technique can be applied whereby POD
coefficients from each velocity field snapshot are used to determine the phase angle.
This has since been applied to this data by Perrin et al. [110], and an improvement
in the quality of the data (in particular a cleaner separation of the coherent and inco-
herent contributions to the motion) was reported.

Despite these limitations, it can nonetheless be said that the prediction of the phase-
averaged motion by the DES agrees well with the experiment to within the level of
accuracy that can be obtained statistically. The validation of the DES with experi-
mental data for the cylinder case can hence be considered extremely successful, with
impressive levels of agreement achieved in the integral drag coefficient, the mean ve-
locity field and Reynolds stresses, the vortex shedding frequency, the velocity spectra
and the phase-averaged data.

One caveat must however be acknowledged. In Sect. 6.5, a serious limitation in the
comparability between experiment and simulations was identified for this test case,
namely the neglection of the appreciable levels of free stream turbulence (the effect

161
8 Demonstration and Analysis

of which was described in Sect. 2.5.2). In the following Sect. 8.6, an extreme sensi-
tivity of the flow prediction to the numerical time step size will be presented and
discussed for the cylinder case. Because only two values of time step have been com-
puted, there is no reason to suppose that the finer time step applied here corresponds
to time step convergence. It is highly possible therefore, that a further refinement of
the time step would in fact give rise to a degradation of the agreement with the ex-
periment. It is thought that a cancellation of errors is in effect here, i.e. the neglected
FST effects are compensated for by a residual level of time filtering. Dissatisfying
though this is to the purist, the value of these results is not diminished entirely: They
show that provided the correct separation point and shear layer transition behaviour
occurs, excellent wake flow prediction can be achieved by the adopted simulation
methodology.

8.6 Time step sensitivity

Besides the detailed validation of DES presented in the preceding section, the cylin-
der test case was also employed to determine the sensitivity of the solution to the
numerical time step size. As has been established in Sect. 3.5.2, a time step size cor-
responding to a CFL number [31] of unity is a common guideline for LES. This there-
fore applies in the LES mode region of DES, and indeed this guideline is invoked for
DES by Spalart [154]. From the point of view of industrial practice however, such
limitation of the time step gives rise to a considerable increase in the computational
cost of DES. Particularly in comparison to URANS, for which significantly coarser
time steps can be applied, the LES time step criterion often results in surprisingly
high computational costs for DES (as the focus when estimating computational cost
often appears to be on the grid cell count). It is of great practical interest therefore
to quantify the influence of a coarser time step in DES. Can an approximately valid
solution be obtained at a fraction of the computational cost in this manner?

In addition to the time step Δt = 0.03D/|U | ∞ applied in Sect. 8.5, a slightly coarser
time step of Δt = 0.05D/|U | ∞ has been used. In all other respects, the simulation
setup is identical. As mentioned in Sect. 6.5, the same physical time sample length
has been computed for statistical evaluation.

An idea of the impact of the time step size can be obtained from the instantaneous
field. Figure 8.26 shows the resolution of vortical structures and the model eddy
viscosity in the early shear layer for an individual snapshot for each time step size.
Whereas the shear layer for the finer time step shows the development of instabilities,
the shear layer for the coarser time step remains smooth and stable. Moreover, al-
though the resolved turbulence is significantly reduced by the coarser time step, the
level of modelled turbulence (represented by the eddy viscosity ratio in Figs. 8.26(c)
& 8.26(d)) remains very similar.

It appears then as if the coarser temporal resolution does not allow the formation of

162
 8.6 Time step sensitivity

(a) Ωz , dt = 0.05D/ |U | ∞ (b) Ωz , dt = 0.03D/ |U | ∞

(c) νt /ν, dt = 0.05D/ |U | ∞ (d) νt /ν, dt = 0.03D/ |U | ∞

Figure 8.26: Effect of time filtering on resolved shear layer structures and modelled turbulence
for coarse and fine time steps (contours of spanwise vorticity Ωz and eddy viscos-
ity ratio νt /ν on a slice at the spanwise mid-section).

these structures in the shear layer, and that a time filtering effect occurs. To investi-
gate the potential of the CFL number as a rule of thumb, a temporal filter width is
constructed as Δ t = |U | Δt. When the ratio of this to the local spatial filter width is
considered, a kind of CFL number emerges:
Δt Δt |U |
CFL =
= , (8.13)
Δx Δx
where |U | is the local velocity magnitude and Δ x is the DES grid length scale. The
precise definition of the quantities making up the CFL number is in some respects
open to interpretation. The choice of quantities here is motivated by the interest in
balancing the spatial and temporal capability to resolve turbulence, hence the use
of the largest grid dimension for Δ x . The expectation is therefore that a criterion of
CFL ≤ 1 is necessary in all regions of resolved turbulent flow.

Figure 8.27 shows the distribution of the CFL number for the coarse and fine time
steps. It is seen that the criterion is met throughout the domain for the finer time
step, whereas values greater than unity emerge, notably at the edge of the early shear
layer, for the coarser time step. From these observations, it seems as if this kind of
CFL criterion could indeed offer a sound basis for the judgement of the sufficiency
of the time step size for DES.

To provide a quantitative assessment of the time step influence, the time-averaged

and phase-averaged flow field is considered. The time-averaged streamlines are

163
8 Demonstration and Analysis

(a) CFL, dt = 0.05D/ |U | ∞ (b) CFL, dt = 0.03D/ |U | ∞

Figure 8.27: Extent of time filtering of resolved scales depicted using a CFL number for the
coarse and fine time steps (slice at the spanwise mid-section).

(a) PIV (b) DES, dt = 0.05D/U∞

Figure 8.28: Comparison of the time-averaged streamlines for the DES with the coarse time
step and experiment.

compared with the PIV data in Fig. 8.28. Recall that an excellent agreement was
observed between the experiment and the fine time step in Fig. 8.20, whereas the
coarse time step exhibits an excessive recirculation length. The recirculation length
(measured from the rear surface of the cylinder at x/D = 0.5) is 1.36, in compari-
son to 0.78 for the experiment and 0.85 for the fine time step. This is accompanied
by a strong under-prediction of the pressure drag coefficient at the centre plane at
Cd,p = 1.12 in comparison to 1.45 and 1.48 for the experiment and fine time step, re-
spectively. It is interesting to note that all of the cases presented by Travin et al. [171]
in their cylinder flow DES study suffered from an excessive recirculation length. As
these employed the same time step size as the coarse time step computed here, this
suggests (at least for higher Re values) that insufficient time resolution may have
been a contributory factor.

164
 8.6 Time step sensitivity

Figure 8.29: Phase-averaged spanwise vorticity  Ωz  for a single phase angle ϕ = 45◦ . Contour
shading from DES with finer time step. Lines depict values of  Ωz  = ±1 for the
PIV (– · –·), DES with finer time step (—) and coarser time step (– – –).

Turning to the phase-averaged flow, contours of the phase-averaged spanwise vor-

ticity Ωz  are compared in Fig. 8.29 for the single phase angle ϕ = 45◦ between both
time step sizes and the PIV data. As established in greater detail in Sect. 8.5, the po-
sition of the shed vortices is predicted very well by the finer time step. The roll-up of
the shear layer can however be seen to be significantly delayed for the coarser time
step, resulting in a downstream displacement of the vortex locations.

In summary then, these results indicate that the coarse time step hinders the develop-
ment of turbulent structures in the early shear layer, and that this deficit in resolved
turbulence is not compensated for by an increase in modelled turbulence. Even this
relatively mild coarsening of Δt causes a drastic degradation in the prediction of the
wake flow, which serves to underscore the importance of sufficient time resolution.
As a result, any application of DES with a coarser time step for pragmatic reasons
should not be recommended .

The fact that the deficit in LES turbulence caused by the time filtering is not compen-
sated by RANS turbulence is caused by the dependency of the LES/RANS switching
on the spatial resolution alone. As such, some inclusion of the temporal filter width
in the DES length scale definition, such as that proposed by Bush & Mani (Sect. 3.7.6)
may hold some promise to reduce the time step dependency. The goal of such a
modification would be to provide a more gradual and predictable decay of solution
quality with increasing Δt. In the limit case of a very coarse time step, such a mod-
ified DES formulation should tend to a URANS solution. This would significantly
enhance the robustness of the method in an industrial context, perhaps making the
coarsening of the time step a justifiable means to sacrifice a limited degree of solution
quality in the interests of a reduced computational burden.

It could be argued that this case is unusually sensitive to time step effects, and that
the strong warning issued regarding practical applications could be exaggerated: Es-

165
8 Demonstration and Analysis

pecially at this precritical Reynolds number (Sect. 2.5.2), even small changes in the
shear layer turbulence level would have a very strong impact on the global flow.
However, it is believed that this case is highly representative of natural DES applica-
tions to bluff bodies in general. The highest CFL numbers in the flow can be expected
to occur at the region of highest velocity, which would likely coincide with the sepa-
ration point. For natural DES the upstream turbulent boundary layer structures are
modelled entirely by RANS, meaning that the LES resolved structures in the shear
layer must develop via a quasi-transition mechanism (as noted in the description of
the grey area problem, Sect. 3.7.3). A high sensitivity to the time step size can hence
be anticipated for all such flows, even at turbulent separation Reynolds numbers.
Only low Reynolds number flows where the early shear layer is laminar should be
expected to be more tolerant of coarser time steps.

8.7 Comparison of DES with experiment for the bump test case
Excellent agreement with reliable and comprehensive experimental data has been
reported in Sect. 8.5 for the type of flow for which DES was intended, namely mas-
sively separated bluff body flow. In this section, a comparison with experimental
data will be carried out for a contrasting flow for which it can be expected that DES
will encounter many problems. The flow phenomena occurring in the bump test case
and the challenges that these present to DES are described in detail in Sect. 6.6. Much
discussion of the results for this test case has already been offered in Sect. 8.4.2, which
will be referred to whenever the differences occurring between the SAE and CEASM
background models become relevant. Only the results from the shielded DES vari-
ants (the CEASM-DES with model-specific GIS shield, Sect. 5.5.4, and the SAE-DDES
formulation) will be discussed here: The effect of the IDDES modifications are to be
treated in Sect. 8.9.

Beginning with the pressure coefficient plots of Fig. G.2, it can indeed be seen that the
agreement with experiment is very poor: A significantly under-proportioned separa-
tion region is predicted by both models, although the CEASM-DES results are better
than those of the SAE-based DES. The CEASM-DES outperforms the SAE-DDES in
a further respect, namely the level of spanwise homogeneity. Both the experiment
and the CEASM-DES return nearly identical pressure profiles at z = 0, z = 0.085
and z = 0.17, whereas those of the SAE-DDES differ fairly strongly. In Sect. 8.4.2
this higher three-dimensionality was explained with reference to Fig. G.7 in terms of
the stronger separated streamwise vortices predicted by the SAE-DDES. Although
no transverse measurements are available from the experiment downstream of the
bump, it can be supposed that the strength of the side vortices is over-predicted by
both simulations. This would agree with a common observation that RANS models
predict excessively strong corner separations15 . Further evidence for this is offered
15
Examples include the computation of the A-airfoil including spanwise end plates and the excessive stream-
wise vortices suppressing flow separation in the symmetry plane for the idealised Ahmed car body case,
both of which are reported in the FLOMANIA project [55].

166
 8.7 Comparison of DES with experiment for the bump test case

Figure 8.30: Ratio of resolved to total kinetic energy at the symmetry plane for CEASM-DES.

by the computations of Davidson for this test case [33]: In order to reduce the grid
size, a thinner spanwise domain extent was computed and the side walls were ne-
glected. These results achieved, in some respects, the best agreement with the exper-
imental data of all the DESider computations [56].

The mean velocity profiles at the symmetry plane are compared between the ex-
perimental LDV and the computations in Fig. G.3. Again, strong discrepancies are
observed with the experimental data together with a clear trend in favour of the
CEASM-DES over the SAE-DDES. The comparison of the Reynolds shear stress in
Fig. G.4 suggests that both models predict the correct maximum level throughout
the duct, once the distortion of the profiles by the mean flow discrepancy is taken
into account. The stations further downstream exhibit larger discrepancies in the
shear stress, and the higher values for the experiment are believed to reflect the flow
reattachment much further downstream than the computations.

The high level of accuracy in the prediction of the peak Reynolds shear stresses in
the shear layer (e.g. at x = 0), together with the rapid development of the resolved
component and its dominance over the modelled component, suggest that the grey
area problem (Sect. 3.7.3) is not an important issue for this flow. Further evidence
for this is obtained from Fig. G.5, where a very rapid development of resolved tur-
bulence is seen (this plot is shown only for the SAE-DDES as there is no notable
difference between the models). This is perhaps contrary to expectation: The thin,
shielded, high-Reynolds-number oncoming boundary layer contains entirely mod-
elled turbulence and the level of recirculation of resolved turbulence was expected to
be weak. However, the strong recirculation regions seen in the corners downstream
of the bump (Fig. G.6) buffet the shear layer with resolved structures, which likely
drives the development of resolved turbulence following separation. To give an es-
timate of the extent of the grey area region, the ratio of resolved to total turbulent
kinetic energy (computed as described in Appendix E) is plotted in Fig. 8.30. Indeed,
assuming the region kres /k tot ≥ 0.8 to correspond to a sufficient level of resolved
LES content (refer to the 80% estimate of Pope [116], discussed in Sect. 3.5.2), a grey
area length of around 30% of the bump height can be estimated. Whether this can be
considered long or short would require a broader comparison with other test cases.

167
8 Demonstration and Analysis

DES therefore can be said to encounter significant problems with this test case, for
both of the the background models chosen. Of the two supposed failure mechanisms,
namely inaccuracy of the RANS solution region and a protracted grey area problem,
the former is believed to be principally responsible. The strong degree of model
dependency established in Sect. 8.4.2 adds weight to this conclusion, underlining
both the weaknesses of RANS models in such complex flows and the sensitivity of
the global flow to the upstream boundary layer prediction.

8.8 Investigation of wall-modelled LES of channel flow

A thorough investigation of the IDDES method (published by Travin et al. [170] and
Shur et al. [149]) on the basis of the canonical turbulent channel flow test case will
be presented. The method has been introduced in Section 3.7.5 and implemented
as described in Section 5.5.5, and the numerical setup and description of the chan-
nel flow can be found in Section 6.2. Background information concerning the flow
physics is presented in Section 2.4. An impression of the improvement offered by
IDDES over a DES97 simulation is given in Section 8.8.1 followed by a more detailed
examination of the IDDES blending function behaviour in Section 8.8.2. The grid re-
quirements and numerical cost scaling of WMLES are then discussed in Section 8.8.3.
Finally, the scope for easing the wall-normal grid resolution requirements by em-
ploying the hybrid-adaptive wall boundary condition of Rung et al. [135] (presented
in Appendix C) is investigated in Section 8.8.4 for the extreme case of cubic grid cells.

To reduce repetition of figures and to enhance comparability between different cases,

visualisations of the same quantity have been grouped into single combined figures.
As a result, some figures referred to in the earlier sections also contain results per-
taining to later discussions.

8.8.1 Comparison of DES97 and IDDES

To provide an initial impression of the IDDES implementation before a more detailed
analysis of the blending functions in Section 8.8.2, a comparison is made with a com-
putation employing DES97. Both simulations were carried out on a wall-refined grid
at Reτ = 4000 (i.e. case WR-4000 from Tab. 6.1) and employed the CEASM-based
DES implementations. One modification was however made for the CEASM-DES97
computation
2 compared to the standard formulation, namely the filter was defined
as Δ = 3 Δx Δy Δz instead of the standard maximum edge length expression. This
was motivated by the consideration that the cube-root definition is widely accepted
as common practice for fully-resolved LES of wall-bounded flows. This means that
the comparison is not ideal for the purposes of this discussion: The improvement
achieved by IDDES is of course best demonstrated relative to a DDES computation.
The reader is referred to the papers of the method authors [149, 170] for this. The in-
fluence of this modification can be assessed by comparison with the results published
by Nikitin et al. [101] and Piomelli et al. [114], who employed the DES-standard max-

168
 8.8 Investigation of wall-modelled LES of channel flow

35 4
Reichardt i=x, j=x, IDDES
30 IDDES i=x, j=x, IDDES
DES97 3
i=z, j=z, IDDES
25 i=x, j=y, IDDES

/ Uτ
as above, DES97
2
20

1/2
+
U

(u’iu’j )
15
1

10
0
5

0 0 1 2 3 4
-1
10 10 10 10 10 0,0 0,2 0,4 0,6 0,8 1,0
+
y y/δ

(a) Mean velocity profiles (b) Resolved Reynolds stresses

Figure 8.31: Comparison of IDDES with DES97 for fully-developed channel flow at Reτ =
4000, CEASM model.

imum expression for Δ in their studies of DES97 for channel flow.

The mean velocity profiles and resolved Reynolds stresses are compared in Fig. 8.31.
A strong displacement of the log-law region at y+ values greater than around 400 is
seen for the DES97 computation, which is connected relatively smoothly to the vis-
cous sublayer region by what appears to be an extended buffer layer. The impact of
this log law displacement can be seen in Tab. 8.8 and Fig. 8.33 where the skin friction
coefficient is compared with the value from Dean’s correlation. A significant devia-
tion of −28.59% is seen for the DES97.

The displacement of the log law region agrees well with the results obtained by
Nikitin et al. [101] at similar Reynolds numbers, however the transition region is
more compact in their simulations. Indeed, two distinct log law regions can be dis-
cerned in the Nikitin et al. profiles, one provided by the RANS model that reaches up
to y+ ≈ 100, and one from the LES resolution of the core flow. These are separated by
what is termed a “super buffer layer”. The reason for this discrepancy is clearly the
Δ formulation, as the cube-root variant pushes the LES-RANS interface much closer
to the wall for the high aspect-ratio cells there.

By comparison, the IDDES velocity profile shows much improvement with very
good agreement relative to the Reichardt correlation and Dean’s skin friction cor-
relation (Tab. 8.8) achieved. Only a minimum of log layer mismatch (LLM) remains,
visible in the form of a small kink in the profile at y+ ≈ 300. Considering that the
method was adopted directly from the published SST-variant without any additional
tuning of the model-specific Cl and Ct parameters, this is highly satisfactory. When
the resolved Reynolds stresses are examined, it can be seen that the principal dis-
crepancy occurs in the near wall region at y/δ < 0.1 (corresponding to y+ < 400).

169
8 Demonstration and Analysis

3
10
RANS
2
IDDES
10 DES97

1
10

νt / ν
0
10

-1
10

-2
10 0 1 2 3 4
10 10 10 10 10
+
y

Figure 8.32: Profiles of time-averaged eddy viscosity ratio for RANS, IDDES and DES97 for
fully-developed channel flow at Re τ = 4000, CEASM model.

Here, the streamwise Reynolds stresses continue to be resolved closer to the wall
by the DES97, whereas these are fairly strongly damped by the IDDES. This trend
can also be observed, albeit less markedly, for the resolved shear stress profiles. The
good agreement between the methods for the Reynolds stresses and log law slope in
the channel core lends weight to the interpretation that the DES97 operates correctly
in LES-mode when sufficient grid resolution is available and that the problem arises
from the “boundary condition” provided to the LES core by the near wall region.

Some insight into the model activity can be obtained from Fig. 8.32, in which profiles
of the time-averaged eddy viscosity ratio are compared between RANS, DES97 and
IDDES. The near-wall eddy viscosity of the IDDES follows the RANS values very
closely before switching fairly sharply to values approximately one order of mag-
nitude lower at y+ ≈ 300. Although the eddy viscosity levels for the DES97 are
approximately similar in the LES region, a strong depletion of the model activity
(relative to the RANS and IDDES curves) is observed right down to the viscous sub-
layer. The nature of the IDDES profile provides evidence of the “sharpening” of the
RANS-LES interface achieved by the IDDES blending functions mentioned in Sec-
tion 3.7.5.

The behaviour of this DES97 simulation contrasts strongly to the DDES simulations
for channel flow presented for comparison in the publications of the IDDES method
by Travin et al. [170] and Shur et al. [149]. The action of the DDES shield function
in fact generates higher eddy viscosity levels near the wall, which are damped less
sharply in the transition to the LES region.

170
 8.8 Investigation of wall-modelled LES of channel flow

0,01
Dean (extrapolated)
0,008
Dean (valid range)
DNS: Moser et al.
DNS: Hoyas et al.
CEASM-IDDES, WR
0,006 SAE-IDDES, WR
CEASM-IDDES, C
Cf

CEASM-DES97, WR
0,004

0,002

0 3 4 5 6 7 8
10 10 10 10 10 10
Reb

Figure 8.33: C f dependency on Re b compiled for all simulation data, compared with the DNS
data of Moser et al. [100] and Hoyas et al. [63] and with the empirical correlation
of Dean [34].

+ +
Model Case Uτ Ub Uδ C f deviation (%)
CEASM-DES97 WR-4000 0.9992 29.02 31.05 −28.59
CEASM-IDDES WR-4000 1.0036 23.86 25.97 0.66
CEASM-IDDES WR-18000 0.9964 27.48 29.62 14.34
SAE-IDDES WR-395 1.0010 18.53 20.90 −12.23
SAE-IDDES WR-18000 0.9993 27.49 29.55 14.36
CEASM-IDDES C-590 0.9959 19.10 21.34 −8.08
CEASM-IDDES C-4000 0.9988 24.48 26.59 −3.83
CEASM-IDDES C-18000 0.9980 28.28 30.71 8.74
CEASM-IDDES C-100000 0.9979 32.36 34.75 31.90

Table 8.8: Global characteristics of the channel flow simulations and percent deviation in skin
friction coefficient relative to the empirical correlation of Dean [34] (given by [( C f −
C f ,Dean ) /C f ,Dean ] × 100).

171
8 Demonstration and Analysis

The comparison of DES97 with IDDES can be rounded up by inspection of instanta-

neous flow field snapshots. The x-vorticity contours at slices normal to the stream-
wise direction (Fig. 8.34) show the higher damping of resolved structures in the very
near wall region by IDDES, which is caused by the higher eddy viscosity levels there
(Fig. 8.35). The y-vorticity contours at a near-wall tangential slice (Fig. 8.36) show the
occurrence of excessively long “super streaks” in the DES97 resolved field, which are
also stronger compared to the IDDES streaks. The occurrence of such structures and
the phenomenon of LLM in general are acknowledged by Baggett [5] and Piomelli et
al. [114] to be a common feature to many hybrid RANS-LES methods for WMLES.

8.8.2 Examination of IDDES functionality

Having gained an initial impression of the IDDES performance for WMLES of chan-
nel flow, a more detailed examination of the blending functions will be carried out.
The response of CEASM-IDDES to a variation of the Reynolds number will first be
considered, before the level of dependency on the background RANS model is as-
sessed by comparison between the SAE-IDDES and CEASM-IDDES variants.

IDDES blending functions with varying Reτ

Figure 8.37 presents a comparison of the mean velocity and resolved Reynolds stress
profiles for the WR-4000 and WR-18000 cases with CEASM-IDDES. The grids are de-
tailed in Tab. 6.1, however it is reminded that only the wall-normal grid spacing is
adjusted between the Reynolds numbers. The velocity profile at Reτ = 18000 shows
a similar behaviour to that of Reτ = 4000, although the residual LLM is perhaps
slightly weaker for the higher Reynolds number case. The transition region between
RANS and LES remains at a similar y/δ position (as can be determined from the in-
flection of the streamwise resolved Reynolds stress component), which corresponds
to a higher y+ ≈ 2000 for the Reτ = 18000 case. The skin friction coefficient shows a
much increased deviation relative to Dean’s correlation (Tab. 8.8 and Fig. 8.33) at the
higher Reynolds number. However, the corresponding bulk Reynolds number lies
outside the quoted validity range of Dean’s correlation, so the extent to which this
represents a genuine inaccuracy is unclear.

The instantaneous vorticity snapshots of Fig. 8.34 show the resolution of finer struc-
tures further into the channel core for the higher Reynolds number. The instanta-
neous eddy viscosity ratio plots of Fig. 8.35 indicate a similar distribution with el-
evated levels in the near-wall RANS region and at the coarser grid in the channel
core. The general level of νt /ν is furthermore higher for Reτ = 18000. The instanta-
neous vorticity parallel to the wall indicates similar streaky structures in both cases,
which are perhaps more densely distributed at Reτ = 18000. A direct comparison
is potentially misleading however, as the slices are at different y+ values in each case.

The mean eddy viscosity ratio variation with Reynolds number, plotted in Fig. 8.38,
demonstrates the scaling of the interface region to higher y+ with increasing Reτ . In

172
 8.8 Investigation of wall-modelled LES of channel flow

(a) CEASM-DES97, WR-4000 (b) CEASM-IDDES, C-590

(c) CEASM-IDDES, WR-4000 (d) CEASM-IDDES, C-4000

(e) CEASM-IDDES, WR-18000 (f) CEASM-IDDES, C-18000

(h) CEASM-IDDES, C-100000

Figure 8.34: Comparison of instantaneous x-vorticity at a constant x-slice between various sim-
ulations for fully-developed channel flow. Solid and dashed lines denote positive
and negative values, respectively.

173
8 Demonstration and Analysis

(a) CEASM-DES97, WR-4000 (b) CEASM-IDDES, C-590

(c) CEASM-IDDES, WR-4000 (d) CEASM-IDDES, C-4000

(e) CEASM-IDDES, WR-18000 (f) CEASM-IDDES, C-18000

(h) CEASM-IDDES, C-100000

Figure 8.35: Comparison of instantaneous eddy viscosity ratio at a constant x-slice between
various simulations for fully-developed channel flow.

174
 8.8 Investigation of wall-modelled LES of channel flow

(a) CEASM-DES97, WR-4000 (b) CEASM-IDDES, C-590

(c) CEASM-IDDES, WR-4000 (d) CEASM-IDDES, C-4000

(e) CEASM-IDDES, WR-18000 (f) CEASM-IDDES, C-18000

(h) CEASM-IDDES, C-100000

Figure 8.36: Comparison of instantaneous y-vorticity at a slice y/δ = 0.05 between various
simulations for fully-developed channel flow. Solid and dashed lines denote posi-
tive and negative values, respectively.

both cases, the IDDES eddy viscosity follows that of the pure RANS simulation fairly
closely up to the interface region, however some deficit is apparent just inside of the
interface, on the RANS side. It is supposed that this could contribute to the residual
LLM observed in the velocity profiles, a theme to be returned to in the following.

Some important IDDES blending functions are plotted for the CEASM-IDDES com-
putations at Reτ = 4000 and Reτ = 18000 in Fig. 8.39. It must be noted that these

175
8 Demonstration and Analysis

40 4

35
Reichardt i=x, j=x, IDDES, WR-4000
CEASM-IDDES, WR-4000 i=x, j=x, IDDES, WR-4000
CEASM-IDDES, WR-18000 3
30 i=z, j=z, IDDES, WR-4000
i=x, j=y, IDDES, WR-4000

/ Uτ
25 as above, IDDES, WR-18000
2

1/2
+

20
U

(u’iu’j )
15 1

10
0
5

0 0 1 2 3 4
-1
10 10 10 10 10 0.0 0.2 0.4 0.6 0.8 1.0
+
y y/δ

(a) Mean velocity profiles (b) Resolved Reynolds stresses

Figure 8.37: Variation of Reynolds number for CEASM-IDDES for fully-developed channel
flow at Reτ = 4000 and Reτ = 18000.

10
4 CEASM-RANS, WR-4000
CEASM-RANS, WR-18000
3 CEASM-IDDES, WR-4000
10 CEASM-IDDES, WR-18000
2
10
νt / ν

1
10

0
10

-1
10

-2
10 0 1 2 3 4
10 10 10 10 10
+
y

Figure 8.38: Profiles of time-averaged eddy viscosity ratio for RANS and IDDES of fully-
developed channel flow at Reτ = 4000 and Re τ = 18000, CEASM model.

plots are extracted from a single wall-normal grid line and time step, so some fluc-
tuation of the solution-dependent quantities must be borne in mind. The near wall
reduction of the grid filter definition Δ, Eq. (3.23), is portrayed by its ratio to the
maximum cell length hmax . This purely grid-dependent quantity is essentially iden-
tical for both simulations, as both exhibit similar wall-normal stretching ratios in the
region of k ≤ 1 + Cw and the same tangential grid resolution. The filter width is
therefore reduced compared to the standard DES Δ (which is equal to hmax ) below
around y/δ = 0.7, and reaches a minimum near-wall value of Cw hmax .

The quantity 1 − f dt shows the activity of the modelled log layer sensor, which as-

176
 8.8 Investigation of wall-modelled LES of channel flow

1 1

0.8 0.8

0.6 Δ / hmax 0.6 Δ / hmax

1 - fdt 1 - fdt
0.4 ~ 0.4 ~
fd fd
frestore frestore
0.2 LIDDES / LRANS 0.2 LIDDES / LRANS

0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
y/δ y/δ

(a) WR-4000 (b) WR-18000

Figure 8.39: Instantaneous profiles of selected IDDES blending functions for CEASM-IDDES at
Reτ = 4000 and Reτ = 18000.

sumes a value of unity within the RANS log layer region. This, together with the grid
dependent f step function (portrayed in Fig. 3.9), makes up the IDDES RANS-to-LES
switching control f˜d , which selects the maximum of each quantity, Eq. (3.35). f˜d is
seen to follow the 1 − f dt branch at Reτ = 4000, which occurs at greater y/δ than the
f step transition (not shown). In the Reτ = 18000 snapshot, f˜d follows 1 − f dt initially,
switching to f step near to the end of the transition region.

The final component contributing to the IDDES length scale is the boosting function
f restore, Eq. (3.26). This is inactive in the Reτ = 4000 snapshot and weakly active at
Reτ = 18000, giving rise to an amplification of LIDDES just on the RANS side of the
transition region. In comparison to the plots of these quantities published by Shur et
al. [149], f restore appears to be insufficiently active in the CEASM-IDDES. This func-
tion is designed to counteract the damping of eddy viscosity on the RANS side of
the interface by diffusive effects of the lower values on the LES side. As a result, it is
supposed that this under-activity of f restore is the cause of the reduced eddy viscosity
relative to the RANS value observed in Fig. 8.38, which was in turn identified as a
possible reason for the residual LLM observed. A calibration of the model-specific
Cl and Ct parameters is therefore seen as a promising route to improving the perfor-
mance of the IDDES implementation.

The effect of the constituent blending functions on the IDDES length scale can hence
be traced from the plots of Fig. 8.39, where the ratio LIDDES /LRANS is shown16 . The
switching behaviour is determined by f˜d and f restore, whereas the profile in the LES-
mode region is determined by the Δ function. The reduced near-wall value of Δ gives
16
The quantity LRANS could possibly be misinterpreted – this is value of the RANS model length scale arising
from the quantities k and ε present locally in the IDDES simulation. This is not the value of LRANS that
would be obtained from a RANS simulation.

177
8 Demonstration and Analysis

40 4

35
Reichardt i=x, j=x, SAE-IDDES
SAE-RANS, WR-18000 i=x, j=x, SAE-IDDES
SAE-IDDES, WR-18000 3
30 i=z, j=z, SAE-IDDES
CEASM-RANS, WR-18000 i=x, j=y, SAE-IDDES
CEASM-IDDES, WR-18000

/ Uτ
25 as above, CEASM-IDDES
2

1/2
+

20
U

(u’iu’j )
15 1

10
0
5

0 0 1 2 3 4
-1
10 10 10 10 10 0.0 0.2 0.4 0.6 0.8 1.0
+
y y/δ

(a) Mean velocity profiles (b) Resolved Reynolds stresses

Figure 8.40: Comparison of SAE-IDDES and CEASM-IDDES for fully-developed channel flow
at Reτ = 18000.

rise to the trough in the LIDDES /LRANS distribution just on the LES-side of the inter-
face. The peak-and-trough nature of the IDDES length scale around the interface is
characteristic of the sharpened RANS-LES blending of IDDES compared to previous
versions. The general level of LIDDES /LRANS is seen to rise as the Reynolds number
increases. The LIDDES /LRANS profiles are qualitatively very similar to those of νt /ν
(Figs. 8.35 and 8.38), indicating the strong linkage of these quantities.

Comparison of the SAE and CEASM IDDES variants

The mean velocity and resolved Reynolds stress profiles from SAE-IDDES and CE-
ASM-IDDES simulations at Reτ = 18000 are compared in Fig. 8.40. The velocity
profiles are highly similar above y+ ≈ 600 indicating that the RANS-LES blending
and LES mode functionality of the implementations is equivalent. A similar level of
residual LLM is also seen for the SAE-IDDES and the skin friction coefficient values
give a very similar discrepancy relative to Dean’s correlation (Tab. 8.8 and Fig. 8.33).
The inclusion of the RANS velocity profiles for each model in Fig. 8.40 elucidates
the deviation around the early log law region. In line with expectation therefore,
the RANS-mode region of IDDES is dominated by the particular RANS formulation
employed. The C f similarity however indicates that the near-wall differences exert a
minimal influence on the global characteristics.

The resolved Reynolds stress profiles also reveal very similar behaviour in the LES-
mode region. Minor differences are seen in the near wall region y/δ < 0.1, where the
damping behaviour of the resolved turbulence appears to differ slightly between the
models. The eddy viscosity distributions in Fig. 8.41 show a high degree of compa-
rability, although the blending occurs at slightly higher y+ for the SAE-IDDES and
slightly higher values are seen in the LES-mode region. A possible explanation for

178
 8.8 Investigation of wall-modelled LES of channel flow

4
10
CEASM-IDDES, WR-18000
3 SAE-IDDES, WR-18000 1.2
10
1
2
10
0.8 Δ / hmax
νt / ν

10
1 1 - fdt
0.6 ~
fd
10
0 frestore
0.4
Ψ
-1
LIDDES / LRANS
10 0.2

-2 0
10 0 1 2 3 4 0.0 0.2 0.4 0.6 0.8 1.0
10 10 10 10 10
+ y/δ
y

(a) νt /ν, SAE-IDDES and CEASM-IDDES (b) SAE-IDDES blending functions

Figure 8.41: IDDES functionality for the SAE implementation at Reτ = 18000. Profiles of time-
averaged eddy viscosity ratio in comparison to CEASM-IDDES and selected ID-
DES blending functions.

the higher LES-mode eddy viscosity is offered by the higher equivalent Smagorinsky
constant for SAE-based DES compared to CEASM-based DES, derived in Section 8.2.
The fact that the velocity profiles are nonetheless so similar echoes the observation
by Nikitin et al. [101] of a low sensitivity to CDES 17 .

The SAE-IDDES blending function behaviour is also depicted in Fig. 8.41, from which
a number of differences to the CEASM-IDDES (Fig. 8.39) are noteworthy. Firstly,
a Ψ function is implemented in this model, as described in Section 5.5.2 and Ap-
pendix D. As could be expected at such a high Reynolds number (with νt /ν > 25
in the LES mode region), the snapshot profile indicates that this is hardly active. A
second difference concerns the f restore behaviour, which, if this snapshot can be con-
sidered representative, exhibits somewhat stronger activity than for CEASM-IDDES.
Indeed, a different set of Cl and Ct parameters is used, those for SAE-IDDES adopt-
ing the published values for SA-IDDES [170, 149]. Finally, strongly reduced values
of LIDDES /LRANS are seen compared to CEASM-IDDES. This is however predom-
inantly a cosmetic difference arising from the fact that LRANS = dw for the SAE-
model18 .

The issue of the residual LLM apparent in the velocity profiles will be returned to.
One possible route to improving this behaviour has already been identified, namely a
tuning of the model-specific Cl and Ct parameters, which has as yet not been carried
out. Another hypothetical explanation for the residual LLM exists however, which

17
This furthermore fuels the discussion in Sect. 8.2 concerning the expected value of a dynamic DES formu-
lation.
18
This is loosely related to the discussion of the DES-interface behaviour in Section 8.1.2. The RANS model
length scales are summarised in Tab. 5.2.

179
8 Demonstration and Analysis

offers a further line of investigation and will be briefly described. The discussion of
the LLM phenomenon in hybrid RANS-LES methods by Baggett [5] (mentioned in
Sect. 3.7.3) noted that the super-streak structures associated with LLM scale with the
grid resolution, i.e. that these correspond to the smallest structures resolvable by the
grid and numerical scheme applied. Part of the task allocated to the IDDES blending
functions is therefore to assess the grid resolution and to displace the RANS/LES
interface further from the wall until this is deemed sufficient. The development of
the IDDES functions was strongly driven by empirical calibration, which was car-
ried out using the flow solver of NTS and its fourth order central difference convec-
tion scheme. Such a scheme can of course accurately represent smaller vortices on a
given grid than the second order CDS applied here. This difference in the numerical
order of accuracy between the NTS solver and the ELAN solver therefore provides
a likely explanation for the residual LLM observed in the ELAN implementation.
Two avenues could therefore be pursued to resolve the remaining LLM problem:
The implementation of 4th order numerics in ELAN and the modification of the rel-
evant IDDES blending functions to push the RANS/LES interface further from the
wall for 2nd order CDS. Should the latter route prove successful, this would reveal
an important numerics dependency of the method. In this case, clear guidelines for
code-specific tuning would be required.

8.8.3 Numerical cost of WMLES and grid considerations

Some comments can be made at this juncture concerning the numerical cost of WM-
LES and the reduction relative to fully-resolved LES achieved by the tangential coars-
ening it allows. To illustrate the saving achieved, the number of grid points and time
steps that would be required by pure LES will be estimated for the highest Reynolds
number IDDES computation presented above, i.e. Reτ = 18000.

The estimate must assume a maximum permissible tangential grid spacing for pure
LES, and as commented in Section 3.5 the literature exhibits some variability in the
guidelines. The values Δxmax+ = 40 and Δz + = 20 will be assumed here, together
max
with the same domain extent of L x = 6.4δ and L z = 3.2δ from the WMLES. This
gives rise to 2880 grid points in each tangential direction, and assuming the same
wall-normal point distribution as the WMLES over a billion (1 × 109 ) cells in total.
Concerning the time step, this scales with Δx in order to maintain CFL ≤ 1, resulting
in a factor of 45 increase in the number of time steps. Combining the spatial and
temporal cost factors and assuming a constant numerical expense per grid cell and
time step, the cost saving of WMLES compared to LES is around a factor 90000 at
this Reynolds number.

This illustrative cost factor calculation must be accompanied with the acknowledge-
ment that the potential cost reduction of WMLES increases with Reynolds number.
Furthermore, a variation in the tangential refinement has not been carried out, and
it is possible that a further cost reduction could be achievable. In tests maintaining

180
 8.8 Investigation of wall-modelled LES of channel flow

Δz = 0.05δ, Shur et al. [149] demonstrate that IDDES maintains a reasonable solution
quality when Δx/Δz is increased up to 4.

Although the tangential resolution of WMLES is unlimited in terms of wall units

[149], this implies that there is no upper limit on the Reynolds number that can be
computed and should not be interpreted to mean that an arbitrarily coarse grid can
be used. Whereas fully-resolved LES is dictated by criteria for resolving the near-wall
structures, WMLES resolution criteria are based on resolution of the outer boundary
layer turbulence. This indicates a fixed number of points per boundary layer thick-
ness, and indeed Piomelli et al. [114] cite the criterion of 15 to 20 points per δ derived
theoretically by Chapman [26] and observed numerically by Nikitin et al. [101].

Finally, it should be remarked upon that the estimate of the numerical expense of a
fully-resolved LES assumes structured grids. As a consequence of keeping the wall-
normal distribution from WMLES but refining tangentially, highly-stretched colum-
nar cells would result in the channel core. Obtaining optimal cubic cells with a struc-
tured solver would require a limitation of Δy = max(Δx ; Δz) and a significant
further increase in the grid density. Alternatively, the adoption of an unstructured
grid methodology would potentially allow a coarsening of the tangential grid fur-
ther from the wall and a corresponding reduction of the numerical expense.

The fixation of the grid resolution relative to δ for WMLES results in an identical tan-
gential grid for any Reynolds number in the case of fully-developed internal flows.
For external flows however, the boundary layer thickness decreases as Re increases.
For this reason, Piomelli et al. [114] quote a scaling of the necessary number of grid
points proportional to Re0.4 .

8.8.4 Combination with the hybrid-adaptive boundary condition

The WMLES concept embodied in the IDDES method has been demonstrated to re-
lieve the tangential resolution requirements of wall-bounded LES. The wall-normal
grid spacing must however still be adapted to the Reynolds number to provide
y+ ≈ 1 in the first cell, as required by the background RANS model treated with
standard low-Re boundary conditions. The hybrid-adaptive boundary condition of
Rung et al. [135] was formulated to relieve this problem for RANS simulations, pro-
viding a seamless blend of high-Re and low-Re boundary conditions. This boundary
condition therefore allows any value of y+ at the wall, and its formulation is de-
scribed in Appendix C.

The attractiveness of combining the hybrid-adaptive boundary condition with ID-

DES is not so much oriented towards a further reduction of computational expense
for WMLES, as might at first be expected. Indeed, Nikitin et al. [101] point out that
“the economic incentive to use ‘wall functions’ to relax the Δy+ ≈ 1 requirement is weak”, as
“a factor of 10 increase in Reτ adds only 17 grid layers for each wall: the evolution is logarith-
mic”. Nonetheless, such functionality is attractive from the perspective of complex

181
8 Demonstration and Analysis

25 4
DNS, Moser et al. i=x, j=x, IDDES
CEASM-IDDES, C-590 i=x, j=x, IDDES
20 3
i=z, j=z, IDDES
i=x, j=y, IDDES

/ Uτ
as above, DNS
15 2

1/2
+
U

(u’iu’j )
10 1

5 0

0 0 1 2 3
-1
10 10 10 10 0.0 0.2 0.4 0.6 0.8 1.0
+
y y/δ

(a) Mean velocity profiles (b) Resolved Reynolds stresses

Figure 8.42: Comparison of CEASM-IDDES employing the hybrid-adaptive wall function with
the DNS of Moser et al. [100] for fully-developed channel flow at Reτ = 590, cubic
grid.

applications: in particular, the need to know the τw distribution in advance in order

to safeguard the correct specification of y+ often leads to iterative grid generation.
The practical advantages of the hybrid-adaptive boundary condition for RANS are
described in Section 3.4.4 and Appendix C and demonstrated by Schmidt et al. [145]
and Mockett et al. [95]. It is expected that many of these benefits would carry over to
WMLES.

In principle, the law of the wall is only valid in a time-averaged sense, so it is unclear
whether its application in a simulation of resolved, unsteady turbulence is valid. This
has also been remarked upon by Deardorff in his pioneering LES study of channel
flow [35], which employed essentially an equivalent of a high-Re boundary condi-
tion to parameterise the near wall profile. In a similar vein to Deardorff’s investi-
gation, an equidistant grid spacing is applied as a challenging test for the hybrid-
adaptive boundary condition with IDDES – the cubic grid is not however suggested
for practical application19 . A secondary point of interest is the behaviour of the grid-
dependent blending functions in the IDDES formulation, in particular the Δ defini-
tion that anticipates a stretched wall-normal grid: Does the IDDES method respond
in a robust manner when the equivalent grid essentially deactivates these functions?

As described in Section 6.2, the cubic grid resolution is Δx = Δy = Δz = 0.05δ (i.e. 20

points per boundary layer thickness) and this has been computed with the CEASM-
IDDES formulation at various Reynolds numbers. An impression of the performance
will first be obtained for the Reτ = 590 computation in comparison with the DNS of
Moser et al. [100], before the behaviour with increasing Reynolds number is exam-
19
(although it would be highly relevant if used in conjunction with the immersed boundary method,
whereby complex geometries are “immersed” in a Cartesian grid instead of using body-fitted meshing)

182
 8.8 Investigation of wall-modelled LES of channel flow

40 4
i=x, j=x, IDDES, C-590
Reichardt i=x, j=x, IDDES, C-590
35 CEASM-IDDES, C-590 3 i=z, j=z, IDDES, C-590
CEASM-IDDES, C-4000 i=x, j=y, IDDES, C-590
30 CEASM-IDDES, C-18000
as above, IDDES, C-4000
CEASM-IDDES, C-100000 2 as above, IDDES, C-18000

/ Uτ
25 as above, IDDES, C-100000

1/2
+

20 1
U

(u’iu’j )
15
0
10
-1
5

0 0 1 2 3 4 5
-2
10 10 10 10 10 10 0,0 0,2 0,4 0,6 0,8 1,0
+
y y/δ

(a) Mean velocity profiles (b) Resolved Reynolds stresses

Figure 8.43: Reynolds number sweep of cubic grid calculations with CEASM-IDDES employ-
ing the hybrid-adaptive wall function (n.b. the Reynolds stress profiles for Reτ =
18000 and Reτ = 100000 are almost identical).

ined.

The profiles of mean velocity and resolved Reynolds stress components at Reτ = 590
are plotted in Fig. 8.42. The mean velocity at the first cell centre (y+ ≈ 15) agrees per-
fectly with the DNS data, illustrating the correct functionality of the hybrid-adaptive
boundary condition even within the buffer layer, where standard high-Reynolds for-
mulations would be invalid. The velocity profile then continues to follow the DNS
log law profile fairly closely, although with a slightly excessive intercept and varying
gradient. Considering the extremely coarse wall-normal resolution of the near wall
region however, the agreement is highly satisfactory. The skin friction coefficient de-
viates by −8.08% from Dean’s correlation (Tab. 8.8), however the Dean correlation
itself deviates from the DNS data by 4.21% in the same sense (as indicated by the
comparison of collated DNS data with Dean’s correlation in Fig. 2.8).

The resolved Reynolds stress profiles show very good agreement with the DNS in the
channel core, however this deteriorates nearer to the wall, below around y/δ ≈ 0.4.
This is interpreted as a consequence of under-resolution in this region (the over-
prediction of the streamwise and under-prediction of the wall-normal and spanwise
components is a typical indication of this [17]). An alarming oscillation of the shear
stress profile is also observed, for which two hypothetical explanations are offered.
One possibility is the instability of the 2nd order CDS convection scheme with exces-
sively coarse grids (Sect. 4.2.2, [106]). The second possible cause is related to the Rhie
& Chow interpolation used to damp the pressure-velocity decoupling of the SIM-
PLE method for collocated numerics (Section 4.3). As shown by Choi [28], the Rhie
& Chow damping can become inactive for the combination of a coarse grid and a
fine time step. Apart from the resolved shear stress however, this instability does not

183
8 Demonstration and Analysis

10
4 CEASM-IDDES, C-590
CEASM-IDDES, C-4000
3 CEASM-IDDES, C-18000
10 CEASM-IDDES, C-100000
2
10

νt / ν
1
10

0
10

-1
10

-2
10 0 1 2 3 4 5
10 10 10 10 10 10
+
y

Figure 8.44: Profiles of time-averaged eddy viscosity ratio for CEASM-IDDES of fully-
developed channel flow for a range of Reτ . Cubic grid calculations employing
the hybrid-adaptive wall function.

appear to impact the statistical quantities, and furthermore a systematic oscillation

cannot be detected in the instantaneous flow snapshots of Figs. 8.34 and 8.36.

Turning to the cubic grid computations at higher Reynolds numbers, the mean ve-
locity and resolved Reynolds stresses are plotted in Fig. 8.43. The satisfactory agree-
ment with the Reichardt correlation is also exhibited at higher Reynolds numbers,
even improving at the very highest values. The poorest agreement is in fact given by
the intermediate Reτ = 4000 calculation, with a slight over-prediction of the log-law
intercept. The skin friction coefficient values (Tab. 8.8, Fig. 8.33) deviate increasingly
from Dean’s correlation as Reτ increases above 18000. However, the corresponding
bulk Reynolds numbers lie outside the range of validity specified for the empirical
correlation (Sect. 2.4.2), so the comparison is perhaps invalid. The resolved Reynolds
stresses exhibit a similar behaviour at the higher Reynolds numbers, although the
symptoms of under-resolution (excessive streamwise stresses, insufficient spanwise
and normal stresses) appear to diminish as Reτ is increased, reaching however an
apparent convergence at Reτ = 18000. The oscillation of the shear stress profile is
also apparent for all Reynolds numbers, but is significantly stronger at the interme-
diate Reτ = 4000.

Attention will now be turned to the instantaneous snapshots of Figs. 8.34, 8.35 and
8.36. With increasing Reτ , the streamwise vorticity reveals an increase in the number
of fine vortical structures and in their strength. The wall-normal snapshots similarly
depict a decrease in the size of the streaks with increasing Reτ . The near wall streaks
are also strongly represented, despite the fact that this slice is located at the first near
wall cell vertex. For both the wall-normal and streamwise structures, a saturation
of the grid resolution appears to occur at around Reτ = 18000 (i.e., no appreciable

184
 8.8 Investigation of wall-modelled LES of channel flow

1 1
Δ / hmax
0.8 1 - fdt 0.8
~
fd
frestore
0.6 0.6
LIDDES / LRANS
0.4 0.4

0.2 0.2

0 0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
y/δ y/δ

(a) C-590 (b) C-4000

1 1

0,8 0,8

0,6 0,6

0,4 0,4

0,2 0,2

0 0
0,0 0,2 0,4 0,6 0,8 1,0 0,0 0,2 0,4 0,6 0,8 1,0
y/δ y/δ

(c) C-18000 (d) C-100000

Figure 8.45: Instantaneous profiles of selected IDDES blending functions for cubic grid calcula-
tions with CEASM-IDDES employing the hybrid-adaptive wall function at various
Reynolds numbers.

increase in the resolution of fine structures is observed at Reτ = 100000). This is

most probably the reason for the similarity of the resolved Reynolds stress profiles at
Reτ = 18000 and 100000 commented on previously.

The instantaneous eddy viscosity ratio plots of Fig. 8.35 show a similar near-wall
peak as for the wall-refined grids at Reτ ≥ 4000, due to the RANS activity there.
The coarse grid resolution however appears to shift this peak slightly further from
the wall. The eddy viscosity distributions in the channel core exhibit an inverse be-
haviour to the wall-refined cases, due to the uniform grid spacing. The general level
of the eddy viscosity ratio increases considerably with Reτ , reflecting a higher pro-
portion of subgrid turbulent activity modelled.

This latter observation is more objectively portrayed by the mean eddy viscosity ratio
profiles, plotted in Fig. 8.44. An identical distribution is seen in this dual-logarithmic

185
8 Demonstration and Analysis

representation, only shifted to higher y+ and νt /ν values at higher Reτ . A com-

parison with the corresponding plot from the wall-refined grids (Fig. 8.38) reveals
a stark contrast in the character of the profiles. The RANS-LES transition region is
very poorly resolved by the cubic grid and the profiles show an opposite gradient in
the LES-mode region.

Some explanation of this behaviour is given by the instantaneous snapshots of the

IDDES blending functions, plotted in Fig. 8.45. When comparing with their coun-
terpart for the wall-refined grids (Fig. 8.39), the most notable difference can be seen
in the quantity Δ/hmax . This remains equal to 1 across the channel due to the ab-
sence of grid stretching, as a consequence of which the trough in LIDDES /LRANS is
absent, and a relatively flat profile in the LES region is retained. The general value
of LIDDES /LRANS increases with Reτ , reflecting the increased subgrid-scale activity
commented upon above. An interesting question is therefore the behaviour at still
higher Reynolds numbers: would LIDDES /LRANS exceed unity, giving rise to RANS
across the channel?

Finally, an important observation should be reported concerning the computational

expense of these simulations. Although identical in terms of grid resolution and time
step size, a computational cost increase nonetheless occurs with increasing Reynolds
number. Although employing the same initialisation method reported in Sect. 6.2
(with a separate RANS profile computed for each Reynolds number), the higher
Reynolds number calculations required significantly more time steps before u τ had
settled to oscillation around unity and a stable value of U b had been obtained. The
reason for this is not entirely clear, however this could be because the same fluctu-
ation strength was applied at all Reynolds numbers. Perhaps if the superimposed
fluctuations were amplified at higher Reynolds numbers, this effect could be dimin-
ished.

8.9 IDDES application to complex flow

Having examined the performance and functionality of the IDDES method for the
canonical and geometrically simple channel flow case, a more complex flow will be
studied. For this purpose, the separating–reattaching bump test case will be em-
ployed, which is described in Sect. 6.6. Because of the steady-state inlet, RANS at-
tached turbulent boundary layers, LES separated flow region and the impingement
of the resolved turbulence onto the wall at reattachment, this case represents an ideal
and challenging test of the IDDES method operating in a “mixed mode” simula-
tion. This functionality was similarly demonstrated by the method’s authors for a
backward-facing step flow [170, 149]. Results from this test case have already been
presented in the assessment of the model-dependency of DES in Sect. 8.4.2 and in
the comparison of shielded DES with experiment in Sect. 8.7, and reference will be
made to these sections where appropriate. As no good agreement with the experi-

186
 8.9 IDDES application to complex flow

mental data can be expected (due to the limitations of the upstream RANS portion,
as established in Sect. 8.7), the aim is primarily to test the performance of the sensors
and blending functions of IDDES rather than to assess the predictive accuracy: Does
the model automatically and appropriately switch between its DDES and WMLES
branches?

A collection of instantaneous visualisations of flow quantities and various blending

functions are given in Fig. 8.47 for a lateral plane at z = 0 and a longitudinal plane at
x = 0 (located at the bump foot). The vorticity magnitude and eddy viscosity ratio
plots serve to illustrate the flow and simulation characteristics at the slices chosen.
The oncoming boundary layers and the earlier upper wall boundary layer exhibit a
RANS nature, with smooth profiles and high values of eddy viscosity. Downstream
of the bump however, LES and WMLES character is evident through finely-resolved
turbulent structures and low values of eddy viscosity, except very near the wall.

The ratio Δ/hmax is again employed to portray the filter width damping according
to Eq. (3.23). Up to x ≈ 0.3, the damping is not active and the filter width is equal
to hmax . This is because the tangential grid spacing is fine (see the grid depiction in
Fig. 6.10), so hmax < max (Cw dw ; Cw hmax ; hwn ). This reflects the almost cubic grid
cells in this region and mimics the behaviour for the cubic channel grids in Sect. 8.8.4.
Examination of the longitudinal cross-section however reveals Δ damping in the cor-
ners caused by the compression of the grid cells in two directions simultaneously
(whilst the streamwise spacing stays constant). Returning to the lateral cross-section,
only when the streamwise grid spacing begins to coarsen after x ≈ 0.3 does a damp-
ing of the filter width relative to hmax start to occur. The strength and extent of this
increases downstream, accompanying the increase in streamwise grid spacing and
anisotropy.

The log-layer sensor visualised by (1 − f dt ) together with the hybrid blending func-
tion f˜d can be used to determine the active branch of the IDDES. In the inflow and
early separated regions, f˜d = (1 − f dt ), meaning that the DDES functionality is in
effect. Further downstream, at around x > 0.8, f˜d = f step is seen to dominate, and
the WMLES branch is active. The IDDES model therefore appropriately selects the
correct DDES and WMLES functionality locally, allowing such mixed-mode compu-
tations. Figure 8.46 compares the eddy viscosity ratios between the DDES and IDDES
computations. This confirms that an identical behaviour is apparent up to x ≈ 0.45,
whereas strong differences in the level of model activity occurs downstream of this.
The activation of the WMLES mode leads to much reduced eddy viscosity levels
here.

The influence of the IDDES on the results in comparison to DDES can be evaluated
from the cross-plots in Appendix G. In line with expectation, the effect on the flow in
the early separated region is minor as both models act in DDES mode there. As the
flow has been demonstrated to depend strongly on the upstream separation predic-

187
8 Demonstration and Analysis

(a) SAE-DDES

(b) SAE-IDDES

Figure 8.46: Instantaneous contours of profiles of eddy viscosity ratio compared between
DDES and IDDES at a lateral cross-section z = 0 throughout the entire compu-
tational domain.

tion, this gives rise to a negligible difference between SAE-IDDES and SAE-DDES
for global flow characteristics such as time-averaged streamlines (Fig. G.1), pres-
sure coefficient profiles (Fig. G.2(a)), separation and reattachment locations (Tab. G.1)
and time-averaged flow topology (Fig. G.6). The discrepancies between IDDES and
DDES grow further downstream, in the region where the WMLES branch of the ID-
DES is active. In particular, the streamwise velocity profile exhibits greater values
near to the wall (Fig. G.3) and a reduced level of modelled and resolved shear stress
is apparent in the downstream boundary layers (Fig. G.4).

These results fall short of demonstrating any improvement in predictive accuracy for
the IDDES compared to the DDES, due to the problem of comparing with experimen-
tal data when the flow topology disagrees so strongly (Sect. 8.7). For this purpose,
the backward facing step flow (with a geometry-induced separation point) reported
by Travin et al. [170] and Shur et al. [149] was much more successful. However, these
results demonstrate important and encouraging features of the mixed-mode oper-
ation of IDDES as well as its general applicability to complex flows. At the same
time however, some concerning new issues arise regarding the transition between
different turbulence treatment paradigms. Two examples occur in this flow: Firstly,
a transition between WMLES and RANS occurs in the lower wall boundary layer
due to gradual streamwise coarsening. Secondly, the inverse transition of a RANS
to a WMLES boundary layer is brought about by the transmission of flow field fluc-

188
 8.9 IDDES application to complex flow

tuations from the opposite WMLES boundary layer20 . It is clear that each of the
limit cases (i.e. RANS and WMLES) are treated appropriately by IDDES. It is the
protracted transition between these states where the uncertainty arises. This issue
can hence be interpreted as another form of grey area (Sect. 3.7.3). These issues are
indeed unimportant for the flow in question and for many other flows, however ap-
plications are conceivable for which such issues may represent serious problems. In
any case, some targeted investigations should be carried out. Canonical flow cases
such as developing channel flows and fully-developed channel flow with gradual
streamwise coarsening come to mind.

20
These issues were brought to light by M. Leschziner in private communication.

189
8 Demonstration and Analysis

(a) | Ω| (b) νt /ν

(c) Δ/h max (d) 1 − f dt

(e) f˜d (f) f restore

Figure 8.47: Instantaneous contours of selected IDDES blending functions for the bump test
case calculated with SAE-IDDES. The slices z = 0 (symmetry plane) and x = 0
(bump foot) are portrayed.

190
9 Conclusion

A diverse and comprehensive set of investigations into the performance of DES for
a wide range of flow cases has been documented. It is hoped that these further the
understanding of the approach, both in terms of its capabilities as well as its limi-
tations. Much effort has been invested in presenting a complete monograph, with
all physical and numerical background information upon which the arguments are
based summarised in the literature review chapters of Part I. A concise description of
the numerical solver and implementation details has been given in Part II, together
with collated descriptions of the numerous test cases employed. A thorough demon-
stration, validation and calibration of key features of the implementation has been
presented in Chap. 7, which provides a foundation for the investigations reported in
Chap. 8. These partly provide an initial overall impression of typical DES solutions
and partly target a deeper insight into more advanced topics and several avenues for
further exploration are outlined. The findings will be summarised and some closing
comments will be given in the following two sections.

9.1 Summary of findings

A bullet point list of the principal findings will be given, which are formulated in a
deliberately concise manner. References to the relevant sections direct the reader to
more detailed descriptions.

• An original approach to taxonomy of hybrid RANS-LES methods has been pro-

posed based on four identified ambitions and ignoring formulation issues. This
provides clarity and overview particularly to the industrial end user concern-
ing an expansive field of research. It furthermore elaborates the definition of
DES and elucidates its differentiation from other approaches. This could form
a useful basis for the formulation of industrial recommendations as well as
providing a link between different hybrid method formulations and a range of
target applications.
⇒ Section 3.6

• Tests employing inviscid isotropic turbulence demonstrate that the CDS con-
vection scheme conserves kinetic energy, at least on the equidistant Cartesian
grids considered. The ELAN numerical setup is hence deemed suitable for
LES. The UDS and TVD schemes cannot be applied, even when combined with
reduced CDES values.
⇒ Section 7.1

191
9 Conclusion

• The Ψ functions derived for the SAE and SALSA models function correctly. No
such correction is required for the WCX model and none could be derived for
the LLR and CEASM models due to their inherent complexity. DIT tests expose
this as a problem for LLR-DES, whereas CEASM-DES functions satisfactorily.
Use of LLR-based DES is therefore advised against.
⇒ Section 7.4 & Appendix D

• An analytical framework for deriving CDES as a function of the Smagorinsky

parameter is demonstrated to hold for SA-family models, but over-predicts
CDES for the two-equation models tested. This emphasises the importance of
empirical calibration.
⇒ Section 8.2

• Arguments for and against the usefulness of a dynamic DES enhancement are
presented, together with a suggested formulation approach based on David-
son’s dynamic one-equation SGS model. Comparison of the Smagorinsky and
Germano/Lilly dynamic SGS models for DIT indicate that dynamic DES could
reduce the dependency exhibited by CDES on the grid resolution.
⇒ Sections 3.5.1 & 8.2

• DES formulations involving alternative substitutions of the length scale pro-

duce equivalent results for a natural DES application once the dissipativity is
balanced via calibration of CDES .
⇒ Section 8.3

• Results for jet flow are cited indicating that a dual substitution of the DES
length scale reduces the grey area problem and an explanatory mechanism has
been proposed in terms of the behaviour of the substituted model terms.
⇒ Section 8.3

• The sensitivity of DES to the underlying RANS model is negligible for massive-
ly-separated flow with geometry-induced separation. For flows with sensitive
turbulent separation from smooth surfaces however the model sensitivity is
high. This elucidates the importance of RANS modelling in a DES context.
⇒ Section 8.4

• Excellent agreement with experiment is demonstrated in an in-depth compar-

ison for cylinder flow. The predictive capability of the LES-mode of DES for
massively-separated wakes is thereby validated. The suspected cancellation of
errors does not diminish this.
⇒ Section 8.5

192
 9.1 Summary of findings

• Natural DES is shown to exhibit a strong sensitivity to the time step size, with a
relatively minor coarsening leading to a dramatic degeneration of the solution
quality. It is argued that this effect can be expected for all DES of bluff bod-
ies with turbulent separation or laminar separation near the critical Reynolds
number.
⇒ Section 8.6

• The implementation of IDDES is thoroughly validated and examined for chan-

nel flow. The results are generally very promising, however a small degree of
log-layer mismatch occurs and hypothetical explanations for this are proposed.
⇒ Section 8.8

• The WMLES modelling paradigm enabled by IDDES is demonstrated to elim-

inate the grid spacing requirements in terms of wall friction units that limit
LES to very low Reynolds numbers. An estimated factor 90000 reduction in
computational cost is achieved.
⇒ Sections 3.5.2 & 8.8.3

• The model sensitivity of WMLES is limited to the near wall RANS region and
negligible in terms of skin friction.
⇒ Section 8.8.2

• The combination of IDDES with hybrid-adaptive RANS wall functions suc-

cessfully removes the dependency of the wall-normal grid resolution on the
Reynolds number and increases the robustness of the method.
⇒ Section 8.8.4

• The combined IDDES and wall function methodology produces acceptable re-
sults on cubic grids, thereby demonstrating the robustness of IDDES to uncon-
ventional mesh construction.
⇒ Section 8.8.4

• IDDES successfully enables mixed DDES and WMLES operation within a sin-
gle solution domain and automatically selects the appropriate modelling frame-
work locally. The IDDES method hence extends DES to include WMLES as an
additional hybrid RANS-LES capability.
⇒ Section 8.9

• Some topics of concern have been documented regarding regions of transition

in IDDES between WMLES and RANS boundary layers and vice-versa.
⇒ Section 8.9

193
9 Conclusion

9.2 Closing comments

In general, the relatively new family of hybrid RANS-LES methods presents the engi-
neer with a significantly enhanced set of tools for the simulation of turbulent flows.
Although generally too computationally expensive to be considered for everyday
application, the expected increases in computational capability will inevitably lead
to increasing reliance on hybrid RANS-LES in the near future. The additional cost
must of course be justified by corresponding benefits for the application in ques-
tion. Hence, two key scenarios are anticipated for the industrial application of hybrid
RANS-LES approaches: Situations in which (U)RANS methods provide poor results
(in particular strongly-separated flows) and applications for which the resolution of
unsteady turbulent motion is essential (a notable example being the field of compu-
tational aeroacoustics).

Among the hybrid RANS-LES approaches, DES has in particular achieved wide-
spread popularity, accounting for the majority of industrial hybrid RANS-LES to
date. It is by now clearly established that excellent results can be achieved for suit-
able applications, which is further supported by the findings of this work. Although
a recent proposition, the IDDES enhancement appears to successfully extend the
range of suitable applications of DES through the inclusion of WMLES capability.
This work corroborates the findings of the method authors in this respect, although
more experience and confidence must be gained from further application to a wider
range of test cases.

Despite these considerably positive perspectives and the significant achievements in

the method development since 1997, a number of fundamental issues remain to be
resolved. It is perhaps important to note that these are inherent to hybrid RANS-
LES methods in general: The grey area problem arises in any formulation interfacing
RANS with LES and the requirement of a sufficiently fine temporal resolution is a
property inherited from LES. An amelioration or resolution of these issues is a clear
direction for future research.

From an industrial perspective, just as important as continuing method development

is accessibility of the technology, for which many factors play a role. In particular,
user education and the availability of concise and comprehensive guidelines are es-
sential if a long term move away from RANS is targeted. Examples of the kind of
practical questions that these must address include the choice of an appropriate hy-
brid RANS-LES strategy for the problem in question, the generation of suitable grids,
the choice of time step and the length of simulation time required for statistics at the
desired level of accuracy. A further key technlology to be pursued in the industri-
alisation of hybrid RANS-LES is a dedicated strategy for automated grid generation
and adaptation.

194
 Part IV

Appendices

195
A The governing equations of fluid motion

A brief description of the origin and properties of the governing equations of fluid
flow will first of all be given, which is a summary of the introductions given by
Pope [116]. The first necessary physical assumption is the continuum hypothesis, i.e.
that the fluid is a continuum of differential variables, and that the discrete nature of
the fluid at molecular scales can be ignored. Considering a fixed inertial, Eulerian
reference frame xi , the variation of a physical quantity φ for a moving fluid particle
is given by the substantial derivative
Dφ ∂φ ∂φ
= + ui , (A.1)
Dt ∂t ∂xi
where u i is the velocity vector.

A.1 Conservation of mass

The conservation of mass, or continuity equation is given by
∂ρ ∂ρu i
+ =0. (A.2)
∂t ∂xi
For all the flows considered in this work, the density ρ is assumed to be constant.
This incompressible flow assumption is approximately valid for aerodynamic flows
at Mach numbers of M ≤ 0.3, and results in considerable simplification of the gov-
erning equations and the numerical implementation. For incompressible flows, the
continuity equation can be rewritten as
∂u i
=0. (A.3)
∂xi
The velocity field is therefore seen to be divergence free or solenoidal, and this assump-
tion will be implicit in the remaining presentation of the fluid flow equations and the
work as a whole.

A.2 Conservation of momentum

The momentum equation is derived from Newton’s second law, relating the accel-
eration of a fluid particle Du i /Dt to the surface and body forces it experiences. The
surface forces are molecular in origin and represented by the stress tensor τij , whereas
the body forces Ψ represent external forces with only gravity considered here.

197
A The governing equations of fluid motion

Du j ∂τij ∂Ψ
ρ = −ρ (A.4)
Dt ∂xi ∂x j
To obtain the expression for the surface forces, the further assumption of a Newto-
nian, constant property fluid is made, giving
 
∂u i ∂u j
τij = − p̃δij + μ + , (A.5)
∂x j ∂xi
where p̃ is the pressure, acting as an isotropic normal force, and μ is the dynamic vis-
cosity coefficient relating the shear force to the velocity gradients. The velocity gra-
dient tensor can be decomposed into isotropic, symmetric-deviatoric and antisym-
metric-deviatoric parts as
∂u i 1 ∂u i
= + Sij + Ωij . (A.6)
∂x j 3 ∂u i
The isotropic part is zero for incompressible flows according to Eq. (A.3). The sym-
metric-deviatoric rate of strain tensor and antisymmetric-deviatoric rate of rotation
tensor are given by
   
1 ∂u i ∂u j 1 ∂u i ∂u j
Sij = + and Ωij = − (A.7)
2 ∂x j ∂xi 2 ∂x j ∂xi
respectively. The stress tensor can therefore be reformulated as

τij = − p̃δij + 2μSij . (A.8)

Some degree of simplification is obtained when the “modified pressure” p is con-
sidered to include the isotropic body forces, i.e. p = p̃ + ρΨ. Substituting the stress
tensor expression (A.8) into the momentum equation (A.4) with uniform ρ and μ and
the modified pressure, the incompressible Navier–Stokes equations are obtained:
Du i 1 ∂p ∂
=− + (2νSij ) , (A.9)
Dt ρ ∂xi ∂x j
where ν is the kinematic viscosity given by ν = μ/ρ.

198
B RANS models

The purpose of this appendix is to list the formulations of the RANS models ap-
plied in the study, with discussions of the model properties and derivations kept to
a minimum. In cases where variations of the formulations exist, the version as it is
implemented in the flow solver will be quoted and the existence of variations noted.
The following definitions for the scalar norms of the deformation tensor will be used
throughout:
 
S ∗ = 2Sij Sij , Ω∗ = 2Ωij Ωij , (B.1)
where Sij and Ωij are defined in Eq. (A.7).

B.1 One-equation models

B.1.1 Spalart–Allmaras
The Spalart-Allmaras (SA) one-equation model [157] takes the form of a transport
equation for a modified eddy viscosity, ν̃ as follows:

D ν̃
= Pν̃ + Dν̃ − ν̃ ,
Dt

where the production, diffusion and dissipation terms are

   
1 ∂ ∂ν̃ ∂ν̃ ∂ν̃
Pν̃ = Cb1 [1 − f t2 ] S̃ν̃ , Dν̃ = (ν + ν̃ ) + Cb2 ,
σ ∂xk ∂xk ∂xk ∂xk
  2
Cb1 ν̃
ν̃ = Cw1 f w − f t2 , (B.2)
κ2 dw
respectively. ν̃ is related to the eddy viscosity, νt by
χ3 ν̃
νt = ν̃ f v1 , f v1 = , χ= . (B.3)
χ3 + Cv1
3 ν

The modified scalar norm of the deformation tensor is given by

ν̃ χ
S̃ = Ω∗ + f v2 , f v2 = 1 − . (B.4)
κ2 d2w 1 + χ f v1

199
B RANS models

The wall-damping function in the destruction term is defined by

 1/6
1 + Cw3
6
ν̃
fw = g , g = r + Cw2 (r6 − r ) , r= . (B.5)
g6 + Cw3
6 S̃κ2 d2w

The remaining terms with subscript t are intended to enable a smooth transition
between a user-defined laminar region and the modelled turbulent region. As this
functionality is not employed, these terms are ignored, i.e. f t1 = f t2 = 0. The values
of the model constants are

Cb1 1 + Cb2
Cb1 = 0.1355 , Cb2 = 0.622 , σ = 2/3 , Cw1 = + ,
κ2 σ
Cw2 = 0.3 , Cw3 = 2 , κ = 0.41 , Cv1 = 7.1 . (B.6)

B.1.2 Edwards modification

The Edwards modification [39] to the original SA model, referred to as SAE, was
proposed to address stability problems in the semi-viscous region. This involves a
modified formulation of S̃, which guarantees a positive production term and further-
more employs the usual S ∗ rather than Ω∗ :
  
1
S̃ = S ∗ + f v1 . (B.7)
χ
A modified definition of r is also proposed:
 
ν̃
r = 1.313 tanh . (B.8)
S̃κ2 d2w

B.1.3 SALSA
The strain-adaptive linear variant of the Spalart–Allmaras (SALSA) model of Rung
et al. [134] is based on the SAE formulation, and involves a modification of the pro-
duction and dissipation terms to sensitise these to non-equilibrium flows. The mod-
ification is effected by a replacement of the constant Cb1 with a function f Cb1 and by
a slightly modified r, as follows:

√
f Cb1 = Cb1 Γ , Γ = min [1.25; max (γ; 0.75)] , γ = max (α1 ; α2 ) ,
 0.65 #  χ $0.65
ν̃
α1 = 1.01 ∗ 2 2 , α2 = max 0; 1 − tanh , (B.9)
S κ dw 68

200
 B.2 Two-equation models
 
ν̃
r = 1.6 tanh 0.7 . (B.10)
S̃κ2 d2w

B.2 Two-equation models

B.2.1 Wilcox k − ω
The Wilcox k − ω model [181] comprises two transport equations for k and ω, where
the turbulent frequency is defined as
ε
ω= . (B.11)
Cμ k
The k and ω transport equations are given in Eqs. B.12 and B.13 respectively.

Dk
= Pk + Dk − k ,
Dt
  
∂ νt ∂k
Pk = νt S ∗2 , Dk = ν+ , k = Cμ kω . (B.12)
∂x j σk ∂x j

Dω
= Pω + Dω − ω ,
Dt
  
ω ∂ νt ∂ω
Pω = Cω1 P , Dω = ν+ , ω = Cω2 ω 2 . (B.13)
k k ∂x j σω ∂x j

The parameters k and ω give an eddy viscosity defined by

k
νt = (B.14)
ω
and the model constants are

Cμ = 0.09 , Cω1 = 5/9 , Cω2 = 3/40 , σk = σω = 2 . (B.15)

This is the original version of the model, published in 1988. Wilcox continued to de-
velop the model to include re-calibrated constants and additional terms to address
laminar to turbulent transition [182]. These modifications have not been incorpo-
rated in the implemented version.

B.2.2 LLR k − ω
The local linear realisable (LLR) model of Rung et al. [136] has been derived from
non-linear models in order to sensitise the k − ω model to strain and rotation effects
and improve behaviour in the wall region. The eddy viscosity level depends on
the strain and rotation rates while the model coefficients depend on the turbulent
Reynolds number. Note in particular that, in contrast to the Wilcox k − ω model, the

201
B RANS models

turbulent frequency is defined without Cμ (or its LLR variable counterpart, f Cμ ):

ε
. ω= (B.16)
k
The remaining changes compared to the Wilcox k − ω model are as follows:

k = f β k kω (B.17)

 
 f Cμ S ∗
∗2
Pω = S f Cμ f 1 f C1 − , ω = f β ω ω 2 (B.18)
ω

k
νt = f Cμ (B.19)
ω

 

1
f Cμ = f μ max 0.04; min 0.1; , AS = 2.12 ,
4 + AS Ũ
2  
0.5 ( S ∗2 + Ω∗2 ) 1 + Re
μ Ret α
Ũ = , f μ = 80 , Reμ = ,
ω 1 + Reμ 70
 
k Ret
Ret = , α = 1 + 0.9 sign 1; − 0.9 ,
νω 70
 2
  + Re
1
S ∗ /ω
t
9070
0.83
+ Rek
f C1 = max 0.43; ∗ , f1 =  2 , f βk = 3
,
S /ω + 4.265 1 + Rek
1 + Re
70
t

 2.5  0.5   
Ret Ret Ret
Rek = A∗ + (1 − A ∗ ) , A∗ = tanh 0.5 ,
100 100 100
⎡ ⎛⎞ −1 ⎤

⎢  ⎥
⎜  f Cμ ⎟
f βω = max ⎢
⎣1.83 ⎝1 + νt ⎠ − 1 ; 0⎥
⎦ . (B.20)
1+
ν
This is the version of the LLR model as it is implemented in ELAN, although some
modifications exist as reported by Xue [184], Schatz [141], Franke [46] and Rung
et al. [136]. In particular, it should be noted that according to Schatz, the constant

202
 B.2 Two-equation models

value of AS = 2.12 is an approximation of a more complex function valid for two-

dimensional flows only. The full variant is in fact
⎡ ⎛ ⎞⎤
√
⎢ ⎜ 6S S S
ij jk ki ⎟⎥
AS = 3 cos ⎣0.5 cos−1 ⎝  3/2 ⎠⎦ . (B.21)
Sij Sij
The effect of this modification has not been studied in the current work.

B.2.3 LL k − ε

The Lien–Leschziner (LL) k − ε model [81] is used as the background model for the
CEASM (Section B.3), and consists of two transport equations for k and ε combined
with damping terms to allow resolution directly to the wall. The transport equations
are given in Eqs. B.22 and B.23.

Dk
= Pk + Dk − k ,
Dt
  
∂ νt ∂k
Pk = νt S ∗2 , Dk = ν+ , k = ε . (B.22)
∂xi σk ∂xi

Dε
= Pε + Dε − ε ,
Dt
  
ε ∂ νt ∂ε ε
Pε = C̃ P , Dε = ν+ , ε = C̃ ε . (B.23)
k ε1 k ∂xi σε ∂xi k ε2
The eddy viscosity is given by
k2
νt = f μ Cμ (B.24)
ε
and the coefficients and constants of the model are

   
P C̃ε2 k3/2 −α d Re2k
C̃ε2 = Cε2 1 − 0.3e−Ret , P =
2
C̃ε1 = Cε1 1 + , e ,
Pk Cε1 L ε
√  
k2 kdw
Ret = , Rek = , L ε = κCμ3/4 dw 1 − e−α ε Rek ,
νε ν
1 − e−α μ Rek
fμ = , (B.25)
1 − e−α ε Rek

203
B RANS models

σk = 1 , σω = 1.3 , Cε1 = 1.44 , Cε2 = 1.92 , Cμ = 0.09 ,

αd = 0.00222 , αε = 0.263 , αμ = 0.016 . (B.26)

B.3 CEASM model

The compact explicit algebraic stress model (CEASM) of Lübcke [84] was developed
for application to strongly three-dimensional flows and derived by simplification of
the Reynolds stress transport equations under equilibrium assumptions. The model
uses a restricted set of tensors in the representation of Pope [115] (which is not re-
peated here), for which the coefficients are

4A3 f Cμ A2 f Cμ I IS
β1 = −2 f Cμ , β2 = − , β4 = − β ,
g g 2 8
  
12A2 f Cμ A3 − A2 − II IIΩS 12A2 f Cμ
β5 = , β8 = − ,
−2g2 + A23 I IS + A22 I IΩ I IS g

A1 = −0.472 , A2 = −0.775 , A3 = −0.375 ,

− A1 g
f Cμ = , (B.27)
g2 − 3 A3 I IS − 2A2 I IΩ
2 2 2

with g given by
I IS 2k2
g = C1 − 1 + √ . (B.28)
4 + 1.83 0.4I IS − 1.6I IΩ ε2
These coefficients provide a description of the anisotropy tensor based on the mean
flow gradients. Alongside these, a description of the turbulent length and time
scales must be provided – this is the role of the background model in explicit al-
gebraic stress models. Although the CEASM framework could be applied to any
two-equation model, the LL k − ε model was chosen by Lübcke for its robustness
and ability to resolve the viscous sublayer.

204
C Hybrid-adaptive wall boundary condition

The existence of two distinct wall boundary conditions has its origins in the physical
nature of turbulent boundary layer flow: as detailed in Section 2.4, distinct zones
are seen to exist at various distances to the wall. One class of boundary conditions
(widely known as “low-Reynolds number” boundary conditions) targets the full res-
olution of the boundary layer profile to the wall, including the viscous sublayer. Best
practice guidelines for RANS grid design [23] hence stipulate that low-Re boundary
conditions require the placement of the first point at a distance of y+ ≈ 1 in order
to sufficiently resolve the strong flow field gradients here. The alternative boundary
condition type, referred to as “high-Reynolds number” or “wall functions”, exploit
the self-similarity of turbulent boundary layers to bridge the viscous sublayer and
buffer layer. The first grid cell is therefore placed within the log-law region and
the corresponding best practice recommendation is that the first point be located at
y+ > 20. Although this allows the use of much coarser wall-normal grid resolution,
the self-similar boundary layer profiles are only valid for flows with weak tangen-
tial pressure gradients and low surface curvature where approximate local turbulent
equilibrium can be assumed. When the simulated flow departs significantly from
these assumptions, such as in the vicinity of shocks or flow separation and reattach-
ment, a considerable loss of solution quality can be expected.

Different families of RANS models exhibit different “inherent” wall boundary condi-
tions - whereas the k − ε models are of the high-Re type, unmodified k − ω model so-
lutions must be resolved entirely using a low-Re formulation. The Spalart–Allmaras
and related models are similarly designed for integration directly to the wall, al-
though reliable solutions are reported with y+ values as large as five [157]. Nonethe-
less, damping terms can be incorporated into k − ε formulations in order to incorpo-
rate the viscous sublayer into the solution (e.g. the Lien–Leschziner k − ε model [81])
and high-Re boundary conditions can be added to k − ω models [145].

Inspection of the y+ criteria for low-Re and high-Re boundary conditions reveals a
further important problem with this state of affairs: a gap is present in the region
1 < y+ < 20 for which neither formulation is valid. A simple switch between the
alternative formulations depending on the local y+ value is therefore not possible.
In order to address these problems, a more general unified boundary condition that
can deliver robust and reliable solutions on arbitrary grids is highly desirable. Such a
generic boundary condition has been proposed by Rung et al. [135], the formulation
of which shall be presented.

The hybrid adaptive boundary condition represents an extension of the high-Re for-
mulation such that it blends seamlessly into a low-Re formulation inside the viscous

205
C Hybrid-adaptive wall boundary condition

sublayer. The dimensionless wall-normal distance in the log-law region is given by

the following rearrangement of Eq. (2.20):
+
e(κU )
y+ = , (C.1)
E
where E = eκB . A Taylor series expansion of the numerator of Eq. (C.1) for small
+ +
values of κU about the origin combined with the low-Re value of y+ (i.e. y+ = U )
delivers the hybrid expression for the dimensionless wall distance
  + +

+ + 1 (κU + ) + (κU )2 (κU )3
yhyb = U + e − 1 + κU + + + ... . (C.2)
E 2! 3!
This incidentally corresponds to the law of the wall formula published by Spald-
ing [161]. Armed with this relatively simple expression for y+
hyb , the hybrid adap-
tive boundary condition can be formulated. This is activated in place of the high-
Re boundary condition when the criterion y+ ≤ 15 is met. The mechanism of the
boundary condition is as follows:

 +

+ 1 +
+ (κU )9
y+
hyb =U + e κU
− 1 − κU − ... −
E 9!

  ln( Ey+ )
+ −0.14y + hyb
U = 1−e hyb
κ
 2 μ ρκU τ
−0.09y +
ϕ = 1−e hyb , θ = (1 − ϕ ) +ϕ , τw = θ (ΔU w ) (C.3)
dw ln( Ey+
hyb )

The blending function ϕ varies smoothly between near-zero in the viscous sublayer
to unity deep inside the log law region, at approximately y+ > 100. This serves to
define the blending between laminar and turbulent viscosity in the low-Re and high-
Re regions respectively, which is effective in the parameter θ. This in turn gives the
effective viscosity divided by the wall normal distance for determination of the shear
stress τw in conjunction with the velocity magnitude relative to the wall at the first
point, ΔU w . The velocity profile resulting from Eqs. (C.3) is plotted in Fig. C.1 and
shows good agreement with the analytical linear and logarithmic distributions.

A demonstration of the effectiveness of the hybrid-adaptive boundary condition for-

mulation of Rung et al. [135] for a range of aerodynamic validation cases and indus-
trial aerodynamic applications is given by Schmidt et al. [145] and Mockett et al. [95].

Knopp [74] has recently proposed a similar but more elaborate boundary condition,
which also allows placement of the first grid point at any value of y+ . It also utilises
Spalding’s formula, Eq. (C.2), but combines this with elements of Reichardt’s correla-

206
 30
+ +
U =y
25 + -1 +
U = κ ln y + B
hybrid-adaptive BC
20

+
15

U 10

0 0 1 2 3 4
10 10 10 10 10
+
y

Figure C.1: Mean velocity profile given by the hybrid-adaptive boundary condition of
Eqs. (C.3) compared to the linear and log-law profiles of Eq. 2.19 and Eq. 2.20 re-
spectively.

tion, Eq. (2.21), and terms derived specifically for different families of RANS models.
Knopp’s method is hence model-specific and gives grid-independent solutions up to
the limitations of numerical errors. The effects of streamwise pressure gradients are
furthermore incorporated to give a more generalised formulation, which is particular
important for flows with boundary layer separation and attachment.

207
C Hybrid-adaptive wall boundary condition

208
D Derivation of model-specific Ψ functions

The derivation of the Ψ functions for each DES model follows the procedure outlined
by Spalart et al. [158] and presented in more detail by Garbaruk et al. in an internal
DESider project report [50].

D.1 Spalart–Allmaras model without “trip terms”

Although this model has not been applied in the presented investigations, the deriva-
tion of the low-Reynolds correction function is presented as an initial step, based
upon which the derivation of ΨSAE is presented in the next section. The low-Rey-
nolds term correction function derived by Spalart et al. [157] and discussed in Section
3.7.3 was obtained by consideration of the original SA model, including the terms
allowing a user-specified laminar region. As these terms are not implemented, an al-
ternative version of the low-Reynolds term correction function, ΨSA needs to be de-
rived. A Smagorinsky-like form for the SA model without trip terms can be derived
by employing the local equilibrium assumption (i.e. a balance of production and de-
struction terms in the transport equation). Adopting the LES length scale including
the correction term (3.18) in the derivation gives rise to the following Smagorinsky
form:

νt = ASA (ΨSA CDES Δ )2 S ∗ , (D.1)

with
Cb1 f v1 S̃
ASA = . (D.2)
Cw1 f w S ∗
As observed in Section 3.7.3, this coefficient term is not a constant. In order to re-
turn to the desired Smagorinsky behaviour, the correction function must cancel this
tendency, i.e. ASA Ψ2SA = const. Examining the variation of ASA with eddy viscosity
shows that the function has an asymptotic value ASA ∗ for ν → ∞. Correspondingly,
t
we seek a function such that the asymptotic correction function ΨSA ∗ = 1, which
gives the following expression for the correction function:
∗
ASA
Ψ2SA = . (D.3)
ASA
Rearranging the expression for ν̃ obtained using local turbulent equilibrium gives
ν̃ Cb1
= , (D.4)
S̃ (ΨSA CDES Δ )2 Cw1 f w
which together with the definition of r (B.5) in LES-mode (i.e. dw = ΨSA CDES Δ)
gives

209
D Derivation of model-specific Ψ functions

Cb1
r= . (D.5)
Cw1 f w κ2
Substituting (D.5) in the definition of S̃ in LES-mode, one obtains
S∗ f C
= 1 − 2 v2 b1 . (D.6)
S̃ κ Cw1 f w
The asymptotic value ASA ∗ shall now be derived by considering the dependencies
of the constituent functions on the eddy viscosity ratio, χ. Clearly, as νt ν and
χ → ∞, S̃ → S ∗ and f v1 → 1. The resulting function is:

∗ Cb1
ASA = , (D.7)
Cw1 f w∗
where f w∗ denotes the asymptotic value of f w , for which an iterative solution deliv-
ers f w∗ = 0.4241. Unlike the full SA model, however, this is not a function of χ, so
f w∗ = f w .

The correction function can now be derived from (D.3) and (D.6), and is presented in
(D.8). As also reported by Spalart and coworkers [158], the value of ΨSA is limited
to 10. Furthermore, in line with expectation, the expression is equal to the function
derived for DDES [158] with the trip terms deactivated (i.e. f t1 = f t2 = 0).
 
1 f C
Ψ2SA = min 102 ; − ∗ v2 2b1 (D.8)
f v1 f w f v1 κ Cw1

D.2 Spalart–Allmaras model with Edwards modification

For the SAE model [39], the derivation of the DES correction function follows the
same method as for the SA model without trip terms, described above. The same
Smagorinsky form is obtained, with an identical expression for the coefficient term,
i.e. ASAE = ASA . However, due to the different definition of S̃, the resulting correc-
tion function differs; this definition gives
  −1
S∗ 1
= + f v1 , (D.9)
S̃ χ
which together with the definition of the correction function (D.3) and the expression
for the Smagorinsky coefficient (D.2) and its asymptotic limit (D.7) results in:
  −1
fw 1
Ψ2SAE = ∗ + f v1 . (D.10)
f w f v1 χ
Due to the modified definition of r, a slightly different value of f ∗ = 0.4685 is ob-
tained for SAE. As for the SA model without trip terms, f w∗ = f w , which together
with an analogous limitation of the function and the denominator used by Spalart et
al. [158], gives:

210
 D.3 SALSA model

⎧   −1 ⎫
⎨
max( f v1 , 10−10 ) ⎬
Ψ2SAE = min 102 ; + f v1
2
. (D.11)
⎩ max(χ , 10−10 ) ⎭

The behaviour of ΨSAE with eddy viscosity ratio is plotted in Fig. D.1.

10

8
ΨSAE

0
0.01 0.1 1 10 100
νt / ν

Figure D.1: Profile of ΨSAE with νt /ν.

D.3 SALSA model

As mentioned in Appendix B, the SALSA model is implemented including the Ed-
wards correction. There are two relevant changes to the formulation compared to the
SAE model, namely the SA and SAE constant Cb1 is replaced by a function denoted
f Cb1 and a change is introduced to the expression for r. f Cb1 is present in the destruc-
tion term directly as well as in Cw1 , which no longer being a constant is termed
f Cb1 1 + Cb2
f Cw1 =
+ . (D.12)
κ2 σ
Correspondingly the coefficient term becomes
f Cb1 f v1 S̃
ASALSA = . (D.13)
f Cw1 f w S ∗
Examination of the asymptotic value of the coefficient term for high values of eddy
viscosity is made slightly more complex by the additional functions, however the
asymptotic values can be obtained when assumed values are chosen for S ∗ and dw ,
which appear in the α2 component of f Cb1 (n.b. dw is not substituted by the DES
length scale in this expression). Furthermore, the development of f Cb1 , f Cw1 and f w
as a function of χ can be examined, and it emerges that all of these functions re-
main bounded within specific values for all χ and for any values of dw and S ∗ . The

211
D Derivation of model-specific Ψ functions

minimum, maximum and asymptotic values of each function are summarised in Ta-
ble D.1.

Minimum: Maximum: χ → 0: χ → ∞:
f Cb1 0.1174 0.1515 0.1355 (= Cb1 ) 0.1515
f Cw1 3.1311 3.3342 3.2391 (= Cw1 ) 3.3342
fw 0.7032 0.7055 0.7042 0.7055

Table D.1: Bounding and asymptotic values of the variable terms appearing in ASALSA .

The correction function for the low-Reynolds terms is therefore derived to be

  −1
f C∗ f Cw1 f w 1 1
Ψ2SALSA = b1 ∗ + f . (D.14)
f Cb1 f Cw1 f w∗ f v1 χ v1

As the functions shown in Table D.1 are seen to vary only weakly with χ, the follow-
ing simplification is justifiable:
f C∗b1 f Cw1 fw
= = ∗ =1. (D.15)
f Cb1 f C∗w1 fw
Together with the implemented limits used for the SAE model, this gives
⎧   −1 ⎫
⎨ max ( f , 10 −10 ) ⎬
v1
Ψ2SALSA = min 102 ; + f 2
= Ψ2SAE . (D.16)
⎩ max(χ, 10−10 ) v1 ⎭

D.4 Wilcox k − ω model

The Smagorinsky form can be simply derived from the original (1988) Wilcox k − ω
model [181] by the assumption of local equilibrium in the k equation (in LES-mode)
and ω equation respectively:

k3/2
νt S ∗2 = (D.17)
ΨWCX CDES Δ
Cω2
νt S ∗2 =
kω . (D.18)
Cω 1
Substituting ω = k/νt in (D.18) gives the following expression for k:
"
Cω1
k= νt S ∗ , (D.19)
Cω2

212
 D.5 LLR k − ω model

which together with (D.17) gives the Smagorinsky form of the Wilcox subgrid scale
model

νt = AWCX (ΨWCX CDES Δ )2 S ∗ , (D.20)

where
 
Cω2 3/2
AWCX = . (D.21)
Cω1
As this is a constant, no low-Reynolds number correction term need be implemented:

ΨWCX = 1 . (D.22)

D.5 LLR k − ω model

Due to the complex terms constituting the LLR k − ω model formulation, the deriva-
tion of ΨLLR is not as straightforward as for other models. Indeed, it is difficult to
reach a conclusion concerning the necessity of this by analytical means alone, as will
become apparent. The local equilibrium assumption delivers

k3/2
νt S ∗2 = (D.23)
ΨLLR CDES Δ
for the k-equation in LES mode, and
 
 f Cμ S ∗
∗2
S f C μ f 1 f C1 − = f βω ω 2 (D.24)
ω
for the ω equation. Substituting ω = f Cμ k/νt in (D.24) gives the following expression
for k
  

 f f Cμ S ∗
k = 3/21 f C1 − νt S ∗ , (D.25)
f Cμ f β ω ω

which when substituted in (D.23) gives the Smagorinsky form of the LLR k − ω
model in LES-mode

νt = ALLR (ΨLLR CDES Δ )2 S ∗ (D.26)

with
⎡  ⎤2/3
f1 f Cμ S ∗
ALLR = ⎣ f C1 − ⎦ . (D.27)
f C3/2 f βω ω
μ

It can be seen that ALLR is a function of νt , k and ω as well as the deformation tensor
in the form of S ∗ and the argument of the variable Ũ (a function of S ∗ and Ω∗ ).
This lack of a simple dependency on νt complicates the situation considerably, and

213
D Derivation of model-specific Ψ functions

a function ΨLLR cannot be derived1 . A parameter variation study reveals that ALLR
can even have an amplifying effect as well as damping. Furthermore, the damping
effect is weak in comparison to the SA model family, with the coefficient minimum
value roughly 0.5. This, together with the difficulty of derivation mean that

ΨLLR = 1 (D.28)
is implemented. Unfortunately, it is demonstrated empirically in Sect. 7.4 that the
subgrid viscosity is indeed damped to an unacceptable extent for fine grids in the
simulation of DIT. An effective ΨLLR function would hence be required in order for
LLR-based DES to be considered valid and its use is not advised.

D.6 LL k − ε model
The properties of the LL k − ε model will be investigated due to the role this plays
as background model for the CEASM-based DES. For the k and ε equations in LES-
mode, the local equilibrium assumption delivers

k3/2
νt S ∗2 = (D.29)
ΨLL CDES Δ
and

C̃ε1 νt S ∗2 = C̃ε2 ε , (D.30)

respectively. A rearrangement of the eddy viscosity expression gives

f μ Cμ k 2
ε= , (D.31)
νt
which when substituted in (D.30) gives
"
C̃ε1
k= νt S ∗ . (D.32)
C̃ε2 f μ Cμ
Substituting (D.32) in (D.29) delivers the Smagorinsky form

νt = ALL (ΨLL CDES Δ )2 S ∗ , (D.33)

with
 3/2
C̃ε2 f μ Cμ
ALL = . (D.34)
C̃ε1
Inspection reveals that ALL is not a constant, rather a function of k, ε, dw and S ∗ . The
strongest damping behaviour is however a function principally of k, with the other
1
It cannot of course be ruled out that a researcher with greater algebraic ability may be able to derive an
effective ΨLLR .

214
 D.7 CEASM model

variables adjusting only the profile of the curve. Asymptotic behaviour is observed
for k → ∞ independent of the other variables: C̃ε1 → Cε1 , C̃ε2 → Cε2 and f μ → 1.
The asymptotic limit ALL → ALL ∗ is therefore

 
∗ Cε2 Cμ 3/2
ALL = . (D.35)
Cε1
Applying the usual definition Ψ2LL = ALL ∗ /A , the correction function including
LL
numerical limitations is obtained:
⎧  3/2 ⎫
⎨ C C̃ ⎬
ε2 ε1
Ψ2LL = min 102 ; . (D.36)
⎩ −
max( f μ ; 10 ) C̃ε2 Cε1
10 ⎭

D.7 CEASM model

Although a low-Reynolds term correction function could be derived above for the
LL k − ε model, the background model for the CEASM, analysis of the CEASM is
made difficult by the complex dependencies on the velocity field gradients. These
are partly manifest in an adjustment of the Cμ parameter which is a constituent of
ALL (and hence ACEASM ), as well as determining an anisotropic Reynolds stress ten-
sor. A straightforward derivation of a correction function analogous to the LL model
is therefore considered impossible, partly due to the algebraic complexity of the mod-
ified Cμ , and partly because the Reynolds stress anisotropy will also be active in LES
mode although not “captured” by the local equilibrium based analysis. Empirical
testing of the LES mode performance of CEASM-based DES is hence considered es-
sential.

The results of such a test employing decaying isotropic turbulence are shown in Sec-
tion 7.4. It is seen that despite the indications from the LL k − ε analysis, which sug-
gest the necessity of a ΨLL , and despite the complex dependencies outlined in the
above paragraph, CEASM-based DES delivers reliable results without a Ψ function.

215
D Derivation of model-specific Ψ functions

216
E Obtaining resolved and modelled turbulent statistics

A short description of the method used to collect resolved and modelled Reynolds
stress data from DES simulations is given. This involves a combination of the meth-
ods used for RANS and LES, which will therefore be described first.

E.1 Method for RANS

If a non-linear RANS model (such as the CEASM) is used, the Reynolds stress com-
ponents are a direct output of the model. For linear eddy viscosity models however,
the Boussinesq hypothesis ([13], Section 3.4.2) must be invoked to provide values of
the approximated Reynolds stress components:
 
  2 ∂Ui ∂Uj
u i u j = δij k − νt + . (E.1)
3 ∂x j ∂xi
This is written out in full for the u  u  and v v normal stress components and the u  v
component of the shear stress:

 
2 ∂U ∂U
u u = k − νt + (E.2)
3 ∂x ∂x
 
2 ∂V ∂V
v v = k − νt + (E.3)
3 ∂y ∂y
 
∂U ∂V
u  v = − νt + . (E.4)
∂y ∂x
If a model is used that does not directly include the parameter k (for example the SA
and derivative models), this can be approximated from the eddy viscosity employing
Bradshaw’s hypothesis ([15], Section 3.4.3):
νt S ∗
k≈ 2 . (E.5)
Cμ

E.2 Method for LES

Whereas the RANS Reynolds stresses are purely modelled quantities emerging from
a steady-state solution, the majority of the turbulent stresses in LES are resolved
directly in the flow field. These must therefore be obtained statistically over a large
number of time steps N as follows:

217
E Obtaining resolved and modelled turbulent statistics

% 7
N
1
U=
N ∑U (E.6)
t =1
% 7
N
1
u i u j =
N ∑ Ui Uj − Ui Uj (E.7)
t =1

It is assumed that the time averaging is begun at a time step t = 1 where the flow field
is fully-developed and free of initial transient character. Note that the summation
part of the computation (inside the curly brackets) can be done cumulatively during
the simulation, and the division by N and subtraction of mean values at the final
output stage. This makes it unnecessary to store the entire flow domain at every time
step. From the resolved Reynolds stresses, the resolved turbulent kinetic energy can
be computed as:
1  
kres =
uu . (E.8)
2 i i
The resolved Reynolds stresses obtained above are however only part of the total
Reynolds stresses in LES. Because the modelled subgrid-scale stresses should be
much smaller than the resolved stresses in LES (Section 3.5.2), these are often ne-
glected. It is however simple to calculate the modelled component: Assuming that
a subgrid-scale model is used that produces an eddy viscosity (e.g. the Smagorin-
sky model), the modelled part of the Reynolds stress can be obtained using the same
method as for the RANS. Unlike the RANS however, these quantities must also be
time averaged. Comparing the relative levels of resolved and modelled Reynolds
stress is a useful means of determining the sufficiency of the grid resolution: the re-
solved stresses should be much larger than the modelled stresses.

E.3 Method for DES

In DES, pure RANS and pure LES coexist in different regions of the flow. It is there-
fore imperative that both the modelled and resolved turbulent quantities are com-
puted, requiring a combination of the methods for RANS and LES. As mentioned for
the modelled Reynolds stresses in the LES case, the same method as for RANS can
be used although the quantities must be time averaged.

218
F Phase averaging of quasi-periodic turbulent flows

For many types of flow, for example bluff body wakes, oscillating airfoils or pul-
sating jets employed in active flow control, the fluctuations consist of a coherent,
quasi-periodic part and an incoherent, turbulent part. In many cases, the coherent
fluctuations exist at much larger scales than the turbulent eddies with which they are
overlaid, and it can be argued that these are non-turbulent in nature. For example,
as described in Section 2.5, a laminar vortex street behind a circular cylinder exists at
Reynolds numbers much lower than those where turbulent flow first emerges [189]
and the oscillation of a shock over a transonic airfoil is an acoustically-driven phe-
nomenon that occurs for both laminar and turbulent boundary layers [36].

In such cases, pure Reynolds averaging of the velocity signals would deliver an over-
estimation of the turbulent stresses, as the contribution of the non-turbulent coherent
fluctuations to the unsteadiness would be included in the budget. A more appropri-
ate method of analysis is therefore the triple decomposition, proposed by Reynolds
& Hussain [122], into a mean component, a quasi-periodic coherent fluctuation and
a random fluctuation:

u i ( xi , t) = U i ( xi ) + ũ i ( xi , t) + u i ( xi , t) . (F.1)
A pictorial representation of this process is given in Fig. F.1 for the simple example
of a sinusoidal signal ũ.

u(t) U ~
u(t) u’’(t)

= + 0 + 0
t t
0 0
t t

Figure F.1: Triple decomposition of an unsteady velocity signal u into a constant mean compo-
nent U, a periodic fluctuation ũ and an incoherent turbulent fluctuation u  [122].

The notation u i is adopted for the incoherent fluctuation in order to differentiate it
from the fluctuation u i obtained by Reynolds averaging (described in Section 3.4.1).
The fluctuating velocity in the sense of Reynolds averaging can hence be expressed
as

u i = ũ i + u i (F.2)
and the Reynolds stress tensor as the sum of the coherent and incoherent contribu-
tions

219
F Phase averaging of quasi-periodic turbulent flows

u i u j = ũ i ũ j + u i u j . (F.3)

To realise the triple decomposition practically, the conditional averaging technique

of phase averaging is applied. The phase-averaged velocity is defined as

Ui  = U i + ũ i . (F.4)
The coherent fluctuation ũ i ( xi , t) is by definition periodic, repeating with a period τ
that can be divided into discrete phase angles 0 < ϕ < 2π. Once a method has been
established to determine the corresponding phase angle for each instant t (such as the
methods discussed in Section 6.5), conditional averaging of the signals with respect
to ϕ can be conducted. The quantities Ui ( xi , ϕ) and u i u j ( xi , ϕ) can hence be
obtained, where u i u j  represents the phase-averaged incoherent turbulent stresses.
From these phase-averaged quantities the time-independent quantities

U ( xi ) = Ui  , (F.5)
  
ũ i ũ j ( xi ) = Ui  − U i Uj  − U j and (F.6)

u i u j ( xi ) = u i u j  (F.7)

can be obtained by ensemble averaging over all ϕ.

220
G Figures for the bump test case

(a) CEASM-DES

(b) SAE-DDES

(c) SAE-IDDES

Figure G.1: Comparison of time-averaged streamlines between computations.

221
G Figures for the bump test case

1,5 1,5 1,5

1 1 1

0,5 0,5 0,5

0 0 0
Cp

Cp

Cp
Exp, z = 0 m Exp, z = 0 m
-0,5 Exp
-0,5 Exp, z = 0.085 m -0,5 Exp, z = 0.085 m
CEASM-DES Exp, z = 0.17 m Exp, z = 0.17 m
-1 -1 -1
SAE-DDES CEASM-DES, z = 0 m SAE-DDES, z = 0 m
-1,5 SAE-IDDES -1,5 CEASM-DES, z = 0.085 m -1,5 SAE-DDES, z = 0.085 m
CEASM-DES, z = 0.17 m SAE-DDES, z = 0.17 m
-2 -2 -2
-0,5 0 0,5 1 -0,5 0 0,5 1 -0,5 0 0,5 1
x x x

(a) z = 0, all models (b) various z, CEASM-DES (c) various z, SAE-DDES

Figure G.2: Comparison of time-averaged pressure coefficient along the bump floor at z = 0,
z = 0.085 and z = 0.17. The reference pressure is taken at x = 0.01.

0,3
LDV
CEASM-DES
0,2 SAE-DDES
SAE-IDDES
y

0,1

-0,2 0 0,2 0,4 0,6 0,8 1 1,2

x, (U / 30) + x

0,3

0,2
y

0,1

-0,2 0 0,2 0,4 0,6 0,8 1 1,2

x, (V / 12) + x

Figure G.3: Comparison of time-averaged velocity profiles at z = 0.

222
 0,3

0,2
y

0,1

-0,2 0 0,2 0,4 0,6 0,8 1 1,2

x, u’v’mod + x

0,3

0,2
y

0,1

-0,2 0 0,2 0,4 0,6 0,8 1 1,2

x, (u’v’res / 10) + x

0,3

0,2
y

0,1

-0,2 0 0,2 0,4 0,6 0,8 1 1,2

x, (u’v’tot / 10) + x

Figure G.4: Comparison of Reynolds shear stress profiles at z = 0.

Δt NΔt t avg xs1 |z=0 xr1 |z=0 xs2 |z=0 xr2 |z=0
CEASM-DES 8 × 10−4 s 9000 7.2s −0.188 −0.156 0.043 0.347
SAE-DDES 5 × 10−4 s 21000 10.5s −0.187 −0.128 −0.023 0.256
SAE-IDDES 5 × 10−4 s 23300 11.65s −0.186 −0.131 −0.026 0.233
Experiment — — — — — — 0.630

Table G.1: Summary of time step size, number of time steps collected for averaging and the
equivalent physical time. The right hand four columns give the x positions of the
first and second separation and reattachment points in the centre plane.

223
G Figures for the bump test case

Figure G.5: Instantaneous vortex core structures visualised as isosurfaces of the λ2 criterion
shaded with streamwise velocity (SAE-DDES).

(a) CEASM-DES (b) SAE-DDES (c) SAE-IDDES

Figure G.6: Comparison of the mean flow topology using an isosurface of U = −0.0001 to de-
pict reverse flow regions and stream ribbons seeded in the lower incoming bound-
ary layer.

(a) CEASM-DES (b) SAE-DDES (c) SAE-IDDES

Figure G.7: Contours of streamwise velocity (thicker lines, dashed lines denote negative val-
ues) and cross-flow streamlines (thinner lines, obtained from the tangential V and
W components only) at the plane x = 0 located at the bump foot.

224
 (a) Inflow, x = −0.367 (b) CEASM-DES, x = (c) SAE-DDES, x =
−0.2 −0.2

Figure G.8: Downstream development of the inflow plane corner vortices for CEASM-DES and
SAE-DDES at x = −0.2. Contours of instantaneous x-vorticity (dashed lines indi-
cate negative values).

(a) 1 − tanh( A2L − ν), CEASM-DES

(b) 1 − f d , SAE-DDES

(c) f˜d , SAE-IDDES

Figure G.9: Instantaneous contours of the GIS-shield functions for CEASM-DES and SAE-
DDES and the hybrid blending function for SAE-IDDES along the centre plane
z = 0.

225
G Figures for the bump test case

226
Bibliography
[1] F. Abernathy. Flow over an inclined plate. Transactions of the ASME. Series D, Journal of Basic Engineering,
61:380–388, 1962.
[2] Y. Addad, U. Gaitonde, D. Laurence, and S. Rolfo. Optimal unstructured meshing for large eddy
simulations. In Meyers, Guerts, and Sagaut, editors, Quality and Reliability of Large-Eddy Simulations,
volume 12 of ERCOFTAC Series, pages 93–103. Springer, 2008.
[3] B. Aupoix. Application of two-point closures to subgrid scale modelling for homogeneous, 3D turbulence.
Lecture Series 1989-03, Turbulent Shear Flows. von Karman Institute for Fluid Dynamics, 1989.
[4] B. Aupoix, P. Barricau, Y. Egorov, C. Geiler, A. Gilliot, F. Menter, J. Monnier, G. Pailhas, L. Perret,
M. Stanislas, and Y. Touvet. The DESider bump experiment. In Proceedings of the 7th International
Symposium on Engineering Turbulence Modelling and Measurements (ETMM7), pages 249–254, Limassol,
Cyprus, 2008.
[5] J. Baggett. On the feasibility of merging LES with RANS for the near-wall region of attached turbulent
flows. Annual Research Briefs, Center for Turbulence Research, Stanford University, 1998.
[6] B. Baldwin and H. Lomax. Thin-layer approximation and algebraic model for separated turbulent
flows. In 16th Aerospace Sciences Meeting, Huntsville, Alabama, USA. AIAA Paper 78-257.
[7] P. Bearman. On vortex shedding from a circular cylinder in the critical Reynolds number régime.
Journal of Fluid Mechanics, 37(3):577–585, 1969.
[8] P. Bearman and T. Morel. Effect of free stream turbulence on the flow around bluff bodies. Progress in
Aerospace Sciences, 20:97–123, 1983.
[9] S. Benhamadouche. Large eddy simulation with the unstructured collocated arrangement. PhD thesis, School
of Mechanical, Aerospace and Civil Engineering, University of Manchester, 2006.
[10] S. Benhamadouche and D. Laurence. Global kinetic energy conservation with unstructured meshes.
International Journal for Numerical Methods in Fluids, 40:561–571, 2002.
[11] M.S. Bloor. The transition to turbulence in the wake of a circular cylinder. Journal of Fluid Mechanics,
19:290–304, 1964.
[12] J Boris and D Book. Flux-corrected transport I. SHASTA, a fluid transport algorithm that works. Journal
of Computational Physics, 11:38–69, 1973.
[13] J. Boussinesq. Theorie de l’écoulement tourbillonnant et tumultueux des liquides dans les lits rectilignes a
grande section. Gauthier-Villars et fils, Paris, France, 1897.
[14] P. Bradshaw. An introduction to turbulence and its measurement. Pergamon Press, 1971.
[15] P. Bradshaw, D. Ferriss, and N. Atwell. Calculation of boundary-layer development using the turbulent
energy equation. Journal of Fluid Mechanics, 28:593–616, 1967.
[16] P. Bradshaw and G. Huang. The law of the wall in turbulent flow. Proceedings of the Royal Society of
London A, 451:165–188, 1995.
[17] M. Breuer. Direkte Numerische Simulation und Large-Eddy Simulation turbulenter Strömungen auf Hochleis-
tungsrechnern. Shaker Verlag, 2002.
[18] M. Breuer, N. Jovičić, and K. Mazaev. Comparison of DES, RANS and LES for the separated flow
around a flat plate at high incidence. International Journal for Numerical Methods in Fluids, 41:357–388,
2003.
[19] U. Bunge. Numerische Simulation turbulenter Strömungen im Kontext der Wechselwirkung zwischen Fluid
und Struktur. PhD thesis, Technische Universität Berlin, Universitätsbibliothek (Diss.-Stelle), 2004.
[20] U. Bunge, C. Mockett, and F. Thiele. Calibration of different models in the context of detached-eddy
simulation. In AG STAB Mitteilungen, Göttingen, 2003. DGLR.
[21] U. Bunge, C. Mockett, and F. Thiele. Guidelines for implementing detached-eddy simulation using
different models. Aerospace Science and Technology, 11:376–385, 2007.

227
Bibliography

[22] R. Bush and M. Mani. A two-equation large eddy stress model for high sub-grid shear. In Proceedings
of the 15th AIAA Computational Fluid Dynamics Conference, Anaheim, California, USA, June 11–14 2001.
[23] M. Casey and T. Wintergerste. Best Practice Guidelines. ERCOFTAC Special Interest Group on Quality
and Trust in Industrial CFD, 2000.
[24] I. Çelik, Z. Cehreli, and I. Yavuz. Index of resolution quality for large eddy simulations. Journal of
Fluids Engineering, 127(5):949–958, 2005.
[25] T. Cebeci and A. Smith. Analysis of turbulent boundary layers. Series in Applied Mathematics and
Mechanics, No. 15), 1974. 418 p, 1974.
[26] D. Chapman. Computational aerodynamics development and outlook. AIAA Journal, 17(12):1293–
1313, 1979.
[27] H. Choi and P. Moin. Effects of the computational time step on numerical solutions of turbulent flow.
Journal of Computational Physics, 113:1–4, 1994.
[28] S. Choi. Note on the use of momentum interpolation method for unsteady flows. Numerical Heat
Transfer, Part A, 36:545–550, 1999.
[29] C. Cichowlas, P. Bonaïti, F. Debbasch, and M. Brachet. Effective dissipation and turbulence in spectrally
truncated euler flows. Physical Review Letters, 95(264502), 2005.
[30] G. Comte-Bellot and S. Corrsin. Simple Eulerian time correlation of full- and narrow-band velocity
signals in grid generated, ”isotropic“ turbulence. Journal of Fluid Mechanics, 48:273–337, 1971.
[31] R. Courant, K. Friedrichs, and H. Lewy. Über die partiellen Differenzengleichungen der mathematis-
chen Physik. Mathematische Annalen, 100(1):32–74, 1928.
[32] L. Davidson. Large eddy simulation: a dynamic one-equation subgrid model for three-dimensional
recirculating flow. In Proceedings of the 11th International Symposium on Turbulent Shear Flow, volume 3,
pages 26.1–26.6, Grenoble, 1997.
[33] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions for flows with recirculation. In S.-H.
Peng and W. Haase, editors, Advances in hybrid RANS-LES modelling, Papers contributed to the 2007
Symposium of Hybrid RANS-LES Methods, Corfu, Greece, volume 97 of Notes on Numerical Fluid Mechanics
and Multidisciplinary Design, pages 55–66. Springer Verlag, 2008.
[34] R. Dean. Reynolds number dependence of skin friction and other bulk flow variables in two-
dimensional rectangular duct flow. Journal of Fluids Engineering, 100:215–223, 1978.
[35] J. Deardorff. A numerical study of three-dimensional turbulent channel flow at large Reynolds num-
bers. Journal of Fluid Mechanics, 41:453–480, 1970.
[36] S. Deck. Numerical simulation of transonic buffet over a supercritical airfoil. AIAA Journal, 43(7):1556–
1566, 2005.
[37] H. Djeridi, M. Braza, R. Perrin, G. Harran, E. Cid, and S. Cazin. Near-wake turbulence properties
around a circular cylinder at high Reynolds number. Flow, Turbulence and Combustion, 71(1):19–34,
2003.
[38] P. Durbin. Separated flow computations with the k − ε − v2 model. AIAA Journal, 33(4):659–664, 1995.
[39] J. Edwards and S. Chandra. Comparison of eddy viscosity-transport turbulence models for three-
dimensional, shock-separated flowfields. AIAA Journal, 34(4):756–763, 1996.
[40] T. Esch, F. Menter, and W. Vieser. Heat transfer predictions based on two-equation turbulence models.
In Proceedings of the 6th ASME-JSME Thermal Engineering Joint Conference, number TED-AJ03-542, 2003.
[41] A. Fage and F. Johansen. The structure of vortex streets. Philosophical Magazine, Series 7, 5(28):417–441,
1928.
[42] E. Fares and W. Schröder. A general one-equation turbulence model for free shear and wall-bounded
flows. Flow, Turbulence and Combustion, 73(3):187–215, 2005.
[43] F. Felten and T. Lund. Kinetic energy conservation issues associated with the collocated mesh scheme
for incompressible flow. Journal of Computational Physics, 215:465–484, 2006.
[44] J. Ferziger and M. Perić. Computational methods for fluid dynamics. Springer Verlag, 2002.

228
 Bibliography

[45] O. Flachsbart. Messungen an ebenen und gewölbten Platten. Ergebnisse der Aerodynamischen Versuch-
sanstalt zu Göttingen, 4:96–100, 1932.
[46] M. Franke. Untersuchung zum Potential höherwertiger Turbulenzmodelle für den aerodynamischen Entwurf.
PhD thesis, Technische Universität Berlin, 2003.
[47] J. Fröhlich and D. von Terzi. hybrid LES/RANS methods for the simulation of turbulent flows. Progress
in Aerospace Sciences, 44:349–377, 2008.
[48] C. Fureby. Large eddy simulation of rearward-facing step flow. AIAA Journal, 37(11):1401–1410, 1999.
[49] B. Ganapathisubramani, E. Longmire, and I. Marusic. Characteristics of vortex packets in turbulent
boundary layers. Journal of Fluid Mechanics, 478:35–46, 2003.
[50] A. Garbaruk, M. Shur, M. Strelets, and A. Travin. D3.2-12, Preliminary report on DES method in-
vestigation. Technical report, New Technologies and Services, 2004. Unpublished DESider project
deliverable.
[51] M. Germano, U. Piomelli, P. Moin, and W. Cabot. A dynamic subgrid-scale eddy viscosity model.
Physics of Fluids A, 3(7):1760–1765, 1991.
[52] S. Ghosal, T. Lund, P. Moin, and K. Akselvoll. A dynamic localization model for large-eddy simulation
of turbulent flows. Journal of Fluid Mechanics, 286:229–255, 1995.
[53] F. Grinstein, L. Margolin, and W. Rider, editors. Implicit large eddy simulation: computing turbulent fluid
dynamics. Cambridge University Press, 2007.
[54] D. Guenot. Simulation des effets instationnaires à grande échelle dans les écoulements décollés. PhD thesis,
SUPAERO, Toulouse, 2004.
[55] W. Haase, B. Aupoix, U. Bunge, and D. Schwamborn, editors. FLOMANIA - A European initiative on
flow physics modelling, volume 94 of Notes on Numerical Fluid Mechanics and Multidisciplinary Design.
Springer Verlag, 2006.
[56] W. Haase, M. Braza, and A. Revell, editors. DESider – A European effort on hybrid RANS-LES modelling,
volume 103 of Notes on Numerical Fluid Mechanics and Multidisciplinary Design. Springer Verlag, 2009.
[57] F. Hamba. An approach to hybrid rans/les calculation of channel flows. In W. Rodi and N. Fueyo,
editors, Engineering Turbulence Modelling and Experiments - 5, pages 297–305. Elsevier, 2002.
[58] K. Hanjalić, M. Hadžiabdić, L. Temmerman, and M. Leschziner. Merging LES and RANS strategies:
Zonal or seamless coupling? In R. Friedrich, B. Guerts, and O. Métais, editors, Direct and Large-Eddy
Simulation V, pages 451–464. Springer, 2004.
[59] A. Harten. High resolution schemes for hyperbolic conservation laws. Journal of Computational Physics,
49:357–393, 1983.
[60] S. Hickel and N. Adams. On implicit subgrid-scale modeling in wall-bounded flows. Physics of Fluids,
19(105106), 2007.
[61] J. Hinze. Turbulence. McGraw-Hill Book Company, 1975.
[62] S. Hoerner. Fluid-dynamic drag. Hoerner Fluid Dynamics, 1965.
[63] S. Hoyas and J. Jiménez. Scaling of the velocity fluctuations in turbulent channels up to Reτ = 2003.
Physics of Fluids, 18(011702), 2006.
[64] J. Humphreys. On a circular cylinder in a steady wind at transition Reynolds numbers. Journal of Fluid
Mechanics, 9:603–612, 1960.
[65] H. Jasak. Error analysis and estimation for the finite volume method with applications to fluid flows. PhD
thesis, Imperial College of Science, Technology and Medicine, London, 1996.
[66] J. Jeong and F. Hussain. On the identification of a vortex. Journal of Fluid Mechanics, 285:69–94, 1995.
[67] J. Jiménez and S. Hoyas. Turbulent fluctuations above the buffer layer of wall-bounded flows. Journal
of Fluid Mechanics, 2007. Submitted for publication.
[68] J. Jiménez and P. Moin. The minimal flow unit in near-wall turbulence. Journal of Fluid Mechanics,
225:213–240, 1991.
[69] J. Jiménez, A. Wray, P. Saffman, and R. Rogallo. The structure of intense vorticity in isotropic turbu-
lence. Journal of Fluid Mechanics, 255:65–90, 1993.

229
Bibliography

[70] D. Johnson and L. King. A mathematically simple turbulence closure model for attached and separated
turbulent boundary layers. AIAA Journal, 23(11):1684–1692, 1985.
[71] W. Jones and B. Launder. The prediction of laminarization with a two-equation model of turbulence.
International Journal of Heat and Mass Transfer, 15(2):301–314, 1972.
[72] K. Karki and S. Patankar. Pressure based calculation procedure for viscous flows at all speeds. AIAA
Journal, 27:1167–1174, 1989.
[73] P. Klebanoff. Characteristics of turbulence in a boundary layer with zero pressure gradient. Report
1247, NACA, 1955.
[74] T. Knopp. On grid-independence of RANS predictions for aerodynamic flows using model-consistent
universal wall-functions. In P. Wesseling, E. Oñate, and J. Périaux, editors, Proceedings of the European
Conference on Computational Fluid Dynamics ECCOMAS CFD 2006, Egmond aan Zee, The Netherlands,
2006.
[75] J. Kok. Resolving the dependence on freestream values for the k − ω turbulence model. AIAA Journal,
38(7):1292–1295, 2000.
[76] J. Kok, H. Dol, B. Oskam, and H. van der Ven. Extra-large eddy simulation of massively separated
flows. In Proceedings of the 42nd AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, USA, 5–8
January 2004.
[77] A. Kolmogorov. The local strcuture of turbulence in incompressible viscous fluid for very large
reynolds numbers. Dokl. Akad. Nauk SSSR, 30:299–303, 1941.
[78] R. Kraichnan. Inertial ranges in two-dimensional turbulence. Physics of Fluids, 10(7):1417–1423, 1967.
[79] H. Lamb. In an address to the British Association for the Advancement of Science, 1932.
[80] M. Leschziner and D. Drikakis. Turbulence modelling and turbulent-flow computation in aeronautics.
Aeronautical Journal, 106(1061):349–383, 2002.
[81] F. Lien and M. Leschziner. Computational modelling of 3D turbulent flow in S-diffuser and transition
ducts. In W. Rodi and F. Martelli, editors, Engineering Turbulence Modelling and Experiments 2, pages
217–228. Elsevier, 1993.
[82] D. Lilly. On the application of the eddy viscosity concept in the inertial sub-range of turbulence. NCAR
Manuscript 123, National Center for Atmospheric Research, Boulder, Colorado, USA, 1966.
[83] D. Lilly. A proposed modification of the germano subgrid-scale closure method. Physics of Fluids A,
4(3):633–635, 1992.
[84] H. Lübcke. Entwicklung expliziter Darstellungen zweiter statistischer Momente zur numerischen Simulation
turbulenter Strömungen. PhD thesis, Technische Universität Berlin, 2001.
[85] H. Lübcke, T. Rung, and F. Thiele. Prediction of the spreading mechanisms of 3D turbulent wall
jets with explicit Reynolds-stress closures. In W. Rodi and N. Fueyo, editors, Engineering Turbulence
Modelling and Experiments 5, pages 127–145. Elsevier, 2002.
[86] K. Mahesh, G. Constantinescu, and P. Moin. A numerical method for large-eddy simulation in complex
geometries. Journal of Computational Physics, 197:215–240, 2004.
[87] B. Massey and J. Ward-Smith. Mechanics of Fluids. Stanley Thornes (Publishers) Ltd, 7th edition, 1998.
[88] F. Menter. Performance of popular turbulence models for attached and separated adverse pressure
gradient flows. AIAA Journal, 30(8):2066–2072, 1992.
[89] F. Menter. Eddy viscosity transport equations and their relation to the k − ε model. Technical Memo-
randum 108854, NASA Ames Research Center, 1994.
[90] F. Menter. Two-equation eddy-viscosity turbulence models for engineering applications. AIAA Journal,
32(8):1598–1605, 1994.
√
[91] F. Menter, Y. Egorov, and D. Rusch. Steady and unsteady flow modelling using the k − kl model. In
K. Hanjalić, Y. Nagano, and S. Jakirlic, editors, 5th International Symposium on Turbulence, Heat and Mass
Transfer, pages 403–406, Dubrovnik, Croatia, 2006.
[92] F. Menter and M. Kuntz. The Aerodynamics of Heavy Vehicles: Trucks, Buses, and Trains, volume 19
of Lecture notes in applied and computational mechanics, chapter Adaption of eddy-viscosity turbulence
models to unsteady separated flow behind vehicles. Springer Verlag, 2004.

230
 Bibliography

[93] U. Michel, D. Eschricht, B. Greschner, T. Knacke, C. Mockett, Ĺ. Panek, F. Thiele, and J. Yan. Simulation
of the sound radiation of turbulent flows with des. In Proceedings of the West-East High Speed Flow Field
Conference, Moscow, Russia, 2007.
[94] C. Mockett, U. Bunge, and F. Thiele. Turbulence modelling in application to the vortex shedding of
stalled airfoils. In W. Rodi and M. Mulas, editors, Engineering Turbulence Modelling and Experiments 6,
pages 617–626. Elsevier, 2005.
[95] C. Mockett, O. Frederich, T. Schmidt, and F. Thiele. Numerical prediction of the aerodynamic interfer-
ence of twin-sting model supports on empennage measurements. In Proceedings of the 7th International
Symposium on Engineering Turbulence Modelling and Measurements (ETMM7), pages 811–816, Limassol,
Cyprus, 2008.
[96] C. Mockett, B. Greschner, T. Knacke, R. Perrin, J. Yan, and F. Thiele. Demonstration of improved
DES methods for generic and industrial applications. In S.-H. Peng and W. Haase, editors, Advances
in hybrid RANS-LES modelling, Papers contributed to the 2007 Symposium of Hybrid RANS-LES Methods,
Corfu, Greece, volume 97 of Notes on Numerical Fluid Mechanics and Multidisciplinary Design, pages 222–
231. Springer Verlag, 2008.
[97] C. Mockett, R. Perrin, T. Reimann, M. Braza, and F. Thiele. Analysis of detached-eddy simulation for
the flow around a circular cylinder with reference to PIV data. In Proceedings of the IUTAM Symposium
on Unsteady Separated Flows and their Control, Corfu, Greece, June 18–22 2007.
[98] C. Mockett, R. Perrin, T. Reimann, M. Braza, and F. Thiele. Analysis of detached-eddy simulation for
the flow around a circular cylinder with reference to PIV data. Flow, Turbulence and Combustion, 2009.
To be published.
[99] C. Mockett and F. Thiele. Overview of detached-eddy simulation for external and internal turbulent
flow applications. In F. G. Zhuang and J. C. Li, editors, New Trends in Fluid Mechanics Research, Proceed-
ings of the 5th International Conference on Fluid Mechanics, pages 79–82, Shanghai, China, 2007. Tsinghua
University Press & Springer.
[100] R. Moser, J. Kim, and N. Mansour. Direct numerical simulation of turbulent channel flow up to Reτ =
590. Physics of Fluids, 11(4):943–945, April 1999.
[101] N. Nikitin, F. Nicoud, B. Wasistho, K. Squires, and P. Spalart. An approach to wall modeling in large-
eddy simulations. Physics of Fluids, 12(7):1629–1632, 2000.
[102] C. Norberg. Effects of Reynolds number and a low-intensity freestream turbulence on the flow around
a circular cylinder. Technical Report Publikation Nr 87/2, Department of Applied Thermodynamics
and Fluid Mechanics, Chalmers University of Technology, Göteborg, Sweden, 1987.
[103] C. Norberg. Fluctuating lift on a circular cylinder: review and new measurements. Journal of Fluids and
Structures, 17:57–96, 2003.
[104] H. Nyquist. Certain topics in telegraph transmission theory. Transactions of the American Institute of
Electrical Engineers, 47:617–644, April 1928.
[105] S. Obi, M. Perić, and G. Scheuerer. Second-moment calculation procedure for turbulent flows with
collocated variable arrangement. AIAA Journal, 29(4):585–590, 1991.
[106] S. Patankar. Numerical heat transfer and fluid flow. McGraw-Hill, New York, 1980.
[107] M. Perić. A finite volume method for the prediction of three-dimensional fluid flow in complex ducts. PhD
thesis, Imperial College of Science, Technology and Medicine, London, 1985.
[108] R. Perrin. Analyse physique et modélisation d’ecoulements incompressibles instationnaire turbulents autor d’un
cylindre circulaire à grand nombre de Reynolds. PhD thesis, Institut National Polytechnique de Toulouse,
5th July 2005.
[109] R. Perrin, M. Braza, E. Cid, S. Cazin, A. Barthet, A. Sevrain, C. Mockett, and F. Thiele. Phase averaged
turbulence properties in the near wake of a circular cylinder at high Reynolds number using POD.
In Proceedings of the 13th International Symposium on Applications of Laser Techniques to Fluid Mechanics,
Lisbon, Portugal, 2006.
[110] R. Perrin, M. Braza, E. Cid, S. Cazin, A. Barthet, A. Sevrain, C. Mockett, and F. Thiele. Obtaining phase
averaged turbulence properties in the near wake of a circular cylinder at high Reynolds number using
POD. Experiments in Fluids; Springer Verlag, 43(2-3):341–355, 2007.

231
Bibliography

[111] R. Perrin, M. Braza, E. Cid, S. Cazin, P. Chassaing, C. Mockett, T. Reimann, and F. Thiele. Coherent and
turbulent process analysis in the flow past a circular cylinder at high Reynolds number. In Proceedings
of the IUTAM Symposium on Unsteady Separated Flows and their Control, Corfu, Greece, June 18–22 2007.
[112] R. Perrin, C. Mockett, M. Braza, E. Cid, S. Cazin, A. Sevrain, P. Chassaing, and F. Thiele. Joint numeri-
cal and experimental investigation of the flow around a circular cylinder at high Reynolds number. In
A. Schröder and C. E. Willert, editors, Particle Image Velocimetry – New Developments and Recent Applica-
tions, volume 112 of Topics in Applied Physics. Springer Verlag, 2008.
[113] A. Perry and T. Lim. Coherent structures in coflowing jets and wakes. Journal of Fluid Mechanics,
88:451–463, 1978.
[114] U. Piomelli, E. Balaras, H. Pasinato, K. Squires, and P. Spalart. The inner–outer layer interface in large-
eddy simulations with wall-layer models. International Journal of Heat and Fluid Flow, 24:538–550, 2003.
[115] S. Pope. A more general effective-viscosity hypothesis. Journal of Fluid Mechanics, 72:331–340, 1975.
[116] S. Pope. Turbulent Flows. Cambridge University Press, 2000.
[117] L. Prandtl. Bericht über Untersuchungen zur ausgebildeten Turbulenz. Zeitschrift für angewandte Math-
ematik und Mechanik, 5:136–139, 1925.
[118] H. Reichardt. Vollständige Darstellung der turbulenten Geschwindigkeitsverteilung in glatten Leitun-
gen. Zeitschrift für angewandte Mathematik und Mechanik, 31:208–219, 1951.
[119] O. Reynolds. An experimental investigation of the circumstances which determine whether the motion
of water shall be direct or sinuous, and of the law of resistance in parallel channels. Philosophical
Transactions of the Royal Society of London, 174:935–982, 1883.
[120] O. Reynolds. On the dynamical theory of incompressible viscous flows and the determination of the
criterion. Philosophical Tranactions of the Royal Society of London A, 186:123–161, 1894.
[121] W. Reynolds. The potential and limitations of direct and large eddy simulations. In J. Lumley, editor,
Whither Turbulence? Turbulence at the Crossroads: Proceedings of a Workshop Held at Cornell University,
volume 357 of Lecture Notes in Physics, pages 313–343, Ithaca, NY, USA, 1990. Springer.
[122] W. Reynolds and A. Hussain. The mechanics of an organized wave in turbulent shear flow. Part 3:
Theoretical models and comparisons with experiments. Journal of Fluid Mechanics, 54:263–288, 1972.
[123] L. Richardson. Weather Prediction by Numerical Process. Cambridge University Press, 1922.
[124] A. Richter and E. Naudascher. Fluctuating forces on a rigid circular cylinder in confined flow. Journal
of Fluid Mechanics, 78(3):561–576, 1976.
[125] J. Riou, E. Garnier, S. Deck, and C. Basdevant. An extension of DDES applied to the computation of a
generic missile fin in transonic regime. In Proceedings of the 7th International Symposium on Engineering
Turbulence Modelling and Measurements (ETMM7), pages 732–737, Limassol, Cyprus, 2008.
[126] J. Riou, E. Garnier, S. Deck, and C. Basdevant. Improvement of delayed-detached eddy simulation
applied to separated flow over missile fin. AIAA Journal, 47(2):345–360, 2009.
[127] W. Rodi. Experience with two-layer models combining the k − ε model with a one-equation model
near the wall. In 29th Aerospace Sciences Meeting, Reno, Nevada, USA, Jan 7th-10th 1991. AIAA Paper
1991-216.
[128] R. Rogallo and P. Moin. Numerical simulation of turbulent flows. Annual Review of Fluid Mechanics,
16:99–137, 1984.
[129] A. Roshko. On the wake and drag of bluff bodies. Journal of the Aeronautical Sciences, 22:124–132, 1955.
[130] J. Rotta. Über eine Methode zur Berechnung turbulenter Strömung. Mitteilungen der Aerodynamischen
Versuchsanstalt Göttingen, Rep 69 A 14, 1968.
[131] T. Rung. Erweiterung von Eingleichungs-Turbulenzmodellen für lokales Nichtgleichgewicht. Internal
report 3-98, Technische Universität Berlin, 1998.
[132] T. Rung. Entwicklung anisotroper Wirbelzähigkeitsbeziehungen mit Hilfe von Projektionstechniken. PhD
thesis, Technische Universität Berlin, 2000.
[133] T. Rung. Statistische Turbulenzmodellierung. Notes accompanying the lecture course “statistische
Turbulenzmodellierung” at the Institute of Fluid Mechanics and Engineering Acoustics, Technische
Universität Berlin, http://www.cfd.tu-berlin.de, 2004.

232
 Bibliography

[134] T. Rung, U. Bunge, M. Schatz, and F. Thiele. Restatement of the Spalart–Allmaras eddy-viscosity model
in strain-adaptive formulation. AIAA Journal, 41(7):1396–1399, 2003.
[135] T. Rung, H. Lübcke, and F. Thiele. Universal wall-boundary conditions for turbulence-transport mod-
els. Zeitschrift für angewandte Mathematik und Mechanik, 81(1):1756–1758, 2000.
[136] T. Rung and F. Thiele. Computational modelling of complex boundary-layer flows. In Proceedings of
the 9th International Symposium on Transport Phenomena in Thermal-Fluid Engineering, Singapore, 1996.
[137] T. Rung, L. Xue, J. Yan, M. Schatz, and F. Thiele. Numerische Methoden der Thermo- und Fluid-
dynamik. Notes accompanying the lecture course “CFD1” at the Institute of Fluid Mechanics and
Engineering Acoustics, Technische Universität Berlin, http://www.cfd.tu-berlin.de.
[138] P. Sagaut. Large eddy simulation for incompressible flows, an introduction. Springer, 2nd edition, 2004.
[139] P. Sagaut, S. Deck, and M. Terracol. Multiscale and multiresolution approaches in turbulence. Imperial
College Press, London, 2006.
[140] M. Schatz. Konvektionschemata in FAN/ELAN2/ELAN3. Internal report, Technische Universität
Berlin, 2003.
[141] M. Schatz. Numerische Simulation der Beeinflussung instationärer Strömungsablösung durch frei bewegliche
Rückstromklappen auf Tragflügeln. PhD thesis, Technische Universität Berlin, 2003.
[142] G. Schewe. On the force fluctuations acting on a circular cylinder in crossflow from subcritical up to
transcritical Reynolds numbers. Journal of Fluid Mechanics, 133:265–285, 1983.
[143] H. Schlichting. Boundary-Layer Theory. McGraw-Hill, 7th edition, 1979.
[144] S. Schmidt. Grobstruktursimulation turbulenter Strömungen in komplexen Geometrien und bei hohen Reynold-
szahlen. PhD thesis, Technische Universität Berlin, 2000.
[145] T. Schmidt, C. Mockett, and F. Thiele. Adaptive wall function for the prediction of turbulent flows. In
N. Kroll, D. Schwamborn, K. Becker, H. Rieger, and F. Thiele, editors, MEGADESIGN and MegaOpt -
Aerodynamic Simulation and Optimization in Aircraft Design, Notes on Numerical Fluid Mechanics and
Multidisciplinary Design. Springer Verlag, 2008. To be published.
[146] M. Shur, P. Spalart, K. Squires, M. Strelets, and A. Travin. Three dimensionality in Reynolds-averaged
Navier–Stokes solutions around two-dimensional geometries. AIAA Journal, 43(6):1230–1242, 2005.
[147] M. Shur, P. Spalart, M. Strelets, and A. Travin. Navier–Stokes simulation of shedding turbulent flow
past a circular cylinder and a cylinder with a backward splitter plate. In G. Desideri, C. Hirsch, P. Le
Tallec, M. Pandolfi, and J. Périaux, editors, Proceedings of the Third European Conference on Computa-
tional Fluid Dynamics ECCOMAS CFD 1996, pages 676–682, Paris, France, 1996. John Wiley & Sons,
Chichester.
[148] M. Shur, P. Spalart, M. Strelets, and A. Travin. Detached-eddy simulation of an airfoil at high angle of
attack. In W. Rodi and D. Laurence, editors, Engineering Turbulence Modelling and Experiments 4, pages
669–678. Elsevier, 1999.
[149] M. Shur, P. Spalart, M. Strelets, and A. Travin. A hybrid RANS-LES approach with delayed DES and
wall-modeled LES capabilities. International Journal of Heat and Fluid Flow, 29:1638–1649, 2008.
[150] J. Smagorinsky. General circulation experiments with the primitive equations: I. the basic experiment.
Monthly Weather Review, 91:99–164, 1963.
[151] A. Sohankar, L. Davidson, and C. Norberg. Large eddy simulation of flow past a square cylinder:
comparison of different subgrid scale models. Journal of Fluids Engineering, 122:39–47, 2000.
[152] A. Sohankar, C. Norberg, and L. Davidson. Low–Reynolds–number flow around a square cylinder at
incidence: Study of blockage, onset of vortex shedding and outlet boundary condition. International
Journal for Numerical Methods in Fluids, 26, 1998.
[153] P. Spalart. Strategies for turbulence modelling and simulations. International Journal of Heat and Fluid
Flow, 21(3):252–263, 2000.
[154] P. Spalart. Young person’s guide to detached-eddy simulation grids. NASA contractor report
NASA/CR-2001-211032, 2001.
[155] P. Spalart. The uses of DES: natural, extended, and improper. Invited presentation
at the DESider Hybrid RANS-LES Symposium, Stockholm, Sweden, 14th-15th July 2005.
http://cfd.mace.manchester.ac.uk/desider/index2.html.

233
Bibliography

[156] P. Spalart. Detached-eddy simulation. Annual Review of Fluid Mechanics, 41:181–202, 2009.
[157] P. Spalart and S. Allmaras. A one-equation turbulence model for aerodynamic flows. In Proceedings of
the 30th AIAA Aerospace Sciences Meeting and Exhibit, Reno, Nevada, USA, 1992.
[158] P. Spalart, S. Deck, M. Shur, K. Squires, M. Strelets, and A. Travin. A new version of detached-eddy
simulation, resistant to ambiguous grid densities. Theoretical and Computational Fluid Dynamics, 20:181–
195, 2006.
[159] P. Spalart, W. Jou, M. Strelets, and S. Allmaras. Comments on the feasibility of LES for wings, and on
a hybrid RANS/LES approach. Advances in DNS/LES, 1, 1997.
[160] P. Spalart and C. Rumsey. Effective inflow conditions for turbulence models in aerodynamic calcula-
tions. AIAA Journal, 45(10):2544–2553, 2007.
[161] D. Spalding. A single formula for the law of the wall. Journal of Applied Mechanics, 28(3):444–458, 1961.
[162] H. Stone. Iterative solution of implicit approximations of multidimensional partial differential equa-
tions. SIAM Journal on Numerical Analysis, 5:530–558, 1968.
[163] M. Strelets. Detached eddy simulation of massively separated flows. In Proceedings of the 39th AIAA
Aerospace Sciences Meeting and Exhibit, Reno, Nevada, USA, 2001.
[164] M. Strelets and M. Shur. Private communication.
[165] V. Strouhal. Über eine besondere Art der Tonerregung. Annalen der Physik und Chemie, Neue Folge,
5(10):216–251, 1878.
[166] K. Swalwell. The effect of turbulence on stall of horizontal axis wind turbines. PhD thesis, Dept. of Mechan-
ical Engineering, Monash University, 2005.
[167] K. Swalwell, J. Sheridan, and W. Melbourne. Frequency analysis of surface pressure on an airfoil after
stall. In Proceedings of the 21st AIAA Applied Aerodynamics Conference, 2004.
[168] S. Szepessy and P Bearman. Aspect ratio and end plate effects on vortex shedding from a circular
cylinder. Journal of Fluid Mechanics, 234:191–217, 1992.
[169] H. Tennekes and J.L. Lumley. A First Course in Turbulence. MIT Press, 1972.
[170] A. Travin, M. Shur, P. Spalart, and M. Strelets. Improvement of delayed detached-eddy simulation for
LES with wall modelling. In P. Wesseling, E. Oñate, and J. Périaux, editors, Proceedings of the European
Conference on Computational Fluid Dynamics ECCOMAS CFD 2006, Egmond aan Zee, The Netherlands,
2006.
[171] A. Travin, M. Shur, M. Strelets, and P. Spalart. Detached-eddy simulations past a circular cylinder.
Flow, Turbulence and Combustion, 63(1):293–313, 2000.
[172] A. Travin, M. Shur, M. Strelets, and P. Spalart. Physical and numerical upgrades in the detached-eddy
simulation of complex turbulent flows. In Proceedings of the 412th Euromech Colloquium on LES and
Complex Transitional and Turbulent Flows, Munich, Germany, 2000.
[173] E. van Driest. On turbulent flow near a wall. Journal of the Aeronautical Sciences, 23(11):1007–1011, 1956.
[174] M. Vanella, U. Piomelli, and E. Balaras. Effect of local grid refinement on large-eddy simulation statis-
tics and flow-fields. In Proceedings of the 7th International Symposium on Engineering Turbulence Modelling
and Measurements (ETMM7), pages 287–296, Limassol, Cyprus, 2008.

[175] T. von Kármán. Mechanische Ähnlichkeit und Turbulenz. In Proceedings of the Third International
Congress on Applied Mechanics, pages 85–105, Stockholm, Sweden, 1930.
[176] T. von Kármán and H. Rubach. über den Mechanismus des flüssigkeits- und luftwiderstandes.
Physikalische Zeitschrift, 13(2):49–59, 1912.

[177] S. Šarić, S. Jakirlić, A. Djugum, and C. Tropea. Computational analysis of locally forced flow over a
wall-mounted hump at high-re number. International Journal of Heat and Fluid Flow, 27(4):707–720, 2006.
[178] S. Wallin and A. Johansson. An explicit algebraic Reynolds stress model for incompressible and com-
pressible turbulent flows. Journal of Fluid Mechanics, 2000.

234
 Bibliography

[179] K. Weinman, H. van der Ven, C. Mockett, T. Knopp, J. Kok, R. Perrin, and F. Thiele. A study of grid
convergence issues for the simulation of the massively separated flow around a stalled airfoil using
DES and related methods. In P. Wesseling, E. Oñate, and J. Périaux, editors, Proceedings of the European
Conference on Computational Fluid Dynamics ECCOMAS CFD 2006, Egmond aan Zee, The Netherlands,
2006.
[180] P. Welch. The use of fast Fourier transform for the estimation of power spectra: a method based on
time averaging over short, modified periodograms. IEEE Transactions on Audio and Electroacoustics,
15(2):70–73, June 1976.
[181] D. Wilcox. Reassessment of the scale-determining equation for advanced turbulence models. AIAA
Journal, 26:1299–1310, 1988.
[182] D. Wilcox. Turbulence modeling for CFD. DCW Industries Inc, 1993.
[183] A. Wray. Unpublished DNS data. Available on AGARD database “Test Cases for the Validation of Large-
Eddy Simulations of Turbulent Flows”, ftp://torroja.dmt.upm.es, 1997.
[184] L. Xue. Entwicklung eines effizienten parallelen Lösungsalgorithmus zur dreidimensionalen Simulation kom-
plexer turbulenter Strömungen. PhD thesis, Technische Universität Berlin, Universitätsbibliothek (Diss.-
Stelle), 1998.
[185] J. Yan. Effiziente Simulation komplexer Strömungen auf semi-strukturierten Gittern. PhD thesis, Technische
Universität Berlin, 2003.
[186] J. Yan, C. Mockett, and F. Thiele. Investigation of alternative length scale substitutions in detached-
eddy simulation. Flow, Turbulence and Combustion, 74(1):85–102, 2005.
[187] J. Yan, K. Tawackolian, U. Michel, and F. Thiele. Computation of jet noise using a hybrid approach. In
Proceedings of the 13th AIAA/CEAS Aeroacoustics Conference (28th AIAA Aeroacoustics Conference), Rome,
Italy, 2007. AIAA Paper 2007-3621.
[188] A. Yoshizawa and K. Horiuti. A statistically-derived subgrid-scale kinetic energy model for the large-
eddy simulation of turbulent flows. Journal of the Physical Society of Japan, 54(8):2834–2839, 1985.
[189] M. Zdravkovich. Flow around circular cylinders, volume 1: Fundamentals. Oxford University Press,
1997.

235