Fluid mechanics, turbulent flow and turbulence

modeling
Lars Davidson
Division of Fluid Dynamics
Department of Applied Mechanics
Chalmers University of Technology
SE-412 96 Göteborg, Sweden
http://www.tfd.chalmers.se/˜lada
[email protected]
June 28, 2017

Abstract
This course material is used in two courses in the International Master’s pro-
gramme Applied Mechanics at Chalmers. The two courses are TME225 Mechanics of fluids
(Chapters 1-10), and MTF270 Turbulence Modeling (Chapters 11-27). MSc stu-
dents who follow these courses are supposed to have taken one basic course in
fluid mechanics.
The lecture notes are also used in the three-day course
Unsteady Simulations for Industrial Flows: LES, DES, hybrid LES-RANS and URANS
(Chapters 16-27).

This document can be downloaded at

http://www.tfd.chalmers.se/˜lada/MoF/lecture notes.html

and

http://www.tfd.chalmers.se/˜lada/comp turb model/lecture notes.html

The Fluid courses in the MSc programme are presented at

http://www.tfd.chalmers.se/˜lada/msc/msc-programme.html

The MSc programme is presented at

http://www.chalmers.se/en/education/programmes/masters-info/Pages/Applied-Mechanics.aspx

1
Contents
1 Motion, flow 11
1.1 Eulerian, Lagrangian, material derivative . . . . . . . . . . . . . . . 11
dv2 ∂v2
1.2 What is the difference between and ? . . . . . . . . . . . . 12
dt ∂t
1.3 Viscous stress, pressure . . . . . . . . . . . . . . . . . . . . . . . . 13
1.4 Strain rate tensor, vorticity . . . . . . . . . . . . . . . . . . . . . . . 14
1.5 Product of a symmetric and antisymmetric tensor . . . . . . . . . . 16
1.6 Deformation, rotation . . . . . . . . . . . . . . . . . . . . . . . . . 17
1.7 Irrotational and rotational flow . . . . . . . . . . . . . . . . . . . . 19
1.7.1 Ideal vortex line . . . . . . . . . . . . . . . . . . . . . . . . 20
1.7.2 Shear flow . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8 Eigenvalues and and eigenvectors: physical interpretation . . . . . . 22

2 Governing flow equations 24

2.1 The Navier-Stokes equation . . . . . . . . . . . . . . . . . . . . . . 24
2.1.1 The continuity equation . . . . . . . . . . . . . . . . . . . . 24
2.1.2 The momentum equation . . . . . . . . . . . . . . . . . . . 24
2.2 The energy equation . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.3 Transformation of energy . . . . . . . . . . . . . . . . . . . . . . . 26
2.4 Left side of the transport equations . . . . . . . . . . . . . . . . . . 27
2.5 Material particle vs. control volume (Reynolds Transport Theorem) . 28

3 Solutions to the Navier-Stokes equation: three examples 30

3.1 The Rayleigh problem . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.2 Flow between two plates . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.1 Curved plates . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.2.2 Flat plates . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.2.3 Force balance . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.4 Balance equation for the kinetic energy . . . . . . . . . . . . 37
3.3 Two-dimensional boundary layer flow over flat plate . . . . . . . . . 38

4 Vorticity equation and potential flow 42

4.1 Vorticity and rotation . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.2 The vorticity transport equation in three dimensions . . . . . . . . . 43
4.3 The vorticity transport equation in two dimensions . . . . . . . . . . 47
4.3.1 Boundary layer thickness from the Rayleigh problem . . . . 47
4.4 Potential flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.4.1 The Bernoulli equation . . . . . . . . . . . . . . . . . . . . 49
4.4.2 Complex variables for potential solutions of plane flows . . . 50
4.4.3 Analytical solutions as f ∝ z n . . . . . . . . . . . . . . . . 51
4.4.3.1 Parallel flow . . . . . . . . . . . . . . . . . . . . . 52
4.4.3.2 Stagnation flow . . . . . . . . . . . . . . . . . . . . 52
4.4.3.3 Flow over a wedge and flow in a concave corner . . 53
4.4.4 Analytical solutions for a line source . . . . . . . . . . . . . 54
4.4.5 Analytical solutions for a vortex line . . . . . . . . . . . . . 55
4.4.6 Analytical solutions for flow around a cylinder . . . . . . . . 56
4.4.7 Analytical solutions for flow around a cylinder with circulation 59
4.4.7.1 The Magnus effect . . . . . . . . . . . . . . . . . . 61
4.4.8 The flow around an airfoil . . . . . . . . . . . . . . . . . . . 63

2
 3

5 Turbulence 66
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5.2 Turbulent scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 Energy spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
5.4 The cascade process created by vorticity . . . . . . . . . . . . . . . 72

6 Turbulent mean flow 76

6.1 Time averaged Navier-Stokes . . . . . . . . . . . . . . . . . . . . . 76
6.1.1 Boundary-layer approximation . . . . . . . . . . . . . . . . 78
6.2 Wall region in fully developed channel flow . . . . . . . . . . . . . 78
6.3 Reynolds stresses in fully developed channel flow . . . . . . . . . . 83
6.4 Boundary layer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

7 Probability density functions 87

8 Transport equations for turbulent kinetic energy 90

8.1 Rules for time averaging . . . . . . . . . . . . . . . . . . . . . . . . 90
8.1.1 What is the difference between v1′ v2′ and v1′ v2′ ? . . . . . . . 90
2
8.1.2 What is the difference between v1′2 and v1′ ? . . . . . . . . . 91
8.1.3 Show that v̄1 v1′2 = v̄1 v1′2 . . . . . . . . . . . . . . . . . . . 91
8.1.4 Show that v̄ 1 = v̄1 . . . . . . . . . . . . . . . . . . . . . . . 92
8.2 The Exact k Equation . . . . . . . . . . . . . . . . . . . . . . . . . 92
8.2.1 Spectral transfer dissipation εκ vs. “true” viscous dissipation, ε 96
8.3 The Exact k Equation: 2D Boundary Layers . . . . . . . . . . . . . 96
8.4 Spatial vs. spectral energy transfer . . . . . . . . . . . . . . . . . . 98
8.5 The overall effect of the transport terms . . . . . . . . . . . . . . . . 99
8.6 The transport equation for v̄i v̄i /2 . . . . . . . . . . . . . . . . . . . 99

9 Transport equations for Reynolds stresses 102

9.1 Source terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
9.2 Reynolds shear stress vs. the velocity gradient . . . . . . . . . . . . 107

10 Correlations 110
10.1 Two-point correlations . . . . . . . . . . . . . . . . . . . . . . . . . 110
10.2 Auto correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
10.3 Taylor’s hypothesis of frozen turbulence . . . . . . . . . . . . . . . 113

11 Reynolds stress models and two-equation models 116

11.1 Mean flow equations . . . . . . . . . . . . . . . . . . . . . . . . . . 116
11.1.1 Flow equations . . . . . . . . . . . . . . . . . . . . . . . . 116
11.1.2 Temperature equation . . . . . . . . . . . . . . . . . . . . . 117
11.2 The exact vi′ vj′ equation . . . . . . . . . . . . . . . . . . . . . . . . 117
11.3 The exact vi′ θ′ equation . . . . . . . . . . . . . . . . . . . . . . . . 119
11.4 The k equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
11.5 The ε equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
11.6 The Boussinesq assumption . . . . . . . . . . . . . . . . . . . . . . 123
11.7 Modeling assumptions . . . . . . . . . . . . . . . . . . . . . . . . . 124
11.7.1 Production terms . . . . . . . . . . . . . . . . . . . . . . . 124
11.7.2 Diffusion terms . . . . . . . . . . . . . . . . . . . . . . . . 125
11.7.3 Dissipation term, εij . . . . . . . . . . . . . . . . . . . . . 126
 4

11.7.4 Slow pressure-strain term . . . . . . . . . . . . . . . . . . . 127

11.7.5 Rapid pressure-strain term . . . . . . . . . . . . . . . . . . 130
11.7.6 Wall model of the pressure-strain term . . . . . . . . . . . . 135
11.8 The k − ε model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
11.9 The modeled vi′ vj′ equation with IP model . . . . . . . . . . . . . . 138
11.10 Algebraic Reynolds Stress Model (ASM) . . . . . . . . . . . . . . . 138
11.11 Explicit ASM (EASM or EARSM) . . . . . . . . . . . . . . . . . . 139
11.12 Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 140
11.12.1 The linear assumption of aij . . . . . . . . . . . . . . . . . 145
11.13 Boundary layer flow . . . . . . . . . . . . . . . . . . . . . . . . . . 145

12 Reynolds stress models vs. eddy-viscosity models 147

12.1 Stable and unstable stratification . . . . . . . . . . . . . . . . . . . 147
12.2 Curvature effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
12.3 Stagnation flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151
12.4 RSM/ASM versus k − ε models . . . . . . . . . . . . . . . . . . . 152

13 Realizability 153
13.1 Two-component limit . . . . . . . . . . . . . . . . . . . . . . . . . 154

14 Non-linear Eddy-viscosity Models 156

15 The V2F Model 159

15.1 Modified V2F model . . . . . . . . . . . . . . . . . . . . . . . . . . 162
15.2 Realizable V2F model . . . . . . . . . . . . . . . . . . . . . . . . . 163
15.3 To ensure that v 2 ≤ 2k/3 . . . . . . . . . . . . . . . . . . . . . . . 163

16 The SST Model 164

17 Overview of RANS models 169

18 Large Eddy Simulations 170

18.1 Time averaging and filtering . . . . . . . . . . . . . . . . . . . . . . 170
18.2 Differences between time-averaging (RANS) and space filtering (LES) 171
18.3 Resolved & SGS scales . . . . . . . . . . . . . . . . . . . . . . . . 172
18.4 The box-filter and the cut-off filter . . . . . . . . . . . . . . . . . . 173
18.5 Highest resolved wavenumbers . . . . . . . . . . . . . . . . . . . . 174
18.6 Subgrid model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
18.7 Smagorinsky model vs. mixing-length model . . . . . . . . . . . . . 176
18.8 Energy path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
18.9 SGS kinetic energy . . . . . . . . . . . . . . . . . . . . . . . . . . 177
18.10 LES vs. RANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
18.11 The dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . 178
18.12 The test filter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
18.12.1 2D filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 179
18.12.2 3D filtering . . . . . . . . . . . . . . . . . . . . . . . . . . 180
18.13 Stresses on grid, test and intermediate level . . . . . . . . . . . . . . 181
18.14 Numerical dissipation . . . . . . . . . . . . . . . . . . . . . . . . . 182
18.15 Scale-similarity Models . . . . . . . . . . . . . . . . . . . . . . . . 183
18.16 The Bardina Model . . . . . . . . . . . . . . . . . . . . . . . . . . 184
18.17 Redefined terms in the Bardina Model . . . . . . . . . . . . . . . . 184
 5

18.18 A dissipative scale-similarity model. . . . . . . . . . . . . . . . . . 184

18.19 Forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
18.20 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
18.20.1 RANS vs. LES . . . . . . . . . . . . . . . . . . . . . . . . 187
18.21 One-equation ksgs model . . . . . . . . . . . . . . . . . . . . . . . 189
18.22 Smagorinsky model derived from the ksgs equation . . . . . . . . . 189
18.23 A dynamic one-equation model . . . . . . . . . . . . . . . . . . . . 189
18.24 A Mixed Model Based on a One-Eq. Model . . . . . . . . . . . . . 191
18.25 Applied LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
18.26 Resolution requirements . . . . . . . . . . . . . . . . . . . . . . . . 191

19 Unsteady RANS 193

19.1 Turbulence Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 196
19.2 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196

20 DES 198
20.1 DES based on two-equation models . . . . . . . . . . . . . . . . . . 199
20.2 DES based on the k − ω SST model . . . . . . . . . . . . . . . . . 200
20.3 DDES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201

21 Hybrid LES-RANS 202

21.1 Momentum equations in hybrid LES-RANS . . . . . . . . . . . . . 204
21.2 The one-equation hybrid LES-RANS model . . . . . . . . . . . . . 204

22 The SAS model 205

22.1 Resolved motions in unsteady . . . . . . . . . . . . . . . . . . . . 205
22.2 The von Kármán length scale . . . . . . . . . . . . . . . . . . . . . 205
22.3 The second derivative of the velocity . . . . . . . . . . . . . . . . . 207
22.4 Evaluation of the von Kármán length scale in channel flow . . . . . 207

23 The PANS Model 210

23.1 PANS as a hybrid LES-RANS model . . . . . . . . . . . . . . . . . 213
23.1.1 The interface conditions at the RANS-LES interface . . . . . 213
23.2 Zonal PANS: different treatments of the RANS-LES interface . . . . 215
23.2.1 The Interface Condition . . . . . . . . . . . . . . . . . . . . 215
23.2.2 Modeling the Interface . . . . . . . . . . . . . . . . . . . . 217
23.2.2.1 Interface Model 1 . . . . . . . . . . . . . . . . . . 217
23.2.2.2 Interface Model 2 . . . . . . . . . . . . . . . . . . 218
23.2.2.3 Interface Model 3 . . . . . . . . . . . . . . . . . . 218
23.2.2.4 Interface Model 4 . . . . . . . . . . . . . . . . . . 219

24 The PITM model 220

24.1 RANS mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220
24.2 LES mode . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 220

25 Hybrid LES/RANS for Dummies 224

25.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
25.1.1 Reynolds-Averaging Navier-Stokes equations: RANS . . . . 224
25.1.2 Large Eddy Simulations: LES . . . . . . . . . . . . . . . . 224
25.1.3 Zonal LES/RANS . . . . . . . . . . . . . . . . . . . . . . . 225
25.2 The PANS k − ε turbulence model . . . . . . . . . . . . . . . . . . 225
 6

25.3 Zonal LES/RANS: wall modeling . . . . . . . . . . . . . . . . . . . 226

25.3.1 The interface conditions . . . . . . . . . . . . . . . . . . . . 226
25.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
25.4 Zonal LES/RANS: embedded LES . . . . . . . . . . . . . . . . . . 227
25.4.1 The interface conditions . . . . . . . . . . . . . . . . . . . . 227
25.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

26 Commutation terms in the k and ω equations 230

27 Inlet boundary conditions 234

27.1 Synthesized turbulence . . . . . . . . . . . . . . . . . . . . . . . . 234
27.2 Random angles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
27.3 Highest wave number . . . . . . . . . . . . . . . . . . . . . . . . . 235
27.4 Smallest wave number . . . . . . . . . . . . . . . . . . . . . . . . . 235
27.5 Divide the wave number range . . . . . . . . . . . . . . . . . . . . 235
27.6 von Kármán spectrum . . . . . . . . . . . . . . . . . . . . . . . . . 236
27.7 Computing the fluctuations . . . . . . . . . . . . . . . . . . . . . . 236
27.8 Introducing time correlation . . . . . . . . . . . . . . . . . . . . . . 236
27.9 Anisotropic Synthetic Turbulent Fluctuations . . . . . . . . . . . . . 240
27.9.1 Eigenvalues and eigenvectors . . . . . . . . . . . . . . . . . 241
27.9.2 Synthetic fluctuations in the principal coordinate system . . . 242

28 Overview of LES, hybrid LES-RANS and URANS models 243

29 Best practice guidelines (BPG) 246

29.1 EU projects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
29.2 Ercoftac workshops . . . . . . . . . . . . . . . . . . . . . . . . . . 246
29.3 Ercoftac Classical Database . . . . . . . . . . . . . . . . . . . . . . 247
29.4 ERCOFTAC QNET Knowledge Base Wiki . . . . . . . . . . . . . . 247

A TME225: ǫ − δ identity 248

B TME225 Assignment 1 in 2016: laminar flow in a channel 249

B.1 Fully developed region . . . . . . . . . . . . . . . . . . . . . . . . 249
B.2 Wall shear stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
B.3 Inlet region . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
B.4 Wall-normal velocity in the developing region . . . . . . . . . . . . 251
B.5 Vorticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
B.6 Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
B.7 Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
B.8 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
B.9 Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

C TME225 Assignment 1 in 2017: laminar flow in a boundary layer 253

C.1 Velocity profiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
C.2 Boundary layer thickness . . . . . . . . . . . . . . . . . . . . . . . 254
C.3 Velocity gradients . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
C.4 Skinfriction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
C.5 Vorticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.6 Deformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
C.7 Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
 7

C.8 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256

C.9 Eigenvectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 257
C.10 Terms in the v1 equation . . . . . . . . . . . . . . . . . . . . . . . . 257

D TME225: Fourier series 258

D.1 Orthogonal functions . . . . . . . . . . . . . . . . . . . . . . . . . 258
D.2 Trigonometric functions . . . . . . . . . . . . . . . . . . . . . . . . 259
D.2.1 “Length” of ψk . . . . . . . . . . . . . . . . . . . . . . . . 260
D.2.2 Orthogonality of φn and ψk . . . . . . . . . . . . . . . . . . 260
D.2.3 Orthogonality of φn and φk . . . . . . . . . . . . . . . . . . 260
D.2.4 “Length” of φk . . . . . . . . . . . . . . . . . . . . . . . . 260
D.3 Fourier series of a function . . . . . . . . . . . . . . . . . . . . . . 261
D.4 Derivation of Parseval’s formula . . . . . . . . . . . . . . . . . . . 261
D.5 Complex Fourier series . . . . . . . . . . . . . . . . . . . . . . . . 263

E TME225: Why does the energy spectrum, E, have such strange dimensions?264
E.1 Energy spectrum for an ideal vortex . . . . . . . . . . . . . . . . . . 264
E.2 An example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
E.3 An example: Matlab code . . . . . . . . . . . . . . . . . . . . . . . 266

F TME225 Assignment 2: turbulent flow 270

F.1 Time history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
F.2 Time averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
F.3 Mean flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
F.4 The time-averaged momentum equation . . . . . . . . . . . . . . . 271
F.5 The gradient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
F.6 Wall shear stress . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
F.7 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . 272
F.8 Fluctuating wall shear stress . . . . . . . . . . . . . . . . . . . . . . 272
F.9 Production terms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
F.10 Pressure-strain terms . . . . . . . . . . . . . . . . . . . . . . . . . . 273
F.11 Dissipation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
F.12 Do something fun! . . . . . . . . . . . . . . . . . . . . . . . . . . . 274

G TME225 Learning outcomes 2016 275

H MTF270: Some properties of the pressure-strain term 287

I MTF270: Galilean invariance 288

J MTF270: Computation of wavenumber vector and angles 290

J.1 The wavenumber vector, κnj . . . . . . . . . . . . . . . . . . . . . . 290
J.2 Unit vector σin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

K MTF270: 1D and 3D energy spectra 292

K.1 Energy spectra from two-point correlations . . . . . . . . . . . . . . 293

L MTF270, Assignment 1: Reynolds averaged Navier-Stokes 295

L.1 Part I: Data of Two-dimensional flow . . . . . . . . . . . . . . . . . 295
L.1.1 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
L.1.2 The momentum equations . . . . . . . . . . . . . . . . . . . 296
 8

L.1.3 The turbulent kinetic energy equation . . . . . . . . . . . . 297

L.1.4 The Reynolds stress equations . . . . . . . . . . . . . . . . 297
L.2 Compute derivatives on a curvi-linear mesh . . . . . . . . . . . . . . 299
L.2.1 Geometrical quantities . . . . . . . . . . . . . . . . . . . . 300
L.3 Part II: StarCCM+ . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
L.3.1 Backward Facing Step Tutorial . . . . . . . . . . . . . . . . 301
L.3.2 Asymmetric diffuser case . . . . . . . . . . . . . . . . . . . 301
L.3.3 Brief instruction to begin the asymmetric diffuser case . . . 302

M MTF270, Assignment 2: LES 306

M.1 Time history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
M.2 Time averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
M.3 Auto correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
M.4 Probability density/histogram . . . . . . . . . . . . . . . . . . . . . 308
M.5 Frequency spectrum . . . . . . . . . . . . . . . . . . . . . . . . . . 309
M.6 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
M.7 Resolved production and pressure-strain . . . . . . . . . . . . . . . 311
M.8 Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
M.9 SGS stresses: Smagorinsky model . . . . . . . . . . . . . . . . . . 312
M.10 SGS stresses: WALE model . . . . . . . . . . . . . . . . . . . . . . 313
M.11 Dissipations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313
M.12 Test filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
M.13 Near-wall behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
M.14 Two-point correlations . . . . . . . . . . . . . . . . . . . . . . . . . 316
M.15 Energy spectra, optional . . . . . . . . . . . . . . . . . . . . . . . . 317
M.16 Something fun . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 317

N MTF270: Compute energy spectra from LES/DNS data using Matlab 318
N.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318
N.2 An example of using FFT . . . . . . . . . . . . . . . . . . . . . . . 318
N.3 Energy spectrum from the two-point correlation . . . . . . . . . . . 321
N.4 Energy spectra from the autocorrelation . . . . . . . . . . . . . . . . 323

O MTF270, Assignment 3: Zonal PANS, DES, DDES and SAS 324

O.1 Time history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324
O.2 Mean velocity profile . . . . . . . . . . . . . . . . . . . . . . . . . 325
O.3 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . 325
O.4 Turbulent kinetic energy . . . . . . . . . . . . . . . . . . . . . . . . 326
O.5 The modelled turbulent shear stress . . . . . . . . . . . . . . . . . . 326
O.6 Location of interface in DES and DDES . . . . . . . . . . . . . . . 326
O.7 SAS turbulent length scales . . . . . . . . . . . . . . . . . . . . . . 326

P MTF270, Assignment 4: Embedded LES with PANS 328

P.1 Time history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
P.2 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
P.3 Turbulent viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . 329
P.4 Modeled stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
P.5 Turbulent SGS dissipation . . . . . . . . . . . . . . . . . . . . . . . 330

Q MTF270, Assignment 5, recirculating flow: PANS, DES and DDES 332

 9

Q.1 Discretization schemes . . . . . . . . . . . . . . . . . . . . . . . . 333

Q.2 The modeled turbulent shear stress . . . . . . . . . . . . . . . . . . 333
Q.3 The turbulent viscosity . . . . . . . . . . . . . . . . . . . . . . . . 334
Q.4 Location of interface in DES and DDES . . . . . . . . . . . . . . . 335
Q.5 Compute fk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

R MTF270, Assignment 6: Hybrid LES-RANS 336

R.1 Time history . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 336
R.2 Mean velocity profile . . . . . . . . . . . . . . . . . . . . . . . . . 337
R.3 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . . 337
R.4 Turbulent kinetic energy . . . . . . . . . . . . . . . . . . . . . . . . 338
R.5 The modelled turbulent shear stress . . . . . . . . . . . . . . . . . . 338
R.6 Location of interface in DES and DDES . . . . . . . . . . . . . . . 338
R.7 Turbulent length scales . . . . . . . . . . . . . . . . . . . . . . . . 338

S MTF270: Transformation of tensors 339

S.1 Rotation from x1∗ − x2∗ to x1 − x2 . . . . . . . . . . . . . . . . . 339
S.2 Rotation from x1 − x2 to x1∗ − x2∗ . . . . . . . . . . . . . . . . . 340
S.3 Transformation of a velocity gradient . . . . . . . . . . . . . . . . . 341

T MTF270: Green’s formulas 342

T.1 Green’s first formula . . . . . . . . . . . . . . . . . . . . . . . . . . 342
T.2 Green’s second formula . . . . . . . . . . . . . . . . . . . . . . . . 342
T.3 Green’s third formula . . . . . . . . . . . . . . . . . . . . . . . . . 342
T.4 Analytical solution to Poisson’s equation . . . . . . . . . . . . . . . 345

U MTF270: Learning outcomes for 2017 346

V References 355
 10

TME225 Mechanics of fluids

L. Davidson
Division of Fluid Dynamics, Department of Applied Mechanics
Chalmers University of Technology, Göteborg, Sweden
http://www.tfd.chalmers.se/˜lada, [email protected]
This report can be downloaded at
http://www.tfd.chalmers.se/˜lada/MoF/
1. Motion, flow 11

T (Xi , t1 )

Xi
T (x1i , t1 )
T (Xi , t2 )

T (x2i , t2 )

Figure 1.1: The temperature of a fluid particle described in Lagrangian, T (Xi , t), or
Eulerian, T (xi , t), approach.

1 Motion, flow
1.1 Eulerian, Lagrangian, material derivative
also [1], Chapt. 3.2.
S EE

Assume a fluid particle is moving along the line in Fig. 1.1. We can choose to study
its motion in two ways: Lagrangian or Eulerian. In the Lagrangian approach we keep
track of its original position (Xi ) and follow its path which is described by xi (Xi , t).
For example, at time t1 the temperature of the particle is T (Xi , t1 ), and at time t2 its
temperature is T (Xi , t2 ), see Fig. 1.1. This approach is not used for fluids because it
is very tricky to define and follow a fluid particle. It is however used when simulating
movement of particles in fluids (for example soot particles in gasoline-air mixtures in
combustion applications). The speed of the particle is then expressed as a function of
time and its position at time zero, i.e. vi = vi (Xi , t).
In the Eulerian approach we pick a position, e.g. x1i , and watch the particle pass
by. This approach is used for fluids. The temperature of the fluid, T , for example, is
expressed as a function of the position, i.e. T = T (xi ), see Fig. 1.1. It may be that the
temperature at position xi , for example, varies in time, t, and then T = T (xi , t).
Now we want to express how the temperature of a fluid particle varies. In the
Lagrangian approach we first pick the particle (this gives its starting position, Xi ).
Once we have chosen a particle its starting position is fixed, and temperature varies
only with time, i.e. T (t) and the temperature gradient can be written dT /dt.
In the Eulerian approach it is a little bit more difficult. We are looking for the
temperature gradient, dT /dt, but since we are looking at fixed points in space we
need to express the temperature as a function of both time and space. From classical
mechanics, we know that the velocity of a fluid particle is the time derivative of its
space location, i.e. vi = dxi /dt. The chain-rule now gives

dT ∂T dxj ∂T ∂T ∂T
= + = + vj (1.1)
dt ∂t dt ∂xj ∂t ∂xj

Note that we have to use partial derivative on T since it is a function of more than one
(independent) variable. The first term on the right side is the local rate of change; by local rate
this we mean that it describes the variation of T in time at position xi . The second term of change
on the right side is called the convective rate of change, which means that it describes Conv. rate
of change
 dv2 ∂v2
1.2. What is the difference between and ? 12
dt ∂t

the variation of T in space when is passes the point xi . The left side in Eq. 1.1 is called
the material derivative and is in this text denoted by dT /dt.
Equation 1.1 can be illustrated as follows. Put your finger out in the blowing wind.
The temperature gradient your finger experiences is ∂T /∂t. Imagine that you are a fluid
particle and that you ride on a bike. The temperature gradient you then experience is
the material derivative, dT /dt.

Exercise 1 Write out Eq. 1.1, term-by-term.

dv2 ∂v2
1.2 What is the difference between and ?
dt ∂t
dv2 ∂v2
Students sometimes get confused about the difference between and . Here we
dt ∂t
give a simple example. Figure 1.2 shows a flow path of the fluid particles which can be
expressed in time as
x1 = exp(t), x2 = exp(−t) (1.2)
and hence x2 = 1/x1 . The flow path is steady in time and it starts at (x1 , x2 ) = (0.5, 2)
and ends at (x1 , x2 ) = (2, 0.5). Equation 1.2 gives the velocities

dx1 dx2
v1L = = exp(t), v2L = = − exp(−t) (1.3)
dt dt
and Eqs. 1.2 and 1.3 give
v1E = x1 , v2E = −x2 (1.4)
(cf. Eq. 4.52). The superscripts E and L denote Eulerian and Lagrangian, respectively.
Note that v1L = v1E and v2L = v2E ; the only difference is that viE is expressed as function
of (t, x1 , x2 ) and viL as function of t (and in general also starting location, X1 , X2 ).
Now we can compute the time derivatives of the v2 velocity as

dv2L
= exp(−t)
dt
(1.5)
dv2E ∂v E ∂v E ∂v E
= 2 + v1E 2 + v2E 2 = 0 + x1 · 0 − x2 · (−1) = x2
dt ∂t ∂x1 ∂x2

dv2 dv E dv L
We find, of course, that = 2 = 2 = x2 = exp(−t).
dt dt dt
Consider, for example, the point (x1 , x2 ) = (1, 1) in Fig. 1.2. The difference bet-
dv2 ∂v2
ween and is now clearly seen by looking at Eq. 1.5. The velocity at the point
dt ∂t
∂v E ∂v2
(x1 , x2 ) = (1, 1) does not change in time and hence 2 = = 0. However, if we
∂t ∂t
sit on a particle which passes the location (x1 , x2 ) = (1, 1), the velocity, v2L , increases
dv L dv2
by time, 2 = = 1 (the velocity, v2 , gets less negative) . Actually it increases
dt dt
dv2
all the time from the starting point where = 2 and v2 = −2 until the end point
dt
dv2
where = 0.5 and v2 = −0.5.
dt
1.3. Viscous stress, pressure 13

x2
2 ▽

x1
0
0 1 2

Figure 1.2: Flow path x2 = 1/x1 . The filled circle shows the point (x1 , x2 ) = (1, 1).
▽: start (t = ln(0.5)); △: end (t = ln(2)).

1.3 Viscous stress, pressure

See also [1], Chapts. 6.3 and 8.1.

We have in Part I [2] derived the balance equation for linear momentum which
reads
ρv̇i − σji,j − ρfi = 0 (1.6)

Switch notation for the material derivative and derivatives so that

dvi ∂σji
ρ = + ρfi (1.7)
dt ∂xj

where the first and the second term on the right side represents, respectively, the net
force due to surface and volume forces (σij denotes the stress tensor). Stress is force
per unit area. The first term includes the viscous stress tensor, τij . As you have learnt
earlier, the first index relates to the surface at which the stress acts and the second
index is related to the stress component. For example, on a surface whose normal is
ni = (1, 0, 0) act the three stress components σ11 , σ12 and σ13 , see Fig. 1.3a; the
volume force acts in the middle of the fluid element, see Fig. 1.3b.
In the present notation we denote the velocity vector by v = vi = (v1 , v2 , v3 )
and the coordinate by x = xi = (x1 , x2 , x3 ). In the literature, you may find other
notations of the velocity vector such as ui = (u1 , u2 , u3 ). If no tensor notation is used
the velocity vector is usually denoted as (u, v, w) and the coordinates as (x, y, z).
The diagonal components of σij represent the normal stresses and the off-diagonal
components of σij represent the shear stresses. In Part I [2] you learnt that the pressure
is defined as minus the sum of the normal stress, i.e.

P = −σkk /3 (1.8)
1.4. Strain rate tensor, vorticity 14

(ê1 )
ti
σ12

σ11
σ13
x2 fi
x2

x1 x1
(a) Stress components and stress vector on a surface. (b) Volume force, fi = (0, −g, 0), acting in
the middle of the fluid element.

(ê1 )
Figure 1.3: Stress tensor, volume (gravitation) force and stress vector, ti , see
Eq. B.2.

The pressure, P , acts as a normal stress. In general, pressure is a thermodynamic

property, pt , which can be obtained – for example – from the ideal gas law. In that
case the thermodynamics pressure, pt , and the mechanical pressure, P , may not be the
same but Eq. 1.8 is nevertheless used. The viscous stress tensor, τij , is obtained by
subtracting the trace, σkk /3 = −P , from σij ; the stress tensor can then be written as

σij = −P δij + τij (1.9)

τij is the deviator of σij . The expression for the viscous stress tensor is found in Eq. 2.4
at p. 24. The minus-sign in front of P appears because the pressure acts into the surface.
When there is no movement, the viscous stresses are zero and then of course the normal
stresses are the same as the pressure. In general, however, the normal stresses are the
sum of the pressure and the viscous stresses, i.e.

σ11 = −P + τ11 , σ22 = −P + τ22 , σ33 = −P + τ33 , (1.10)

Exercise 2 Consider Fig. 1.3. Show how σ21 , σ22 , σ23 act on a surface with normal
vector ni = (0, 1, 0). Show also how σ31 , σ32 , σ33 act on a surface with normal vector
ni = (0, 0, 1).

Exercise 3 Write out Eq. 1.9 on matrix form.

1.4 Strain rate tensor, vorticity

See also [1], Chapt. 3.5.3, 3.6.

We need an expression for the viscous stresses, τij . They are needed in the mo-
mentum equations, Eq. 1.7 (see also Eq. 1.9). They will be expressed in the velocity
∂vi
gradients, ∂x j
. Hence we will now discuss the velocity gradients.
1.4. Strain rate tensor, vorticity 15

The velocity gradient tensor can be split into two parts as

 
∂vi 1  ∂vi ∂vi ∂vj ∂vj 
=  + + − 
∂xj 2 ∂xj ∂xj ∂xi ∂xi 
=0 (1.11)
2∂vi /∂xj
   
1 ∂vi ∂vj 1 ∂vi ∂vj
= + + − = Sij + Ωij
2 ∂xj ∂xi 2 ∂xj ∂xi
where
Sij is a symmetric tensor called the strain-rate tensor Strain-rate
tensor
Ωij is a anti-symmetric tensor called the vorticity tensor vorticity ten-
The vorticity tensor is related to the familiar vorticity vector which is the curl of sor
the velocity vector, i.e. ω = ∇ × v, or in tensor notation
∂vk
ωi = ǫijk (1.12)
∂xj
where ǫijk is the permuation tensor, see lecture notes of Toll & Ekh [2]. If we set, for
example, i = 3 we get
ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 . (1.13)
The vorticity represents rotation of a fluid particle. Inserting Eq. 1.11 into Eq. 1.12
gives
ωi = ǫijk (Skj + Ωkj ) = ǫijk Ωkj (1.14)
since ǫijk Skj = 0 because the product of a symmetric tensor (Skj ) and an anti-
symmetric tensor (εijk ) is zero. Let us show this for i = 1 by writing out the full
equation. Recall that Sij = Sji (i.e. S12 = S21 , S13 = S31 , S23 = S32 ) and
ǫijk = −ǫikj = ǫjki etc (i.e. ε123 = −ε132 = ε231 . . . , ε113 = ε221 = . . . ε331 = 0)

ε1jk Skj = ε111 S11 + ε112 S21 + ε113 S31

+ ε121 S12 + ε122 S22 + ε123 S32
+ ε131 S13 + ε132 S23 + ε133 S33
= 0 · S11 + 0 · S21 + 0 · S31 (1.15)
+ 0 · S12 + 0 · S22 + 1 · S32
+ 0 · S13 − 1 · S23 + 0 · S33
= S32 − S23 = 0
Now let us invert Eq. 1.14. We start by multiplying it with εiℓm so that
εiℓm ωi = εiℓm ǫijk Ωkj (1.16)
The ε-δ-identity gives (see Table A.1 at p. 248)
εiℓm ǫijk Ωkj = (δℓj δmk − δℓk δmj )Ωkj = Ωmℓ − Ωℓm = 2Ωmℓ (1.17)
This can easily be proved by writing out all the components, see Table A.1 at p. 248.
Now Eqs. 1.16 and 1.17 give
1 1 1
Ωmℓ = εiℓm ωi = εℓmi ωi = − εmℓi ωi (1.18)
2 2 2
1.5. Product of a symmetric and antisymmetric tensor 16

or, switching indices

1
Ωij = − εijk ωk (1.19)
2
It turns out that is is much easier to go from Eq. 1.14 to Eq. 1.19 by writing out the
components of Eq. 1.14 (here we do it for i = 1)

ω1 = ε123 Ω32 + ε132 Ω23 = Ω32 − Ω23 = −2Ω23 (1.20)

and we get
1
Ω23 = − ω1 (1.21)
2
which indeed is identical to Eq. 1.19.

Exercise 4 Write out the second and third component of the vorticity vector given in
Eq. 1.12 (i.e. ω2 and ω3 ).

Exercise 5 Complete the proof of Eq. 1.15 for i = 2 and i = 3.

Exercise 6 Write out Eq. 1.20 also for i = 2 and i = 3 and find an expression for Ω12
and Ω13 (cf. Eq. 1.21). Show that you get the same result as in Eq. 1.19.

Exercise 7 In Eq. 1.21 we proved the relation between Ωij and ωi for the off-diagonal
components. What about the diagonal components of Ωij ? What do you get from
Eq. 1.11?

Exercise 8 From your course in linear algebra, you should remember how to compute
a vector product using Sarrus’ rule. Use it to compute the vector product
 
ê1 ê2 ê3
ω = ∇ × v =  ∂x ∂
1
∂
∂x2
∂
∂x3

v1 v2 v3

Verify that this agrees with the expression in tensor notation in Eq. 1.12.

1.5 Product of a symmetric and antisymmetric tensor

In this section we show the proof that the product of a symmetric and antisymmetric
tensor is zero. First, we have the definitions:

• A tensor aij is symmetric if aij = aji ;

• A tensor bij is antisymmetric if bij = −bji .

It follows that for an antisymmetric tensor all diagonal components must be zero;
for example, b11 = −b11 can only be satisfied if b11 = 0.
The (inner) product of a symmetric and antisymmetric tensor is always zero. This
can be shown as follows
aij bij = aji bij = −aji bji ,
where we first used the fact that aij = aji (symmetric), and then that bij = −bji
(antisymmetric). Since the indices i and j are both dummy indices we can interchange
them in the last expression (−aji bji = −aij bij ), which gives

aij bij = −aij bij

1.6. Deformation, rotation 17

This expression says that A = −A which can be only true if A = 0 and hence aij bij =
0.
This can of course also be shown be writing out aij bij on component form, i.e.

aij bij = a11 b11 + a12 b12 + a13 b13

I II
+ a21 b21 +a22 b22 + a23 b23
I III
+ a31 b31 + a32 b32 +a33 b33 = 0
II III

The underlined terms are zero (b11 = b22 = b33 = 0); terms I cancel each other
(a12 = a21 , b12 = −b21 ) as do terms II and III.

1.6 Deformation, rotation

See also [1], Chapt. 3.3.

The velocity gradient can, as shown above, be divided into two parts: Sij and Ωij .
We have shown that the latter is connected to rotation of a fluid particle. During rotation rotation
the fluid particle is not deformed. This movement can be illustrated by Fig. 1.4. The
vertical movement (v2 ) of the lower-right corner (x1 + ∆x1 ) of the particle in Fig. 1.4
is estimated as follows. The velocity at the lower-left corner is v2 (x1 ). Now we need
the velocity at the lower-right corner which is located at x1 + ∆x1 . It is computed
using the first term in the Taylor series as1
∂v2
v2 (x1 + ∆x1 ) = v2 (x1 ) + ∆x1
∂x1
It is assumed that the fluid particle in Fig. 1.4 is rotated the angle α during the
time ∆t. The angle rotation per unit time can be estimated as dα/dt = −∂v1 /∂x2 =
∂v2 /∂x1 ; if the fluid element does not rotate as a solid body, the rotation is computed
as the average, i.e. dα/dt = (∂v2 /∂x1 − ∂v1 /∂x2 )/2. The vorticity is computed as
ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 = −2Ω12 = 2dα/dt, see Eq. 1.13 and Exercise 4. Hence,
the vorticity ω3 can be interpreted as twice the average rotation per unit time of the
horizontal edge (∂v2 /∂x1 ) and vertical edge (−∂v1 /∂x2 ).
Next let us have a look at the deformation caused by Sij . It can be divided into two
parts, namely shear and elongation (also called extension or dilatation). The deforma-
tion due to shear is caused by the off-diagonal terms of Sij . In Fig. 1.5, a pure shear de-
formation by S12 = (∂v1 /∂x2 + ∂v2 /∂x1 )/2 is shown. The deformation due to elon-
gation is caused by the diagonal terms of Sij . Elongation caused by S11 = ∂v1 /∂x1 is
illustrated in Fig. 1.6.
In general, a fluid particle experiences a combination of rotation, deformation and
elongation as indeed is given by Eq. 1.11.

Exercise 9 Consider Fig. 1.4. Show and formulate the rotation by ω1 .

Exercise 10 Consider Fig. 1.5. Show and formulate the deformation by S23 .

Exercise 11 Consider Fig. 1.6. Show and formulate the elongation by S22 .
1 this corresponds to the equation for a straight line y = kx + ℓ where k is the slope which is equal to the

derivative of y, i.e. dy/dx, and ℓ = v2 (x1 )

1.6. Deformation, rotation 18

∆x1

∂v1
− ∆x2 ∆t
∂x2

∆x2

∂v2
∆x1 ∆t
α ∂x1

x2

x1

Figure 1.4: Rotation of a fluid particle during time ∆t. Here ∂v1 /∂x2 = −∂v2 /∂x1
so that −Ω12 = ω3 /2 = ∂v2 /∂x1 > 0.

∆x1

∂v1
∆x2 ∆t
∂x2

∆x2

∂v2
∆x1 ∆t
α ∂x1

x2

x1

Figure 1.5: Deformation of a fluid particle by shear during time ∆t. Here ∂v1 /∂x2 =
∂v2 /∂x1 so that S12 = ∂v1 /∂x2 > 0.
1.7. Irrotational and rotational flow 19

∆x1

∂v1
∆x1 ∆t
∂x1

∆x2

x2

x1

Figure 1.6: Deformation of a fluid particle by elongation during time ∆t.

ti dℓ

x2

x1

Figure 1.7: The surface, S, is enclosing by the line ℓ. The vector, ti , denotes the unit
tangential vector of the enclosing line, ℓ.

1.7 Irrotational and rotational flow

In the previous subsection we introduced different types of movement of a fluid parti-
cle. One type of movement was rotation, see Fig. 1.4. Flows are often classified based
on rotation: they are rotational (ωi 6= 0) or irrotational (ωi = 0); the latter type is also
called inviscid flow or potential flow. We will talk more about that later on, see Sec-
tion 4.4. In this subsection we will give examples of one irrotational and one rotational
flow. In potential flow, there exists a potential, Φ, from which the velocity components
can be obtained as
∂Φ
vk = (1.22)
∂xk
Before we talk about the ideal vortex line in the next section, we need to introduce
the concept circulation. Consider a closed line on a surface in the x1 − x2 plane, see
Fig. 1.7. When the velocity is integrated along this line and projected onto the line we
1.7. Irrotational and rotational flow 20

obtain the circulation I

Γ= vm tm dℓ (1.23)

Using Stokes’s theorem we can relate the circulation to the vorticity as

I Z Z Z
∂vk
Γ = vm tm dℓ = εijk ni dS = ωi ni dS = ω3 dS (1.24)
S ∂xj S S

where ni = (0, 0, 1) is the unit normal vector of the surface S. Equation 1.24 reads in
vector notation
I Z Z Z
Γ = v · tdℓ = (∇ × v) · ndS = ω · ndS = ω3 dS (1.25)
S S S

The circulation is useful in, for example, aeronautics and windpower engineering
where the lift of a 2D section of an airfoil or a rotorblade is expressed in the circulation
for that 2D section. The lift force is computed as (see Eqs. 4.85 and 4.86)

L = ρV Γ (1.26)

where V is the velocity around the airfoil (for a rotorblade it is the relative velocity,
since the rotorblade is rotating). In an PhD project, an inviscid simulation method
(based on the circulation and vorticity sources) is used to compute the aerodynamic
loads for windturbine rotorblades [3].

Exercise 12 In potential flow ωi = εijk ∂vk /∂xj = 0. Multiply Eq. 1.22 by εijk and
derivate with respect to xk (i.e. take the curl of) and show that the right side becomes
zero as it should, i.e. εijk ∂ 2 Φ/(∂xk ∂xj ) = 0.

1.7.1 Ideal vortex line

The ideal vortex line is an irrotational (potential) flow where the fluid moves along
circular paths, see Fig. 1.8. The governing equations are derived in Section 4.4.5. The
velocity field in polar coordinates reads
Γ
vθ = , vr = 0 (1.27)
2πr
where Γ is the circulation. Its potential reads
Γθ
Φ= (1.28)
2π
The velocity, vθ , is then obtained as
1 ∂Φ Γ
vθ = = (1.29)
r ∂θ 2πr
To transform Eq. 1.27 into Cartesian velocity components, consider Fig. 1.9. The
Cartesian velocity vectors are expressed as
x2 x2
v1 = −vθ sin(θ) = −vθ = −vθ 2
r (x1 + x22 )1/2
x1 x1 (1.30)
v2 = vθ cos(θ) = vθ = vθ 2
r (x1 + x22 )1/2
1.7. Irrotational and rotational flow 21

1.5

0.5

x2
0

−0.5

−1
−1.5 −1 −0.5 0 0.5 1 1.5
x1
Figure 1.8: Ideal vortex. The fluid particle (i.e. its diagonal, see Fig. 1.4) does not
rotate. The locations of the fluid particle is indicated by black, filled squares. The
diagonales are shown as black dashed lines. The fluid particle is shown at θ = 0, π/4,
3π/4, π, 5π/4, 3π/2 and −π/6.

Inserting Eq. 1.30 into Eq. 1.27 we get

Γx2 Γx1
v1 = − , v2 = . (1.31)
2π(x21 + x22 ) 2π(x21 + x22 )

To verify that this flow is a potential flow, we need to show that the vorticity, ωi =
εijk ∂vk /∂xj is zero. Since it is a two-dimensional flow (v3 = ∂/∂x3 = 0), ω1 =
ω2 = 0, we only need to compute ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 . The velocity derivatives
are obtained as
∂v1 Γ x21 − x22 ∂v2 Γ x22 − x21
=− , = (1.32)
∂x2 2π (x21 + x22 )2 ∂x1 2π (x21 + x22 )2

and we get
Γ 1
ω3 = (x2 − x21 + x21 − x22 ) = 0 (1.33)
2π (x21 + x22 )2 2
which shows that the flow is indeed a potential flow, i.e. irrotational (ωi ≡ 0). Note
that the deformation is not zero, i.e.
 
1 ∂v1 ∂v2 Γ x22
S12 = + = (1.34)
2 ∂x2 ∂x1 2π (x21 + x22 )2

Hence a fluid particle in an ideal vortex does deform but it does not rotate (i.e. its
diagonal does not rotate, see Fig. 1.8).
It may be little confusing that the flow path forms a vortex but the flow itself has no
vorticity. Thus one must be very careful when using the words “vortex” and ”vorticity”. vortex vs.
By vortex we usually mean a recirculation region of the mean flow. That the flow has vorticity
no vorticity (i.e. no rotation) means that a fluid particle moves as illustrated in Fig. 1.8.
As a fluid particle moves from position a to b – on its counter-clockwise-rotating path
– the particle itself is not rotating. This is true for the whole flow field, except at the
center where the fluid particle does rotate. This is a singular point as is seen from
Eq. 1.27 for which vθ → ∞.
Note that generally a vortex has vorticity, see Section 4.2. The ideal vortex is a very
special flow case.
1.8. Eigenvalues and and eigenvectors: physical interpretation 22

vθ
θ
x2

r
θ x1

Figure 1.9: Transformation of vθ into Cartesian components.

v1 v1
b c
α

x2
α

x3 x1

Figure 1.10: A shear flow. The fluid particle rotates. v1 = cx22 .

1.7.2 Shear flow

Another example – which is rotational – is the lower half of fully-developed channel
flow for which the velocity reads (see Eq. 3.28)
v1 4x2  x2 
= 1− , v2 = 0 (1.35)
v1,max h h

where x2 < h/2, see Fig. 1.10. The vorticity vector for this flow reads
 
∂v2 ∂v1 4 2x2
ω1 = ω2 = 0, ω3 = − =− 1− (1.36)
∂x1 ∂x2 h h

When the fluid particle is moving from position a, via b to position c it is indeed
rotating. It is rotating in clockwise direction. Note that the positive rotating direction
is defined as the counter-clockwise direction, indicated by α in Fig. 1.10. This is why
the vorticity, ω3 , in the lower half of the channel (x2 < h/2) is negative. In the upper
half of the channel the vorticity is positive because ∂v1 /∂x2 < 0.

1.8 Eigenvalues and and eigenvectors: physical interpretation

See also [1], Chapt. 2.5.5.
PSfrag

1.8. Eigenvalues and and eigenvectors: physical interpretation 23

σ23 v̂1
v̂2
σ21 λ2
σ12 λ1
σ11
x2′ x1′
x2

x1

Figure 1.11: A two-dimensional fluid element. Left: in original state; right: rotated to
principal coordinate directions. λ1 and λ2 denote eigenvalues; v̂1 and v̂2 denote unit
eigenvectors.

Consider a two-dimensional fluid (or solid) element, see Fig. 1.11. In the left figure
it is oriented along the x1 − x2 coordinate system. On the surfaces act normal stresses
(σ11 , σ22 ) and shear stresses (σ12 , σ21 ). The stresses form a tensor, σij . Any tensor has
eigenvectors and eigenvalues (also called principal vectors and principal values). Since
σij is symmetric, the eigenvalues are real (i.e. not imaginary). The eigenvalues are
obtained from the characteristic equation, see [1], Chapt. 2.5.5 or Eq. 13.5 at p. 153.
When the eigenvalues have been obtained, the eigenvectors can be computed. Given
the eigenvectors, the fluid element is rotated α degrees so that its edges are aligned
with the eigenvectors, v̂1 = x̂1′ and v̂2 = x̂2′ , see right part of Fig. 1.11. Note that the
sign of the eigenvectors is not defined, which means that the eigenvectors can equally
well be chosen as −v̂1 and/or −v̂2 . In the principal coordinates x1′ − x2′ (right part
of Fig. 1.11), there are no shear stresses on the surfaces of the fluid element. There
are only normal stresses. This is the very definition of eigenvectors. Furthermore, the
eigenvalues are the normal stresses in the principal coordinates, i.e. λ1 = σ1′ 1′ and
λ2 = σ2′ 2′ .
2. Governing flow equations 24

2 Governing flow equations

EE also [1], Chapts. 5 and 8.1.
S
2.1 The Navier-Stokes equation
2.1.1 The continuity equation
The first equation is the continuity equation (the balance equation for mass) which
reads [2]
ρ̇ + ρvi,i = 0 (2.1)
Change of notation gives
dρ ∂vi
+ρ =0 (2.2)
dt ∂xi
For incompressible flow (ρ = const) we get
∂vi
=0 (2.3)
∂xi

2.1.2 The momentum equation

The next equation is the momentum equation. We have formulated the constitutive law
for Newtonian viscous fluids [2]
2
σij = −P δij + 2µSij − µSkk δij
3 (2.4)
2
τij = 2µSij − µSkk δij
3
Inserting Eq. 2.4 into the balance equations, Eq. 1.7, we get
 
dvi ∂P ∂τji ∂P ∂ 2 ∂vk
ρ =− + + ρfi = − + 2µSij − µ δij + ρfi (2.5)
dt ∂xi ∂xj ∂xi ∂xj 3 ∂xk
where µ denotes the dynamic viscosity. This is the Navier-Stokes equations (sometimes
the continuity equation is also included in the name “Navier-Stokes”). It is also called
the transport equation for momentum. If the viscosity, µ, is constant it can be moved
outside the derivative. Furthermore, if the flow is incompressible the second term in
the parenthesis on the right side is zero because of the continuity equation. If these two
requirements are satisfied we can also re-write the first term in the parenthesis as
 
∂ ∂ ∂vi ∂vj ∂ 2 vi
(2µSij ) = µ + =µ (2.6)
∂xj ∂xj ∂xj ∂xi ∂xj ∂xj
because of the continuity equation, i.e.
   
∂ ∂vj ∂ ∂vj
µ =µ = 0. (2.7)
∂xj ∂xi ∂xi ∂xj
Equation 2.5 can now – for constant µ and incompressible flow – be written
dvi ∂P ∂ 2 vi
ρ =− +µ + ρfi (2.8)
dt ∂xi ∂xj ∂xj
2.2. The energy equation 25

In inviscid (potential) flow, there are no viscous (friction) forces. In this case, the
Navier-Stokes equation reduces to the Euler equations Euler
equations
dvi ∂P
ρ =− + ρfi (2.9)
dt ∂xi
Exercise 13 Equation 1.7 states that mass times acceleration is equal to the sum of
forces forces (per unit volume). Write out the momentum equation (without using the
summation rule) for the x1 direction and show the surface forces and the volume force
on a small, square fluid element (see lecture notes of Toll & Ekh [2]). Now repeat it for
the x2 direction.

Exercise 14 Formulate the Navier-Stokes equation for incompressible flow but non-
constant viscosity.

2.2 The energy equation

See also [1], Chapts. 6.4 and 8.1.

We have in Part I [2] derived the energy equation which reads

ρu̇ − vi,j σji + qi,i = ρz (2.10)

where u denotes internal energy. qi denotes the conductive heat flux and z the net
radiative heat source. For simplicity, we neglect the radiation from here on. Change of
notation gives
du ∂vi ∂qi
ρ = σji − (2.11)
dt ∂xj ∂xi
In Part I [2] we formulated the constitutive law for the heat flux vector (Fourier’s
law)
∂T
qi = −k (2.12)
∂xi
Inserting the constitutive laws, Eqs. 2.4 and 2.12, into Eq. 2.11 gives
 
du ∂vi 2 ∂ ∂T
ρ = −P + 2µSij Sij − µSkk Sii + k (2.13)
dt ∂xi 3 ∂xi ∂xi
Φ

where we have used Sij ∂vi /∂xj = Sij (Sij + Ωij ) = Sij Sij because the product of a
symmetric tensor, Sij , and an anti-symmetric tensor, Ωij , is zero. Two of the viscous
terms (denoted by Φ) represent irreversible viscous heating (i.e. transformation of
kinetic energy into thermal energy); these terms are important at high-speed flow2 (for
example re-entry from outer space) and for highly viscous flows (lubricants). The first
term on the right side represents reversible heating and cooling due to compression and
expansion of the fluid. Equation 2.13 is the transport equation for (internal) energy, u.
Now we assume that the flow is incompressible (i.e. the velocity should be smaller
than approximately 1/3 of the speed of sound) for which

du = cp dT (2.14)
2 High-speed flows relevant for aeronautics will be treated in detail in the course “TME085 Compressible

flow” in the MSc programme.

2.3. Transformation of energy 26

where cp is the heat capacity (see Part I) [2] so that Eq. 2.13 gives (cp is assumed to be
constant)  
dT ∂ ∂T
ρcp =Φ+ k (2.15)
dt ∂xi ∂xi
The dissipation term is simplified to Φ = 2µSij Sij because Sii = ∂vi /∂xi = 0. If we
furthermore assume that the heat conductivity coefficient is constant and that the fluid
is a gas or a common liquid (i.e. not an lubricant oil), we get

dT ∂2T
=α (2.16)
dt ∂xi ∂xi
where α = k/(ρcp ) is the thermal diffusivity. The Prandtl number is defined as thermal
diffusivity
ν
Pr = (2.17)
α
where ν = µ/ρ is the kinematic viscosity. The physical meaning of the Prandtl number
is the ratio of how well the fluid diffuses momentum to how well it diffuses internal
energy (i.e. temperature).
The dissipation term, Φ, is neglected in Eq. 2.16 when one of two assumptions are
valid:
1. The fluid is a gas with low velocity (lower than 1/3 of the speed of sound); this
assumption was made when we assumed that the fluid is incompressible
2. The fluid is a common liquid (i.e. not an lubricant oil). In lubricant oils the
viscous heating (i.e. the dissipation, Φ) is large. One example is the oil flow in a
gearbox in a car where the temperature usually is more than 100oC higher when
the car is running compared to when it is idle.

Exercise 15 Write out and simplify the dissipation term, Φ, in Eq. 2.13. The first term
is positive and the second term is negative; are you sure that Φ > 0?

2.3 Transformation of energy

Now we will derive the equation for the kinetic energy, k = vi vi /2. Multiply Eq. 1.7
with vi
dvi ∂σji
ρvi − vi − vi ρfi = 0 (2.18)
dt ∂xj
Using the product rule backwards (Trick 2, see Eq. 8.4), the first term on the left side
can be re-written
dvi 1 d(vi vi ) dk
ρvi = ρ =ρ (2.19)
dt 2 dt dt
(vi vi /2 = k) so that
dk ∂σji
ρ = vi + ρvi fi (2.20)
dt ∂xj
Re-write the stress-velocity term so that (Trick 1, see Eq. 8.2)

dk ∂vi σji ∂vi

ρ = − σji + ρvi fi (2.21)
dt ∂xj ∂xj
2.4. Left side of the transport equations 27

This is the transport equation for kinetic energy, k. Adding Eq. 2.21 to Eq. 2.11 gives

d(u + k) ∂σji vi ∂qi

ρ = − + ρvi fi (2.22)
dt ∂xj ∂xi

This is an equation for the sum of internal and kinetic energy, u + k. This is the
transport equation for total energy, u + k.
Let us take a closer look at Eqs. 2.11, 2.21 and 2.22. First we separate the term
σji ∂vi /∂xj in Eqs. 2.11 and 2.21 into work related to the pressure and viscous stresses
respectively (see Eq. 1.9), i.e.
∂vi ∂vi ∂vi
σji = −P + τji (2.23)
∂xj ∂xi ∂xj
a b=Φ

The following things should be noted.

• The physical meaning of the a-term in Eq. 2.23 – which includes the pressure, P
– is heating/cooling by compression/expansion. This is a reversible process, i.e.
no loss of energy but only transformation of energy.
• The physical meaning of the b-term in Eq. 2.23 – which includes the viscous
stress tensor, τij – is a dissipation, which means that kinetic energy is trans-
formed to thermal energy. It is denoted Φ, see Eq. 2.13, and is called viscous
dissipation. It is always positive and represents irreversible heating.
• The dissipation, Φ, appears as a sink term in the equation for the kinetic energy, k
(Eq. 2.21) and it appears as a source term in the equation for the internal energy,
u (Eq. 2.11). The transformation of kinetic energy into internal energy takes
place through this source term. In incompressible flow for which the viscous
term in Navier-Stokes can be simplified (see Eq. 2.8), the viscous term reads
∂vi ∂vi ∂vi
τji =µ (2.24)
∂xj ∂xj ∂xj

This quantity is very important in turbulent flow, cf. Eqs. 8.14 and 8.35.
• Φ does not appear in the equation for the total energy u+k (Eq. 2.22); this makes
sense since Φ represents a energy transfer between u and k and does not affect
their sum, u + k.

Dissipation is very important in turbulence where transfer of energy takes place at

several levels. First energy is transferred from the mean flow to the turbulent fluctua-
tions. The physical process is called production of turbulent kinetic energy. Then we
have transformation of kinetic energy from turbulence kinetic energy to thermal en-
ergy; this is turbulence dissipation (or heating). At the same time we have the usual
viscous dissipation from the mean flow to thermal energy, but this is much smaller than
that from the turbulence kinetic energy. For more detail, see Section 5.

2.4 Left side of the transport equations

So far, the left sides in transport equations have been formulated using the material
derivative, d/dt. Let ψ denote a transported quantity (i.e. ψ = vi , u, T . . .); the left
2.5. Material particle vs. control volume (Reynolds Transport Theorem) 28

side of the equation for momentum, thermal energy, total energy, temperature etc reads

dψ ∂ψ ∂ψ
ρ =ρ + ρvj (2.25)
dt ∂t ∂xj

This is often called the non-conservative form. Using the continuity equation, Eq. 2.2, non-
it can be re-written as conser-
  vative
dψ ∂ψ ∂ψ dρ ∂vj
ρ =ρ + ρvj +ψ +ρ =
dt ∂t ∂xj dt ∂xj
=0 (2.26)
 
∂ψ ∂ψ ∂ρ ∂ρ ∂vj
ρ + ρvj +ψ + vj +ρ
∂t ∂xj ∂t ∂xj ∂xj

The two underlined terms will form a time derivative term, and the other three terms
can be collected into a convective term, i.e.
dψ ∂ρψ ∂ρvj ψ
ρ = + (2.27)
dt ∂t ∂xj

Thus, the left side of the temperature equation and the Navier-Stokes, for example, can
be written in three different ways (by use of the chain-rule and the continuity equation)

dvi ∂vi ∂vi ∂ρvi ∂ρvj vi

ρ =ρ + ρvj = +
dt ∂t ∂xj ∂t ∂xj
(2.28)
dT ∂T ∂T ∂ρvi ∂ρvj T
ρ =ρ + ρvj = +
dt ∂t ∂xj ∂t ∂xj

The continuity equation can also be written in three ways (by use of the chain-rule)

dρ ∂vi ∂ρ ∂ρ ∂vi ∂ρ ∂ρvi

+ρ = + vi +ρ = + (2.29)
dt ∂xi ∂t ∂xi ∂xi ∂t ∂xi
The forms on the right sides of Eqs. 2.28 and 2.29 are called the conservative form. conser-
When solving transport equations (such as the Navier-Stokes) numerically using finite vative
volume methods, the left sides in the transport equation are always written as the ex-
pressions on the right side of Eqs. 2.28 and 2.29; in this way Gauss law can be used
to transform the equations from a volume integral to a surface integral and thus ensur-
ing that the transported quantities are conserved. The results may be inaccurate due
to too coarse a numerical grid, but no mass, momentum, energy etc is lost (provided a
transport equation for the quantity is solved): “what comes in goes out”.

2.5 Material particle vs. control volume (Reynolds Transport The-

orem)
See also lecture notes of Toll & Ekh [2] and [1], Chapt. 5.2.

In Part I [2] we initially derived all balance equations (mass, momentum R and en-
ergy) for a collection of material
R particles. The conservation of mass, d/dt ρdV = 0,
Newton’s second law, d/dt ρvi = Fi etc were derived for a collection of particles in
the volume Vpart , where Vpart is a volume that includes the same fluid particles all the
time. This means that the volume, Vpart , must be moving and it may expand or contract
2.5. Material particle vs. control volume (Reynolds Transport Theorem) 29

(if the density is non-constant), otherwise particles would move across its boundaries.
The equations we have looked at so far (the continuity equation 2.3, the Navier-Stokes
equation 2.8, the energy equations 2.13 and 2.21) are all given for a fixed control vol-
ume. How come? The answer is the Reynolds transport theorem, which converts the
equations from being valid for a moving, deformable volume with a collection of parti-
cles, Vpart , to being valid for a fixed volume, V . The Reynolds transport theorem reads
(first line)
Z Z  
d dΦ ∂vi
ΦdV = +Φ dV
dt Vpart V dt ∂xi
Z   Z  
∂Φ ∂Φ ∂vi ∂Φ ∂vi Φ
= + vi +Φ dV = + dV (2.30)
V ∂t ∂xi ∂xi V ∂t ∂xi
Z Z
∂Φ
= dV + vi ni ΦdS
V ∂t S

where V denotes a fixed non-deformable volume in space. The divergence of the ve-
locity vector, ∂vi /∂xi , on the first line represents the increase or decrease of Vpart
during dt. The divergence theorem was used to obtain the last line and S denotes the
bounding surface of volume V . The last term on the last line represents the net flow
of Φ across the fixed non-deformable volume, V . Φ in the equation above can be ρ
(mass), ρvi (momentum) or ρu (energy). This equation applies to any volume at every
instant and the restriction to a collection of a material particles is no longer necessary.
Hence, in fluid mechanics the transport equations (Eqs. 2.2, 2.5, 2.11, . . . ) are valid
both for a material collection of particles as well as for a volume; the latter is usually
fixed (this is not necessary).
3. Solutions to the Navier-Stokes equation: three examples 30

x2
V0
x1

Figure 3.1: The plate moves to the right with speed V0 for t > 0.

x2 t3

t2
t1

0 0.2 0.4 0.6 0.8 1

v1 /V0
Figure 3.2: The v1 velocity at three different times. t3 > t2 > t1 .

3 Solutions to the Navier-Stokes equation: three exam-

ples
3.1 The Rayleigh problem
MAGINE the sudden motion of an infinitely long flat plate. For time greater than
I zero the plate is moving with the speed V0 , see Fig. 3.1. Because the plate is in-
finitely long, there is no x1 dependency. Hence the flow depends only on x2 and t, i.e.
v1 = v1 (x2 , t) and p = p(x2 , t). Furthermore, ∂v1 /∂x1 = ∂v3 /∂x3 = 0 so that the
continuity equation gives ∂v2 /∂x2 = 0. At the lower boundary (x2 = 0) and at the
upper boundary (x2 → ∞) the velocity component v2 = 0, which means that v2 = 0
in the entire domain. So, Eq. 2.8 gives (no body forces, i.e. f1 = 0) for the v1 velocity
component
∂v1 ∂ 2 v1
ρ =µ 2 (3.1)
∂t ∂x2
We will find that the diffusion process depends on the kinematic viscosity, ν = µ/ρ,
rather than the dynamic one, µ. The boundary conditions for Eq. 3.1 are

v1 (x2 , t = 0) = 0, v1 (x2 = 0, t) = V0 , v1 (x2 → ∞, t) = 0 (3.2)

The solution to Eq. 3.1 is shown in Fig. 3.2. For increasing time (t3 > t2 > t1 ), the
moving plate affects the fluid further and further away from the plate.
It turns out that the solution to Eq. 3.1 is a similarity solution; this means that the similarity
number of independent variables is reduced by one, in this case from two (x2 and t) to solution
one (η). The similarity variable, η, is related to x2 and t as
x2
η= √ (3.3)
2 νt
3.1. The Rayleigh problem 31

If the solution of Eq. 3.1 depends only on η, it means that the solution for a given fluid
will√be the same (“similar”) for many (infinite) values of x2 and t as long as the ratio
x2 / νt is constant. Now we need to transform the derivatives in Eq. 3.1 from ∂/∂t
and ∂/∂x2 to d/dη so that it becomes a function of η only. We get

∂v1 dv1 ∂η x2 t−3/2 dv1 1 η dv1

= =− √ =−
∂t dη ∂t 4 ν dη 2 t dη
∂v1 dv1 ∂η 1 dv1
= = √
∂x2 dη ∂x2 2 νt dη
     
∂ 2 v1 ∂ ∂v1 ∂ 1 dv1 1 ∂ dv1 1 d2 v1
2 = = √ = √ =
∂x2 ∂x2 ∂x2 ∂x2 2 νt dη 2 νt ∂x2 dη 4νt dη 2
(3.4)

We introduce a non-dimensional velocity

v1
f= (3.5)
V0
Inserting Eqs. 3.4 and 3.5 in Eq. 3.1 gives

d2 f df
+ 2η =0 (3.6)
dη 2 dη
We have now successfully transformed Eq. 3.1 and reduced the number of independent
variables from two to one. Now let us find out if the boundary conditions, Eq. 3.2, also
can be transformed in a physically meaningful way; we get

v1 (x2 , t = 0) = 0 ⇒ f (η → ∞) = 0
v1 (x2 = 0, t) = V0 ⇒ f (η = 0) = 1 (3.7)
v1 (x2 → ∞, t) = 0 ⇒ f (η → ∞) = 0

Since we managed to transform both the equation (Eq. 3.1) and the boundary conditions
(Eq. 3.7) we conclude that the transformation is suitable.
Now let us solve Eq. 3.6. Integration once gives
df
= C1 exp(−η 2 ) (3.8)
dη
Integration a second time gives
Z η
f = C1 exp(−η ′2 )dη ′ + C2 (3.9)
0

The integral above is the error function

Z η
2
erf(η) ≡ √ exp(−η ′2 )dη ′ (3.10)
π 0

At the limits, the error function takes the values 0 and 1, i.e. erf(0) = 0 and erf(η →
∞) = 1. Taking into account the boundary √ conditions, Eq. 3.7, the final solution to
Eq. 3.9 is (with C2 = 1 and C1 = −2/ π)

f (η) = 1 − erf(η) (3.11)

3.1. The Rayleigh problem 32

2.5

η 1.5

0.5

0
0 0.2 0.4 0.6 0.8 1
f
Figure 3.3: The velocity, f = v1 /V0 , given by Eq. 3.11.

2.5

η 1.5

0.5

0
−6 −5 −4 −3 −2 −1 0
τ21 /(ρV0 )
Figure 3.4: The shear stress for water (ν = 10−6 ) obtained from Eq. 3.12 at time
t = 100 000.

The solution is presented in Fig. 3.3. Compare this figure with Fig. 3.2 at p. 30; all
graphs in that figure collapse into one graph in Fig. 3.3. To compute the velocity, v1 ,
we pick a time t and insert x2 and t in Eq. 3.3. Then f is obtained from Eq. 3.11 and
the velocity, v1 , is computed from Eq. 3.5. This is how the graphs in Fig. 3.2 were
obtained.
From the velocity profile we can get the shear stress as
∂v1 µV0 df µV0

τ21 = µ = √ = −√ exp −η 2 (3.12)
∂x2 2 νt dη πνt
where we used ν = µ/ρ. Figure 3.4 presents the shear stress, τ21 . The solid line is
obtained from Eq. 3.12 and circles are obtained by evaluating the derivative, df /dη,
numerically using central differences (fj+1 − fj−1 )/(ηj+1 − ηj−1 ). As can be seen
from Fig. 3.4, the magnitude of√the shear stress increases for decreasing η and it is
largest at the wall, τw = −ρV0 / πt
The vorticity, ω3 , across the boundary layer is computed from its definition (Eq. 1.36)
∂v1 V0 df V0
ω3 = − =− √ =√ exp(−η 2 ) (3.13)
∂x2 2 νt dη πνt
From Fig. 3.2 at p. 30 it is seen that for large times, the moving plate is felt further
and further out in the flow, i.e. the thickness of the boundary layer, δ, increases. Often
3.2. Flow between two plates 33

P1

P2 h

P1
x2
V
x1

Figure 3.5: Flow in a horizontal channel. The inlet part of the channel is shown.

the boundary layer thickness is defined by the position where the local velocity, v1 (x2 ),
reaches 99% of the freestream velocity. In our case, this corresponds to the point where
v1 = 0.01V0 . Find the point f = v1 /V0 = 0.01 in Fig. 3.3; at this point η ≃ 1.8 (we
can also use Eq. 3.11). Inserting x2 = δ in Eq. 3.3 gives
δ √
η = 1.8 = √ ⇒ δ = 3.6 νt (3.14)
2 νt

It can be seen that the boundary layer thickness increases with t1/2 . Equation 3.14 can
also be used to estimate the diffusion length. After, say, 10 minutes the diffusion length diffusion
for air and water, respectively, are length

δair = 10.8cm
(3.15)
δwater = 2.8cm

As mentioned in the beginning of this section, the diffusion length is determined by

the kinematic viscosity, ν = µ/ρ rather than by the dynamic one, µ. The diffusion
length can also be used to estimate the thickness of a developing boundary layer, see
Section 4.3.1.

Exercise 16 Consider the graphs in Fig. 3.3. Create this graph with Matlab.

Exercise 17 Consider the graphs in Fig. 3.2. Note that no scale is used on the x2 axis
and that no numbers are given for t1 , t2 and t3 . Create this graph with Matlab for both
air and engine oil. Choose suitable values on t1 , t2 and t3 .

Exercise 18 Repeat the exercise above for the shear stress, τ21 , see Fig. 3.4.

3.2 Flow between two plates

Consider steady, incompressible flow in a two-dimensional channel, see Fig. 3.5, with
constant physical properties (i.e. µ = const).

3.2.1 Curved plates

Provided that the walls at the inlet are well curved, the velocity near the walls is larger
than in the center, see Fig. 3.5. The reason is that the flow (with velocity V ) following
the curved wall must change its direction. The physical agent which accomplish this
3.2. Flow between two plates 34

P2

P1

x2

V x1

Figure 3.6: Flow in a channel bend.

is the pressure gradient which forces the flow to follow the wall as closely as possible
(if the wall is not sufficiently curved a separation will take place). Hence the pressure
in the center of the channel, P2 , is higher than the pressure near the wall, P1 . It is thus
easier (i.e. less opposing pressure) for the fluid to enter the channel near the walls than
in the center. This explains the high velocity near the walls.
The same phenomenon occurs in a channel bend, see Fig. 3.6. The flow V ap-
proaches the bend and the flow feels that it is approaching a bend through an increased
pressure. The pressure near the outer wall, P2 , must be higher than that near the inner
wall, P1 , in order to force the flow to turn. Hence, it is easier for the flow to sneak
along the inner wall where the opposing pressure is smaller than near the outer wall:
the result is a higher velocity near the inner wall than near the outer wall. In a three-
dimensional duct or in a pipe, the pressure difference P2 − P1 creates secondary flow
downstream the bend (i.e. a swirling motion in the x2 − x3 plane).

3.2.2 Flat plates

The flow in the inlet section (Fig. 3.5) is two dimensional. Near the inlet the velocity is
largest near the wall and further downstream the velocity is retarded near the walls due
to the large viscous shear stresses there. The flow is accelerated in the center because
the integrated mass flow (from x2 = 0 to h) at each x1 must be constant because of
continuity. The acceleration and retardation of the flow in the inlet region is “paid for ”
by a pressure loss which is rather high in the inlet region; if a separation occurs because
of sharp corners at the inlet, the pressure loss will be even higher. For large x1 the flow
will be fully developed; the region until this occurs is called the entrance region, and
the entrance length can, for moderately disturbed inflow, be estimated as [4]
x1,e V Dh
= 0.016ReDh ≡ 0.016 (3.16)
Dh ν
where V denotes the bulk (i.e. the mean) velocity, and Dh = 4A/Sp where Dh ,
A and Sp denote the hydraulic diameter, the cross-sectional area and the perimeter,
respectively. For flow between two plates we get Dh = 2h.
Let us find the governing equations for the fully developed flow region; in this
region the flow does not change with respect to the streamwise coordinate, x1 (i.e.
∂v1 /∂x1 = ∂v2 /∂x1 = 0). Since the flow is two-dimensional, it does not depend
on the third coordinate direction, x3 (i.e. ∂/∂x3 ), and the velocity in this direction is
zero, i.e. v3 = 0. Taking these restrictions into account the continuity equation can be
3.2. Flow between two plates 35

simplified as (see Eq. 2.3)

∂v2
=0 (3.17)
∂x2
Integration gives v2 = C1 and since v2 = 0 at the walls, it means that

v2 = 0 (3.18)

across the entire channel (recall that we are dealing with the part of the channel where
the flow is fully developed; in the inlet section v2 6= 0, see Fig. 3.5).
Now let us turn our attention to the momentum equation for v2 . This is the vertical
direction (x2 is positive upwards, see Fig. 3.5). The gravity acts in the negative x2
direction, i.e. fi = (0, −g, 0). The momentum equation can be written (see Eq. 2.8 at
p. 24)
dv2 ∂v2 ∂v2 ∂P ∂ 2 v2
ρ ≡ ρv1 + ρv2 =− + µ 2 − ρg (3.19)
dt ∂x1 ∂x2 ∂x2 ∂x2
Since v2 = 0 we get
∂P
= −ρg (3.20)
∂x2
Integration gives
P = −ρgx2 + C1 (x1 ) (3.21)
where the integration “constant” C1 may be a function of x1 but not of x2 . If we denote
the pressure at the lower wall (i.e. at x2 = 0) as p we get

P = −ρgx2 + p(x1 ) (3.22)

Hence the pressure, P , decreases with vertical height. This agrees with our experience
that the pressure decreases at high altitudes in the atmosphere and increases the deeper
we dive into the sea. Usually the hydrodynamic pressure, p, is used in incompressible hydrodynamic
flow. This pressure is zero when the flow is static, i.e. when the velocity field is zero. pressure
However, when you want the physical pressure, the ρgx2 as well as the surrounding
atmospheric pressure must be added.
We can now formulate the momentum equation in the streamwise direction

dv1 ∂v1 ∂v1 dp ∂ 2 v1

ρ ≡ ρv1 + ρv2 =− +µ 2 (3.23)
dt ∂x1 ∂x2 dx1 ∂x2

where P was replaced by p using Eq. 3.22. Since v2 = ∂v1 /∂x1 = 0 the left side is
zero so
∂ 2 v1 dp
µ 2 = (3.24)
∂x2 dx1
Since the left side is a function of x2 and the right side is a function of x1 , we conclude
that they both are equal to a constant (i.e. Eq. 3.24 is independent of x1 and x2 ) . The
velocity, v1 , is zero at the walls, i.e

v1 (0) = v1 (h) = 0 (3.25)

where h denotes the height of the channel, see Fig. 3.5. Integrating Eq. 3.24 twice and
using Eq. 3.25 gives
h dp  x2 
v1 = − x2 1 − (3.26)
2µ dx1 h
3.2. Flow between two plates 36

0.8

0.6
x2 /h
0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
v1 /v1,max
Figure 3.7: The velocity profile in fully developed channel flow, Eq. 3.28.

The minus sign on the right side appears because the pressure is decreasing for increas-
ing x1 ; the pressure is driving the flow. The negative pressure gradient is constant (see
Eq. 3.24) and can be written as −dp/dx1 = ∆p/L.
The velocity takes its maximum in the center, i.e. for x2 = h/2, and reads
 
h ∆p h 1 h2 ∆p
v1,max = 1− = (3.27)
2µ L 2 2 8µ L

We often write Eq. 3.26 on the form

v1 4x2  x2 
= 1− (3.28)
v1,max h h

The mean velocity (often called the bulk velocity) is obtained by integrating Eq. 3.28
across the channel, i.e.
Z 
v1,max h x2  2
v1,mean = 4x2 1 − dx2 = v1,max (3.29)
h 0 h 3

The velocity profile is shown in Fig. 3.7

Since we know the velocity profile, we can compute the wall shear stress. Equa-
tion 3.26 gives
∂v1 h dp h ∆p
τw = µ =− = (3.30)
∂x2 2 dx1 2 L
Actually, this result could have been obtained by simply taking a force balance of a
slice of the flow far downstream.
This flow is analyzed in Appendix B.

3.2.3 Force balance

We continue to consider fully developed flow between two parallel plates. To formulate
a force balance in the x1 direction, we start with Eq. 1.7 which reads for i = 1

dv1 ∂σj1
ρ = (3.31)
dt ∂xj
3.2. Flow between two plates 37

τw,U

p1 p2
V h
walls

x2
τw,L
x1

Figure 3.8: Force balance of the flow between two plates.

The left hand side is zero since the flow is fully developed. Forces act on a volume and
its bounding surface. Hence we integrate Eq. 3.31 over the volume of a slice (length
L), see Fig. 3.8 Z
∂σj1
0= dV (3.32)
V ∂xj
Recall that this is the form on which we originally derived the momentum balance
(Newton’s second law) in Part I. [2] Now use Gauss divergence theorem
Z Z
∂σj1
0= dV = σj1 nj dS (3.33)
V ∂xj S

The bounding surface consists in our case of four surfaces (lower, upper, left and right)
so that
Z Z Z Z
0= σj1 nj dS + σj1 nj dS + σj1 nj dS + σj1 nj dS (3.34)
Slef t Sright Slower Supper

The normal vector on the lower, upper, left and right are ni,lower = (0, −1, 0), ni,upper =
(0, 1, 0), ni,lef t = (−1, 0, 0), ni,right = (1, 0, 0). Inserting the normal vectors and us-
ing Eq. 1.9 give
Z Z Z Z
0=− (−p + τ11 )dS + (−p + τ11 )dS − τ21 dS + τ21 dS
Slef t Sright Slower Supper
(3.35)
τ11 = 0 because ∂v1 /∂x1 = 0 (fully developed flow). The shear stress at the upper and
lower surfaces, τ21 , have opposite sign because µ(∂v1 /∂x2 )lower = −µ(∂v1 /∂x2 )upper .
Using this and Eq. 3.22 give (p = p(x1 ) and τw is constant and can thus be taken out
in front of the integration)
0 = p1 W h − p2 W h − 2τw LW (3.36)
where τw = µ(∂v1 /∂x2 )lower and W is the width (in x3 direction) of the two plates
(for convenience we set W = 1). With ∆p = p1 − p2 we get Eq. 3.30.

3.2.4 Balance equation for the kinetic energy

In this subsection we will use the equation for kinetic energy, Eq. 2.21. Let us integrate
this equation in the same way as we did for the force balance. The left side of Eq. 2.21
3.3. Two-dimensional boundary layer flow over flat plate 38

is zero because we assume that the flow is fully developed; using Eq. 1.9 gives
∂vi σji ∂vi
0= − σji + ρvi fi
∂xj ∂xj
=0
∂vj p ∂vi τji ∂vi ∂vi (3.37)
=− + + pδij − τji
∂xj ∂xj ∂xj ∂xj
Φ

On the first line vi fi = v1 f1 + v2 f2 = 0 because v2 = f1 = 0. The third term on

the second line pδij ∂vi /∂xj = p∂vi /∂xi = 0 because of continuity. The last term
corresponds to the viscous dissipation term, Φ (i.e. loss due to friction), see Eq. 2.23
(term b). Now we integrate the equation over a volume
Z  
∂pvj ∂τji vi
0= − + − Φ dV (3.38)
V ∂xj ∂xj
Gauss divergence theorem on the two first terms gives
Z Z
0 = (−pvj + τji vi )nj dS − ΦdV (3.39)
S V

where S is the surface bounding the volume. The unit normal vector is denoted by nj
which points out from the volume. For example, on the right surface in Fig. 3.8 it is
nj = (1, 0, 0) and on the lower surface it is nj = (0, −1, 0). Now we apply Eq. 3.39
to the fluid enclosed by the flat plates in Fig. 3.8. The second term is zero on all
four surfaces and the first term is zero on the lower and upper surfaces (see Exercises
below). We replace the pressure P with p using Eq. 3.22 so that
Z Z
(−pv1 + ρgx2 v1 )n1 dS = −(p2 − p1 ) v1 n1 dS
Slef t &Sright Slef t &Sright

= ∆pv1,mean W h
because ρgx2 n1 v1 on the left and right surfaces cancels; p can be taken out of the
integral as it does not depend on x2 . Finally we get
Z
1
∆p = ΦdV (3.40)
W hv1,mean V

3.3 Two-dimensional boundary layer flow over flat plate

The equations for steady, two-dimensional, incompressible boundary layer flow reads
(x1 and x2 denote streamwise and wall-normal coordinates, respectively)
∂v1 ∂v1 ∂ 2 v1
v1 + v2 =ν
∂x1 ∂x2 ∂x22
∂p
=0 (3.41)
∂x2
∂v1 ∂v2
+ =0
∂x1 ∂x2
where the pressure gradient is omitted in the v1 momentum equation because ∂p/∂x =
0 along a flat plate in infinite surroundings. The boundary conditions are
x2 = 0 : v1 = v2 = 0 (at the wall)
(3.42)
x2 → ∞ : v1 → V1,∞ , v2 = 0 (far from the wall)
3.3. Two-dimensional boundary layer flow over flat plate 39

Let’s introduce the streamfunction Ψ, which is useful when re-writing the two- stream-
dimensional Navier-Stokes equations. It is defined as function

∂Ψ ∂Ψ
v1 = , v2 = − (3.43)
∂x2 ∂x1
With the velocity field expressed in Ψ, the continuity equations is automatically satis-
fied which is easily shown by inserting Eq. 3.43 into the continuity equation

∂v1 ∂v2 ∂2Ψ ∂2Ψ

+ = − =0 (3.44)
∂x1 ∂x2 ∂x1 ∂x2 ∂x2 ∂x1
Inserting Eq. 3.43 into the streamwise momentum equation gives

∂Ψ ∂ 2 Ψ ∂Ψ ∂ 2 Ψ ∂3Ψ
− 2 =ν 3 (3.45)
∂x2 ∂x1 ∂x2 ∂x1 ∂x2 ∂x2

The boundary conditions for the streamfunction read

∂Ψ
x2 = 0 : Ψ = = 0 (at the wall)
∂x2
(3.46)
∂Ψ
x2 → ∞ : → V1,∞ (far from the wall)
∂x2
As in Section 3.1 we want to transform the partial differential equation, Eq. 3.45,
into an ordinary differential equation. In Section 3.1 we replaced x1 and t with the new
non-dimensional variable η. Now we want to replace x1 and x2 with a new dimension-
less variable, say ξ. At the same time we define a new dimensionless streamfunction,
g(ξ), as
 1/2
V1,∞ 1/2
ξ= x2 , Ψ = (νV1,∞ x1 ) g (3.47)
νx1
First we need the derivatives ∂ξ/∂x1 and ∂ξ/∂x2
 1/2
∂ξ 1 V1,∞ x2 ξ
=− =−
∂x1 2 νx1 x1 2x1
 1/2 (3.48)
∂ξ V1,∞ ξ
= =
∂x2 νx1 x2

Now we express the first derivatives of Ψ in Eq. 3.45 as derivatives of g, i.e. (g ′

denotes dg/dξ)

∂Ψ ∂  1/2

1/2 ∂ξ
= (νV1,∞ x1 ) g + (νV1,∞ x1 ) g ′
∂x1 ∂x1 ∂x1
 1/2
1 νV1,∞ 1/2 ξ
= g − (νV1,∞ x1 ) g ′
2 x1 2x1
 1/2 (3.49)
1 νV1,∞
= (g − ξg ′ )
2 x1
∂Ψ ∂   ∂ξ ′
= (νV1,∞ x1 )1/2 g + (νV1,∞ x1 )1/2 g = V1,∞ g ′
∂x2 ∂x2 ∂x2
3.3. Two-dimensional boundary layer flow over flat plate 40

The second and third derivatives of Ψ read

 1/2
∂2Ψ ′′ ∂ξ V1,∞ ξ
= V1,∞ g = V1,∞ g ′′ = V1,∞ g ′′
∂x22 ∂x2 νx1 x2
 1/2  2
∂3Ψ V1,∞ ′′′ ∂ξ V1,∞ ′′′ ξ
= V1,∞ g = V1,∞ g = V 1,∞ g ′′′ (3.50)
∂x32 νx1 ∂x2 νx1 x2
∂2Ψ ∂ξ ξ
= V1,∞ g ′′ =− V1,∞ g ′′
∂x1 ∂x2 ∂x1 2x1
Inserting Eqs. 3.49 and 3.50 into Eq. 3.45 gives
 1/2 !  1/2
′ξ 1 νV1,∞ V1,∞
−V1,∞ g V1,∞ g ′′ − ′
(g − ξg ) V1,∞ g ′′
2x1 2 x1 νx1
(3.51)
2
V1,∞
=ν g ′′′
νx1
2
Divide by V1,∞ and multiply by x1 gives

ξ 1
− g ′ g ′′ − (g − ξg ′ ) g ′′ = g ′′′ (3.52)
2 2
so that
1 ′′
gg + g ′′′ = 0 (3.53)
2
This equation was derived (and solved numerically!) by Blasius in his PhD thesis
1907 [5, 6]. The numerical solution is given in Table 3.1. The flow is analyzed in
Appendix C.

Exercise 19 For the fully developed flow, compute the vorticity, ωi , using the exact
solution (Eq. 3.28).

Exercise 20 Show that the first and second terms in Eq. 3.39 are zero on the upper and
the lower surfaces in Fig. 3.8.

Exercise 21 Show that the second term in Eq. 3.39 is zero also on the left and right
surfaces in Fig. 3.8 (assume fully developed flow).

Exercise 22 Using the exact solution, compute the dissipation, Φ, for the fully devel-
oped flow.

Exercise 23 From the dissipation, compute the pressure drop. Is it the same as that
obtained from the force balance (if not, find the error; it should be!).
3.3. Two-dimensional boundary layer flow over flat plate 41

ξ g g′ g′′
0 0.000000000E+00 0.000000000E+00 3.320573362E-01
0.2 6.640999715E-03 6.640779210E-02 3.319838371E-01
0.4 2.655988402E-02 1.327641608E-01 3.314698442E-01
0.6 5.973463750E-02 1.989372524E-01 3.300791276E-01
0.8 1.061082208E-01 2.647091387E-01 3.273892701E-01
1.0 1.655717258E-01 3.297800312E-01 3.230071167E-01
1.2 2.379487173E-01 3.937761044E-01 3.165891911E-01
1.4 3.229815738E-01 4.562617647E-01 3.078653918E-01
1.6 4.203207655E-01 5.167567844E-01 2.966634615E-01
1.8 5.295180377E-01 5.747581439E-01 2.829310173E-01
2.0 6.500243699E-01 6.297657365E-01 2.667515457E-01
2.2 7.811933370E-01 6.813103772E-01 2.483509132E-01
2.4 9.222901256E-01 7.289819351E-01 2.280917607E-01
2.6 1.072505977E+00 7.724550211E-01 2.064546268E-01
2.8 1.230977302E+00 8.115096232E-01 1.840065939E-01
3.0 1.396808231E+00 8.460444437E-01 1.613603195E-01
3.2 1.569094960E+00 8.760814552E-01 1.391280556E-01
3.4 1.746950094E+00 9.017612214E-01 1.178762461E-01
3.6 1.929525170E+00 9.233296659E-01 9.808627878E-02
3.8 2.116029817E+00 9.411179967E-01 8.012591814E-02
4.0 2.305746418E+00 9.555182298E-01 6.423412109E-02
4.2 2.498039663E+00 9.669570738E-01 5.051974749E-02
4.4 2.692360938E+00 9.758708321E-01 3.897261085E-02
4.6 2.888247990E+00 9.826835008E-01 2.948377201E-02
4.8 3.085320655E+00 9.877895262E-01 2.187118635E-02
5.0 3.283273665E+00 9.915419002E-01 1.590679869E-02
5.2 3.481867612E+00 9.942455354E-01 1.134178897E-02
5.4 3.680919063E+00 9.961553040E-0 1 7.927659815E-03
5.6 3.880290678E+00 9.974777682E-0 1 5.431957680E-03
5.8 4.079881939E+00 9.983754937E-0 1 3.648413667E-03
6.0 4.279620923E+00 9.989728724E-01 2.402039844E-03
6.2 4.479457297E+00 9.993625417E-01 1.550170691E-03
6.4 4.679356615E+00 9.996117017E-01 9.806151170E-04
6.6 4.879295811E+00 9.997678702E-01 6.080442648E-04
6.8 5.079259772E+00 9.998638190E-01 3.695625701E-04
7.0 5.279238811E+00 9.999216041E-01 2.201689553E-04
7.2 5.479226847E+00 9.999557173E-01 1.285698072E-04
7.4 5.679220147E+00 9.999754577E-01 7.359298339E-05
7.6 5.879216466E+00 9.999866551E-01 4.129031111E-05
7.8 6.079214481E+00 9.999928812E-01 2.270775140E-05
8.0 6.279213431E+00 9.999962745E-01 1.224092624E-05
8.2 6.479212887E+00 9.999980875E-01 6.467978611E-06
8.4 6.679212609E+00 9.999990369E-01 3.349939753E-06
8.6 6.879212471E+00 9.999995242E-01 1.700667989E-06
8.8 7.079212403E+00 9.999997695E-01 8.462841214E-07

Table 3.1: Blasius numerical solution of laminar flow along a flat plate.
 4. Vorticity equation and potential flow 42

τ21 (x2 + 0.5∆x2 )n2

v1 (x2 )
τ12 (x1 + 0.5∆x1 )n1

P (x1 − 0.5∆x1 ) (x1 , x2 ) P (x1 + 0.5∆x1 )

τ12 (x1 − 0.5∆x1 )n1

τ21 (x2 − 0.5∆x2 )n2

x2

x1

Figure 4.1: Surface forces acting on a fluid particle. The fluid particle is located in the
lower half of fully developed channel flow. The v1 velocity is given by Eq. 3.28 and
v2 = 0. Hence τ11 = τ22 = ∂τ12 /∂x1 = 0 and ∂τ21 /∂x2 > 0. The v1 velocity field
is indicated by dashed vectors.

4 Vorticity equation and potential flow

4.1 Vorticity and rotation
ORTICITY , ωi , was introduced in Eq. 1.12 at p. 15. As shown in Fig. 1.4 at p. 18,
V vorticity is connected to rotation of a fluid particle. Figure 4.1 shows the surface
forces acting on a fluid particle in a shear flow. Looking at Fig. 4.1 it is obvious that
only the shear stresses are able to rotate the fluid particle; the pressure acts through the
center of the fluid particle and is thus not able to affect rotation of the fluid particle.
Note that the v2 momentum equation (see Eqs. 2.4 and 3.32) requires that the vertical
viscous stresses in Fig. 4.1 are in balance. The v1 momentum equation requires that
the horizontal viscous stresses balance the pressure difference.
Let us have a look at the momentum equations in order to show that the viscous
terms indeed can be formulated with the vorticity vector, ωi . In incompressible flow
the viscous terms read (see Eqs. 2.4, 2.5 and 2.6)

∂τji ∂ 2 vi
=µ (4.1)
∂xj ∂xj ∂xj

The right side can be re-written using the tensor identity

 2 
∂ 2 vi ∂ 2 vj ∂ vj ∂ 2 vi
= − −
∂xj ∂xj ∂xj ∂xi ∂xj ∂xi ∂xj ∂xj
 
∂ ∂vj ∂ 2 vk ∂ 2 vk (4.2)
= −εinm εmjk = −εinm εmjk
∂xi ∂xj ∂xj ∂xn ∂xj ∂xn
=0
4.2. The vorticity transport equation in three dimensions 43

where the first on the the second line is zero because of continuity. Let’s verify that
 2 
∂ vj ∂ 2 vi ∂ 2 vk
− = εinm εmjk (4.3)
∂xj ∂xi ∂xj ∂xj ∂xj ∂xn

Use the ε − δ-identity (see Table A.1 at p. 248)

∂ 2 vk ∂ 2 vk ∂ 2 vk ∂ 2 vi
εinm εmjk = (δij δnk − δik δnj ) = − (4.4)
∂xj ∂xn ∂xj ∂xn ∂xi ∂xk ∂xj ∂xj

which shows that Eq. 4.3 is correct. At the right side of Eq. 4.3 we recognize the
vorticity, ωm = εmjk ∂vk /∂xj , so that

∂ 2 vi ∂ωm
= −εinm (4.5)
∂xj ∂xj ∂xn

In vector notation the identity Eq. 4.5 reads

∇2 v = ∇(∇ · v) − ∇ × ∇ × v = −∇ × ω (4.6)

Using Eq. 4.5, Eq. 4.1 reads

∂τji ∂ωm
= −µεinm (4.7)
∂xj ∂xn

Thus, there is a one-to-one relation between the viscous term and vorticity: no viscous
terms means no vorticity and vice versa. An imbalance in shear stresses (left side of
Eq. 4.7) causes a change in vorticity, i.e. generates vorticity (right side of Eq. 4.7).
Hence, inviscid flow (i.e. friction-less flow) has no rotation. (The exception is when
vorticity is transported into an inviscid region, but also in that case no vorticity is
generated or destroyed: it stays constant, unaffected.) Inviscid flow is often called
irrotational flow (i.e. no rotation) or potential flow. The vorticity is always created at potential
boundaries, see Section 4.3.1.
The main points that we have learnt in this section are:

1. The viscous terms are responsible for creating vorticity; this means that the vor-
ticity can not be created or destroyed in inviscid (friction-less) flow
2. The viscous terms in the momentum equations can be expressed in ωi ; consider-
ing Item 1 this was to be expected.

Exercise 24 Prove the first equality of Eq. 4.5 using the ε-δ-identity.

Exercise 25 Write out Eq. 4.7 for i = 1 and verify that it is satisfied.

4.2 The vorticity transport equation in three dimensions

Up to now we have talked quite a lot about vorticity. We have learnt that physically
it means rotation of a fluid particle and that it is only the viscous terms that can cause
rotation of a fluid particle. The terms inviscid (no friction), irrotational and potential
flow all denote frictionless flow which is equivalent to zero vorticity. There is a small friction-
difference between the three terms because there may be vorticity in inviscid flow that less
is convected into the flow at the inlet(s); but also in this case the vorticity is not affected
4.2. The vorticity transport equation in three dimensions 44

once it has entered the inviscid flow region. However, usually no distinction is made
between the three terms.
In this section we will derive the transport equation for vorticity in incompressible
flow. As usual we start with the Navier-Stokes equation, Eq. 2.8 at p.24. First, we
re-write the convective term of the incompressible momentum equation (Eq. 2.8) as
 
∂vi 1
vj = vj (Sij + Ωij ) = vj Sij − εijk ωk (4.8)
∂xj 2
where Eq. 1.19 on p. 16 was used. Inserting Sij = (∂vi /∂xj + ∂vj /∂xi )/2 and
multiplying by two gives
 
∂vi ∂vi ∂vj
2vj = vj + − εijk vj ωk (4.9)
∂xj ∂xj ∂xi
The second term on the right side can be written as (Trick 2, see Eq. 8.4)
∂vj 1 ∂(vj vj ) ∂k
vj = = (4.10)
∂xi 2 ∂xi ∂xi
where k = vj vj /2. Equation 4.9 can now be written as
∂vi ∂k
vj = − εijk vj ωk (4.11)
∂xj ∂xi
rotation
no rotation

The last term on the right side is the vector product of v and ω, i.e. v × ω.
The trick we have achieved is to split the convective term into one term without
rotation (first term on the right side of Eq. 4.11) and one term including rotation (second
term on the right side). Inserting Eq. 4.11 into the incompressible momentum equation
(Eq. 2.8) yields
∂vi ∂k 1 ∂P ∂ 2 vi
+ − εijk vj ωk = − +ν + fi (4.12)
∂t ∂xi ρ ∂xi ∂xj ∂xj
rotation
no rotation

The volume source is in most engineering flows represented by the gravity, i.e. fi = gi .
From Eq. 4.12 we get Crocco’s theorem for steady inviscid flow
   
∂ P ∂ P
εijk vj ωk = + k − fi = +k+φ (4.13)
∂xi ρ ∂xi ρ
P0 /ρ

where ∂φ/∂xi = −fi is the potential of the body force. In vector notation, Eq. 4.13
reads
1
v × ω = ∇(P0 ) (4.14)
ρ
These equations states that the gradient of stagnation pressure, P0 , is orthogonal to
both the velocity and vorticity vector.
Since the vorticity vector in Eq. 4.12 is defined by the cross product εpqi ∂vi /∂xq
(∇ × v in vector notation, see Exercise 8), we start by applying the operator εpqi ∂/∂xq
to the Navier-Stokes equation (Eq. 4.12) so that
∂ 2 vi ∂2k ∂vj ωk
εpqi + εpqi − εpqi εijk
∂t∂xq ∂xi ∂xq ∂xq
(4.15)
1 ∂2P ∂ 3 vi ∂gi
= −εpqi + νεpqi + εpqi
ρ ∂xi ∂xq ∂xj ∂xj ∂xq ∂xq
4.2. The vorticity transport equation in three dimensions 45

where the body force fi was replaced by gi . We know that εijk is anti-symmetric in
all indices, and hence the second term on line 1 and the first term on line 2 are zero
(product of a symmetric and an anti-symmetric tensor). The last term on line 2 is zero
because the gravitation vector, gi , is constant. The last term on line 1 is re-written using
the ε-δ identity (see Table A.1 at p. 248)
∂vj ωk ∂vj ωk ∂vp ωk ∂vq ωp
εpqi εijk = (δpj δqk − δpk δqj ) = −
∂xq ∂xq ∂xk ∂xq
(4.16)
∂ωk ∂vp ∂ωp ∂vq
= vp + ωk − vq − ωp
∂xk ∂xk ∂xq ∂xq

Using the definition of ωi we find that its divergence

 
∂ωi ∂ ∂vk ∂ 2 vk
= εijk = εijk =0 (4.17)
∂xi ∂xi ∂xj ∂xj ∂xi

is zero (product of a symmetric and an anti-symmetric tensor). Using the continuity

equation (∂vq /∂xq = 0) and Eq. 4.17, Eq. 4.16 can be written

∂vj ωk ∂vp ∂ωp

εpqi εijk = ωk − vk (4.18)
∂xq ∂xk ∂xk

The second term on line 2 in Eq. 4.15 can be written as

 
∂ 3 vi ∂2 ∂vi ∂ 2 ωp
νεpqi =ν εpqi =ν (4.19)
∂xj ∂xj ∂xq ∂xj ∂xj ∂xq ∂xj ∂xj

Inserting Eqs. 4.18 and 4.19 into Eq. 4.15 gives finally

dωp ∂ωp ∂ωp ∂vp ∂ 2 ωp

≡ + vk = ωk +ν (4.20)
dt ∂t ∂xk ∂xk ∂xj ∂xj

We recognize the usual unsteady term, the convective term and the diffusive term.
Furthermore, we have got rid of the pressure gradient term. That makes sense, because
as mentioned in connection to Fig. 4.1, the pressure cannot affect the rotation (i.e. the
vorticity) of a fluid particle since the pressure acts through its center. Equation 4.20
has a new term on the right-hand side which represents amplification and bending or
tilting of the vorticity lines. If we write it term-by-term it reads

 ∂v1 ∂v1 ∂v1

 ω1 +ω2 + ω3 , p=1

 ∂x ∂x ∂x
∂vp  ∂v2
1
∂v2
2
∂v2
3
ωk = ω1 +ω2 + ω3 , p=2 (4.21)
∂xk 
 ∂x1 ∂x2 ∂x3

 ∂v ∂v ∂v

 ω1 3 +ω2 3 + ω3 3 , p=3
∂x1 ∂x2 ∂x3
The diagonal terms in this matrix represent vortex stretching. Imagine a slender, Vortex
cylindrical fluid particle with vorticity ωi and introduce a cylindrical coordinate system stretching
with the x1 -axis as the cylinder axis and r2 as the radial coordinate (see Fig. 4.2) so
that ωi = (ω1 , 0, 0). We assume that a positive ∂v1 /∂x1 is acting on the fluid cylinder;
it will act as a source in Eq. 4.20 increasing ω1 and it will stretch the cylinder. The vol-
ume of the fluid element must stay constant during the stretching (the incompressible
continuity equation), which means that the radius of the cylinder will decrease. Hence
4.2. The vorticity transport equation in three dimensions 46

v1 ω1 v1

x2

x1

Figure 4.2: Vortex stretching. Dashed lines denote fluid element before stretching.
∂v1
> 0.
∂x1

ω2 ∂v1
v1 (x2 ) ∂x2 >0

x2

x1

Figure 4.3: Vortex tilting. Dashed lines denote fluid element before bending or tilting.

vortex stretching will either make a fluid element longer and thinner (as in the example
above) or shorter and thicker (when ∂v1 /∂x1 < 0).
The off-diagonal terms in Eq. 4.21 represent vortex tilting. Again, take a slender Vortex
fluid particle, but this time with its axis aligned with the x2 axis, see Fig. 4.3. Assume tilting
is has a vorticity, ω2 , and that the velocity surrounding velocity field is v1 = v1 (x2 ).
The velocity gradient ∂v1 /∂x2 will tilt the fluid particle so that it rotates in clock-wise
direction. The second term ω2 ∂v1 /∂x2 in line one in Eq. 4.21 gives a contribution to
ω1 . This means that vorticity in the x2 direction, through the source term ω2 ∂v1 /∂x2 ,
creates vorticity in the x1 direction..
Vortex stretching and tilting are physical phenomena which act in three dimensions:
fluid which initially is two dimensional becomes quickly three dimensional through
these phenomena. Vorticity is useful when explaining why turbulence must be three-
dimensional, see Section 5.4.
4.3. The vorticity transport equation in two dimensions 47

4.3 The vorticity transport equation in two dimensions

It is obvious that the vortex stretching/tilting has no influence in two dimensions; in
this case the vortex stretching/tilting term vanishes because the vorticity vector is or-
thogonal to the velocity vector (for a 2D flow the velocity vector reads vi = (v1 , v2 , 0)
and the vorticity vector reads ωi = (0, 0, ω3 ) so that the vector ωk ∂vp /∂xk = 0). Thus
in two dimensions the vorticity equation reads

dω3 ∂ 2 ω3
=ν (4.22)
dt ∂xα ∂xα
(Greek indices are used to indicate that they take values 1 or 2). This equation is exactly
the same as the transport equation for temperature in incompressible flow, see Eq. 2.16.
This means that vorticity convects and diffuses in the same way as temperature does. In
fully developed channel flow, for example, the vorticity and the temperature equations
reduce to
∂ 2 ω3
0=ν (4.23a)
∂x22
∂ 2T
0=k 2 (4.23b)
∂x2

For the temperature equation the heat flux is given by q2 = −k∂T /∂x2; with a hot
lower wall and a cold upper wall (constant wall temperatures) the heat flux is constant
for all x2 and goes from the lower wall to the upper wall. We have the same situation for
the vorticity. Its gradient, i.e. the vorticity flux, γ2 = −ν∂ω3 /∂x2 , is constant across
the channel (you have plotted this quantity in TME225 Assignment 1). Equation 4.23
is turned into relations for q2 and γ2 by integration

γwall = γ2 (4.24a)
qwall = q2 (4.24b)

If the wall-normal temperature derivative ∂T /∂x2 = 0 at both walls (adiabatic

walls), the heat flux at the walls, qwall , will be zero and the temperature will be equal to
an arbitrary constant in the entire domain. It is only when the wall-normal temperature
derivative at the walls are non-zero that a temperature field is created in the domain.
The same is true for ω3 : if ν∂ω3 /∂x2 = −γ2 = 0 at the walls, the flow will not include
any vorticity. Hence, vorticity is – in the same way as temperature – generated at the
walls.

4.3.1 Boundary layer thickness from the Rayleigh problem

In Section 3.1 we studied the Rayleigh problem (unsteady diffusion). As shown above,
the two-dimensional unsteady temperature equation is identical to the two-dimensional
unsteady equation for vorticity. The diffusion time, t, or the diffusion length, δ, in
Eq. 3.14 can now be used to estimate the thickness of a developing boundary layer
(recall that the limit between the boundary layer and the outer free-stream region can
be defined by vorticity: inside the vorticity is non-zero and outside it is zero).
In a boundary layer, the streamwise pressure gradient is zero. This means that

∂ 2 v1 
µ 2 =0
∂x2 wall
4.3. The vorticity transport equation in two dimensions 48

V0
x2
1
0
111111111
000000000

0
1
0
1
0
1
0
1
0
1
0
1
0
1
0
1
x1
1111111
0000000
0
1
δ
00000000000000000000000000000000000000000
11111111111111111111111111111111111111111
00000000000000000000000000000000000000000
11111111111111111111111111111111111111111

Figure 4.4: Boundary layer. The boundary layer thickness, δ, increases for increasing
streamwise distance from leading edge (x1 = 0).

because, at the wall, the only non-zero terms in the Navier-Stokes equation are the
streamwise pressure gradient and the wall-normal diffusion term (see, for example,
Eqs. 2.8 and 3.23). Hence, the flux of vorticity
 
∂ω3  ∂ 2 v1 
γ2 = −ν =ν =0 (4.25)
∂x2 wall ∂x22 wall

(recall that (∂v2 /∂x1 )wall = 0) along the wall which means that no vorticity is created
along the boundary. The vorticity in a developing boundary layer is created at the
leading edge of the plate (note that in channel flow, vorticity is indeed created along the
walls because in this case the streamwise pressure gradient is not zero). The vorticity
generated at the leading edge is transported along the wall by convection and at the
same time it is transported by diffusion away from the wall.
Below we will estimate the boundary layer thickness using the expression derived
for the Rayleigh problem. In a boundary layer there is vorticity and outside the bound-
ary layer it is zero (in the Rayleigh flow problem, the vorticity is created at time t = 0+
when the plate instantaneously accelerates from rest to velocity V0 ). Hence, if we can
estimate how far from the wall the vorticity diffuses, this gives us an estimation of the
boundary layer thickness.
Consider the boundary layer in Fig. 4.4. The boundary layer thickness at the end of
the plate is δ(L). The time it takes for a fluid particle to travel from the leading edge of
the plate to x = L is L/V0 (in the Rayleigh problem this corresponds to the flow field
after time t = L/V0 ). During this time vorticity will be transported by diffusion in the
x2 direction the length δ according to Eq. 3.14. If we assume that the fluid is air with
the speed V0 = 3m/s and that the length of the plate L = 2m we get from Eq. 3.14
that δ(L) = 1.2cm.

Exercise 26 Note that the estimate above is not quite accurate because in the Rayleigh
problem we assumed that the convective terms are zero, but in a developing boundary
layer, as in Fig. 4.4, they are not (v2 6= 0 and ∂v1 /∂x1 6= 0). The proper way to
solve the problem is to use Blasius solution, see Section 3.3. Blasius solution gives (see
Eq. C.1)
δ 5 V0 L
= 1/2
, ReL = (4.26)
L Re ν
L

Compute what δ(L) you get from Eq. 4.26.

Exercise 27 Assume that we have a developing flow in a pipe (radius R) or between

two flat plates (separation distance h). We want to find out how long distance it takes
for the the boundary layers to merge. Equation 3.14 can be used with δ = R or h.
Make a comparison with this and Eq. 3.16.
4.4. Potential flow 49

4.4 Potential flow

N potential flow, the velocity vector can be expressed as the gradient of its poten-
I tial Φ, see Eq. 1.22. The vorticity is then zero by definition since the curl of the
divergence is zero. This is easily seen by inserting Eq. 1.22 (vi = ∂Φ/∂xi ) into the
definition of the vorticity, Eq. 1.12, i.e.
∂vk ∂2Φ
ωi = ǫijk = ǫijk =0 (4.27)
∂xj ∂xj ∂xk
since ǫijk is anti-symmetric in indices j and k and ∂ 2 Φ/∂xj ∂xk is symmetric in j and
k. Inserting Eq. 1.22 into the continuity equation, Eq. 2.3, gives
 
∂vi ∂ ∂Φ ∂2Φ
0= = = (4.28)
∂xi ∂xi ∂xi ∂xi ∂xi
i.e. the potential satisfies the Laplace equation. This is of great important since many
analytical methods exist for the Laplace equation.

4.4.1 The Bernoulli equation

The velocity field in potential flow is thus given by the continuity equation, Eq. 4.28
(together with Eq. 1.22). Do we have any use of the Navier-Stokes equation? The
answer is yes: this equation provides the pressure field. We use the Navier-Stokes
equation (Eq. 4.12) with the viscous term expressed as in Eq. 4.5
∂vi ∂k 1 ∂P ∂ωm
+ − εijk vj ωk = − − νεinm + fi (4.29)
∂t ∂xi ρ ∂xi ∂xn
Since ωi = 0 in potential (irrotational) flow, we get (with fi = gi ) and using k =
vi vi /2 = v 2 /2  
∂ ∂Φ 1 ∂v 2 1 ∂P
+ =− + gi (4.30)
∂xi ∂xi 2 ∂xi ρ ∂xi
where vi in the unsteady term was replaced by its potential (Eq. 1.22). The gravity
force can be expressed as a force potential, gi = −∂ Φ̂/∂xi (see Eq. 4.13), because it is
conservative. The gravity force is conservative because when integrating this force, the conservative
work (i.e. the integral) depends only on the starting and ending points of the integral: force
in mathematics this is called an exact differential.
Inserting gi = −∂ Φ̂/∂xi in Eq. 4.30 gives
 
∂ ∂Φ v 2 P
+ + + Φ̂ = 0 (4.31)
∂xi ∂t 2 ρ
Integration gives the famous Bernoulli equation
∂Φ v 2 P
+ + + Φ̂ = C(t) (4.32)
∂t 2 ρ
where Φ̂ = −gi xi . In steady flow, we get
v2 P
+ − g3 x3 = C (4.33)
2 ρ
where gi = (0, 0, g3). Using the height, gh = −g3 x3 , we get the more familiar form
v2 P
+ + gh = C (4.34)
2 ρ
4.4. Potential flow 50

4.4.2 Complex variables for potential solutions of plane flows

Complex analysis is a suitable tool for studying potential flow. We start this section by
repeating some basics of complex analysis. For real functions, the value of a partial
derivative, ∂f /∂x, at x = x0 is defined by making x approach x0 and then evaluating
(f (x+x0 )−f (x))/x0 . The total derivative, df /dt, is defined by approaching the point
x10 , x20 , x30 , t as a linear combination of all independent variables (cf. Eq. 1.1).
A complex derivative of a complex variable is defined as (f (z + z0 ) − f (z))/z0
where z = x+iy and f = u+iv. We can approach the point z0 both in the real coordi-
nate direction, x, and in the imaginary coordinate direction, y. The complex derivative
is defined only if the value of the derivative is independent of how we approach the
point z0 . Hence
df f (z0 + ∆z) − f (z0 )
= lim
dz ∆z→0 ∆z
f (x0 + ∆x, iy0 ) − f (x0 , iy0 ) f (x0 , iy0 + i∆y) − f (x0 , iy0 )
= lim = lim .
∆x→0 ∆x ∆y→0 i∆y
(4.35)
The second line can be written as
∂f 1 ∂f i ∂f ∂f
= = 2 = −i (4.36)
∂x i ∂y i ∂y ∂y
since i2 = −1. Inserting f = u + iv and taking the partial derivative of f we get
∂f ∂u ∂v
= +i
∂x ∂x ∂x
(4.37)
∂f ∂u ∂v
i =i −
∂y ∂y ∂y
Using Eq. 4.36 gives
∂u ∂v ∂u ∂v
= , =− (4.38)
∂x ∂y ∂y ∂x
Equations 4.38 are called the Cauchy-Riemann equations. Another way to derive
Eq. 4.38 is to require that f should depend only on z but not on z̄ [7] (z̄ is the complex
conjugate of z, i.e. z̄ = x − iy).
So far the complex plane has been expressed as z = x+iy. It can also be expressed
in polar coordinates (see Fig. 4.5)

z = reiθ = r(cos θ + i sin θ) (4.39)

Now we return to fluid mechanics and potential flow. Let us introduce a complex
potential, f , based on the streamfunction, Ψ (Eq. 3.43), and the velocity potential, Φ
(Eq. 1.22)
f = Φ + iΨ (4.40)
Recall that for two-dimensional, incompressible flow, the velocity potential satisfies the
Laplace equation, see Eq. 4.28. The streamfunction also satisfies the Laplace equation
in potential flow where the vorticity, ωi , is zero. This is easily seen by taking the
divergence of the streamfunction, Eq. 3.43
∂2Ψ ∂2Ψ ∂v2 ∂v1
+ =− + = −ω3 = 0 (4.41)
∂x21 ∂x22 ∂x1 ∂x2
4.4. Potential flow 51

θ
x

Figure 4.5: The complex plane in polar coordinates. Real and imaginary axes corre-
spond to the horizontal and vertical axes, respectively.

see Eq. 1.13. Hence the complex potential, f , also satisfies the Laplace equation.
Furthermore, f also satisfies the Cauchy-Riemann equations, Eq. 4.38, since
∂Φ ∂Ψ ∂Φ ∂Ψ
= = v1 and =− = v2 (4.42)
∂x ∂y ∂y ∂x
see Eqs. 3.43 and 1.22. Thus we can conclude that f defined as in Eq. 4.40 is differen-
tiable, i.e. df /dz exists.

4.4.3 Analytical solutions as f ∝ z n

Now we will give some examples of f (z) which correspond to useful engineering
flows. We can choose any exponent n and multiply with any constant in order to get a
physical, meaningful flow. The solution
f = C1 z n (4.43)
is one example. Let’s first verify that this is a solution of the Laplace equation. We
write it in polar coordinates

n
f = C1 reiθ = C1 rn einθ = C1 rn (cos(nθ) + i sin(nθ)) (4.44)
The Laplace operator in polar coordinates reads
 
2 1 ∂ ∂f 1 ∂2f
∇ f= r + 2 2 (4.45)
r ∂r ∂r r ∂θ
Taking the first and the second derivatives of Eq. 4.44 with respect to r and θ gives
∂f
= C1 nrn−1 (cos(nθ) + i sin(nθ))
 ∂r

1 ∂ ∂f
r = C1 n2 rn−2 (cos(nθ) + i sin(nθ))
r ∂r ∂r
(4.46)
∂f
= C1 nrn (− sin(nθ) + i cos(nθ))
∂θ
∂2f
= −C1 n2 rn (cos(nθ) + i sin(nθ))
∂θ2
4.4. Potential flow 52

V∞

Figure 4.6: Parallel flow.

When we divide the fourth line with r2 and add it to the second line we find that the
Laplace equation (Eq. 4.45) is indeed zero, i.e.
 
2 1 ∂ ∂f 1 ∂2f
∇ f= r + 2 2 =0 (4.47)
r ∂r ∂r r ∂θ

4.4.3.1 Parallel flow

When we set n = 1 in Eq. 4.43 we get (C1 = V∞ )

f = V∞ z = V∞ (x + iy) (4.48)

The streamfunction, Ψ, is equal to the imaginary part, see Eq. 4.40. Equation 4.42
gives the velocity components
∂Ψ ∂Ψ
v1 = = V∞ and v2 = − =0 (4.49)
∂y ∂x
The flow is shown in Fig. 4.6.

4.4.3.2 Stagnation flow

When we set n = 2 in Eqs. 4.43 and 4.44 we get (inviscid) stagnation flow onto a wall.
The streamfunction, Ψ, corresponds to the imaginary part of f , see Eqs. 4.40 and 4.44,
so that (C1 = 1)

Ψ = r2 sin(2θ) (4.50)
The solution in form of a vector plot and contour plot of the streamfunction is given in
Fig. 4.7. The flow impinges at the wall at x2 = 0. The streamfunction is zero along the
symmetry line, x1 = 0, and it is negative to the left and positive to the right. The polar
velocity components are obtained as (see. Eq. 4.42)
1 ∂Ψ
vr = = 2r cos(2θ)
r ∂θ (4.51)
∂Ψ
vθ = − = −2r sin(2θ)
∂r
4.4. Potential flow 53

1.2 0.8

1 0.6
1
0.4
0.8
0.2
0.6
x2

x2
0.5 0
0.4
−0.2
0.2
−0.4
0
0 −0.6
−0.2 −0.8
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
x1 x1
(a) Vector plot. (b) Streamlines.

Figure 4.7: Potential flow. Stagnation flow.

2.5
2.5
2
2
2

1.5 1.5
1.5
x2

x2

1 1
1

0.5 0.5
α 0.5

0 0
−2 −1 0 1 −1.5 −1 −0.5 0 0.5 1 1.5
x1 x1
(a) Vector plot. (b) Streamlines.

Figure 4.8: Potential flow. The lower boundary for x1 < 0 can either be a wall (concave
corner) or symmetry line (wedge).

and in Cartesian components (see Fig. 1.9)

v1 = vr cos θ − vθ sin θ = 2r cos θ cos(2θ) + 2r sin θ sin(2θ)
= 2r cos θ(1 − 2 sin2 θ) + 4r sin2 θcosθ = 2r cos θ = 2x1
(4.52)
v2 = vr sin θ + vθ cos θ = 2r sin θ cos(2θ) − 2r cos θ sin(2θ)
= 2r sin θ(2 cos2 θ − 1) − 4r sin θ cos2 θ = −2r sin θ = −2x2
Recall that since the flow is inviscid (no friction), the boundary condition on the wall is
slip, i.e. a frictionless wall (same as a symmetric boundary). Note that this flow is the
same as we looked at in Section 1.2 except that the velocities are here twice as large
because we chose C1 = 1.
Figure 4.7 was generated in Matlab by evaluating Eqs. 4.50, 4.51 and 4.52 on a
polar grid.

4.4.3.3 Flow over a wedge and flow in a concave corner

Next we set n = 6/5 in Eqs. 4.43 and 4.44. This gives us (inviscid) flow over a
wedge and flow over a concave corner (n should be in the interval 1 < n < 2). The
streamfunction, the imaginary part of f , is given by (Eqs. 4.40 and 4.44)
Ψ = r6/5 sin(6θ/5) (4.53)
4.4. Potential flow 54

θ
x

Figure 4.9: Line source. ṁ > 0

(C1 = 1) and the velocity components read (see Fig. 4.51)

1 ∂Ψ 6
vr = = r cos(6θ/5)
r ∂θ 5 (4.54)
∂Ψ 6
vθ = − = − r sin(6θ/5)
∂r 5
The velocity vector field and the streamfunction are presented in Fig. 4.8. The stream-
function is zero along the lower boundary. The angle, α, in Fig. 4.8a is given by

(n − 1)π π
α= = (4.55)
n 6

4.4.4 Analytical solutions for a line source

The complex potential for a line source reads
ṁ
f= ln z (4.56)
2π
where ṁ is the strength of the source; the physical meaning of ṁ is volume flow
assuming that the extent of the domain in the third coordinate direction, x3 , is one.
Writing Eq. 4.56 on polar form gives
ṁ
ṁ

ṁ
f= ln reiθ = ln r + ln eiθ = (ln r + iθ) (4.57)
2π 2π 2π
4.4. Potential flow 55

First, we need to make sure that this solution satisfies the Laplace equation, Eq. 4.45.
The first and second derivatives read
∂f ṁ
=
 ∂r  2πr
1 ∂ ∂f
r =0
r ∂r ∂r
(4.58)
∂f ṁ
=i
∂θ 2π
∂2f
=0
∂θ2
which shows that Eq. 4.47 is satisfied.
The streamfunction corresponds to the imaginary part of f and we get (see Eq. 4.51)
1 ∂Ψ ṁ
vr = =
r ∂θ 2πr (4.59)
∂Ψ
vθ = − =0
∂r
We find that the physical flow is in the radial direction, see Fig. 4.9. If ṁ > 0, the flow
is outwards directed and for ṁ < 0 it is going inwards toward origo. When origo is
approached, the velocity, vr , tends to infinity. Hence, Eq. 4.59 gives nonphysical flow
near origo. The reason is that the inviscid assumption (zero viscosity) is not valid in
this region.
It was mentioned above that the physical meaning of ṁ is volume flow. This is
easily seen by integrating vr (Eq. 4.59) over a cylindrical surface as
Z 1 Z 2π Z 1 Z 2π Z Z
ṁ ṁ 1 2π
dx3 vr rdθ = dx3 rdθ = dx3 dθ = ṁ. (4.60)
0 0 0 0 2πr 2π 0 0

4.4.5 Analytical solutions for a vortex line

A line vortex is another example of a complex potential; it is very similar to Eq. 4.56
and reads
Γ
f = −i ln z (4.61)
2π
which on polar form reads (cf. Eq. 4.57)
Γ
f =− (i ln r − θ) (4.62)
2π
From the streamfunction (the imaginary part of f ) we get (cf. Eq. 4.59)
1 ∂Ψ
vr = =0
r ∂θ (4.63)
∂Ψ Γ
vθ = − =
∂r 2πr
This flow was introduced in Section 1.7.1 (where we called it an ideal vortex line) as an
example of a flow with no vorticity. The flow is in the positive θ direction along lines
of constant radius, see Fig. 4.10. The circulation, Γ, appears in the expression of vθ . It
was introduced in Section 1.7. It is defined as a closed line integral along line C, see
Eq. 1.23 and can be expressed as an integral of the vorticity over surface S bounded by
line C, see Eq. 1.25 and Fig. 1.7.
4.4. Potential flow 56

θ
x

Figure 4.10: Vortex line.

x2
V∞ θ
x1

Figure 4.11: Flow around a cylinder of radius r0 .

4.4.6 Analytical solutions for flow around a cylinder

The complex potential for the flow around a cylinder can be found by combining a
doublet and a parallel flow. A doublet consists of a line source (strength ṁ) and sink
(strength −ṁ) separated by a distance ε in the x1 direction (line sources were intro-
duced in Section 4.4.4). Imagine that we move the source and the sink closer to each
other and at the same time we increase their strength |ṁ| so that the product µ = ṁε
stays constant. The resulting complex potential is
µ
f= (4.64)
πz
When adding the complex potential of parallel flow, see Eq. 4.48, we get
 4.4. Potential flow 57

rdθ
x2

θ
x1

Figure 4.12: Flow around a cylinder of radius r0 . Integration of surface pressure.

1 Re50 1
Re100
Re150
0.5 Re200 0.5

0 0
Cp Cp
−0.5 −0.5

−1 −1

−1.5 −1.5
0 30 60 90 120 150 180 0 30 60 90 120 150 180
θ θ
(a) CFD of unsteady laminar flow [8]. The mark- (b) Potential flow, Eq. 4.73
ers show the time-averaged location of separation.

Figure 4.13: Pressure coefficients.

µ
f= + V∞ z (4.65)
πz
Now we define the radius of the cylinder, r0 , as

r02 = µ/(πV∞ ) (4.66)

so that
V∞ r02
f= + V∞ z (4.67)
z
On polar form it reads
 2 
V∞ r02 iθ r0 −iθ iθ
f= + V∞ re = V∞ e + re
reiθ r
 2  (4.68)
r0
= V∞ (cos θ − i sin θ) + r(cos θ + i sin θ)
r

The streamfunction reads (imaginary part)

 
r02
Ψ = V∞ r − sin θ (4.69)
r

Now we can compute the velocity components (see Eq. 4.51)

 
1 ∂Ψ r02
vr = = V∞ 1 − 2 cos θ
r ∂θ r
  (4.70)
∂Ψ r02
vθ = − = −V∞ 1 + 2 sin θ
∂r r
4.4. Potential flow 58

We find that vr = 0 for r = r0 as intended (thanks to the definition in Eq. 4.66). We

are not interested in the solution inside the cylinder (r < r0 ). Furthermore, we see
that the tangential velocity is zero at θ = 0 and π; hence these points correspond to
the stagnation points, see Fig. 4.11. The velocity field at the cylinder surface, r = r0 ,
reads
vr,s = 0
(4.71)
vθ,s = −2V∞ sin θ

where index s denotes surface. Note that the local velocity gets twice as large as the
freestream velocity at the top (θ = π/2) and the bottom (θ = −π/2) of the cylinder.
The surface pressure is obtained from Bernoulli equation (see Eq. 4.34)
2 2 2 2
V∞ P∞ vθ,s ps V∞ − vθ,s
+ = + ⇒ ps = P∞ + ρ (4.72)
2 ρ 2 ρ 2
where we neglected the gravitation term. The surface pressure is usually expressed as
a pressure coefficient
2
ps − P∞ vθ,s
Cp ≡ 2 /2
= 1 − 2
= 1 − 4 sin2 θ (4.73)
ρV∞ V∞

using Eq. 4.71.

It should be stressed that although Eqs. 4.71 and 4.73 are exact they are not realistic
because of the strict requirement that the flow should be inviscid. This requirement is
not valid neither in the boundary layers nor in the wake; the boundary layers may
be thin but the wake is a large part of the domain. Figure 4.13 presents the pressure
coefficient for potential flow and accurate unsteady CFD of two-dimensional viscous
flow [8] (the Reynolds number is sufficiently low for the flow to be laminar); Eqs. 2.3
and 2.8 are solved numerically [8]. The potential solution agrees rather well with
viscous flow up to θ ≃ 20o .
How do we find the lift and drag force? The only force (per unit area) that acts
on the cylinder surface is the pressure (in viscous flow there would also be a viscous
stress, but it is usually much smaller). To find the lift force, FL , we simply integrate
the pressure over the surface. Usually the lift force is expressed as a lift coefficient,
2
CL , which is scaled with the dynamic pressure ρV∞ /2. The lift coefficient is obtained
as
Z 1 Z 2π
FL ps
CL = 2
=− dx3 2 /2
sin θr0 dθ
ρV∞ /2 0 0 ρV∞
Z 1 Z 2π
= −r0 dx3 (1 − 4 sin2 θ) sin θdθ (4.74)
0 0
  2π
1 3
= −r0 − cos θ − 4 cos(3θ) − cos θ =0
12 4 0

The sin θ on the first line appears because we project the pressure force in the vertical
direction (see Fig. 4.12) and minus sign is because pressure acts inwards, see Eq. 1.9
and Fig. 4.1. We assume in Eq. 4.74 that the length of the cylinder in the x3 direction
4.4. Potential flow 59

x2
a b
V∞ θ
x1

FL

Figure 4.14: Flow around a cylinder of radius r0 with additional circulation which give
a (negative) lift force, see Eq. 4.85.

is one. The drag coefficient is computed as

Z 1 Z 2π
FD
CD = 2 /2
=− dx3 (1 − 4 sin2 θ) cos θr0 dθ
ρV∞ 0 0
Z 1 Z 2π  2π (4.75)
4
= −r0 dx3 sin θ − sin3 θ =0
0 0 3 0

The cos θ on the first line appears because we project the pressure force in the hori-
zontal direction (see Fig. 4.12). Equations 4.74 and 4.75 give CL = CD = 0; hence
we find that inviscid flow around a cylinder creates neither lift nor drag. The reason is
that the pressure is symmetric both with respect to x1 = 0 and x2 = 0. The lift force
on the lower surface side cancels the force on the upper side. Same argument for the
drag force: the pressure force on the upstream surface cancels that on the downstream
surface.

4.4.7 Analytical solutions for flow around a cylinder with circulation

We will now introduce a second example of potential flow around cylinders, which is
by far the most important one from engineering point of view. Here we will introduce
the use of additional circulation which alters the locations of the stagnation points and
creates lift. This approach is used in potential methods for predicting flow around
airfoils in aeronautics (mainly helicopters) and windpower engineering.
We add the complex potential of a vortex line (see Eq. 4.61) to Eq. 4.67 so that
V∞ r02 Γ
f= + V∞ z − i ln z (4.76)
z 2π
On polar form it reads (see Eqs. 4.62 and 4.68)
 2 
r0 Γ
f = V∞ (cos θ − i sin θ) + r(cos θ + i sin θ) − (i ln r − θ) (4.77)
r 2π
The imaginary part gives the streamfunction
 
r02 Γ
Ψ = V∞ r − sin θ − ln r (4.78)
r 2π
4.4. Potential flow 60

c
x2
V∞ θ
x1

Figure 4.15: Flow around a cylinder of radius r0 with maximal additional circulation.

We get the velocity components as (see Eqs. 4.63 and 4.70)

 
1 ∂Ψ r2
vr = = V∞ 1 − 02 cos θ
r ∂θ r
  (4.79)
∂Ψ r2 Γ
vθ = − = −V∞ 1 + 02 sin θ +
∂r r 2πr
The effect of the added vortex line is, as expected, to increase vθ while leaving vr
unaffected. The larger the circulation, the larger vθ .
The velocity at the surface, r = r0 , reads
vr,s = 0
Γ (4.80)
vθ,s = −2V∞ sin θ +
2πr0
Now let’s find the location of the stagnation points, i.e. where vθ,s = 0. Equation 4.80
gives  
Γ Γ
2V∞ sin θstag = ⇒ θstag = arcsin (4.81)
2πr0 4πr0 V∞
The two angles that satisfy this equation are located in the the first and second quad-
rants. The two positions are indicated with a and b in Fig. 4.14. For a limiting value of
the circulation, Γmax , the two locations s and b will merge at θ = π/2, denoted with c
in Fig. 4.15,
Γmax = 4πV∞ r0 . (4.82)
This corresponds to the maximum value of the circulation for which there is a stag-
nation point on the cylinder surface. For circulation larger than Γmax , the stagnation
point will be located above the cylinder.
The pressure is obtained from Bernoulli equation as (see Eq. 4.73)
2  2
vθ,s Γ
Cp = 1 − 2 = 1 − −2 sin θ +
V∞ 2πr0 V∞
 2 (4.83)
2 4Γ sin θ Γ
= 1 − 4 sin θ + −
2πr0 V∞ 2πr0 V∞
4.4. Potential flow 61

x2
F

FL
x1

Figure 4.16: Table tennis. The loop uses the Magnus effect. Side view.

We found in Section 4.4.6 that a cylinder without circulation gives neither drag nor
lift, see Eqs. 4.74 and 4.75. What about the present case? Let’s compute the lift. We
found in Eq. 4.74 that the two first terms in Eq. 4.83 give no contribution to the lift.
The last term cannot give any contribution to the lift because it is constant on the entire
surface. Hence we only need to include the third term in Eq. 4.83 so that
Z 1 Z 2π
FL ps
CL = 2 /2
= − dx 3 2
sin θr0 dθ
ρV∞ 0 0 ρV ∞ /2
Z 1 Z 2π
2Γ sin θ
= −r0 dx3 sin θdθ (4.84)
0 0 πr0 V∞
 2π
Γθ Γ 2Γ
=− − sin(2θ) =−
πV∞ 2πV∞ 0 V∞

We find that the lift force on a unit length of the cylinder can be computed from the
circulation as
FL = −ρV∞ Γ (4.85)
This relation is valid for any body and it is called the Kutta-Joukowski law who –
independent of each other – formulated it. The reason to the sign of the lift force can
easily be seen from Fig. 4.14. The stagnation points, where the pressure is largest, are
located at the top of the cylinder and hence the pressure is higher on the top than on the
bottom. The ”lift” force is acting downwards, i.e. in the negative x2 direction.
The drag is, however, still zero. In Eq. 4.75 we found that the first and the second
terms in Eq. 4.83 gives no contribution to drag. Hence, we only need to consider the
third terms. In the drag integral (see Eq. 4.75), this term in Eq. 4.83 gives rise to a term
proportional to sin θ cos θ whose contribution is zero. Hence, the additional circulation
does not give rise to any drag.

4.4.7.1 The Magnus effect

Circulation around a cylinder is very similar to a rotating cylinder. Instead of adding a
circulation, we let the cylinder rotate with speed Ω. A rotating cylinder produces lift.
This has interesting application in sports, for example football, table tennis and golf.
In table tennis, the ball must hit the table on the side of the opponent. One way to
improve the chance that this will happen is to make a loop. This means that you hit the
ball slightly on the top. The ball experiences a force, F , when you hit it (see Fig. 4.16)
and this force makes it rotate with rotation speed Ω (clockwise direction). The rotation
causes a lift, FL , which acts downwards so that the ball drops down quickly and (hope-
fully) hits the table on the other side of the net. The lift force is downwards because
4.4. Potential flow 62

FL
F wall
goal

Figure 4.17: Football. A free-kick uses the Magnus effect. Top view

Ω Vwind
Vship

x2 α

FL
x1

Figure 4.18: Flettner rotor (in blue) on a ship. The relative velocity between the ship
and the wind is Vwind + Vship . The ship moves with speed Vship . Top view.

the stagnation points are located on the upper surface. Recall that the relative velocity
of the air is in the negative x1 direction.
Another example where the Magnus effect is important is golf. Here the object is
often vice versa. You want the ball to go as far as possible. Hence you hit it with a
slice so that it spins with a positive Ω (counter-clockwise). The result is a lift force in
the positive x2 direction which makes the ball go further.
A final sports example is football. Here the lift is used sideways. Imagine there is a
free-kick rather close to the opponents’ goal, see Fig. 4.17. The opponents erects a wall
of players between the goal and the location of the free-kick. The player who makes
the free-kick wants to make the ball go on the left side of the wall; after the wall of
players, the ball should turn right towards the goal. The Magnus effect helps to achieve
this. The player hits the ball with his/her left foot on the left side of the ball which
creates a force F on the ball. This makes the ball rotate clockwise, see Fig. 4.17, and
creates a lift force so that the ball after it has passed the wall turns to the right towards
the goal. The reason that the ball turns to the right first after the wall (and not before)
is that the forward momentum created by F (the player) is much larger than FL .
If you are interested in football you may be pleased to learn that by use of fluid
dynamics it is now scientifically proven that it was much harder to make a good freekick
in 2010 worldcup than in 2014 [9]. Figure 7b in that paper is particularly interesting.

As an experiment, two identical freekicks are made with the football used at
4.4. Potential flow 63

δ(x1 )

V∞
c
x2

x1

Figure 4.19: Airfoil. The boundary layers, δ(x1 ), and the wake illustrated in red.
x1 = 0 and x1 = c at leading and trailing edge, respectively.

the 2013 FIFA Confederations. The freekicks are made 25m from the goal. The
initial velocity of the football is 30 m/s. The result of the two freekicks is that the
two footballs reach the goal three meters from each other in the vertical direction.
Why? Because the ball was rotated 45 degrees before the second freekick (see
Figs. 2c,d) in [9].

Finally we give an engineering example of the use of the Magnus effect. The first
Flettner rotors on ships were produced in 1924. It has recently gained new interest as
the cost of fuel is rising. A Flettner rotor is a rotating cylinder (or many) on a ship,
see Fig. 4.18. The diameter of this rotor can be a couple of meter and have a length
(i.e. height) of 10 − 20 meter. The ship is moving to the right with speed Vship . The
wind comes towards the ship from the left-front (relative wind at an angle of π/4).
The Flettner rotor rotates in the clockwise direction. The Magnus effect creates a force
in the orthogonal direction to the relative windspeed, i.e. at an angle of −π/4. Note
that if the wind comes from the right instead of from the left, the rotor should rotate
in the counter-clockwise direction. The additional propulsion force is FL cos(α). The
Division of Fluid Dynamics recently took part in an EU project where we studied the
flow around rotating cylinders in relation to Flettner rotors [10].

4.4.8 The flow around an airfoil

Flow around airfoils is a good example where potential methods are useful. These
methods are still in use in wind engineering and for helicopters. At the Division of
Fluid Dynamics we have an on-going PhD project where we use potential methods for
computing the aerodynamic loads for windturbine rotorblades [11].
The flow around airfoils is a good example where the flow can be treated as inviscid
in large part of the flow. For low angles of attack (which is the case for, for example, an
aircraft in cruise conditions) the boundary layers and the wake are thin. Outside these
regions the flow is essentially inviscid.
Figure 4.19 (see also Fig. 16.1) shows a two-dimensional airfoil. The boundary
layers and the wake are illustrated in red. The boundary layer is thinner on the pres-
sure (lower) side than on the suction (upper) side. It grows slightly thicker towards the
trailing edge (denoted by δ(x1 ) in Fig. 4.19). When this flow is computed using po-
tential methods, the location of the front stagnation point is reasonably well captured,
4.4. Potential flow 64

V∞

Figure 4.20: Airfoil. Streamlines from potential flow. Rear stagnation point at the
upper surface (suction side).

V∞

Figure 4.21: Airfoil. Streamlines from potential flow with added circulation. Rear
stagnation point at the trailing edge.
4.4. Potential flow 65

see Fig. 4.20. However, the stagnation point near the trailing edge is located on the
suction side which is clearly nonphysical. The flow on the pressure (lower) side cannot
be expected to make a 180o turn at the trailing edge and then go in the negative x1
direction towards the stagnation point located on the suction side.
The solution is to move the stagnation points in the same way as we did for the
cylinder flow in Section 4.4.7. We want to move the rear stagnation point towards
the trailing edge. This is achieved by adding a circulation in the clockwise direction,
see Fig. 4.21. The magnitude of the circulation is determined by the requirement that
the stagnation point should be located at the trailing edge. This is called the Kutta
condition. The added circulation is negative (clockwise). In aeronautics, the sign of
circulation is usually changed so that Γaeronautic = −Γ. The lift of a two-dimensional
airfoil (or a two-dimensional section of a three-dimensional airfoil) is then computed
as (see Eq. 4.85)
FL = ρV∞ Γaeronautic (4.86)
5. Turbulence 66

5 Turbulence
5.1 Introduction
LMOST all fluid flow which we encounter in daily life is turbulent. Typical exam-
A ples are flow around (as well as in) cars, aeroplanes and buildings. The boundary
layers and the wakes around and after bluff bodies such as cars, aeroplanes and build-
ings are turbulent. Also the flow and combustion in engines, both in piston engines
and gas turbines and combustors, are highly turbulent. Air movements in rooms are
turbulent, at least along the walls where wall-jets are formed. Hence, when we com-
pute fluid flow it will most likely be turbulent. In turbulent flow we usually divide the
velocities in one time-averaged part v̄i , which is independent of time (when the mean
flow is steady), and one fluctuating part vi′ so that vi = v̄i + vi′ .
There is no definition on turbulent flow, but it has a number of characteristic fea-
tures (see Pope [12] and Tennekes & Lumley [13]) such as:
I. Irregularity. Turbulent flow is irregular and chaotic (they may seem random,
but they are governed by Navier-Stokes equation, Eq. 2.8). The flow consists of a
spectrum of different scales (eddy sizes). We do not have any exact definition of an
turbulent eddy, but we suppose that it exists in a certain region in space for a certain turbulent
time and that it is subsequently destroyed (by the cascade process or by dissipation, eddy
see below). It has a characteristic velocity and length (called a velocity and length
scale). The region covered by a large eddy may well enclose also smaller eddies. The
largest eddies are of the order of the flow geometry (i.e. boundary layer thickness, jet
width, etc). At the other end of the spectrum we have the smallest eddies which are
dissipated by viscous forces (stresses) into thermal energy resulting in a temperature
increase. Even though turbulence is chaotic it is deterministic and is described by the
Navier-Stokes equations.
II. Diffusivity. In turbulent flow the diffusivity increases. The turbulence increases
the exchange of momentum in e.g. boundary layers, and reduces or delays thereby
separation at bluff bodies such as cylinders, airfoils and cars. The increased diffusivity
also increases the resistance (wall friction) and heat transfer in internal flows such as
in channels and pipes.
III. Large Reynolds Numbers. Turbulent flow occurs at high Reynolds number.
For example, the transition to turbulent flow in pipes occurs that ReD ≃ 2300, and in
boundary layers at Rex ≃ 500 000.
IV. Three-Dimensional. Turbulent flow is always three-dimensional and unsteady.
However, when the equations are time averaged, we can treat the flow as two-dimensional
(if the geometry is two-dimensional).
V. Dissipation. Turbulent flow is dissipative, which means that kinetic energy in
the small (dissipative) eddies are transformed into thermal energy. The small eddies
receive the kinetic energy from slightly larger eddies. The slightly larger eddies receive
their energy from even larger eddies and so on. The largest eddies extract their energy
from the mean flow. This process of transferring energy from the largest turbulent
scales (eddies) to the smallest is called the cascade process, see Fig. M.4. cascade
VI. Continuum. Even though we have small turbulent scales in the flow they are process
much larger than the molecular scale and we can treat the flow as a continuum.
5.2. Turbulent scales 67

flow of kinetic energy

ℓη

ℓ0 ℓ1
....
ℓ2

large scales intermediate scales dissipative scales

Figure 5.1: Cascade process with a spectrum of eddies. The energy-containing eddies
are denoted by v0 ; ℓ1 and ℓ2 denotes the size of the eddies in the inertial subrange such
that ℓ2 < ℓ1 < ℓ0 ; ℓη is the size of the dissipative eddies.

5.2 Turbulent scales

The largest scales are of the order of the flow geometry (the boundary layer thickness,
for example), with length scale ℓ0 and velocity scale v0 . These scales extract kinetic
energy from the mean flow which has a time scale comparable to the large scales, i.e.

∂v̄1
∼ t−1
0 ∼ v0 /ℓ0 (5.1)
∂x2
Part of the kinetic energy of the large scales is lost to slightly smaller scales with which
the large scales interact. Through the cascade process, kinetic energy is in this way
transferred from the largest scale to the smallest scales. At the smallest scales the
frictional forces (viscous stresses) become large and the kinetic energy is transformed
(dissipated) into thermal energy. The kinetic energy transferred per unit time from
eddy-to-eddy (from an eddy to a slightly smaller eddy) is the same for each eddy size.
The dissipation is denoted by ε which is energy per unit time and unit mass (ε =
[m2 /s3 ]). The dissipation is proportional to the kinematic viscosity, ν, times the fluc-
tuating velocity gradient up to the power of two (see Section 8.2). The friction forces
exist of course at all scales, but they are largest at the smallest eddies. In reality a small
fraction is dissipated at all scales. However it is assumed that most of the energy that
goes into the large scales per unit time (say 90%) is finally dissipated at the smallest
(dissipative) scales.
The smallest scales where dissipation occurs are called the Kolmogorov scales
whose velocity scale is denoted by vη , length scale by ℓη and time scale by τη . We
assume that these scales are determined by viscosity, ν, and dissipation, ε. The argu-
ment is as follows.
5.3. Energy spectrum 68

viscosity: Since the kinetic energy is destroyed by viscous forces it is natural to assume
that viscosity plays a part in determining these scales; the larger viscosity, the
larger scales.
dissipation: The amount of energy per unit time that is to be dissipated is ε. The more
energy that is to be transformed from kinetic energy to thermal energy, the larger
the velocity gradients must be.
Having assumed that the dissipative scales are determined by viscosity and dissipation,
we can express vη , ℓη and τη in ν and ε using dimensional analysis. We write

vη = νa εb
2 2 3 (5.2)
[m/s] = [m /s] [m /s ]
where below each variable its dimensions are given. The dimensions of the left and the
right side must be the same. We get two equations, one for meters [m]
1 = 2a + 2b, (5.3)
and one for seconds [s]
−1 = −a − 3b, (5.4)
which give a = b = 1/4. In the same way we obtain the expressions for ℓη and τη so
that
 3 1/4  ν 1/2
1/4 ν
vη = (νε) , ℓη = , τη = (5.5)
ε ε

5.3 Energy spectrum

As mentioned above, the turbulence fluctuations are composed of a wide range of
scales. We can think of them as eddies, see Fig. 5.1. It turns out that it is often conve-
nient to use Fourier series to analyze turbulence. In general, any periodic function, g,
with a period of 2L (i.e. g(x) = g(x + 2L)), can be expressed as a Fourier series, i.e.
X∞
1
g(x) = a0 + (an cos(κn x) + bn sin(κn x)) (5.6)
2 n=1

where x is a spatial coordinate and

nπ 2π
κn = or κ = . (5.7)
L L
κn is called the wavenumber. The Fourier coefficients are given by
Z
1 L
an = g(x) cos(κn x)dx
L −L
Z
1 L
bn = g(x) sin(κn x)dx
L −L
Parseval’s formula states that
Z L X∞
L
g 2 (x)dx = a20 + L (a2n + b2n ) (5.8)
−L 2 n=1
5.3. Energy spectrum 69

∂hv̄i i
−hv̄i′ v̄j′ i
∂xj

I
E

II

εκ

E(
κ)
III ε

∝
κ
−
5/
3

Figure 5.2: Spectrum for turbulent kinetic energy, k. I: Range for the large, energy
containing eddies. II: the inertial subrange. III: Range for small, isotropic scales. For
a discussion of εκ vs. ε, see Section 8.2.1. The wavenumber, κ, is proportional to the
inverse of the length scale of a turbulent eddy, ℓκ , i.e. κ ∝ ℓ−1
κ . For a discussion of εκ
vs. ε, see Section 8.2.1.

For readers not familiar to Fourier series, a brief introduction is given in Appendix D.
An example of a Fourier series and spectra are given in Appendix E. Let g be a fluc-
tuating velocity component, say v1′ . The left side of Eq. 5.8 expresses v1′2 in physical
space (vs. x) and the right side v1′2 in wavenumber space (vs. κn ). The reader who is
not familiar to the term “wavenumber”, is probably more familiar to “frequency”. In
that case, express g in Eq. 5.6 as a series in time rather than in space. In this case the
left side of Eq. 5.8 expresses v1′2 as a function of time and the right side expresses v1′2
as a function of frequency.
The turbulent scales are distributed over a range of scales which extends from the
largest scales which interact with the mean flow to the smallest scales where dissipation
occurs, see Fig. 5.1. Let us think about how the kinetic energy of the eddies varies with
eddy size. Intuitively we assume that large eddies have large fluctuating velocities
which implies large kinetic energy, vi′ vi′ /2. It is convenient to study the kinetic energy
of each eddy size in wavenumber space. In wavenumber space the energy of eddies
can be expressed as

E(κ)dκ (5.9)

where Eq. 5.9 expresses the contribution from the scales with wavenumber between κ
and κ + dκ to the turbulent kinetic energy k. The energy spectrum, E(κ), corresponds
to g 2 (κ) in Eq. 5.8. The dimension of wavenumber is one over length; thus we can think
of wavenumber as proportional to the inverse of an eddy’s diameter, i.e κ ∝ 1/d. The
total turbulent kinetic energy is obtained by integrating over the whole wavenumber
5.3. Energy spectrum 70

space, i.e. Z ∞ X
k= E(κ)dκ = L g 2 (κn ) (5.10)
0
Think of this equation as a way to compute the kinetic energy by first sorting all eddies
by size (i.e. wavenumber), then computing the kinetic energy of each eddy size (i.e.
E(κ)dκ), and finally summing the kinetic energy of all eddy sizes (i.e. carrying out the
integration). Note that the physical meaning of E is kinetic energy per unit wavenum-
ber of eddies of size ℓκ ∝ κ−1 . Hence the dimension of E is v 2 /κ, see Eq. 5.10; for a
discussion on the dimension of E, see Appendix E.
The kinetic energy is the sum of the kinetic energy of the three fluctuating velocity
components, i.e.
1  ′2  1
k= v1 + v2′2 + v3′2 = vi′ vi′ (5.11)
2 2
The spectrum of E is shown in Fig. 5.2. We find region I, II and III which are discussed
below.

I. In this region we have the large eddies which carry most of the energy. These
eddies interact with the mean flow and extract energy from the mean flow. This
energy transfer takes places via the production term, P k , in the transport equation
for turbulent kinetic energy, see Eq. 8.14. Part of the energy extracted per unit
time by the largest eddies is transferred (per unit time) to slightly smaller scales.
The eddies’ velocity and length scales are v0 and ℓ0 , respectively.
III. Dissipation range. The eddies are small and isotropic and it is here that the
dissipation occurs. The energy transfer from turbulent kinetic energy to thermal
energy (increased temperature) is governed by ε in the transport equation for
turbulent kinetic energy, see Eq. 8.14. The scales of the eddies are described by
the Kolmogorov scales (see Eq. 5.5)
II. Inertial subrange. The existence of this region requires that the Reynolds number
is high (fully turbulent flow). The eddies in this region represent the mid-region.
The turbulence is also in this region isotropic. This region is a “transport re-
gion” (i.e. in wavenumber space) in the cascade process. The “transport” in
wavenumber space is called spectral transfer. Energy per time unit, P k = ε, is spectral
coming from the large eddies at the lower part of this range and is transferred transfer
per unit time to the dissipation range at the higher part. Note that the relation
P k = {dissipation at small scales}, see Fig. 5.2, is given by the assumption of
the cascade process, i.e. that the energy transfer per unit time from eddy-size–
to–eddy-size is the same for all eddy sizes.
′ ′
The kinetic energy, kκ = vκ,i vκ,i /2, of an eddy of size (lengthscale), 1/κ, repre-
sents the kinetic energy of all eddies of this size. The kinetic energy of all eddies
(of all size) is computed by Eq. 5.11. The eddies in this region are indepen-
dent of both the large, energy-containing eddies and the eddies in the dissipation
range. One can argue that the eddies in this region should be characterized by
the spectral transfer of energy per unit time (ε) and the size of the eddies, 1/κ.
Dimensional analysis gives

E = κa εb
(5.12)
[m /s2 ] =
3 2 3
[1/m] [m /s ]
5.3. Energy spectrum 71

We get two equations, one for meters [m]

3 = −a + 2b,

and one for seconds [s]

−2 = −3b,
so that b = 2/3 and a = −5/3. Inserted in Eq. 5.12 we get
2 5
E(κ) = CK ε 3 κ− 3 (5.13)

where the Kolmogorov constant CK ≃ 1.5. This is a very important law (Kol-
mogorov spectrum law or the −5/3 law) which states that, if the flow is fully
turbulent (high Reynolds number), the energy spectra should exhibit a −5/3-
decay in the inertial region (region II, Fig. 5.2).

Above we state that the eddies in Region II and III are isotropic. This means that –
in average – the eddies have no preferred direction, i.e. the fluctuations in all directions
are the same so that v1′2 = v2′2 = v3′2 . Note that is not true instantaneously, i.e. in
general v1′ 6= v2′ 6= v3′ . Furthermore, isotropic turbulence implies that if a coordinate isotropic
direction is switched (i.e. rotated 180o), nothing should change. For example if the turbulence
x1 coordinate direction is rotated 180o the v1′ v2′ should remain the same, i.e. v1′ v2′ =
−v1′ v2′ . This is possible only if v1′ v2′ = 0. Hence, all shear stresses are zero in isotropic
turbulence. Using our knowledge in tensor notation, we know that an isotropic tensor
can be written as const. · δij . Hence, the Reynolds stress tensor for small scales can be
written as vi′ vj′ = const.δij which, again, shows us that the shear stresses are zero in
isotropic turbulence.
As discussed on p. 67, the concept of the cascade process assumes that the energy
extracted per unit time by the large turbulent eddies is transferred (per unit time) by
non-linear interactions through the inertial range to the dissipative range where the
kinetic energy is transformed (per unit time) to thermal energy (increased temperature).
The spectral transfer rate of kinetic energy from eddies of size 1/κ to slightly smaller
eddies can be estimated as follows. An eddy loses (part of) its kinetic energy during
one revolution. The kinetic energy of the eddy is proportional to vκ2 and the time for
one revolution is proportional to ℓκ /vκ . Hence, the energy spectral transfer rate, εκ ,
for an eddy of length scale 1/κ can be estimated as (see Fig. 5.2)

vκ2 v2 v3
εκ ∼ ∼ κ ∼ κ (5.14)
tκ ℓκ vκ ℓκ

Kinetic energy is transferred per unit time to smaller and smaller eddies until the trans-
fer takes place by dissipation (i.e. increased temperature) at the Kolmogorov scales. In
the inertial subrange, the cascade process assumes that εκ = ε. Applying Eq. 5.14 for
the large energy-containing eddies gives

v02 v3
ε0 ∼  ∼ 0 ∼ εκ = ε (5.15)
ℓ0 v0 ℓ0

The dissipation at small scales (large wavenumbers) is determined by how much energy
per unit time enters the cascade process at the large scales (small wavenumbers). We
5.4. The cascade process created by vorticity 72

generation x1 x2 x3
1st 1 0 0
2nd 0 1 1
3rd 2 1 1
4th 2 3 3
5th 6 5 5
6th 10 11 11
7th 22 21 21

Table 5.1: Number of eddies at each generation with their axis aligned in the x1 , x2 or
x3 direction, see Fig. 5.3.

can now estimate the ratio between the large eddies (with v0 and ℓ0 ) to the Kolmogorov
eddies (vη and ℓη ). Equations 5.5 and 5.15 give
v0
−1/4 1/4
= (νε)−1/4 v0 = νv03 /ℓ0 v0 = (v0 ℓ0 /ν) = Re1/4
vη
 3 −1/4  3 −1/4  3 −1/4
ℓ0 ν ν ℓ0 ν
= ℓ0 = ℓ 0 = = Re3/4 (5.16)
ℓη ε v03 v03 ℓ30
 −1/2  3 1/2  1/2
τo νℓ0 v0 ℓ0 v0 ℓ0
= τ0 = = = Re1/2
τη v03 νℓ0 v0 ν

where Re = v0 ℓ0 /ν. We find that the ratio of the velocity, length and time scales of the
energy-containing eddies to the Kolmogorov eddies increases with increasing Reynolds
number. This means that the eddy range (wavenumber range) of the intermediate region
(region II, the inertial region) increases with increasing Reynolds number. Hence, the
larger the Reynolds number, the larger the wavenumber range of the intermediate range
where the eddies are independent of both the large scales and the viscosity. or in other
words: the large the Reynolds number, the larger the difference between the largest
and the smallest scales. This is the very reason why it is so expensive (in terms of
computer power) to solve the Navier-Stokes equations. With a computational grid we
must resolve all eddies. Hence, as the Reynolds number increases, the number of grid
cells increases rapidly, see Eq. 28.1.

5.4 The cascade process created by vorticity

The interaction between vorticity and velocity gradients is an essential ingredient to
create and maintain turbulence. Disturbances are amplified by interaction between the
vorticity vector and the velocity gradients; the disturbances are turned into chaotic,
three-dimensional fluctuations, i.e. into turbulence. Two idealized phenomena in this
interaction process can be identified: vortex stretching and vortex tilting.
The equation for the instantaneous vorticity (ωi = ω̄i + ωi′ ) reads (see Eq. 4.20)

∂ωi ∂ωi ∂vi ∂ 2 ωi

+ vj = ωj +ν
∂t ∂xj ∂xj ∂xj ∂xj
(5.17)
∂vk
ωi = ǫijk
∂xj
 5.4. The cascade process created by vorticity 73

b 1

b b

2 2

3 b 3 b b 3 b 3

b 4 b 4 b 4 b 4 4 b 4 b 4 b 4 b

b b b b b b b b b b b b b b b b

5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5

Figure 5.3: Family tree of turbulent eddies (see also Table 5.1). Five generations. The
large original eddy, with axis aligned in the x1 direction, is 1st generation. Adapted
from [14]

As we learnt in Section 4.2 this equation is not an ordinary convection-diffusion equa-

tion: it has an additional term on the right side which represents amplification and
rotation/tilting of the vorticity lines (the first term on the right side). The i = j compo-
nents of this term represent (see Eq. 4.21) vortex stretching. A positive ∂v1′ /∂x1 will Vortex
stretch the cylinder, see Fig. 4.2 and from the requirement that the volume must not stretching
change (incompressible continuity equation) we find that the radius of the cylinder will
decrease. We have neglected the viscosity since viscous diffusion at high Reynolds
number is much smaller than the turbulent one and since viscous dissipation occurs at
small scales (see p. 67). Thus we can assume that there are no viscous stresses acting
on the cylindrical fluid element surface which means that the angular momentum

r2 ω1′ = const. (5.18)

remains constant as the radius of the fluid element decreases. Note that also the cir-
culation, Γ – which is the integral of the tangential velocity round the perimeter, see
Eq. 1.23 – is constant. Equation 5.18 shows that the vorticity increases if the radius
decreases (and vice versa). As was mentioned above, the continuity equation shows
that stretching results in a decrease of the radius of a slender fluid element and an in-
crease of the vorticity component (i.e. the tangential velocity component) aligned with
the element. For example, an extension of a fluid element in one direction (x1 direc-
tion) decreases the length scales in the x2 direction and increases ω1′ , see Fig. 5.4. At
the same time, vortex tilting creates small-scale vorticity in the x2 and x3 direction,
ω2′ and ω3′ . The increased ω1′ means that the velocity fluctuation in the x2 direction
is increased, see Fig. 5.5. The increased v2′ velocity component will stretch smaller
5.4. The cascade process created by vorticity 74

∂v1′
∂x1 >0

ω2′ , ω3′ ω1′ ω2′ , ω3′

v1′ v1′

x2

x1

∂v1′
Figure 5.4: A fluid element is stretched by > 0. Its radius decreases (from dashed
∂x1
line to solid line).

fluid elements aligned in the x2 direction, see Fig. 5.5. This will increase their vortic-
ity ω2′ and decrease their radius. In the same way will the increased ω1′ also stretch a
fluid element aligned in the x3 direction and increase ω3′ and decrease its radius. At
each stage, the length scale of the eddies – whose velocity scale are increased – de-
creases. Figure 5.3 illustrates how a large eddy whose axis is oriented in the x1 axis
in a few generations creates – through vortex stretching – smaller and smaller eddies
with larger and larger velocity gradients. Here a generation is related to a wavenumber
in the energy spectrum (Fig. 5.2); young generations correspond to high wavenumbers.
The smaller the eddies, the less the original orientation of the large eddy is recalled.
In other words, the small eddies “don’t remember” the characteristics of their original
ancestor. The small eddies have no preferred direction. They are isotropic. The cre-
ation of multiple eddies by vortex stretching from one original eddies is illustrated in
Fig. 5.3 and Table 5.1. The large original eddy (1st generation) is aligned in the x1
direction. It creates eddies in the x2 and x3 direction (2nd generation); the eddies in
the x2 direction create new eddies in the x1 and x3 (3rd generation) and so on. For
each generation the eddies become more and more isotropic as they get smaller.
The i 6= j components in the first term on the right side in Eq. 4.21 represent vortex Vortex
tilting. Again, take a slender fluid element, now with its axis aligned with the x2 axis, tilting
Fig. 4.3. The velocity gradient ∂v1 /∂x2 (or ∂v1′ /∂x2 , which is equivalent) will tilt the
fluid element so that it rotates in the clock-wise direction. As a result, the second term
ω2 ∂v1 /∂x2 in line one in Eq. 4.21 gives a contribution to ω1 (and ω1′ ). This shows
how vorticity in one direction is transferred to the other two directions through vortex
tilting.
Vortex stretching and vortex tilting qualitatively explain how interaction between
vorticity and velocity gradient create vorticity in all three coordinate directions from
a disturbance which initially was well defined in one coordinate direction. Once this
process has started it continues, because vorticity generated by vortex stretching and
vortex tilting interacts with the velocity field and creates further vorticity and so on.
The vorticity and velocity field becomes chaotic and three-dimensional: turbulence has
been created. The turbulence is also maintained by these processes.
From the discussion above we can now understand why turbulence always must be
three-dimensional (Item IV on p. 66). If the instantaneous flow is two-dimensional
(x1 − x2 plane) we find that the vortex-stretching/tilting term on the right side of
Eq. 5.17 vanishes because the vorticity vector and the velocity vector are orthogonal.
5.4. The cascade process created by vorticity 75

∂v3′ ω1′ , ω2′

>0
v3′ ∂x3 v3′ ω3′

x2 ω1′

x3

Figure 5.5: The rotation rate of the fluid element (black circles) in Fig. 5.4 increases
∂v ′
and its radius decreases. This creates a positive 3 > 0 which stretches the small red
∂x3
fluid element aligned in the x3 direction and increases ω3′ . The radius of the red fluid
element decreases.

The only non-zero component of vorticity vector is ω3 because

∂v3 ∂v2
ω1 = − ≡0
∂x2 ∂x3
∂v1 ∂v3
ω2 = − ≡ 0.
∂x3 ∂x1
Since v3 = 0, we get ωj ∂vi /∂xj = 0.
6. Turbulent mean flow 76

6 Turbulent mean flow

6.1 Time averaged Navier-Stokes
HEN the flow is turbulent it is preferable to decompose the instantaneous vari-
W ables (for example the velocity components and the pressure) into a mean value
and a fluctuating value, i.e.

vi = v̄i + vi′
(6.1)
p = p̄ + p′

where the bar, ¯·, denotes the time averaged value defined as
Z T
1
v̄ = vdt. (6.2)
2T −T

where T is sufficiently large. When we time average Eq. 6.1 we get

v̄i = v̄ i + vi′ = v̄i + vi′ (6.3)

where we used the fact that v̄i = v̄i , see Section 8.1.4. Hence, Eq. 6.3 gives

vi′ = 0, p′ = 0 (6.4)

One reason why we decompose the variables is that when we measure flow quan-
tities we are usually interested in their mean values rather than their time histories.
Another reason is that when we want to solve the Navier-Stokes equation numerically
it would require a very fine grid to resolve all turbulent scales and it would also require
a fine resolution in time (turbulent flow is always unsteady).
The continuity equation and the Navier-Stokes equation for incompressible flow
with constant viscosity read
∂vi
= 0 (6.5)
∂xi
∂vi ∂vi vj ∂p ∂ 2 vi
ρ +ρ = − +µ (6.6)
∂t ∂xj ∂xi ∂xj ∂xj

The gravitation term, −ρgi , has been omitted which means that the p is the hydro-
dynamic pressure (i.e. when vi ≡ 0, then p ≡ 0, see p. 35). Inserting Eq. 6.1 into the
continuity equation (6.5)

∂v̄i + vi′ ∂v̄ i ∂v ′ ∂v̄ i ∂v̄i

= + i = = (6.7)
∂xi ∂xi ∂xi ∂xi ∂xi

where we used the fact that vi′ = 0 (see Eq. 6.4 and v̄i = v̄i , see section 8.1.4).
Next, we use the decomposition in Navier-Stokes equation (Eq. 6.6)

∂(v̄i + vi′ ) ∂(v̄i + vi′ )(v̄j + vj′ ) ∂(p̄ + p′ ) ∂ 2 (v̄i + vi′ )

ρ +ρ =− +µ (6.8)
∂t ∂xj ∂xi ∂xj ∂xj
I II III IV

Let’s consider the equation term-by-term.

6.1. Time averaged Navier-Stokes 77

Term I:
∂(v̄i + vi′ ) ∂v̄i ∂v ′ ∂v̄i ∂v̄i
= + i = =
∂t ∂t ∂t ∂t ∂t
We assume that the mean flow, v̄i , is steady, and hence the term is zero.
Term II:

∂(v̄i + vi′ )(v̄j + vj′ ) ∂v̄i v̄j + v̄i vj′ + vi′ v̄j + vi′ vj′
=
∂xj ∂xj
∂v̄i v̄j ∂v̄i vj′ ∂v ′ v̄j ∂vi′ vj′
= + + i +
∂xj ∂xj ∂xj ∂xj

• Section 8.1.4 shows that v̄i v̄j = v̄i v̄j .

• Section 8.1.3 shows that v̄i vj′ = v̄ i vj′ = 0 and v̄j vi′ = v̄ j vi′ = 0

Hence, Term II reads

∂v̄i v̄j ∂vi′ vj′
+
∂xj ∂xj

Term III:
∂(p̄ + p′ ) ∂ p̄ ∂p′ ∂ p̄
= + =
∂xi ∂xi ∂xi ∂xi
Term IV:
∂ 2 (v̄i + vi′ ) ∂ 2 v̄i ∂ 2 vi′ ∂ 2 v̄i
= + =
∂xj ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj

Now we van finally write the time averaged continuity equation and Navier-Stokes
equation

∂v̄i
= 0 (6.9)
∂xi
 
∂v̄i v̄j ∂ p̄ ∂ ∂v̄i
ρ = − + µ − ρvi′ vj′ (6.10)
∂xj ∂xi ∂xj ∂xj

It is assumed that the mean flow is steady. This equation is the time-averaged
Navier-Stokes equation and it is often called the Reynolds equation. A new term ρvi′ vj′ Reynolds
appears on the right side of Eq. 6.10 which is called the Reynolds stress tensor. The equations
tensor is symmetric (for example v1′ v2′ = v2′ v1′ ). It represents correlations between
fluctuating velocities. It is an additional stress term due to turbulence (fluctuating ve-
locities) and it is unknown. We need a model for vi′ vj′ to close the equation system in
Eq. 6.10. This is called the closure problem: the number of unknowns (ten: three ve- closure
locity components, pressure, six stresses) is larger than the number of equations (four: problem
the continuity equation and three components of the Navier-Stokes equations).
The continuity equation applies both for the instantaneous velocity, vi (Eq. 6.5),
and for the time-averaged velocity, v̄i (Eq. 6.9); hence it applies also for the fluctuating
velocity, vi′ , i.e.
∂vi′
=0 (6.11)
∂xi
6.2. Wall region in fully developed channel flow 78

x2
v̄1 (x2 ) 2δ

xA
2
x1

x3

Figure 6.1: Flow between two infinite parallel plates. The width (i.e. length in the x3
direction) of the plates, Zmax , is much larger that the separation between the plates,
i.e. Zmax ≫ δ.

6.1.1 Boundary-layer approximation

For boundary-layer type of flow (i.e. boundary layers along a flat plate, channel flow,
pipe flow, jet and wake flow, etc.) the following relations apply
∂v̄1 ∂v̄1
v̄2 ≪ v̄1 , ≪ , (6.12)
∂x1 ∂x2
Assume steady (∂/∂t = 0), two-dimensional (v̄3 = ∂/∂x3 = 0) boundary-layer flow.
First we re-write the left side of Eq. 6.10 using the continuity equation
∂v̄i v̄j ∂v̄i ∂v̄j ∂v̄i
ρ = ρv̄j + ρv̄i = ρv̄j (6.13)
∂xj ∂xj ∂xj ∂xj
=0

Using Eq. 6.13, Eq. 6.10 can be written

 
∂v̄1 ∂v̄1 ∂ p̄ ∂ ∂v̄1 ′ ′
ρv̄1 + ρv̄2 = − + µ − ρv1 v2 (6.14)
∂x1 ∂x2 ∂x1 ∂x2 ∂x2
τ12,tot

x1 and x2 denote the streamwise and wall-normal coordinate, respectively, see Fig. 6.1.
Note that the two terms on the left side are of the same order, because they both include
the product of one large (v̄1 or ∂/∂x2 ) and one small (v̄2 or ∂/∂x1 ) part.
In addition to the viscous shear stress, µ∂v̄1 /∂x2 , an additional turbulent one – a shear
Reynolds shear stress – appears on the right side of Eq. 6.14. The total shear stress is stress
thus
∂v̄1
τ12,tot = µ − ρv1′ v2′ (6.15)
∂x2

6.2 Wall region in fully developed channel flow

The region near the wall is very important. Here the velocity gradient is largest as
the velocity drops down to zero at the wall over a very short distance. One important
quantity is the wall shear stress which is defined as

∂v̄1 
τw = µ (6.16)
∂x2 w
6.2. Wall region in fully developed channel flow 79

From the wall shear stress, we can define a wall friction velocity, uτ , as wall
 1/2 friction
τw velocity
τw = ρu2τ ⇒ uτ = (6.17)
ρ

In order to take a closer look at the near-wall region, let us, again, consider fully
developed channel flow between two infinite plates, see Fig. 6.1. In fully developed
channel flow, the streamwise derivative of the streamwise velocity component is zero
(this is the definition of fully developed flow), i.e. ∂v̄1 /∂x1 = 0. The continuity
equation gives now v̄2 = 0, see Eq. 3.18 at p. 35. The first term on the left side of
Eq. 6.14 is zero because we have fully developed flow (∂v̄1 /∂x1 = 0) and the last term
is zero because v̄2 ≡ 0. The streamwise momentum equation, Eq. 6.14, can now be
written  
∂ p̄ ∂ ∂v̄1 ′ ′
0=− + µ − ρv1 v2 (6.18)
∂x1 ∂x2 ∂x2
We know that the first term is a function only of x1 and the two terms in parenthesis
are functions of x2 only; hence they must be constant (see Eq. 3.24 and the text related
to this equation), i.e.
∂ p̄
− = constant
∂x1
  (6.19)
∂ ∂v̄1 ′ ′ ∂τ12,tot
µ − ρv1 v2 = = constant
∂x2 ∂x2 ∂x2

where the total stress, τ12,tot , is given by Eq. 6.15. Integrating Eq. 6.18 from x2 = 0
to x2 gives
∂ p̄ ∂ p̄  x2 
τ12,tot (x2 ) − τw = x2 ⇒ τ12,tot = τw + x2 = τw 1 − (6.20)
∂x1 ∂x1 δ
At the last step we used the fact that the pressure gradient balances the wall shear stress,
i.e. −∂ p̄/∂x1 = τw /δ, see Eq. 3.30 (note that h = 2δ) and Eq. 6.37.
The wall region can be divided into one outer and one inner region, see Fig. 6.2.
The inner region includes the viscous region, x+ 2 . 5 (dominated by the viscous diffu-
sion), and the logarithmic region, x+ 2 & 30 (dominated by turbulent diffusion); the log-
arithmic region is sometimes called the inertial region, because the turbulent stresses
stem from the inertial (i.e. the non-linear convection) term. The buffer region acts as a
transition region between these two regions where viscous diffusion of streamwise mo-
mentum is gradually replaced by turbulent diffusion. In the inner region, the total shear
stress is approximately constant and equal to the wall shear stress τw , see Fig. 6.3.
Note that the total shear stress is constant only close to the wall (Fig. 6.3b); further
away from the wall it decreases (in fully developed channel flow it decreases linearly
with the distance from the wall, see Eq. 6.20 and Fig. 6.3a). The Reynolds shear stress
vanishes at the wall because v1′ = v2′ = 0, and the viscous shear stress attains its
wall-stress value τw = ρu2τ . As we go away from the wall the viscous stress decreases
and the turbulent one increases and at x+ 2 ≃ 11 they are approximately equal. In the
logarithmic layer the viscous stress is negligible compared to the Reynolds stress.
At the wall, the velocity gradient is directly related to the wall shear stress, i.e. (see
Eq. 6.16 and 6.17) 
∂v̄1  τw ρ 1
= = u2τ = u2τ (6.21)
∂x2 w µ µ ν
6.2. Wall region in fully developed channel flow 80

10−4 10−3 10−2 10−1 1

x2 /δ
outer region
overlap region
log-law region

Centerline
Wall

inner region

buffer region
viscous region

x+
2 = y
+

1 5 10 30 100 1000 10000

Figure 6.2: The wall region (adapted from Ch.7 in [12]) for Reτ = 10 000. δ denotes
half width of the channel, see Fig. 6.1 and x+
2 = x2 uτ /ν denotes the normalized wall
distance.

Integration gives (recall that both ν and u2τ are constant)

1 2
v̄1 = u x2 + C1
ν τ
Since the velocity, v̄1 , is zero at the wall, the integration constant C1 = 0 so that
v̄1 uτ x2
= (6.22)
uτ ν
Equation 6.22 is expressed in inner scaling (or wall scaling) which means that v̄1 and
x2 are normalized with quantities related to the wall, i.e. the friction velocity stemming
from the wall shear stress and the viscosity (here we regard viscosity as a quantity
related to the wall, since the flow is dominated by viscosity). Often the plus-sign (‘ + ‘)
is used to denote inner scaling and equation Eq. 6.22 can then be written

v̄1+ = x+
2 (6.23)

From the friction velocity and the viscosity we can define the viscous length scale, ℓν ,
for the near-wall region as
ν
x+
2 = x2 /ℓν ⇒ ℓν = (6.24)
uτ

Further away from the wall at 30 . x+ 2 . 3000 (or 0.003 . x2 /δ . 0.3), we

encounter the log-law region, see Fig. 6.2. In this region the flow is assumed to be
independent of viscosity. The Reynolds shear stress, ρv1′ v2′ , is in the region x+
2 . 200
(i.e. x2 /δ . 0.1) fairly constant and approximately equal to τw , see Fig. 6.3b. Hence
the friction velocity, uτ , is a suitable velocity scale in the inner logarithmic region; it
is used in the entire region.
What about the length scale? Near the wall, an eddy cannot be larger than the
distance to the wall and it is the distance to the wall that sets an upper limit on the
6.2. Wall region in fully developed channel flow 81

2000 1 200 0.1

0.8
1500

0.6

x2 /δ

x2 /δ
x+

x+
2

2
1000 100 0.05
0.4

500
0.2

0 0 0 0
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
a) b)

Figure 6.3: Reynolds shear stress. Reτ = 2000. a) lower half of the channel; b) zoom
near the wall. DNS (Direct Numerical Simulation) data [15, 16]. : −ρv1′ v2′ /τw ;
: µ(∂v̄1 /∂x2 )/τw .
2000
24

20
1500
16
v̄1 /uτ

x+
2 1000 12

8
500
4

0
0 5 10 15 20 25 1 10 100 1000
a) b)
v̄1 /uτ x+
2
Figure 6.4: Velocity profiles in fully developed channel flow. Reτ = 2000. : DNS
(Direct Numerical Simulation) data [15,16]; : v̄1 /uτ = (ln x+
2 )/0.41 + 5.2; :
v̄1 /uτ = x+
2 .

eddy-size. Hence it seems reasonable to take the wall distance as the characteristic
length scale; a constant, κ, is added so that
ℓ = κx2 . (6.25)
The velocity gradient can now be estimated as
∂v̄1 uτ
= (6.26)
∂x2 κx2
based on the velocity scale, uτ , and the length scale κx2 . Another way of deriving the
expression in Eq. 6.26 is to use the Boussinesq assumption (see Eq. 11.33) in which a
turbulent Reynolds stress is assumed to be equal to the product between the turbulent
viscosity and the velocity gradient as
∂v̄1
− v1′ v2′ = νt (6.27)
∂x2
The turbulent viscosity, νt , represents the turbulence and has the same dimension as ν,
i.e. [m2 /s]. Hence νt can be expressed as a product of a turbulent velocity scale and a
turbulent length scale, and in the log-law region that gives
νt = uτ κx2 (6.28)
6.2. Wall region in fully developed channel flow 82

τ32,tot

x2

x3

Figure 6.5: Symmetry plane of channel flow.

so that Eq. 6.27 gives (inserting −v1′ v2′ = u2τ )

∂v̄1 ∂v̄1 uτ
u2τ = κuτ x2 ⇒ = (6.29)
∂x2 ∂x2 κx2
In non-dimensional form Eqs. 6.26 and 6.29 read

∂v̄1+ 1
= (6.30)
∂x+ 2 κx+
2

Integration gives now

1

v̄1+ = ln x+
2 + B or
κ
(6.31)
v̄1 1  x2 uτ 
= ln +B
uτ κ ν
where B is an integration constant. Equation 6.31 is the logarithmic law due to von log-law
Kármán [17]. The constant, κ, is called the von Kármán constant. The constants in the
log-law are usually set to κ = 0.41 and B = 5.2.
As can be seen in Fig. 6.2 the log-law applies for x+ 2 . 3000 (x2 /δ . 0.3).
Figure 6.4 – where the Reynolds number is lower than in Fig. 6.2 – shows that the log-
law fit the DNS (Direct Numerical Simulation) up to x+ 2 . 500 (x2 /δ . 0.25). Hence,
the upper limit for the validity of the log-law is dependent on Reynolds number; the
larger the Reynolds number, the larger the upper limit.
In the outer region of the boundary layer, the relevant length scale is the boundary
layer thickness. The resulting velocity law is the defect law
v̄1,c − v̄1 x 
2
= FD (6.32)
uτ δ
where c denotes centerline. The velocity in the log-region and the outer region (often
called the wake region) can be written as
v̄1 1 2Π  πx 
2
= ln(y + ) + B + sin2 (6.33)
uτ κ κ 2δ
where κ = 0.38, B = 4.1 and Π = 0.5 are taken from boundary layer flow [18–20].
6.3. Reynolds stresses in fully developed channel flow 83

200 2000
1800

150 1600
1400
1200
x+
2 100 x+
2 1000
800
50 600
400
0 200
−150 −100 −50 0 50 100 150 −2 −1 0 1 2
a) b)

Figure 6.6: Fully developed channel flow. Reτ = 2000. Forces in the v̄1 equation,
see Eq. 6.18. a) near the lower wall of the channel; b) lower half of the channel ex-
cluding the near-wall region. DNS (Direct Numerical Simulation) data [15, 16]. :
−ρ(∂v1′ v2′ /∂x2 )/τw ; : µ(∂ 2 v̄1 /∂x22 )/τw ; : −(∂ p̄/∂x1 )/τw .

6.3 Reynolds stresses in fully developed channel flow

The flow is two-dimensional (v̄3 = 0 and ∂/∂x3 = 0). Consider the x2 − x3 plane,
see Fig. 6.5. Since nothing changes in the x3 direction, the viscous shear stress
 
∂v̄3 ∂v̄2
τ32 = µ + =0 (6.34)
∂x2 ∂x3

because v̄3 = ∂v̄2 /∂x3 = 0. The turbulent part shear stress, ρv2′ v3′ , can be expressed
using the Boussinesq assumption (see Eq. 11.33)
 
′ ′ ∂v̄3 ∂v̄2
− ρv2 v3 = µt + =0 (6.35)
∂x2 ∂x3

and it is also zero since v̄3 = ∂v̄2 /∂x3 = 0. With the same argument, v1′ v3′ = 0.
However note that v3′2 = v32 6= 0. The reason is that although the time-averaged flow
is two-dimensional (i.e. v̄3 = 0), the instantaneous turbulent flow is always three-
dimensional and unsteady. Hence v3 6= 0 and v3′ 6= 0 so that v3′2 6= 0. Consider, for
example, the time series v3 = v3′ = (−0.25, 0.125, 0.125, −0.2, 0.2). This gives

v̄3 = (−0.25 + 0.125 + 0.125 − 0.2 + 0.2)/5 = 0

but
 
v3′2 = v32 = (−0.25)2 + 0.1252 + 0.1252 + (−0.2)2 + 0.22 /5 = 0.03475 6= 0.

Figure 6.3 presents the Reynolds and the viscous shear stresses for fully developed
flow. As can be seen, the viscous shear stress is negligible except very near the wall. It
is equal to one near the wall and decreases rapidly for increasing wall distance. On the
other hand, the Reynolds shear stress is zero at the wall (because the fluctuating veloc-
ities are zero at the wall) and increases for increasing wall distance. The intersection
of the two shear stresses takes place at x+
2 ≃ 11.
Looking at Eq. 6.18 we find that it is not really the shear stress that is interesting,
but its gradient. The gradient of the shear stress, −∂(ρv1′ v2′ )/∂x2 and µ∂ 2 v̄1 /∂x22
represent, together with the pressure gradient, −∂ p̄/∂x1 , the forces acting on the fluid.
6.3. Reynolds stresses in fully developed channel flow 84

∂v1′ v2′ ∂ 2 v̄1

x2 −ρ µ
∂x2 ∂x22
1

∂ p̄ ∂v1′ v2′
− −ρ
∂x1 ∂x2

0
0 1 v̄1

Figure 6.7: Forces in a boundary layer. The red (dashed line) and the blue (solid line)
fluid particle are located at x+ +
2 ≃ 400 and x2 ≃ 20, respectively (see Fig. 6.6).

Figure 6.6 presents the forces. Start by looking at Fig. 6.6b which shows the forces
in the region away from the wall, see the red fluid particle in Fig. 6.7. The pressure
gradient is constant and equal to one: this is the force driving the flow. This agrees
– fortunately – with our intuition. We can imagine that the fluid (air, for example) is
driven by a fan. Another way to describe the behaviour of the pressure is to say that
there is a pressure drop. The pressure must decrease in the streamwise direction so that
the pressure gradient term, −∂ p̄/∂x1 , in Eq. 6.18 takes a positive value which pushes
the flow in the x1 direction. The force that balances the pressure gradient is the gradient
of the Reynolds shear stress. This is the force opposing the movement of the fluid. This
opposing force has its origin at the walls due to the viscous wall force (viscous shear
stress multiplied by area).
Now let’s have a look at the forces in the near-wall region, see Fig. 6.6a. Here the
forces are two orders of magnitude larger than in Fig. 6.6b but they act over a very thin
region (x+ 2 ≤ 40 or x2 /δ < 0.02). In this region the Reynolds shear stress gradient
term is driving the flow and the opposing force is the viscous force, see the blue fluid
particle in Fig. 6.7. We can of course make a force balance for a section of the channel,
as we did for laminar flow, see Eq. 3.36 at p. 37 and Fig. 3.8 at p. 37 which reads

0 = p̄1 Zmax 2δ − p̄2 Zmax 2δ − 2τw LZmax (6.36)

where L is the length of the section. We get

∆p̄ ∂ p̄ τw
=− = (6.37)
L ∂x1 δ
As can be seen the pressure drop is directly related to the wall shear stress. In turbulent
flow the velocity profile in the center region is much flatter than in laminar flow (cf.
Fig. 6.4 and Fig. 3.7 at p. 36). This makes the velocity gradient near the wall (and
the wall shear stress, τw ) much larger in turbulent flow than in laminar flow: Eq. 6.37
6.4. Boundary layer 85

2000 100

80
1500

60
x+
2 1000 x+
2
40

500
20

0 0
0 2 4 6 8 0 2 4 6 8
a) b)

Figure 6.8: Normal Reynolds stresses and turbulent kinetic energy. Reτ = 2000. DNS
(Direct Numerical Simulation) data [15, 16]. : ρv1′2 /τw ; : ρv2′2 /τw ; :
′2 2
ρv3 /τw ; ◦: k/uτ .

2000
24

20
1500
16
v̄2 /uτ

x+
2 1000 12

8
500
4

0
0 5 10 15 20 25 1 10 100 1000
a) b)
v̄2 /uτ x+
2
Figure 6.9: Velocity profiles in a boundary layer along a flat plate. : DNS (Direct
Numerical Simulation) data [21]; : v̄2 /uτ = (ln x+
2 )/0.41 + 5.2; : v̄2 /uτ =
x+
2 .

shows why the pressure drop is larger in the former case compared to the latter; or —
in other words – why a larger fan is required to push the flow in turbulent flow than in
laminar flow.
Figure 6.8 presents the normal Reynolds stresses, ρv1′2 , ρv2′2 and ρv3′2 . As can
be seen, the streamwise stress is largest and the wall-normal stress is smallest. The
former is largest because the mean flow is in this direction; the latter is smallest because
the turbulent fluctuations are dampened by the wall. The turbulent kinetic energy,
k = vi′ vi′ /2, is also included. Note that this is smaller than v1′2 .

6.4 Boundary layer

Up to now we have mainly discussed fully developed channel flow. What is the dif-
ference between that flow and a boundary layer flow? First, in a boundary layer flow
the convective terms are not zero (or negligible), i.e. the left side of Eq. 6.14 is not
zero. The flow in a boundary layer is continuously developing, i.e. its thickness, δ,
increases continuously for increasing x1 . The flow in a boundary layer is described by
Eq. 6.14. Second, in a boundary layer flow the wall shear stress is not determined by
the pressure drop (indeed it is zero); the total shear stress is balanced by the convective
terms. Third, the outer part of the boundary layer is highly intermittent, consisting of
6.4. Boundary layer 86

turbulent/non-turbulent motion.
However, the inner region of a boundary layer (x2 /δ < 0.1) is principally the same
as for the fully developed channel flow, see Fig. 6.9: the linear and the log-law regions
are very similar for the two flows. However, in boundary layer flow the log-law is
valid only up to approximately x2 /δ ≃ 0.1 (compared to approximately x2 /δ ≃ 0.3 in
channel flow)
7. Probability density functions 87

20 1.5 6

15 1 4

0.5 2
10
′
v 5
0 0

−0.5 −2
0
−1 −4

−5
−1.5 −6
10 12 14 16 18 20 10 12 14 16 18 20 10 12 14 16 18 20
t t t
(a) Point 1. vrms = 2.15. (b) Point 2. vrms = 0.23. (c) Point 3. vrms = 1.44.

Figure 7.1: Time history of v ′ . Horizontal red lines show ±vrms .

7 Probability density functions

statistical information is obtained by forming the mean and second moments,
OME
S for example v̄ and v2′2 , as was done in Section 6. The root-mean-square (RMS) can root-mean-
be defined from the second moment as square
 1/2 RMS
vrms = v ′2 (7.1)

The RMS is the same as the standard deviation which is equal to the square-root of the standard
variance. In order to extract more information, probability density function is a useful deviation
statistical tool to analyze turbulence. From the velocity signals we can compute the variance
probability densities (sometimes called histograms). With a probability density, fv , of
the v velocity, the mean velocity is computed as
Z ∞
v̄ = vfv (v)dv (7.2)
−∞

Normalize the probability functions, so that

Z ∞
fv (v)dv = 1 (7.3)
−∞

Here we integrate over v. The mean velocity can of course also be computed by
integrating over time, as we do when we define a time average, (see Eq. 6.1 at p. 76),
i.e. Z T
1
v̄ = vdt (7.4)
2T −T
where T is “sufficiently” large.
Consider the probability density functions of the fluctuations. The second moment
corresponds to the variance of the fluctuations (or the square of the RMS, see Eq. 7.1),
i.e. Z ∞
v ′2 = v ′2 fv′ (v ′ )dv ′ (7.5)
−∞

As in Eq. 7.4, v ′2 is usually computed by integrating in time, i.e.

Z T
1
v ′2 = v ′2 (t)dt
2T −T
 7. Probability density functions 88

3
0.3
2.5 0.25
0.25

2 0.2
0.2

fv′ 0.15 1.5 0.15

0.1 1 0.1

0.05 0.5 0.05

0 0 0
−2 0 2 4 6 8 10 −5 0 5 −4 −2 0 2 4

v ′ /vrms v ′ /vrms v ′ /vrms

(a) Point 1. S = 1.51 and F = (b) Point 2. S = −0.27 and F = (c) Point 3. S = −0.09 and F =
7.68. 5.81. 2.77.

Figure 7.2: Probability density functions of time histories in Fig. 7.1. Vertical red
lines show ±vrms . The skewness, S, and the flatness, F , are given for the three time
histories.

A probability density function is symmetric if positive values are as frequent and

large as the negative values. Figure 7.1 presents the time history of the v ′ history at
three different points in a flow (note that v ′ = 0). The red horizontal lines indicate the
RMS value of v ′ . The resulting probability densities functions are shown in Fig. 7.2.
The red vertical lines show plus and minus RMS of v ′ . Let us analyze the data at the
three points.

Point 1. The time history of the velocity fluctuation (Fig. 7.1a) shows that there ex-
ists large positive values but no large negative values. The positive values are
often larger than +vrms (the peak is actually close to 8vrms ) but the negative
values are seldom smaller than −vrms . This indicates that the distribution of v ′
is skewed towards the positive side. This is confirmed in the PDF distribution,
see Fig. 7.2a.
Point 2. The fluctuations at this point are much smaller and the positive values are as
large the negative values; this means that the PDF should be symmetric which is
confirmed in Fig. 7.2b. The extreme values of v ′ are approximately ±1.5vrms ,
see Figs. 7.1b and 7.2b.
Point 3. At this point the time history (Fig. 7.1c) shows that the fluctuations are clus-
tered around zero and much values are within ±vrms . The time history shows
that the positive and the negative values have the same magnitude. The PDF
function in Fig. 7.2c confirms that there are many value around zero, that the ex-
treme value are small and that positive and negative values are equally frequent
(i.e. the PDF is symmetric).

In Fig. 7.2 we can judge whether the PDF is symmetric, but instead of “looking” at
the probability density functions, we should use a definition of the degree of symmetry,
which is the skewness. It is defined as skewness
Z ∞
v ′3 = v ′3 fv′ (v ′ )dv ′
−∞

3
and is commonly normalized by vrms , so that the skewness, Sv′ , of v ′ is defined as
Z ∞ Z T
1 ′3 ′ ′ 1
Sv ′ = 3
v fv′ (v )dv = 3
v ′3 (t)dt
vrms −∞ 2vrms T −T
7. Probability density functions 89

Note that f must be normalized (see Eq. 7.3).

There is yet another statistical quantity which sometimes is used for describing
turbulent fluctuations, namely the flatness. The variance (the square of RMS) tells us flatness
how large the fluctuations are in average, but it does not tell us if the time history
includes few very large fluctuations or if all are rather close to vrms . The flatness gives
this information, and it is defined computed from v ′4 and normalized by vrms 4
, i.e.
Z ∞
1
F = 4 v ′4 fv′ (v)dv
vrms −∞

The fluctuations at Point 1 (see Fig. 7.1a) includes some samples which are very large
and hence its flatness is large (see caption in Fig. 7.2a), whereas the fluctuation for
Point 3 all mostly clustered within ±2vrms giving a small flatness, see Fig. 7.1c and
the caption in Fig. 7.2c. For a Gaussian distribution
 
1 (v ′ − vrms )2
f (v ′ ) = exp − 2
vrms 2vrms

for which F = 3.
8. Transport equations for turbulent kinetic energy 90

8 Transport equations for turbulent kinetic energy

N this section and Section 9 we will derive various transport equations. There are two
I tricks which often will be used. Both tricks simply use the product rule for derivative
backwards.

Trick 1: Using the product rule we get

∂Ai Bj ∂Bj ∂Ai
= Ai + Bj (8.1)
∂xk ∂xk ∂xk
This expression can be re-written as
∂Bj ∂Ai Bj ∂Ai
Ai = − Bj (8.2)
∂xk ∂xk ∂xk
and then we call it the “product rule backwards”.
Trick 2: Using the product rule we get
 
1 ∂Ai Ai 1 ∂Ai ∂Ai ∂Ai
= Ai + Ai = Ai (8.3)
2 ∂xj 2 ∂xj ∂xj ∂xj
This trick is usually used backwards, i.e.
∂Ai 1 ∂Ai Ai
Ai = (8.4)
∂xj 2 ∂xj

8.1 Rules for time averaging

8.1.1 What is the difference between v1′ v2′ and v1′ v2′ ?
Using 6.2 we get
Z T
1
v1′ v2′ = v1′ v2′ dt.
2T −T

whereas ! !
Z T Z T
1 1
v1′ v2′ = v1′ dt v1′ dt
2T −T 2T −T

We take a numerical example. Let the time series of v1′ and v2′ be

v1′ = [0.2, −0.3, 0.18, −0.08]

v2′ = [0.15, −0.25, 0.04, 0.06]
N
1 X ′
v1′ = v = (0.2 − 0.3 + 0.18 − 0.08)/4 = 0
N n=1 1,n
N
1 X ′
v2′ = v = (0.15 − 0.25 + 0.04 + 0.06)/4 = 0
N n=1 2,n
so that ! !
N N
1 X ′ 1 X ′
v1′ v2′ = v v =0·0=0
N n=1 1,n N n=1 2,n
8.1. Rules for time averaging 91

However, the time average of their product is not zero, i.e.

N
1 X ′ ′
v1′ v2′ = v v = (0.2·0.15+0.3·0.25+0.18·0.04−0.08·0.06)/4 = 0.02685
N n=1 1,n 2,n

2
8.1.2 What is the difference between v1′2 and v1′ ?
Using 6.2 we get
Z T
1
v1′2 = v ′2 dt.
2T −T
whereas !2
Z T
2 1 ′
v1′ = v dt .
2T −T

The numerical example gives

N
1 X ′2
v1′2 = v = (0.22 + 0.32 + 0.182 + 0.082 )/4 = 0.0422
N n=1 1,n
N
1 X ′2
v2′2 = v = (0.152 + 0.252 + 0.042 + 0.062 )/4 = 0.02255
N n=1 2,n

but
N
!2
2 1 X ′ 2
v1′ = v = [(0.2 − 0.3 + 0.18 − 0.08)/4] = 0
N n=1 1,n
N
!2
2 1 X ′ 2
v2′ = v = [(0.15 − 0.25 + 0.04 + 0.06)/4] = 0
N n=1 2,n

8.1.3 Show that v̄1 v1′2 = v̄1 v1′2

Using 6.2 we get
Z T
1
v̄1 v1′2 = v̄1 v1′2 dt
2T −T
and since v̄ does not depend on t we can take it out of the integral as
Z T
1
v̄1 v ′2 dt = v̄1 v1′2
2T −T 1
Now let’s do it with numerical values. Assume that v¯1 = 10.
N N
!
′2 1 X 1 X ′2
v̄1 v1 = v1,m v1,n =
N n=1 N m=1
= (10 · 0.22 + 10 · 0.32 + 10 · 0.182 + 10 · 0.082 )/4 = 0.422
N
! N
!
1 X 1 X
v̄1 v1′2 = v1,n v ′2 =
N n=1 N n=1 1,n
 
= 10 · (0.22 + 0.32 + 0.182 + 0.082 )/4 = 0.422
8.2. The Exact k Equation 92

8.1.4 Show that v̄ 1 = v̄1

Using 6.2 we get
Z T
1
v̄ 1 = v̄1 dt
2T −T
and since v̄ does not depend on t we can take it out of the integral as
Z T
1 1
v̄1 dt = v̄1 2T = v̄1
2T −T 2T
With numerical values we get
N
1 X
v̄ 1 = = (10 + 10 + 10 + 10)/4 = 10 = v̄1
N n=1

8.2 The Exact k Equation

The equation for turbulent kinetic energy, k = 21 vi′ vi′ , is derived from the Navier-Stokes
equation. Again, we assume incompressible flow (constant density) and constant vis-
cosity (cf. Eq. 6.6). We subtract Eq. 6.10 from Eq. 6.6 and divide by density, multiply
by vi′ and time average which gives

∂
vi′ [vi vj − v̄i v̄j ] =
∂xj
(8.5)
1 ∂ ∂2 ∂vi′ vj′ ′
− vi′ [p − p̄] + νvi′ [vi − v̄i ] + v
ρ ∂xi ∂xj ∂xj ∂xj i
Using vj = v̄j + vj′ , the left side can be rewritten as

∂   ∂  ′ 
vi′ (v̄i + vi′ )(v̄j + vj′ ) − v̄i v̄j = vi′ v̄i vj + vi′ v̄j + vi′ vj′ . (8.6)
∂xj ∂xj
Using the continuity equation ∂vj′ /∂xj = 0 (see Eq. 6.11), the first term is rewritten as

∂
∂v̄i
vi′ v̄i vj′ = vi′ vj′ . (8.7)
∂xj ∂xj
For the second term in Eq. 8.6 we start using ∂v̄j /∂xj = 0

∂ ∂v ′
vi′ (vi′ v̄j ) = v̄j vi′ i (8.8)
∂xj ∂xj
Next, we use Trick 2
!  
∂vi′ ∂ 1 ′ ′ ∂ ∂
v̄j vi′ = v̄j vv = v̄j (k) = (v̄j k) (8.9)
∂xj ∂xj 2 i i ∂xj ∂xj

The third term in Eq. 8.6 can be written as (replace v̄j by vj′ and use the same technique
as in Eq. 8.9)
1 ∂

v′ v′ v′ . (8.10)
2 ∂xj j i i
8.2. The Exact k Equation 93

The first term on the right side of Eq. 8.5 is re-written using Trick 1

1 ∂p′ 1 ∂p′ vi′ 1 ∂v ′ 1 ∂p′ vi′

− vi′ =− + p′ i = − (8.11)
ρ ∂xi ρ ∂xi ρ ∂xi ρ ∂xi
where the continuity equation was used at the last step. The second term on the right
side of Eq. 8.5 can be written
   
∂ 2 vi′ ∂ ∂vi′ ∂ ∂vi′ ∂v ′ ∂vi′
νvi′ = νvi′ =ν ′
vi −ν i (8.12)
∂xj ∂xj ∂x2 ∂xj ∂xj ∂xj ∂xj ∂xj

applying Trick 1 (A = vi′ and B = ∂vi′ /∂xj ). For the first term in Eq. 8.12 we use the
same trick as in Eq. 8.9 so that
    
∂ ∂v ′ ∂ 1 ∂v ′ ∂v ′
ν vi′ i = ν vi′ i + vi′ i =
∂xj ∂xj ∂xj 2 ∂xj ∂xj
  ′ ′  (8.13)
∂ 1 ∂vi vi 1 ∂ 2 vi′ vi′ ∂ 2k
ν =ν =ν
∂xj 2 ∂xj 2 ∂xj ∂xj ∂xj ∂xj

The last term on the right side of Eq. 8.5 is zero because it is time averaging of a
fluctuation, i.e. āb′ = āb¯′ = 0. Now we can assemble the transport equation for the
turbulent kinetic energy. Equations 8.7, 8.9, 8.11, 8.12 and 8.13 give
 
∂v̄j k ∂v̄i ∂ 1 ′ ′ 1 ′ ′ ′ ∂k ∂v ′ ∂vi′
= −vi′ vj′ − vj p + vj vi vi − ν −ν i (8.14)
∂xj ∂xj ∂xj ρ 2 ∂xj ∂xj ∂xj
I II III IV

The terms in Eq. 8.14 have the following meaning.

I Convection.
II Production, P k . The large turbulent scales extract energy from the mean flow.
This term (including the minus sign) is almost always positive. It is largest for
the energy-containing eddies, i.e. for small wavenumbers, see Fig. 5.2. This term
originates from the convection term (the first term on the right side of Eq. 8.6).
It can be noted that the production term is an acceleration term, vj′ ∂v̄i /∂xj , mul-
tiplied by a fluctuating velocity, vi′ , i.e. the product of an inertial force per unit
mass (acceleration) and a fluctuating velocity. A force multiplied with a velocity
corresponds to work per unit time. When the acceleration term and the fluctuating
velocity are in opposite directions (i.e. when P k > 0), the mean flow performs
work on the fluctuating velocity field. When the production term is negative, it
means that the fluctuations are doing work on the mean flow field. In this case, vj′
and the acceleration term, vj′ ∂v̄i /∂xj , have the same sign.

III The two first terms represent turbulent diffusion by pressure-velocity fluctua-
tions, and velocity fluctuations, respectively. The last term is viscous diffusion.
The velocity-fluctuation term originates from the convection term (the last term
on the right side of Eq. 8.6).
IV Dissipation, ε. This term is responsible for transformation of kinetic energy at
small scales to thermal energy. The term (excluding the minus sign) is always
8.2. The Exact k Equation 94

A B

δB

x2 δA

x1

Figure 8.1: The size of the largest eddies (dashed lines) for different velocity profiles.

positive (it consists of velocity gradients squared). It is largest for large wavenum-
bers, see Fig. 5.2. The dissipation term stems from the viscous term (see Eq. 8.12)
in the Navier-Stokes equation. It can be written as vi′ ∂τij′ /∂xj , see Eq. 4.1. The
divergence of τij′ is a force vector (per unit mass), i.e. Ti′ = ∂τij′ /∂xj . The
dissipation term can now be written vi′ Ti′ , which is a scalar product between two
vectors. When the viscous stress vector is in the opposite direction to the fluctuat-
ing velocity, the term is negative (i.e. it is dissipative); this means that the viscose
stress vector performs work and transforms kinetic energy into internal energy.

The transport equation for k can also be written in a simplified easy-to-read sym-
bolic form as
C k = P k + Dk − ε (8.15)
where C k , P k , Dk and ε correspond to terms I-IV in Eq. 8.14.
Above, it is stated that the production takes place at the large energy-containing
eddies, i.e. we assume that the large eddies contribute much more to the production
term more than the small eddies. There are two arguments for this:

1. The Reynolds stresses (which appear in P k ) are larger for large eddies than for
small eddies.
2. The mean flow generates large eddies which will have same time scale as the
mean velocity gradient, ∂v̄i /∂xj . In the fully turbulent region of a boundary
layer, for example, both time scales are proportional to κx2 /uτ . The time scale
of the velocity gradient is given by κx2 /uτ , see Eq. 6.26, and the time scale of
a large eddy is also given by ℓ0 /v0 = κx2 /uτ . Figure 8.1 shows how different
velocity profiles create different largest eddies. The largest eddies created by the
velocity profile A are much smaller than those created by the velocity profile
B, because the gradient of profile A acts over a much shorter length than the
gradient of profile B.
8.2. The Exact k Equation 95

E(κ)
ε(κ)
εκ

εκ+dκ

κ + dκ
dκ
κ

Figure 8.2: Zoom of the energy spectrum for a wavenumber located in Region II or III,
see Fig. 5.2.

In the cascade process (see Section 5.3) we assume that the viscous dissipation, ε,
takes places at the smallest scales. How do we know that the majority of the dissipation
takes place at the smallest scales? First, let us investigate how the time scale varies with
eddy size. Consider the inertial subrange. Let’s denote the energy that is transferred
in spectral space (i.e. from eddy-to-eddy) per unit time by εκ . How large is ε, that it
generating heat, at wavenumber κ which we here denote ε(κ) (see Section 8.2.1 and
Fig. 8.2)? Recall that the viscous dissipation, ε, is expressed as the viscosity times
the square of the velocity gradient, see Eq. 8.14. The velocity gradient for an eddy
characterized by velocity vκ and lengthscale ℓκ can be estimated as
 
∂v vκ
1/2
∝ ∝ vκ2 κ (8.16)
∂x κ ℓκ

since ℓκ ∝ κ−1 . Now we know that the energy spectrum (see Eqs. 5.10 and 5.13),

E ∝ kκ /κ ∝ vκ2 /κ ∝ κ−5/3 ⇒ vκ2 ∝ κ−2/3 (8.17)

in the inertial region. Inserting Eq. 8.17 into Eq. 8.16 gives
   1/2
∂v
∝ κ−2/3 κ ∝ κ−1/3 κ ∝ κ2/3 (8.18)
∂x κ

Thus the viscous dissipation at wavenumber κ can be estimated as (see the last term in
Eq. 8.14)
 2
∂v ′ ∂vi′ ∂v
ε=ν i ⇒ ε(κ) ∝ ∝ κ4/3 , (8.19)
∂xj ∂xj ∂x κ
i.e. ε(κ) does indeed increase for increasing wavenumber.
The energy transferred from eddy-to-eddy per unit time in spectral space can also be
used for estimating the velocity gradient of an eddy. The cascade process assumes that
this energy transfer per unit time is the same for each eddy size, i.e. εκ = ε = vκ3 /ℓκ =
8.3. The Exact k Equation: 2D Boundary Layers 96

ℓ2κ /τκ3 = ℓ20 /τ03 , see Eq. 5.14. We find from ℓ2κ /τκ3 = ℓ20 /τ03 that for decreasing eddy
size (decreasing ℓκ ), the time scale, τκ , also decreases, i.e.
 2/3
ℓκ
τκ = τ0 (8.20)
ℓ0

where τ0 and ℓ0 are constants (they are given by the flow we’re looking at, for example
a boundary layer which has the large scales, τ0 and ℓ0 ). Hence
 
∂v vκ
∝ ∝ τκ−1 ∝ ℓκ−2/3 ∝ κ2/3 , (8.21)
∂x κ ℓκ

which is the same as Eq. 8.18.

8.2.1 Spectral transfer dissipation εκ vs. “true” viscous dissipation, ε

As a final note to the discussion in the previous section, it may be useful to look at the
difference between the spectral transfer dissipation εκ , and the “true” viscous dissipa-
tion, ε; the former is the energy transferred from eddy-to-eddy per unit time, and the
latter is the energy transformed per unit time to internal energy (i.e. increased temper-
ature) for the entire spectrum (occurring mainly at the small, dissipative scales), see
Fig. 5.2. Now consider Fig. 8.2 which shows a zoom of the energy spectrum. We as-
sume that no mean flow energy production occurs between κ and κ+ dκ, i.e. the region
may be in the −5/3 region or in the dissipation region. Turbulent kinetic per unit time
energy enters at wavenumber κ at a rate of εκ and leaves at wavenumber κ + dκ a rate
of εκ+dκ . If κ and κ + dκ are located in the inertial region (i.e. the −5/3 region),
then the usual assumption is that εκ ≃ εκ+dκ and that there is no viscous dissipation to
internal energy, i.e. ε(κ) ≃ 0. If there is viscous dissipation at wavenumber κ (which
indeed is the case if the zoomed region is located in the dissipative region), then ε(κ)
is simply obtained through an energy balance per unit time, i.e.

ε(κ) = εκ+dκ − εκ (8.22)

8.3 The Exact k Equation: 2D Boundary Layers

In 2D boundary-layer flow, for which ∂/∂x2 ≫ ∂/∂x1 and v̄2 ≪ v̄1 , the exact k
equation reads

∂v̄1 k ∂v̄2 k ∂v̄1

+ = −v1′ v2′
∂x1 ∂x2 ∂x2
  (8.23)
∂ 1 ′ ′ 1 ∂k ∂v ′ ∂vi′
− p v2 + v2′ vi′ vi′ − ν −ν i
∂x2 ρ 2 ∂x2 ∂xj ∂xj

Note that the dissipation includes all derivatives. This is because the dissipation term
is at its largest for small, isotropic scales for which all derivatives are of the same order
and hence the usual boundary-layer approximation ∂/∂x1 ≪ ∂/∂x2 does not apply
for these scales.
Figure 8.3 presents the terms in Eq. 8.23 for fully developed channel flow. The left
side is – since the flow is fully developed – zero. In the outer region (Fig. 8.3b) all
terms are negligible except the production term and the dissipation term which balance
each other. This is called local equilibrium, see p. 98. Closer to the wall (Fig. 8.3a) the local equilib-
rium
8.3. The Exact k Equation: 2D Boundary Layers 97

0.3 0.04

0.03
0.2
0.02
0.1
0.01

0 0

−0.01
−0.1
−0.02
−0.2
−0.03

−0.04
0 10 20 30 40 100 200 300 400 500
a) b)
x+
2 x+
2
Figure 8.3: Channel flow at Reτ = 2000. Terms in the k equation scaled by u4τ /ν.
Reτ = 2000. a) Zoom near the wall; b) Outer region. DNS (Direct Numerical Sim-
ulation) data [15, 16]. : P k; : −ε; ▽: −∂v ′ p′ /∂x2 ; +: −∂v2′ vi′ vi′ /2/∂x2 ; ◦:
2 2
ν∂ k/∂x2 .

40 40 0.02

35

30

25

x2 /δ
x+
x+
2

2 20 20 0.01

15

10

0 0 0
0 0.2 0.4 0.6 0.8 1 0.2 0.4 0.6 0.8 1
∂v̄1+ /∂x+
2 −v1′ v2′ /u2τ
(a) Velocity gradient (b) Reynolds shear stress

Figure 8.4: Channel flow at Reτ = 2000. DNS (Direct Numerical Simulation)
data [15, 16].

other terms do also play a role. Note that the production and the dissipation terms close
to the wall are two orders of magnitude larger than in the logarithmic region (Fig. 8.3b).
At the wall the turbulent fluctuations are zero which means that the production term is
zero. Since the region near the wall is dominated by viscosity the turbulent diffusion
terms due to pressure and velocity are also small. The dissipation term and the viscous
diffusion term attain their largest value at the wall and they much be equal to each other
since all other terms are zero or negligible.
The turbulence kinetic energy is produced by its main source term, the production
term, P k = −v1′ v2′ ∂v̄1 /∂x2 . The velocity gradient is largest at the wall (see Fig. 8.4a)
where the shear stress is zero (see Fig. 8.4b)); the former decreases and the magnitude
of the latter increases with wall distance and their product takes its maximum at x+ 2 ≃
11. Since P k is largest here so is also k, see Fig. 6.8. k is transported in the x2 direction
by viscous and turbulent diffusion and it is destroyed (i.e. dissipated) by ε.
8.4. Spatial vs. spectral energy transfer 98

8.4 Spatial vs. spectral energy transfer

In Section 5.3 we discussed spectral transfer of turbulent kinetic energy from large to
small eddies (which also applies to the transport of the Reynolds stresses). In Sec-
tion 8.2 we derived the equation for spatial transport of turbulent kinetic energy. How
are the spectral transfer and the spatial transport related? The reason that we in Sec-
tion 5.3 only talked about spectral transfer was that we assumed homogeneous tur-
bulence in which the spatial derivatives of the time-averaged turbulent quantities are homogeneous
zero, for example ∂v1′2 /∂xi = 0, ∂k/∂xi = 0 etc. (Note that the derivatives of the turbulence
instantaneous turbulent fluctuations are non-zero even in homogeneous turbulence, i.e.
∂v1′ /∂xi 6= 0; the instantaneous flow field in turbulent flow is – as we mentioned at the
beginning of this section, p. 66 – always three-dimensional and unsteady). In homoge-
neous turbulence the spatial transport terms (i.e. the convective term, term I, and the
diffusion terms, term III in Eq. 8.14) are zero. Hence, in homogeneous turbulence there
is no time-averaged spatial transport. However, there is spectral transfer of turbulent
kinetic energy which takes place in wavenumber space, from large to small eddies. The
production term (term II in Eq. 8.14) corresponds to the process in which large energy-
containing eddies extract energy from the mean flow. The dissipation term (term IV in
Eq. 8.14) corresponds to transformation of the turbulent kinetic energy at the small ed-
dies to thermal energy. However, real flows are hardly ever homogeneous. Some flows
may have one or two homogeneous directions. Consider, for example, fully developed
channel turbulent flow. If the channel walls are very long and wide compared to the
distance between the walls, 2δ, then the turbulence (and the flow) is homogeneous in
the streamwise direction and the spanwise direction, i.e. ∂v̄1 /∂x1 = 0, ∂vi′2 /∂x1 = 0,
∂vi′2 /∂x3 = 0 etc.
In non-homogeneous turbulence, the cascade process is not valid. Consider a large,
turbulent eddy at a position xA 2 (see Fig. 6.1) in fully developed channel flow. The
′ ′
instantaneous turbulent kinetic energy, kκ = vκ,i vκ,i /2, of this eddy may either be
transferred in wavenumber space or transported in physical (spatial) space, or both. It
may first be transported in physical space towards the center, and there lose its kinetic
energy to smaller eddies. This should be kept in mind when thinking in terms of the
cascade process. Large eddies which extract their energy from the mean flow may not
give their energy to the slightly smaller eddies as assumed in Figs. 5.2 and 5.1, but kκ
may first be transported in physical space and then transferred in spectral space (i.e. to
a smaller eddy).
In the inertial range (Region II), however, the cascade process is still a good ap-
proximation even in non-homogeneous turbulence. The reason is that the transfer of
turbulent kinetic energy, kκ , from eddy-to-eddy, occurs at a much faster rate than the
spatial transport by convection and diffusion. In other words, the time scale of the cas-
cade process is much smaller than that of convection and diffusion which have no time
to transport kκ in space before it is passed on to a smaller eddy by the cascade process.
We say that the turbulence at these scales is in local equilibrium. The turbulence in local
the buffer layer and the logarithmic layer of a boundary layer (see Fig. 6.2) is in local equilibrium
equilibrium. In Townsend [22], this is (approximately) stated as:
the local rates of turbulent kinetic energy (i.e. production and dissipation)
are so large that aspects of the turbulent motion concerned with these pro-
cesses are independent of conditions elsewhere in the flow.
This statement simply means that production is equal to dissipation, i.e. P k = ε, see
Fig. 8.3.
8.5. The overall effect of the transport terms 99

In summary, care should be taken in non-homogeneous turbulence, regarding the

validity of the cascade process for the large scales (Region I).

8.5 The overall effect of the transport terms

The overall effect (i.e. the net effect) of the production term is to increase k, i.e. if we
integrate the production term over the entire domain, V , we get
Z
P k dV > 0 (8.24)
V

Similarly, the net effect of the dissipation term is a negative contribution, i.e.
Z
−εdV < 0 (8.25)
V

What about the overall effect of the transport terms, i.e. convection and diffusion?
Integration of the convection term over the entire volume, V , gives, using Gauss diver-
gence law, Z Z
∂v̄j k
dV = v̄j knj dS (8.26)
V ∂xj S
where S is the bounding surface of V . This shows that the net effect of the convection
term occurs only at the boundaries. Inside the domain, the convection
R merely transports
k with out adding or subtracting anything to the integral of k, V kdV ; the convection
acts as a source term in part of the domain, but in the remaining part of the domain it
acts as an equally large sink term. Similarly for the diffusion term, we get
Z  
∂ 1 ′ ′ ′ 1 ∂k
− vj vk vk + p′ vj′ − ν V
V ∂xj 2 ρ ∂xj
Z   (8.27)
1 ′ ′ ′ 1 ∂k
=− vj vk vk + p′ vj′ − ν nj dS
S 2 ρ ∂xj
The only net contribution occurs at the boundaries. Hence, Eqs. 8.26 and 8.27 show
that the transport terms only – as the word implies – transports k without giving any
net effect except at the boundaries. Mathematically these terms are called divergence
terms, i.e. they can both be written as the divergence of a vector Aj , divergence
terms
∂Aj
(8.28)
∂xj
where Aj for the convection and the diffusion term reads

  v̄j k  convection term
Aj = 1 ′ ′ ′ 1 ′ ′ ∂k (8.29)
 − v v v + pv −ν diffusion term
2 j k k ρ j ∂xj

8.6 The transport equation for v̄i v̄i /2

The equation for K = v̄i v̄i /2 is derived in the same way as that for vi′ vi′ /2. Multiply
the time-averaged Navier-Stokes equations, Eq. 6.10, by v̄i so that

∂v̄i v̄j 1 ∂ p̄ ∂ 2 v̄i ∂vi′ vj′

v̄i = − v̄i + νv̄i − v̄i . (8.30)
∂xj ρ ∂xi ∂xj ∂xj ∂xj
8.6. The transport equation for v̄i v̄i /2 100

0.8

0.6

0.4

0.2

0
0 20 40 60 80 100
x+
2
Figure 8.5: Channel flow at Reτ = 2000. Comparison of mean and fluctuating dissi-
pation terms, see Eqs. 8.36 and 8.37. Both terms are normalized by u4τ /ν. DNS (Direct
Numerical Simulation) data [15, 16]. : ν(∂v̄1 /∂x2 )2 ; : ε.

Using the continuity equation and Trick 2 the term on the left side can be rewritten as
1
∂v̄i v̄j ∂v̄i ∂v̄i v̄i ∂v̄j K
v̄i = v̄j v̄i = v̄j 2 = (8.31)
∂xj ∂xj ∂xj ∂xj

The viscous term in Eq. 8.30 is rewritten in the same way as the viscous term in Sec-
tion 8.2, see Eqs. 8.12 and 8.13, i.e.

∂ 2 v̄i ∂2K ∂v̄i ∂v̄i

νv̄i =ν −ν . (8.32)
∂xj ∂xj ∂xj ∂xj ∂xj ∂xj

Equations 8.31 and 8.32 inserted in Eq. 8.30 gives

∂v̄j K ∂2K v̄i ∂ p̄ ∂v̄i ∂v̄i ∂vi′ vj′

=ν − −ν − v̄i . (8.33)
∂xj ∂xj ∂xj ρ ∂xi ∂xj ∂xj ∂xj

The last term is rewritten using Trick 1 as

∂vi′ vj′ ∂v̄i vi′ vj′ ∂v̄i

− v̄i =− + vi′ vj′ . (8.34)
∂xj ∂xj ∂xj

Note that the first term on the right side differs to the corresponding term in Eq. 8.14
by a factor of two since Trick 2 cannot be used because v̄i 6= vi′ . Inserted in Eq. 8.33
gives (cf. Eq. 8.14)
 
∂v̄j K ∂v̄i v̄i ∂ p̄ ∂ ∂K ∂v̄i ∂v̄i
= vi′ vj′ − − v̄i vi′ vj′ − ν −ν
∂xj ∂xj ρ ∂xi ∂xj ∂xj ∂xj ∂xj (8.35)
−P k , sink source εmean , sink

On the left side we have the usual convective term. On the right side we find:
• loss of energy to k due to the production term
• work performed by the pressure gradient; in channel flow, for example, this term
gives a positive contribution to K since −v̄1 ∂ p̄/∂x1 > 0
• diffusion by velocity-stress interaction
8.6. The transport equation for v̄i v̄i /2 101

∆T

n ε
ea
εm

K Pk k

Figure 8.6: Transfer of energy between mean kinetic energy (K), turbulent kinetic
energy (k) and internal energy (denoted as an increase in temperature, ∆T ). K =
1 1 ′ ′
2 v̄i v̄i and k = 2 vi vi . The cascade process assumes that the term in red is negligible
(see also Figs. 8.2 and M.3).

• viscous diffusion.
• viscous dissipation, εmean . This corresponds to the dissipation term in Eq. 2.21;
if you replace vi with v̄i and use the continuity equation to get rid of the sec-
ond velocity gradient in S̄ij you find that the dissipation term in Eq. 2.21 (see
Eq. 2.24), is identical to εmean .
Note that the first term in Eq. 8.35 is the same as the first term in Eq. 8.14 but with
opposite sign: here we clearly can see that the main source term in the k equation (the
production term) appears as a sink term in the K equation.
In the K equation the dissipation term and the negative production term (represent-
ing loss of kinetic energy to the k field) read

∂v̄i ∂v̄i ∂v̄i

−ν + vi′ vj′ , (8.36)
∂xj ∂xj ∂xj

and in the k equation the production and the dissipation terms read

∂v̄i ∂v ′ ∂vi′
− vi′ vj′ −ν i (8.37)
∂xj ∂xj ∂xj

The gradient of the time-averaged velocity field, v̄i , is much smoother than the gradient
of the fluctuating velocity field, vi′ . Hence, the dissipation by the turbulent fluctuations,
ε, in the turbulent region (logarithmic region and further out from walls), is much larger
than the dissipation by the mean flow (left side of Eq. 8.36). This is seen in Fig. 8.5
(x+2 & 15). The energy flow from the mean flow to internal energy is illustrated in
Fig. 8.6. The major part of the energy flow goes from K to k and then to dissipation.
In the viscous-dominated wall region (x+ 2
2 . 5), the mean dissipation, ν(∂v̄1 /∂x2 ) ,
is much larger than ε, see Fig. 8.5. At the wall, the mean dissipation takes the value
ν = 1/2000 (normalized by u4τ /ν).
9. Transport equations for Reynolds stresses 102

9 Transport equations for Reynolds stresses

N this section we will derive the transport equation for the Reynolds stress tensor.
I This is an unknown in the time-averaged Navier-Stokes equations, Eq. 6.10, which
must be known before Eq. 6.10 can be solved. The most accurate way to find vi′ vj′ is,
of course, to solve a transport equation for it. This is computationally expensive since
we then need to solve six additional transport equations (recall that vi′ vj′ is symmetric,
i.e. v1′ v2′ = v2′ v1′ etc). Often, some simplifications are introduced, in which vi′ vj′ is
modeled by expressing it as the product of a turbulent viscosity and velocity gradients.
Two-equations models are commonly used in these simplified models; no transport
equation for vi′ vj′ is solved.
In Section 8 we derived transport equations for kinetic turbulent energy, k, which
is the trace of the Reynolds stress tensor vi′ vj′ divided by two, i.e. k = vi′ vi′ /2. This
means that k is equal to half the sum of the diagonal components of vi′ vj′ , i.e. k =
0.5(v1′2 + v2′2 + v3′2 ).
Now let’s start to derive the transport equation for vi′ vj′ . This approach is very simi-
lar to that we used when deriving the k equation in Section 8.2. Steady, incompressible
flow with constant density and viscosity is assumed. Subtract Eq. 6.10 from Eq. 6.6
and divide by density, multiply by vj′ and time average and we obtain

∂
vj′ [vi vk − v̄i v̄k ] =
∂xk
(9.1)
1 ∂ ′ ∂ 2 vi′ ∂v ′ v ′
− vj′ p + νvj′ + i k vj′
ρ ∂xi ∂xk ∂xk ∂xk
Equation 6.10 is written with the index i as free index, i.e. i = 1, 2 or 3 so that the
equation is an equation for v1 , v2 or v3 . Now write Eq. 6.10 as an equation for vj and
multiply this equation by vi′ . We get

∂
vi′ [vj vk − v̄j v̄k ] =
∂xk
(9.2)
1 ∂ ′ ∂ 2 vj′ ∂vj′ vk′ ′
− vi′ p + νvi′ + v
ρ ∂xj ∂xk ∂xk ∂xk i
It may be noted that Eq. 9.2 is conveniently obtained from Eq. 9.1 by simply switching
indices i and j. Adding Eqs. 9.1 and 9.2 together gives

∂ ∂
vj′ [vi vk − v̄i v̄k ] + vi′ [vj vk − v̄j v̄k ] =
∂xk ∂xk
1 ∂p′ 1 ∂p′
− vi′ − vj′
ρ ∂xj ρ ∂xi
(9.3)
∂ 2 vj′ ∂ 2 vi′
+νvi′ + νvj′
∂xk ∂xk ∂xk ∂xk
∂vj′ vk′ ′ ∂vi′ vk′ ′
+ v + v
∂xk i ∂xk j
Note that each line in the equation is symmetric: if you switch indices i and j in any
of the lines nothing changes. This is important: since the tensor vi′ vj′ is symmetric, all
9. Transport equations for Reynolds stresses 103

terms in its transport equation must also be symmetric. Furthermore, you can check
that the equation is correct according to the tensor notation rules. Indices i and j appear
once in each term (not more, not less) and index k (the dummy index) appears exactly
twice in each term (implying summation). Note that it is permitted to use any other
index than k in some terms (but you must not use i and j). You could, for example,
replace k with m in the first term and with q in the second term; however, usually we
use the same dummy index in every term.
Using vi = v̄i + vi′ , the first line can be rewritten as

∂ ∂  ′ 
vj′ [v̄i vk′ + vi′ v̄k + vi′ vk′ ] + vi′ v̄j vk + vj′ v̄k + vj′ vk′ (9.4)
∂xk ∂xk
Using the continuity equation the first terms in the two groups are rewritten as

∂v̄i ∂v̄j
vj′ vk′ + vi′ vk′ (9.5)
∂xk ∂xk
We merge the second terms in the two groups in Eq. 9.4.

∂vi′ v̄k ∂vj′ v̄k ∂v ′ ∂vj′

vj′ + vi′ = v̄k vj′ i + v̄k vi′
∂xk ∂xk ∂xk ∂xk
(9.6)
∂vi′ vj′ ∂vi′ vj′ v̄k
= v̄k =
∂xk ∂xk
The continuity equation was used twice (to get the right side on the first line and to get
the final expression) and the product rule was used backwards to get the second line.
Re-writing also the third terms in the two groups in Eq. 9.4 in the same way, the second
and the third terms in Eq. 9.4 can be written

∂vi′ vj′ v̄k ∂vi′ vj′ vk′

+ (9.7)
∂xk ∂xk
The second line in Eq. 9.3 is also re-written using Trick 1

1 ∂ ′ ′ 1 ∂ ′ ′ 1 ′ ∂vi′ 1 ∂vj′
− vi p − vj p + p + p′ (9.8)
ρ ∂xj ρ ∂xi ρ ∂xj ρ ∂xi

It will later turn out that it is convenient to express all derivatives as ∂/∂xk . Therefore
we re-write the derivative in the two first terms as
∂ ∂ ∂ ∂
= δjk and = δik (9.9)
∂xj ∂xk ∂xi ∂xk

so that
1 ∂ ′ ′ 1 ∂ ′ ′ 1 ′ ∂vi′ 1 ∂vj′
− δjk vi p − δik vj p + p + p′ (9.10)
ρ ∂xk ρ ∂xk ρ ∂xj ρ ∂xi
The third line in Eq. 9.3 is also re-written using Trick 1
   
∂ ∂vj′ ∂ ∂vi′ ∂v ′ ∂vj′
ν ′
vi +ν ′
vj − 2ν i
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
9. Transport equations for Reynolds stresses 104

The product rule is used backwards to merge the two first terms so that the third line in
Eq. 9.3 reads
!
∂ ∂vj′ ∂vi′ ∂v ′ ∂vj′
ν ′
vi ′
+ vj − 2ν i
∂xk ∂xk ∂xk ∂xk ∂xk
! (9.11)
∂ ∂vi′ vj′ ∂vi′ ∂vj′ ∂ 2 vi′ vj′ ∂vi′ ∂vj′
=ν − 2ν =ν − 2ν
∂xx ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk

The terms on the fourth line in Eq. 9.3 are zero because āb′ = āb¯′ = 0. We can now
put everything together. Put the first term in Eq. 9.7 on the left side and the second
term on the right side together with Eqs. 9.5, 9.10 and 9.11 so that

∂ ∂v̄i ∂v̄j
(v̄k vi′ vj′ ) = −vj′ vk′ − vi′ vk′
∂xk ∂xk ∂xk
Cij ,I Pij ,II
!
∂ 1 1 ∂vi′ vj′
− vi′ vj′ vk′ + δjk vi′ p′ + δik vj′ p′ − ν
∂xk ρ ρ ∂xk (9.12)
Dij ,III
 
1 ∂vi′ ∂vj′ ∂vi′ ∂vj′
+ p′ + − 2ν
ρ ∂xj ∂xi ∂xk ∂xk
Πij ,V εij ,IV

Note that the manipulation in Eq. 9.9 allows the diffusion (term III) to be written on a
more compact form. After a derivation, it is always useful to check that the equation is
correct according to the tensor notation rules.
• Every term – or group of terms – should include the free indices i and j (only
once);
• Every term – or group of terms – should be symmetric in i and j;
• A dummy index (in this case index k) must appear exactly twice (=summation)
in every term
Equation 9.12 can also be written in a simplified easy-to-read symbolic form as

Cij = Pij + Dij + Πij − εij (9.13)

where Πij denotes the pressure-strain term

 
p′ ∂vi′ ∂vj′
Πij = + (9.14)
ρ ∂xj ∂xi

Equation 9.12 is the (exact) transport equation of the Reynolds stress, vi′ vj′ . It is called
the Reynolds stress equations. It is an equation for a second-order tensor which con- Reynolds
sists of nine equations, but since it is symmetric we only need to consider six of them. stress
Compare Eq. 9.12 with the equation for turbulent kinetic energy, Eq. 8.14. An alter- equations
native – and maybe easier – way to derive Eq. 8.14 is to first derive Eq. 9.12 and then
take the trace (setting i = j) and divide by two. In both the k and the vi′ vj′ equations
9. Transport equations for Reynolds stresses 105

0.5 0.1

0.05

0 0

−0.05

−0.5 −0.1
0 10 20 30 40 100 200 300 400 500
a) b)
x+
2 x+
2
Figure 9.1: Channel flow at Reτ = 2000. Terms in the v1′2 equation scaled by u4τ /ν. a)
Zoom near the wall; b) Outer region. DNS (Direct Numerical Simulation) data [15,16].
: P11 ; : −ε11 ; : Π11 ; +: −∂(v2′ v1′2 )/∂x2 ; ◦: ν∂ 2 v1′2 /∂x22 .

there is a convection term (I), a production term (II), a diffusion term (III) and a dis-
sipation term (IV). In the vi′ vj′ equation there is a fifth term (V), see Eq. 9.14, which
is called the pressure strain term. The physical meaning of this term is to redistribute pressure
energy between the normal stress components (if we transform Eq. 9.12 to the princi- strain
pal coordinates of vi′ vj′ there are no shear stresses, only normal stresses). The average
of the normal stresses is vav ′2 = v ′ v ′ /3. For a normal stress that is larger than v ′2 , the
i i av
pressure-strain term is negative and vice-versa. It is often called the Robin Hood term Robin Hood
because it – as Robin Hood – “takes from the rich and gives to the poor”. Note that the
trace of the pressure-strain term is zero, i.e.
 
1 ∂vi′ ∂v ′
Πii = p′ + i =0 (9.15)
ρ ∂xi ∂xi

because of the continuity equation and this is the reason why this term does not appear
in the k equation.
For 2D boundary layer flow, Eq. 9.12 reads

∂ ∂ ∂v̄i ∂v̄j
(v̄1 vi′ vj′ ) + (v̄2 vi′ vj′ ) = −vj′ v2′ − vi′ v2′
∂x1 ∂x2 ∂x2 ∂x2
!
∂v ′ v′
∂ 1 1 i j
− vi′ vj′ v2′ + δj2 vi′ p′ + δi2 vj′ p′ − ν (9.16)
∂x2 ρ ρ ∂x2
 ′ 
1 ∂vi ∂vj′ ∂v ′ ∂vj′
+ p′ + − 2ν i
ρ ∂xj ∂xi ∂xk ∂xk

Now let’s look at this equation for fully developed channel flow for which

v̄2 = v̄3 = 0
∂(·) ∂(·) (9.17)
= =0
∂x1 ∂x3
Note that on the second line the streamwise (x1 ) and tne spanwise (x3 ) derivatives
operate on time-averaged quantities; those that operate on instantaneous quantities,
such as in εij and Πij , are not zero.
9.1. Source terms 106

wall wall
x1

Figure 9.2: One-dimensional unsteady heat conduction. In the middle there is a heat
source, Q.

9.1 Source terms

In order to analyze the Reynolds stress equation, Eq. 9.16, we will now look at the
source terms. A positive source term in a transport equation, for example Φ, increases
the value of Φ. A simple example is the one-dimensional unsteady heat conduction
equation (Eq. 2.16 with vi ≡ 0)
∂T ∂2T
=α 2 +Q (9.18)
∂t ∂x1
where Q is a heat source, see Fig. 9.2. If Q is positive, T will increase and vice-versa.
Now we will look at an important source term in the vi′ vj′ equation, namely the
production term. The production term in Eq. 9.16 reads
∂v̄i ∂v̄j
Pij = −vj′ v2′ − vi′ v2′ (9.19)
∂x2 ∂x2

For the v1′2 (i = j = 1), v2′2 (i = j = 2), v3′2 (i = j = 3) and v1′ v2′ (i = 1, j = 2)
equations we get
∂v̄1
P11 = −2v1′ v2′ (9.20a)
∂x2
∂v̄2
P22 = −2v2′ v2′ =0 (9.20b)
∂x2
∂v̄3
P33 = −2v3′ v2′ =0 (9.20c)
∂x2
∂v̄1 ∂v̄2 ∂v̄1
P12 = −v2′ v2′ − v1′ v2′ = −v2′2 (9.20d)
∂x2 ∂x2 ∂x2
using Eq. 9.17.
Figure 9.1 presents the terms in the v1′2 equation (Eq. 9.16 with i = j = 1). As
we saw for the k equation, the production term, P11 , reaches its maximum at x2 ≃ 11
where also v1′2 takes its maximum (Fig. 6.8). The pressure-strain term, Π11 , and the
dissipation term act as sink terms. In the outer region (Fig. 9.1b) the production term
balances the pressure-strain term and the dissipation term.
The terms in the wall-normal stress equation, v2′2 , are shown in Fig. 9.4. Here we
find – as expected – that the pressure-strain term, Π22 , acts as the main source term.
As mentioned previously, Π22 – the “Robin Hood” term – takes from the “rich” v1′2
equation and gives to the “poor” v2′2 equation energy because v1′2 is large and v2′2 is
small.
9.2. Reynolds shear stress vs. the velocity gradient 107

Figure 9.4 presents the terms in the v3′2 equation. As for the v2′2 equation, the main
source term is the pressure-strain term, but it may be noted that here it is positive
near the wall; for the v2′2 equation it goes negative near the wall since the pressure-
strain term dampens v2′2 near the wall. Another difference is that the pressure diffusion
term, 2∂v3′ p′ /∂x2 , is zero (as it is in the v1′2 equation), whereas it near the wall gives
an important contribution in the v2′2 equation (it balances the negative pressure-strain
term).
Figure 9.6 presents the terms in the v1′ v2′ equation. The production term – which
should be a source term – is here negative. Indeed it should be. Recall that v1′ v2′ is
here negative and hence its source must be negative; or, rather, the other way around:
v1′ v2′ is negative because its production term, P12 = −v2′2 ∂v̄1 /∂x2 , is negative since
∂v̄1 /∂x2 > 0. Note that in the upper half of the channel ∂v̄1 /∂x2 < 0 and hence P12
and v1′ v2′ are positive. Furthermore, note that the dissipation, ε12 , is zero. This is be-
cause dissipation takes place at the smallest scales and they are isotropic. That implies
there is no correlation between two fluctuating velocity components, e.g. v1′ v2′ = 0 (in
general, for i 6= j, the stresses vi′ vj′ in isotropic turbulence are zero). Hence, also their
gradients are zero so that
∂v ′ ∂v ′
ε12 = 2ν 1 2 = 0 (9.21)
∂xk ∂xk
However, very close to the wall, x+ 2 ≤ 10, ε12 6= 0 because here the wall affects
the dissipative scales making them non-isotropic; ε12 is positive since v1′ v2′ < 0, see
Fig. 9.6. When looking at the energy spectrum, neither the production term nor the
dissipation of mean kinetic energy enters the spectrum, see Fig. 9.3
The main sink term in the v1′ v2′ equation is the pressure-strain term, π12 , see Fig. 9.6.
But since v1′ v2′ < 0 in the lower half of the channel, it means that it is positive. In order
to understand the sign of π12 , we can look at the pressure-strain term in the principal
coordinate directions and transform it to the x1 − x2 coordinate system, see Eq. 11.52.
If you want to learn more how to derive transport equations of turbulent quantities,
see [23] which can be downloaded here
http://www.tfd.chalmers.se/˜lada/allpaper.html

9.2 Reynolds shear stress vs. the velocity gradient

In boundary-layer type of flow, the Reynolds shear stress and the velocity gradient
∂v̄1 /∂x2 have nearly always opposite signs. For channel flow, for example, Eq. 9.20
shows that P12 is negative (and hence also v1′ v2′ ) in the lower half because ∂v̄1 /∂x2 > 0
and it is positive in the upper half because ∂v̄1 /∂x2 < 0. It can be summarized as:
∂v̄1
• P12 = −v2′2 is the source term in v1′ v2′ equation and it is large
∂x2
⇒ P12 and v1′ v2′ have the same sign. Compare Fig. 9.2 (temperature in o C and
T = 0 at the boundaries): a negative source, Q, gives negative temperature
and vice versa;
∂v̄1
⇒ v1′ v2′ and have opposite sign;
∂x2
∂v̄1
⇒ The production term in the k equation, P k = −v1′ v2′ > 0;
∂x2
• P k is always positive in fully-developed channel flow;
9.2. Reynolds shear stress vs. the velocity gradient 108

∂v̄i ∂v̄i
ν
∂xj ∂xj

∂v̄i
−vi′ vj′
∂xj
∂vi′ ∂vi′
ν
∂xj ∂xj
E

Figure 9.3: Energy spectrum. Transfer of kinetic energy. The cascade process assumes
that the term in red are negligible (see also Fig. 8.2). The term in blue show the viscous
dissipation of the mean flow.

• In general flows, P k is almost always positive.

The fact that v1′ v2′ and ∂v̄1 /∂x2 almost always have different signs can also be
shown by physical argumentation. Consider the flow in a boundary layer, see Fig. 9.7.
A fluid particle is moving downwards (particle drawn with solid line) from x2,B to
x2,A with (the turbulent fluctuating) velocity v2′ . At its new location the v1 velocity is
in average smaller than at its old, i.e. v̄1 (x2,A ) < v̄1 (x2,B ). This means that when the
particle at x2,B (which has streamwise velocity v1 (x2,B )) comes down to x2,A (where
the streamwise velocity is v1 (x2,A )) it has an excess of streamwise velocity compared
to its new environment at x2.A . Thus the streamwise fluctuation is positive, i.e. v1′ > 0
and the correlation between v1′ and v2′ is in average negative (v1′ v2′ < 0).
If we look at the other particle (dashed line in Fig. 9.7) we reach the same con-
clusion. The particle is moving upwards (v2′ > 0), and it is bringing a deficit in v1
so that v1′ < 0. Thus, again, v1′ v2′ < 0. If we study this flow for a long time and
average over time we get v1′ v2′ < 0. If we change the sign of the velocity gradient so
that ∂v̄1 /∂x2 < 0 we will find that the sign of v1′ v2′ also changes.
In cases where the shear stress and the velocity gradient have the same sign (for
example, in a wall jet) the reason is that the other terms (usually the transport terms)
9.2. Reynolds shear stress vs. the velocity gradient 109

0.05 0.02

0.015

0.01

0.005

0 0

−0.005

−0.01

−0.015

−0.05 −0.02
0 10 20 30 40 100 200 300 400 500
a) b)
x+
2 x+
2
Figure 9.4: Channel flow at Reτ = 2000. Terms in the v2′2 equation scaled by u4τ /ν. a)
Zoom near the wall; b) Outer region. DNS (Direct Numerical Simulation) data [15,16].
: −ε22 ; ▽ : −2∂v2′ p′ /∂x2 ; : Π22 ; +: −∂(v2′ v2′2 )/∂x2 ; ◦: ν∂ 2 v2′2 /∂x22 .

0.02
0.15
0.015
0.1
0.01
0.05 0.005

0 0

−0.05 −0.005

−0.01
−0.1
−0.015
−0.15
−0.02
0 10 20 30 40 100 200 300 400 500
a) b)
x+
2 x+
2
Figure 9.5: Channel flow at Reτ = 2000. Terms in the v3′2 equation scaled by u4τ /ν. a)
Zoom near the wall; b) Outer region. DNS (Direct Numerical Simulation) data [15,16].
: −ε33 ; : Π33 ; +: −∂(v2′ v3′2 )/∂x2 ; ◦: ν∂ 2 v3′2 /∂x22 .

are more important than the production term.

10. Correlations 110

0.15 0.05

0.1

0.05

0 0

−0.05

−0.1

−0.05
0 10 20 30 40 100 200 300 400 500
a) b)
x+
2 x+
2
Figure 9.6: Channel flow at Reτ = 2000. Terms in the v1′ v2′ equation scaled by u4τ /ν.
a) Zoom near the wall; b) Outer region. DNS (Direct Numerical Simulation) data [15,
16]. : P12 ; : −ε12 ; ▽ : −∂v1′ p′ /∂x2 ; : Π12 ; +: −∂(v1′ v2′2 )/∂x2 ; ◦:
2 ′ ′ 2
ν∂ v1 v2 /∂x2 .

x2,B

v2′ < 0 v2′ > 0

v̄1 (x2 )
x2,A

x2

x1
Figure 9.7: Sign of the Reynolds shear stress −ρv1′ v2′ in a boundary layer.

10 Correlations
10.1 Two-point correlations
WO -point correlations are useful when describing some characteristics of the tur-
T bulence. By “correlation”, we mean the tendency for two values or variables to
change together, in either the same or opposite way. Pick two points along the x1 axis,
say xA C
1 and x1 , and sample the fluctuating velocity in, for example, the x1 direction.
We can then form the correlation of v1′ at these two points as

B11 (xA C ′ A ′ C
1 , x1 ) = v1 (x1 )v1 (x1 ) (10.1)

Often, it is expressed as

B11 (xA ′ A ′ A
1 , x̂1 ) = v1 (x1 )v1 (x1 + x̂1 ) (10.2)

where x̂1 = xC A
1 − x1 is the separation distance between point A and C.
It is obvious that if we move point A and C closer to each other, B11 increases;
when the two points are moved so close that they merge, then B11 = v ′2 (xA1 ), see
 10.1. Two-point correlations 111

B11

|xA C
1 − x1 |
Figure 10.1: Two-point correlation.

Fig. 10.1. If, on the other hand, we move point C further and further away from point
A, then B11 will go to zero. It is convenient to normalize B11 so that it varies between
−1 and +1. The normalized two-point correlation reads
norm A 1
B11 (x1 , x̂1 ) = v1′ (xA ′ A
1 )v1 (x1 + x̂1 ) (10.3)
v1,rms (xA A
1 )v1,rms (x1 + x̂1 )
where subscript rms denotes root-mean-square, which for v1′ , for example, is defined
as   1/2
v1,rms = v1′2 (10.4)
RMS is the same as standard deviation (Matlab command std) which is the square-
root of the variance (Matlab command var).
Consider a flow where the largest eddies have a length scale of Lint , see Fig. 10.2.
We expect that the two point correlation, B11 , approaches zero for separation distance,
|xA C A B
1 − x1 | > Lint because for separation distances larger than |x1 − x1 | there is no
correlation between v1′ (xA
1 ) and v ′
1 1(x C
). Hence, flows with large eddies will have a
two-point correlation function which decreases slowly with separation distance. For
flows with small eddies, the two-point correlation, B11 , decreases rapidly with x̂1 .
If the flow is homogeneous (see p.98) in the x1 direction, the two-point correlation
does not depend on the location of xA 1 , but it is only dependent on the separation of the
two points, x̂1 , i.e.
norm 1
B11 (x̂1 ) = v ′ (x1 )v1′ (x1 + x̂1 ) (10.5)
v1,rms 1
2

From the two-point correlation, B11 , an integral length scale, Lint , can be com- integral
puted which is defined as the integral of B11 over the separation distance, i.e. length scale
Z ∞
B11 (x1 , x̂1 )
Lint (x1 ) = A C
dx̂1 (10.6)
0 v1,rms v1,rms

The integral length scale represents the length scale of the large energy-containing
eddies. If the flow is homogeneous in the x1 direction then Lint does not depend on
x1 , and the integral length scale is then computed as
Z ∞
norm
Lint = B11 (x̂1 )dx̂1 (10.7)
0
10.1. Two-point correlations 112

Lint
xA
1

B11

x1
(a) Small integral length scale

xA
1
Lint

B11

x1

(b) Large integral length scale

Figure 10.2: Schematic relation between the two-point correlation, the largest eddies
(thick lines) and the integral length scale, Lint .
10.2. Auto correlation 113

10.2 Auto correlation

Auto correlation is a “two-point correlation” in time, i.e. the correlation of a turbulent
fluctuation with a separation in time. If we again choose the v1′ fluctuation, the auto
correlation reads
B11 (tA , t̂ ) = v1′ (tA )v1′ (tA + t̂ ) (10.8)
where t̂ = tC − tA , is the time separation distance between time A and C. If the mean
flow is steady, the “time direction” is homogeneous and B11 is independent on tA ; in
this case the auto-correlation depends only on time separation, t̂, i.e.

B11 (t̂ ) = v1′ (t)v1′ (t + t̂ ) (10.9)

where the right side is time-averaged over t.

The normalized auto-correlation reads

norm 1
B11 (t̂ ) = v ′ (t)v1′ (t + t̂ ) (10.10)
v1,rms 1
2

In analogy to the integral length scale, Lint , the integral time scale, Tint , is defined integral
as (assuming steady flow) time scale
Z ∞
norm
Tint = B11 (t̂)dt̂ (10.11)
0

The integral time scale represents the time scale of the large energy-containing eddies.

10.3 Taylor’s hypothesis of frozen turbulence

norm
The autocorrelation, B11 (t̂), is much easier to measure than the two-point correla-
norm
tion, B11 (x̂1 ). Let’s try to obtain the two-point correlation from the autocorrelation.
Consider the velocity fluctuation at point A (i.e. v1′A (t)) and B (i.e. v1′B (t)), see
Fig. 10.3. Assume that the turbulent fluctuation at point A is transported by the mean
velocity, v̄1 , in a frozen state to point B. This is called Taylor’s hypothesis of frozen
turbulence. This asumption is better the smaller the turbulence intensity, v1,rms /v̄1 . It
takes x̂1 /v̄1 seconds for the fluid particle at point A to reach point B. Based on Taylor’s
hypothesis we can estimate v1′ at point A by measuring v1′ at point B x̂1 /v̄1 seconds
later. This gives

norm 1 1
B11 (x̂1 ) = v1′A (t)v1′B (t) ≃ 2 v ′B (t + x̂1 /v̄1 )v1′B (t)
2
v1,rms v1,rms 1

Based on Taylor’s hypohesis, the integral length scale can in Fig. 10.3 be estimated
from the integral time scale as
Lint = v̄1 Tint (10.12)
10.3. Taylor’s hypothesis of frozen turbulence 114

v̄1
A B
x2

x̂1
x1

Figure 10.3: Two-point correlation and frozen turbulence.

10.3. Taylor’s hypothesis of frozen turbulence 115

MTF270 Turbulence Modeling

L. Davidson
Division of Fluid Dynamics, Department of Applied Mechanics
Chalmers University of Technology, Göteborg, Sweden
http://www.tfd.chalmers.se/˜lada, [email protected]
This report can be downloaded at
http://www.tfd.chalmers.se/˜lada/comp turb model/lecture notes.html
11. Reynolds stress models and two-equation models 116

11 Reynolds stress models and two-equation models

11.1 Mean flow equations
11.1.1 Flow equations
For incompressible turbulent flow, all variables are divided into a mean part (time av-
eraged) and fluctuating part. For the velocity vector this means that vi is divided into
a mean part v̄i and a fluctuating part vi′ so that vi = v̄i + vi′ . Time average and we get
(see Eq. 6.9 at. p. 77):
∂v̄i
=0 (11.1)
∂xj
∂ρ0 v̄i ∂ ∂ p̄ ∂ 2 v̄i ∂τij
+ (ρ0 v̄i v̄j ) = − +µ − − βρ0 (θ̄ − θ0 )gi (11.2)
∂t ∂xj ∂xi ∂xj ∂xj ∂xj
(note that θ denotes temperature) where ρ0 is a constant reference density, the volume
force fi = −β(θ̄ − θ0 )gi and the turbulent stress tensor (also called Reynolds stress Reynolds
tensor) is written as: stress
τij = ρ0 vi′ vj′ (11.3) tensor
The pressure, p̄, denotes the hydrodynamic pressure, see Eq. 3.22, which means that
when the flow is still (i.e. v̄i ≡ 0), then the pressure is zero (i.e. p̄ ≡ 0). We have
assumed that the temperature variations are small (typically smaller than 10 o C) so
that the density variations can be neglected (using ρ0 ) in all terms except the gravity
term. This is called the Boussinesq approximation. Boussinesq
The body force fi – which was omitted for convenience in Eq. 6.9 – has here been approximation
re-introduced. The body force in Eq. 11.2 is due to buoyancy, i.e. density differences.
The basic form of the buoyancy force is fi = gi where gi denotes gravitational acceler-
ation. Since the pressure, p̄, is defined as the hydrodynamic pressure we have re-written
the buoyancy source as

ρ0 fi → (ρ − ρ0 )gi (11.4)
so that p̄ ≡ 0 when v̄i ≡ 0 (note that the true pressure decreases upwards as ρg∆h
where ∆h denotes change in height). If we let density depend on pressure and temper-
ature, differentiation gives
   
∂ρ ∂ρ
dρ = dθ + dp (11.5)
∂θ p ∂p θ
Our flow is incompressible, which means that the density does not depend on pressure,
i.e. ∂ρ/∂p = 0; it may, however, depend on temperature and mixture composition.
Hence the last term in Eq. 11.5 is zero and we introduce the volumetric thermal expan-
sion, β, so that
 
1 ∂ρ
β=− ⇒
ρ0 ∂θ p (11.6)
dρ = −ρ0 βdθ ⇒ ρ − ρ0 = −βρ0 (θ − θ0 )
where β is a physical property which is tabulated in physical handbooks. For a perfect
gas it is simply β = θ−1 (with θ in degrees Kelvin). Now we can re-write the buoyancy
term as
(ρ − ρ0 )gi = −ρ0 β(θ̄ − θ0 )gi (11.7)
11.2. The exact vi′ vj′ equation 117

which is the last term in Eq. 11.2. Consider the case where x3 is vertically upwards.
Then gi = (0, 0, −g) and a large temperature in Eq. 11.7 results in a force vertically
upwards, which agrees well with our intuition.

11.1.2 Temperature equation

The instantaneous temperature, θ, is also decomposed into a mean and a fluctuating
component as θ = θ̄ + θ′ . The transport equation for θ reads (see Eq. 2.16 where
temperature was denoted by T )

∂θ ∂vi θ ∂2θ
+ =α (11.8)
∂t ∂xi ∂xi ∂xi

where α = k/(ρcp ), see Eq. 2.16 on p. 26. Introducing θ = θ̄ + θ′ we get

∂ θ̄ ∂v̄i θ̄ ∂ 2 θ̄ ∂v ′ θ′
+ =α − i (11.9)
∂t ∂xi ∂xi ∂xi ∂xi
The last term on the right side is an additional term whose physical meaning is turbulent
heat flux vector. This is similar to the Reynolds stress tensor on the right side of the
time-averaged momentum equation, Eq. 11.2. The total heat flux vector – viscous plus
turbulent – in Eq. 11.9 reads (cf. Eq. 2.12)

qi,tot qi qi,turb ∂ θ̄
= + =α − vi′ θ′ (11.10)
ρcp ρcp ρcp ∂xi

11.2 The exact vi′ vj′ equation

Now we want to solve the time-averaged continuity equation (Eq. 11.1) and the three
momentum equations (Eq. 11.2). Unfortunately there are ten unknowns; the four usual
ones (v̄i , p̄) plus six turbulent stresses, vi′ vj′ . We must close this equation system; it is
called the closure problem. We must find some new equations for the turbulent stresses. closure
We need a turbulence model. problem
The most comprehensive turbulence model is to derive exact transport equations
for the turbulent stresses. An exact equation for the Reynolds stresses can be derived
from the Navies-Stokes equation. It is emphasized that this equation is exact; or, rather,
as exact as the Navier-Stokes equations. The derivation follows the steps below.

• Set up the momentum equation for the instantaneous velocity vi = v̄i + vi′ →
Eq. (A)
• Time average → equation for v̄i , Eq. (B)
• Subtract Eq. (B) from Eq. (A) → equation for vi′ , Eq. (C)
• Do the same procedure for vj → equation for vj′ , Eq. (D)
• Multiply Eq. (C) with vj′ and Eq. (D) with vi′ , time average and add them together
→ equation for vi′ vj′

In Section 9 at p. 102 these steps are given in some detail. More details can also be
found in [23] (set the SGS tensor to zero, i.e. τija = 0).
11.2. The exact vi′ vj′ equation 118

The final vi′ vj′ -equation (Reynolds Stress equation) reads (see Eq. 9.12)
 
∂vi′ vj′ ∂vi′ vj′ ′ ′ ∂v̄j ′ ′ ∂v̄i p′ ∂vi′ ∂vj′
+ v̄k = −vi vk − vj vk + +
∂t ∂xk ∂xk ∂xk ρ ∂xj ∂xi
Cij Pij Πij
" #
2
∂ p′ vj′ p′ vi′ ∂ vi′ vj′
− vi′ vj′ vk′ + δik + δjk + ν
∂xk ρ ρ ∂xk ∂xk (11.11)
D Dij,ν
ij,t

∂vi′ ∂vj′
−gi βvj′ θ′ − gj βvi′ θ′ − 2ν
∂xk ∂xk
Gij
εij

where Dij,t and Dij,ν denote turbulent and viscous diffusion, respectively. The total
diffusion reads Dij = Dij,t + Dij,ν . This is analogous to the momentum equation
where we have gradients of viscous and turbulent stresses which correspond to viscous
and turbulent diffusion. Equation 11.11 can symbolically be written

Cij = Pij + Πij + Dij + Gij − εij

where
Cij Convection
Pij Production
Πij Pressure-strain
Dij Diffusion
Gij Buoyancy production
εij Dissipation
Which terms in Eq. 11.11 are known and which are unknown? First, let’s think
about which physical quantities we solve for.
v̄i is obtained from the momentum equation, Eq. 11.2
vi′ vj′ is obtained from the modeled RSM equation, Eq. 11.101
Hence the following terms in Eq. 11.11 are known (i.e. they do not need to be modeled)
• The left side
• The production term, Pij
ν
• The viscous part of the diffusion term, Dij , i.e. Dij

• The buoyancy term, Gij (provided that a transport equation is solved for vi′ θ′ ,
Eq. 11.22; if not, vi′ θ′ is obtained from the Boussinesq assumption, Eq. 11.35)
11.3. The exact vi′ θ′ equation 119

11.3 The exact vi′ θ′ equation

If temperature variations occur we must solve for the mean temperature field, see
Eq. 11.9. Then we need the unknown turbulent heat fluxes, vi′ θ′ . To derive its transport
equation, start with the equation for the fluctuating temperature. Subtract Eq. 11.9 from
Eq. 11.8

∂θ′ ∂ ∂ 2 θ′ ∂v ′ θ′
+ (vk′ θ̄ + v̄k θ′ + vk′ θ′ ) = α + k (11.12)
∂t ∂xk ∂xk ∂xk ∂xk
To get the equation for the fluctuating velocity, vi′ , subtract the equation for the mean
velocity v̄i (Eq. 11.2) from the equation for the instantaneous velocity, vi (Eq. 6.6) so
that
∂vi′ ∂ 1 ∂p′ ∂ 2 vi′ ∂v ′ v ′
+ (vk′ v̄i + v̄k vi′ + vk′ vi′ ) = − +ν + i k − gi βθ′ (11.13)
∂t ∂xk ρ ∂xi ∂xk ∂xk ∂xk
Multiply Eq. 11.12 with vi′ and multiply Eq. 11.13 with θ′ , add them together and
time average

∂vi′ θ′ ∂ ∂
+ vi′ (vk′ θ̄ + v̄k θ′ + vk′ θ′ ) + θ′ (v̄i vk′ + v̄k vi′ + vi′ vk′ )
∂t ∂xk ∂xk
(11.14)
θ′ ∂p′ ∂ 2 θ′ ∂ 2 vi′
=− + αvi′ + νθ′ − gi βθ′ θ′
ρ ∂xi ∂xk ∂xk ∂xk ∂xk
The Reynolds stress term in Eq. 11.13 multiplied by θ′ and time averaged is zero, i.e.

∂vi′ vj′ ∂vi′ vj′

θ′ = θ′ = 0
∂xk ∂xk
If you have forgotten the rules for time-averaging, see Section 8.1.
The first term in the two parentheses on line 1 in Eq. 11.14 are combined into two
production terms (using the continuity equation, ∂vk′ /∂xk = 0)

∂ θ̄ ∂v̄
vi′ vk′ + vk′ θ′ (11.15)
∂xk ∂xk
The second term in the two parenthesis on the first line of Eq. 11.14 are re-written using
the continuity equation
!
′ ∂v̄k θ′ ′
∂v̄k vi′ ′ ∂θ′ ′
∂vi′
vi +θ = v̄k vi +θ (11.16)
∂xk ∂xk ∂xk ∂xk

Now the two terms can be merged (product rule backwards, Trick 1)
!
∂θ′ ∂vi′ ∂v ′ θ′ ∂v̄k vi′ θ′
v̄k vi′ +θ ′ = v̄k i = (11.17)
∂xk ∂xk ∂xk ∂xk

where we used the continuity equation to obtain the right side. The last two terms
on the first line in Eq. 11.14 are re-cast into turbulent diffusion terms using the same
procedure as in Eqs. 11.16 and 11.17

∂ ∂ ∂v ′ v ′ θ′
vi′ (vk′ θ′ ) + θ′ (vi′ vk′ ) = i k (11.18)
∂xk ∂xk ∂xk
11.3. The exact vi′ θ′ equation 120

The viscous diffusion terms on the right side are re-written using the product rule back-
wards (Trick 1, see p. 90)
 ′  
′ ∂ 2 θ′ ′ ∂ ∂θ ∂ ′ ∂θ′ ∂θ′ ∂vk′
αvi = αvi =α vi −α
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
 ′ ! (11.19)
′
2
∂ vi ′
′
∂ ∂vi ∂ ′
′
∂vi ∂θ′ ∂vk′
νθ = νθ =ν θ −ν
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
Inserting Eqs. 11.15, 11.17, 11.18 and 11.19 into Eq. 11.14 gives the transport
equation for the heat flux vector vi′ θ′
∂vi′ θ′ ∂ ∂ θ̄ ∂v̄i θ′ ∂p′ ∂ ′ ′ ′
+ v̄k vi′ θ′ = −vi′ vk′ − vk′ θ′ − − v vθ
∂t ∂xk ∂xk ∂xk ρ ∂xi ∂xk k i
Piθ Πiθ Diθ,t
    (11.20)
∂ ∂θ′ ∂ ∂v ′ ∂v ′ ∂θ′
+α vi′ +ν θ′ i − (ν + α) i −gi βθ′2
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
Giθ
Diθ,ν εiθ

where Piθ , Πiθ and Diθ,t denote the production, scramble and turbulent diffusion term,
respectively. The production term includes one term with the mean velocity gradient
and one with the mean temperature gradient. On the last line, Diθ,ν , εiθ and Giθ denote
viscous diffusion, dissipation and buoyancy term, respectively. The unknown terms –
Πiθ , Diθ , εiθ , Giθ – have to be modeled as usual; this is out of the scope of the present
course but the interested reader is referred to [24].
It can be noted that there is no usual viscous diffusion term in Eq. 11.20. The
reason is that the viscous diffusion coefficients are different in the vi equation and
the θ equation (ν in the former case and α in the latter). However, if ν ≃ α (which
corresponds to a Prandtl number of unity, i.e. P r = ν/α ≃ 1, see Eq. 2.17), the
diffusion term in Eq. 11.20 assumes the familiar form
   
∂ ′ ∂θ′ ∂ ′
∂vi′
α vi +ν θ
∂xk ∂xk ∂xk ∂xk
!  
2 ′ ′
∂ vi θ ∂ ′
∂v ′
i ∂ 2 vi′ θ′ ∂ ′ ∂θ′
=α −α θ +ν −ν vi
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
!   (11.21)
 ν  ∂ 2 vi′ θ′ ∂ ∂vi′ ∂ ∂θ′
≃ ν+ −ν θ ′ −ν ′
vi
P r ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
 ν  ∂ 2 vi′ θ′
= ν+ − Diθ,ν
P r ∂xk ∂xk
where Diθ,ν cancels the corresponding term in Eq. 11.20 if α = ν. Often the viscous
diffusion is simplified in this way. Hence, if α ≃ ν the transport equation for vi′ θ′ can
be simplified as
∂vi′ θ′ ∂ ∂ θ̄ ∂v̄i θ′ ∂p′ ∂ ′ ′ ′
+ v̄k vi′ θ′ = −vi′ vk′ − vk′ θ′ − − v vθ
∂t ∂xk ∂xk ∂xk ρ ∂xi ∂xk k i
Piθ Πiθ Diθ,t
 (11.22)
ν  ∂ 2 vi′ θ′ ∂v ′ ∂θ′
+ ν+ − (ν + α) i −gi βθ′2
P r ∂xk ∂xk ∂xk ∂xk
Giθ
Diθ,ν εiθ
11.4. The k equation 121

The same question arises as for the vi′ vj′ equation: which terms need to be modeled
in Eq. 11.22? The following quantities are known:
v̄i is obtained from the momentum equation, Eq. 11.2
θ̄ is obtained from the temperature equation, Eq. 11.9
vi′ vj′ is obtained from the modeled RSM equation, Eq. 11.101

vi′ θ′ is obtained from the modeled vi′ θ′ equation

Hence the following terms in Eq. 11.22 are known (i.e. they do not need to be modeled)
• The left side
• The production term, Piθ
• The viscous diffusion term, Diθ,ν

• The buoyancy term, Giθ (provided that a transport equation is solved for θ′2 ; if
not, θ′2 is usually modeled via a relation to k)

11.4 The k equation

The turbulent kinetic energy is the sum of all normal Reynolds stresses, i.e.
1  ′2  1
k= v1 + v2′2 + v3′2 ≡ vi′ vi′
2 2
By taking the trace (setting indices i = j) of the equation for vi′ vj′ and dividing by two
we get the equation for the turbulent kinetic energy:

∂k ∂k ∂v̄i ∂v ′ ∂vi′
+ v̄j = − vi′ vj′ −ν i
∂t ∂xj ∂xj ∂xj ∂xj
Ck Pk ε

( (11.23)
 ) 2
∂ p′ 1 ∂ k
− vj′ + vi′ vi′ +ν −gi βvi′ θ′
∂xj ρ 2 ∂xj ∂xj k
G
Dtk Dνk

where – as in the vi′ vj′ equation – Dtk and Dνk denotes turbulent and viscous diffusion,
respectively. The total diffusion reads Dk = Dtk + Dνk . Equation 11.23 can symboli-
cally be written:
C k = P k + Dk + Gk − ε (11.24)
The known quantities in Eq. 11.23 are:
v̄i is obtained from the momentum equation, Eq. 11.2
k is obtained from the modeled k equation, Eq. 11.97
Hence the following terms in Eq. 11.23 are known (i.e. they do not need to be modeled)
• The left side
11.5. The ε equation 122

k
• The viscous diffusion term, Di,ν

• The buoyancy term, Gij (provided that a transport equation is solved for vi′ θ′ ,
Eq. 11.22; if not, vi′ θ′ is obtained from the Boussinesq assumption, Eq. 11.35)

11.5 The ε equation

Two quantities are usually used in eddy-viscosity model to express the turbulent vis-
cosity. In the k − ε model, k and ε are used. The turbulent viscosity is estimated –
using dimensional analysis – as the product of a turbulent velocity, U, and length scale,
L,
νt ∝ UL (11.25)
The velocity scale is taken as k 1/2 and the length scale as k 3/2 /ε which gives
k2
νt = Cµ
ε
where Cµ = 0.09. An exact equation for the transport equation for the dissipation

∂vi′ ∂vi′
ε=ν
∂xj ∂xj
can be derived (see, e.g., [25]), but it is very complicated and in the end many terms
are found negligible. It is much easier to look at the k equation, Eq. 11.24, and to setup
a similar equation for ε. The transport equation should include a convective term, C ε ,
a diffusion term, Dε , a production term, P ε , a production term due to buoyancy, Gε ,
and a destruction term, Ψε , i.e.
C ε = P ε + Dε + Gε − Ψε (11.26)
The production and destruction terms, P k and ε, in the k equation are used to for-
mulate the corresponding terms in the ε equation. The terms in the k equation have
the dimension [m2 /s3 ] (look at the unsteady term, ∂k/∂t) whereas the terms in the ε
equation have the dimension [m2 /s4 ] (cf. ∂ε/∂t). Hence, we must multiply P k and ε
by a quantity which has the dimension [1/s]. One quantity with this dimension is the
mean velocity gradient which might be relevant for the production term, but not for the
destruction. A better choice should be ε/k = [1/s]. Hence, we get
ε

P ε + Gε − Ψε = cε1 P k + cε1 Gk − cε2 ε (11.27)
k
where we have added new unknown coefficients in front of each term. The turbulent
diffusion term is expressed in the same way as that in the k equation (see Eq. 11.40)
but with its own turbulent Prandtl number, σε (see Eq. 11.37), i.e.
  
∂ νt ∂ε
Dε = ν+ (11.28)
∂xj σε ∂xj
The final form of the ε transport equation reads
  
∂ε ∂ε ε ∂ νt ∂ε
+ v̄j = (cε1 P k + cε1 Gk − cε2 ε) + ν+ (11.29)
∂t ∂xj k ∂xj σε ∂xj
Note that this is a modeled equation since we have modeled the production, destruction
and turbulent diffusion terms.
11.6. The Boussinesq assumption 123

11.6 The Boussinesq assumption

In the Boussinesq assumption an eddy (i.e. a turbulent) viscosity is introduced to model
the unknown Reynolds stresses in Eq. 11.2. Consider the diffusion terms in the incom-
pressible momentum equation in the case of non-constant viscosity (see Eq. 2.5)
   
∂ ∂v̄i ∂v̄j
ν + − vi′ vj′ (11.30)
∂xj ∂xj ∂xi

Now we want to replace the Reynolds stress tensor, vi′ vj′ , by a turbulent viscosity, νt ,
so that the the diffusion terms can be written
  
∂ ∂v̄i ∂v̄j
(ν + νt ) + (11.31)
∂xj ∂xj ∂xi
Identification of Eqs. 11.30 and 11.31 gives
 
′ ′ ∂v̄i ∂v̄j
− vi vj = νt + (11.32)
∂xj ∂xi
This is identical to the assumption for the Newtonian, viscous stress for incompressible
flow, see Eq. 2.4. Equation 11.32 is not valid upon contraction 3 (the right side will be
zero due to continuity, but not the left side). Hence we add the trace of the left side to
the right side so that
 
′ ′ ∂v̄i ∂v̄j 1 2
vi vj = −νt + + δij vk′ vk′ = −2νt s̄ij + δij k (11.33)
∂xj ∂xi 3 3
Now the equation is valid also when it is contracted (i.e taking the trace); after contrac-
tion both left and right side are equal (as they must be) to vi′ vi′ = 2k. When Eq. 11.33
is included in Eq. 11.2 we replace six turbulent stresses with one new unknown (the
turbulent viscosity, νt ). This is of course a drastic simplification. With the Boussinesq
assumption the momentum equation reads (see Eq. 11.2 and 11.33)
∂ρ0 v̄i ∂
+ (ρ0 v̄i v̄j )
∂t ∂xj
   (11.34)
∂ p̄B ∂ ∂v̄i ∂v̄j
=− + (µ + µt ) + − βρ0 (θ̄ − θ0 )gi
∂xi ∂xj ∂xj ∂xi
where the turbulent kinetic energy (last term in Eq. 11.33) has been incorporated in the
pressure, i.e. p̄B = p̄ + 2k/3.
If the mean temperature equation, Eq. 11.9, is solved for, we need an equation for
the heat flux vector, vi′ θ′ . One option is to solve its transport equation, Eq. 11.22.
However, it is more common to used an eddy-viscosity model for the heat flux vector.
The Boussinesq assumption reads
∂ θ̄
vi′ θ′ = −αt (11.35)
∂xi
where αt denotes the turbulent thermal diffusivity. Note that this is the same assump-
tion as Fourier’s law for a Newtonian flux, see Eq. 2.12. The turbulent thermal diffu-
sivity, αt , is usually obtained from the turbulent viscosity as
νt
αt = (11.36)
σθ
3 contraction means that i is set to j
11.7. Modeling assumptions 124

where σθ is the turbulent Prandtl number; it is an empirical constant which is usually

set to 0.7 ≤ σθ ≤ 0.9. The physical meaning of the turbulent Prandtl number, σθ ,
is analogous to the physical meaning of the usual Prandtl number, see Eq. 2.17; it
defines how efficient the turbulence transports (by diffusion) momentum compared to
how efficient it transports thermal energy, i.e.
νt
σθ = (11.37)
αt
It is important to recognize that the viscosity (ν), the Prandtl number (P r), the
thermal diffusivity (α) are physical parameters which depend on the fluid (e.g. water
or air) and its conditions (e.g. temperature). However, the turbulent viscosity (νt ), the
turbulent thermal diffusivity (αt ) and the turbulent Prandtl number (σθ ) depend on the
flow (e.g. mean flow gradients and turbulence).

11.7 Modeling assumptions

Now we will compare the modeling assumptions for the unknown terms in the vi′ vj′ ,
vi′ θ′ , k and ε equations and formulate modeling assumptions for the remaining terms in
the Reynolds stress equation. This will give us the Reynolds Stress Model [RSM] (also
called the Reynolds Stress Transport Model [RSTM]) where a (modeled) transport
equation is solved for each stress. Later on, we will introduce a simplified algebraic
model, which is called the Algebraic Stress Model [ASM] (this model is also called
Algebraic Reynolds Stress Model, ARSM)
Summary of physical meaning:

Pij , Piθ and P k are production terms of vi′ vj′ , vi′ θ′ and k

Gij , Giθ and Gk are production terms of vi′ vj′ , vi′ θ′ and k due to buoyancy

Dij,t , Diθ,t , Dtk are the turbulent diffusion terms of vi′ vj′ , vi′ θ′ and k

Πiθ is the pressure-scramble terms of vi′ θ′

Πij is the pressure-strain correlation term, which promotes isotropy of the tur-
bulence
εij , εiθ and ε are dissipation of vi′ vj′ , vi′ θ′ and k, respectively. The dissipation
takes place at the small-scale turbulence.

11.7.1 Production terms

In RSM and ASM the production terms are computed exactly

∂v̄j ∂v̄i 1 ∂v̄i

Pij = −vi′ vk′ − vj′ vk′ , Pk = Pii = −vi′ vj′
∂xk ∂xk 2 ∂xj
(11.38)
∂ θ̄ ∂v̄i
Piθ = −vi′ vk′ − vk′ θ′
∂xk ∂xk
k is usually not solved for in RSM but a length-scale equation (i.e. ε or ω) is always
part of an RSM and that equation includes P k .
11.7. Modeling assumptions 125

In the k − ε model, the Reynolds stresses in the production term are computed
using the Boussinesq assumption, which gives
 
′ ′ ∂v̄i ∂v̄j 2
−vi vj = νt + − δij k
∂xj ∂xi 3
   
k ∂v̄i ∂v̄j 2 ∂v̄i
P = νt + − δij k
∂xj ∂xi 3 ∂xj
 
∂v̄i ∂v̄j ∂v̄i
= νt + = νt 2s̄ij (s̄ij + Ωij ) = 2νt s̄ij s̄ij
∂xj ∂xi ∂xj
   
1 ∂v̄i ∂v̄j 1 ∂v̄i ∂v̄j ∂v̄i
s̄ij = + , Ωij = − , = s̄ij + Ωij
2 ∂xj ∂xi 2 ∂xj ∂xi ∂xj
(11.39)
where on the third line we used the fact that s̄ij Ωij = 0 because the product between a
symmetric tensor (s̄ij ) and an asymmetric tensor (Ωij ) is zero. The incompressibility
condition, ∂v̄i /∂xi = 0, was used to obtain the third line.

11.7.2 Diffusion terms

The diffusion terms in the k and ε-equations in the k − ε model are modeled using the
standard gradient hypothesis which reads
  
∂ νt ∂k
Dk = ν+
∂xj σk ∂xj
   (11.40)
∂ νt ∂ε
Dε = ν+
∂xj σε ∂xj
The gradient hypothesis simply assumes that turbulent diffusion acts as to even out
all inhomogeneities. In other words, it assumes that the turbulent diffusion term, Dtk ,
transports k from regions where k is large to regions where k is small. The turbulent
diffusion flux of k is expressed as
1 ′ ′ ′ νt ∂k
dkj,t = v vv =− (11.41)
2 j i i σk ∂xj
Note that this is the same assumption as Fourier’s law for a Newtonian flux, see
Eq. 2.12. Only the triple correlations are included since the pressure diffusion usu-
ally is negligible (see Fig. 8.3 at p. 97). Taking the divergence of Eq. 11.41 (including
the minus sign in Eq. 11.23) gives the turbulent diffusion term in Eq. 11.40.
Solving the equations for the Reynolds stresses, vi′ vj′ , opens possibilities for a more
advanced model of the turbulent diffusion terms. Equation 11.41 assumes that if the
gradient is zero in xj direction, then there is no diffusion flux in that direction. A more
general gradient hypothesis can be formulated without this limitation, e.g.
∂k
dkj,t,G ∝ vj′ vk′ (11.42)
∂xk
which is called the general gradient diffusion hypothesis (GGDH). It was derived in
[26] from the transport equation of the triple correlation vj′ vi′ vi′ . In GGDH the turbulent
flux dk1,t,G , for example, is computed as
∂k ∂k ∂k
dk1,t,G ∝ v1′ v1′ + v1′ v2′ + v1′ v3′ (11.43)
∂x1 ∂x2 ∂x3
11.7. Modeling assumptions 126

Hence, even if ∂k/∂x1 = 0 the diffusion flux dk1,t,G may be non-zero. A quantity of
dimension [s] must be added to get the correct dimension, and as in Eq. 11.27 we take
k/ε so that
k ∂k
dkj,t,G = ck vj′ vk′ (11.44)
ε ∂xk
The diffusion term, Dtk , in the k equation is obtained by taking the divergence of this
equation
 
∂dkj,t,G ∂ k ∂k
Dtk = = ck vj′ vk′ (11.45)
∂xj ∂xj ε ∂xk
This diffusion model may be used when the k equation is solved in an RSM or an ASM.
The corresponding diffusion terms for the ε and vi′ vj′ equations read
 
ε ∂ ′ ′ k ∂ε
Dt = cε vj vk
∂xj ε ∂xk
! (11.46)
∂ ′ k ∂vi′ vj′
Dij,t = ck vk vm ′
∂xk ε ∂xm

Equation 11.46 often causes numerical problems. A more stable alternative is to model
the diffusion terms as in 11.40 which for vi′ vj′ reads
!
∂ νt ∂vi′ vj′
Dij,t = (11.47)
∂xm σk ∂xm

11.7.3 Dissipation term, εij

The dissipation term εij (see Eq. 11.11) is active for the small-scale turbulence. Be-
cause of the cascade process and vortex stretching (see Figs. 5.2 and 5.3) the small-
scale turbulence is isotropic. This means that the velocity fluctuations of the small-
scale turbulence have no preferred direction, see p. 71. This gives:

1. v1′2 = v2′2 = v3′2 .

2. All shear stresses are zero, i.e.
vi′ vj′ = 0 if i 6= j
because the fluctuations in two different coordinate directions are not correlated.
What applies for the small-scale fluctuations (Items 1 and 2, above) must also apply
for the gradients of the fluctuations, i.e.

∂v1′ ∂v1′ ∂v2′ ∂v2′ ∂v3′ ∂v3′

= =
∂xk ∂xk ∂xk ∂xk ∂xk ∂xk
′
(11.48)
∂vi ∂vj
′
= 0 if i 6= j
∂xk ∂xk
The relations in Eq. 11.48 are conveniently expressed in tensor notation as
2
εij = εδij (11.49)
3
where the factor 2/3 is included so that ε = 12 εii is satisfied, see Eqs. 11.11 and 11.23.
11.7. Modeling assumptions 127

v1′ v1′
O

x2

x1

Figure 11.1: Physical illustration of the pressure-strain term.

11.7.4 Slow pressure-strain term

The pressure-strain term, Πij , makes a large contribution to the vi′ vj′ equation. In
Section 9 it was shown that for channel flow it is negative for the streamwise equation,
v1′2 , and positive for the wall-normal, v2′2 , and spanwise, v3′2 , equations. Furthermore,
it acts as a sink term for the shear stress equation. In summary, it was shown that the
term acts as to make the turbulence more isotropic, i.e. decreasing the large normal
stresses and the magnitude of the shear stress and increasing the small normal stresses.
The pressure-strain term is often called the Robin Hood terms, because it “takes from
the rich and gives to the poor”.
The role of the pressure strain can be described in physical terms as follows. As-
sume that two fluid particles with fluctuating velocities v1′ bounce into each other at O
so that ∂v1′ /∂x1 < 0, see Fig. 11.1. As a result the fluctuating pressure p′ increases at
O so that
∂v ′
p′ 1 < 0
∂x1
The fluid in the x1 direction is performing work, moving fluid particles against the
pressure gradient. The kinetic energy lost in the x1 direction is transferred to the x2
and x3 directions and we assume that the collision makes fluid particles move in the
other two directions, i.e.
∂v2′ ∂v3′
> 0, >0 (11.50)
∂x2 ∂x3
Indeed, if ∂v1′ /∂x1 < 0, the continuity equation gives ∂v2′ /∂x2 + ∂v3′ /∂x3 > 0.
However, in Eq. 11.50 we assume that not only their sum is positive but also that they
both are positive. If this is to happen the kinetic energy in the x1 direction must be
larger than that in the x2 and x3 direction, i.e. v1′2 > v2′2 and v1′2 > v3′2 . If v3′2 ≃ v1′2 ,
the pressure strain re-distributes kinetic energy from both v1′2 and v3′2 to v2′2 .
Now let’s assume that v1′2 > v2′2 and v1′2 > v3′2 . The amount of kinetic energy
transferred from the x1 direction to the x2 and x3 directions, should be proportional to
11.7. Modeling assumptions 128

the difference of their energies, i.e.

∂v1′ ρ 1 h ′2   i
p′ ∝− v1 − v2′2 + v1′2 − v3′2
∂x1 t2
       
ρ 1 ′2 ρ 3 ′2 1  ′2 ρ 3 ′2
= − v1′2 − v2 + v3′2 = − v1 − v1 + v2′2 + v3′2 = − v1 − k
t 2 t 2 2 t 2
(11.51)

where t denotes a turbulent timescale. The expression in Eq. 11.51 applies only to
the normal stresses, i.e. the principal axis of vi′ vj′ . Let us show that by transform-
ing the fluctuations to a coordinate system which is rotated an angle α = π/4 then
p′ (∂v1′ /∂x2 + ∂v2′ /∂x1 ) ∝ −v1′ v2′ (α = π/4 corresponds to the special case when the
normal stresses are equal). We express Eq. 11.51 in principal coordinates, (x1∗ , x2∗ ),
and then transform the equation to (x1 , x2 ) by rotating it angle α = π/4, see Appendix
S.1. Replacing u12 in Eq. S.6b by v1′ v2′ we get
 
v1′ v2′ = 0.5 v1∗′2 − v ′2
2∗ (11.52)

′ v′
since v1∗ ′ ′
2∗ = v2∗ v1∗ . Now we have transformed the right side of Eq. 11.51 (the
right side on the first line). Next step is to transform the left side, i.e. the velocity
gradients. We use Eqs. S.6b and S.6c: replacing u12 and u21 by ∂v1′ /∂x2 and ∂v2′ /∂x1 ,
respectively, and adding them gives
∂v2′ ∂v ′ ′
∂v1∗ ∂v ′
+ 1 = − 2∗ (11.53)
∂x1 ∂x2 ∂x1∗ ∂x2∗
the pressure-strain term in Eqs. 11.11 and 11.51 can be written
 ′   ′ 
′
∂v2 ∂v1′ ′
∂v1∗ ′
∂v2∗
p + =p − (11.54)
∂x1 ∂x2 ∂x1∗ ∂x2∗
Now we apply Eq. 11.51 in the x1∗ and −x2∗ directions (looking at the right side of
Eq. 11.54) so that
 ′ 
∂v1∗ ∂v ′ 3ρ  ′2 
p′ − 2∗ ∝ − v1∗ − v2∗′2 (11.55)
∂x1∗ ∂x2∗ 2t
Inserting Eqs. 11.52 and 11.54 into Eq. 11.55 gives finally
 ′ 
∂v2 ∂v ′ 3ρ
p′ + 1 ∝ − ρv1′ v2′ (11.56)
∂x1 ∂x2 4t
This shows that the pressure-strain term acts as a sink term in the shear stress equation.
Thus, Eqs. 11.51 and 11.56 lead as to write
   
∂vi′ ∂vj′ ε 2
Φij,1 ≡ p′ + = −c1 ρ vi′ vj′ − δij k (11.57)
∂xj ∂xi k 3
where Φ denotes the modeled pressure-strain term and subscript 1 means the slow part;
the concept “slow” and “rapid” is discussed at p. 130. We have introduced the turbulent
time scale t = k/ε and a constant c1 . This pressure-strain model for the slow part was
proposed by Rotta in 1951 [27].
11.7. Modeling assumptions 129

v̄1

x2 x2

a) x1 b) x3

Figure 11.2: Decaying grid turbulence. The circles (a) and the thin rectangles (b)
illustrates part of the grid which consists of a mesh of circular cylinders.

Let us investigate how Eq. 11.57 behaves for decaying grid turbulence, see Fig. 11.2.
Flow from left with velocity v̄1 passes through a grid. The grid creates velocity gra-
dients behind the grid which generates turbulence. Further downstream the velocity
gradients are smoothed out and the mean flow becomes constant. From this point
and further downstream the flow represents homogeneous turbulence which is slowly
approaching isotropic turbulence; furthermore the turbulence is slowly dying (i.e. de-
caying) due to dissipation. The exact vi′ vj′ equation for this flow reads (no production
or diffusion because of homogeneity)
 
dvi′ vj′ p′ ∂vi′ ∂vj′
v̄1 = + − εij (11.58)
dx1 ρ ∂xj ∂xi
Rotta’s pressure-strain model is supposed to reduce anisotropy. Thus it should be in-
teresting to re-write Eq. 11.58 expressed in the normalized anisotropy Reynolds stress
tensor which is defined as
vi′ vj′ 2
aij = − δij (11.59)
k 3
Note that when the turbulence is isotropic, then aij = 0. We introduce aij (Eq. 11.59),
Rotta’s model (Eq. 11.57) and the model for the dissipation tensor (11.49) into Eq. 11.58
so that  
d(kaij ) 2 ∂k 2
v̄1 + δij = −c1 εaij − δij ε (11.60)
dx1 3 ∂x1 3
Analogously to Eq, 11.58, the k equation in decaying grid turbulence reads
dk
v̄1 = −ε (11.61)
dx1
Inserting Eq. 11.61 in Eq. 11.60 and dividing by k we obtain
daij ε 2 ε ε 2 ε ε
v̄1 = −c1 aij − δij + aij + δij = aij (1 − c1 ) (11.62)
dx1 k 3 k k 3 k k
11.7. Modeling assumptions 130

Provided that c1 > 1 Rotta’s model does indeed reduce non-isotropy as it should.
The model of the slow pressure-strain term in Eq. 11.57 can be extended by in-
cluding terms which are non-linear in vi′ vj′ . To make it general it is enough to include
terms which are quadratic in vi′ vj′ , since according to the Cayley-Hamilton theorem, a
second-order tensor satisfies its own characteristic equation (see Section 1.20 in [28]);
3
this means that terms that are cubic in vi′ vj′ (i.e. vi′ vj′ = vi′ vk′ vk′ vm
′ v ′ v ′ ) can be
m j
′ ′
expressed in terms that are linear and quadratic in vi vj . The most general form of Φij,1
can be formulated as [29]
  
1
Φij,1 = −c1 ρ εaij + c′1 aik akj − δij akℓ aℓk
3
(11.63)
′ ′
vi vj 2
aij = − δij
k 3
aij is an anisotropy tensor whose trace is zero. In isotropic flow all its components are
zero. Note that the right side is trace-less (i.e. the trace is zero). This should be so
since the exact form of Φij is trace-less, i.e. Φii = 2p′ ∂vi′ /∂xi = 0.

11.7.5 Rapid pressure-strain term

Above a model for the slow part of the pressure-strain term was developed using phys-
ical arguments. Here we will carry out a mathematical derivation of a model for the
rapid part of the pressure-strain term.
The notation “rapid” comes from a classical problem in turbulence called the rapid
distortion problem, where a very strong velocity gradient ∂v̄i /∂xj is imposed so that
initially the second term (the slow term) can be neglected, see Eq. 11.65. It is assumed
that the effect of the mean gradients is much larger than the effect of the turbulence,
i.e.  
 ∂v̄i 
 
 ∂xj  (ε/k) → ∞ (11.64)

Thus in this case it is the first term in Eq. 11.65 which gives the most “rapid”
response in p′ . The second “slow” term becomes important first at a later stage when
turbulence has been generated.
Now we want to derive an exact equation for the pressure-strain term, Πij . Since
it includes the fluctuating pressure, p′ , we start by deriving an exact equation for p′
starting from Navier-Stokes equations.
1. Take the divergence of the incompressible
 Navier-Stokes
 equation assuming con-
∂ ∂vi
stant viscosity (see Eq. 6.6) i.e. vj = . . . ⇒ Equation A.
∂xi ∂xj
2. Take the divergence of the incompressible time-averaged
 Navier-Stokes
 equation
∂ ∂v̄i
assuming constant viscosity (see Eq. 6.10) i.e. v̄j = . . . ⇒ Equation
∂xi ∂xj
B.
Subtraction of Equation B from Equation A gives a Poisson equation for the fluc-
tuating pressure p′
1 ∂ 2 p′ ∂v̄i ∂vj′ ∂2  ′ ′ 
= −2 − vi vj − vi′ vj′ (11.65)
ρ ∂xj ∂xj ∂xj ∂xi ∂xi ∂xj
rapid term slow term
11.7. Modeling assumptions 131

y
V

x2 , y2 x

x1 , y1

Figure 11.3: The exact solution to Eq. 11.66. The integral is carried out for all points,
y, in volume V .

The factor two in the rapid term appears because when taking the divergence of the
convective term there are two identical terms, see right-side of Eq. 8.6. For a Poisson
equation
∂2ϕ
=f (11.66)
∂xj ∂xj
there exists an exact analytical solution given by Green’s formula, see Appendix T (it
is derived from Gauss divergence law)
Z
1 f (y)dy1 dy2 dy3
ϕ(x) = − (11.67)
4π V |y − x|

where the integrals at the boundaries vanish because it is assumed that f → 0 at the
boundaries, see Fig. 11.3. Applying Eq. 11.67 on Eq. 11.65 gives
ρ
p′ (x) =
 4π 
Z ′   (11.68)
 ∂v̄i (y) ∂vj (y) ∂2  dy3
2 + vi′ (y)vj′ (y) − vi′ (y)vj′ (y) 
 ∂yj ∂yi ∂yi ∂yj  |y − x|
V
rapid term slow term

where dy3 = dy1 dy2 dy3 . Now make two assumptions in Eq. 11.68:

i) the turbulence is homogeneous (i.e. the spatial derivative of all time-averaged

fluctuating quantities is zero). This means that the last term in square brackets
is zero. This requirement is not as drastic as it may sound (although very few
turbulent flows are homogeneous). This term is indeed very small compared to
the second derivative of the instantaneous fluctuations, vi′ (y)vj′ (y).
ii) the variation of ∂v̄i /∂xj in space is small. The same argument can be used as
above: the mean gradient ∂v̄i /∂xj varies indeed much more slowly than the
instantaneous velocity gradient, ∂vj′ (y)/∂yi
11.7. Modeling assumptions 132

Assumption i) means that the last term in the integral in Eq. 11.68 is zero, i.e.
∂ 2 vi′ vj′
=0
∂yi ∂yj
Assumption ii) means that the mean velocity gradient can be taken outside the integral.
Now multiply Eq. 11.68 with ∂vi′ /∂xj + ∂vj′ /∂xi . Since this term is not a function of
y it can be moved in under the integral. We obtain after time averaging
 ′ 
1 ′ ∂vi (x) ∂vj′ (x)
p (x) +
ρ ∂xj ∂xi
Z  ′ 
∂v̄k (x) 1 ∂vi (x) ∂vj′ (x) ∂vℓ′ (y) dy3
= +
∂xℓ 2π V ∂xj ∂xi ∂yk |y − x|
Mijkℓ
(11.69)

Z  
1 ∂vi′ (x) ∂vj′ (x) ∂2 dy3
+ + (vk′ (y)vℓ′ (y))
4π V ∂xj ∂xi ∂yk ∂yℓ |y − x|
Aij

Note that the mean velocity gradient, ∂v̄/∂xℓ , is taken at point x because it has been
moved out of the integral. In order to understand this better, consider the integral
Z L
g(ξ)dξ
f (x) = (11.70)
0 |x − ξ|
Note that x and ξ are coordinates along the same axis (think of them as two different
points along the x axis). If the two points, x and ξ, are far from each other, then the
denominator is large and the contribution to the integral is small. Hence, we only need
to consider ξ points which are close to x. If we assume that g(ξ) varies slowly with ξ,
g(ξ) can be moved out of the integral and since x is close to ξ, Eq. 11.70 can be written
as Z L
dξ
f (x) = g(x) (11.71)
0 |x − ξ|
Going from Eq. 11.70 to Eq. 11.71 corresponds to moving the mean velocity gradient
out of the integral. Equation 11.69 can be written on shorter form as
 
p′ ∂vi′ ∂vj′ ∂v̄k
+ = Aij + Mijkℓ = Φij,1 + Φij,2 (11.72)
ρ ∂xj ∂xi ∂xℓ
where the first term represents the slow term, Φij,1 (see Eq. 11.57), and second term is
the rapid term, Φij,2 (index 2 denotes the rapid part).
Now we will take a closer look at the rapid part (i.e. the second term) of Mijkℓ .
The second term of Mijkℓ in the integral in Eq. 11.69 can be rewritten as
!
∂vj′ (x) ∂vℓ′ (y) ∂ ∂vj
′ (x) ∂ 2 vj′ (x)
= vℓ′ (y) − vℓ′ (y)
∂xi ∂yk ∂yk ∂xi ∂yk ∂xi
!
∂2  ′ ′
 ∂ ′ ∂vℓ′ (y) (11.73)
= v (y)vj (x) − vj (x)
∂yk ∂xi ℓ ∂yk ∂xi
∂2  ′ 
= vℓ (y)vj′ (x)
∂yk ∂xi
11.7. Modeling assumptions 133

∂ 2 vj′ (x)/∂yk ∂xi on line 1 is zero because vj′ (x) is not a function of y. For the same
reason the last term on line 2 is zero.
Note that the terms above as well as in Eq. 11.69 are two-point correlations, the
two points being x and y. Introduce the distance vector between the two points

ri = yi − xi (11.74)

Differentiating Eq. 11.74 gives

∂ ∂ ∂
= − (11.75)
∂ri ∂yi ∂xi
Equation 11.74 is a coordinate transformation where we replace xi and yi with
I. xi and ri , or
II. yi and ri .
Assumption i) at p. 131 gives that ∂/∂xi = 0 (Item I) or ∂/∂yi = 0 (Item II). In other
words, the two-point correlations are independent of where in space the two points are
located; they are only dependent on the distance between the two points (i.e. ri ). Hence
we can replace the spatial derivative by the distance derivative, i.e.
∂ ∂
=−
∂xi ∂ri
(11.76)
∂ ∂
=
∂yi ∂ri
We can now write Mijkℓ in Eq. 11.69, using Eqs. 11.73 and 11.76, as
Z  
1 ∂2  ′ ′  ∂ 2  ′ ′  dr3
Mijkℓ = − vv + vv
2π V ∂rk ∂ri ℓ j ∂rk ∂rj ℓ i |r| (11.77)
= aijkℓ + ajikℓ

It can be shown that aijkℓ is symmetric with respect to index j and ℓ (recall that vℓ′ and
vj′ are not at the same point but separated by ri ), i.e.

aijkℓ = aiℓkj (11.78)

see Appendix H on p. 287. Furthermore, Eq. 11.77 is independent of in which order

the two derivatives are taken, so that aijkℓ is symmetric with respect to i and k, i.e.

aijkℓ = akjiℓ (11.79)

Now let us formulate a general expression of aijkℓ which is linear in vi′ vj′ and
symmetric in (j, ℓ) and (i, k). We get

aijkℓ = c1 δik vj′ vℓ′

+ c2 δjℓ vi′ vk′
+ c3 (δij vk′ vℓ′ + δkj vi′ vℓ′ + δiℓ vk′ vj′ + δkℓ vi′ vj′ ) (11.80)
+ c4 δjℓ δik k
+ c5 (δij δkℓ + δjk δiℓ )k
11.7. Modeling assumptions 134

Each line is symmetric in (j, ℓ) and (i, k). For example, on line 3, term 1 & term 3 and
term 2 & term 4 are symmetric with respect to j and ℓ and term 1 & term 2 and term 3
& term 4 are symmetric with respect to i and k.
Consider Eq. 11.69. Here it is seen that if i = j then Mijkℓ = 0 due to the
continuity equation; looking at Eq. 11.77 we get

aiikℓ = 0 (11.81)

Applying this condition to Eq. 11.80 gives

0 = c1 δik vi′ vℓ′ + c2 δiℓ vi′ vk′ + c3 (3vk′ vℓ′ + δki vi′ vℓ′ + δiℓ vk′ vi′ + δkℓ vi′ vi′ )
+ c4 δiℓ δik k + c5 (3δkℓ + δik δiℓ )k
= c1 vk′ vℓ′ + c2 vℓ′ vk′ + c3 (3vk′ vℓ′ + vk′ vℓ′ + vk′ vℓ′ + 2δkℓ k) (11.82)
+ c4 δkℓ k + c5 (3δkℓ + δkℓ )k
= vk′ vℓ′ (c1 + c2 + 5c3 ) + kδkℓ (c4 + 2c3 + 4c5 )

Green’s third formula reads (see Appendix H on p. 287)

aijiℓ = 2vj′ vℓ′ (11.83)

Using Eq. 11.83 in Eq. 11.80 gives

2vj′ vℓ′ = 3c1 vj′ vℓ′ + c2 δjℓ vi′ vi′ + c3 (δij vi′ vℓ′ + δij vi′ vℓ′ + δiℓ vi′ vj′ + δiℓ vi′ vj′ )
+ (3c4 δjℓ + c5 (δij δiℓ + δji δiℓ ))k
(11.84)
= 3c1 vj′ vℓ′ + 2c2 δjℓ k + 4c3 vj′ vℓ′ + (3c4 + 2c5 )δjℓ )k
= vj′ vℓ′ (3c1 + 4c3 ) + δjℓ k(2c2 + 3c4 + 2c5 )

Equations 11.82 and 11.84 give four equations

c1 + c2 + 5c3 = 0, c4 + 2c3 + 4c5 = 0

(11.85)
3c1 + 4c3 − 2 = 0, 2c2 + 3c4 + 2c5 = 0

for the five unknown constants. Let us express all constants in c2 which gives
4c2 + 10 3c2 + 2 50c2 + 4 20c2 + 6
c1 = , c3 = − , c4 = − , c5 = (11.86)
11 11 55 55
Inserting Eq. 11.86 into Eq. 11.80 and 11.72 gives
∂v̄k ∂v̄k
φij,2 = Mijkℓ = (aijkℓ + ajikℓ )
∂xℓ ∂xℓ
   
∂v̄i ∂v̄j ∂v̄k ∂v̄k
= c1 vj′ vℓ′ + vi′ vℓ′ + c2 vi′ vk′ + vj′ vk′
∂xℓ ∂xℓ ∂xj ∂xi
  (11.87)
∂v̄k ∂v̄j ∂v̄i ∂v̄k ∂v̄k
+c3 2δij vk′ vℓ′ + vi′ vℓ′ + vj′ vℓ′ + vk′ vj′ + vk′ vi′
∂xℓ ∂xℓ ∂xℓ ∂xi ∂xj
   
∂v̄i ∂v̄j ∂v̄j ∂v̄i
+c4 k + + c5 k +
∂xj ∂xi ∂xi ∂xj
We find that the c1 term and the second and third part of the c3 term can be merged.
Furthermore, the c2 term and the third and fourth part of the c3 term can be merged as
11.7. Modeling assumptions 135

LRR model LRR-IP model

c1 (Eq. 11.57 1.5 1.5
c2 (Eq. 11.89) 0.4 −
c2 (Eq. 11.90) − 0.6

Table 11.1: Constants in the LRR and LRR-IP pressure-strain models.

well as the c4 and c5 terms; using Eq. 11.85 we get

c2 + 8 8c2 − 2 6c2 + 4 k 4 − 60c2
φij,2 = − Pij − Dij + P + ks̄ij
11 11 11 55
∂v̄k ∂v̄k (11.88)
Dij = −vi′ vk′ − vj′ vk′
∂xj ∂xi
Finally we re-write this equation so that it is expressed in trace-less tensors
 
c2 + 8 2
Φij,2 = −ρ Pij − δij P k
11 3
  (11.89)
8c2 − 2 2 60c2 − 4
−ρ Dij − δij P k − ρks̄ij
11 3 55
where c2 = 0.4. Note that Φii = 0 as we required in Eq. 11.81. This pressure-strain
model is called the LRR model and it was proposed in [30].
All three terms in Eq. 11.89 satisfy continuity and symmetry conditions. It might
be possible to use a simpler pressure-strain model using one or any two terms. Since
the first term is the most important one, a simpler model has been proposed [30, 31]
 
2
Φij,2 = −c2 ρ Pij − δij P k (11.90)
3
It can be noted that there is a close similarity between the Rotta model and Eq. 11.90:
both models represent “return-to-isotropy”, the first expressed in vi′ vj′ and the second
in Pij . The model in Eq. 11.90 is commonly called the IP model (IP=Isotropization
by Production) . Since two terms are omitted we should expect that the best value of
γ should be different than (c2 + 8)/11; a value of γ = 0.6 (c2 = −1.4) was found to
give good agreement with experimental data. Since Eq. 11.90 is a truncated form of
Eq. 11.89 it does not satisfy all requirements that Eq. 11.89 do. Equation 11.90 does
satisfy symmetry condition and continuity but it does not satisfy the integral condition
in Eq. 11.83. Although Eq. 11.90 is a simpler, truncated version of Eq. 11.89, it is
often found to give more accurate results [32]. Since the IP model is both simpler and
seems to be more accurate than Eq. 11.89, it is one of the most popular models of the
rapid pressure-strain term. The coefficients for the slow and rapid terms in the LRR
and LRR-IP models are summarized in Table 11.1

11.7.6 Wall model of the pressure-strain term

When we derived the rapid pressure-strain model using Green’s function in Eq. 11.68
we neglected the influence of any boundaries. In wall-bounded domains it turns out
that the effect of the walls must be taken into account. Both the rapid term in the LRR
model and the IP model must be modified to include wall modeling.
The effect of the wall is to dampen turbulence. There are two main effects whose
underlying physics are entirely different.
11.7. Modeling assumptions 136

x2

x1

Figure 11.4: Modeling of wall correction in pressure-strain terms.

1. Viscosity. Close to the wall the viscous processes (viscous diffusion and dissi-
pation) dominate over the turbulent ones (production and turbulent diffusion).
2. Pressure. When a fluid particle approaches a wall, the presence of the wall is felt
by the fluid particle over a long distance. This is true for a fluid particle carried
by the wind approaching a building as well as for a fluid particle carried by a
fluctuating velocity approaching the wall in a turbulent boundary layer. In both
cases it is the pressure that informs the fluid particle of the presence of the wall.

Since the pressure-strain term includes the fluctuating pressure, it is obviously the
second of these two processes that we want to include in the wall model. Up to now
we have introduced two terms for modeling the pressure-strain term, the slow and the
fast term. It is suitable to include a slow and a fast wall model term, i.e.

Φij = Φij,1 + Φij,2 + Φij,1w + Φij,2w (11.91)

where subscript w denotes wall modeling.

Consider a wall, see Fig. 11.4. The pressure fluctuations dampens the wall-normal
fluctuations. Furthermore, the damping effect of the wall should decrease for increasing
wall distance. We need to scale the wall-normal distance with a relevant quantity and
the turbulent length scale, k 3/2 /ε, seems to be a good candidate. For the wall-normal
fluctuations, the IP wall model reads [33]
ε
Φ22,1w = −2c1w v2′2 f
k
k2
3 (11.92)
f=
2.55|ni,w (xi − xi,w )|ε
where ni,w (xi − xi,w ) denotes the distance to the wall. As explained above, this damp-
ing is inviscid (due to pressure) and affects the turbulent fluctuations well into the
log-region. It has nothing to do with viscous damping. Away from the wall, in the
fully turbulent region, the damping function goes to zero since the distance to the wall,
|ni,w (xi − xi,w )|, increases faster than the turbulence length scale, k 3/2 /ε. In the vis-
cous region the wall model term, Φ22,1w , is not relevant and should be zero since it
should account only for inviscid damping. Moreover, function f should not exceed
one.
The IP wall model for the wall-parallel fluctuations reads
ε
Φ11,1w = Φ33,1w = c1w v2′2 f (11.93)
k
11.8. The k − ε model 137

The requirement that the sum of the pressure strain term should be zero. i.e. Φii,1w =
0, is now satisfied since Φ11,1w + Φ22,1w + Φ33,1w = 0.
The wall model for the shear stress is set as
3 ε
Φ12,1w = − c1w v1′ v2′ f (11.94)
2 k
The factor 3/2 is needed to ensure that Φii,1w = 0 is satisfied when the coordinate sys-
tem is rotated. You can prove this by rotating the matrix [Φ11,1w , Φ12,1w ; Φ21,1w , Φ22,1w ]
and taking the trace of Φ in the principal coordinates system (i.e. taking the sum of the
eigenvalues).
The general formula for a wall that is not aligned with a Cartesian coordinate axis
reads [33]
 
ε 3 ′ ′ 3 ′ ′
Φij,1w = c1w vk′ vm
′ n
k,w nm,w δij − vk vi nk,w nj,w − vk vj ni,w nk,w f
k 2 2
(11.95)
An analogous wall model is used for the rapid part which reads
 
3 3
Φij,2w = c2w Φkm,2 nk,w nm,w δij − Φki,2 nk,w nj,w − Φkj,2 ni,w nk,w f
2 2
(11.96)

11.8 The k − ε model

The exact k equation is given by Eq. 11.23. By inserting the model assumptions for
the turbulent diffusion (Eq. 11.40), the production (Eq. 11.39) and the buoyancy term
(Eqs. 11.35 and 11.36) we get the modeled equation for k
 
∂k ∂k ∂v̄i ∂v̄j ∂v̄i νt ∂ θ̄
+ v̄j = νt + + gi β
∂t ∂xj ∂xj ∂xi ∂xj σθ ∂xi
   (11.97)
∂ νt ∂k
−ε + ν+
∂xj σk ∂xj

In the same way, the modeled ε equation is obtained from Eq. 11.29
 
∂ε ∂ε ε ∂v̄i ∂v̄j ∂v̄i
+ v̄j = cε1 νt +
∂t ∂xj k ∂xj ∂xi ∂xj
   (11.98)
ε νt ∂ θ̄ ε2 ∂ νt ∂ε
+ cε1 gi − cε2 + ν+
k σθ ∂xi k ∂xj σε ∂xj

The turbulent viscosity is computed as

k2
νt = c µ (11.99)
ε
The standard values for the coefficients read

(cµ , cε1 , cε2 , σk , σε ) = (0.09, 1.44, 1.92, 1, 1.3) (11.100)

For details on how to obtain these constants are obtained, see Section 3 in Introduction to turbulence models.
In that report, details on wall-functions and low-Reynolds number models can be found
in Sections 3 and 4, respectively.
11.9. The modeled vi′ vj′ equation with IP model 138

11.9 The modeled vi′ vj′ equation with IP model

With the models for diffusion, pressure-strain and dissipation we get

∂vi′ vj′
= (unsteady term)
∂t
∂vi′ vj′
v̄k = (convection)
∂xk
∂v̄j ∂v̄i
−vi′ vk′ − vj′ vk′ (production)
∂xk ∂xk
 
ε 2
−c1 vi′ vj′ − δij k (pressure strain, slow part)
k 3
 
2
−c2 Pij − δij P k (pressure strain, rapid part)
3
ε ′ ′ 3 ′ ′
+c1w ρ [ vk vm nk nm δij − vi vk nk nj
k 2
3 ′ ′
− vj vk nk ni ]f (pressure strain, wall, slow part)
2
3
+c2w [ Φkm,2 nk nm δij − Φik,2 nk nj
2
3
− Φjk,2 nk ni ]f (pressure strain, wall, rapid part)
2
∂ 2 vi′ vj′
+ν (viscous diffusion)
∂xk ∂xk
" #
∂ νt ∂vi′ vj′
+ (turbulent diffusion)
∂xm σk ∂xm
−gi βvj′ θ′ − gj βvi′ θ′ (buoyancy production)
2
− εδij (dissipation)
3
(11.101)

11.10 Algebraic Reynolds Stress Model (ASM)

The Algebraic Reynolds Stress Model is a simplified Reynolds Stress Model. The
RSM and k − ε models are written in symbolic form (see p. 118 & 121) as:

RSM : Cij − Dij = Pij + Φij − εij

(11.102)
k − ε : C k − Dk = P k − ε

In ASM we assume that the transport (convective and diffusive) of vi′ vj′ is related to
that of k, i.e.
vi′ vj′

Cij − Dij = C k − Dk
k
Inserting Eq. 11.102 into the equation above gives

vi′ vj′

Pij + Φij − εij = Pk − ε (11.103)
k
11.11. Explicit ASM (EASM or EARSM) 139

Thus the transport equation (PDE) for vi′ vj′ has been transformed into an algebraic
equation based on the assumption in Eq. 11.102.
Now we want to re-write this equation as an equation for vi′ vj′ . Insert the IP models
for Φij,1 (Eq. 11.57) and Φij,2 (Eq. 11.90) and the isotropic model for εij (Eq. 11.49)
in Eq. 11.103 and multiply by k/ε so that
   
k ′ ′ 2 k 2 k 2
Pij − c1 vi vj − δij k − c2 Pij − δij P − δij k
ε 3 ε 3 3
k vi′ vj′

+ (Φij,1w + Φij,2w ) = Pk − ε
ε ε
Collect all vi′ vj′ terms so that
 
Pk
vi′ vj′
− 1 + c1 =
ε
   
k 2 2
Pij − c2 Pij − δij P k + Φij,1w + Φij,2w + δij k(−1 + c1 )
ε 3 3
"   #
k 2 2 2
= Pij −δij P k − c2 Pij − δij P k
+ Φij,1w + Φij,2w + δij k(P k /ε − 1 + c1 )
ε 3 3 3

where (2/3)δij P k k/ε was added and subtracted at the last line (shown in boxes). Di-
viding both sides by P k /ε − 1 + c1 gives finally


2 k (1 − c2 ) Pij − 23 δij P k + Φij,1w + Φij,2w
vi′ vj′ = δij k + (11.104)
3 ε c1 + P k /ε − 1
In boundary layer flow Eq. 11.104 reads (without any wall terms, i.e. Φij,1w =
Φij,2w = 0)
2 c1 − 1 + c2 P k /ε k 2 ∂v̄1
−v1′ v2′ = (1 − c2 )
3 (c1 − 1 + P k /ε) ε ∂y
cµ
As can be seen, this model can be seen as an extension of an eddy-viscosity model
where the cµ constant is made a function of the ratio P k /ε.

11.11 Explicit ASM (EASM or EARSM)

Equation 11.104 is an implicit equation for vi′ vj′ , i.e. the Reynolds stresses appear both
on the left and the right side of the equation. It would of course be advantageous to
be able to get an explicit expression for the Reynolds stresses. Pope [34] managed
to derive an explicit expression for ASM in two dimensions. He assumed that the
Reynolds stress tensor can be expressed in the strain-rate tensor, s̄ij , and the vorticity
tensor, Ωij . Furthermore, he showed that the coefficients, G(n) , in that expression can
be a function of not more than the following five invariants
(k 2 /ε2 )s̄ij s̄ji , (k 2 /ε2 )Ω̄ij Ω̄ji , (k 3 /ε3 )s̄ij s̄jk s̄ki
(11.105)
(k 3 /ε3 )Ω̄ij Ω̄jk s̄ki , (k 4 /ε4 )Ω̄ij Ω̄jk s̄km s̄mi
There are five invariants because when s̄ij and Ωij are transformed to principal coordi-
nates, there are three eigenvalue for each of them. Furthermore, s̄ii = 0 which means
there are only five independent invariants.
11.12. Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 140

In two dimension the expression reads

2 k2 k3
vi′ vj′ = kδij + G(1) s̄ij + G(2) 2 (s̄ik Ω̄kj − Ω̄ik s̄kj ) (11.106)
3 ε ε
In general three-dimensional flow, the Reynolds stress tensor depends on 10 ten-
sors, Tijn [34], i.e.

X10
2
vi′ vj′ − kδij = G(n) Tijn
3 n=1
1
Tij1 = s̄ij , Tij2 = s̄ik Ω̄kj − s̄jk Ω̄ki , Tij3 = s̄ik s̄kj − δij s̄ik s̄ki
3
1
Tij4 = Ω̄ik Ω̄kj − δij Ω̄ik Ω̄ki , Tij5 = Ω̄ik s̄km s̄mj − s̄im s̄mk Ω̄kj
3
2
Tij6 = Ω̄im Ω̄mk s̄kj + s̄ik Ω̄km Ω̄mj − δij Ω̄pm Ω̄mk s̄kp
3
Tij7 = Ω̄im s̄mk Ω̄kn Ω̄nj − Ω̄im Ω̄mk s̄kn Ω̄nj , Tij8 = s̄im Ω̄mk s̄kn s̄nj − s̄im s̄mk Ω̄kn s̄nj
2
Tij9 = Ω̄im Ω̄mk s̄kn s̄nj − s̄im s̄mk Ω̄kn Ω̄nj − δij Ω̄pm Ω̄mk s̄kn s̄np
3
Tij10 = Ω̄im s̄mk s̄kn Ω̄np Ω̄pj − Ω̄im Ω̄mk s̄kn s̄np Ω̄pj
(11.107)

where Tijn may depend on the five invariants in Eq. 11.105. Equation 11.107 is a general
form of a non-linear eddy-viscosity model. Any ASM may be written on the form of
Eq. 11.107.
It may be noted that Eq. 11.107 includes only linear and quadratic terms of s̄ij
and Ω̄ij . That is because of Cayley-Hamilton theorem which states that a second-
order tensor satisfies its own characteristic equation (see Section 1.20 in [28]); hence
cubic terms or higher can recursively be expressed in linear (s̄ij ) and quadratic tensors
(s̄ik s̄kj ). Furthermore, note that all terms in Eq. 11.107 are symmetric and traceless as
required by the left side, vi′ vj′ − 2δij k/3.

11.12 Derivation of the Explicit Algebraic Reynolds Stress Model

(EARSM)
The algebraic stress model (ASM) is given by Eq. 11.104. This equation is implicit,
since the Reynolds stresses appear on the right side (in the production and the rapid
pressure-strain terms). In this section we will derive an explicit algebraic Reynolds
stress model (EARSM). The derivation presented here is based on [35]. Whereas the
ASM employs the IP model (Eq. 11.90) for the rapid pressure-strain term, the EARSM
is based on the LRR model (Eq. 11.89). Thus we start with Eq. 11.103 using the Rotta
model for the slow part (Eq. 11.57) and the LRR model (Eq. 11.89) for the rapid part.

 
2 c2 + 8 2
(aij + δij )(P k − ε) = Pij − c1 εaij − Pij − δij P k
3 11 3
 
8c2 − 2 2 60c2 − 4 2
− Dij − δij P k − ks̄ij − δij ε
11 3 55 3
(11.108)
11.12. Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 141

where the anisotropy tensor, aij , in Eq. 11.59 is used on the left side. The wall correc-
tion terms are neglected (as they usually are in the LRR model). Equation 11.108 is
re-arranged as
 
2 c2 + 8 2
aij (P k + c1 ε − ε) = Pij − δij P k − Pij − δij P k
3 11 3
  (11.109)
8c2 − 2 2 60c2 − 4
− Dij − δij P k − ks̄ij
11 3 55

Now we introduce the anisotropy tensor, aij , also on the right side. Start by ex-
pressing the production term, Pij (see Eq. 11.11) in aij , s̄ij and Ω̄ij (see Eq. 9.12)
2 2
Pij = −k(aik + δik )(s̄jk + Ω̄jk ) − k(ajk + δjk )(s̄ik + Ω̄ik )
3 3
4
= − ks̄ij − kaik (s̄jk + Ω̄jk ) − kajk (s̄ik + Ω̄ik ) (11.110)
3
4
= − ks̄ij − k(s̄jk aik + ajk s̄ki ) + k(aik Ω̄kj − Ω̄ik akj )
3
The production term, P k , is equal to 0.5Pii , and Eq. 11.110 gives

P k = −ks̄ik aik (11.111)

so that we can express the P k terms on the right side in Eq. 11.109 as
 
2 k c2 + 8 8c2 − 2 2 9c2 − 5
δij P −1 + + = − δij ks̄ik aik (11.112)
3 11 11 3 11

Dij is the same thing as Pij except that the indices on the velocity gradients (i.e.
the tensors a and Ω̄ in Eq. 11.110 are switched), see Eq. 11.88. Hence we get (cf.
Eq. 11.110)
4
Dij = − ks̄ij − k(s̄jk aki + ajk s̄ki ) − k(aik Ω̄kj − Ω̄ik akj ) (11.113)
3
Collect all terms including Pij , s̄ij and Dij in Eq. 11.109
 
c2 + 8 8c2 − 2 60c2 − 4 2 9c2 − 5
Pij 1 − − Dij − ks̄ij − δij ks̄ik aik
11 11 55 3 11

Inserting Eqs. 11.110 and 11.113 gives

  
c2 + 8 4
k 1− − s̄ij − (s̄jk aik + ajk s̄ki ) + (aik Ω̄kj − Ω̄ik akj )
11 3
 
8c2 − 2 4
−k − s̄ij − (s̄jk aki + ajk s̄ki ) − (aik Ω̄kj − Ω̄ik akj ) (11.114)
11 3
2 9c2 − 5 60c2 − 4
− δij ks̄ik aik − ks̄ij
3 11 55
Gathering all terms including s̄ij gives
 
4 c2 + 8 8c2 − 2 3 60c2 − 4 8
− ks̄ij 1 − + − = − ks̄ij (11.115)
3 11 11 4 55 15
11.12. Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 142

The terms including the product of the tensors a and s in Eq. 11.114 read
 
c2 + 8 8c2 − 2
−k 1 − (s̄jk aik + ajk s̄ki ) + k(s̄jk aki + ajk s̄ki ) =
11 11
(11.116)
k
(s̄jk aki + ajk s̄ki )(9c2 − 5)
11
and the product of the tensors a and Ω̄ in Eq. 11.114 read
 
c2 + 8
8c2 − 2
k 1− aik Ω̄kj − Ω̄ik akj + k (aik Ω̄kj − Ω̄ik akj ) =
11 11
(11.117)
k
(aik Ω̄kj − Ω̄ik akj )(1 + 7c2 )
11
Using Eqs. 11.115, 11.116, 11.117 and the underlined term in Eq. 11.114, Eq. 11.114
can now be written
8 1 + 7c2
−k s̄ij + k (aik Ω̄kj − Ω̄ik akj )
15 11  (11.118)
9c2 − 5 2
+k s̄jk aki + ajk s̄ki − δij ks̄ik aik
11 3
Equation 11.118 is the right side of Eq. 11.109. Insert Eq. 11.118 into Eq. 11.109 and
divide by ε
 k 
P k 8 k 1 + 7c2
aij + c1 − 1 = − s̄ij + (aik Ω̄kj − Ω̄ik akj )
ε ε 15 ε 11
  (11.119)
k 9c2 − 5 2
+ s̄jk aki + ajk s̄ki − δij ks̄ik aik
ε 11 3
The coefficient, c2 , in the LRR model is usually set to c2 = 0.4, see Table 11.1.
In [36–38], they noted that the relation in Eq. 11.119 is substantially simplified if c2 =
5/9. This assumption is made in EARSM [35], which gives
 k 
P 8 4
aij + c1 − 1 = − ŝij + (aik Ω̂kj − Ω̂ik akj ) (11.120)
ε 15 9
where the strain-rate and vorticity tensors are made non-dimensional
k k
ŝij = s̄ij , Ω̂ij = Ωij (11.121)
ε ε
Equation 11.120 can now be written as
6
N aij = − ŝij + (aik Ω̂kj − Ω̂ik akj ) (11.122)
5

9P k 9
N= + c′1 , c′1 = (c1 − 1) (11.123)
4ε 4
The most general form of aij is given by Eq. 11.107. In two-dimensional flow, we
will later show (see Section 11.12.1) that it is sufficient to include only the two first
terms, i.e.
aij = β1 ŝij + β4 (ŝim Ω̂mj − Ω̂im ŝmj ) (11.124)
11.12. Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 143

where we now denote the coefficients by β1 and β4 as in [35]. In order to solve

Eq. 11.122, insert Eq. 11.124 which gives
6
N (β1 ŝij + β4 (ŝik Ω̂kj − Ω̂ik ŝkj )) = − ŝij + (β1 ŝik + β4 (ŝim Ω̂mk − Ω̂im ŝmk ))Ω̂kj
5  
−Ω̂ik β1 ŝkj + β4 (ŝkm Ω̂mj − Ω̂km ŝmj )
6
= − ŝij + β1 (ŝik Ω̂kj − Ω̂ik ŝkj ) + β4 (ŝim Ω̂mk Ω̂kj − 2Ω̂im ŝmk Ω̂kj + Ω̂ik Ω̂km ŝmj )
5
(11.125)

The last tern including β4 can be considerably simplified. Recall that Ω̂11 = Ω̂22 = 0
and Ω̂12 = −Ω̂21 , see Eq. 1.11. We get for the 11 component of Eq. 11.125

ŝ1m Ω̂mk Ω̂k1 − 2Ω̂1m ŝmk Ω̂k1 + Ω̂1k Ω̂km ŝm1

(11.126)
= ŝ11 Ω̂12 Ω̂21 − 2Ω̂12 ŝ22 Ω̂21 + Ω̂12 Ω̂21 ŝ11 = 4ŝ11 Ω̂12 Ω̂21

since ŝ11 = −ŝ22 ( ŝii = 0 due to continuity). In the same way we get 4ŝ22 Ω̂12 Ω̂21 for
the 22 component. The 12 component (and the 21 component) read

ŝ1m Ω̂mk Ω̂k2 − 2Ω̂1m ŝmk Ω̂k2 + Ω̂1k Ω̂km ŝm2

= ŝ12 Ω̂21 Ω̂12 − 2Ω̂12 ŝ21 Ω̂12 + Ω̂12 Ω̂21 ŝ12 (11.127)
= ŝ12 Ω̂21 Ω̂12 + 2Ω̂12 ŝ21 Ω̂21 + Ω̂12 Ω̂21 ŝ12 = 4Ω̂12 Ω̂21 ŝ12

We find that the last term including β4 in Eq. 11.125 can be written as 2IIΩ ŝij where
IIΩ = Ω̂km Ω̂mk = Ω̂12 Ω̂21 + Ω̂21 Ω̂12 = 2Ω̂12 Ω̂21 . Equation 11.125 can now be
re-written as
6
N (β1 ŝij + β4 (ŝik Ω̂kj − Ω̂ik ŝkj )) = − ŝij + β1 (ŝik Ω̂kj − Ω̂ik ŝkj ) + 2β4 IIΩ ŝij
5
(11.128)

Separating ŝij and (ŝik Ω̂kj − Ω̂ik ŝkj ) we get two equations for β1 and β4

6
N β1 = − + 2β4 IIΩ
5 (11.129)
N β4 = β1

so that
6 1
β4 = −
5 N 2 − 2IIΩ
(11.130)
6 N
β1 = −
5 N 2 − 2IIΩ
In order to get the final equation for N , multiply Eq. 11.124 by ŝjk and then take the
trace (which is equal to the production P k /ε = aij ŝji , see Eq. 11.111), i.e.

Pk
≡ −aij ŝji = −β1 IIS + β4 (ŝim Ω̂mj − Ω̂im ŝmj )ŝji (11.131)
ε
11.12. Derivation of the Explicit Algebraic Reynolds Stress Model (EARSM) 144

where IIS = ŝij ŝji . The β4 term reads

(ŝ12 Ω̂21 − Ω̂12 ŝ21 )ŝ11

+(ŝ11 Ω̂12 − Ω̂12 ŝ22 )ŝ21
(11.132)
+(ŝ22 Ω̂21 − Ω̂21 ŝ11 )ŝ12
+(ŝ21 Ω̂12 − Ω̂21 ŝ12 )ŝ22 = 0

since line 2 and 3 are zero and line 1 and 4 cancel each other (Ω̂12 = −Ω̂21 , ŝ12 = ŝ21
and ŝ11 = −ŝ22 ). β1 is now obtained from Eq. 11.131 as

Pk
β1 = − (11.133)
IIS ε
Inserting β1 in Eq. 11.130 gives

Pk 6 N
= IIS (11.134)
ε 5 N 2 − 2IIΩ
Equations 11.123 and 11.134 gives finally an equation for N
96 N
N= IIS + c′1 (11.135)
4 5 N 2 − 2IIΩ
which is re-written as
27
N (N 2 − 2IIΩ ) − IIS N − c′1 (N 2 − 2IIΩ ) = 0
10
so that  
27
N 3 − c′1 N 2 − IIS + 2IIΩ N + 2c′1 IIΩ = 0.
10
The analytical solution for the positive root reads [35]
c′1  p 1/3  p   p 1/3
N= + P1 + P2 + sign P1 − P2 P1 − P2  , P2 ≥ 0
3 " !#
c′1
1/6 1 P 1
N= + 2 P12 − P2 cos arccos p 2 , P2 < 0
3 3 P1 − P2
(11.136)

where 0 ≤ ξ ≤ π in arccos(ξ) and

 
1 ′2 9 2
P1 = c + IIS − IIΩ c′1
27 1 20 3
 3 (11.137)
1 ′2 9 2
P2 = P12 − c1 + IIS + IIΩ
9 10 3
Equation 11.137 is valid for two-dimensional flow. For three-dimensional flow,
Eq. 11.124 includes more (six) of the terms in 11.107. This derivation is given in [35].
It results in a 6th -order equation for N which must be solved numerically.
In the original LRR model, c1 = 1.5 and c2 = 0.4 (see Table 11.1). In the EARSM,
c1 = 1.8 and c2 = 5/9; recall that this choice of c2 simplifies the rapid pressure-strain
model (cf. Eqs. 11.119 and 11.120).
11.13. Boundary layer flow 145

n
x2
v̄1 (x2 )
s x1
Xx
Xx

Figure 11.5: Boundary layer flow.

11.12.1 The linear assumption of aij

In Eq. 11.124 we made an assumption for the relation of aij expressed in ŝij and Ω̂ij .
The argument is as follows [39]. We start from the assumption (cf. Eq. 11.122)

aij = ŝij + (aim Ω̂mj − Ω̂im amj ) (11.138)

Since ŝij and Ω̂ij are independent, aij must include one ŝij term, a0ij = ŝij . Insert this
on the right side of Eq. 11.138

aij = ŝij + (ŝim Ω̂mj − Ω̂im ŝmj ) (11.139)

which gives the term a1ij = ŝim Ω̂mj − Ω̂im ŝmj . We continue to iterate. Inserting a1ij
on the right side of Eq. 11.138 gives (a1im = ŝik Ω̂km − Ω̂ik ŝkm )

aij = ŝij + (ŝik Ω̂km − Ω̂ik ŝkm )Ω̂mj − Ω̂im (ŝmk Ω̂kj − Ω̂mk ŝkj ) = ŝij − 2IIΩ ŝij
(11.140)
see Eqs. 11.125–11.128. We get

a2ij ∝ ŝij = a0ij (11.141)

which shows that it is sufficient to include a0ij and a1ij and hence Eq. 11.124 is complete.
In 3D, we do not get the relation in Eq. 11.141 but another two iterations are re-
quired which gives five independent arij (i.e. r = 4).

11.13 Boundary layer flow

In order to better understand the Reynolds stress equation, Eq. 11.101, it is useful to
look at its source terms which to a large degree govern the magnitude of vi′ vj′ . A
large source term in the equation for the v1′2 equation, for example, will increase v1′2
and vice versa, see Section 9.1. Let us study boundary layer flow (Fig. 11.5) where
v̄2 = 0, v̄1 = v̄1 (x2 ). The production Pij has the form:

∂v̄j ∂v̄i
Pij = −vi′ vk′ − vj′ vk′
∂xk ∂xk
11.13. Boundary layer flow 146

In this special case we get:

∂v̄1
P11 = −2v1′ v2′
∂x2

∂v̄1
P12 = −v2′2
∂x2

P22 = 0

Is v2′2 zero because its production term P22 is zero? No! The sympathetic term Φij ,
which takes from the rich (i.e. v1′2 ) and gives to the poor (i.e. v2′2 ), saves the unfair
situation! The IP model for Φij,1 and Φij,2 (Eq. 11.57) and Φij,2 (Eq. 11.90) gives
 
ε 2
Φ22,1 = c1 k − v2′2 > 0
k 3
1 2 ∂v̄1
Φ22,2 = c2 P11 = −c2 v1′ v2′ >0
3 3 ∂x2

Note also that the dissipation term for the v1′ v2′ is zero, but it takes the value 32 ε for
the v1′2 and v2′2 equations (see p. 126). Since the modeled v1′ v2′ does not have any
dissipation term, the question arises: what is the main sink term in the v1′ v2′ equation?
The answer is, again, the pressure strain term Φ12,1 and Φ12,2 .
12. Reynolds stress models vs. eddy-viscosity models 147

θ̄2 > θ̄0

2
ρ2 < ρ0

0 ∂ θ̄/∂x3 > 0
∂ρ/∂x3 < 0

F x3

θ̄1 < θ̄0

ρ1 > ρ0 1

Figure 12.1: Stable stratification due to positive temperature gradient ∂ θ̄/∂x3 > 0.

12 Reynolds stress models vs. eddy-viscosity models

In this section we present three fundamental physical processes which Reynolds stress
models are able to handle whereas eddy-viscosity models fail. The reason for the
superiority of the former model is in all cases that the production term is treated exactly,
whereas in eddy-viscosity models it is modeled.

12.1 Stable and unstable stratification

In flows where buoyancy is dominating, the temperature has a large effect on the tur-
bulence through the buoyancy term Gij , see Eq. 11.11. If the temperature increases
upwards (i.e. ∂ θ̄/∂x3 > 0), then the flow is stably stratified. This is illustrated in
Fig. 12.1. Consider ∂ θ̄/∂x3 > 0. This means that the density decreases with increas-
ing vertical height, i.e. ∂ρ/∂x3 < 0. If a fluid particle is displaced from its equilibrium
level 0 up to level 2, see Fig. 12.1, it is heavier then the surrounding at this new level
(ρ0 > ρ2 ). Hence, the buoyancy forces the particle back to its original position 0. In
this way the vertical turbulent fluctuations are dampened. Similarly if a particle origi-
nating at level 0, is moved down to level 1. Here it is lighter than its new environment,
and buoyancy takes it back to its original level 0.
For the case of unstable stratification, the situation is reversed. Cold fluid is
located on top of hot fluid, i.e. ∂ θ̄/∂x3 < 0 and ∂ρ/∂x3 > 0. In Fig. 12.1 we would
then have ρ2 > ρ0 . If a fluid particle at level 0 is displaced upwards to level 2, it is
at this location lighter than its new environment; hence it continues to move upwards.
If it is moved down to level 1 it is heavier than its new environments and it will then
continue downwards. Hence, turbulent fluctuations are enhanced. This flow situation
is called unstable stratification.
Now we will investigate how the Reynolds stress model behaves in stable con-
ditions, i.e. when ∂ θ̄/∂x3 > 0. The production term due to buoyancy reads (see
Eq. 11.11)
G33 = 2gβv3′ θ′ (12.1)
since gi = (0, 0, −g). From the equation for the turbulent heat flux, v3′ θ′ (i.e. Eq. 11.22
12.2. Curvature effects 148

with i = 3), we find the production term for v3′ θ′

∂ θ̄ ∂v̄3
P3θ = −v3′ vk′ − vk′ θ′ (12.2)
∂xk ∂xk
In the case illustrated in Fig. 12.1, the production term due to temperature gradient
reads P3θ = −v3′2 ∂ θ̄/∂x3 < 0 (recall that we assume that buoyancy dominates so that
the first term in Eq. 12.2 is much larger than the second one). Since the main source
term in the v3′ θ′ equation, P3θ , is negative, it makes v3′ θ′ < 0 so that G33 < 0 (see
Eq. 12.1). Thus, for the case illustrated in Fig. 12.1, we find that the production term,
G33 , due to buoyancy yields a damping of the vertical fluctuations as it should. Note
that the horizontal turbulent fluctuations are not affected by the buoyancy term, Gij ,
since G11 = G22 = 0 because the gravity is in the x3 direction (i.e. g1 = g2 = 0).
If the situation in Fig. 12.1 is reversed so that ∂ θ̄/∂x3 < 0 the vertical fluctuations
are instead augmented. This is called unstably stratified conditions.
When eddy-viscosity models are used, transport equations are usually not solved
for vi′ θ′ . Instead the heat flux tensor is modeled with an eddy-viscosity assumption
using the Boussinesq assumption, see Eq. 11.35. The buoyancy term, Gk , in the k
equation reads, see Eq. 11.11 (take the trace of Gij and divide by two)

Gk = 0.5Gii = −gi βvi′ θ′ (12.3)

For gi = (0, 0, −g), it reads Gk = gβv3′ θ′ which with Eq. 11.35 gives

νt ∂ θ̄
Gk = −gβ (12.4)
σθ ∂x3
Hence it is seen that in stably stratified conditions, Gk < 0 as required. The differ-
ence between an eddy-viscosity model and a Reynolds stress model, is that the former
reduces k whereas the latter reduces only the vertical fluctuations.

12.2 Curvature effects

When the streamlines in boundary layer flow have a convex curvature, the turbulence
is stabilized. This dampens the turbulence [40, 41], especially the shear stress and
the Reynolds stress normal to the wall. Concave curvature destabilizes the turbu-
lence. The ratio of boundary layer thickness δ to curvature radius R is a common
parameter for quantifying the curvature effects on the turbulence. The work reviewed
by Bradshaw [40] demonstrates that even such small amounts of convex curvature as
δ/R = 0.01 can have a significant effect on the turbulence. In [42] they carried out an
experimental investigation on a configuration simulating the flow near a trailing edge
of an airfoil, where they measured δ/R ≃ 0.03. They reported a 50 percent decrease
of ρv2′2 (Reynolds stress in the normal direction to the wall) owing to curvature. The
reduction of ρv1′2 and −ρv1′ v2′ was also substantial. In addition they reported significant
damping of the turbulence in the shear layer in the outer part of the separation region.
An illustrative model case is curved boundary layer flow, see Fig. 12.2. A polar
coordinate system r − θ with θ̂ locally aligned with the streamline is introduced. As
vθ = vθ (r) (with ∂vθ /∂r > 0 and vr = 0), the radial inviscid momentum equation
(i.e. the terms including viscosity are omitted) degenerates to
ρvθ2 ∂p
− =0 (12.5)
r ∂r
12.2. Curvature effects 149

A
A
0
0
B B
r vθ (r)

Figure 12.2: Flow in a polar coordinate system illustrating streamline curvature. The
streamline is aligned with the θ axis.

v̄1 streamline
x2
θ
x1

Figure 12.3: Streamline curvature occurring when the flow approaches, for example, a
separation region or an obstacle.

Here the variables are instantaneous or laminar. The centrifugal force exerts a force in
the normal direction (outward) on a fluid following the streamline, which is balanced
by the pressure gradient. Since we have assumed that ∂vθ /∂r > 0, Eq. 12.5 shows that
the pressure gradient increases with r. If the fluid by some disturbance (e.g. turbulent
fluctuation) is displaced outwards to level A, it encounters a pressure gradient larger
than that to which it was accustomed at r = r0 , since (vθ )A > (vθ )0 , which from
Eq. 12.5 gives (∂p/∂r)A > (∂p/∂r)0 . Hence the fluid is forced back to r = r0 .
Similarly, if the fluid is displaced inwards to level B, the pressure gradient is smaller
here than at r = r0 and cannot keep the fluid at level B. Instead the centrifugal force
drives it back to its original level.
It is clear from the model problem above that convex curvature, when ∂vθ /∂r > 0,
has a stabilizing effect on (turbulent) fluctuations, at least in the radial direction. It is
discussed below how the Reynolds stress model responds to streamline curvature.
Assume that there is a flat-plate boundary layer flow, see Fig. 12.3. The ratio of
the normal stresses ρv1′2 to ρv2′2 is typically 5 (or more). At one x1 station, the flow
is deflected upwards. How will this affect turbulence? Let us study the effect of con-
cave streamline curvature. The production terms Pij owing to rotational strains (i.e.
12.2. Curvature effects 150

∂Vθ /∂r > 0 ∂Vθ /∂r < 0

convex curvature stabilizing destabilizing

concave curvature destabilizing stabilizing

Table 12.1: Effect of streamline curvature on turbulence.

v̄1 (x2 )

x2

x1

Figure 12.4: The velocity profile for a wall jet.

∂v̄1 /∂x2 , ∂v̄2 /∂x1 ) can be written as (see Eq. 11.11):

∂v̄1
RSM, v1′2 − eq. : P11 = −2v1′ v2′ (12.6a)
∂x2
∂v̄2 ∂v̄1
RSM, v1′ v2′ − eq. : P12 = −v1′2 − v2′2 (12.6b)
∂x1 ∂x2
∂v̄2
RSM, v2′2 − eq. : P22 = −2v1′ v2′ (12.6c)
∂x1
!2
k ∂v̄1 ∂v̄2
k−ε P = νt + (12.6d)
∂x2 ∂x1

The terms in boxes appear because of the streamline curvature.

As long as the streamlines are parallel to the wall, all production is a result of
∂v̄1 /∂x2 . However as soon as the streamlines are deflected, there are more terms
resulting from ∂v̄2 /∂x1 . Even if ∂v̄2 /∂x1 is much smaller than ∂v̄1 /∂x2 it will still
contribute non-negligibly to P12 as ρv1′2 is much larger than ρv2′2 . Thus the magnitude
of P12 will increase (P12 is negative) as ∂v̄2 /∂x1 > 0. An increase in the magnitude of
P12 will increase −v1′ v2′ , which in turn will increase P11 and P22 . This means that ρv1′2
and ρv2′2 will be larger and the magnitude of P12 will be further increased, and so on.
It is seen that there is a positive feedback, which continuously increases the Reynolds
stresses. The turbulence is destabilized owing to concave curvature of the streamlines.
Note that eddy-viscosity models such as k − ε and k − ω models cannot account for
streamline curvature since the two rotational strains, ∂v̄1 /∂x2 and ∂v̄2 /∂x1 , in the
production term are multiplied by the same coefficient (the turbulent viscosity).
Above, we have assumed concave curvature and positive velocity gradient. There
are two other options.
1. If the flow (concave curvature) is a wall jet flow where ∂v̄1 /∂x2 < 0 in the
12.3. Stagnation flow 151

x1
x2

x2
x1

x1
x2

Figure 12.5: The flow pattern for stagnation flow.

outer part (see Fig. 12.4) the situation will be reversed: the turbulence will be
stabilized.
2. If the streamline (and the wall) is deflected downwards, the situation will be as
follows: the turbulence is stabilizing when ∂v̄1 /∂x2 > 0, and destabilizing for
∂v̄1 /∂x2 < 0.

The stabilizing or destabilizing effect of streamline curvature depends on the type

of curvature (convex or concave), and whether there is an increase or decrease in tan-
gential momentum with radial distance from its origin (i.e. the sign of ∂Vθ /∂r). For
convenience, these cases are summarized in Table 12.1. It should be noted that con-
cave or convex depends on from which point the streamline is viewed. The streamline
in Fig. 12.3, for example, is concave when viewed from the wall but convex when
viewed from the origin of the circle with radius r.

12.3 Stagnation flow

The k − ε model does in this type of flow not model the normal stresses properly,
whereas ASM/RSM do. The production term in the k equations for RSM/ASM and
k − ε model in stagnation flow (see Fig. 12.5) due to ∂v̄1 /∂x1 and ∂v̄2 /∂x2 (which
are the dominant strains) is:
∂v̄1 ∂v̄2 ∂v̄1 ′2
RSM : 0.5 (P11 + P22 ) = −v1′2 − v2′2 =− (v − v2′2 ) (12.7)
∂x1 ∂x2 ∂x1 1
( 2  2 )
k ∂v̄1 ∂v̄2
k − ε : P = 2νt + (12.8)
∂x1 ∂x2

where continuity ∂v̄1 /∂x1 = −∂v̄2 /∂x2 has been employed. In RSM, the two terms
are added with sign. In the k − ε model, however, the production will be large because
the difference in sign of the two terms is not taken into account because it includes the
square of the velocity gradients.
12.4. RSM/ASM versus k − ε models 152

12.4 RSM/ASM versus k − ε models

• Advantages with k − ε models (or eddy viscosity models):

i) simple due to the use of an isotropic eddy (turbulent) viscosity

ii) stable via stability-promoting second-order gradients in the mean-flow equa-
tions
iii) work reasonably well for a large number of engineering flows

• Disadvantages:

i) isotropic, and thus not good in predicting normal stresses (v1′2 , v2′2 , v3′2 )
ii) as a consequence of i) it is unable to account for curvature effects
iii) as a consequence of i) it is unable to account for irrotational strains (stag-
nation flow)
iv) in boundary layers approaching separation, the production due to normal
stresses is of the same magnitude as that due to shear stresses [43].

• Advantages with ASM/RSM:

i) the production terms do not need to be modeled

ii) thanks to i) it can selectively augment or damp the stresses due to cur-
vature effects (RSM is better than ASM because the convective terms are
accounted for) and it is more accurate for boundary layers approaching
separation, buoyancy etc.

• Disadvantages with ASM/RSM:

i) RSM is complex and difficult to implement, especially implicit ASM

ii) numerically unstable because small stabilizing second-order derivatives in
the momentum equations (only viscous diffusion)
iii) CPU time consuming
13. Realizability 153

13 Realizability
There are a number of realizability constraints. The usual two ones are that all normal
stresses should stay positive and that the correlation coefficient for the shear stress
should not exceed one, i.e.

vi′2 ≥ 0 for all i

|vi′ vj′ | (13.1)
 1/2 ≤ 1 no summation over i and j, i 6= j
vi′2 vj′2

These criteria are seldom used in RSMs. However, satisfying the first criteria is actually
of importance for eddy-viscosity models in stagnation flow [44]. Assume that the flow
is in the x1 direction and that it approaches the wall (see Fig. 12.5). The Boussinesq
assumption for the normal stress v1′2 reads (cf. Eq. 12.7)

2 ∂v̄1 2
v1′2 = k − 2νt = k − 2νt s̄11 (13.2)
3 ∂x1 3

It is seen that if s̄11 gets too large then v1′2 < 0 which is nonphysical, i.e. non-realizable.
Let’s now briefly repeat the concept “invariants”. This means something that is
independent of the coordinate system. Here we mean independent of rotation of the
coordinate system. If a tensor is symmetric, then we know that it has real eigenvalues
which means that we can rotate the coordinate system so that the off-diagonal compo-
nents vanish (see, e.g., [28]). This corresponds to the principal coordinate directions.
For the strain tensor it means that the off-diagonal components of s̄ij vanish and this
is the coordinate system where the diagonal components become largest (e.g. s̄11 in
Eq. 13.2). This is thus the coordinate system in which the danger of negative v1′2 from
Eq. 13.2 is largest. The equation for finding the eigenvalues of a tensor Cij is (see
e.g. [28] or [45])
|Cij − δij λ| = 0 (13.3)
which gives, in 2D,  
 C11 − λ C12 
 =0 (13.4)
 C21 C22 − λ 
The resulting equation is

λ2 − I12D λ + I22D = 0
I12D = Cii (13.5)
1
I22D = (Cii Cjj − Cij Cij ) = det(Cij )
2
Since Eq. 13.5 is the same irrespectively of the orientation of the original coordinate
system, it follows that its coefficients I12D and I22D are invariants.
In 3D, Eq. 13.3 gives
 
 C11 − λ C12 C13 

 C21 C22 − λ C23  = 0 (13.6)

 C31 C32 C33 − λ 
13.1. Two-component limit 154

which gives

λ3 − I13D λ2 + I23D λ − I33D = 0

I13D = Cii
1 (13.7)
I23D = (Cii Cjj − Cij Cij )
2
3D 1
I3 = (2Cij Cjk Cki − 3Cij Cji Ckk + Cii Cjj Ckk ) = det(Cij )
6
The invariants are I13D , I23D and I23D .
Let’s go back to Eq. 13.2 and assume incompressible 2D flow. The first invariant
reads (cf. Eq. 13.5)

I12D = s̄ii = s̄11 + s̄22 = λ1 + λ2 = 0 (13.8)

It is zero due to the continuity equation. The second invariant of s̄ij reads

I22D = −s̄ij s̄ij /2, (13.9)

(see Eq. 13.5) which is independent of the orientation of the coordinate system (hence
the name ”invariant”). The solution to Eq. 13.5, using Eq. 13.8, is

1/2  s̄ s̄ 1/2
ij ij
λ1,2 = ± −I22D =± (13.10)
2
The eigenvalues of s̄ij correspond to the strains in the principal axis. As discussed
above, we apply Eq. 13.2 in the principal coordinate directions of s̄ij . Hence, s̄11 in
Eq. 13.2 is replaced by the largest eigenvalue so that
2
v1′2 = k − 2νt λ1 (13.11)
3
The requirement v1′2 ≥ 0 gives now together with Eq. 13.11
 1/2
k k 2
νt ≤ = (13.12)
3|λ1 | 3 s̄ij s̄ij
In 3D, Eq. 13.7 instead of Eq. 13.5 is used, and Eq. 13.10 is replaced by [44]
 1/2
2s̄ij s̄ij
|λk | = (13.13)
3
This is a simple modification of an eddy-viscosity model, and it ensures that the normal
stresses stay positive.

13.1 Two-component limit

Another realizability constraint is to require that when vi′2 approaches zero near walls,
it should do so smoothly. One way to ensure this is to require that the derivative of vi′2
should go to zero when vi′2 goes to zero, i.e.

dvi′2
vi′2 → 0 ⇒ →0 (13.14)
dt
13.1. Two-component limit 155

where d/dt denotes the material derivative (think of Eq. 13.14 in Lagrangian coordi-
nates, i.e. we follow a fluid particle as it approaches the wall). Equation 13.14 requires
that when vi′2 approaches zero, the left side (and thus also the right side) of the transport
equation of vi′2 should also do so. Since we are here concerned about the pressure-strain
term, we’ll take a look at how it behaves near walls when vi′2 → 0. This is of some rel-
evance in near-wall turbulence where the wall-normal stress goes to zero faster than the
wall-parallel ones: this state of turbulence is called the two-component limit [46]. The
wall-normal component goes to zero as v2′2 = O(x42 ) whereas v1′2 and v3′2 go to zero as
O(x22 ), see Section 4 in Introduction to turbulence models. Neither the form of Φij,2
in Eq. 11.90 nor Eq. 11.89 satisfy the requirement that Φ22,2 = 0 when v2′2 = 0 [29].
In Eq. 11.90, for example,
2
Φ22,2 → c2 δij P k 6= 0 (13.15)
3

even if v2′2 = 0. Very complex forms of Φij,2 have been proposed [47] [CL96] which
include terms cubic in vi′ vj′ . The CL96 model does satisfy the two-component limit.
Another advantage of the CL96 model is that it does not need any wall distances, which
is valuable in complex geometries.
The models of the slow pressure-strain in Eq. 11.57 (linear model) and Eq. 11.63
(non-linear model) do also not satisfy the two-component limit. The Rotta model
(Eq. 11.57), for example, gives
2ε
Φ22,1 → c1 ρ 6= 0 (13.16)
3

when v2′2 → 0. The only way to ensure that Φ22,1 → 0 is to make c1 → 0 when
the wall is approached. A convenient parameter proposed in [46] is A which is an
expression of A2 and A3 (the second and third invariant of aij , respectively), i.e.

9
A2 = aij aji , A3 = aij ajk aki , A = 1 − (A2 − A3 ) (13.17)
8
The parameter A = 0 in the two-component limit and A = 1 in isotropic turbulence.
Thus A is a suitable parameter to use when damping the constant c1 as the wall is
approached.
 14. Non-linear Eddy-viscosity Models 156

14 Non-linear Eddy-viscosity Models

In traditional eddy-viscosity models the turbulent stress vi′ vj′ is formulated from the
Boussinesq assumption, i.e.
s̄ij
aij = −2νt
 k  (14.1)
1 ∂v̄i ∂v̄j
s̄ij = +
2 ∂xj ∂xi
where the anisotropy tensor is defined as

vi′ vj′ 2
aij ≡ − δij (14.2)
k 3
The relation between the stress vi′ vj′ and the velocity gradient in Eq. 14.1 is, as can
be seen, linear. One way to make eddy-viscosity models more general is to include
non-linear terms of the strain-rate (i.e. the velocity gradient) [34]. A subset of the most
general form (given by 11.107) reads [48]

aij = −2cµ τ s̄ij

 
1

2
+ c1 τ s̄ik s̄kj − s̄ℓk s̄ℓk δij + c2 τ 2 Ω̄ik s̄kj − s̄ik Ω̄kj
3
 
1

+ c3 τ Ω̄ik Ω̄jk − Ω̄ℓk Ω̄ℓk δij + c4 τ 3 s̄ik s̄kℓ Ω̄ℓj − Ω̄iℓ s̄ℓk s̄kj
2
3
  (14.3)
3 2
+ c5 τ Ω̄iℓ Ω̄ℓm s̄mj + s̄iℓ Ω̄ℓm Ω̄mj − Ω̄mn Ω̄nℓ s̄ℓm δij
3
+ c6 τ 3 s̄kℓ s̄kℓ s̄ij + c7 τ 3 Ω̄kℓ Ω̄kℓ s̄ij
 
1 ∂v̄i ∂v̄j
Ω̄ij = −
2 ∂xj ∂xi
where τ is a turbulent time scale; for a non-linear k − ε model τ = k/ε, and for a non-
linear k − ω model τ = 1/ω. The tensor groups correspond to a subset of Eq. 11.107:

Line 1: Tij1 ,

Line 2: Tij3 and Tij2

Line 3: Tij4 and Tij5

Line 4: Tij6

Line 5: Tij1 multiplied by the invariants s̄kℓ s̄kℓ and Ω̄kℓ Ω̄kℓ

The expression in Eq. 14.3 is cubic in ∂v̄i /∂xj . However, note that it is only
quadratic in s̄ij and Ω̄ij . This is due to Cayley-Hamilton theorem which states that a
tensor is only linearly independent up to quadratic terms, see p. 130; this means that,
for example, s̄3ij = s̄ik s̄kℓ s̄ℓj can be expressed as a linear combination of s̄2ij = s̄ik s̄kj
and s̄ij .
aij is symmetric and its trace is zero; it is easily verified that the right side of
Eq. 14.3 also has these properties. Examples of non-linear models (sometimes also
14. Non-linear Eddy-viscosity Models 157

called explicit algebraic Reynolds stress models, EARSM) in the literature are the mod-
els presented in [35,48–50]. EARSMs are very popular — especially the model in [35]
— in the aeronautical community where explicit time-marching solvers are used. They
are computationally cheap, more accurate than linear eddy-viscosity models and they
do not give rise to any numerical instabilities as in implicit solvers (like SIMPLE).
In implicit solvers a large turbulent viscosity in the diffusion term of the momentum
equations is needed to stabilize the solution procedure.
Let’s take a closer look on Eq. 14.3 in fully developed channel flow (v̄2 = v̄3 =
∂/∂x1 = ∂/∂x3 ≡ 0); we obtain
 2
1 2 ∂v̄1
a11 = τ (c1 + 6c2 + c3 )
12 ∂x2
 2
1 2 ∂v̄1
a22 = τ (c1 − 6c2 + c3 )
12 ∂x2
 2 (14.4)
1 2 ∂v̄1
a33 =− τ (c1 + c3 )
6 ∂x2
 3
∂v̄1 1 ∂v̄1
a12 = −cµ τ + τ3 (−c5 + c6 + c7 )
∂x2 4 ∂x2

Using values on the constants as in [48], i.e c1 = −0.05, c2 = 0.11, c3 = 0.21,

c4 = −0.8 c5 = 0, c6 = −0.5 and c7 = 0.5 we get
 2  2
0.82 2 ∂v̄1 2 0.82 2 ∂v̄1
a11 = τ ⇒ v1′2 = k + kτ
12 ∂x2 3 12 ∂x2
 2  2
−0.5 2 ∂ ū1 ′2 2 0.5 2 ∂v̄1
a22 = τ ⇒ v2 = k − kτ
12 ∂x2 3 12 ∂x2
 2  2 (14.5)
−0.16 2 ∂v̄1 2 0.16 2 ∂v̄1
a33 = τ ⇒ v3′2 = k − kτ
12 ∂x2 3 12 ∂x2
k ∂v̄1
a12 = −cµ
ε ∂x2
We find that indeed the non-linear model gives anisotropic normal Reynolds stresses.
In Eqs. 14.4 and 14.5 we have assumed that the only strain is ∂v̄1 /∂x2 . When
we discussed streamline curvature effects at p. 151 we found that it is important to
investigate the effect of secondary strains such as ∂v̄2 /∂x1 . Let’s write down Eq. 14.3
for the strain ∂v̄2 /∂x1
 2
1 2 ∂v̄2
a11 = τ (c1 − 6c2 + c3 )
12 ∂x1

2
1 2 ∂v̄2
a22 = τ (c1 + 6c2 + c3 )
12 ∂x1
 2 (14.6)
1 ∂v̄2
a33 = − τ2 (c1 + c3 )
6 ∂x1
 3
1 3 ∂v̄2
a12 =− τ (c5 + c6 + c7 )
4 ∂x1
14. Non-linear Eddy-viscosity Models 158

Inserting the constants from [48] (see above) we obtain

 2
0.5 2 ∂v̄2
a11 =− τ
12 ∂x1
 2
0.82 2 ∂v̄2
a22 = τ (14.7)
12 ∂x1
 2
0.16 2 ∂v̄2
a33 =− τ , a12 = 0
12 ∂x1

As can be seen the coefficient for a22 is larger than that in Eq. 14.5, and hence the model
is slightly more sensitive to the secondary strain, ∂v̄2 /∂x1 , than to the primary one,
∂v̄1 /∂x2 . Thus, the non-linear models are able to account for streamline curvature, but
due to the choice of constants c5 + c6 + c7 = 0 this effect is weak.
15. The V2F Model 159

15 The V2F Model

In the V2F model of [44, 51, 52] two additional equations, apart from the k and ε-
equations, are solved: the wall-normal stress v2′2 and a function f . This is a model
which aims at improving modeling of wall effects on the turbulence.
Walls affect the fluctuations in the wall-normal direction, v2′2 , in two ways. The wall
damping of v2′2 is felt by the turbulence fairly far from the wall (x+ 2 . 200) through the
pressure field (i.e. the pressure-strain term) whereas the viscous damping takes place
within the viscous and buffer layer (x+ 2 . 10). In usual eddy-viscosity models both
these effects are accounted for through damping functions. The damping of v2′2 is in
the RSM accounted for through the modeled pressure-strain terms Φ22,1w and Φ22,2w
(see Eqs. 11.95 and Eq. 11.96). They go to zero far away from the wall (x+ 2 & 100).
In the V2F model the problem of accounting for the wall damping of v2′2 is simply
resolved by solving its transport equation. The v2′2 equation in boundary-layer form
reads (see Eq. 9.16 at p. 105)
" #
∂ρv̄1 v2′2 ∂ρv̄2 v2′2 ∂ ∂v2′2 ∂p′
+ = (µ + µt ) − 2v2′ − ρε22 (15.1)
∂x1 ∂x2 ∂x2 ∂x2 ∂x2

in which the diffusion term has been modeled with an eddy-viscosity assumption, see
Eq. 11.47 at p. 126. Note that the production term P22 = 0 because in boundary-layer
approximation v̄2 ≪ v̄1 and ∂/∂x1 ≪ ∂/∂x2 . The model for the dissipation term,
ε22 , is taken as
v ′2
εmodel
22 = 2 ε
k
This is a more elaborate model than in RSM (see Eq. 11.49). Add and subtract εmodel
22
on the right side of Eq. 15.1 yields

∂ρv̄1 v2′2 ∂ρv̄2 v2′2

+ =
∂x1 ∂x2
" # (15.2)
∂ ∂v ′2 ∂p′ v ′2 v ′2
(µ + µt ) 2 − 2v2′ − ρε22 + ρ 2 ε − ρ 2 ε
∂x2 ∂x2 ∂x2 k k

In the V2F model P is now defined as

2 ∂p′ v ′2
P = − v2′ − ε22 + 2 ε (15.3)
ρ ∂x2 k
so that Eq. 15.2 can be written as
" #
∂ρv̄1 v2′2 ∂ρv̄2 v2′2 ∂ ∂v2′2 v ′2
+ = (µ + µt ) + ρP − ρ 2 ε (15.4)
∂x1 ∂x2 ∂x2 ∂x2 k

P is the source term in the v2′2 -equation above, and it includes the velocity-pressure
gradient term and the difference between the exact and the modeled dissipation. Note
that this term is commonly split into a diffusion term and the pressure-strain term as

∂p′ ∂v ′ p′ ∂v ′
v2′ = 2 − p′ 2
∂x2 ∂x2 ∂x2
15. The V2F Model 160

Physically, the main agent for generating wall-normal stress is indeed the pressure-
strain term via re-distribution, see example in Section 11.13.
A new variable f = P/k is defined and a relaxation equation is formulated for f
as
!
2 ′2
∂ f Φ 22 1 v 2
L2 2 − f = − − 2
−
∂x2 ρk T k 3
  ν 1/2 
k
T = max , CT
ε ε
!  2 (15.5)
Φ22 C1 2 v2′2 νt ∂v̄1
= − + C2
ρk T 3 k k ∂x2
(  1/4 )
k 3/2 ν3
L = CL max , Cη
ε ε
where Φ22 is the IP model of the pressure-strain term, see Eqs. 11.57 and 11.90, the
first term being the slow term, and the second the rapid term. The constants are given
the following values: cµ = 0.23, CT = 6, cε1 = 1.44, cε2 = 1.9, σk = 0.9, σε =
1.3, C1 = 1.3, C2 = 0.3, CL = 0.2, Cη = 90.
The boundary condition for f is obtained from the v2′2 equation. Near the wall, the
′2
v2 equation reads
∂ 2 v2′2 v ′2
0=ν 2 + fk − 2 ε (15.6)
∂x2 k
The first and the last term behave as O(x22 ) as x2 → 0 because Taylor analysis gives
v2′2 = O(x42 ), ε = O(x02 ) and k = O(x22 ), see Section 4 Introduction to turbulence
Furthermore, ε = 2νk/x22 [53]; using this expression to replace k in Eq. 15.6 gives
∂ 2 v2′2 f εx22 2v2′2
0= 2 + − (15.7)
∂x2 2ν 2 x22
Assuming that f and ε are constant very close to the wall, this equation turns into an
ordinary second-order differential equation with the solution
B x4
v2′2 = Ax22 + − εf 2 2
x2 20ν
Since v2′2 = O(x42 ) as x2 → 0, both constants must be zero, i.e. A = B = 0, so we get
20ν 2 v2′2
f =− (15.8)
ε x42
For more details, see [54].
Above we have derived the v2′2 equation in boundary layer form assuming that x2
is the wall-normal coordinate. In general, three-dimensional flow it reads
 
∂ρv̄j v 2 ∂ ∂v 2 v2
= (µ + µt ) + ρf k − ρ ε
∂xj ∂xj ∂xj k
2
 2 
∂ f Φ22 1 v 2
L2 −f =− − − (15.9)
∂xj ∂xj ρk T k 3
 
Φ22 C1 2 v 2 Pk
= − + C2
ρk T 3 k k
15. The V2F Model 161

1.5

L=0.04,S=1
L=0.08,S=1
f 1
L=0.12,S=1
L=0.2,S=1
0.5
L=0.2, S=2
L=0.2, S=0.5

0
0 0.2 0.4 0.6 0.8 1
x2

Figure 15.1: Illustration of Eq. 15.12

In the V2F model a transport equation for the normal stress normal to walls is solved
for. If the wall lies in the x1 − x3 plane, then v 2 = v2′2 . However, if a wall lies in
the x2 − x3 plane, for example, this means that the transport equation for v2′2 is turned
into an equation for v1′2 , i.e. v 2 = v1′2 . This is done automatically since in the general
formulation in Eq. 15.9, ∂v̄1 /∂x2 in the expression for Φ22 is replaced by P k . If the
wall lies in the x2 − x3 plane the largest velocity gradient will be ∂v̄2 /∂x1 or ∂v̄3 /∂x1 .
Why does the right side of Eq. 15.5 has the form it has? Far from the wall,
the source term in the v2′2 -equation simplifies to Φ22 plus isotropic dissipation (see
Eq. 15.1). This is what happens, because far from the wall when ∂ 2 f /∂x22 ≃ 0, so that
Eq. 15.5 can be written (T = k/ε)

kf ≡ P → Φ22 + ε(v2′2 /k − 2/3) (15.10)

When this expression is inserted in Eq. 15.4 we get

" #
∂ρv̄1 v2′2 ∂ρv̄2 v2′2 ∂ ∂v2′2 2
+ = (µ + µt ) + ρΦ22 − ρε (15.11)
∂x1 ∂x2 ∂x2 ∂x2 3

which is the usual form of the modeled v2′2 -equation with isotropic dissipation. Near
the wall, the diffusion term ∂ 2 f /∂x22 makes f go from the value of its source term to
its (negative) wall value (see Eq. 15.8) over the lengthscale L. This is how the influence
of the source term P in Eq. 15.4 is reduced as the wall is approached. The behavior of
the equation for f (Eq. 15.5) is illustrated in Fig. 15.1 where the equation

∂2f
L2 −f +S =0 (15.12)
∂x22

has been solved with f = 0 at the wall and with different L and S.
As can be seen, f is, as required, reduced as the wall is approached. Furthermore,
f approaches the value of the source term as x2 > L. The influence of the lengthscale
L is nicely illustrated: the larger L, the further away from the wall does f go to its
far-field value.
15.1. Modified V2F model 162

In the V2F model the turbulent viscosity is computed from

νt = Cµ v2′2 T (15.13)
The k and ε-equations are also solved (without damping functions). For conve-
nience, the boundary conditions are given again

k = 0, v2′2 = 0
ε = 2νk/x22
(15.14)
20ν 2 v2′2
f =−
εx42
The boundary condition for f makes the equation system numerically unstable.
One way to get around that problem is to solve both the k, ε and v2′2 , f equations
coupled [54]. An alternative is to use the ζ − f model [55] which is more stable. In
this model they solve for the ratio v2′2 /k instead of for v2′2 which gives a simpler wall
boundary condition for f , namely f = 0.

15.1 Modified V2F model

In [56] they proposed a modification of the V2F model allowing the simple explicit
boundary condition f = 0 at walls. They introduced a new variable
f ∗ = f − 5εv 2 /k 2
and they neglected the term
 
∂2 εv 2
−5L2
∂xj ∂xj k2

The resulting v2′2 and f ∗ -equation read [56]

 
∂v̄j v 2 ∂ ∂v 2 v2
= (ν + νt ) + kf ∗ − 6 ε (15.15)
∂xj ∂xj ∂xj k
 
∂2f ∗ 1 v2 2 Pk
−L2 + f∗ = − (C1 − 6) − (C1 − 1) + C2
∂xj ∂xj T k 3 k
 
∂v̄i ∂v̄ j ∂v̄ i
P k = νt +
∂xj ∂xi ∂xj
 
k  ν 1/2 (15.16)
T = max ,6
ε ε
(  3 1/4 )
3/2
k ν
L = CL max , Cη
ε ε
Boundary conditions at the walls are
k = 0, v 2 = 0
ε = 2νk/x22
f∗ = 0
This modified model is numerically much more stable. Note that the modified model
is identical to the original model far from the wall.
15.2. Realizable V2F model 163

15.2 Realizable V2F model

The realizable condition for stagnation flow (see p. 153) is used also for the V2F model,
and they read [56]
" #
k 0.6k
T = min ,√ 1/2
ε 6Cµ v 2 (s̄ij s̄ij )
" # (15.17)
k 3/2 k 3/2
L = min ,√
ε 6Cµ v 2 (2s̄ij s̄ij )1/2

These realizable conditions have been further investigated by Sveningsson [54,57–60],

and it was found that the limitation on T is indeed important, whereas that for L is
not. Furthermore, it was found that it is important to impose the limitation on T in
a consistent manner. For instance, if the limit is used in the f equation, it must for
consistency also be used for ε/k in Eq. 15.15.

15.3 To ensure that v 2 ≤ 2k/3

In the V2F model, v 2 denotes the generic wall-normal stress. Thus it should be the
smallest one. This is not ensured in the V2F models presented above. Below the
simple modification proposed by [61] is presented.
The source term kf in the v 2 -equation (Eq. 15.15) includes the modeled velocity-
pressure gradient term which is dampened near walls as f goes to zero. Since v 2
represents the wall-normal normal stress, it should be the smallest normal stress, i.e.
v2′2 ≤ v1′2 and v2′2 ≤ v3′2 , and thus v2′2 should be smaller than or equal to 32 k. In
the homogeneous region far away from the wall, the Laplace term is assumed to be
negligible i.e. ∂ 2 f /∂xj ∂xj → 0. Then Eq. 15.16 reduces to f = right side.
It turns out that in the region far away from the wall, the Laplace term is not negli-
gible, and as a consequence v 2 gets too large so that v 2 > 23 k. A simple modification
is to use the right side of Eq. 15.16 as an upper bound on the source term kf in the
v 2 -equation, i.e.
   
2 ε 2 2k k
vsource = min kf, − (C1 − 6)v − (C1 − 1) + C2 P (15.18)
k 3

This modification ensures that v 2 ≤ 2k/3. For more details, see [61].
16. The SST Model 164

Figure 16.1: Flow around an airfoil. Pressure contours. Red: high pressure; blue: low
pressure

16 The SST Model

The SST (Shear Stress Transport) model of [62] is an eddy-viscosity model which
includes two main novelties:
1. It is combination of a k − ω model (in the inner boundary layer) and k − ε model
(in the outer region of the boundary layer as well as outside of it);
2. A limitation of the shear stress in adverse pressure gradient regions.
The k − ε model has two main weaknesses: it over-predicts the shear stress in
adverse pressure gradient flows because of too large length scale (due to too low dis-
sipation) and it requires near-wall modification (i.e. low-Re number damping func-
tions/terms). Various k−ω models are presented in Section 4 Introduction to turbulence
One example of adverse pressure gradient is the flow along the surface of an airfoil,
see Fig. 16.1. Consider the upper surface (suction side). Starting from the leading edge,
the pressure decreases because the velocity increases. At the crest (at x/c ≃ 0.15)
the pressure reaches its minimum and increases further downstream as the velocity
decreases. This region is called the adverse pressure gradient (APG) region.
The k − ω model is better than the k − ε model at predicting adverse pressure
gradient flow and the standard k − ω model of [63] (see also [53]) does not use any
damping functions. However, the disadvantage of the standard k − ω model is that it is
dependent on the free-stream value of ω [64, 65].
In order to improve both the k − ε and the k − ω model, it was suggested in [62]
to combine the two models. Before doing this, it is convenient to transform the k − ε
model into a k − ω model using the relation ω = ε/(β ∗ k), where β ∗ = cµ . The left-
hand side of the ω equation will consist of the convection term, dω/dt, which denotes
the material derivative, assuming steady flow, see Eq. 2.25. Let us express the left-
hand side of the ω equation as a combination of the left-hand sides of the ε and the k
16. The SST Model 165

equations by using the chain rule, i.e.

 
dω d ε 1 dε ε d(1/k)
= ∗
= ∗ + ∗
dt dt β k β k dt β dt
(16.1)
1 dε ε dk 1 dε ω dk
= ∗ − ∗ 2 = ∗ −
β k dt β k dt β k dt k dt
Now we have transformed the left side of the ω equation. The right side should be
transformed in the same manner. For example, the production of the ω equation will
consist of two terms, one term from the ε equation
1
Pε (the first term at the right side in Eq. 16.1) (16.2)
β∗k
and one from the k equation
ω k
− P (the second term at the right side in Eq. 16.1) (16.3)
k
In the same way we transform the entire right side inserting the modeled equations for
k and ε so that
   
Dω 1 ω 1 ω
= ∗ Pε − P k − ∗ Ψε − Ψk +
Dt β k k β k k
Production, Pω Destruction, Ψω
  " # (16.4)
1 ω ν ∂2ε νω ∂ 2 k
DεT − DkT + ∗ 2 −
β∗k k β k ∂xj k ∂x2j
Turbulent diffusion, DωT Viscous diffusion, Dων

• Production term
1 ω ε ω
Pω = Pε − P k = Cε1 ∗ 2 P k − P k
β∗k k β k k (16.5)
ω k
= (Cε1 − 1) P
k
• Destruction term
1 ω ε2 ω
Ψω = Ψε −
Ψk = Cε2 − ε
β∗k k k k (16.6)
= (Cε2 − 1)β ∗ ω 2

• Viscous diffusion term

ν ∂2ε νω ∂ 2 k ν ∂ 2 ωk νω ∂ 2 k
Dων = 2 − 2 = −
∗
β k ∂xj k ∂xj k ∂x2j k ∂x2j
  
ν ∂ ∂k ∂ω ω ∂2k
= ω +k −ν
k ∂xj ∂xj ∂xj k ∂x2j
" # (16.7)
ν ∂ω ∂k ∂2k ∂k ∂ω ∂2ω ω ∂2k
= +ω 2 + +k 2 −ν
k ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj k ∂x2j
 
2ν ∂ω ∂k ∂ ∂ω
= + ν
k ∂xj ∂xj ∂xj ∂xj
16. The SST Model 166

The turbulent diffusion term is obtained as (the derivation is found in [66] which
can be downloaded from http://www.tfd.chalmers.se/˜lada)
 
2νt ∂k ∂ω ω νt νt ∂ 2 k
DωT = + − +
σε k ∂xj ∂xj k σε σk ∂x2j
    (16.8)
ω 1 1 ∂νt ∂k ∂ νt ∂ω
+ − +
k σε σk ∂xj ∂xj ∂xj σε ∂xj

In the standard k − ε model we have σk = 1 and σε = 1.3. If we assume that σk = σε

in the second and third term of the right-hand side, we can considerably simplify the
turbulence diffusion so that
 
T 2νt ∂k ∂ω ∂ νt ∂ω
Dω = + (16.9)
σε k ∂xj ∂xj ∂xj σε ∂xj

We can now finally write the ε equation formulated as an equation for ω

  
∂ ∂ νt ∂ω ω
(v̄j ω) = ν+ + α P k − βω 2
∂xj ∂xj σε ∂xj k
 
2 νt ∂k ∂ω (16.10)
+ ν+
k σε ∂xi ∂xi
α = Cε1 − 1 = 0.44, β = (Cε2 − 1)β ∗ = 0.0828

Since the k − ε model will be used for the outer part of the boundary layer, the vis-
cous part of the cross-diffusion term (second line) is usually neglected (the viscous
terms are negligible in the outer region). The turbulent viscosity is computed as (using
dimensional analysis)
k
νt = (16.11)
ω
In the SST model the coefficients are smoothly switched from k − ω values in the
inner region of the boundary layer to k − ε values in the outer region. Functions of the
form
" ( √ ) #
4 k 500ν 4σω2 k
F1 = tanh(ξ ), ξ = min max , , (16.12)
β ∗ ωy y 2 ω CDω y 2

are used. F1 = 1 in the near-wall region and F1 = 0 in the outer region. The β-
coefficient, for example, is computed as

βSST = F1 βk−ω + (1 − F1 )βk−ε (16.13)

where βk−ω = 0.075 and βk−ε = 0.0828. Since the standard k − ω model does
not include any cross-diffusion term, the last term in the ω equation (second line in
Eq. 16.10) should only be active in the k − ε region; hence it is multiplied by (1 − F1 ).
At p. 164 it was mentioned that the k − ω model is better than the k − ε model
in predicting adverse pressure-gradient flows because it predicts a smaller shear stress.
Still, the predicted shear stress is too large. This brings us to the second modification
(see p. 164). When introducing this second modification, the author in [62] noted that
a model (the Johnson - King model [JK]) which is based on a transport equation of
the main shear stress v1′ v2′ , predicts adverse pressure gradient flows much better than
16. The SST Model 167

the k − ω model. In the JK model, the v1′ v2′ transport equation is built on Bradshaw’s
assumption [67]
− v1′ v2′ = a1 k (16.14)
1/2
where a1 = cµ = β ∗1/2 . In boundary layer flow, the Boussinesq assumption can be
written as
"  2 #1/2  k 1/2
k ∂v̄1 cµ k 2 ∂v̄1 1/2 cµ k 2 ∂v̄1 1/2 P
−v1′ v2′ = = = cµ k 2
= cµ k
ω ∂x2 ε ∂x2 ε ∂x2 ε
νt ,k−ω νt ,k−ε
(16.15)

It is found from experiments that in boundary layers of adverse pressure gradient flows,
1/2
the production is much larger than the dissipation (P k ≫ ε) and that −v1′ v2′ ≃ cµ k.
When the Boussinesq assumption is used in the k − ω model for adverse pressure
gradient flows, P k ≫ ε, and hence (see Eq. 16.15)

v1′ v2′
− 1/2
≫1 (16.16)
cµ k
which explains why the Boussinesq assumption over-predicts the shear stress and works
1/2
poorly in this type of flow (recall that according to experiments −v1′ v2′ ≃ cµ k). To
′ ′
reduce |v1 v2 | in Eq. 16.15 in adverse pressure gradient flow, [62] proposed to re-define
the turbulent eddy viscosity including the expression in Eq. 16.14. We have two ex-
pressions for the turbulent viscosity (expressing a Johnson-King-turbulent viscosity
with the Boussinesq assumption)
1/2
−v1′ v2′ cµ k
νt = = Johnson-King (16.17a)
Ω̄ Ω̄
1/2
k cµ k
νt = = 1/2 k − ω model (16.17b)
ω cµ ω

where Ω̄ is the absolute vorticity (in boundary layer flow Ω̄ = ∂v̄1 /∂x2 ); in (a) the
Boussinesq assumption together with Eq. 16.14 were used and (b) is taken from the
k − ω model. We want (a) to apply only in the boundary layer and hence we multiply
it with a function F2 (similar to F1 ) which is 1 near walls and zero elsewhere. Then we
take the minimum of (a) and (b) so that
1/2
cµ k
νt = 1/2
(16.18)
max(cµ ω, F2 Ω̄)

When the production is large (i.e. when Ω̄ is large), Eq. 16.18 reduces νt according to
the Johnson - King model, i.e. Eq. 16.17a. It is important to ensure that this limitation
is not active in usual boundary layer flows where P k ≃ ε. It can be seen that νt is
1/2
reduced only in regions where P k > ε, because if P k < ε then Ω̄ < cµ ω since
1 ω ω
Ω̄2 = νt Ω̄2 = P k < ε = cµ ω 2 (16.19)
νt k k
Hence, in regions where P k < ε, Eq. 16.18 returns to νt = k/ω as it should.
To summarize the SST modification:
16. The SST Model 168

1/2
• the second part, cµ k/Ω in Eq. 16.18 (which mimics the Johnson-King model),
is active in APG flow
• the first part, k/ω in Eq. 16.18 (which corresponds to the usual Boussinesq
model), should be active in the remaining part of the flow. Equation 16.19 shows
that (it is likely that) the second part is active only in APG regions and not else-
where.
Today, the SST model has been slightly further developed. Two modifications have
been introduced [68]. The first modification is that the absolute vorticity Ω̄ in Eq. 16.18
has been replaced by |s̄| = (2s̄ij s̄ij )1/2 which comes from the production term using
the Boussinesq assumption (see Eq. 11.39), i.e.
 
∂v̄i ∂v̄j ∂v̄i
|s̄|2 = + = 2s̄ij (s̄ij + Ω̄ij ) = 2s̄ij s̄ij
∂xj ∂xi ∂xj
  (16.20)
1 ∂v̄i ∂v̄j
Ω̄ij = −
2 ∂xj ∂xi

where s̄ij Ω̄ij = 0 because s̄ij is symmetric and Ω̄ij is anti-symmetric. Equation 16.18
with |s̄| limits νt in stagnation regions similar to Eq. 13.12. The second modification
in the SST model is that the production term in the new SST model is limited by 10ε,
i.e.

Pk,new = min P k , 10ε (16.21)
The final form of the SST model is given in Eq. 20.4 at p. 200.
17. Overview of RANS models 169

17 Overview of RANS models

This section presents a short overview of the presented RANS models. First the models
can be classified as models based on eddy viscosity (i.e. turbulent viscosity) or models
in which equations are solved (algebraic or differential) to obtain the Reynolds stress
tensor, vi′ vj′ . Eddy-viscosity models, which are based on the Boussinesq assumption, Eddy-
see Eq. 11.33, are viscosity
models
• standard k − ε (see Section 11.8) and k − ω models
• combination of k − ε and k − ω models such as the SST model, see Section 16.

There are more elaborate eddy-viscosity models such as

non-linear
• non-linear models, see Section 14 models

Models not based on the eddy-viscosity are the Reynolds stress models such as
• the Reynolds stress transport model (RSM or RSTM), in which transport equa- RSM
tions are solved for vi′ vj′ , see Eq. 11.101.
• the Algebraic Reynolds stress model, ASM, in which algebraic equations are ASM
solved for vi′ vj′ , see Eq. 11.104.
explicit ASM
• explicit Algebraic Reynolds stress models, in which explicit algebraic equations
are solved for vi′ vj′ , see Section 11.11.
Finally, there is a class of models which are somewhere in between two-equation eddy-
viscosity models and Reynolds stress models, and that is
V2F
• V2F models, see Section 15
The V2F model is an eddy-viscosity model and the model is based on four transport
equations.
18. Large Eddy Simulations 170

0.5

0.45

0.4

0.35

v1 , v¯1
0.3

0.25

0.2
no filter
0.15 one filter
0.1
2 2.5 3 3.5 4
x1

Figure 18.1: Filtering the velocity.

18 Large Eddy Simulations

18.1 Time averaging and filtering
N CFD we time average our equations to get the equations in steady form. This is
I called Reynolds time averaging:
Z T
1
hΦi = Φ(t)dt, Φ = hΦi + Φ′ (18.1)
2T −T
(note that we use the notation h.i for time averaging). In LES we filter (volume average)
the equations. In 1D we get (see Fig. 18.1)

Z x+0.5∆x
1
Φ̄(x, t) = Φ(ξ, t)dξ
∆x x−0.5∆x
′′
Φ = Φ̄ + Φ
Since in LES we do not average in time, the filtered variables are functions of space
and time. The equations for the filtered variables have the same form as Navier-Stokes,
i.e.
∂v̄i ∂ 1 ∂ p̄ ∂ 2 v̄i ∂τij
+ (v̄i v̄j ) = − +ν −
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj
(18.2)
∂ v¯i
=0
∂xi
where the subgrid stresses are given by
τij = vi vj − v̄i v̄j (18.3)
It should be noted that it is formally incorrect to denote τij as a stress since the density
is omitted (see, e.g., the text below Eq. 6.10 and Eq. 11.3). Contrary to Reynolds time
averaging where hvi′ i = 0, we have here
vi′′ 6= 0

v̄ i 6= v̄i
18.2. Differences between time-averaging (RANS) and space filtering (LES) 171

This is true for box filters. Note that for the spectral cut-off filter v̄i = v̄ i , see
p. 173. However, in finite volume methods, box filters are always used. In this course
we use box filters, if not otherwise stated.
Let’s look at the filtering of Eq. 18.2 in more detail. The pressure gradient term, for
example, reads
Z
∂p 1 ∂p
= dV
∂xi V V ∂xi
Now we want to move the derivative out of the integral. When is that allowed? The
answer is “if the integration region is not a function of xi ”, i.e. if V is constant. In finite
volume methods, the filtering volume, V , is (almost always) identical to the control
volume. In general, the size of the control volume varies in space. Fortunately, it can
be shown that if V is a function of xi , the error we do when moving the derivative out
of the integral is proportional to (∆x)2 [69], i.e. it is an error of second order. Since
this is the order of accuracy of our finite volume method anyway, we can accept this
error. Now let’s move the derivative out of the integral, i.e.
 Z 
∂p ∂ 1
∂ p̄

= pdV + O (∆x)2 = + O (∆x)2
∂xi ∂xi V V ∂xi

All linear terms are treated in the same way.

Now we take a look at the non-linear term in Eq. 18.2, i.e. the convective term.
First we filter the term and move the derivative out of the integral, i.e.
 Z 
∂vi vj ∂ 1
∂

= vi vj dV + O (∆x)2 = (vi vj ) + O (∆x)2
∂xj ∂xj V V ∂xj

There is still a problem with the formulation of this term: it includes an integral of a
product, i.e. vi vj ; we want it to appear like a product of integrals, i.e. v̄i v̄j . To achieve
this we simple add the term we want (v̄i v̄j ) and subtract the one we don’t want ( vi vj )
on both the right and left side. This is how we end up with the convective term and the
SGS term in Eq. 18.2.
The filtering in Eq. 18.2 is usually achieved through the finite-volume discretiza-
tion. This means that no additional filtering is done. This is called implicit filtering.
Hence, the discretized momentum equations in RANS and LES (Eqs. 6.6 and 18.2)
are identical. The only difference is the magnitude of the modeled Reynolds stresses,
which are much larger in RANS than in LES.

18.2 Differences between time-averaging (RANS) and space filter-

ing (LES)
In RANS, if a variable is time averaged twice (hhvii), it is the same as time averaging
once (hvi). This is because hvi is not dependent on time. From Eq. 18.1 we get
Z T
1 1
hhvii = hvidt = hvi2T = hvi
2T −T 2T

This is obvious if the flow is steady, i.e. ∂hvi/∂t = 0. If the flow is unsteady, we must
assume a separation in time scales so that the variation of hvi during the time interval
T is negligible, i.e. ∂/∂t ≪ 1/T . In practice this requirement is rarely satisfied.
In LES, v̄ 6= v̄ (and since v = v̄ + v ′′ we get v ′′ 6= 0).
18.3. Resolved & SGS scales 172

∆x

x x

I −1 A I B I +1

Figure 18.2: Box filter illustrated for a control volume.

Let’s filter v̄I once more (filter size ∆x, see Fig. 18.2. For simplicity we do it in
1D. (Note that subscript I denotes node number.)
Z ∆x/2 Z 0 Z ∆x/2 !
1 1
v̄ I = v̄(ξ)dξ = v̄(ξ)dξ + v̄(ξ)dξ =
∆x −∆x/2 ∆x −∆x/2 0
 
1 ∆x ∆x
= v̄A + v̄B .
∆x 2 2

The trapezoidal rule, which is second-order accurate, was used to estimate the integrals.
v̄ at locations A and B (see Fig. 18.2) is estimated by linear interpolation, which gives
   
1 1 3 3 1
v̄ I = v̄I−1 + v̄I + v̄I + v̄I+1
2 4 4 4 4
(18.4)
1
= (v̄I−1 + 6v̄I + v̄I+1 ) 6= v̄I
8

18.3 Resolved & SGS scales

The basic idea in LES is to resolve (large) grid scales (GS), and to model (small)
subgrid-scales (SGS). Even if LES is always made unsteady, we are usually inter-
ested in the time-averaged results. This means that we time-average all quantities for
post-processing, e.g. hv̄i i where the brackets denote time-averaging. From the LES
equations, Eq. 18.2, we get the resolved (GS) velocities, v̄i . Then we can compute the
resolved fluctuations as v̄i′ = v̄i − hv̄i i.
The limit (cut-off) between GS and SGS is supposed to take place in the inertial
subrange (II), see Fig. 18.3.
I: large, energy-containing scales
II: inertial subrange (Kolmogorov −5/3-range)
III: dissipation subrange
18.4. The box-filter and the cut-off filter 173

I
E
II

εsgs

III

resolved cut-off SGS

scales κ
κc

Figure 18.3: Spectrum of velocity.

18.4 The box-filter and the cut-off filter

The box filtering is formally defined as (1D)
Z ∞
v̄(x) = GB (r)v(x − r)dr
−∞

1/∆, if r ≤ ∆/2
GB (r) = (18.5)
0, if r > ∆/2
Z ∞
GB (r)dr = 1
−∞

However, it is often convenient to study the filtering process in the spectral space.
The filter in spectral space is particular simple: we simply set the contribution from
wavenumbers larger than cut-off to zero. Hence the cut-off filter filters out all scales
with wavenumber larger than the cut-off wavenumber κc = π/∆. It is defined as

1 if κ ≤ κc
ĜC (κ) = (18.6)
0 otherwise
We now look at what happens when we filter twice in spectral space. It appears that
the calculations for this gets much easier if we first introduce the Fourier transform.
The Fourier transform is defined as (cf. Eq. D.28)
Z ∞
v̂(κ) = v(r) exp(−2πıκr)dr (18.7)
0

and its inverse Z ∞

v(r) = v̂(κ) exp(2πıκr)dκ (18.8)
0
√
where κ denotes the wavenumber and ı = −1. A slightly different definition has
been choosen here compared to Eq. D.28. It is chosen since no factor 2π appears in
18.5. Highest resolved wavenumbers 174

front of the integral in Eq. 18.8 (the Fourier par is symmetric); as a consequence it is
also absent in Eq. 18.9.
Note that it is physically meaningful to use Fourier transforms only in a homoge-
neous coordinate direction; in non-homogeneous directions the Fourier coefficients –
which are not a function of space – have no meaning. Using the convolution theorem
(saying that the integrated product of two functions is equal to the product of their
Fourier transforms), we can filter v̂ using Eqs. 18.7 and 18.5
Z ∞
v̂(κ) = b̄
v(κ) = v̄(η) exp(−2πıκη)dη
0
Z ∞Z ∞
= exp(−2πıκη)G(α)v(η − α)dαdη
0 0
Z ∞Z ∞ (18.9)
= exp(−2πıκα) exp(−2πıκ(η − α))G(α)v(η − α)dαdη
0 0
Z ∞ Z ∞
= exp(−2πıκα) exp(−2πıκξ)G(α)v(ξ)dξdα = Ĝ(κ)v̂(κ)
0 0

(in the last line we used ξ = η − α). If we use the cut-off filter and filter twice we get

v̂ = ĜC ĜC v̂ = ĜC v̂ (18.10)

since Ĝ2C = ĜC , see Eq 18.6. Since v̂¯ = v̂¯ for the Fourier transform v̂, we know that
¯ Thus, contrary to the box-filter (see Eq. 18.4),
– when using the cuf-off fiter – v̄¯ = v̂.
nothing happens when we filter twice in spectral space. The box filter is sharp in
physical space but not in wavenumber space; for the cut-off filter it is vice versa.
In finite volume methods box filtering is always used. Furthermore implicit filtering
is employed. This means that the filtering is the same as the discretization (=integration
over the control volume is equal to the filter volume, see Eq. 18.14).

18.5 Highest resolved wavenumbers

Any function can be expressed as a Fourier series such as Eq. 18.8 (see Section 5.3,
Eq. D.28 and Section E) provided that the coordinate direction is homogeneous. Let’s
choose the fluctuating velocity in the x1 direction, i.e. v1′ , and let it be a function of x1 .
We require it to be homogeneous, i.e. its RMS, v1,rms , does not vary with x1 . Now we
ask the question: on a given grid, what is the highest wavenumber that is resolved? Or,
in other words, what is the cut-off wavenumber?
The wave shown in Fig. 18.4a reads

v1′ = 0.25 [1 + 0.8 sin(κ1 x1 )] (18.11)

and it covers two cells (∆x1 /L = 0.5). If we define this as the cut-off wavenumber
we get κ1,c L = κ1,c 2∆x1 = 2π (i.e. sin(κ1,c 2∆x1 ) = sin(2π); recall that 2π is one
period) so that
κ1,c = 2π/(2∆x1 ) = π/∆x1 (18.12)
It is of course questionable if v1′ in Fig. 18.4a really is resolved since the sinus wave
covers only two cells. However this is the usual definition of the cut-off wavenumber.
If we require that the highest resolved wavenumber should be covered by four cells
(∆x1 /L = 0.25), as in Fig. 18.4b, then the cut-off wavenumber is given by κ1,c =
2π/(4∆x1 ) = π/(2∆x1 ).
 18.6. Subgrid model 175

0.4 0.4

v1′ 0.2 v1′ 0.2

0 0
node 1 2 3 1 2 3 4 5
−0.2 −0.2
0 0.5 1 1.5 0 0.5 1 1.5
x1 /L x1 /L
(a) One sinus period covering two cells (b) One sinus period covering four cells

Figure 18.4: Physical and wavenumber space. Sinus curves with different wavenum-
bers illustrated in physical space.

18.6 Subgrid model

We need a subgrid model to model the turbulent scales which cannot be resolved by
the grid and the discretization scheme.
The simplest model is the Smagorinsky model [70]:
 
1 ∂v̄i ∂v̄j
τij − δij τkk = −νsgs + = −2νsgs s̄ij
3 ∂xj ∂xi (18.13)
2 p 2
νsgs = (CS ∆) 2s̄ij s̄ij ≡ (CS ∆) |s̄|
and the filter-width is taken as the local grid size
1/3
∆ = (∆VIJK ) (18.14)
The scalar |s̄| is the norm (i.e. the “length”) of ∂v̄i /∂xj + ∂v̄j /∂xi in the Boussinesq
assumption, see Eq. 16.20. The Smagorinsky model is derived as follows. The turbu-
lent viscosity may – as in RANS – be obtained through dimensional analysis as (see
Eq. 11.25)
νsgs = UL. (18.15)
The turbulent velocity scale, U, is obtained as the first term in Taylor expansion,
L∂v̄/∂x, and the length scale , L, is taken as CS ∆ which gives
νsgs = (CS ∆)2 |s̄| (18.16)
where the one-dimensional velocity gradient, ∂v̄/∂x was replaced by the general three-
dimensional gradient, |s̄|. Near the wall, the SGS viscosity becomes quite large since
the velocity gradient is very large at the wall. However, because the SGS turbulent
fluctuations near a wall go to zero, so must the SGS viscosity. A damping function fµ
is added to ensure this
fµ = 1 − exp(−x+ 2 /26) (18.17)
A more convenient way to dampen the SGS viscosity near the wall is simply to use
the RANS length scale as an upper limit, i.e.
n o
1/3
∆ = min (∆VIJK ) , κn (18.18)

where n is the distance to the nearest wall.

Disadvantage of Smagorinsky model: the “constant” CS is not constant, but it is
flow-dependent. It is found to vary in the range from CS = 0.065 [71] to CS =
0.25 [72].
18.7. Smagorinsky model vs. mixing-length model 176

∂hv̄i i
−hv̄i′ v̄j′ i
∂xj

int
e
rti
E

al
su
br an
ge
ε sg

,E
s
=

∝
P sg

κ
k s

−
5/
3
ε

κ
κc

Figure 18.5: Energy spectrum.

18.7 Smagorinsky model vs. mixing-length model

The eddy viscosity according to the mixing length theory reads in boundary-layer
flow [73, 74]  
 
2  ∂v̄1 
νt = ℓ  .
∂x2 
Generalized to three dimensions, we have
  1/2
∂v̄i ∂v̄j ∂v̄i
νt = ℓ 2 + = ℓ2 (2s̄ij s̄ij )1/2 ≡ ℓ2 |s̄|.
∂xj ∂xi ∂xj

In the Smagorinsky model the SGS turbulent length scale corresponds to ℓ = CS ∆ so

that
νsgs = (CS ∆)2 |s̄|
which is the same as Eq. 18.13

18.8 Energy path

The path of kinetic energy is illustrated in Fig. 18.5. At cut-off, SGS kinetic energy is
dissipated
εsgs = −τij s̄ij = 2νsgs s̄ij s̄ij (18.19)
from the resolved turbulence. This energy is transferred to the SGS scales and act as
production term (Pksgs ) in the ksgs equation. The SGS kinetic energy is then trans-
ferred to higher wave-numbers via the cascade effect and the kinetic energy is finally
dissipated (ε=physical dissipation) in the dissipation range. It should be mentioned that
18.9. SGS kinetic energy 177

this process is an idealized one. We assume that ALL dissipation takes place in the dis-
sipation range. This is a good approximation, but in reality dissipation (i.e. transfer of
energy from kinetic energy to internal energy which corresponds to an increase in tem-
perature) takes place at all wave numbers, and the dissipation increases for increasing
wave number, see Eq. 8.19.

18.9 SGS kinetic energy

The SGS kinetic energy ksgs can be estimated from the Kolmogorov −5/3 law. The
total turbulent kinetic energy is obtained from the energy spectrum as
Z ∞
k= E(κ)dκ
0

Changing the lower integration limit to wavenumbers larger than cut-off (i.e. κc ) gives
the SGS kinetic energy Z ∞
ksgs = E(κ)dκ (18.20)
κc

The Kolmogorov −5/3 law now gives

Z ∞
ksgs = CK κ−5/3 ε2/3 dκ
κc

where CK = 1.5 (Note that for these high wavenumbers, the Kolmogorov spectrum
ought to be replaced by the Kolmogorov-Pau spectrum in which an exponential de-
caying function is added for high wavenumbers [73, Chapter 3]). Carrying out the
integration and replacing κc with π/∆ we get
 2/3
3 ∆ε
ksgs = CK (18.21)
2 π
In the same way as ksgs can be computed from Eq. 18.20, the resolved turbulent kinetic
energy, kres , is obtained from
Z κc
kres = E(κ)dκ
0

18.10 LES vs. RANS

LES can handle many flows which RANS (Reynolds Averaged Navier Stokes) cannot;
the reason is that in LES large, turbulent scales are resolved. Examples are:
o Flows with large separation
o Bluff-body flows (e.g. flow around a car); the wake often includes large, un-
steady, turbulent structures
o Transition
• In RANS all turbulent scales are modeled ⇒ inaccurate
• In LES only small, isotropic turbulent scales are modeled ⇒ accurate
• LES is very much more expensive than RANS.
18.11. The dynamic model 178

grid filter
test filter
E I

II
SGS grid scales

SGS test scales

III
resolved on test filter

resolved on grid filter

κ
z{ κc = π/∆
κ = π/ ∆

Figure 18.6: Energy spectrum with grid and test filter.

18.11 The dynamic model

In this model of [75] the constant C is not arbitrarily chosen (or optimized), but it is
computed.
If we apply two filters to Navier-Stokes [grid filter and a second, coarser filter (test
z{
filter, denoted by z{
. )] where ∆ = 2∆ we get
z{   z{ z{
∂ v̄ i ∂ z{ z{ 1 ∂ p̄ ∂ 2 v̄ i ∂Tij
+ v̄ i v̄ j = − +ν − (18.22)
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj
where the subgrid stresses on the test level now are given by
z { z{ z{
Tij = vi vj − v̄ i v̄ j (18.23)

z{   z{ z{ z{
∂ v̄ i ∂ z{ z{ 1 ∂ p̄ ∂ 2 v̄ i ∂ τ ij
+ v̄ i v̄ j = − +ν −
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj (18.24)
 
∂ z { z{ z{
− v̄i v̄j − v̄ i v̄ j
∂xj
Identification of Eqs. 18.22 and 18.24 gives
z { z{ z{ z{
v̄i v̄j − v̄ i v̄ j + τ ij = Tij (18.25)

The dynamic Leonard stresses are now defined as

z { z{ z{ z{
Lij ≡ v̄i v̄j − v̄ i v̄ j = Tij − τ ij (18.26)
18.12. The test filter 179

test filter grid filter

W P E
∆x ∆x

Figure 18.7: Control volume for grid and test filter.

The trace of this relation reads

z{
Lii ≡ Tii − τ ii
With this expression we can re-formulate Eq. 18.26 as
 
1 1 z{ 1 z{
Lij − δij Lkk = Tij − δij Tkk − τ ij − δij τ kk (18.27)
3 3 3
In the energy spectrum, the test filter is located at lower wave number than the grid
filter, see Fig. 18.6.

18.12 The test filter

z{
The test filter is twice the size of the grid filter, i.e. ∆ = 2∆.
The test-filtered variables are computed by integration over the test filter. For ex-
z{ z {
ample, the 1D example in Fig. 18.7 v̄ is computed as (∆x = 2∆x)
Z E Z P Z E !
z{ 1 1
v̄P = v̄dx = v̄dx + v̄dx
2∆x W 2∆x W P
 
1 1 v̄W + v̄P v̄P + v̄E (18.28)
= (v̄w ∆x + v̄e ∆x) = +
2∆x 2 2 2
1
= (v̄W + 2v̄P + v̄E )
4

18.12.1 2D filtering
In 2D, we do as in 1D: first compute the value at the center of the four squares (or
rectangles) surrounding the node (I, J, K), marked by x, see Fig. 18.8. For the lower-
left square, for example, we get
f 1
v̄I−1/2,J−1/2,K = (v̄I−1,J−1,K + v̄I−1,J,K + v̄I,J−1,K + v̄I,J,K ) (18.29)
4
Next, the test filtered variable is computed as
z{ 1 f f f f
v̄ I,J,K = (v̄I−1/2,J−1/2,K + v̄I−1/2,J+1/2,K + v̄I+1/2,J−1/2,K + v̄I+1/2,J+1/2,K )
4
(18.30)
18.12. The test filter 180

(I, J + 1, K)

(I −
+ 1/2, J + 1/2, K)
x x

(I, J, K) (I + 1, J, K)
(I − 1, J, K)

x x
(I − 1/2, J − 1/2, K)

x1
(I, J − 1, K)
x1

Figure 18.8: A 2D test filter control volume.

18.12.2 3D filtering
In 3D we have eight cubes. Filtering at the test level is carried out in the same way
by integrating over the test cell assuming linear variation of the variables [76]. For the
bottom-lower-left cube, for example, we get

f 1
v̄I−1/2,J−1/2,K−1/2 = (v̄I−1,J−1,K + v̄I−1,J,K + v̄I,J−1,K + v̄I,J,K
8 (18.31)
+v̄I−1,J−1,K−1 + v̄I−1,J,K−1 + v̄I,J−1,K−1 + v̄I,J,K−1 )

Next, the test filtered variable is computed as

z{ 1 f f
v̄ I,J,K = (v̄I−1/2,J−1/2,K−1/2 + v̄I+1/2,J−1/2,K−1/2
8
f f
+v̄I−1/2,J+1/2,K−1/2 + v̄I+1/2,J+1/2,K−1/2
(18.32)
f f
+v̄I−1/2,J−1/2,K+1/2 + v̄I+1/2,J−1/2,K+1/2
f f
+v̄I−1/2,J+1/2,K+1/2 + v̄I+1/2,J+1/2,K+1/2 )
18.13. Stresses on grid, test and intermediate level 181

18.13 Stresses on grid, test and intermediate level

The stresses on the grid level, test level and intermediate level (dynamic Leonard
stresses) have the form

τij = vi vj − v̄i v̄j stresses with ℓ < ∆

z { z{ z{ z{
Tij = vi vj − v̄ i v̄ j stresses with ℓ < ∆
z{ z{
Lij = Tij − τ ij stresses with ∆ < ℓ < ∆

Thus the dynamic Leonard stresses represent the stresses with lengthscale, ℓ, in the
z{
range between ∆ and ∆.
Assume now that the same functional form for the subgrid stresses that is used at the
grid level (τij ) also can be used at the test filter level (Tij ). If we use the Smagorinsky
model we get
1
τij − δij τkk = −2C∆2 |s̄|s̄ij (18.33)
3
1 z{2 z{ z{
Tij − δij Tkk = −2C ∆ | s̄ | s̄ ij (18.34)
3
where  z{ z{   
z{ 1  ∂ v̄ i ∂ v̄ j  z{ z{ z{ 1/2
s̄ ij = + , | s̄ | = 2 s̄ ij s̄ ij
2 ∂xj ∂xi

Note that C in Eq. 18.33 is not squared (cf. the Smagorinsky model, Eq. 18.13 at
p.175). Hence, C should be compared with CS2 . Applying the test filter to Eq. 18.33
(assuming that C varies slowly), substituting this equation and Eq. 18.34 into Eq. 18.27
gives !
1 z{2 z{ z{ z {
2
Lij − δij Lkk = −2C ∆ | s̄ | s̄ ij − ∆ |s̄|s̄ij (18.35)
3
Note that the “constant” C really is a function of both space and time, i.e. C =
C(xi , t).
Equation 18.35 is a tensor equation, and we have five (s̄ij is symmetric and trace-
less) equations for C. Lilly [77] suggested to satisfy Eq. 18.35 in a least-square sense.
Let us define the error as the difference between the left-hand side and the right-hand
side of Eq. 18.35 raised to the power of two, i.e.
 2
1
Q= Lij − δij Lkk + 2CMij (18.36a)
3
!
z{2 z{ z{ z {
2
Mij = ∆ | s̄ | s̄ ij − ∆ |s̄|s̄ij (18.36b)

The error, Q, has a minimum (or maximum) when ∂Q/∂C = 0. Carrying out the
derivation of 18.36a gives
 
∂Q 1
= 4Mij Lij − δij Lkk + 2CMij = 0 (18.37)
∂C 3
18.14. Numerical dissipation 182

v̄I

I −1 I +1
I

Figure 18.9: Numerical dissipation.

Since ∂ 2 Q/∂C 2 = 8Mij Mij > 0 it is a minimum. Equation 18.36 is re-written so

that
Lij Mij
C=− (18.38)
2Mij Mij
It turns out that the dynamic coefficient C fluctuates wildly both in space and time.
This causes numerical problems, and it has been found necessary to average C in homo-
geneous direction(s). Furthermore, C must be clipped to ensure that the total viscosity
stays positive (ν + νsgs ≥ 0).
In real 3D flows, there is no homogeneous direction. Usually local averaging and
clipping (i.e. requiring that C stays within pre-defined limits) of the dynamic coeffi-
cient is used.
Use of one-equation models solve these numerical problems (see p. 189).

18.14 Numerical dissipation

The main function of an SGS model is to dissipate (i.e. to dampen) resolved turbulent
fluctuations. The SGS model is – hopefully – designed to give a proper amount of
dissipation. This is the reason why in LES we should use a central differencing scheme,
because this class of schemes does not give any numerical dissipation. All upwind
schemes give numerical dissipation in addition to the modeled SGS dissipation. Indeed,
there are LES-methods in which upwind schemes are used to create dissipation and
where no SGS model is used at all (e.g. MILES [78]). However, in this course we
focus on ensuring proper dissipation through an SGS model rather than via upwind
differencing. It can be shown using Neumann stability analysis that all upwind schemes
are dissipative (see Further reading at
http://www.tfd.chalmers.se/˜lada/comp turb model/). Below it is
shown that first-order upwind schemes are dissipative.
The first-derivative in the convective term is estimated by first-order upwind differ-
encing as (finite difference, see Fig. 18.9)
   
∂v̄ v̄I − v̄I−1
v̄ = v̄I + O (∆x) (18.39)
∂x exact ∆x

where we have assumed v̄I > 0. Taylor expansion gives

∂v̄ 1 ∂ 2 v̄

v̄I−1 = v̄I − ∆x + (∆x)2 2 + O (∆x)3
∂x 2 ∂x
so that
v̄I − v̄I−1 ∂v̄ 1 ∂ 2 v̄

= − ∆x 2 + O (∆x)2
∆x ∂x 2 ∂x
18.15. Scale-similarity Models 183

Insert this into Eq. 18.39

 
∂v̄ ∂v̄ 1 ∂ 2 v̄

v̄ = v̄ − ∆xv̄ 2 +v̄O (∆x)2 (18.40)
∂x exact ∂x 2 ∂x
O(∆x)

where the second term on the right side corresponds to the error term in Eq. 18.39.
When this expression is inserted into the LES momentum equations, the second term
on the right-hand side will act as an additional (numerical) diffusion term. The total
diffusion term will have the form
 
∂ ∂v̄
diffusion term = (ν + νsgs + νnum ) (18.41)
∂x ∂x

where the additional numerical viscosity, νnum = 0.5|v̄|∆x, see Eq. 18.40. This means
that the total dissipation due to SGS viscosity and numerical viscosity is (cf. Eq. 18.19)

εsgs+num = 2(νsgs + νnum )s̄ij s̄ij

For more details on derivation of equations transport equations of turbulent kinetic

energies, see [23].

18.15 Scale-similarity Models

In the models presented in the previous sections (the Smagorinsky and the dynamic
models) the total SGS stress τij = vi vj − v̄i v̄j was modeled with an eddy-viscosity
hypothesis. In scale-similarity models the total stress is split up as

τij = vi vj − v̄i v̄j = (v̄i + vi′′ )(v̄j + vj′′ ) − v̄i v̄j

= v̄i v̄j + v̄i vj′′ + v̄j vi′′ + vi′′ vj′′ − v̄i v̄j
h i
= (v̄i v̄j − v̄i v̄j ) + v̄i vj′′ + v̄j vi′′ + vi′′ vj′′

where the term in brackets is denoted the Leonard stresses, the term in square brackets
is denoted cross terms, and the last term is denoted the Reynolds SGS stress. Thus

τij = Lij + Cij + Rij

Lij = v̄i v̄j − v̄i v̄j
(18.42)
Cij = v̄i vj′′ + v̄j vi′′
Rij = vi′′ vj′′ .

Note that the Leonard stresses, Lij , are computable, i.e. they are exact and don’t need
to be modeled.
In scale-similarity models the main idea is that the turbulent scales just above cut-
off wavenumber, κc , (scales smaller than ∆) are similar to the ones just below κc
(scales larger than ∆); hence the word ”scale-similar”. Looking at Eq. 18.42 it seems
natural to assume that the cross term is responsible for the interaction between resolved
scales (v̄i ) and modeled scales (vi′′ ), since Cij includes both scales.
18.16. The Bardina Model 184

18.16 The Bardina Model

In the Bardina model the Leonard stresses Lij are computed explicitly, and the sum of
the cross term Cij and the Reynolds term is modeled as [79, 80]
M
Cij = cr (v̄i v̄j − v̄ i v̄ j ) (18.43)
M
and Rij = 0 (superscript M denotes Modeled). It was found that this model was not
sufficiently dissipative, and thus a Smagorinsky model was added
M
Cij = cr (v̄i v̄j − v̄ i v̄ j )
M
(18.44)
Rij = −2CS2 ∆2 |s̄|s̄ij

18.17 Redefined terms in the Bardina Model

In [80] it was found that the Leonard term Lij and the cross term Cij are not Galilean
invariant by themselves, but only the sum Lij + Cij is (see Appendix I). As a con-
sequence, if the cross term is neglected, the Leonard stresses must not be computed
explicitly, because then the modeled momentum equations do not satisfy Galilean in-
variance.
The stresses in the Bardina model can be redefined to make them Galilean invariant
for any value cr A modified Leonard stress tensor Lmij is defined as [81]

τijm = τij = Cij

m
+ Lm m
ij + Rij
Lm
ij = cr (v̄i v̄j − v̄ i v̄ j )
m (18.45)
Cij =0
m
Rij = Rij = vi′′ vj′′

Note that the modified Leonard stresses is the same as the “unmodified” one plus
the modeled cross term Cij in the Bardina model with cr = 1 (right-hand side of
Eq. 18.43), i.e.
Lm
ij = Lij + Cij
M

In order to make the model sufficiently dissipative a Smagorinsky model is added,

and the total SGS stress τij is modeled as

τij = v̄i v̄j − v̄ i v̄ j − 2(CS ∆)2 |s̄|s̄ij (18.46)

Below we verify that the modified Leonard stress is Galilean invariant.

1 m∗ ∗ ∗
L = v̄i∗ v̄j∗ − v̄ i v̄ j = (v̄i + Vi )(v̄j + Vj ) − (v̄i + Vi ) (v̄j + Vj )
cr ij
= v̄i v̄j + v̄i Vj + v̄j Vi − v̄ i v̄ j − v̄ i Vj − Vi v̄ j (18.47)
1
= v̄i v̄ j − v̄ i v̄ j = Lm
cr ij

18.18 A dissipative scale-similarity model.

Above it was mentioned that when the first scale-similarity model was proposed it
was found that it is not sufficiently dissipative [79]. An eddy-viscosity model has to
18.18. A dissipative scale-similarity model. 185

E(κ)
10

0
ε+
SGS
−10
ε−
SGS

−20

−30

0 50 100 150
κc κ
x+
2
Figure 18.10: Dissipation terms and production term from DNS data. 963 mesh data
filtered onto a 483 mesh. Reτ = 500. : −ε+SGS ; : −ε−
SGS ; +: −εSGS .

be added to make the model sufficiently dissipative; these models are called mixed
models. Ref. [82] presents and evaluates a dissipative scale-similarity model.
The filtered Navier-Stokes read
dv̄i 1 ∂ p̄ ∂ 2 v̄i ∂τik
+ =ν − (18.48)
dt ρ ∂xi ∂xk ∂xk ∂xk

where d/dt and τik denote the material derivative and the SGS stress tensor, respec-
tively. The SGS stress tensor is given by

τik = vi vk − v̄i v̄k . (18.49)

When it is modeled with the standard scale-similarity model, it is not sufficiently dis-
sipative. Let us take a closer look at the equation for the resolved, turbulent kinetic
energy, k = hvi′ vi′ i/2, which reads
 2 ′ 
dk ∂hv̄i i ∂hp̄′ v̄i′ i 1 ∂hv̄k′ v̄i′ v̄i′ i ∂ v̄i ′
+ hv̄k′ v̄i′ i + + =ν v̄i −
dt ∂xk ∂xi 2 ∂xk ∂xk ∂xk
     2 ′   
∂τik ∂τik ∂ v̄ ∂τik ′
− v̄i′ = ν i
v̄i′ − v̄i =
∂xk ∂xk ∂xk ∂xk ∂xk (18.50)
2
 ′ ′
  
∂ k ∂v̄i ∂v̄i ∂τik ′
ν −ν − v̄
∂xk ∂xk ∂xk ∂xk ∂xk i
ε εSGS

The first term on the last line is the viscous diffusion term and the second term, ε, is
the viscous dissipation term which is always positive. The last term, εSGS , is a source
term arising from the SGS stress tensor, which can be positive or negative. When it is
positive, forward scattering takes place (i.e. it acts as a dissipation term); when it is
negative, back scattering occurs.
Figure 18.10 presents SGS dissipation, εSGS in Eq. 18.50, computed from filtered
DNS data. The forward scatter, ε+ −
SGS , and back scatter, εSGS , SGS dissipation are
defined as the sum of all instants when εSGS is positive and negative, respectively. As
can be seen, the scale-similarity model is slightly dissipative (i.e. εSGS > 0) , but the
forward and back scatter dissipation are both much larger than εSGS .
18.19. Forcing 186

One way to make the SGS stress tensor strictly dissipative is to set the back scatter
to zero, i.e. max(εSGS , 0). This could be achieved by setting ∂τik /∂xk = 0 when its
sign is different from that of v̄i′ (see the last term in Eq. 18.50). This would work if we
were solving for k. Usually we do not, and the equations that we do solve (the filtered
Navier-Stokes equations) are not directly affected by the dissipation term, εSGS .
Instead we have to modify the SGS stress tensor as it appears in the filtered Navier-
Stokes equations, Eq. 18.48. The second derivative on the right side is usually called a
diffusion term because it acts like a diffusion transport term. When analyzing the sta-
bility properties of discretized equations to an imposed disturbance, v̄ ′ , using Neumann
analysis (see, for example, Chapter 8 in [83]), this term is referred to as a dissipation
term. In stability analysis the concern is to dampen numerical oscillations; in con-
nection with SGS models, the aim is to dampen turbulent resolved fluctuations. It is
shown in Neumann analysis that the diffusion term in the Navier-Stokes equations is
dissipative, i.e. it dampens numerical oscillations. However, since it is the resolved
turbulent fluctuations, i.e. k in Eq. 18.50, that we want to dissipate, we must consider
the filtered Navier-Stokes equations for the fluctuating velocity, v̄i′ . The viscous diffu-
sion term in the momentum equations appears in the first term on the right side (first
line) in Eq. 18.50. To ensure that εSGS > 0, we set −∂τik /∂xk to zero when its sign
is different from that of the viscous diffusion term (cf. the two last terms on the second
line in Eq. 18.50). This is achieved by defining a sign function; for details, see [82].

18.19 Forcing
An alternative way to modify the scale-similarity model is to omit the forward scatter,
i.e. to include instants when the subgrid stresses act as counter-gradient diffusion. In
hybrid LES-RANS, the stresses can then be used as forcing at the interface between
URANS and LES. This new approach is the focus of [84].

18.20 Numerical method

A numerical method based on an implicit, finite volume method with collocated grid
arrangement, central differencing in space, and Crank-Nicolson (α = 0.5) in time is
briefly described below. The discretized momentum equations read
 
n+1/2 n+1/2
v̄i = v̄in + ∆tH v̄ n , v̄i
(18.51)
∂ p̄n+1/2 ∂ p̄n
−α∆t − (1 − α)∆t
∂xi ∂xi
where H includes convective, viscous and SGS terms. In SIMPLE notation this equa-
tion reads
n+1/2
X ∂ p̄n+1/2
aP v̄i = anb v̄ n+1/2 + SU − α∆t ∆V
∂xi
nb

where SU includes all source terms except the implicit pressure. The face velocities
n+1/2 n+1/2 n+1/2
v̄f,i = 0.5(v̄i,j + v̄i,j−1 ) (note that j denotes node number and i is a tensor
index) do not satisfy continuity. Create an intermediate velocity field by subtracting
18.20. Numerical method 187

the implicit pressure gradient from Eq. 18.51, i.e.

  ∂ p̄n
n+1/2
v̄i∗ = v̄in + ∆tH v̄ n , v̄i − (1 − α)∆t (18.52a)
∂xi
n+1/2
n+1/2 ∂ p̄
⇒ v̄i∗ = v̄i + α∆t (18.52b)
∂xi
n+1/2
Take the divergence of Eq. 18.52b and require that ∂v̄f,i /∂xi = 0 so that
∗
∂ 2 p̄n+1 1 ∂v̄f,i
= (18.53)
∂xi ∂xi ∆tα ∂xi

The Poisson equation for p̄n+1 is solved with an efficient multigrid method [85]. In the
3D MG we use a plane-by-plane 2D MG. The face velocities are corrected as

n+1 ∗ ∂ p̄n+1
v̄f,i = v̄f,i − α∆t (18.54)
∂xi
A few iterations (typically two) solving the momentum equations and the Poisson pres-
sure equation are required each time step to obtain convergence. More details can be
found [86].

1. Solve the discretized filtered Navier-Stokes equation, Eq. 18.52a, for v̄1 , v̄2 and
v̄3 .
2. Create an intermediate velocity field v̄i∗ from Eq. 18.52b.
3. Use linear interpolation to obtain the intermediate velocity field, v̄f,i , at the face
4. The Poisson equation (Eq. 18.53) is solved with an efficient multigrid method [85].
5. Compute the face velocities (which satisfy continuity) from the pressure and the
intermediate face velocity from Eq. 18.54
6. Step 1 to 4 is performed till convergence (normally two or three iterations) is
reached.
7. The turbulent viscosity is computed.
8. Next time step.

Since the Poisson solver in [85] is a nested MG solver, it is difficult to parallelize

with MPI (Message Passing Interface) on large Linux clusters. Hence, when we do
large simulations (> 20M cells) we use a traditional SIMPLE method.

18.20.1 RANS vs. LES

Above a numerical procedure suitable for LES was described. However, in general,
any numerical procedure for RANS can also be used for LES; for example pressure-
correction methods such as SIMPLE [87, 88] are often used for LES. What are the
specific requirements to carry out LES with a finite volume code? If you have a RANS
finite volume code, it is very simple to transform that into an LES code. An LES code
is actually simpler than a RANS code. Both the discretization scheme and and the
turbulence model are simpler in LES than in RANS, see Table 18.1.
18.20. Numerical method 188

RANS LES
Domain 2D or 3D always 3D
Time domain steady or unsteady always unsteady
Space discretization 2nd order upwind central differencing
Time discretization 1st order 2nd order (e.g. C-N)
Turbulence model more than two-equations zero- or one-equation

Table 18.1: Differences between a finite volume RANS and LES code.
v̄1

t
t1 : start t2 : end

Figure 18.11: Time averaging in LES.

It is important to use a non-dissipative discretization scheme which does not intro-

duce any additional numerical dissipation, see Section 18.14; hence a second-order (or
higher) central differencing scheme should be employed.
The time discretization should also be non-dissipative. The Crank-Nicolson scheme
is suitable.
As mentioned above, turbulence models in LES are simple. There are two reasons:
first, only the small-scale turbulence is modeled and, second, no equation for the tur-
bulent length scale is required since the turbulent length scale can be taken as the filter
width, ∆.
In LES we are doing unsteady simulations. The question then arises, when can we
start to time average and for how long? This is exactly the same question we must ask
ourselves whenever doing an experiment in, for example, a windtunnel. We start the
windtunnel: when has the flow (and the turbulence) reached fully developed conditions
so that we can start to measure the flow? Next question: for how long should we carry
out the measurements?
Both in LES and the windtunnel, the recorded time history of the v̄1 velocity at a
point may look like in Fig. 18.11. Time averaging can start at time t1 when the flow
seems to have reached fully developed conditions. It is difficult to judge for how long
one should carry out time averaging. Usually it is a good idea to form a characteristic
time scale from a velocity, V (free-stream or bulk velocity), and a length scale, L
(width of a wake or a body, length of a recirculation region), and use this to estimate
the required averaging time; 100 time units, i.e. 100L/V , may be a suitable averaging
time for the flow around a bluff body; a value of 10 may be sufficient if L is the
length of a recirculation region. The theoretical statistical error varies with number of
independent samples, N , as N −1/2 , see Eq. M.2 on p. 307.
18.21. One-equation ksgs model 189

18.21 One-equation ksgs model

A one-equation model can be used to model the SGS turbulent kinetic energy. The
equation can be written on the same form as the RANS k-equation, i.e.
 
∂ksgs ∂ ∂ ∂ksgs
+ (v̄j ksgs ) = (ν + νsgs ) + Pksgs − ε
∂t ∂xj ∂xj ∂xj
3/2
(18.55)
1/2 ksgs
νsgs = ck ∆ksgs , Pksgs = 2νsgs s̄ij s̄ij , ε = Cε
∆
Note that the production term, Pksgs , is equivalent to the SGS dissipation in the equa-
tion for the resolved turbulent kinetic energy (look at the flow of kinetic energy dis-
cussed at the end of [89]).

18.22 Smagorinsky model derived from the ksgs equation

We can use the one-equation model to derive the Smagorinsky model, Eq. 18.13. The
length scale in the Smagorinsky model is the filter width, ∆ ∝ κ−1
II , see Fig. 18.12. The
cut-off takes place in the inertial subrange where diffusion and convection in the ksgs
equation are negligible (their time scales are too large so they have no time to adapt
to rapid changes in the velocity gradients, s̄ij ). Hence, production and dissipation in
Eq. 18.55 are in balance so that

Pksgs = 2νsgs s̄ij s̄ij = ε (18.56)

Let us replace ε by SGS viscosity and ∆. We can write the SGS viscosity as

νsgs = εa (CS ∆)b (18.57)

Dimensional analysis yields a = 1/3, b = 4/3 so that

νsgs = (CS ∆)4/3 ε1/3 . (18.58)

Eq. 18.56 substituted into Eq. 18.58 gives

3
νsgs = (CS ∆)4 ε = (CS ∆)4 νsgs (2s̄ij s̄ij )
⇒ νsgs = (CS ∆)2 |s̄| (18.59)
1/2
|s̄| = (2s̄ij s̄ij )

which is the Smagorinsky model.

18.23 A dynamic one-equation model

One of the drawbacks of the dynamic model of [75] (see p. 178) is the numerical
instability associated with the negative values and large variation of the C coefficient.
Usually this problem is fixed by averaging the coefficient in some homogeneous flow
direction. In real applications ad-hoc local smoothing and clipping is used. Below
a dynamic one-equation model is presented. The main object when developing this
model was that it should be applicable to real industrial flows. Furthermore, being a
dynamic model, it has the great advantage that the coefficients are computed rather than
being prescribed.
18.23. A dynamic one-equation model 190

E(κ) I

II

III

κ
Figure 18.12: Spectrum for k. I: Range for the large, energy containing eddies; II:
the inertial subrange for isotropic scales, independent of the large scales (ℓ) and the
dissipative scales (ν); III: Range for small, isotropic, dissipative scales.

The equation for the subgrid kinetic energy reads [90, 91] (see also [92, 93])
  3/2
∂ksgs ∂ ∂ ∂ksgs ksgs
+ (v̄j ksgs ) = Pksgs + νef f − C∗
∂t ∂xj ∂xj ∂xj ∆ (18.60)
1
Pksgs = −τija v̄i,j , τija = −2C∆ksgs
2
s̄ij
1
with νef f = ν + 2Chom ∆ksgs 2
. The C in the production term Pksgs is computed
dynamically (cf. Eq. 18.38). To ensure numerical stability, a constant value (in space)
of C (Chom ) is used in the diffusion term in Eq. 18.60 and in the momentum equations.
Chom is computed by requiring that Chom should yield the same total production of
ksgs as C, i.e.
1 1
h2C∆ksgs
2
s̄ij s̄ij ixyz = 2Chom h∆ksgs
2
s̄ij s̄ij ixyz
The dissipation term εksgs is estimated as:
3/2
ksgs
εksgs ≡ νTf (vi,j , vi,j ) = C∗ . (18.61)
∆
Now we want to find a dynamic equation for C∗ . The equations for ksgs and K read in
symbolic form
3/2
ksgs
T (ksgs ) ≡ Cksgs − Dksgs = Pksgs − C∗
∆
3/2 (18.62)
K K
T (K) ≡ CK − DK = P − C∗ z{
∆
Since the turbulence on both the grid level and the test level should be in local equilib-
rium (in the inertial −5/3 region), the left-hand side of the two equations in Eq. 18.62
should be close to zero. An even better approximation should be to assume T (ksgs ) =
T (K), i.e.
z{ 1z {3/2 K 3/2
P ksgs − C∗ ksgs = P K − C∗ z{ ,
∆
∆
so that
z{ z {! z{
n+1 K 1 n 3/2 ∆
C∗ = P − P ksgs + C∗ ksgs 3 . (18.63)
∆ K2
18.24. A Mixed Model Based on a One-Eq. Model 191

The idea is to put the local dynamic coefficients in the source terms, i.e. in the produc-
tion and the dissipation terms of the ksgs equation (Eq. 18.60). In this way the dynamic
coefficients C and C∗ don’t need to be clipped or averaged in any way. This is a big
advantage compared to the standard dynamic model of Germano (see discussion on
p. 182).

18.24 A Mixed Model Based on a One-Eq. Model

Recently a new dynamic scale-similarity model was presented by [94]. In this model a
dynamic one-equation SGS model is solved, and the scale-similarity part is estimated
in a similar way as in Eq. 18.46.

18.25 Applied LES

At the Department we used LES for applied flows such as flow around a cube [95,
96], the flow and heat transfer in a square rotating duct [97, 98], the flow around a
simplified bus [96, 99], a simplified car [100–102] and the flow around an airfoil [103,
104], detailed SUV [105], trains and buses subjected to sidewinds and wind gusts [106–
108]. We have also done some work on buoyancy-affected flows [109–115].

18.26 Resolution requirements

The near-wall grid spacing should be about one wall unit in the wall-normal direction.
This is similar to the requirement in RANS (Reynolds-Averaged Navier-Stokes) using
low-Re number models. The resolution requirements in wall-parallel planes for a well-
resolved LES in the near-wall region expressed in wall units are approximately 100
(streamwise) and 30 (spanwise). This enables resolution of the near-wall turbulent
structures in the viscous sub-layer and the buffer layer consisting of high-speed in-
rushes and low-speed ejections [116], often called the streak process. At low to medium
Reynolds numbers the streak process is responsible for the major part of the turbulence
production. These structures must be resolved in an LES in order to achieve accurate
results. Then the spectra of the resolved turbulence will exhibit −5/3 range, see figure
on p. 69.
In applied LES, this kind of resolution can hardly ever be afforded. In outer scaling
(i.e. comparing the resolution to the boundary layer thickness, δ), we can afford δ/∆x1
and δ/∆x3 in the region of 10 − 20 and 20 − 40, respectively. In this case, the spectra
in the boundary layer will look something like that shown in Fig. 18.13 [117]. Energy
spectra are actually not very reliable to judge if a LES simulation is well resolved or
not. In [117,118] different ways to estimate the resolution of an LES were investigated.
The suggestion in these works is that two-point correlations is the best way to estimate
if an LES is sufficiently resolved or not.
Even if the turbulence in boundary layer seldom can be resolved, the flow in re-
circulation regions and shear layer can. In [119] the flow (Re ≃ 106 ) over a bump
was computed. The geometry is shown in Fig. 18.14. The turbulence in the bound-
ary layer on the bump was very poorly resolved: ∆x1 /δin = 0.33, ∆x3 /δin = 0.44,
∆x+ +
1 = 1300 and ∆x3 = 1800. Nevertheless, the turbulence in the recirculation re-
gion and in the shear layer downstream the bump turned out to be well resolved, see
Fig. 18.15.
For wall-bounded flows at high Reynolds numbers of engineering interest, the
computational resource requirement of accurate LES is prohibitively large. Indeed,
18.26. Resolution requirements 192

Thick dashed line −5/3 slope

−1
10

E33 (k3 )
−2
10

−3
10

−4
10 0 1 2
10 10 10
κ3 = 2πk3 /x3,max
Figure 18.13: Energy spectra in fully developed channel flow [117]. δ denotes half
channel width. Number of cells expressed as (δ/∆x1 , δ/∆x3 ). : (10, 20); :
(20, 20); : (10, 40); ◦: (5, 20); +: (10, 10).
L1

V1,in

y H

h
x

L2
L

Figure 18.14: Onera bump. Computational domain (not to scale).

the requirement of near-wall grid resolution is the main reason why LES is too ex-
pensive for engineering flows, which was one of the lessons learned in the LESFOIL
project [120, 121].
19. Unsteady RANS 193

−1
10

E33 (κ3 )
−2
10

−3
10
10 100 200
κ3 = 2πk3 /x3,max

Figure 18.15: Energy spectra E33 (κ3 ) in the recirculation region and the shear layer
downstream the bump (x1 /H = 1.2). Thick dashed line shows −5/3 slope. :
x2 /H = 0.0035 (near the wall); : x2 /H = 0.13; : x2 /H = 0.34 (in the
shear layer).

19 Unsteady RANS
O perform an accurate LES, a very fine mesh must be used. This causes problems,
T for example, near walls. LES is very good for wake flow, where the flow is gov-
erned by large, turbulent structures, which can be captured by a fairly coarse mesh.
However, if attached boundary layers are important, LES will probably give poor pre-
dictions in these regions, unless fine grids are used.
An alternative to LES for industrial flows can be unsteady RANS (Reynolds-
Averaged Navier-Stokes), often denoted URANS (Unsteady RANS). In URANS the
usual Reynolds decomposition is employed, i.e.
Z t+T
1
v̄(t) = v(t)dt, v = v̄ + v ′′ (19.1)
2T t−T
The URANS equations are the usual RANS equations, but with the transient (unsteady)
term retained; on incompressible form they read

∂v̄i ∂ 1 ∂ p̄ ∂ 2 v̄i ∂vi′′ vj′′

+ (v̄i v̄j ) = − +ν −
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj (19.2)
∂ v¯i
=0
∂xi
Note that the dependent variables are now not only function of the space coordinates,
but also function of time, i.e. v̄i = v̄i (x1 , x2 , x3 , t), p̄ = p̄(x1 , x2 , x3 , t) and vi′′ vj′′ =
vi′′ vj′′ (x1 , x2 , x3 , t).
Even if the results from URANS are unsteady, one is often interested only in the
time-averaged flow. We denote here the time-averaged velocity as hv̄i, which means
that we can decompose the results from an URANS as a time-averaged part, hv̄i, a
resolved fluctuation, v̄ ′ , and the modeled, turbulent fluctuation, v ′′ , i.e.

v = v̄ + v ′′ = hv̄i + v̄ ′ + v ′′ (19.3)
19. Unsteady RANS 194

see Fig. 19.1. The modeled turbulent fluctuation, v ′′ , is not shown in the figure; when
this is added to hv̄i + v̄ ′ we obtain v.
What type of turbulence model should be used in URANS? That depends on type
of flow. If the flow has strong vortex shedding, the standard high-Re number k − ε
model can be used, i.e.
  
∂ρk ∂ρv̄j k ∂ µt ∂k
+ = µ+ + P k − ρε (19.4)
∂t ∂xj ∂xj σk ∂xj
  
∂ρε ∂ρv̄j ε ∂ µt ∂ε ε

+ = µ+ + c1ε P k − cε2 ρε (19.5)
∂t ∂xj ∂xj σε ∂xj k
k2
µt = c µ ρ (19.6)
ε
With an eddy-viscosity, the URANS equations read
 
∂ρv̄i ∂ρv̄i v̄k ∂ p̄ ∂ ∂v̄i
+ =− + (µ + µt ) (19.7)
∂t ∂xk ∂xi ∂xk ∂xk

Usually the standard k − ε model is not a good URANS model because it gives too
much modeled dissipation (i.e. too large turbulent viscosity), which dampens the re-
solved fluctuations too much. The V2F and non-linear eddy-viscosity models are bet-
ter.
So we are doing unsteady simulations, but still we time average the equations. How
is this possible? The theoretical answer is that the time, T , in Eq. 19.1 should be much
smaller than the resolved time scale, i.e. the modeled turbulent fluctuations, v ′′ , should
have a much smaller time scale than the resolved ones, v̄ ′ . This is called scale separa-
tion. In practice this requirement is often not satisfied [86]. On the other hand, how do
the momentum equation, Eq. 19.7, know how they were time averaged? Or if they were
volume filtered? The answer is that they don’t. The URANS momentum equation and
the LES momentum equation are exactly the same, except that we denote the turbulent
viscosity in the former case by νt and in the latter case by νsgs . In URANS, much
more of the turbulence is modeled than in LES, and, hence, the turbulent viscosity, νt ,
is much larger than the SGS viscosity, νsgs .
The common definition of URANS is that the turbulent length scale is not deter-
mined by the grid, whereas in LES it is. In URANS we do usually not care about scale
separation. What we care about is that the turbulence model and the discretization

v, v̄, hv̄i

hv̄i

Figure 19.1: Decomposition of velocities in URANS. : v̄; : v; hv̄i.

19. Unsteady RANS 195

Figure 19.2: Configuration of the flow past a triangular flameholder. Flow from left to
right

Figure 19.3: 2D URANS k − ε simulations [122]. One cycle of the v̄2 velocity in a
cell near the upper-right corner of the flameholder.

scheme should not be too dissipative, i.e. they should not kill the resolved fluctuations,
v̄ ′ .
The standard k − ε model (Eq. 19.4 and 19.5) was used in [122] for two-dimensional
URANS simulations computing the flow around a triangular flame-holder in a channel,
see Fig. 19.2. This flow has a very regular vortex shedding. and the flow actually has a
scale separation. In Fig. 19.3 the v̄2 velocity in a point above the flame-holder is shown
and it can be seen that the velocity varies with time in a sinusoidal manner.
When we’re doing URANS, the question arises how the results should be time
averaged, i.e. when should we start to average and for how long. This issue is the same
when doing LES, and this was discussed in connection to Fig. 18.11.
19.1. Turbulence Modeling 196

Figure 19.4: 2D URANS k − ε simulations compared with experiment [122]. Solid

lines: total turbulent kinetic energy; dashed lines: resolved turbulent kinetic energy:
∗: experimental data. Left figure: x = 0.43H; right figure: x = 1.1H (x = 0 at the
downstream vertical plane of the flame-holder).

19.1 Turbulence Modeling

In URANS, part of the turbulence is modeled (v ′′ ) and part of the turbulence is resolved
(v̄ ′ ). If we want to compare computed turbulence with experimental turbulence, we
must add these two parts together. Profiles downstream the flameholder are shown in
Fig. 19.4. It can be seen that here the resolved and the modeled turbulence are of the
same magnitude.
If the turbulence model in URANS generates ”too much” eddy viscosity, the flow
may not become unsteady at all, because the unsteadiness is dampened out; the reason
for this is that the turbulence model is too dissipative. It was found in [123, 124] when
using URANS for the flow around a surface-mounted cube and around a car, that the
standard k − ε model was too dissipative. Non-linear models like that of [125] was
found to be less dissipative, and was successfully applied in URANS-simulations for
these two flows.

19.2 Discretization
In LES it is well-known that non-dissipative discretization schemes should be used.
The reason is that we don’t want to dampen out resolved, turbulent fluctuations. This is
to some extent true also for URANS. In the predictions on the flame-holder presented
above, the hybrid discretization scheme for the convective terms was used together
with fully implicit first-order discretization in time; this gives first-order accuracy in
both space and time. The turbulence model that was used was the standard k − ε
model. Thus, both the discretization and the turbulence model have high dissipation.
The reason why the unsteadiness in these computations was not dampened out is that
the vortex shedding in this flow is very strong.
In general, a discretization scheme which has little numerical dissipation should be
19.2. Discretization 197

Figure 19.5: URANS simulations of the flow around a surface-mounted cube.

used. How dissipative a scheme needs to be in order to be stable is flow dependent; for
some simple flows, it may work with no dissipation at all (i.e. central differencing),
whereas for industrially complex flows maybe a bounded second-order scheme must
be used. For time discretization, the second-order accurate Crank-Nicolson works in
most cases.
In [123] LES and URANS simulations were carried out of the flow around a surface-
mounted cube (Fig. 19.5) with a coarse mesh using wall-functions. Two different dis-
cretization schemes were used: the central scheme and the Mars scheme (a blend be-
tween central differencing and a bounded upwind scheme of second-order accuracy).
In Fig. 19.6 the time-averaged velocity profile upstream of the cube (x1 = −0.6H)
using URANS and LES with central differencing are shown together with URANS and
the Mars scheme. It is seen that with LES and central differencing nonphysical oscil-
lations are present (this was also found in [95]). However, LES with the Mars scheme
(in which some numerical dissipation is present) and URANS with the central scheme
(where the modeling dissipation is larger than in LES) no such nonphysical oscillations
are present. The main reason to the nonphysical oscillations is that the predicted flow
in this region does not have any resolved fluctuations. If turbulent unsteady inlet fluc-
tuations are used, the nonphysical oscillations do usually not appear, even if a central
differencing scheme is used. In this case the turbulent, resolved fluctuations dominate
over any numerical oscillations.
20. DES 198

Figure 19.6: URANS simulations of the flow around a surface-mounted cube. Velocity
profiles upstream the cube [123].

20 DES
ES (Detached Eddy Simulation) is a mix of LES and URANS. The aim is to treat
D the boundary layer with RANS and capture the outer detached eddies with LES.
Since the flow in the boundary layer will be strongly influenced by the unsteady LES
in the outer region, the flow in the boundary layer will also be unsteady. Hence the
boundary layer is treated with unsteady RANS (URANS). The DES was originally
developed for wings at very high angles of attack.
The RANS model that was originally used was the one-equation model proposed
in [126]. It can be written [120, 126, Sect. 4.6]
 
∂ρν̃t ∂ρv̄j ν̃t ∂ µ + µt ∂ ν̃t Cb2 ρ ∂ ν̃t ∂ ν̃t
+ = + +P −Ψ
∂t ∂xj ∂xj σν˜t ∂xj σν˜t ∂xj ∂xj (20.1)
νt = ν̃t f1
The production term P and the destruction term Ψ have the form
 
ν̃t
P = Cb1 ρ s̄ + 2 2 f2 ν̃t
κ d
 2 (20.2)
ν̃t
s̄ = (2s̄ij s̄ij )1/2 , Ψ = Cw1 ρfw
d
d in the RANS SA model is equal to the distance to the nearest wall.
In DES [127], d is taken as the minimum of the RANS turbulent length scale d and
the cell length ∆ = max(∆xξ , ∆xη , ∆xζ ), i.e.

d˜ = min(d, Cdes ∆). (20.3)

∆xξ , ∆xη and ∆xζ denote the cell length in the three grid directions ξ, η and ζ. The
constant Cdes is usually set to 0.65.
20.1. DES based on two-equation models 199

In the boundary layer d < Cdes ∆ and thus the model operates in RANS mode.
Outside the turbulent boundary layer d > Cdes ∆ so that the model operates in LES
mode. The modeled length scale is reduced and the consequence is that the destruction
term Ψ increases, which gives a reduction in the turbulent viscosity ν̃t . A reduced ν̃t
gives a smaller production term P so that the turbulent viscosity is further reduced.
At first sight it may seem that as the model switches from RANS mode to LES
mode thus reducing d, this would give rise to an increased production term P through
the second term (see Eq. 20.2). However, this second term is a viscous term and is
active only close to the wall. This term is sometimes neglected [128].

20.1 DES based on two-equation models

The model described above is a one-equation model. In RANS mode it takes its length
scale from the wall distance, which in many situations is not a relevant turbulent length
scale. Recently, DES models based on two-equation models were proposed [129–131].
In these models the turbulent length scale is either obtained from the two turbulent
quantities (e.g. k 3/2 /ε or k 1/2 /ω) or the filter width ∆. A model based on the k − ε
model can read
  
∂k ∂ ∂ νt ∂k
+ (v̄j k) = ν+ + P k − εT
∂t ∂xj ∂xj σk ∂xj
  
∂ε ∂ ∂ νt ∂ε ε
+ (v̄j ε) = ν+ + (C1 P k − C2 ε)
∂t ∂xj ∂xj σε ∂xj k
P k = 2νt s̄ij s̄ij , νt = k 1/2 ℓt

The turbulent length scale, ℓt , and the turbulent dissipation, εT , are computed as [131,
132]
 
k 3/2
ℓt = min Cµ , Ck ∆
ε
 
k 3/2
εT = max ε, Cε
∆

In other models [68, 129] only the dissipation term, εT is modified. When the grid
is sufficiently fine, the length scale is taken as ∆. The result is that the dissipation in
the k equation increases so that k decreases which gives a reduced νt . In regions where
the turbulent length scales are taken from ∆ (LES mode) the ε-equation is still solved,
but ε is not used. However, ε is needed as soon as the model switches to RANS model
again. A third alternative is to modify only the turbulent length scale appearing in the
turbulent viscosity [132].
A rather new approach is to reduce the destruction term in the ε equation as in
PANS [133, 134] (Partially Averaged Navier-Stokes, see Section 23) and PITM [135]
(Partially Integrated Transport Modeling, see Section 24). In these models ε increases
because of its reduced destruction term which decreases both k and νt . A low-Reynolds
number PANS was recently proposed [134] in which the near-wall modifications were
taken from the AKN model [136]. In [137] different ways of treating the interface
between the URANS and LES regions were evaluated.
In the RANS mode the major part of the turbulence is modeled. When the model
switches to LES mode, the turbulence is supposed to be represented by resolved tur-
bulence. This poses a major problem with this type of models. If the switch occurs at
20.2. DES based on the k − ω SST model 200

location x1 , say, it will take some distance L before the momentum equations start to
resolve any turbulence. This is exactly what happens at an inlet in an LES simulation
if no real turbulence is given as inlet boundary conditions. One way to get around this
is to impose turbulence fluctuations as forcing conditions [89, 138–141] at the location
where the model switches from RANS mode to LES mode. The forcing is added in the
form of a source term (per unit volume) in the momentum equations.

20.2 DES based on the k − ω SST model

The standard k − ω model SST reads [62, 68]

  
∂k ∂ ∂ νt ∂k
+ (v̄j k) = ν+ + P k − β ∗ kω
∂t ∂xj ∂xj σk ∂xj
  
∂ω ∂ ∂ νt ∂ω Pk
+ (v̄j ω) = ν+ +α − βω 2
∂t ∂xj ∂xj σω ∂xj νt
1 ∂k ∂ω
+ 2(1 − F1 )σω2
ω ∂xi ∂xi
" ( √ ) #
4 k 500ν 4σωk−ε k
F1 = tanh(ξ ), ξ = min max , (20.4)
β ∗ ωd d2 ω CDω d2
 
1 ∂k ∂ω −10
CDω = max 2σωk−ε , 10
ω ∂xi ∂xi
a1 k
νt =
max(a1 ω, |s̄|F2 )
 1/2 
2 2k 500ν
F2 = tanh(η ), η = max ,
β ∗ ωd d2 ω
where d is the distance to the closest wall node. The SST model behaves as a k − ω
model near the wall where F1 = 1 and a k − ε model far from walls (F1 = 0). All
coefficients are blended between the k − ω and the k − ε model using the function F1 ,
for α, for example,
α = F1 αk−ω + (1 − F1 )αk−ε (20.5)
The constants take the following values:
β ∗ = 0.09, a1 = 0.3
αk−ω = 5/9, βk−ω = 3/40, σk,k−ω = 0.85, σω,k−ω = 0.5 (20.6)
αk−ε = 0.44, βk−ε = 0.0828, σk,k−ε = 1, σω,k−ε = 0.856.
In DES the dissipation term in the k equation is modified as [68]
 
∗ ∗ Lt
β kω → β kωFDES , FDES = max ,1
CDES ∆
(20.7)
k 1/2
∆ = max {∆x1 , ∆x2 , ∆x3 } , Lt = ∗
β ω
where CDES = 0.61.
Again, the DES modification is meant to switch the turbulent length scale from a
RANS length scale (∝ k 1/2 /ω) to a LES length scale (∝ ∆) when the grid is suf-
ficiently fine. When FDES is larger than one, the dissipation term in the k equation
20.3. DDES 201

increases which in turn decreases k and thereby also the turbulent viscosity. With a
smaller turbulent viscosity in the momentum equations, the modeled dissipation (i.e
the damping) is reduced and the flow is induced to go unsteady. The result is, hope-
fully, that a large part of the turbulence is resolved rather than being modeled.
Equation 20.7 shows that it is the grid that determines the location where the model
switches between RANS and LES. Hence it is crucial to generate an appropriate grid.
The larger the maximum cell size (usually ∆x or ∆z) is made, the further out from the
wall does the switch take place.

20.3 DDES
In some flows it may occur that the FDES term switches to LES in the boundary layer
because ∆x or ∆z (∆x1 or ∆x3 ) are too small (smaller than the boundary layer thick-
ness, δ). This means that the flow in the boundary layer is treated in LES mode with too
a coarse mesh. This results in a poorly resolved LES and hence inaccurate predictions.
Different proposals have been made [142, 143] to protect the boundary layer from the
LES mode  
Lt
FDES = max (1 − FS ), 1 (20.8)
CDES ∆
where FS is taken as F1 or F2 (see Eq. 20.4) of the SST model. In [68] F = F2 . This
is called DDES (Delayed DES).
21. Hybrid LES-RANS 202

21 Hybrid LES-RANS
HEN simulating bluff body flows, LES (Large Eddy Simulation) is the ideal
W method. Bluff body flows are dominated by large turbulent scales that can be
resolved by LES without too fine a resolution and accurate results can thus be obtained
at an affordable cost. On the other hand, it is a challenging task to make accurate pre-
dictions of wall-bounded flows with LES. The near-wall grid spacing should be about
one wall unit in the wall-normal direction. This is similar to the requirement in RANS
using low-Re number models. The resolution requirements in wall-parallel planes for
a well-resolved LES in the near-wall region expressed in wall units are approximately
100 (streamwise) and 30 (spanwise). This enables resolution of the near-wall turbu-
lent structures in the viscous sub-layer and the buffer layer consisting of high-speed
in-rushes and low-speed ejections [116], often called the streak process.
An event of a high-speed in-rush is illustrated in Fig. 21.1. In the lower part of
the figure the spanwise vortex line is shown. Initially it is a straight line, but due to a
disturbance – e.g. a turbulent fluctuation – the mid-part of the vortex line is somewhat
lifted up away from the wall. The mid-part of the vortex line experiences now a higher
v̄1 velocity (denoted by U in the figure) than the remaining part of the vortex line.
As a result the mid-part is lifted up even more and a tip of a hairpin vortex is formed.
The vorticity of the legs of the hairpin lift each other through self-induction which helps
lifting the tip even more. In the x1 −x2 plane (upper part of Fig. 21.1) the instantaneous
and mean velocity profiles (denoted by U and Ū in the figure, respectively) are shown
as the hairpin vortex is created. It can be seen that an inflection point is created in the
instantaneous velocity profile, U , and the momentum deficit in the inner layer increases
for increasing x1 . Eventually the momentum deficit becomes too large and the high-
speed fluid rushes in compensating for the momentum deficit. The in-rush event is also
called a sweep. There are also events which occurs in the other direction, i.e. low-
speed fluid is ejected away from the wall. These events are called bursts or ejections.
The spanwise separation between sweeps and bursts is very small (approximately 100
viscous units, see Fig. 21.1). This is the main reason why the grid must be very fine
in the spanwise direction. The streamwise distance between the events is related to
the boundary layer thickness (4δ, see Fig. 21.1). The process by which the events are
formed is similar to the later stage in the transition process from laminar to turbulent
flow. Figure 21.2 presents the instantaneous field of the streamwise velocity fluctuation,
v1′ in the viscous wall region. As can be seen, the turbulent structures very elongated
in the streamwise direction.
At low to medium Reynolds numbers the streak process is responsible for the major
part of the turbulence production. These structures must be resolved in an LES in order
to achieve accurate results. Thus, for wall-bounded flows at high Reynolds numbers
of engineering interest, the computational resource requirement of accurate LES is
prohibitively large. Indeed, the requirement of near-wall grid resolution is the main
reason why LES is too expensive for engineering flows, which was one of the lessons
learned in the LESFOIL project [120, 121].
The object of hybrid LES-RANS (and of DES) is to eliminate the requirement of
high near-wall resolution in wall-parallel planes. In the near-wall region (the URANS
region), a low-Re number RANS turbulence model (usually an eddy-viscosity model)
is used. In the outer region (the LES region), the usual LES is used, see Fig. 21.3.
The idea is that the effect of the near-wall turbulent structures should be modeled by
the RANS turbulence model rather than being resolved. In the LES region, coarser
21. Hybrid LES-RANS 203

Figure 21.1: Illustration of near-wall turbulence (taken from [73]).

x3

x1

Figure 21.2: Fluctuating streamwise velocity in a wall-parallel plane at x+

2 = 5. DNS
of channel flow [89].
wall

URANS region

LES region

x2 URANS region
x+
2,ml

wall
x1
Figure 21.3: The LES and URANS region.
21.1. Momentum equations in hybrid LES-RANS 204

URANS region LES region

−3/4
ℓ κcµ n[1 − exp(−0.2k 1/2 n/ν)] ℓ=∆
1/4
νT κcµ k 1/2 n[1 − exp(−0.014k 1/2n/ν)] 0.07k 1/2 ℓ
Cε 1.0 1.05

Table 21.1: Turbulent viscosity and turbulent length scales in the URANS and LES
regions. n and κ denote the distance to the nearest wall and von Kármán constant
(= 0.41), respectively. ∆ = (δV )1/3
.

grid spacing in wall-parallel planes can be used. The grid resolution in this region is
presumably dictated by the requirement of resolving the largest turbulent scales in the
flow (which are related to the outer length scales, e.g. the boundary layer thickness)
rather than the near-wall turbulent processes. The unsteady momentum equations are
solved throughout the computational domain. The turbulent RANS viscosity is used in
the URANS region, and the turbulent SGS viscosity is used in the LES region.
Hybrid LES-RANS is similar to DES (Detached Eddy Simulations) [127,143,144].
The main difference is that the original DES aims at covering the whole attached
boundary layer with URANS, whereas hybrid LES-RANS aims at covering only the
inner part of the boundary layer with URANS. In later work DES has been used as a
wall model [139, 145] – called wall-modelled LES – and, in this form, DES is similar
hybrid LES-RANS.

21.1 Momentum equations in hybrid LES-RANS

The incompressible Navier-Stokes equations with an added turbulent/SGS viscosity
read  
∂v̄i ∂ 1 ∂ p̄ ∂ ∂v̄i
+ (v̄i v̄j ) = − + (ν + νT ) (21.1)
∂t ∂xj ρ ∂xi ∂xj ∂xj
where νT = νt (νt denotes the turbulent RANS viscosity) for x2 ≤ x2,ml (the URANS
region, see Fig. 21.3) and, for x2 > x2,ml (the LES region), νT = νsgs .

21.2 The one-equation hybrid LES-RANS model

When a one-equation model is employed in both the URANS region and the LES
region, it reads
  3/2
∂kT ∂ ∂ ∂kT k
+ (v̄j kT ) = (ν + νT ) + PkT − Cε T
∂t ∂xj ∂xj ∂xj ℓ (21.2)
PkT = −τij s̄ij , τij = −2νT s̄ij
In the inner region (x2 ≤ x2,ml ) kT corresponds to the RANS turbulent kinetic energy,
k; in the outer region (x2 > x2,ml ) it corresponds to the subgrid-scale kinetic turbulent
energy (ksgs ). No special treatment is used in the equations at the matching plane ex-
cept that the form of the turbulent viscosity and the turbulent length scale are different
in the two regions, see Table 21.1. At the walls, kT = 0.
When prescribing the location of the RANS-LES interface, the velocity profile may
show unphysical behaviour near the interface because of the rapid variation of the tur-
bulence viscosity. Much work on forcing have been presented in order to alleviate this
problem [89, 117, 139, 140].
22. The SAS model 205

22 The SAS model

22.1 Resolved motions in unsteady
HEN doing URANS or DES, the momentum equations are triggered through in-
W stabilities to go unsteady in regions where the grid is fine enough. In URANS
or in DES operating in RANS mode, high turbulent viscosity often dampens out these
instabilities. In many cases this is an undesired feature, because if the flow wants to go
unsteady, it is usually a bad idea to force the equations to stay steady. One reason is that
there may not be any steady solution. Hence, the equations will not converge. Another
reason is that if the numerical solution wants to go unsteady, the large turbulent scales
— i.e. part of the turbulent spectrum — will be resolved instead of being modeled.
This leads to a more accurate prediction of the flow.
One way to improve a RANS model’s ability to resolve large-scale motions is to
use the SAS (Scale- Adaptive Simulation) model

22.2 The von Kármán length scale

The von Kármán length scale
 
 ∂hv̄i/∂x2 
LvK,1D 
= κ 2  (22.1)
∂ hv̄i/∂x2  2

which includes the second velocity gradient is a suitable length scale for detecting
unsteadiness. The von Kármán length scale is smaller for an instantaneous velocity
profile than for a time averaged velocity, see Fig. 22.1. This is interesting because, as
noted in [146], the von Kármán length scale decreases when the momentum equations
resolve (part of) the turbulence spectrum.
The first and second derivatives in Eq. 22.1 are given in boundary layer form. We
want to extend this expression to a general one, applicable in three dimensions. In the
same way as in, for example, the Smagorinsky model, we take the first derivative as
|s̄| = (2s̄ij s̄ij )1/2 . The second derivative can be generalized in a number of ways. In
the SAS model it is taken as
 2 0.5
∂ v̄i ∂ 2 v̄i
U ′′ = (22.2)
∂xj ∂xj ∂xk ∂xk

There are other options how to computed this second derivative, see Eq. O.5 at p. 327.
Hence, the general three-dimensional expression for the von Kármán length scale reads

|s̄|
LvK,3D = κ (22.3)
|U ′′ |

In [147] they derived a one-equation νt turbulence model where the von Kármán
length scale was used. The model was called the SAS model. Later, based on the
k − k 1/2 L model of Rotta [148], Menter & Egorov [146] derived a new k − kL model
using the von Kármán length scale. Finally, in [149] they modified the k − ω-SST
model to include the SAS features; they called this model the SST-SAS model. This
model is described in more detail below.
22.2. The von Kármán length scale 206

0.8

0.6
x2 LvK,3D
LvK,1D
0.4

0.2

0
0 5 10 15 20
v̄1

Figure 22.1: Velocity profiles from a DNS of channel flow. Solid line: time-averaged
velocity with length scale Lx,1D , Eq. 22.1 ; dashed line: instantaneous velocity with
length scale LvK,3D , Eq. 22.3.

The SST-SAS model

The k − ω SST model is given in Eq. 20.4 at p. 200 (see also the section starting at
p. 164) Now, Menter & Egorov [149] introduced a SAS-term in the ω equation. The
object of this term is to decrease the turbulent viscosity when unsteadiness is detected,
i.e. when the von Kármán length scale becomes small. The production term in the ω
equation in the k−ω-SST model reads Pω = αP k /νt ∝ |s̄|2 . To decrease the turbulent
viscosity we should increase ω. Thus it seems reasonable to add a new production term
proportional to Pω Lt /LvK,3D where Lt denotes a RANS length scale. The additional
term reads
Lt k 1/2
ζ̃2 κ|s̄|2 , Lt = 1/4 (22.4)
LvK,3D ωcµ
When unsteadiness occurs — i.e. when the momentum equations attempt to resolve
part of the turbulence spectrum — this term reacts as follows:
• Local unsteadiness will create velocity gradients which decrease the turbulent
length scale, see Fig. 22.1
• This results in a decrease in the von Kármán length scale, LvK,3D
• As a consequence the additional source, Eq. 22.4, in the ω equation increases
• This gives an increase in ω and hence a decrease in νt
• The decreased turbulent viscosity will allow the unsteadiness to stay alive and,
perhaps, grow.
The last item in the list above is the main object of the SAS model. The reaction
to local unsteadiness in a eddy-viscosity model without the SAS feature is as follows:
the increased local velocity gradients will create additional production of turbulent ki-
netic energy and give an increased turbulent viscosity which will dampen/kill the local
unsteadiness. As mentioned in the introduction to this chapter, this is an undesirable
feature.
22.3. The second derivative of the velocity 207

When incorporating the additional production term (Eq. 22.4) in the k − ω-SST
model, the last term in the ω equation is replaced by (for further details, see [149])

PSAS = FSAS max (T1 − T2 , 0)

L
T1 = ζ̃2 κS 2
LvK,3D
 
2k 1 ∂ω ∂ω 1 ∂k ∂k (22.5)
T2 = max ,
σΦ ω 2 ∂xj ∂xj k 2 ∂xj ∂xj
k 1/2
L= 1/4
ωcµ

Note that the term T1 is the “real” additional SAS term; T2 is included to make sure
that the model in steady flow works as a k − ω SST model.

22.3 The second derivative of the velocity

To compute U ′′ in Eq. 22.2, we need to compute the second velocity gradients. In finite
volume methods there are two main options for computing second derivatives.
Option I: compute the first derivatives at the faces
   
∂v vj+1 − vj ∂v vj − vj−1
= , =
∂x2 j+1/2 ∆x2 ∂x2 j−1/2 ∆x2
and then  
∂2v vj+1 − 2vj + vj−1 (∆x2 )2 ∂ 4 v
⇒ = +
∂x22 j (∆x2 )2 12 ∂x42
Option II: compute the first derivatives at the center
   
∂v vj+2 − vj ∂v vj − vj−2
= , =
∂x2 j+1 2∆x2 ∂x2 j−1 2∆x2
and then  
∂2v vj+2 − 2vj + vj−2 (∆x2 )2 ∂ 4 v
⇒ = +
∂x22 j 4(∆x2 )2 3 ∂x42
In [150], Option I was used unless otherwise stated.

22.4 Evaluation of the von Kármán length scale in channel flow

In Fig. 22.2 the turbulent length scale, hLvK,3D i, is evaluated using DNS data of fully
developed channel flow. When using DNS data only viscous dissipation of resolved tur-
bulence affects the equations. This implies that the smallest scales that can be resolved
are related to the grid scale. The von Kármán length scale based on instantaneous ve-
locities, hLvK,3D i, is presented in Fig. 22.2. For x2 > 0.2, its magnitude is close to
∆x2 which confirms that the von Kármán length scale is related to the smallest resolv-
able scales. Closer to the wall, hLvK,3D i increases slightly whereas ∆x2 continues to
decrease.
The von Kármán length scale, LvK,1D , based on the averaged velocity profile
hv̄1 i = hv̄1 i(x2 ) is also included in Fig. 22.2, and as can be seen it is much larger than
hLvK,3D i. Near the wall LvK,1D increases because the time-average second derivative,
 22.4. Evaluation of the von Kármán length scale in channel flow 208

x+
2 x+
2
0 100 200 300 400 500 0 5 10 15 20 25
0.1

0.3
0.08
0.25

0.06
0.2

0.15
0.04

0.1
0.02
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.01 0.02 0.03 0.04 0.05
x2 /δ x2 /δ
Figure 22.2: Turbulent length scales in fully developed channel flow. Left: global
view; right: zoom. DNS. 963 mesh. Reτ = 500. ∆x1 /δ = 0.065, ∆x3 /δ = 0.016,
x2 -stretching of 9%. : hLvK,3D i; : LvK,1D ; : (∆x1 ∆x2 ∆x3 )1/3 ; ◦:
∆x2 .
x+2 x+
2
0 500 1500 2000 0 20 40 60 80 100
0.1

0.25
0.08

0.2
0.06
0.15
0.04
0.1

0.02
0.05

0 0
0 0.2 0.4 0.6 0.8 1 0 0.01 0.02 0.03 0.04 0.05
x2 /δ x2 /δ
Figure 22.3: Turbulent length scales in fully developed channel flow. Hybrid LES-
RANS. Left: global view; right: zoom. 32 × 64 × 32 mesh. Reτ = 2000. ∆x1 /δ =
0.39, ∆x3 /δ = 0.19, x2 -stretching of 17%. : hLvK,3D i; : LvK,1D ; :
1/3 0.5 1/4
(∆x1 ∆x2 ∆x3 ) ; ◦: ∆x2 ; +: ℓk−ω = k /(cµ ω).

∂ 2 hv̄1 i/∂x22 , goes to zero as the wall is approached. No such behavior is seen for the
three-dimensional formulation, hLvK,3D i.
In Fig. 22.3, data from hybrid LES-RANS are used (taken from [89]). When using
hybrid LES-RANS, part of the turbulence is resolved and part of the turbulence is
modeled. The resolved turbulence is dissipated by a modeled dissipation, −2hνT s̄ij s̄ij i
(νT denotes SGS or RANS turbulent viscosity), and νT ≫ ν. As a result, the length
scale of the smallest resolved turbulence is larger in hybrid LES-RANS than in DNS.
Close to the wall in the URANS region (x2 < 0.031δ), the resolved turbulence is
dampened by the high turbulent viscosity, and as a results hLvK,3D i follows closely
LvK,1D .
The RANS turbulent length scale, ℓk−ω , from a 1D RANS simulation at Reτ =
22.4. Evaluation of the von Kármán length scale in channel flow 209

2000 with the k − ω SST model is also included in Fig. 22.3. In the inner region
(x2 < 0.5δ), its behavior is close to that of the von Kármán length scale, LvK,1D . In the
center region the RANS turbulent length scale continues to increase which is physically
correct. However, the von Kármán length scale, LvK,1D , goes to zero because the
velocity derivative goes to zero.
Two filter scales are included in Figs. 22.2 and 22.3. In the DNS-simulations,
∆x2 < (∆x1 ∆x2 ∆x3 )1/3 near the wall, whereas far from the wall ∆x2 > (∆x1 ∆x2 ∆x3 )1/3
because of the stretching in the x2 direction and because of small ∆x1 and ∆x3 . In the
hybrid simulations, it can be noted that the three-dimensional filter width is more that
twice as large as the three-dimensional formulation of the von Kármán length scale,
i.e. (∆x1 ∆x3 ∆x3 )1/3 > 2hLvK,3D i.
In [150], the SST-SAS model has been evaluated in channel flow, flow in an asym-
metric diffusor and flow over an axi-symmetric hill.
23. The PANS Model 210

23 The PANS Model

HE PANS method uses the so-called “partial averaging” concept, which corresponds
T to a filtering operation for a portion of the fluctuating scales [151].
For an instantaneous flow variable, F , we use f¯ to denote the partially-averaged
part, namely f¯ = P(F ), where P denotes the partial-averaging operator. We consider
incompressible flows. Applying the partial averaging to the governing equations gives
∂v̄i
=0 (23.1)
∂xi
 
∂v̄i ∂(v̄i v̄j ) 1 ∂ p̄ ∂ ∂v̄i
+ =− + ν − τij (23.2)
∂t ∂xj ρ ∂xi ∂xj ∂xj
where τij is the central second moment resulting from the partial averaging for the
nonlinear terms, that is τij = P(vi vj ) − v̄i v̄j , where vi indicates instantaneous veloc-
ity components. This term is similar to the Reynolds stress tensor resulting from the
Reynolds averaging in RANS or to the subgrid-scale (SGS) stress tensor after the spa-
tial filtering in LES. For simplicity, we also use the terminology of Reynolds stresses
for the term τij in Eq. 23.2.
To close the system of the partially-averaged Navier-Stokes equations, as in RANS
and LES, a model is needed for τij . In [151] they proposed using the conventional
eddy viscosity concept so that τij = −2νu s̄ij , where s̄ij is the strain-rate tensor of the
computed flow and νu is the PANS eddy viscosity.
In order to formulate the PANS eddy viscosity, they defined in [151] another two
quantities, the partially-averaged turbulent kinetic energy, ku and its dissipation rate εu ,
so that νu = Cµ ku2 /εu . In the derivation of the transport equations for ku and εu , two
parameters, fk and fε , have been introduced, relating the unresolved to the resolved
fluctuating scales. Parameter fk defines the ratio of unresolved (partially-averaged)
turbulent kinetic energy (ku ) to the total kinetic energy (k), and fε is the ratio between
the unresolved (εu ) and the total (ε) dissipation rates. These give
ku εu
k= and ε = (23.3)
fk fε
The extent of the resolved part is now determined by fk and fε . Usually fε = 1; fε < 1
implies that dissipative scales are resolved. In [151, 152] they employed the standard
k − ε model as the base model.
The ku equation is derived by multiplying the RANS k equation (Eq. 11.97) in the
k − ε model by fk , i.e. (for simplicity we omit the buoyancy term)
     
∂k ∂k ∂ νt ∂k
fk + V̄j = fk P k − ε + ν+ (23.4)
∂t ∂xj ∂xj σk ∂xj
where Vi denotes the RANS velocity. The left side can be re-written
 
∂k ∂k ∂ku ∂ku ∂ku ∂ku ∂ku
fk + V̄j = + V̄j = + v̄j + (V̄j − v̄j ) (23.5)
∂t ∂xj ∂t ∂xj ∂t ∂xj ∂xj
Note that we have assumed that fk is constant. The convective term must be expressed
in v̄j (the PANS averaged velocity) rather than in V̄j (the RANS averaged velocity),
because it is v̄j that transports ku because v̄j represents the PANS resolved part of vj .
Anyway, the last term on the right side in Eq. 23.5 is usually neglected.
23. The PANS Model 211

The diffusion term is re-written using Eq. 23.3

      
∂ νt ∂k ∂ νt ∂ku
fk ν+ = ν+
∂xj σk ∂xj ∂xj σk ∂xj
   (23.6)
∂ νu ∂ku
= ν+
∂xj σku ∂xj

where
fk2
σku = σk (23.7)
fε
and
ku2
νu = c µ (23.8)
εu
The source terms in Eq. 23.4 are replaced by Pu and εu , i.e.


fk P k − ε = Pu − εu (23.9)

This relation implies

1 εu
Pk = (Pu − εu ) + (23.10)
fk fε
Using Eqs. 23.5, 23.6 and 23.9 the final transport equation for ku can now be written
as
  
∂ku ∂(ku v̄j ) ∂ νu ∂ku
+ = ν+ + Pu − εu (23.11)
∂t ∂xj ∂xj σku ∂xj

where the production term, Pu , is expressed in terms of the PANS eddy viscosity, νu ,
and the strain rate of PANS-resolved flow field, i.e.
 
∂v̄i ∂v̄j ∂v̄i
Pu = νu + (23.12)
∂xj ∂xi ∂xj

The εu equation is derived by multiplying the RANS ε equation by fε , i.e.

 
∂εu ∂(εu v̄j ) ∂ε ∂(εV̄j )
+ = fε +
∂t ∂xj ∂t ∂xj
     (23.13)
∂ νt ∂ε ε ε2
= fε ν+ + Cε1 P k − Cε2
∂xj σε ∂xj k k

where fε is assumed to be constant. The diffusion term is re-written using Eq. 23.3
      
∂ νt ∂ε ∂ νt ∂εu
fε ν+ = ν+
∂xj σε ∂xj ∂xj σε ∂xj
   (23.14)
∂ νu ∂εu
= ν+
∂xj σεu ∂xj

where
fk2
σεu = σε (23.15)
fε
23. The PANS Model 212

In the same way, the production and destruction terms are re-formulated as (using
Eqs. 23.3 and 23.10)
   
ε ε2 εu fk 1 εu ε2 fk
fε Cε1 P k − Cε2 = Cε1 (Pu − εu ) + − Cε2 u
k k ku fk fε fε ku
2 2 2
εu ε ε fk ε fk
= Cε1 Pu − Cε1 u + Cε1 u − Cε2 u (23.16)
ku ku ku fε fε ku
2
εu ∗ εu
= Cε1 Pu − Cε2
ku ku
where
∗ fk
Cε2 = Cε1 + (Cε2 − Cε1 ) (23.17)
fε
The εu equation in the PANS model now takes the following form
   2
∂εu ∂(εu v̄j ) ∂ νu ∂εu εu ∗ εu
+ = ν+ + Cε1 Pu − Cε2 (23.18)
∂t ∂xj ∂xj σεu ∂xj ku ku

As in the ku equation, the the additional term (V̄j − v̄j )∂εu /∂xj has been neglected.
The PANS equation for ku , Eq. 23.11, was derived by multiplying the RANS equa-
tion for k by fk which was assumed to be constant in space and in time. By referring
to Eqs. 23.6, 23.8 and 23.7, the turbulent diffusion term was obtained as
   
∂ νt ∂k ∂ νt ∂ku
fk = (23.19a)
∂xj σk ∂xj ∂xj σk ∂xj
 
∂ νu ∂ku
= (23.19b)
∂xj σku ∂xj

The expression on the right-hand side of Eq. 23.19(a) suggests that the turbulent trans-
port for the PANS-modeled turbulent kinetic energy, ku , is actually formulated in terms
of the RANS turbulent viscosity from the base model. This is different from the turbu-
lent diffusion in subgrid scale (SGS) modeling of LES with a one-equation ksgs model,
which reads  
∂ νsgs ∂ksgs
(23.20)
∂xj σk ∂xj
In Eq. 23.20 the SGS turbulent viscosity is invoked for the transport of ksgs , whereas
on the right-hand side of Eq. 23.19(a) the total (i.e. the RANS) turbulent viscosity has
been used for ku . Equation 23.19(a) suggests that, when used as an SGS model, the
modeled turbulent diffusion in the PANS formulation is a factor of σk /σku = fε /fk2
larger than in Eq. 23.20, see Eqs. 23.11 and 23.19(b). With fε = 1 and fk = 0.4, for
example, this factor is larger than six. The modification of the diffusion coefficient,
σku , is a unique property of the PANS model. In other models, such as DES [153],
X-LES [131] and PITM [135], the sink term in the k, ε or ω equation is modified, but
not the diffusion term. The only difference between PANS and PITM is that in the
former model the diffusion coefficients in the k and ε are modified.
23.1. PANS as a hybrid LES-RANS model 213

A Low Reynolds number PANS model was presented in [134, 154] which reads
  
∂ku ∂(ku v̄j ) ∂ νu ∂ku
+ = ν+ + (Pu − εu )
∂t ∂xj ∂xj σku ∂xj
   2
∂εu ∂(εu v̄j ) ∂ νu ∂εu εu ∗ εu
+ = ν+ + Cε1 Pu − Cε2
∂t ∂xj ∂xj σεu ∂xj ku ku
k2 ∗ fk (23.21)
νu = Cµ fµ u , Cε2 = Cε1 + (Cε2 f2 − Cε1 )
εu fε

fk2 f2
σku ≡ σk , σεu ≡ σε k
fε fε

The modifications introduced by the PANS modeling as compared to its parent

RANS model are highlighted by boxes. The model constants take the same values as
in the LRN model [136], i.e.

Cε1 = 1.5, Cε2 = 1.9, σk = 1.4, σε = 1.4, Cµ = 0.09 (23.22)

23.1 PANS as a hybrid LES-RANS model

A new approach to use the partially averaged Navier-Stokes (PANS) model as a hybrid
RANS-LES model was presented by Davidson [155]. It was evaluated in fully devel-
oped channel flow and embedded LES in a hump flow. For the channel flow, the two
RANS-LES interfaces are parallel to the walls. In the URANS region, fk is set to one.
In the LES region, fk is set to a constant value (the baseline value is fk = 0.4) or it
is computed. It is found that the new model gives good results for channel flow for a
large span of Reynolds numbers (4 000 ≤ Reτ ≤ 32 000). In the channel flow sim-
ulations, three different grids are used in the wall-parallel planes, 322 , 642 and 1282 ,
and the model yields virtually grid-independent flow fields and turbulent viscosities.
Embedded LES is used for the hump flow which is well predicted. The RANS-LES
interface is normal to the flow from the inlet. RANS is used upstream of the interface.
Downstream this interface, RANS is used near the wall and LES is used away from the
wall.

23.1.1 The interface conditions at the RANS-LES interface

The interface plane separates the URANS region near the wall and the LES region in
the outer region. In the former region, the turbulent viscosity, νt , should be a RANS

kint , εint LES, fk < 1

URANS, fk = 1.0 yint

y
wall

x
Figure 23.1: The URANS and the LES regions.
23.1. PANS as a hybrid LES-RANS model 214

viscosity and in the latter region it should be an SGS viscosity. Hence νt must decrease
rapidly when going from the URANS region to the LES region. This is achieved by
setting the usual convection and diffusion fluxes of k at the interface to zero. New
fluxes are introduced in which the interface condition is set to kint = fk kRAN S , where
kRAN S is the k value in the cell located in the URANS region adjacent to the interface.
Unless otherwise stated, no modification is made for the convection and diffusion of ε
across the interface. The implementation is presented in some detail below. We write
the discretized equation in the y direction (see Figs. 23.1 and 23.2) as [88]

aP kP = aN kN + aS kS + SU , a P = a S + a N − SP

where aS and aN are related to the convection and diffusion through the south and
north face, respectively, and SU and SP kP include the production and the dissipation
term, respectively. For a cell in the LES region adjacent to the interface (cell P ), aS
is set to zero, setting the usual convection and diffusion fluxes to zero. New fluxes,
including fk , are incorporated in additional source terms as

SU = (Cs + Ds )fk kS , SP = −(Cs + Ds )

µtot As (23.23)
Cs = max (v̄s As , 0) , Ds =
∆y
where Cs and Ds denote convection (first-order upwind) and diffusion, respectively,
through the south face, and As is the south area of the cell. As can be seen, the kS
is multiplied by fk and hence the new convective flux is a factor fk smaller than the
original one. Also the diffusion flux is smaller; it is Ds (fk kS − kP ) compared with the
original flux Ds (kS − kP ).
The interface is defined along gridlines. The approach presented above is also
applicable when the location is automatically computed where the extent of the RANS
region varies along the wall. The convective and diffusive fluxes are modified in exactly
the same way.
The method of adding a new flux in the k equation is similar to the method proposed
in [156]. In that work they take into account the spatial variation of fk . A decrease of
fk in space – as occurs at the interface in Fig. 23.1 – means that the turbulent kinetic
energy should be transferred from modeled to resolved. This is done by introducing

P
y

x Interface ∆y
As

Figure 23.2: Control volume, P , in the LES region adjacent to the interface.
23.2. Zonal PANS: different treatments of the RANS-LES interface 215

a sink term in the k equation which decreases the turbulent viscosity. [156] also add
a turbulent diffusion term in the momentum equation which appears as a source term
in the equation for turbulent resolved kinetic energy. No source term is used in the
momentum equations in this method. A further difference is that this method modifies
the equations only in the LES region (in which k is reduced), not in the RANS region.
The reason is that this effect is believed to be the most critical; it is important that the
resolved turbulence on the LES side of the interface is activated as soon as possible
(i.e. as close as possible to the interface). The total turbulent kinetic energy across the
interface is not conserved, but as shown below, the total turbulent kinetic energy that is
lost across the interface is not large. For more details and results, see Davidson [155].

23.2 Zonal PANS: different treatments of the RANS-LES interface

In [156], a method was proposed to include the effect of the gradient of fk in PANS.
This approach was used at RANS-LES interfaces by Davidson [137]. Four different
interface models are evaluated in fully developed channel flow and embedded LES of
channel flow; in both cases, PANS is used as a zonal model with fk = 1 in the URANS
region and fk = 0.4 in the LES region. In fully developed channel flow, the RANS-
LES interface is parallel to the wall (horizontal) and in embedded LES it is parallel to
the inlet (vertical).
The importance of the location of the horizontal interface in fully developed chan-
nel flow is also investigated. It is found that the location – and the choice of the treat-
ment at the interface – may be critical at low Reynolds number or if the interface is
placed too close to the wall. The reason is that the modeled turbulent shear stress at
the interface is large and hence the relative strength of the resolved turbulence is small.
In RANS, the turbulent viscosity – and consequently also the modeled Reynolds shear
stress – is only weakly dependent on Reynolds number. It is found that that also applies
in the URANS region.

23.2.1 The Interface Condition

The commutation error in PANS was recently addressed in [156]. In PANS, the equa-
tion for the modeled turbulent kinetic energy, k, is derived by multiplying the ktot
equation (ktot = kres + k) by fk where kres denotes the resolved turbulent kinetic
energy. The convective term in the k equation with constant fk is then obtained as
dktot d(fk ktot ) dk
fk = = (23.24)
dt dt dt
where
k
fk = . (23.25)
ktot
Now, if fk varies in space, we get instead
dktot d(fk ktot ) dfk dk dfk
fk = − ktot = − ktot (23.26)
dt dt dt dt dt
Since fk here is constant in time, dfk /dt = v̄i ∂fk /∂xi . The second term on the right
side of Eq. 23.26 is the commutation term; it represents (excluding the minus sign)
energy transfer from resolved to modeled turbulence. It can be written (on the right
side of the k equation)
dfk dfk dfk hv̄ ′ v̄ ′ i dfk
ktot = (k + kres ) =k + i i (23.27)
dt dt dt 2 dt
23.2. Zonal PANS: different treatments of the RANS-LES interface 216

LES, fk = 0.4

URANS, fk = 1 yint
y
wall

x
Figure 23.3: Fully developed channel flow. The URANS and the LES regions.

xI xtr

0.4 < fk < 1

v̄ ′ , v̄ ′ , w̄′
2δ
RANS LES
y fk = 1 fk = 0.4

L
x

Figure 23.4: Embedded LES. Vertical thick line shows the interface at xI = 0.95. fk
varies linearly in the gray area (width xtr ) from 1 to 0.4. δ = 1.

The commutation term in the kres equation is the same but with opposite sign, i.e.
dfk hv̄ ′ v̄ ′ i dfk
−k − i i (23.28)
dt 2 dt
The question is now which term should be added to the momentum equations to get
the commutation term in the kres equation. We start by the second term in Eq. 23.28.
This term can be represented by the source term
1 dfk
Si1 = − v̄i′ (23.29)
2 dt
in the momentum equation. To show that this term corresponds to the commutation
term in Eq. 23.28, consider the momentum equation for the fluctuating velocity, v̄i′ .
Multiplying Si1 by v̄i′ = v̄i − hv̄i i and time-averaging gives the source term in the kres
equation as
   
1 dfk 1 ′ ′ dfk 1 dfk
− (v̄i − hv̄i i) v̄i′ =− v̄i v̄i = − hv̄i′ v̄i′ i (23.30)
2 dt 2 dt 2 dt
Equation 23.30 is equal to the time average of the second term in Eq. 23.28 as it should.
A term corresponding to the first term in Eq. 23.28 can be added as a source term
in the momentum equation as
kv̄i′ dfk
Si2 = − ′ v̄ ′ i dt
(23.31)
hv̄m m
23.2. Zonal PANS: different treatments of the RANS-LES interface 217

Multiplying Si2 by v̄i′ and time averaging gives

hkv̄i′ v̄i′ i dfk dfk

− ′ v̄ ′ i dt
≃ −hki (23.32)
hv̄m m dt

where we assume that the correlation between k and v̄i′ is weak. Equation 23.32 is
equal to the time average of the first term in Eq. 23.28 as it should.
In [156], the second term on the right side of Eq. 23.26 is represented by introducing
an additional turbulent viscosity, νtr , in a diffusion term in the momentum equation as
 
∂ 1 ∂v̄i ∂v̄j
(νtr s̄ij ) , s̄ij = + (23.33)
∂xj 2 ∂xj ∂xi

where
Pktr
νtr = . (23.34)
|s̄|2

The term, Pktr , is computed as

dfk
Pktr = ktot (23.35)
dt
Pktr is an additional production term in the k equation, see Eq. 23.21. In [156], Pktr is
re-written using Eq. 23.25 as
k dfk
Pktr = (23.36)
fk dt

23.2.2 Modeling the Interface

The gradient of fk across a RANS-LES interface gives rise to an additional term in
the momentum equations and the k equation. These terms are included only when the
flow goes from a RANS region to an LES region. The effect of these terms will reduce
k and act as a forcing term in the momentum equations. Four different models are
investigated.

23.2.2.1 Interface Model 1

This is based on the approach suggested in [156]. The additional turbulent viscosity,
νtr , gives an additional production term, Pktr , in the k equation, see Eq. 23.36. Since
we are interested in stimulating resolved turbulence in the LES region adjacent to the
RANS region, only negative values of νtr are included. A negative νtr means physi-
cally transfer of kinetic energy from modeled to resolved. It is found that the magnitude
of the positive values of νtr is actually larger than the magnitude of the negative ones,
which means that hνtr i > 0. The negative values correspond to dfk /dt < 0 (see
Eqs. 23.34, 23.35 and 23.36), i.e. when a fluid particle in a RANS region passes the
interface into an adjacent LES region. However, νtr takes such large (negative) values
that νt + νtr < 0. To stabilize the simulations, it was found necessary to introduce a
limit νt + νtr > 0 in the diffusion term in the momentum equation. No such limit is
used in the k equation, and hence Pk + Pktr is allowed to go negative.
23.2. Zonal PANS: different treatments of the RANS-LES interface 218

23.2.2.2 Interface Model 2

This model is identical to Model 1 except that Eq. 23.35 is used instead of Eq. 23.36.
ktot in Eq. 23.35 is defined as
1
ktot = k + hv̄i′ v̄i′ ir.a (23.37)
2
where subscript r.a. denotes running average. Is is averaged in all homogeneous direc-
tions including time (note that time is not a homogeneous direction in decaying grid
turbulence). Since ktot is mostly larger than k/fk [155], this approach will give a
larger magnitude of the (negative) production than Model 1. It is found in [137] that
this modification is of utmost importance.

23.2.2.3 Interface Model 3

The right side of Eq. 23.27 is added to the k equation, but only when the flow goes
from a RANS region to an LES region (i.e. dfk /dt < 0), i.e.
 
dfk
Pktr = ktot min ,0 (23.38)
dt

where ktot is computed as in Eq. 23.37. The production term Pktr < 0 which means
that it reduces k as it should. The sum of the Si1 and Si2 terms (see Eqs. 23.29 and
23.31) in the momentum equations read
  
dfk k
Si = − min ,0 0.5 + ′ ′ v̄i′ (23.39)
dt hv̄m v̄m ir.a.

Since dfk /dt < 0, the source Si has the same sign as v̄i′ ; this means that the source
enhances the resolved turbulence as it should.
It may be noted that dfk /dt in Eq. 23.38 assumes the correct (i.e. negative) sign
irrespectively of the orientation of the RANS-LES interface. Consider, for example,
the RANS-LES interfaces at the lower and upper wall in fully-developed channel flow
(Fig. 23.3). For these interfaces, the gradient of fk in Eq. 23.38 reads v̄∂fk /∂y. When
a fluid particles at the lower interface goes from the RANS region to the LES region,
v̄ > 0 and ∂fk /∂y < 0 so that v̄∂fk /∂y < 0 as intended. Also for the upper interface
we get v̄∂fk /∂y < 0 since v̄ < 0 and ∂fk /∂y > 0. For non-cartesian grids the
material derivative, dfk /dt < 0, has to be formulated in local grid coordinates.
It is however found that the forcing often becomes too strong when Si is added
to the momentum equation. The effect of adding or neglecting Si in the momentum
equations is evaluated, see [137].
The differences between Model 3 and 2 are that
• When Si = 0, no explicit modification is made in the momentum equation in
Model 3 (recall that −ν < νtr < 0 is used in the momentum equation in Models
1 and 2).
• Model 2 (and Model 1) may need regularization in case |s̄| → 0 in the denomi-
nator of νttr in Eq. 23.35. No such regularization, however, is used here. It can
be argued that the commutation term in Model 3 is introduced in a more physical
way compared to Models 1 and 2 where artificially negative viscosities are used.
23.2. Zonal PANS: different treatments of the RANS-LES interface 219

23.2.2.4 Interface Model 4

This interface model was developed in [155] for horizontal interfaces. The modeled
turbulent kinetic energy in the LES region adjacent to the interface is reduced by setting
the usual convection and diffusion fluxes of k at the interface to zero. New fluxes are
introduced in which the interface condition is set to kint = fk kRAN S (fk = 0.4),
where kRAN S is the k value in the cell located in the URANS region adjacent to the
interface. No modification is made for the convection and diffusion of ε across the
interface. The implementation is presented in Section 23.1.1.
24. The PITM model 220

24 The PITM model

ITM is an acronym for Partially Integrated Transport Model [135, 157].
P
24.1 RANS mode
Consider homogeneous turbulence, e.g. decay of grid generated turbulence, see Section
3.4 in [53] and Fig. 11.2. The k and ε equations can for this flow be simplified as
dk
= Pk − ε (24.1)
dt
dε ε
= (cε1 P k − cε2 ε) (24.2)
dt k
where d/dt = v̄1 d/dx. The dissipation can be estimated as

k 3/2
ε=c (24.3)
Lt
Differentiation of Eq. 24.3 gives

dε 3 k 1/2 dk k 3/2 dLt

=c −c 2 (24.4)
dt 2 Lt dt Lt dt
Inserting Eq. 24.1 into Eq. 24.4

dε 3 k 1/2 k k 3/2 dLt

=c (P − ε) − c 2
dt 2 Lt Lt dt
and using Eq. 24.3 gives

dε 3ε k ε dLt 3 Pk − ε T dLt ε2
= (P − ε) − = − (24.5)
dt 2k Lt dt 2 T Lt dt k
where T = k/ε. Comparing Eqs. 24.5 and 24.2 we find that
T dLt
Cε1 = 1.5 and Cε2 = Cε1 + (24.6)
Lt dt

24.2 LES mode

Now let’s transform the RANS k − ε model above to an SGS k − ε model. In the
RANS model, k represents all turbulence. In an SGS model, the modeled turbulent
kinetic energy, ksgs , represents scales with wavelength larger that the cut-off, κc , see
Fig. 18.5. The viscous dissipation is the same in RANS and LES. The production
P k is replaced by Pksgs = εsgs (see Fig. 18.5). The time scale, T , is replaced by
Tsgs = ksgs /ε since the time scale of the small scales is different (it is smaller) than
that of the large scales. Now, Eq. 24.5 can be written as

dε εεsgs ε2
= Cε1,sgs − Cε2,sgs
dt ksgs ksgs
(24.7)
3 Tsgs dLt 3 T dLt ksgs
Cε2,sgs = + = +
2 Lt dt 2 Lt dt ktot
24.2. LES mode 221

E
Fc

Ec ε

κ−5/3

dκ

κc κd κ

Figure 24.1: Spectral energy balance in PITM. Homogeneous turbulence.

where ktot denotes the total turbulence (i.e. resolved and SGS). Note that L is not
replaced by Lsgs , because Eq. 24.3 is still a valid estimate ε.
Let’s introduce a new dissipative length scale of wavenumber κd which is much
larger than κc
ε
κd − κc = ξ 3/2 (24.8)
ksgs
The constant is very large since κc ≪ κd ; thus we can write
ε
κd ≃ ξ 3/2
(24.9)
ksgs
Consider spectral balance of ksgs in homogeneous turbulence, see Fig. 24.1. When the
cut-off does not move we get simply

Fc = ε (24.10)

where Fc = F (κc ) denotes the spectral transfer from wavenumber range [0, κ− c ] into
[κ+
c , ∞]; note that κ d ≃ ∞. Recall that ksgs is the area below E in Fig. 24.1. see
Eq. 18.20. When the cut-off does not move in time (which implies that dksgs /dt), then

Fc = εsgs = Pksgs (24.11)

When κc is time dependent as, for example, in decaying homogeneous grid turbulence,
see Eq. 24.1, we can set-up the balance of ksgs when κc moves dκc during the time
interval dt, i.e.
dksgs = Fc dt − Ec dκc − εdt (24.12)
where Ec dκ denotes the spectral energy in the slice dκc . Taking the time derivative of
Eq. 24.12 we get
dksgs dκc
= Fc − Ec −ε (24.13)
dt dt
Comparing Eq. 24.13 with Eq. 24.10 and 24.11, we find the spectral energy transfer,
εsgs , at the cut-off, κc , is
dκc
εsgs = Fc − Ec (24.14)
dt
24.2. LES mode 222

We can write a similar equation as Eq. 24.14 at wavenumber κd , i.e.

dκd
ε = Fd − Ed (24.15)
dt
where ε is the energy transfer rate across κd . Now derivate Eq. 24.9 with respect to
time and insert Eq. 24.15 !
d ε Fd − ε
ξ 3/2 = (24.16)
dt ksgs Ed
It may be noted that both the nominator and denominator of the right side go to zero as
κd → ∞. The left side is re-written as
" #
1 dε 3 −5/2 dksgs
ξ 3/2 − ksgs ε (24.17)
ksgs dt 2 dt

With Eq. 24.13 we get

"  #
1 dε 3 −5/2 dκc
ξ 3/2 − ksgs ε Fc − Ec −ε (24.18)
ksgs dt 2 dt
Replacing the left side of Eq. 24.16 with Eq. 24.18 gives
"  #
1 dε 3 −5/2 dκc Fd − ε
ξ 3/2 − ksgs ε Fc − Ec −ε = (24.19)
ksgs dt 2 dt Ed

We get (using first Eq. 24.14 and then Eq. 24.9)

dε 3/2 Fd − ε 3 ε Fd − ε 3 ε
= ksgs + (εsgs − ε) = ε + (εsgs − ε) (24.20)
dt ξEd 2 ksgs κd Ed 2 ksgs
This equation can now be written
  
dε 3 ε ε2 3 ksgs Fd
= εsgs − − −1 (24.21)
dt 2 ksgs
|{z} ksgs 2 κd Ed ε
| {z }
Cε1,sgs Cε2,sgs

When κc → 0 (RANS), then ksgs → ktot so that

Cε1,sgs → Cε1 (24.22)

 
3 ktot Fd
Cε2,sgs → Cε2 = − −1 (24.23)
2 κd Ed ε
It may be noted that both the term in the parenthesis and Ed go to zero as κd → ∞.
Inserting Eq. 24.23 in the expression for Cε2,sgs in Eq. 24.21 gives
   
3 ksgs ktot Fd 3 ksgs 3
Cε2,sgs = − −1 = + Cε2 −
2 ktot κd Ed ε 2 ktot 2
(24.24)
ksgs
= Cε1 + (Cε2 − Cε1 )
ktot
When ksgs /ktot → 0 (only a small part of the turbulent kinetic energy is modeled),
the PITM model is in LES mode, i.e. Cε2,sgs → Cε1 . Then Cε2,sgs is reduced which
24.2. LES mode 223

increases ε and reduces the modeled turbulent kinetic energy and the turbulent viscos-
ity. On the other hand, when ksgs /ktot → 1 (everything is modeled), then the PITM
model is in RANS mode (i.e. a usual RANS k − ε model).
It should be noted that this model is very similar to the PANS model. Indeed,
∗
Eq. 24.24 is identical to the expression for Cε2 , see Eq. 23.17 (provided that no dissi-
pative scales are resolved so that fε = 1), i.e.

Cε2,sgs = Cε1 + fk (Cε2 − Cε1 ) (24.25)

In both models the Cε2 coefficients are modified. The only difference between the
PANS and PITM models is that the diffusion coefficients in the ku and εu are modified
in the former model, see Eq. 23.21.
In [135] they do not use ksgs /ktot to compute Cε2,sgs . Instead they compute it
using the ratio of the integral and the modeled lengthscales as
0.42
Cε2 = Cε1 +
1 + γ(Lt /∆)2/3
3/2
where Lt = ktot /ε and γ = 0.42.
25. Hybrid LES/RANS for Dummies 224

25 Hybrid LES/RANS for Dummies

25.1 Introduction
LUID flow problems are governed by the Navier-Stokes equations
F ∂vi ∂vi vj 1 ∂p ∂ 2 vi
+ =− +ν (25.1)
∂t ∂xj ρ ∂xi ∂xj ∂xj
where vi denotes the velocity vector, p is the pressure and ν and ρ are the viscosity and
density of the fluid, respectively. In turbulent flow, the velocity and pressure are un-
steady and vi and p include all turbulent motions, often called eddies. The spatial scale
of these eddies vary widely in magnitude where the largest eddies are proportional to
the size of the largest physical length (for example the boundary layer thickness, δ, in
case of a boundary layer). The smallest scales are related to the eddies where dissipa-
tion takes place, i.e. where the kinetic energy of the eddies is transformed into internal
energy causing increased temperature. The ratio of the largest to the smallest eddies
increases with Reynolds number, Re = |vi |δ/ν. This has the unfortunate consequence
– unless one is a fan of huge computer centers – that it is computationally extremely
expensive to solve the Navier-Stokes equations for large Reynolds numbers.

25.1.1 Reynolds-Averaging Navier-Stokes equations: RANS

In order to be able to solve the Navier-Stokes equations with a reasonable computa-
tional cost, the velocity vector and the pressure are split into a time-averaged part (hvi i
and hpi) and a fluctuating part (vi′ and p′ ), i.e. vi = hvi i + vi′ , p = hpi + p′ . The
resulting equation is called the RANS (Reynolds-Averaging Navier-Stokes) equations
∂hvi ihvj i 1 ∂hpi ∂ 2 hvi i ∂hvi′ vj′ i
=− +ν −
∂xj ρ ∂xi ∂xj ∂xj ∂xj
  (25.2)
1 ∂hpi ∂ ∂hvi i
=− + (ν + νt )
ρ ∂xi ∂xj ∂xj
The last term on the first line is called the Reynolds stress and it is unknown and must
be modeled. All turbulent fluctuation are modeled with a turbulence model and the
results when solving Eq. 25.2 are highly dependent on the accuracy of the turbulence
model. On the right side of Eq. 25.2 the unknown Reynolds stresses are expressed by
a turbulence model in which a new unknown variable is introduced which is called the
turbulent viscosity, νt . The ratio of νt to ν may be of the order of 1000 or larger. In in-
dustry today, CFD (Computationally Fluid Dynamics) based on finite volume methods
is used extensively to solve the RANS equations, Eq. 25.2.

25.1.2 Large Eddy Simulations: LES

A method more accurate than RANS is LES (Large Eddy Simulations) in which only
the small eddies (fluctuations whose eddies are smaller than the computational cell) are
modeled with a turbulence model. The LES equations read
∂v̄i ∂v̄i v̄j 1 ∂ p̄ ∂ 2 v̄i ∂τij
+ =− +ν −
∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj
  (25.3)
1 ∂ p̄ ∂ ∂v̄i
=− + (ν + νsgs )
ρ ∂xi ∂xj ∂xj
25.2. The PANS k − ε turbulence model 225

Note that the time dependence term (the first term on the left side of the first line) has
been retained, because the large, time dependent turbulent (i.e. the resolved) fluctua-
tions are part of v̄i and p̄ and are not modeled with the turbulence model. The last term
on the first line includes the Reynolds stresses of the small eddies, which are called
SGS (sub-grid stresses). This term must also – as in Eq. 25.2 – be modeled, and at the
second line it has been modeled with a SGS turbulent viscosity, νsgs . The difference
of νsgs compared to νt in Eq. 25.2 is that it includes only the effect of the small eddies.
The ratio of νsgs to ν is of the order of 1 to 100. However, the ratio of the resolved
to the modeled turbulence, |v̄i′ v̄j′ |/|τij | (see Eqs. 25.2 and 25.3) is much larger than
one. Hence, LES is much more accurate than RANS because only a small part of the
turbulence is modeled with the turbulence SGS model whereas in RANS all turbulence
is modeled. The disadvantage of LES is that it is much more expensive than RANS
because a finer mesh must be used and because the equations are solved in four dimen-
sions (time and three spatial directions) whereas RANS can be solved in steady state
(no time dependence).
When the flow near walls is of importance, is turns out that LES is prohibitively
expensive because very fine cells must be used there. The reason is entirely due to
physics: near the walls, the spatial scales of the “large” turbulent eddies which should
be resolved by LES are in reality rather small. Furthermore, their spatial scales get
smaller for increasing Reynolds number. Much research has the last ten years been
carried out to circumvent this problem. All proposed methods combines RANS and
LES where RANS is used near walls and LES is used some distance away from the
walls, see Fig. 25.1. These methods are called Detached Eddy Simulation (DES), hy-
brid LES/RANS or zonal LES/RANS. The focus here is zonal LES/RANS.

25.1.3 Zonal LES/RANS

Equations 25.2 and 25.3 can be written in a same form as
 
∂v̄i ∂v̄i v̄j 1 ∂ p̄ ∂ ∂v̄i
+ =− + (ν + νT ) (25.4)
∂t ∂xj ρ ∂xi ∂xj ∂xj

Near the walls, a RANS turbulence model is used for the turbulent viscosity, i.e. νT =
νt and away from the walls an LES turbulence model is employed, i.e. νT = νsgs .
Note that the time dependence term is now retained also in the RANS region: near the
wall we are using an unsteady RANS, i.e. URANS.
Above, we have describe how to use the zonal LES/RANS method for flows near
walls. Another form of zonal LES/RANS is embedded LES, in which an LES region is
embedded in a RANS region. One example is prediction of aeroacoustic noise created
by the turbulence around an external mirror on a vehicle [105]. The flow around the ve-
hicle can be computed with RANS, but in order to predict the noise in the region of the
external mirror we must predict the large turbulence fluctuations and hence LES must
be used in this region. In Section 25.4 we will present simulations using embedded
LES in a simplified configuration represented by the flow in a channel in which RANS
is used upstream of the interface and LES is used downstream of it, see Fig. 25.4.

25.2 The PANS k − ε turbulence model

The PANS k − ε model can be used to simulate wall-bounded flow at high Reynolds
number as well as embedded LES. The turbulence model reads [133,134], see Eq. 23.21
25.3. Zonal LES/RANS: wall modeling 226

URANS, fk = 1.0
interface
2
LES, fk = 0.4
x2 interface
URANS, fk = 1.0

x1 3.2

Figure 25.1: The LES and URANS regions. Fully developed channel flow. Periodic
boundary conditions are applied at the left and right boundaries.

(here in a slightly simplified form to enhance readability)

  
∂k ∂kv̄j ∂ νT ∂k
+ = ν+ + Pk − ε (25.5)
∂t ∂xj ∂xj σk ∂xj
   2
∂ε ∂εv̄j ∂ νT ∂ε ε ∗ ε
+ = ν+ + Cε1 P k − Cε2 (25.6)
∂t ∂xj ∂xj σε ∂xj k k
∗
Cε2 = Cε1 + fk (Cε2 − Cε1 ), Cε1 = 1.5, Cε2 = 1.9 (25.7)
k2
νT = Cµ , Cµ = 0.09 (25.8)
ε
Note that k and ε are always positive. The key elements in the present use of the PANS
k − ε model are highlighted in red. When fk in Eq. 25.7 is equal to one, the model
acts as a standard k − ε RANS model giving a large turbulent viscosity. When fk is
∗
decreased (to 0.4 in the present study), Cε2 in Eq. 25.7 decreases. As a result

• ε increases because the destruction term (last term in Eq. 25.6 which is the main
sink term) in the ε equation decreases,
• k decreases because ε (last term in Eq. 25.5) is the main sink term in the k
equation increases, and
• νT in Eq. 25.8 decreases because k decreases and ε increases.

Hence, the turbulence model in Eqs. 25.5–25.8 acts as a RANS turbulence model
(large turbulent viscosity) when fk = 1 and it acts as an LES SGS turbulence model
(small turbulent viscosity) when fk = 0.4.

25.3 Zonal LES/RANS: wall modeling

25.3.1 The interface conditions
The interface plane (see Fig. 25.1) separates the URANS regions near the walls and the
LES region in the core region. In the LES region fk = 0.4 and in the URANS region
fk = 1. In the former region, the turbulent viscosity νT should be an SGS viscosity and
in the latter region it should be an RANS viscosity. Hence νT must decrease rapidly
when going from the URANS region to the LES region. This is achieved by setting
the usual convection and diffusion fluxes of k at the interface to zero. New fluxes are
introduced using smaller SGS values [155].
25.4. Zonal LES/RANS: embedded LES 227

0
30
−0.2
25

hv1′ v2′ i+
20 −0.4
U+

15
−0.6
10
−0.8
5

0 −1
1 100 1000 30000 0 0.05 0.1 0.15 0.2
x+ x2
2
(a) Velocities (b) Resolved shear stresses

Figure 25.2: Velocities and resolved shear stresses. (Nx × Nz ) = (64 × 64) :
Reτ = 4 000; : Reτ = 8 000; : Reτ = 16 000; ////: Reτ = 32 000.

25.3.2 Results
Fully developed channel flow is computed for Reynolds numbers Reτ = uτ δ/ν =
4 000, 8 000, 16 000 and 32 000. The baseline mesh has 64 × 64 cells in the streamwise
(x1 ) and spanwise (x3 ) directions. The size of the domain is x1,max = 3.2, x2,max = 2
and x3,max = 1.6 (δ = uτ = 1). The grid in the x2 direction varies between 80 and
128 cells depending on Reynolds number. The interface is set to x+ 2 ≃ 500 for all
grids.
The velocity profiles and the resolved shear stresses are presented in Fig. 25.2. As
can be seen, the predicted velocity profiles are in good agreement with the log-law
which represents experiments. Figure 25.2b presents the resolved shear stresses. The
interface is shown by thick dashed lines and it moves towards the wall for increasing
Reynolds number since it is located at x+ 2 ≃ 500 for all Reynolds numbers.
The turbulent viscosity profiles are shown in Fig. 25.3 for three different resolutions
in the x1 − x3 plane. It is interesting to note that the turbulent viscosity is not affected
by the grid resolution. Hence, the model yields grid independent results contrary to
other LES/RANS models.
The turbulent viscosity (Fig. 25.3) is sharply reduced when going across the in-
terface from the URANS region to the LES region and the resolved fluctuations (the
Reynolds shear stress in Fig. 25.2b) increase. This shows that the model is switch-
ing from RANS mode to LES mode as it should. More detailed results can be found
in [155].

25.4 Zonal LES/RANS: embedded LES

25.4.1 The interface conditions
The interface plane is now vertical, see Fig. 25.4. The interface conditions for k and
ε are treated in the same way as in Section 25.3.1. The difference is now that “inlet”
turbulent fluctuations must be added to the LES v̄i equations (Eq. 25.3) to trigger the
flow into turbulence-resolving mode. Anisotropic synthetic turbulent fluctuations are
used [158, 159].
 25.4. Zonal LES/RANS: embedded LES 228

0.015 0.015

0.01 0.01

νT /(uτ δ)
νT /(uτ δ)

0.005 0.005

0 0
0 0.05 0.1 0.15 0.2 0.25 0.3 0 0.2 0.4 0.6 0.8 1
x2 x2
(a) Reτ = 4000. : (Nx × Nz ) = (64 × (b) (Nx × Nz ) = (64 × 64). : Reτ =
64); : (Nx × Nz ) = (32 × 32); : 4 000; : Reτ = 8 000; : Reτ =
(Nx × Nz ) = (128 × 128) 16 000; ////: Reτ = 32 000.

Figure 25.3: Turbulent viscosity.

Interface

RANS, fk = 1.0 LES, fk = 0.4

2
x2

x1
1 2.2

Figure 25.4: The LES and RANS regions. The left boundary is an inlet and the right
boundary is an outlet.

4 100
30
1/2

25
(maxx2 hu′1 u′1 i)

hνT /νimax

20
U+

15 2 50

10

0 0 1 2
0 0
10 10 10 0 0.5 1 1.5 2 2.5 3 3.5
x1
x+
2
(a) Velocities. x1 = 0.19; x1 = (b) : Turbulent viscosity (right x2 axis);
1.25; x1 = 3. The log-law is plotted in : maximum streamwise fluctuations (left x2
symbols. axis)

Figure 25.5: Channel flow with inlet and outlet. (a) Velocities; (b) maximum resolved
streamwise turbulent fluctuations and turbulent viscosity versus x1 .
25.4. Zonal LES/RANS: embedded LES 229

25.4.2 Results
The Reynolds number for the channel flow is Reτ = 950. With a 3.2 × 2 × 1.6 domain,
a mesh with 64 × 80 × 64 cells is used in the streamwise (x1 ), the wall-normal (x2 )
and the spanwise (x3 ) direction, see Fig. 25.4. Inlet conditions at x = 0 are created by
computing fully developed channel flow with the PANS k − ε model in RANS mode
(i.e. with fk = 1).
Figure 25.5a presents the mean velocity and the resolved shear stresses at three
streamwise locations, x1 = 0.19, 1.25 and 3 (recall that the interface is located at
x1 = 1). At x1 = 3, the predicted velocity agrees very well with the experimental
log-law profile.
The resolved streamwise velocity fluctuations are zero in the RANS region, as they
should (Fig. 25.5b), and the maximum resolved values increase sharply over the in-
terface thanks to the imposed synthetic turbulent “inlet” fluctuations. The turbulent
viscosity is reduced at the interface from its peak RANS value of approximately 80
to a small LES value of approximately one (these values are both fairly low because
of the low Reynolds number). Hence, it is seen that the present model successfully
switches from RANS to LES across the interface. The results are presented in more
detail in [155].
26. Commutation terms in the k and ω equations 230

26 Commutation terms in the k and ω equations

In embedded LES, part of the flow is treated by RANS and part by LES. Here we
are interested in a configuration where the upstream region is treated by RANS and
the downstream by LES, and the interface is vertical, parallel to the inlet. We denote
the interface between the two regions as “the RANS-LES interface” located at xR−L ,
see Figs.25.4, 26.1 and 26.2. The transition region in which the predicted flow is in
between RANS and LES is often called the grey area, a problem described in [160].
When the flow goes from a RANS region to an LES region through the RANS-LES
interface, it should – in order to minimize the grey area – switch as quickly as possible
from RANS mode to LES mode.
Hamda [161] showed that, when the filter size (i.e. the grid) is non-uniform, a
commutation error appears in SGS models based on transport equations. He found by
analyzing DNS channel data that this commutation term is large at interfaces between
RANS and LES. Here we show how to apply the commutation term derived in [161] at
RANS-LES interfaces. We present a method in which commutations terms are added
in the k and ω equations in order to quickly switch from RANS te LES.
The two-equation zonal k − ω hybrid RANS-LES model of [162] is employed. In
the LES region, the model reads
  
∂k ∂v̄i k k k 3/2 ∂ νt ∂k
+ = P − fk + ν+ (26.1)
∂t ∂xi ℓt ∂xj σk ∂xj
  
∂ω ∂v̄i ω ω k 2 ∂ νt ∂ω νt ∂k ∂ω
+ = Cω1 fω P − Cω2 ω + ν+ + Cω
∂t ∂xi k ∂xj σω ∂xj k ∂xj ∂xj
  (26.2)
k k ∂v̄i ∂v̄j ∂v̄i
νt = f µ , P = νt + , ℓt = ΨP DH CLES ∆dw
ω ∂xj ∂xi ∂xj
"  3/4 #
fω (26.3)
ΨP DH = min 10, fk , ∆max = max{∆x, ∆y, ∆z}
fµ
∆dw = min (max [Cdw dw , Cw ∆max , ∆nstep ] , ∆max )
where dw denotes the distance to the nearest wall and ∆step is the grid step size in the
wall-normal direction. The damping functions read
"   # "  1/2 #
4
Rt Rt
fk = 1 − 0.722 · exp − , fω = 1 + 4.3 · exp −
10 1.5
( "   #) ( "  2 #)
3/4
Rt 0.001 Rt
fµ = 0.025 + 1 − exp − 0.975 + · exp − .
10 Rt 200

The turbulent Reynolds number is defined as Rt = k/(νω). The length scale, ∆dw ,
is taken from the IDDES model [153]. In the RANS regions, ℓt = k 1/2 /(Ck ω). The
constants read σk = 0.8, σw = 1.35, Ck = 0.09 Cω1 = 0.42, Cω2 = 0.075, Cω =
0.75, CLES = 0.7 and Cdw = 0.15.
The difference between Eqs. 26.1 and 26.2 in the RANS and LES regions can be
summarized as follows: in RANS regions, the RANS lengthscale, ℓt = k 1/2 /(Ck ω),
is used in the dissipation term in the k equation and, in LES regions, the filter length
scale, ∆dw , is used.
26. Commutation terms in the k and ω equations 231

xtr

RANS

6= 0
v̄ ′ , v̄ ′ , w̄′
LES 2δ

∂x1
∂∆
y RANS

L
x

Figure 26.1: Channel flow.

xtr
6= 0

LES
∂x1
∂∆

′ ′ ′
v̄ , v̄ , w̄
δin
y RANS

L
x

Figure 26.2: Boundary layer flow.

When the filter size in LES varies in space, an additional term appears in the mo-
mentum equation because the spatial derivatives and the filtering do not commute. For
the convective term in Navier-Stokes, for example, we get

∂vi vj ∂

= (vi vj ) + O (∆x)2 .
∂xj ∂xj

Ghosal & Moin [69] show that the error is proportional to (∆x)2 and, since this error is
of the same order as the discretization error of most finite volume methods, it is usually
neglected.
However, in zonal4 hybrid RANS-LES, the length scale at the RANS-LES interface
changes abruptly from a RANS length scale to an LES length scale. Hamda [161]
estimated the commutation error at RANS-LES interfaces and found that it is large.
The commutation term for the divergence of a flux, qi , reads

∂qi ∂ q¯i ∂∆ ∂ q¯i

= −
∂xi ∂xi ∂xi ∂∆
4 by “zonal”, we imply that the interface is chosen at a location where the RANS and LES length scales

differ
26. Commutation terms in the k and ω equations 232

For the k equation, the commutation term reads [161]

∂ui k ∂v̄i k ∂∆ ∂v̄i k

= − (26.4)
∂xi ∂xi ∂xi ∂∆
Consider a fluid particle in a RANS region moving in the x1 direction and passing
across a RANS-LES interface. The filter width decreases across the interface, i.e.
∂∆ ∆LES − ∆RAN S
≃ <0 (26.5)
∂x1 ∆x1
and
∂v̄1 k u1 kLES − u1 kRAN S
≃ >0 (26.6)
∂∆ ∆LES − ∆RAN S
which means that the last term on the right-hand side of Eq. 26.4 gives a positive
contribution on the left side of the k equation, Eq. 26.1; on the right side of Eq. 26.1,
it gives a negative contribution. Hence, the additional term in Eq. 26.4 at a RANS-
LES interface reduces k, as expected. To obtain the right hand of Eqs. 26.5 and 26.6,
the derivatives on the left side of the equations have been estimated by simple finite-
difference expressions, i.e.
df ∆f
≃ (26.7)
dx ∆x
It may be noted that the idea of adding an additional source term in the k equation
due to a commutation error is similar to the proposal in [156]; they use a commutation
term based on the gradient of fk in the PANS model. This idea was later used by the
present author at RANS-LES interfaces [137].
To find the corresponding term in the ω equation, let us start by looking at the ε
equation. What happens with ε when a fluid particle moves from a RANS region into
an LES region? The answer is, nothing. The dissipation is the same in a RANS region
as in an LES region. This is best seen by looking at the ksgs equation
  
∂ksgs ∂v̄i ksgs ∂ νsgs ∂ksgs
+ = P ksgs + ν+ − ε. (26.8)
∂t ∂xi ∂xj σk ∂xj
The dissipation term, ε, in Eq. 26.8 is the same as the dissipation term in Eq. 26.1 unless
the resolution is very fine (close to DNS). Then much of the dissipation is resolved,
reducing the production term, P ksgs . However, this kind of resolution is not realistic.
Now consider the ω equation. It is derived by transformation of the k and ε equa-
tions to an ω equation as
 
dω d ε 1 dε ε d(1/k) 1 dε ω dk
= = + = − (26.9)
dt dt Ck k Ck k dt Ck dt Ck k dt k dt
The right-hand side shows that the source terms in the ω equation correspond to those in
the ε equation multiplied by 1/(Ck k) together with those in the k equation multiplied
by −ω/k. Hence, the source term due to the commutation error in the ω equation is
the commutation term in Eq. 26.4 multiplied by −ω/k so that

∂ui ω ∂v̄i ω ∂∆ ∂v̄i ω ∂v̄i ω ω ∂∆ ∂v̄i k

= − = + (26.10)
∂xi ∂xi ∂xi ∂∆ ∂xi k ∂xi ∂∆
Assuming again a flow in the x1 direction from a RANS region to an LES region,
we find that the second term on the right-hand side of Eq. 26.10 is negative since
26. Commutation terms in the k and ω equations 233

∂∆/∂x1 < 0 (see Eq. 26.5) and ∂v̄1 k/∂∆ > 0 (see Eq. 26.6) so that the commutation
term in Eq. 26.10 is positive/negative on the right/left-side of the ω equation (Eq. 26.2).
This means that the commutation term in Eq. 26.10 will increase ω when moving from
a RANS region to an LES region. Hence the source terms in the k and ω equations both
contribute to reducing the turbulent viscosity, which is an effect we are looking for at
RANS-LES interfaces: a reduced turbulent viscosity will promote growth of resolved
turbulence on the LES side of an interface.
The results using the commutation terms are presented in [163, 164].
27. Inlet boundary conditions 234

27 Inlet boundary conditions

N RANS it is sufficient to supply profiles of the mean quantities such as velocity
I and temperature plus the turbulent quantities (e.g. k and ε). However, in unsteady
simulations (LES, URANS, DES . . . ) the time history of the velocity and temperature
need to be prescribed; the time history corresponds to turbulent, resolved fluctuations.
In some flows it is critical to prescribe reasonable turbulent fluctuations, but in many
flows it seems to be sufficient to prescribe constant (in time) profiles [119, 165].
There are different ways to create turbulent inlet boundary conditions. One way is
to use a pre-cursor DNS or well resolved LES of channel flow. This method is limited
to fairly low Reynolds numbers and it is difficult (or impossible) to re-scale the DNS
fluctuations to higher Reynolds numbers.
Another method based partly on synthesized fluctuations is the vortex method [166].
It is based on a superposition of coherent eddies where each eddy is described by a
shape function that is localized in space. The eddies are generated randomly in the
inflow plane and then convected through it. The method is able to reproduce first and
second-order statistics as well as two-point correlations.
A third method is to take resolved fluctuations at a plane downstream of the inlet
plane, re-scale them and use them as inlet fluctuations.
Below we present a method of generating synthesized inlet fluctuations.

27.1 Synthesized turbulence

The method described below was developed in [89, 167, 168] for creating turbulence
for generating noise. It was later further developed for inlet boundary conditions [158,
169, 170].
A turbulent fluctuating velocity fluctuating field (whose average is zero) can be
expressed using a Fourier series, see Section 5.3 and Eq. D.17. Let us re-write this
formula as
an cos(nx) + bn sin(nx) =
(27.1)
cn cos(αn ) cos(nx) + cn sin(αn ) sin(nx) = cn cos(nx − αn )
where an = cn cos(α) , bn = cn sin(αn ). The new coefficient, cn , and the phase angle,
αn , are related to an and bn as
 

1/2 bn
cn = a2n + b2n αn = arctan (27.2)
an
A general form for a turbulent velocity field can thus be written as
N
X
v′ (x) = 2 ûn cos(κn · x + ψ n )σ n (27.3)
n=1

where ûn , ψ n and σin are the amplitude, phase and direction of Fourier mode n. The
synthesized turbulence at one time step is generated as follows.

27.2 Random angles

The angles ϕn and θn determine the direction of the wavenumber vector κ, see Eq. 27.3
and Eq. 27.1; αn denotes the direction of the velocity vector, v′ . For more details, see
Appendix J.
27.3. Highest wave number 235

ξ2n σin

αn
x3
ξ1n
ξ3n
ξ3n

ξ1n
θ n κn
i

0
x2

x1 ϕn

Figure 27.1: The wave-number vector, κni , and the velocity unit vector, σin , are orthog-
onal (in physical space) for each wave number n.

27.3 Highest wave number

Define the highest wave number based on mesh resolution κmax = 2π/(2∆) (see
Section 18.5), where ∆ is the grid spacing. Often the smallest grid spacing near the
wall is too small, and then a slightly larger values may be chosen. The fluctuations
are generated on a grid with equidistant spacing (or on a weakly stretched mesh),
∆η = x2,max /N2 , ∆x3 = x3,max /N3 , where η denotes the wall-normal direction
and N2 and N3 denote the number of cells in the x2 and x3 direction, respectively. The
fluctuations are set to zero at the wall and are then interpolated to the inlet plane of the
CFD grid (the x2 − x3 plane).

27.4 Smallest wave number

Define the smallest wave number from κ1 = κe /p, where κe = α9π/(55Lt), α =
1.453. The turbulent length scale, Lt , may be estimated in the same way as in RANS
simulations, i.e. Lt ∝ δ where δ denotes the inlet boundary layer thickness. In [158,
169, 170] it was found that Lt ≃ 0.1δin is suitable.
Factor p should be larger than one to make the largest scales larger than those
corresponding to κe . A value p = 2 is suitable.

27.5 Divide the wave number range

Divide the wavenumber space, κmax − κ1 , into N modes, equally large, of size ∆κ.
27.6. von Kármán spectrum 236

27.6 von Kármán spectrum

A modified von Kármán spectrum is chosen, see Eq. 27.4 and Fig. 27.2. The amplitude
ûn of each mode in Eq. 27.3 is then obtained from

ûn = (E(κ)∆κ)1/2
u2rms (κ/κe )4 2
E(κ) = cE e[−2(κ/κη ) ] (27.4)
κe [1 + (κ/κe )2 ]17/6
κ = (κi κi )1/2 , κη = ε1/4 ν −3/4

The coefficient cE is obtained by integrating the energy spectrum over all wavenumbers
to get the turbulent kinetic energy, i.e.
Z ∞
k= E(κ)dκ (27.5)
0

which gives [73]

4 Γ(17/6)
cE = √ ≃ 1.453 (27.6)
π Γ(1/3)
where Z ∞
′
Γ(z) = e−z xz−1 dz ′ (27.7)
0

27.7 Computing the fluctuations

Having ûn , κnj , σin and ψ n , allows the expression in Eq. 27.3 to be computed, i.e.

N
X
v1′ =2 ûn cos(β n )σ1
n=1
N
X
v2′ = 2 ûn cos(β n )σ2
n=1
(27.8)
N
X
v3′ = 2 ûn cos(β n )σ3
n=1
β n = k1n x1 + k2n x2 + k3n x3 + ψ n

where ûn is computed from Eq. 27.4.

In this way inlet fluctuating velocity fields (v1′ , v2′ , v3′ ) are created at the inlet x2 −x3
plane.
The code for generating the isotropic fluctuations can be downloaded here
http://www.tfd.chalmers.se/˜lada/projects/inlet-boundary-conditions/proright.ht

27.8 Introducing time correlation

A fluctuating velocity field is generated each time step as described above. They are
independent of each other and their time correlation will thus be zero. This is non-
physical. To create correlation in time, new fluctuating velocity fields, V1′ , V2′ , V3′ , are
27.8. Introducing time correlation 237

E(κ)
κe
κ1

E(κn ) E(κ) ∝ κ−5/3

∆κn

κn log(κ)
Figure 27.2: Modified von Kármán spectrum

computed based on an asymmetric time filter

(V1′ )m = a(V1′ )m−1 + b(v1′ )m

(V2′ )m = a(V2′ )m−1 + b(v2′ )m (27.9)
(V3′ )m = a(V3′ )m−1 + b(v3′ )m

where m denotes the time step number and

a = exp(−∆t/Tint ) (27.10)

where ∆t and Tint denote the computational time step and the integral time scale,
respectively. The second coefficient is taken as

b = (1 − a2 )0.5 (27.11)

which ensures that hV1′2 i = hv1′2 i (h·i denotes averaging). The time correlation of will
be equal to
exp(−t̂/Tint ) (27.12)
where t̂ is the time separation and thus Eq. 27.9 is a convenient way to prescribe the
turbulent time scale of the fluctuations. For more detail, see Section 27.8. The inlet
boundary conditions are prescribed as (we assume that the inlet is located at x1 = 0
and that the mean velocity is constant in the spanwise direction, x3 )

v̄1 (0, x2 , x3 , t) = V1,in (x2 ) + u′1,in (x2 , x3 , t)

′
v̄2 (0, x2 , x3 , t) = V2,in (x2 ) + v2,in (x2 , x3 , t) (27.13)
′
v̄3 (0, x2 , x3 , t) = V3,in (x2 ) + v3,in (x2 , x3 , t)
′
where v1,in = (V1′ )m , v2,in
′
= (V2′ )m and v3,in
′
= (V3′ )m (see Eq. 27.9). The mean
inlet profiles, V1,in , V2,in , V3,in , are either taken from experimental data, a RANS
27.8. Introducing time correlation 238

0.8

0.6

B(τ )
0.4

0.2

0 0.2 0.4 0.6 0.8 1

τ

Figure 27.3: Auto correlation, B(τ ) = hv1′ (t)v1′ (t − τ )t (averaged over time, t). :
Eq. 27.12; : computed from synthetic data, (V1′ )m , see Eq. 27.9.

solution or from the law of the wall; for example, if V2,in = V3,in = 0 we can estimate
V1,in as [171]
 +
 x2 x+2 ≤ 5
+
V1,in = −3.05 + 5 ln(x2 ) 5 < x+
+
2 < 30 (27.14)
 1 + +
κ ln(x2 ) + B x 2 ≥ 30

where κ = 0.4 and B = 5.2.

The method to prescribed fluctuating inlet boundary conditions have been used for
channel flow [158], for diffusor flow [165] as well as for the flow over a bump and an
axisymmetric hill [172].
Equation 27.9 introduces a time correlation with an integral time scale Tint . In
order to understand Eq. 27.10, Equation 27.9 is written for m = N . . . 1 where N
denotes number of time steps
′ ′ ′
UN = aUN −1 + buN
′ ′ ′
UN −1 = aUN −2 + buN −1
′ ′ ′
UN −2 = aUN −3 + buN −2
.. (27.15)
.
U3 = aU2′ + bu′3
′

U2′ = aU1′ + bu′2

U1′ = aU0′ + bu′1
27.8. Introducing time correlation 239

With U0′ = 0 we get

U1′ = bu′1
U2′ = abu′1 + bu′2
U3′ = a2 bu′1 + abu′2 + bu′3 (27.16)
U4′ = a bu′1
3
+ a bu′2
2
+ abu′3 + bu′4
U5′ = a4 bu′1 + a3 bu′2 + a2 bu′3 + abu′4 + bu′5

which can be written as

m
X
′
Um =b am−k u′k (27.17)
k=1

using Eq. 27.10 with t̂ = t/Tint . Let’s show that hU ′2 i = hu′2 i. Equation 27.17 gives
m m m
1 X ′ ′ b2 X m−k ′ X m−i ′
hU ′2 i = Uk Uk = 2 a uk a ui (27.18)
m m i=1
k=1 k=1

Recall that the synthetic fields u′k and u′i are independent which means that hu′k u′i i = 0
for k 6= i. Hence
m m
b2 X 2(m−k) ′ ′ X
hU ′2 i = a uk uk = hu′2 i(1 − a2 ) a2k (27.19)
m
k=1 k=1

replacing a with b using Eq. 27.11. The requirement is that hU ′2 i = hu′ i and hence
m
X m
X
(1 − a2 ) a2k = (1 − a2 ) ck = 1 (27.20)
k=1 k=1
Pm
must be satisfied. The sum k=1 ck has for large m the value Sm = 1/(1 − c) =
1/(1 − a2 ) which shows that the requirement in Eq. 27.20 is satisfied.
The time correlation between many time steps with time step separation m − n =
r ≥ 0 reads
Xn m−r
X
′ ′ b2 n−p ′
hU U ir = a up am−k u′k
(m − r)n p=1
k=1
(27.21)
2 m−r
X
b
= a2(m−k) a−r u′2
k
′ ′
= hu u ia −r
m−r
k=1

using Eqs. 27.19 and 27.20. Inserting Eq. 27.10 into Eq. 27.21 we get the normalized
autocorrelation
hU ′ U ′ ir
B11 (t̂)norm = = exp(−t̂/Tint ) (27.22)
hu′2 i
where t̂ = r∆t. The integral time scale can be computed from the autocorrelation as
(see Eq. 10.11)
Z ∞ Z ∞
norm
Tint = B11 (t̂)dt̂ = exp(−t̂/Tint )dt̂
0 0 (27.23)
 ∞
= −Tint exp(−t̂/Tint ) 0 = −Tint (0 − 1) = Tint
27.9. Anisotropic Synthetic Turbulent Fluctuations 240

This shows that the one-sided filter in Eq. 27.9 introduces an integral time scale Tint
as prescribed by the definition of a in Eq. 27.10. The integral time scale can be taken
as Tint = Lint /Ub using Taylor’s hypothesis, where Ub and Lint denote the inlet bulk
velocity and the integral length scale, respectively.

27.9 Anisotropic Synthetic Turbulent Fluctuations

Isotropic fluctuations are generated above. However, turbulence is generally anisotropic.
The method for generating anisotropic turbulence is presented in [89,159,167,173] and
below it is described in detail.
The method can be summarized by the following steps.

1. A Reynolds stress tensor, hvi′ vj′ i, is taken from DNS data for turbulent channel
flow. Since the generated turbulence is homogeneous, it is sufficient to choose
one location of the DNS data. The Reynolds stresses at y + ≃ 16 of the DNS
channel data at Reτ = 590 [174] are used. Here hv1′ v1′ i – and hence the degree
of anisotropy – is largest. The stress tensor reads
 
7.67 −0.662 0
hvi′ vj′ i =  −0.662 0.32 0  (27.24)
0 0 1.50

2. The principal directions (the eigenvectors, see Fig. 27.4), ê∗i , are computed for
the hvi′ vj′ i tensor, see Section 27.9.1. The eigenvalues are normalized so that
their sum is equal to three. This ensures that the kinetic energy of the synthetic
fluctuations does not change during transformation.
3. Isotropic synthetic fluctuations, u′i,iso , are generated in the principal directions
of hu′i u′j i.
4. Now make the isotropic fluctuations non-isotropic according to hvi′ vj′ i. This is
done by multiplying the isotropic synthetic fluctuations
√ in the ê∗i directions by
the square-root of the normalized eigenvalues, λ (i)
√ norm . In Section 27.9.2 this
∗n
is done by multiplying the unit vector σi by λ(i) norm , see Eq. 27.28 and
Fig. 27.5. This gives a new field of fluctuations
q
(1)
′
(u1 )aniso = λnorm (u′1 )iso
q
(2) (27.25)
(u′2 )aniso = λnorm (u′2 )iso
q
(3)
(u′3 )aniso = λnorm (u′3 )iso
√
The wavenumber vector, κj , is divided by λ(i) to ensure that the (u′i )aniso
velocity field is divergence free, i.e.
q s
∗ ∗ (j) 1
σj κj = λnorm σj (j)
κj = σj κj = 0, (27.26)
λnorm

see Eq. J.3. Note that there is still no correlation between the (u′i )aniso fluctua-
tions, which means that the shear stresses are zero (for example, h(u′1 u′2 )aniso i =
0).
27.9. Anisotropic Synthetic Turbulent Fluctuations 241

5. The (u′i )aniso fluctuations are transformed to the computational coordinate sys-
tem, xi ; these anisotropic fluctuations are denoted (u′i )synt . The transformation
reads
(u′i )synt = Rij (u′j )aniso (27.27)
where Rij is the transformation matrix which is defined be the eigenvectors, see
Section 27.9.1. The (u′i )synt fluctuations are now used in Eq. 27.9 replacing
′
vi.in .
The Reynolds stress tensor of the synthetic anisotropic fluctuations is now iden-
tical to the DNS Reynolds stress tensor, i.e. h(u′i u′j )synt i = hvi′ vj′ i
6. Since the u′i,synt are homogeneous, the Reynolds stresses, hu′i,synt u′j,synt i, have
constant values in the inlet plane. One can choose to scale the inlet fluctuations
by a k profile taken from DNS, experiments or a RANS simulation.

The Matlab code for generating isotropic or anisotropic fluctuations can be down-
loaded [175].

27.9.1 Eigenvalues and eigenvectors

The normalized eigenvalues and eigenvectors may conveniently be computed with Mat-
lab (or the freeware Octave) as

stress=[7.6684 -6.6206e-01 0; ...

-6.6206e-01 3.1974e-01 0; ...
0 0 1.4997];
diag_sum=trace(stress)/3
stress=stress/diag_sum % ensures that the sum
% of the eigenvalues=3

[R,lambda] = eig(stress)

v_1_temp=[R(1,1);R(2,1); R(3,1)];
v_2_temp=[R(1,2);R(2,2); R(3,2)];
v_3_temp=[R(1,3);R(2,3); R(3,3)];

lambda_1_temp=lambda(1,1);
lambda_2_temp=lambda(2,2);
lambda_3_temp=lambda(3,3);

where stress is taken from Eq. 27.24. Matlab defines the smallest eigenvalue as the
first one and the largest as the last. Here we define the first eigenvalue (streamwise
direction) as the largest and the second (wall-normal direction) as the smallest, i.e.

v1=v3_temp;
v2=v1_temp;
v3=v2_temp;

lambda_1=lambda_3_temp;
lambda_2=lambda_1_temp;
lambda_3=lambda_2_temp;
27.9. Anisotropic Synthetic Turbulent Fluctuations 242

ê∗2
ê∗1

x2

x1

Figure 27.4: Eigenvectors ê∗1 and ê∗2 . x1 and x2 denote streamwise and wall-normal
direction, respectively.
√
λ(1) ê∗1

√
λ(2) ê∗2
x2

x1

Figure 27.5: Eigenvectors multiplied by the √ x1 and x2 denote streamwise

√ eigenvalues.
and wall-normal direction, respectively. λ(1) and λ(2) define RMS of (u′1 )aniso
and (u′2 )aniso , respectively.

Make sure that Matlab has defined the first eigenvector in the first or the third
quadrant, and the second eigenvector in the second or the fourth quadrant, see Fig. 27.4.
If not, change sign on some of the eigenvector components.

% switch sign on 12 and 21 to fix the above requirements

v1_new(2)=-v1_new(2);
v2_new(1)=-v2_new(1);

27.9.2 Synthetic fluctuations in the principal coordinate system

The equation for generating the synthetic fluctuations in the principal coordinate sys-
tem, (u′i )aniso , is similar to Eq. 27.3. The difference is that we now do it in the trans-
formed coordinate system, and hence we have to involve the eigenvector matrix, Rij ,
and the eigenvalues, λ(i) . The equation reads [167, 173]
N
X
u′ aniso (x∗ ) = 2 ûn cos(κ∗n · x∗ + ψ n )σ ∗n
n=1 (27.28)
r
1 p
κ∗ = κ∗i = Rji κj , σ ∗n = σi∗n = λ(i) Rji σjn
λ(i)
The superscript ∗ denotes the principal coordinates.
28. Overview of LES, hybrid LES-RANS and URANS models 243

28 Overview of LES, hybrid LES-RANS and URANS

models
IRECT Numerical Simulation (DNS) is the most accurate method available. In DNS
D DNS we solve the unsteady, 3D Navier-Stokes equations (see Eq. 2.8) numerically
without any turbulence model. This gives the exact solution of the flow field in time
and space. We can afford to do DNS only for low Reynolds number. The higher the
Reynolds number gets, the finer the grid must be because the smallest turbulent length
and time scales, which we must resolve in DNS, decrease with Re−3/4 and Re−1/2 ,
respectively, see Eq. 5.16. Hence the cell size in each coordinate direction of our CFD
grid must decrease with Re−3/4 and the time step must decrease with Re−1/2 . Let’s
take an example. If the Reynolds number increases by a factor of two, the number of
cells increases by

∆Re3/4 ∆Re3/4 ∆Re3/4 ∆Re1/2 = ∆Re11/4 = 211/4 = 6.7 (28.1)

x-direction y-direction z-direction time

Above we assume that the lengthscales are reduced when the Reynolds numbers
is increased. This implies that we assume that the Reynolds number is increased due
to an increase in velocity or a decrease in viscosity. We can, of course, also consider
the change of Reynolds number by changing the size of the object. For example, it is
affordable to compute the flow around a small car such as those we played with as kids
(for this car of, say, length of 5cm, the Reynolds number is very small). As we increase
the size of the car we must increase the number of cells (the smallest cells cannot be
enlarged, because the smallest turbulent scales will not increase). Also the time step
cannot be increased, but we must compute longer time (i.e. increase the number of
timesteps) in order to capture the largest time scales (assuming that the velocity of the
small and the large car is the same).
Having realized that DNS is not feasible, we turn to LES, see Section 18. Here, LES
the smallest scales are modeled, and only the eddies that are larger than the grid are
resolved by the (filtered) Navier-Stokes. With LES, we can make the smallest grid cells
somewhat larger (the cell side, say, 2 − 3 times larger).
However, it is found that LES needs very fine resolution near walls, see Section 21.
To find an approximate solution to this problem we use RANS near the walls and LES
away from the walls. The models which we have looked at are DES (Section 20),
hybrid LES-RANS (Section 21), SAS (Section 22) and PANS (Section 23); see also
Section 25 where PANS and Zonal PANS are discussed.
As stated above, the LES must at high Reynolds number be combined with a
URANS treatment of the near-wall flow region. In the literature, there are different
methods for bridging this problem such as Detached Eddy Simulation (DES) [127,
144, 153] hybrid LES/RANS [176] and Scale-Adapted Simulations (SAS) [177, 178]
(for a review, see [179]). The two first classes of models take the SGS length scale
from the cell size whereas the last (SAS) involves the von Kármán lengthscale.
The DES, hybrid LES/RANS and the SAS models have one thing in common: in
the LES region, the turbulent viscosity is reduced. This is achieved in different ways. In
some models, the turbulent viscosity is reduced indirectly by reducing the dissipation
term in the k equation, see Eq. 20.7, as in two-equation DES [180]. In other models,
such as the two-equation XLES model [131] and in the one-equation hybrid LES-
RANS [89], it is accomplished by reducing the length scale in both the expression for
28. Overview of LES, hybrid LES-RANS and URANS models 244

the turbulent viscosity as well as for the dissipation term in the k equation, see Eq. 21.2
and Table 21.1.
In the partially averaged Navier-Stokes (PANS) model [133] and the Partially In-
tegrated Transport Model (PITM) [135, 157], the turbulent viscosity is reduced by de-
creasing the destruction term in the dissipation (ε) equation which increases ε, see
Eq. 23.17. This decreases the turbulent viscosity in two ways: first, the turbulent vis-
cosity is reduced because of the enhancement of ε (νt = cµ k 2 /ε, see Eq. 23.8), and,
second, the turbulent kinetic energy (k) decreases because of the increased dissipation
term, ε. In the SAS model based on the k − ω model, the turbulent viscosity is reduced
by an additional source term, PSAS , see Eq. 22.5, in the ω equation. The source term
is activated by resolved turbulence; in steady flow it is inactive. When the momentum
equations are in turbulence-resolving mode, PSAS increases which increases ω. This
decreases the turbulent viscosity in two ways: first, directly, because ω appears in the
denominator in the expression for the turbulent viscosity, νt , and, second, because k is
reduced due to the increased dissipation term β ∗ kω.
The PANS model and the PITM models are very similar to each other although
their derivations are completely different. The only difference in the models is that
in the PANS model the turbulent diffusion coefficients in the k and ε equations are
modified. These two models do not use the filter width, and can hence be classified as
URANS models. On the other hand, a large part of the turbulence spectrum is usually
resolved which is in contrast to standard URANS models. PANS and PITM models
have in [179] been classified as second-generation URANS models, or 2G-URANS
models.
A short description of the models are given here.
DES. A RANS models is used near the walls and LES is used away from the walls.
The interface is usually defined automatic. In the original DES the entire bound-
ary layer is covered by RANS. However, when the grid is refined in streamwise
and spanwise directions, the interface moves closer to the wall. When a large part
of the boundary layer is covered by LES, it is called WM-LES (Wall-Modeled WM-LES
LES). The LES lengthscale is the filterwidth.
Hybrid LES-RANS. The difference between DES and hybrid LES-RANS is that the
original DES covers the entire boundary layer by RANS whereas hybrid LES-
RANS treats most of the boundary layer in LES mode. Hybrid LES-RANS and
WM-LES can be considered to be the same thing. The LES lengthscale is the
filterwidth.

PANS, PITM. These models are able to operate both in LES and RANS mode. In
LES mode the models do not use the filterwidth as a lengthscale. Hence they are
usually defined as an URANS model (defined below). Since the models usually
aim at resolving a substantial part of the turbulence spectrum, they can be defined
as a second-generation URANS model (a 2G-URANS model [179]).
SAS. This is also a model that can operate both in LES and RANS mode. In unsteady
mode the model does not use the filterwidth as a lengthscale. In unsteady mode
this model usually resolved less turbulence than the other models mentioned
above; hence it can be classified as an URANS model (first generation).
URANS. A RANS model is used in unsteady mode. In unsteady mode the model does
not use the filterwidth as a lengthscale. Unless the flow is prone to go unsteady,
28. Overview of LES, hybrid LES-RANS and URANS models 245

the results will be steady (i.e. same as RANS). Usually a RANS model developed
for steady flow is used. Only a small part of the turbulence is resolved.
The models listed above can be ranked in terms of accuracy and CPU cost:

1. Hybrid LES-RANS, PANS, PITM, WM-LES. Highest accuracy and CPU cost
2. DES.
3. SAS.
4. URANS. Lowest accuracy and CPU cost
29. Best practice guidelines (BPG) 246

29 Best practice guidelines (BPG)

In the early days of CFD, different CFD codes used to give different results. Even
if the same grid and the same turbulence model were used, there could be substantial
differences between the results. The reasons to these differences could be that the
turbulence model was not implemented in exactly the same way in the two codes, or
that the discretization scheme in one code was more diffusive than in the other. There
could be small differences in the implementation of the boundary conditions in the two
codes.
Today the situation is much improved. Two different CFD codes usually give the
same results on the same grid. The main reason for this improved situation is because
of workshops and EU projects where academics, engineers from industry and CFD
software vendors regularly meet and discuss different aspects of CFD. Test cases with
mandatory grids, boundary conditions, turbulence models etc are defined and the par-
ticipants in the workshops and EU projects carry out CFD simulations for these test
cases. Then they compare and discuss their results.

29.1 EU projects
Four EU projects in which the author has taken part can be mentioned

LESFOIL: Large Eddy Simulation of Flow Around Airfoils

http://www.tfd.chalmers.se/˜lada/projects/lesfoil/proright.html
FLOMANIA: Flow Physics Modelling: An Integrated Approach
http://cfd.mace.manchester.ac.uk/flomania/
DESIDER: Detached Eddy Simulation for Industrial Aerodynamics
http://cfd.mace.manchester.ac.uk/desider
ATAAC: Advanced Turbulence Simulation for Aerodynamic Application Challenges
http://cfd.mace.manchester.ac.uk/ATAAC/WebHome

29.2 Ercoftac workshops

Workshops are organized by Ercoftac (European Research Community On Flow, Turbulence
And Combustion). The Special Interest Group Sig15 is focused on evaluating turbu-
lence models. The outcome from all workshop are presented
here
http://www.ercoftac.org/fileadmin/user upload/bigfiles/sig15/database/index.html
Ercoftac also organizes workshops and courses on Best Practice Guidelines. The
publication Industrial Computational Fluid Dynamics of Single-Phase Flows can be
ordered on
Ercoftac www page
http://www.ercoftac.org/publications/ercoftac best practice guidelines/single-phase flows spf/
29.3. Ercoftac Classical Database 247

29.3 Ercoftac Classical Database

A Classical Database, which includes some 100 experimental investigations, can be
found at
Ercoftac’s www page
http://www.ercoftac.org/products and services/classic collection database

29.4 ERCOFTAC QNET Knowledge Base Wiki

The QNET is also the responsibility of Ercoftac. Here you find descriptions of how
CFD simulations of more than 60 different flows were carried out. The flows are di-
vided into

Application Areas. These are sector disciplines such as Built Environment, Chemical
and Process Engineering, External Aerodynamics, Turbomachinery, Combustion
and Heat Transfer etc. Each Application Area is comprised of Application Chal-
lenges. These are realistic industrial test cases which can be used to judge the
competency and limitations of CFD for a given Application Area.
Underlying Flow Regimes. These are generic, well-studied test cases capturing im-
portant elements of the key flow physics encountered across the Application Ar-
eas.

For more information, visit

ERCOFTAC QNET Knowledge Base Wiki
http://www.ercoftac.org/products and services/wiki/
A. TME225: ǫ − δ identity 248

A TME225: ǫ − δ identity
HE ǫ − δ identity reads
T ǫinm ǫmjk = ǫmin ǫmjk = ǫnmi ǫmjk = δij δnk − δik δnj
In Table A.1 the components of the ǫ − δ identity are given.

i n j k ǫinm ǫmjk δij δnk − δik δnj

1 2 1 2 ǫ12m ǫm12 = ǫ123 ǫ312 = 1 · 1 = 1 1−0=1
2 1 1 2 ǫ21m ǫm12 = ǫ213 ǫ312 = −1 · 1 = −1 0 − 1 = −1
1 2 2 1 ǫ12m ǫm21 = ǫ123 ǫ321 = 1 · −1 = −1 0 − 1 = −1

1 3 1 3 ǫ13m ǫm13 = ǫ132 ǫ213 = −1 · −1 = 1 1−0=1

3 1 1 3 ǫ31m ǫm13 = ǫ312 ǫ213 = 1 · −1 = −1 0 − 1 = −1
1 3 3 1 ǫ13m ǫm31 = ǫ132 ǫ231 = −1 · 1 = −1 0 − 1 = −1

2 3 2 3 ǫ23m ǫm23 = ǫ231 ǫ123 = 1 · 1 = 1 1−0=1

3 2 2 3 ǫ32m ǫm23 = ǫ321 ǫ123 = −1 · 1 = −1 0 − 1 = −1
2 3 3 2 ǫ23m ǫm32 = ǫ231 ǫ132 = 1 · −1 = −1 0 − 1 = −1

Table A.1: The components of the ǫ − δ identity which are non-zero.

B. TME225 Assignment 1 in 2016: laminar flow in a channel 249

B TME225 Assignment 1 in 2016: laminar flow in a

channel
OU will get results of a developing two-dimensional channel flow (i.e. flow be-
Y tween two parallel plates), see Fig. B.1. The flow is steady and incompressible.
The simulations have been done with Calc-BFC [181]. The inlet boundary condition
(left boundary) is v1 = Vin = 0.9. The height of the channel is h = 0.01m and
L = 0.6385m; the fluid is air of 20o C. You will use Matlab to analyze the data. You
can also use Octave. Octave is a Matlab clone which can be downloaded for free.

• First, find out and write down the governing equations (N.B:. you cannot assume
that the flow is fully developed).

From the course www page http://www.tfd.chalmers.se/˜lada/MoF/,

download the data file channel flow data.dat and the m-file channel flow.m
which reads the data and plot some results. Open Matlab and execute channel flow.
Open channel flow.m in an editor and make sure that you understand it. There
are three field variables, v1 , v2 and p; the corresponding Matlab arrays are v1 2d,
v2 2d and p 2d. The grid is 199 × 28, i.e. ni = 199 grid points in the x1 direction
and nj = 28 grid points in the x2 direction. The field variables are stored at these grid
points. We denote the first index as i and the second index as j, i.e. v1 2d(i,j).
Hence in

v1 2d(:,1) are the v1 values at the lower wall;

v1 2d(:,nj) are the v1 values at the upper wall;
v1 2d(1,:) are the v1 values at the inlet;
v1 2d(ni,:) are the v1 values at the outlet;

The work should be carried out in groups of two (you may also do it on your
own, but we don’t recommend it). At the end of this Assignment the group should
write and submit a report (in English). Divide the report into sections corresponding
to the sections B.1 – B.9. In some sections you need to make derivations; these should
clearly be described and presented. Present the results in each section with a figure
(or a numerical value). The results should also be discussed and – as far as you can –
explained.
It is recommended (but the not required) that you use LATEX(an example of how to
write in LATEXis available on the course www page). It is available on Linux. On Win-
dows you can use, for example, ShareLatex (https://www.sharelatex.com)
Lyx (www.lyx.org) or MikTex (www.miktex.org).

B.1 Fully developed region

Fully developed conditions mean that the flow does not change in the streamwise di-
rection, i.e. ∂v1 /∂x1 = 0. If we define “fully developed” as the location where the
velocity gradient in the center becomes smaller than 0.01, i.e. |∂v1 /∂x1 | < 0.01, how
long distance from the inlet does the flow become fully developed?
Another way to define fully developed conditions can be the x1 position where the
centerline velocity has reached, for example, 99% of its final value. What x1 value do
you get?
B.2. Wall shear stress 250

h
x2
x1

Figure B.1: Flow between two plates (not to scale).

In Section 3.2.2, a distance taken from the literature is given. How well does this
agree with your values?
In the fully developed region, compare the velocity profile with the analytical pro-
file (see Section 3.2.2).
Look at the vertical velocity component, v2 . What value should it take in the fully
developed region (see Section 3.2.2)? What value does it take (at x2 = h/4, for
example)?

B.2 Wall shear stress

On the lower wall, the wall shear stress, τw,L (index L denotes Lower), is computed as

∂v1 
τw,L ≡ τ21,w,L = µ (B.1)
∂x2 L

Recall that τ12 = µ(∂v1 /∂x2 + ∂v2 /∂x1 ) (see Eqs. 2.4 and 1.9) but at the wall
∂v2 /∂x1 = 0; Skk = 0 because of the continuity equation, Eq. 2.3. Plot τw,L ver-
sus x1 . Why does it behave as it does?
Now we will compute the wall shear stress at the upper wall, τw,U . If you use
Eq. B.1, you get the incorrect sign. Instead, use Cauchy’s formula (see Fig. 1.3 and [1],
Chapt. 4.2)
(n̂)
ti = τji nj (B.2)
which is a general way to compute the stress vector on a surface whose (outward point-
ing) normal vector is n̂ = nj . The expression for τij can be found in Eqs. 1.9 and
2.4; recall that the flow in incompressible. On the top wall, the normal vector points
out from the surface (i.e. nj = (0, −1, 0)). Use Eq. B.2 to compute the wall shear
stress at the upper wall. Plot the two wall shear stresses in the same figure. How do
they compare? In the fully developed region, compare with the analytical value (see
Eq. 3.30).

B.3 Inlet region

In the inlet region the flow is developing from its inlet profile (v1 = V = 0.9) to
the fully developed profile somewhere downstream. The v1 velocity is decelerated
in the near-wall regions, and hence the v1 velocity in the center must increase due
to continuity. Plot v1 in the center and near the wall as a function of x1 . Plot also
B.4. Wall-normal velocity in the developing region 251

∂v1 /∂x1 . If you, for a fixed x1 , integrate v1 , i.e.

Z h
ξ(x1 ) = v1 (x1 , x2 )dx2
0

what do you get? How does ξ(x1 ) vary in the x1 direction? How should it vary?

B.4 Wall-normal velocity in the developing region

In Section B.3 we found that, in the developing region, v1 near the walls decreases for
increasing x1 . What about v2 ? How do you explain the behaviour of v2 ?

B.5 Vorticity
Do you expect the flow to be irrotational anywhere? Let’s find out by computing the
vorticity vector ωi , see Section 1.4 (note that only one component of ωi is non-zero).
Plot it in the fully developed region as ω3 vs. x2 . Where is it largest? Plot the vorticity
also in the inlet and developing regions; what happens with the vorticity in the inlet
region? Now, is the flow rotational anywhere? Why? Why not?

B.6 Deformation
In Section 1.6, we divided the velocity gradient into a strain-rate tensor, Sij , and a vor-
ticity tensor, Ωij . Since the flow is two-dimensional, we have only two off-diagonal
terms (which ones?). Plot and compare one of the off-diagonal term of Sij and Ωij .
Where are they largest? Why? What is the physical meaning of Sij and Ωij , re-
spectively? Compare Ωij with the vorticity, ωi , you plotted in Section B.5. Are they
similar? Any comment?

B.7 Dissipation
Compute and plot the dissipation, Φ = τji ∂vi /∂xj , see Eq. 2.13. What is the physical
meaning of the dissipation? Where do you expect it to be largest? Where is it largest?
Any difference it its behaviour in the inlet region compared to in the fully developed
region?
The dissipation appears as a source term in the equation for internal energy, see
Eq. 2.13. This means that dissipation increases the internal energy, i.e. the temperature.
This is discussed in some detail at p. 27.
Use Eq. 2.15 to compute the temperature increase that is created by the flow (i.e. by
dissipation). Start by integrating the dissipation over the entire computational domain.
Next, re-write the left side on conservative form (see Section 2.4) and then apply the
Gauss divergence theorem. Assume that the upper and the lower wall are adiabatic;
furthermore we can neglect the heat flux by conduction, q1 , (see Eq. 2.12) at the inlet
and outlet. Now you can compute the increase in bulk temperature, Tb , from inlet to
outlet. The bulk temperature is defined as
Rh
v1 T dx2
Tb = R0 h (B.3)
0 v1 dx2
B.8. Eigenvalues 252

When you compute the convective flux in Eq. 2.10 at the outlet, for example, you get
Z h
v1 T dx2
0

which indeed is very similar to the bulk temperature in Eq. B.3.

B.8 Eigenvalues
Compute and plot the eigenvalues of the viscous stress tensor, τij . Use the Matlab
command eig. If you have computed the four elements of the τij matrix you can use
the following commands:
tau=[tau_11 tau_12; tau_21 tau_22];
[n,lambda]=eig(tau);
where n and lambda denote eigenvalues and eigenvectors, respectively. Note that
tau 11, tau 12, tau 21, tau 22 are scalars and hence the coding above must
be inserted in for loops.
What is the physical meaning of the eigenvalues (see Chapter 1.8)? Pick an x1 loca-
tion where the flow is fully developed. Plot one eigenvalue as a x− y graph (eigenvalue
versus x2 ). Plot also the four stress components, τij , versus x2 . Is (Are) anyone(s) neg-
ligible? How does the largest component of τij compare with the largest eigenvalue?
Any thoughts? And again: what is the physical meaning of the eigenvalues?

B.9 Eigenvectors
Compute and plot the eigenvectors of τij . Recall that at each point you will get two
eigenvectors, perpendicular to each other. It is enough to plot one of them. An eigen-
vector is, of course, a vector. Use the Matlab command quiver to plot the field of the
eigenvectors. Recall that the sign of the eigenvector is not defined (for example, both
v̂1 and −v̂1 in Fig. 1.11 at p. 23 are eigenvectors). Try to analyze why the eigenvectors
behave as they do.
C. TME225 Assignment 1 in 2017: laminar flow in a boundary layer 253

C TME225 Assignment 1 in 2017: laminar flow in a

boundary layer
OU will get results of a developing two-dimensional boundary-layer flow (i.e. flow
Y along a flat plate, see Fig. C.1). The flow is steady and incompressible. The
simulations have been done with CALC-BFC [181] using the QUICK discretizatin
scheme [182]. The height of the domain is H = 1.289m, L = 2.461m and x0 =
0.19m; the fluid is air of 20o C and V1,∞ = 1m/s . The mesh has 252 × 200 grid
points, i.e. ni = 252 points in the x1 direction and nj = 200 points in the x2 direction.

Inlet boundary, x1 = −x0 (i=1)

v1 = V1,∞ = 1, v2 = 0.

Lower boundary, x2 = 0 (j=1)

x1 ≤ 0, i≤ i=i lower sym: v2 = ∂v1 /∂x2 = 0, i.e. a symmetry boundary con-
dition
x1 > 0, i> i=i lower sym: v1 = v2 = 0, i.e. a wall

Outlet, x1 = L (i=ni)
Neumann boundary conditions are used, i.e. ∂v1 /∂x1 = ∂v2 /∂x1 = 0.

Upper boundary, x2 = H (j=nj)

v2 = ∂v1 /∂x2 = 0, i.e. a symmetry boundary condition

You will use Matlab to analyze the data. You can also use Octave. Octave is a
Matlab clone which can be downloaded for free.

• First, find out and write down the governing equations

From the course www page http://www.tfd.chalmers.se/˜lada/MoF/,

download the data files boundary layer data.dat, xc.dat, yc.dat and blasius.dat
and the m-file boundary layer.m which reads the data and plot some results.
Download also the file with the Blasius solution, blasius.dat. Open Matlab and
execute boundary layer.m
Open boundary layer.m in an editor and make sure that you understand it.
There are three field variables, v1 , v2 and p; the corresponding Matlab arrays are u2d,
v2d and p2d. The field variables are stored at the cell centers denoted xp,yp. We
denote the first index as i and the second index as j, i.e. u2d(i,j). Hence

• u2d(:,1) are the v1 values at the lower boundary

– i≤ i=i lower sym: symmetric boundary condition

– i> i=i lower sym: wall

• u2d(:,nj) are the v1 values at the upper boundary;

C.1. Velocity profiles 254

V∞

H V∞ boundary layer
x2 x1

δ(x1 )

x0 L

Figure C.1: Flow along a flat plate (not to scale).

• u2d(1,:) are the v1 values at the inlet;

• u2d(ni,:) are the v1 values at the outlet;

The work should be carried out in groups of two (you may also do it on your own,
but we don’t recommend it). At the end of this Assignment the group should write
and submit a report (in English). Divide the report into sections corresponding to the
sections C.1 – C.10. In some sections you need to make derivations; these should
clearly be described and presented. Present the results in each section with a figure
(or a numerical value). The results should also be discussed and – as far as you can –
explained.
It is recommended (but the not required) that you use LATEX(an example of how to
write in LATEXis available on the course www page). It is available on Linux. On Win-
dows you can use, for example, ShareLatex (https://www.sharelatex.com)
Lyx (www.lyx.org) or MikTex (www.miktex.org).

C.1 Velocity profiles

The flat plate starts at x1 = 0. What happens with the streamwise velocity, v1 , near
the wall as you go downstream from x1 = 0? Why does v1 behave as it does? What
happens with v2 ? Does the behavior of v2 has anything to do with v1 ?
Plot profiles of v1 vs. x2 at a couple of x1 locations.
Now plot v1 /V1,∞ vs. ξ, see Eq. 3.47 and compare with Blasius solution, see
Table 3.1 (the Blasius solution is loaded in boundary layer.m). Note that when
you compute ξ you must use x1 , x2 and ν from the Navier-Stokes solution. Plot also
v2 /V1,∞ .
Compare v1 /V1,∞ and v2 /V1,∞ with Blasius solution; v1,Blasius /V1,∞ is taken
from the data file and v2,Blasius /V1,∞ is obtained from Eqs. 3.43 and 3.49.

C.2 Boundary layer thickness

Look at the v1 profiles vs. x2 you plotted at different x1 locatations in Section C.1.
The boundary layer gets thicker and thicker as you go downstream, right? The bound-
ary layer thickness, δ, can be defined in many ways. One way is the boundary layer
thickness at the x2 value where the local velocity is 99% of the freestream velocity;
this boundary layer thickness is denoted δ99 .
C.3. Velocity gradients 255

Compute and plot δ99 vs. x1 and compare it with the Blasius solution
 1/2
νx1
δ99,Blasius = 5 (C.1)
V1,∞

The disadvantage of δ99 is that this definition is entirely dependent on the behavior
of the velocity profile near the edge of the boundary layer. There are two other, more
stable, definitions of boundary layer thickness, the displacement thickness, δ ∗ , and the
momentum thickness, θ; they are defined as
Z ∞ 
∗ v1
δ = 1− dx2
0 V1,∞
Z ∞   (C.2)
v1 v1
θ= 1− dx2
0 V1,∞ V1,∞

The upper limit of the integrals is infinity. If you look carefully at the velocity profiles
which you plotted in Section C.1, you find that v1 has a maximum outside the boundary
layer. We don’t want to carry out the integration in Eq. C.2 outside of this maximum,
∗
because this will give an incorrect estimate of δBlasius and θBlasius . Hence, terminate
the integration when u2d(i,j+1) < u2d(i,j). Plot δ ∗ and θ vs. x1 and compare
with the Blasius solution
 1/2
∗ νx1
δBlasius = 1.721
V1,∞
 1/2
νx1
θBlasius = 0.664
V1,∞

C.3 Velocity gradients

Compute and plot ∂v1 /∂x2 vs. x2 at a couple of x1 locations. How does this velocity
derivative change as you move downstream along the plate? Do you see any connection
with the change of the boundary layer thickness with x1 and the change of ∂v1 /∂x2
with x1 ?
In boundary layer approximations, we assume that v2 ≪ v1 and ∂v1 /∂x2 ≫
∂v1 /∂x1 : are these assumptions satisfied in this flow?

C.4 Skinfriction
The skinfriction, Cf , is a dimensionless wall shear stress defined as
τw
Cf = 1 2
(C.3)
2 ρV1,∞

The wall shear stress, τw , is the same as τ12 = τ21 , see Eq. 2.4, at the wall. When
using Eq. 2.4, recall that the flow is incompressible. Furthermore, what about ∂v2 /∂x1
in Eq. 2.4? how large is it at the wall? Plot Cf vs. x1 and compare it with the Blasius
solution
0.664
Cf,Blasius = p
V1,∞ x1 /ν
Cf decreases vs. x1 and δ increases vs. x1 : do you see any connection?
C.5. Vorticity 256

C.5 Vorticity
Do you expect the flow to be irrotational anywhere? Let’s find out by computing the
vorticity vector, ωi , see Section 1.4 (note that only one component of ωi is non-zero).
Plot the vorticity above the plate vs. x2 at a couple of x1 locations. Where is it largest?
Plot the vorticity also upstream of the plate. Now, is the flow rotational anywhere?
Why? Why not?

C.6 Deformation
In Section 1.6, we divided the velocity gradient into a strain-rate tensor, Sij , and a vor-
ticity tensor, Ωij . Since the flow is two-dimensional, we have only two off-diagonal
terms (which ones?). Plot and compare one of the off-diagonal term of Sij and Ωij .
Where are they largest? Why? What is the physical meaning of Sij and Ωij , respec-
tively? Compare Ωij with the vorticity, ωi , which you plotted in Section C.5. Are they
similar? Any comment?

C.7 Dissipation
Compute and plot the dissipation, Φ = τji ∂vi /∂xj , see Eq. 2.13. What is the physical
meaning of the dissipation? Where do you expect it to be largest? Where is it largest?
How large is is upstream the plate?
The dissipation appears as a source term in the equation for internal energy, see
Eq. 2.13. This means that dissipation increases the internal energy, i.e. the temperature.
This is discussed in some detail at p. 27.
Use Eq. 2.15 to compute the temperature increase that is created by the flow (i.e. by
dissipation). Start by integrating the dissipation over the entire computational domain.
Next, re-write the left side on conservative form (see Section 2.4) and then apply the
Gauss divergence theorem. Assume that the lower (x2 = 0) and the upper (x2 =
H) boundaries are adiabatic; furthermore we can neglect the heat flux by conduction,
q1 , (see Eq. 2.12) at the inlet and outlet. Now you can compute the increase in bulk
temperature, Tb , from inlet to outlet. The bulk temperature is defined as
Rh
v1 T dx2
Tb = R0 h (C.4)
0 v1 dx2

When you compute the convective flux in Eq. 2.10 at the outlet, for example, you get
Z h
v1 T dx2
0

which indeed is very similar to the bulk temperature in Eq. C.4.

C.8 Eigenvalues
Compute and plot the eigenvalues of the viscous stress tensor, τij . Use the Matlab
command eig. If you have computed the four elements of the τij matrix you can use
the following commands:
tau=[tau_11 tau_12; tau_21 tau_22];
[n,lambda]=eig(tau);
C.9. Eigenvectors 257

where n and lambda denote eigenvalues and eigenvectors, respectively. Note that
tau 11, tau 12, tau 21, tau 22 are scalars and hence the coding above must
be inserted in for loops.
What is the physical meaning of the eigenvalues (see Chapter 1.8)? Pick an x1
location above the plate. Plot one eigenvalue as a x − y graph (eigenvalue versus x2 ).
Plot also the four stress components, τij , versus x2 . Is (Are) anyone(s) negligible?
How does the largest component of τij compare with the largest eigenvalue? Any
thoughts? And again: what is the physical meaning of the eigenvalues?

C.9 Eigenvectors
Compute and plot the eigenvectors of τij . Recall that at each point you will get two
eigenvectors, perpendicular to each other. It is enough to plot one of them. An eigen-
vector is, of course, a vector. Use the Matlab command quiver to plot the field of the
eigenvectors. Recall that the sign of the eigenvector is not defined (for example, both
v̂1 and −v̂1 in Fig. 1.11 at p. 23 are eigenvectors). Try to analyze why the eigenvectors
behave as they do.

C.10 Terms in the v1 equation

The Navier-Stokes equation for v1 includes only three terms, see Eq. 3.41. The left
side and the right side of this equation must be equal. Let’s verify that. Plot the three
terms vs. x2 at a couple of x1 locations. Plot also the sum of the terms, i.e. check if
the left side and the right side of Eq. 3.41 are equal: are they? If not, what could be
the reason? Recall that in Eq. 3.41 we assume that ∂p/∂x1 = 0. Maybe it is not zero?
Plot it and check! If it’s not zero, what could be the reason?
How large is ∂ 2 v1 /∂x22 at the wall? How large should it be? (Hint: which other
terms in the v1 equation are non-zero?). How large is the gradient of vorticity, ∂ω/∂x2 ,
at the wall? Look at the discussion in connection to Eq. 4.25.
D. TME225: Fourier series 258

V2

(V2 |T)

V1
(V1 |T)

Figure D.1: Scalar product.

D TME225: Fourier series

a brief introduction to Fourier series extracted from [183] is given.
H ERE

D.1 Orthogonal functions

Consider three vectors, V1 , V2 , V3 , in physical space which form an orthogonal base
in R3 (i.e. their scalar products are zero). Let us call them basis functions. Any vector,
T, in R3 can now be expressed in these three vectors, i.e.

T = c1 V1 + c2 V2 + c3 V3 (D.1)

see Fig. D.1. Now define the scalar product of two vectors, a and b, as a · b = (a|b).
The coordinates, ci , can be determined by making a scalar product of Eq. D.1 and Vi
which gives

(T|Vi ) = (c1 V1 |Vi ) + (c2 V2 |Vi ) + (c3 V3 |Vi )

= (c1 V1 |V1 ) + (c2 V2 |V2 ) + (c3 V3 |V3 ) (D.2)
2 2 2 2
= c1 |V1 | + c2 |V2 | + c3 |V3 | = ci |Vi |

where |Vi | denotes the length of Vi ; the second line follows because of the orthogo-
nality of Vi . Hence the coordinates, ci , are determined by

ci = (T|Vi )/|Vi |2 (D.3)

D.2. Trigonometric functions 259

Now let us define an infinite (∞-dimensional) functional space, B, with orthogonal

basis functions {g}∞1 . The “scalar product” of two functions, f and gn , is defined as
Z b
(f |gn ) = f (x)gn (x)dx (D.4)
a

Then, in a similar way to Eq. D.1, any function can, over the interval [a, b], be expressed
as
X∞
f= cn g n (D.5)
n=1

As above, we must now find the “coordinates”, cn (cf. the coordinates, ci , in Eq. D.1).
Multiply, as in Eq. D.2, f with the basis functions, gi , i.e.
∞
X
(f |gi ) = cn (gn |gi ) (D.6)
n=1

Since we know that all gn are orthogonal, Eq. D.6 is non-zero only if i = n, i.e.

(f |gi ) = (c1 g1 |gi ) + (c2 g2 |gi ) . . . ci (gi |gi ) . . . ci+1 (gi+1 |gi ) . . . =
(D.7)
= ci (gi |gi ) = ci ||gi ||2

Similar to Eq. D.3, the “coordinates” can be found from (switch from index i to n)

cn = (f |gn )/||gn ||2 (D.8)

The “coordinates”, cn , are called the Fourier coefficients to f in system {g}∞ 1 and
||gn || is the “length” of gn (cf. |Vi | which is the length of Vi in Eq. D.3), i.e.

Z !1/2
b
||gn || = (gn |gn )1/2 = gn (x)gn (x)dx (D.9)
a

Let us now summarize and compare the basis functions in physical space and the
basis functions in functional space.
1. Any vector in R3 can be expressed in 1. Any function in [a, b] can be ex-
the orthogonal basis vectors Vi pressed in the orthogonal basis func-
tions gn
2. The length of the basis vector, Vi , is 2. The length of the basis function, gn ,
|Vi | is ||gn ||
3. The coordinates of Vi are computed 3. The coordinates of gn are computed
as ci = (T|Vi )/|Vi |2 as cn = (f |gn )/||gn ||2

D.2 Trigonometric functions

Here we choose gn as trigonometric functions which are periodic in [−π, π]. The
question is now how to choose the orthogonal function system {g}∞ 1 on the interval
[−π, π]. In mathematics, we usually start by doing an intelligent “guess”, and then we
prove that it is correct. So let us “guess” that the trigonometric series

[1, sin x, cos x, sin(2x), . . . , sin(nx), cos(nx), . . .] (D.10)

D.2. Trigonometric functions 260

is an orthogonal system. The function system in Eq. D.10 can be defined as


φk (x), for n = 2k = 2, 4, . . .
gn (x) = (D.11)
ψk (x), for n = 2k + 1 = 1, 3, . . .

where φk (x) = sin(kx) (k = 1, 2, . . .) and ψk (x) = cos(kx) (k = 0, 1, . . .). Now we

need to show that they are orthogonal, i.e. that the integral of the product of any two
functions φk and ψk is zero on B[−π, π] and we need to compute their “length” (i.e.
their norm).

Orthogonality of ψn and ψk
Z π Z
1 π
(ψn |ψk ) = cos(nx) cos(kx)dx = [cos((n + k)x) + cos((n − k)x)] dx
−π 2 −π
 π
1 1 1
= sin((n + k)x) + sin((n − k)x) = 0 for k 6= n
2 n+k n−k −π
(D.12)

D.2.1 “Length” of ψk
Z π  π
2 2 x 1
(ψk |ψk ) = ||ψk || = cos (kx)dx = + sin(2kx) = π for k > 0
−π 2 4k −π
Z π
2
(ψ0 |ψ0 ) = ||ψ0 || = 1 · dx = 2π
−π
(D.13)

D.2.2 Orthogonality of φn and ψk

Z π Z
1 π
(φn |ψk ) = sin(nx) cos(kx)dx = [sin((n + k)x) + sin((n − k)x)] dx
−π 2 −π
 π
1 1 1
=− cos((n + k)x) + cos((n − k)x) =0
2 n+k n−k −π
(D.14)
because cos((n + k)π) = cos(−(n + k)π) and cos((n − k)π) = cos(−(n − k)π).

D.2.3 Orthogonality of φn and φk

Z π Z
1 π
(φn |φk ) = sin(nx) sin(kx)dx = [cos((n − k)x) − cos((n + k)x)] dx
−π 2 −π
 π
1 1 1
= sin((n − k)x) − sin((n + k)x) = 0 for k 6= n
2 n−k n+k −π
(D.15)

D.2.4 “Length” of φk
Z π  π
2 2 x 1
(φk |φk ) = ||φk || = sin (kx)dx = − sin(2kx) = π for k ≥ 1
−π 2 4k −π
(D.16)
D.3. Fourier series of a function 261

D.3 Fourier series of a function

Now that we have proved that {g}∞
1 in Eq. D.11 forms an orthogonal system of func-
tions, we know that we can express any periodic function, f (with a period of 2π) in
{g}∞1 as
X∞
f (x) = c + (an cos(nx) + bn sin(nx)) (D.17)
n=1
where x is a spatial coordinate. The Fourier coeffients are given by
Z
1 π
bn = (f |φn )/||φn ||2 = f (x) sin(nx)dx (D.18a)
π −π
Z π
1
an = (f |ψn )/||ψn ||2 = f (x) cos(nx)dx (D.18b)
π −π
Z π
1
c = (f |ψ0 )/||ψ0 ||2 = f (x)dx (D.18c)
2π −π

where n > 0. If we set c = a0 /2, then a0 is obtained from Eq. D.18b, i.e.
∞
a0 X
f (x) = + (an cos(nx) + bn sin(nx)) (D.19a)
2 n=1
Z
1 π
bn = (f |φn )/||φn ||2 = f (x) sin(nx)dx (D.19b)
π −π
Z
1 π
an = (f |ψn )/||ψn ||2 = f (x) cos(nx)dx (D.19c)
π −π

Note that a0 /2 corresponds to the average of f . Taking the average of f (i.e.

integrating f from −π to π) and dividing with the integration length, 2π, gives (see
Eq. D.19a) Z π
1 1 a0 a0
f¯ = f (x)dx = · 2π = (D.20)
2π −π 2π 2 2
Hence, if f¯ = 0 then a0 = 0.

D.4 Derivation of Parseval’s formula

Parseval’s formula reads
Z π X∞
π 2
(f (x))2 dx = a0 + π (a2n + b2n ) (D.21)
−π 2 n=1

We will try to prove this formula. Assume that we want to approximate the function
f as well as possible with an orthogonal series
∞
X
an g n (D.22)
n=1

Now we want to prove that the Fourier coefficients are the best choice to minimize the
difference
XN
||f − an gn || (D.23)
n=1
D.4. Derivation of Parseval’s formula 262

Later we will let N → ∞. Using the definition of the norm and the laws of scalar
product we can write
 !
XN N
X  XN
2 
||f − an gn || = f − an gn f − ak g k

n=1 n=1 k=1
N
X N
X N X
X N
= (f |f ) − an (f |gn ) − ak (f |gk ) + an ak (gn |gk ) = (D.24)
n=1 k=1 n=1 k=1
N
X N
X
= (f |f ) − 2 an (f |gn ) + a2n (gn |gn )
n=1 n=1

because of the orthogonality of the function system, {g}N1 . Expressing f in the second
term using the Fourier coefficients cn (see Eqs. D.5 and D.8) gives
N
X N
X
(f |f ) − 2 an cn (gn |gn ) + a2n (gn |gn )
n=1 n=1
N
X

= ||f ||2 + ||gn ||2 a2n − 2an cn (D.25)
n=1
N
X N
X
2
= ||f ||2 + ||gn ||2 (an − cn ) − ||gn ||2 c2n
n=1 n=1

The left side of Eq. D.24 is thus minimized if the coefficients an are chosen as the
Fourier coefficients, cn so that
N
X N
X
||f − an gn ||2 = ||f ||2 − ||gn ||2 c2n (D.26)
n=1 n=1

The left side must always be positive and hence

N
X Z π
||gn ||2 c2n ≤ ||f ||2 = (f (x))2 dx for all N (D.27)
n=1 −π

As N is made larger, the magnitude of the left side increases, and its magnitude gets
closer and closer to that of the right side, but it will always stay smaller than ||f ||2 .
This means that the series on the left side is convergent. Using the Fourier coefficients
in Eq. D.19 and letting N → ∞ it can be shown that we get equality of the left and
right side, which gives Parseval’s formula,
Z π N
X   N
X
2 2 a20
||f || ≡ (f (x)) dx = ||gn ||2 c2n = ||ψ0 || + ||ψn ||a2n + ||φn ||b2n
−π n=1
2 n=1
  N
X ∞
X
a20 π 2
= 2π +π a2n + b2n = a +π (a2n + b2n )
2 n=1
2 0 n=1

Note that 2π and π on the second line are the “length” of ||gn ||, i.e. the length of ||ψ0 ||,
||ψn || and ||φn || (see Sections D.2.1 and D.2.4).
Appendix N describes in detail how to create energy spectra from two-point corre-
lations.
D.5. Complex Fourier series 263

D.5 Complex Fourier series

Equation D.19 gives the Fourier series of a real function. It is more convenient to
express a Fourier series in complex variables even if the function f itself is real. On
complex form it reads
∞
X
f (x) = cn exp(ınx)) (D.28a)
n=−∞
Z π
1
cn = f (x) exp(−ınx)dx (D.28b)
2π −π

where the Fourier coefficients, cn , are complex. Below we verify that if f is real, then
Eq. D.28 is equivalent to Eq. D.19. The Fourier coefficients, cn , read – assuming that
f is real – according to Eq. D.28
Z π
1 1
cn = f (x)(cos(nx) − ı sin(nx))dx = (an − ıbn ), n > 0 (D.29)
2π −π 2

where an and bn are given by Eq. D.19. For negative n in Eq. D.28 we get
Z π
∗ 1 1
c−n = cn = f (x)(cos(nx) + ı sin(nx))dx = (an + ıbn ), n > 0 (D.30)
2π −π 2

where c∗n denotes the complex conjugate. For n = 0, Eq. D.28 reads
Z π
1 1
c0 = f (x)dx = a0 (D.31)
2π −π 2

see Eq. D.19. Inserting Eqs. D.29, D.30 and D.31 into Eq. D.28 gives
∞
1 1X
f (x) = a0 + (an − ıbn ) exp(ınx) + (an + ıbn ) exp(−ınx)
2 2 n=1
∞
1 1X
= a0 + (an − ıbn )(cos(nx) + ı sin(nx)) + (an + ıbn )(cos(nx) − ı sin(nx))
2 2 n=1
X∞ X∞
1 1
= a0 + an cos(nx) − ı2 bn sin(nx) = a0 + an cos(nx) + bn sin(nx)
2 n=1
2 n=1
(D.32)

which verifies that the complex Fourier series for a real function f is indeed identical
to the usual formulation in Eq. D.19 although the Fourier coefficients, cn , are complex.
One advantage of Eq. D.28 over the formulation in Eq. D.19 is that we don’t need any
special definition for the first Fourier coefficient, a0 . The trick in the formulation in
Eq. D.28 is that the imaginary coefficients for negative and positive n cancel whereas
the real coefficients add. This means that the real coefficients are multiplied by a factor
two except the first coefficient, a0 , which makes up for the factor 21 in front of a0 in
Eq. D.19.
E. TME225: Why does the energy spectrum, E , have such strange dimensions? 264

E TME225: Why does the energy spectrum, E, have

such strange dimensions?
HE energy spectrum, E, has the strange dimension v 2 /ℓ. The reason is that it is
T a spectral density so that the kinetic energy, k = vi′ vi′ , is computed by integrating
over all wavenumbers, see Eq. 5.10. The energy spectrum is a spectral density function
in a similar way as fv (v) in Eq. 7.2; the difference is that fv (v) in Eq. 7.2 is the first
moment. Equation 7.5 defines the second moment. The dimension of fv (v) in Eq. 7.2
and fv′ (v ′ ) in Eq. 7.5 is one over velocity.
Since we have chosen to express the energy spectrum as a function of the wavenum-
ber, dimension analysis gives E ∝ κ−5/3 , see Eqs. 5.12 and 5.13. A similar dimension
analysis for the kinetic energy of v ′ gives vκ′2 ∝ κ−2/3 . However, integrating vκ′2 over
all wavenumbers does not give any useful integral quantity.
The integral of the energy spectrum in the inertial region can be estimated as (see
Eqs. 5.10 and 5.13) Z κ2
2 5
k = CK ε 3 κ− 3 dκ (E.1)
κ1
We could also express E as a function of the turbulent length scale of the eddies, ℓκ
(κ = 2π/ℓκ , see Eq. 5.7). The energy spectrum is then integrated as
Z κ2 Z κ2
2 5
k= E(κ)dκ = CK ε 3 κ− 3 dκ
κ1 κ1 κ−2/3
[m2 ]
Z κ2 Z ℓ2
2 dκ 5 2 5
= CK ε 3 κ− 3
dℓκ = −CK 2πε 3 κ− 3 ℓ−2
κ dℓκ (E.2)
κ1 dℓ κ ℓ1
 ε  23 Z ℓ1 1 Z ℓ1
−
= CK ℓκ 3 dℓκ = E ℓ (ℓκ )dκ
2π ℓ2 2/3 ℓ 2 2
ℓκ [m ]

As can be seen, the energy spectrum E ℓ (ℓκ ) obeys the −1/3 law and E(κ) obeys the
2/3
−5/3 law. However, as mentioned above, vκ′2 varies as κ−2/3 (or ℓκ ).

E.1 Energy spectrum for an ideal vortex

The velocity field for an ideal vortex is given by vθ = Γ/(2πr), vr = 0, see Eq. 1.27
and Fig. 1.8, where Γ denotes the circulation. The kinetic energy, k, for the vortex
between r = R1 and r = R2 is obtained by integrating vθ2 /2 from R1 to R2 , i.e.
Z R2
1 Γdr
k= (E.3)
8π 2 (R2 − R1 ) R1 r2
From this we can define a kinetic energy spectrum as
Γ
E ideal (r) = (E.4)
8π 2 (R2 − R1 )r2
so that Z R2
k= E ideal (r)dr (E.5)
R1

Equation E.4 shows that E ideal ∝ r−3 . As in the previous section, we find that the
dependence of k on the radius is one degree higher, i.e. k = vθ2 /2 ∝ r−2 .
E.2. An example 265

E.2 An example
Let’s generate a fluctuating velocity, v ′ , using a Fourier series. For simplicity we make
it symmetric so that only the cosine part needs to be used. We make it with four terms.
It reads then (see Eq. D.19)
   
′ 0.5 2π 0.5 2π
v (x) = a1 cos x + a2 cos x
L/1 L/2
    (E.6)
0.5 2π 0.5 2π
+ a3 cos x + a4 cos x
L/3 L/4

where the wavenumber n in Eq. D.19 corresponds to 2π/(L/m) with m = (1, 2, 3, 4).
The first coefficient, a0 = 0, in Eq. D.19 because the mean of the fluctuation v ′ is zero,
i.e. Z L
v ′ dx = 0 (E.7)
0

Equation Eq. E.6 is continuous, i.e. it is given for any x/L = [0, 1]. In unsteady
CFD simulations we are always dealing with discrete points, i.e. a computational grid.
Hence, let’s express Eq. E.6 for N = 16 discrete points with ∆x = 1/(N − 1) as
   
′ 0.5 2π(n − 1)) 0.5 2π(n − 1))
v (x) = a1 cos + a2 cos
N/1 N/2
    (E.8)
0.5 2π(n − 1)) 0.5 2π(n − 1))
+ a3 cos + a4 cos
N/3 N/4

where (n − 1)/N = x/L and n = [1, N ].

Now we want v ′ in Eq. E.8 to have an energy spectrum of −5/3. Parseval’s formula,
Eq. D.21, tells us that the kinetic energy of an eddy of wavenumber κ is simply the
square of its Fourier coefficient. Hence we let the ratio of the ak coefficients in Eq. E.8
decrease as m−5/3 , i.e.

a1 = 1, a2 = 2−5/3 , a3 = 3−5/3 , a4 = 4−5/3 (E.9)

Figure E.1 shows how v ′ varies over x/L. The four terms in E. E.8 shown in Fig. E.1
can be regarded as the velocity fluctuations at one time instant of four eddies of length-
scale L, L/2, L/3 and L/4. The period of the four terms is L, L/2, L/3, and L/4
corresponding to wavenumber 2π/L, 2 · 2π/L, 3 · 3π/L and 4 · 2π/L.
Now let’s make a DFT of v ′ to get the energy spectrum (see Matlab code in Section
E.3). In DFT, the integral in Eq. D.18
Z Z  
1 π ′ 1 π ′ 2πmx
am = v (x) cos(κx)dx = v (x) cos dx (E.10)
π −π π −π L

is replaced by a summation over discrete points, i.e.

N  
1 X ′ 2πm(n − 1)
Am = v (x) cos (E.11)
N n=1 N

where (n − 1)/N = x/L (note that m = 0 corresponds to the mean, which is zero, see
Eq. E.7).
E.3. An example: Matlab code 266

m = 4, L/4
m = 3, L/3
m = 2, L/2
m = 1, L
2

0 0.2 0.4 0.6 0.8 1

x/L
Figure E.1: v ′ in Eq. E.8 vs. x/L. : term 1 (m = 1); : term 2 (m = 2); :
term 3 (m = 3); ◦: term 4 (m = 4); thick line: v ′ . Matlab code is given in Section E.3.
0 0
10 10
Evv

Evv

−1
ℓ

10

−2
10 1 −1 0
10 2π(m−1) 10 10
κ= L
ℓκ
(a) Dashed thick line shows κ−5/3 . (b) Dashed thick line shows κ−1/3 .

Figure E.2: Energy spectrum of v ′ . ◦: Evv = A2m . Matlab code is given in Section
E.3.

Now plot the energy spectrum, Evv = A2m versus wavenumber, see Fig. E.2a.
It can be seen that it decays as κ−5/3 as expected (recall that we chose the Fourier
coefficients, am , to achieve this). The total energy is now computed as
N
X N
X
hv ′2 ix = A2m = Evv (m) (E.12)
m=1 m=1

where h·ix denotes averaging over x.

Figure E.2b shows the computed energy spectrum, E ℓ (ℓκ ), versus eddy size, see
−1/3
Matlab code in Section E.3. It decays as ℓκ as it should, see Eq. E.2.

E.3 An example: Matlab code

close all
clear all

% number of cells
N=16;
E.3. An example: Matlab code 267

L=1;

n=1:1:N;
x_over_L=(n-1)/N;

% E_vv=mˆ(-5/3)
a1=1;
a2=2ˆ(-5/3);
a3=3ˆ(-5/3);
a4=4ˆ(-5/3);

for i=1:N
arg2(i)=2*pi*(i-1)/N;
arg2(i)=2*pi*x_over_L(i);
v(i)=a1ˆ0.5*cos(arg2(i))+a2ˆ0.5*cos(2*arg2(i))+a3ˆ0.5*cos(3*arg2(i))+a4ˆ0.5*c
end

% take DFT
W_cos=zeros(1,N);
W_sin=zeros(1,N);
for m=1:N
for i=1:N
a=v(i);
arg1=2*pi*(m-1)*(i-1)/N;
W_cos(m)=W_cos(m)+a*cos(arg1)/N;
W_sin(m)=W_sin(m)+a*sin(arg1)/N;
end
end

% Note that all elements of W_sin are zero since v(i) is symmetric

%*******************************************************************
figure(1)
f1=a1ˆ0.5*cos(arg2);
f2=a2ˆ0.5*cos(2*arg2);
f3=a3ˆ0.5*cos(3*arg2);
f4=a4ˆ0.5*cos(4*arg2);
plot(x_over_L,f1,’linew’,2)
hold
plot(x_over_L,f2,’r--’,’linew’,2)
plot(x_over_L,f3,’k-.’,’linew’,2)
plot(x_over_L,f4,’o’,’linew’,2)
plot(x_over_L,w,’k-’,’linew’,4)

h=gca
set(h,’fontsi’,[20])
xlabel(’x’)
ylabel(’y’)
E.3. An example: Matlab code 268

axis([0 1 -1 3])
print vprim_vs_L.ps -depsc2

%
%*******************************************************************
%
figure(2)
%
% the power spectrum is equal to W*conj(W) = W_cosˆ2+W_sinˆ2
PW=W_cos.ˆ2+W_sin.ˆ2;

mx=2*pi*(n-1)/L;

% plot power spectrum; plot only one side of the symmetric spectrum and
% multiply by two so that all energy is accounted for
plot(mx(1:N/2),2*PW(1:N/2),’bo’,’linew’,2)
hold
h=gca
set(h,’xscale’,’log’)
set(h,’yscale’,’log’)
axis([5 50 0.01 1])

% plot -5/3 line

xxx=[5 50];
yynoll=1;
yyy(1)=yynoll;
yyy(2)=yyy(1)*(xxx(2)/xxx(1))ˆ(-5/3);
plot(xxx,yyy,’r--’,’linew’,4)

% compute the average of energy in physical space

int_phys=0;
for i=1:N
int_phys=int_phys+v(i).ˆ2/N;
end

% compute the average of energy in wavenumber space

int_wave=0;
for i=1:N
int_wave=int_wave+PW(i);
end

set(h,’fontsi’,[20])
xlabel(’x’)
ylabel(’y’)
print spectra_vs_kappa.ps -depsc2
%
%*******************************************************************
E.3. An example: Matlab code 269

%
figure(3)

% compute the length corresponding the wavenumber

lx=2*pi./mx;

% multiply PW by lxˆ(-2) to get the energy spectrum E(lx)

PW_L=PW.*lx.ˆ(-2);

% plot power spectrum

plot(lx(1:N/2),2*PW_L(1:N/2),’bo’,’linew’,2)
hold
h=gca
set(h,’xscale’,’log’)
set(h,’yscale’,’log’)

axis([0.1 1.2 0.4 1])

% plot -1/3 line

xxx=[0.1 1]
yynoll=0.9;
yyy(1)=yynoll;
yyy(2)=yyy(1)*(xxx(2)/xxx(1))ˆ(-1/3);
plot(xxx,yyy,’r--’,’linew’,4);

set(h,’fontsi’,[20])
xlabel(’x’)
ylabel(’y’)
print spectra_vs_L.ps -depsc2
F. TME225 Assignment 2: turbulent flow 270

F TME225 Assignment 2: turbulent flow

N this exercise you will use data from a Direct Numerical Simulation (DNS) for
I fully developed channel flow. In DNS the unsteady, three-dimensional Navier-Stokes
equations are solved numerically. The Re number based on the friction velocity and
the half channel width is Reτ = uτ h/ν = 500 (h = ρ = uτ = 1 so that ν = 1/Reτ ).
A 96 × 96 × 96 mesh has been used. The streamwise, wall-normal and spanwise
directions are denoted by x (x1 ), y (x2 ) and z (x3 ) respectively. The cell size in x and
z directions are ∆x = 0.0654 and ∆z = 0.0164. Periodic boundary conditions were
applied in the x and z direction (homogeneous directions). All data have been made
non-dimensional by uτ and ρ.
You can do the assignment on your own or in a group of two. You should write a
report where you analyze the results following the heading F.1–F.12. It is recommended
(but the not required) that you use LATEX(an example of how to write in LATEXis available
on the course www page). It is available on Linux. On Windows you can use, for
example, MikTex (www.miktex.org) which is free to download.

F.1 Time history

At the course home page http://www.tfd.chalmers.se/˜lada/MoF/ you
find a file u v time 4nodes.dat with the time history of v1 and v2 . The file
has eight columns of v1 and v2 at four nodes: x2 /δ = 0.0039, x2 /δ = 0.0176,
x2 /δ = 0.107 and x2 /δ = 0.47. With uτ = 1 and ν = 1/Reτ = 1/500 this cor-
respond to x+ + + +
2 = 1.95, x2 = 8.8, x2 = 53.5 and x2 = 235. The sampling time step
is ∆t = 0.0033 (every second time step). The four points are located in the viscous
sublayer, the buffer layer and in the logarithmic layer, see Fig. 6.2 at p. 80.
Use the Matlab. You can also use Octave. Octave is a Matlab clone which can be
downloaded for free. Start the program pl time.m which loads and plots the time
history of v1 . Start Matlab and run the program pl time. Recall that the velocities
have been scaled with the friction velocity uτ , and thus what you see is really v1 /uτ .
The time history of v1 at x2 /δ = 0.0176 and x2 /δ = 0.107 are shown. Study the time
history of the blue line (x2 /δ = 0.0176) more in detail. Make a zoom between, for
example, t = 10 and t = 11 and v1,min = 3 and v1,min = 21. This is conveniently
done with the command

axis([10 11 3 21])

In order to see the value at each sampling time step, change the plot command to

plot(t,u2,’b-’,t,u2,’bo’)

Use this technique to zoom, to look at the details of the time history. Alternatively,
you can use the zoom buttons above the figure.
Plot v1 for all four nodes. How does the time variation of v1 vary for different
positions? Plot also v2 at the four different positions. What is the differences between
v1 and v2 ?

F.2 Time averaging

Compute the average of the v1 velocity at node 2. Add the following code (before the
plotting section)
F.3. Mean flow 271

umean=mean(u2)

Here the number of samples is n = 5000 (the entire u2 array). Find out how many
samples must be used to get a correct mean value. Start by trying with 100 samples as

umean_100=mean(u2(1:100))

What is the maximum and minimum value of v1 ? Compare those to the mean.
Do the same exercise for the other three nodes.
Compute and plot also the instantaneous fluctuations; v1′ at node 1, for example, is
computed as

u1_mean=mean(u1);
u1_fluct=u1-u1_mean;

F.3 Mean flow

All data in the data files below have been stored every 10th time step.
Download the file uvw inst small.mat, y.dat and the Matlab file pl vel.m
which reads the data files. The data file includes v1 , v2 and v3 from the same DNS as
above, but now you are given the time history of all x2 nodes at one chosen x1 and x3
node. There are nj = 98 nodes in the x2 direction; node 1 and nj are located at the
lower and upper wall, respectively.
Your data are instantaneous. Compute the mean velocity. Plot it both as linear-
linear plot and a log-linear plot (cf. Fig. 6.4).
In the log-linear plot, use x+ + +
2 for the wall distance. Include the linear law, v1 = x2 ,
+ −1 +
and the log law, v1 = κ ln x2 + B (κ = 0.41 is the von Kármán constant and
B = 5.2). How far out from the wall does the velocity profile follow the linear law?
At what x+ 2 does it start to follow the log-law?
Compute the bulk velocity
Z 2h
1
V1,b = v̄1 dx2 (F.1)
2h 0
(recall that h denote half the channel width) What is the Reynolds number based on
V1,b and centerline velocity, V1,c , respectively?

F.4 The time-averaged momentum equation

Let us time average the streamwise momentum equation. Since the flow is fully devel-
oped and two dimensional we get

1 ∂ p̄ ∂ 2 v̄1 ∂v ′ v ′
0=− +ν 2 − 1 2 (F.2)
ρ ∂x1 ∂x2 ∂x2
This equation is very similar to fully developed laminar flow which you studied in
Assignment 1, see Eq. 3.24; the difference is that we now have an additional term
which is the derivative of the Reynolds shear stress. Recall that all terms in the equa-
tion above represent forces (per unit mass). Let us investigate how these forces (the
pressure gradient, the viscous term and the Reynolds stress term) affect fluid parti-
cles located at different x2 locations. Compute and plot the three terms. (the file
uvw inst small.mat does not include p̄; set ∂ p̄/∂x = −1.)
F.5. The gradient 272

If a term is positive it means that it pushes the fluid particle in the positive x1
direction. What about the viscous term? Is it always negative? Where is it largest? At
that point? which term balances it? How large is the third term? The pressure term
should be a driving force. Where is the Reynolds shear stress positive and where is it
negative?

F.5 The gradient

In the Section F.4 you needed to compute the gradient of the velocities, e.g. ∂v̄1 /∂x2 .
Presumably, you used the Matlab command

dudy=gradient(u1d,y);

Now you should compute ∂v̄1 /∂x2 yourself, i.e. without using the Matlab com-
mand gradient. Make a for loop and compute ∂v̄1 /∂x2 as ∆v̄1 /∆x2 . Recall that
the grid is non-equidistant, i.e. ∆x2 is not constant. Make sure that your ∂v̄1 /∂x2
gives exactly the same values as the Matlab command gradient, also at the bound-
aries (j = 1 and j = nj).
Hint: compute first the face values of v̄1 (v̄1,n and v̄1,s ) using linear interpola-
tions (you cannot use weight factors of 0.5 since ∆x2 is not constant). Then compute
∆v̄1 /∆x2 as (v̄1,n − v̄1,s )/(x2,n − x2,s ) where subscript n and s denote north and
south, respectively. Recall that the variable yc in pl vel.m denotes x2 at the faces.
Compute also the second derivate, ∂ 2 v̄1 /∂x22 , without using the Matlab command
gradient. Compare this result with what you get with the gradient, i.e.

dudy=gradient(u1d,y);
d2udy2=gradient(dudy,y);

Hint: compute first the face values of ∂v̄1 /∂x2 .

F.6 Wall shear stress

Compute the wall shear stress at both walls. They should be equal. Are they?

F.7 Resolved stresses

In Section F.3 you computed the mean velocities. From the instantaneous and the mean
velocity, you can compute the fluctuations as

vi′ = vi − v̄i (F.3)

Now you can easily compute all stresses vi′ vj′ . Plot the normal stresses in one figure
and the shear stresses in one figure (plot the stresses over the entire channel, i.e. from
x2 = 0 to x2 = 2h). Which shear stresses are zero?

F.8 Fluctuating wall shear stress

In the same way as the velocity, the wall shear stress can be decomposed into a mean
value and a fluctuation, i.e.
∂v1 ∂v̄1
τw,inst = µ , τw = µ (F.4)
∂x2 ∂x2
F.9. Production terms 273

where the first expression is the instantanous wall shear stress and the second is the
time-averaged. In general, any fluctuating variable, φ, can be decomposed into a mean
and fluctuation as φ = φ̄ + φ′ . The root-mean-square (RMS) is then defined as
 1/2
φrms = φ′2 (F.5)

Compute the RMS of the wall shear stress. This is a measure of the fluctuating tan-
gential force on the wall due to turbulence. If heat transfer is involved, the fluctuating
temperature at the wall inducing fluctuating heat transfer may be damaging to the ma-
terial of the walls causing material fatigue. This is probably the most common form of
fluid-solid interaction.

F.9 Production terms

In order to understand why a stress is large, it is useful to look at its transport equation,
see Eq. 9.12. Usually, a stress is large when its production term, Pij , is large (there
may be exceptions when other terms, such as the diffusion term, are largest). Plot the
production terms for all non-zero stresses across the entire channel. Which ones are
zero (or close to)? Does any production term change sign at the centerline? If so, what
about the sign of the corresponding shear stress plotted in Section F.7?

F.10 Pressure-strain terms

The pressure-strain term reads (see Eq. 9.14)
 
p′ ∂vi′ ∂vj′
Πij = + (F.6)
ρ ∂xj ∂xi

Our data are obtained from incompressible simulations, in which the pressure may
vary unphysically in time (∂p/∂t does not appear in the equations). Hence, we prefer
to compute the velocity-pressure gradient term

vi′ ∂p′ vj′ ∂p′

Πpij = − − , (F.7)
ρ ∂xj ρ ∂xi

see the second line in Eq. 9.3. The pressure diffusion term in the v2′2 equation – which
is the difference between Eqs. F.6 and F.7 (the two first terms in Eq. 9.8) – is small
except very close to the wall (see Figs. 9.4 and 9.6). Hence, the difference between Πpij
and Πij is small.
Download the data file p inst small.mat and the Matlab file pl press strain.m
which reads the data file. The time histories of the pressure along five x2 lines [(x1 , x2 , x3 ),
(x1 ± ∆x1 , x2 , x3 ) and (x1 , x2 , x3 ± ∆x3 )] are stored in this file. This allows you to
compute all the three spatial derivatives of p′ . Using the velocities stored in uvw inst small.mat
(see Section F.3), you can compute all the terms in Eq. F.7.
Plot the pressure strain, Πpij , for the three normal stresses and the shear stress across
the channel. For which stresses is it negative and positive? Why?
Which term Πpij is the largest source and sink term, respectively?
F.11. Dissipation 274

F.11 Dissipation
The physical meaning of dissipation, ε, is transformation of turbulent kinetic energy
into internal energy, i.e. increased temperature.
Download the files y half.dat, diss inst.mat and the Matlab file pl diss.m
which reads it. The data file includes the time history of the velocities along five x2
lines [(x1 , x2 , x3 ), (x1 ±∆x1 , x2 , x3 ) and (x1 , x2 , x3 ±∆x3 )] so that you can compute
all spatial derivatives. The data cover only the lower half of the channel. Compute and
plot
∂v ′ ∂vi′
ε=ν i (F.8)
∂xk ∂xk
see Eq. 8.14. Where is it largest? In which equation does this quantity appear?
Let us now consider the equations for the mean kinetic energy, K = v̄i v̄i /2 (Eq. 8.35)
and turbulent kinetic energy, k = vi′ vi′ /2 (Eq. 8.14). The dissipation in the K equation
reads
∂v̄i ∂v̄i
εmean = ν (F.9)
∂xk ∂xk
The flow of kinetic energy between K, k and ∆T is illustrated in Fig. 8.6 The dissipa-
tions, ε and εmean , are defined in Eqs. F.8 and F.9, respectively. Compute and plot also
εmean and P k . Which is large and which is small? How is the major part of the kinetic
energy transformed from K to ∆T ? Is it transformed via k or directly from K to ∆T ?

F.12 Do something fun!

You have been provided with a lot of data which you have analyzed in many ways.
Now think of some other way to analyze the data. There are many interesting things
yet to be analyzed!
G. TME225 Learning outcomes 2016 275

G TME225 Learning outcomes 2016

TME225 Learning outcomes 2016: week 1
1. Explain the difference between Lagrangian and Eulerian description of the mo-
tion of a fluid particle.
2. Consider the flow in Section 1.2. Show that ∂v1 /∂t is different from dv1 /dt.
3. Watch the on-line lecture Eulerian and Lagrangian Description, part 1 – 3 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. Part 1 describes the difference between Lagrangian and Eulerian points and
velocities.
∂T ∂T
ii. The formula + vi is nicely explained in Part 2
∂t ∂xi
4. Show which stress components, σij , that act on a Cartesian surface whose normal
vector is ni = (1, 0, 0). Show also the stress vector, tn̂i . (see Eq. B.2 and Fig. 1.3
and the Lecture notes of Ekh [2])
5. Show that the product of a symmetric and an antisymmetric tensor is zero.
6. Explain the physical meaning of diagonal and off-diagonal components of Sij
7. Explain the physical meaning of Ωij
8. What is the definition of irrotational flow?
9. What is the physical meaning of irrotational flow?
10. Derive the relation between the vorticity vector and the vorticity tensor
11. Start from Eq. 1.16 and express the vorticity tensor as a function of the vorticity
vector (Eq. 1.18)
12. Explain the physical meaning of the eigenvectors and the eigenvalues of the
stress tensor (see Section 1.8 and the Lecture notes of Ekh [2])
13. Watch the on-line lecture Vorticity, part 1 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. What is a vortex line?
ii. The teacher talks of ωA and ωB , where ω = 0.5(ωA + ωB ); what does it
denote? (cf. Fig. 1.4 in the lecture notes)
iii. The teacher shows the rotating tank (after 3 minutes into the movie). He
puts the vorticity meter into the tank. The flow in the tank moves like a
solid body. How does the vorticity meter move? This is a curved flow with
vorticity.
iv. The teacher puts the vorticity meter into a flow in a straight channel (near a
wall). What happens with the vorticity meter? (cf. Fig. 1.10)
G. TME225 Learning outcomes 2016 276

v. After 4:20 minutes, the teacher shows the figure of a boundary layer. He
says that one of the “vorticity legs” (ωA in Item 13ii above) is parallel to the
wall; what does he say about ωB ? What conclusion does is draw about ω?
This is a straight flow with vorticity.
vi. After 4:30 minutes, the teacher introduces a spiral vortex tank (in the lecture
notes this is called an ideal vortex). How does the vortex meter behave? (cf.
Fig. 1.8 in the lecture notes). How does the teacher explain that the vorticity
is zero (look at the figure he talks about after 6 minutes); the explanation
uses the fact that the tangential velocity, vθ , is inversely proportional to the
radius, r, see Eq. 1.29 in the lecture notes.
vii. After 8:40 minutes, the teacher puts the vorticity meter at different locations
in the boundary layer; he puts it near the solid wall, a bit further out and
finally at the edge of the boundary layer. How does the vorticity meter move
at the different locations? Where is the vorticity smallest/largest? Explain
why.
viii. After 10:35, the vortex meter is shown in the spiral vortex tank. What hap-
pens with the vorticity when we get very close to the center? Does it still
remain zero? What happens with the tangential velocity? (see Eq. 1.29)
ix. The teachers explains the concept of circulation, Γ, and its relation to vor-
ticity (cf. Eqs. 1.23 and 1.25).
x. What is a vortex core?
xi. How large is the vorticity and the circulation in the rotating tank?
xii. How large is the vorticity and the circulation in the spiral vortex tank? Does
it matter if you include the center?
xiii. The teacher presents a two-dimensional wing. Where is the pressure low and
high, respectively? The Bernoulli equation gives then the velocity; where is
it high and low, respectively? The velocity difference creates a circulation,
Γ.
xiv. After 15:25 minutes, the teacher looks at the rotating tank again. He starts
to rotate the tank; initially there is only vorticity near the outer wall. As time
increases, vorticity (and circulation) spread toward the center. Finally, the
flow in the entire tank has vorticity (and circulation). This illustrates that as
long as there is an imbalance in the shear stresses, vorticity (and circulation)
is changed (usually created), see Figs. 1.10 and 4.1.
14. Equation 1.7 states that mass times acceleration is equal to the sum of forces (per
unit volume). Write out the momentum equation (without using the summation
rule) for the x1 direction and show the surface forces and the volume force on a
small, square fluid element (see lecture notes of Ekh [2]). Now repeat it for the
x2 direction.
15. Derive the Navier-Stokes equation, Eq. 2.5 (use the formulas in the Formula
sheet which can be found on the course www page)
G. TME225 Learning outcomes 2016 277

TME225 Learning outcomes 2016: week 2-3

1. Simplify the Navier-Stokes equation for incompressible flow and constant vis-
cosity (Eq. 2.8)
2. Derive the transport equation for the internal energy, u, Eq. 2.13 (again, use the
Formula sheet). What is the physical meaning of the different terms?
3. Simplify the the transport equation for internal energy to the case when the flow
is incompressible (Eq. 2.16).
4. Derive the transport equation for kinetic energy, vi vi /2, Eq. 2.21. What is the
physical meaning of the different terms?
5. Explain the energy transfer between kinetic energy, k, and internal energy, u
6. Show how the left side of the transport equations can be written on conservative
and non-conservative form
7. Starting from the Navier-Stokes equations (see Formula sheet), derive the flow
equation governing the Rayleigh problem expressed in f and η; what are the
boundary conditions in time (t) and space (x2 ); how are they expressed in the
similarity variable η?
8. Show how the boundary layer thickness can be estimated from the Rayleigh
problem using f and η (Fig. 3.3)
9. Explain the flow physics at the entrance (smooth curved walls) to a plane channel
(Fig. 3.5).
10. Explain the flow physics in a channel bend (Fig. 3.6).
11. Derive the flow equations for fully developed flow between two parallel plates,
i.e. fully developed channel flow (Eqs. 3.18, 3.22 and 3.26)
12. Show that the continuity equation is automatically satisfied in 2D when the ve-
locity is expressed in the streamfunction, Ψ
13. Starting from Eq. 3.41, derive the equation for two-dimensional boundary-layer
flow expressed in the streamfunction (Eq. 3.45). Start from Eq. 3.41
14. Derive the Blasius equation, Eq. 3.53. Start from Eq. 3.45
15. Explain (using words and a figure) why vorticity can be created only by an im-
balance (i.e. a gradient) of shear stresses. Explain why pressure and the gravity
force cannot create vorticity.
16. The Navier-Stokes equation can be re-written on the form

∂vi ∂k 1 ∂p ∂ 2 vi
+ − εijk vj ωk = − +ν + fi
∂t ∂x | {z } ρ ∂xi ∂xj ∂xj
|{z}i rotation
no rotation

Derive the transport equation (3D) for the vorticity vector, Eq. 4.20
17. Show that the divergence of the vorticity vector, ωi , is zero
G. TME225 Learning outcomes 2016 278

18. Explain vortex stretching and vortex tilting

19. Show that the vortex stretching/tilting term is zero in two-dimensional flow
20. Derive the 2D equation transport equation for the vorticity vector from the 3D
transport equation, Eq. 4.22
21. Show the similarities between the vorticity and temperature transport equations
in fully developed flow between two parallel plates
r r
δ ν 1
22. Use the diffusion of vorticity to show that ∝ = (see also
ℓ Uℓ Re
Eq. 3.14).
23. Watch the on-line lecture Boundary layers parts 1 at
http://www.tfd.chalmers.se/˜lada/flow viz.html

i. Consider the flow over the flat plate. How does the boundary layer thickness
change when we move downstream?
ii. What value does the fluid velocity take at the surface? What is this boundary
conditions called: slip or no-slip? How do they define the boundary layer
thickness?
iii. How is the wall shear stress defined? How does it change when we move
downstream? (how does this compare with the channel flow in TME225
Assignment 1?)
iv. How is the circulation, Γ, defined? (cf. with Eq. 1.23) How is it related to
vorticity? How do they compute Γ for a unit length (> δ) of the boundary
layer? How large is it? How does it change when we move downstream on
the plate?
v. Where is the circulation (i.e. the vorticity) created? Where is the vortic-
ity created in “your” channel flow (TME225 Assignment 1)? The vorticity
is created at different locations in the flat-plate boundary layer and in the
channel flow: can you explain why? (hint: in the former case

∂p ∂ 2 v1 
= µ 2  = 0,
∂x1 ∂x2 wall

but not in the latter; this has an implication for γ2,wall [see Section 4.3])
vi. How do they estimate the boundary layer thickness? (cf. Section. 4.3.1)
G. TME225 Learning outcomes 2016 279

Week 4
1. Derive the Bernoulli equation (Eq. 4.32)
Hint: The gravitation term is first expressed as a potential gi = −∂ Φ̂/∂xi .
2. Consider the derivative of the complex function (f (z + z0 ) − f (z))/z0 where
z = x + iy and f = u + iv. The derivative of f must be independent in which
coordinate direction the derivative is taken (either along the real or the imaginary
axis). Show that this leads to the Cauchy-Riemann equations
∂u ∂v ∂u ∂v
= , =−
∂x ∂y ∂y ∂x

3. Show that in inviscid flow, both the velocity potential and the streamfunction
satisfy the Laplace equation.
4. Consider the complex potential f = Φ + iΨ. Show that f = C1 z n satisfies the
Laplace equation. Derive the velocity polar components for n = 1 and n = 2.
What physical flow do these two cases correspond to?
5. Derive the polar velocity components for the complex potential f = ṁ ln z/(2π)
and f = −iΓ ln z/(2π) (Γ denotes circulation). What does the physical flow
look like? Show that they satisfy the Laplace equation.
6. A doublet is a combination of a radial sink and source and its complex potential
reads f = µ/(πz). Combine it with the potential for parallel flow (z = V∞ z).
Derive the resulting velocity field around a cylinder (in polar components).
Hint: the radius is defined as r0 = µ/(πV∞ ).
7. Consider the potential flow around a cylinder. Show that the radial velocity is
zero at the surface. Use the Bernoulli equation to get the surface pressure. Show
that the drag and lift forces are zero. Where are the stagnation points located?
8. Add the complex potential of a vortex line, f = −iΓ ln z/(2π) (Γ denotes cir-
culation) to the complex potential for cylinder flow. Compute the polar velocity
components. Where are the stagnation points located? What happens with the lo-
cation of the stagnation point(s) when the circulation becomes very large? How
is the lift of the cylinder computed (which applies for any body).
9. What is the Magnus effect? Explain the three applications in the text: why is it
efficient to use loops in table tennis? Why does the Magnus effect help a football
player get the ball around the wall (of players) when making a free-kick? How
does the Magnus effect help propulsing a ship using Flettner rotors (to learn more
about Flettner rotors, google “flettner rotors”).
10. Consider the inviscid flow around an airfoil, see Fig. 4.19. In inviscid theory it
would look like Fig. 4.20. What has been done to achieve the flow in Fig. 4.21?
How is the lift computed?
11. Watch the on-line lecture Boundary layers part 2 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. How does the boundary layer thickness change at a given x when we in-
crease the velocity? Explain why.
G. TME225 Learning outcomes 2016 280

ii. Consider the flow in a contraction: what happens with the boundary layer
thickness after the contraction?
iii. Why is the vorticity level higher after the contraction?
iv. Is the wall shear stress lower or higher after the contraction? Why?
v. Consider the flow in a divergent channel (a diffusor): what happens with the
boundary layer thickness and the wall shear stress?
vi. What happens when the angle of the diffusor increases?
vii. What do we mean by a “separated boundary layer”? How large is the wall
shear stress at the separation point?
viii. The second part of the movie deals with turbulent flow: we’ll talk about that
in the next lecture (and the remaining ones).
12. Watch the on-line lecture Boundary layers parts 2 (second half) & 3 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. The flow is “tripped” into turbulence. How?
ii. When the flow along the lower wall of the diffusor is tripped into turbulent
flow, the separation region is suppressed. Try to explain why.
iii. Two boundary layers – one on each side of the plate – are shown. The upper
one is turbulent and the lower one is laminar. What is the difference in the
two velocity profiles? (cf. my figures in the ’summary of lectures’) Explain
the differences.
iv. Why is the turbulent wall shear stress larger for the turbulent boundary
layer? What about the amount of circulation (and vorticity) in the laminar
and turbulent boundary layer? How are they distributed?
v. Consider the airfoil: when the boundary layer on the upper (suction) side
is turbulent, stall occurs at a higher angle of incidence compared when the
boundary layer is laminar. Why?
vi. Vortex generator are place on the suction side in order prevent or delay sep-
aration. Try to explain why separation is delayed.
13. What characterizes turbulence? Explain the characteristics. What is a turbulent
eddy?
14. Explain the cascade process. How large are the largest scales? What is dissi-
pation? What dimensions does it have? Which eddies extract energy from the
mean flow? Why are these these eddies “best” at extracting energy from the
mean flow?
15. What are the Kolmogorov scales? Use dimensional analysis to derive the expres-
sion for the velocity scale, vη , the length scale, ℓη and the time scale, τη .
16. Make a figure of the energy spectrum. The energy spectrum consists of thee
subregions: which? describe their characteristics. Show the flow of turbulent
kinetic energy in the energy spectrum. Given the energy spectrum, E(κ), how
is the turbulent kinetic energy, k, computed? Use dimensional analysis to derive
the −5/3 Kolmogorov law.
17. What does isotropic turbulence mean? What about the shear stresses?
G. TME225 Learning outcomes 2016 281

Week 5
1. How is the energy transfer from eddy-to-eddy, εκ , estimated? Show how the ratio
of the large eddies to the dissipative eddies depends on the Reynolds number (see
Eq. 5.16).
2. Describe the cascade process created by vorticity. Write the vortex stretch-
ing/tilting term in tensor notation. What is its physical meaning? Describe the
physical process of vortex stretching which creates smaller and smaller eddies.
Show and discuss the family tree of turbulence eddies and their vorticity. Show
that in 2D flow the vortex stretching/tilting term vanishes.
3. Watch the on-line lecture Turbulence part 1 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. Why does the irregular motion of wave on the sea not qualify as turbulence?
ii. How is the turbulence syndrome defined?
iii. The movie shows laminar flow in a pipe. The viscosity is decreased, and the
pressure drop (i.e. the resistance, the drag, the loss) decreases. Why? The
viscosity is further decreased, and the pressure drop increases. Why? How
does the characteristics of the water flow coming out of the pipe change due
to the second decrease of viscosity?
iv. It is usually said that the flow in a pipe gets turbulent at a Reynolds number
of 2300. In the movie they show that the flow can remain laminar up to
8 000. How do they achieve that?
v. Dye is introduced into the pipe. For laminar flow, the dye does not mix with
the water; in turbulent flow it does. When the mixing occurs, what happens
with the pressure drop?
4. Watch the on-line lecture Turbulence part 2 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. Draw a laminar and turbulent velocity profile for pipe flow. What is the
∂v̄1
main difference? In which flow is the wall shear stress τw = µ largest,
∂x2
laminar or turbulent?
ii. In turbulent flow, the velocity near the wall is larger than in laminar flow.
Why?
iii. Discuss the connection between mixing and the cross-stream (i.e. v2′ ) fluc-
tuations.
iv. Try to explain the increased pressure drop in turbulent flow with increased
mixing.
v. The center part of the pipe is colored with blue dye and the wall region is
colored with red dye: by looking at this flow, try to explain how turbulence
creates a Reynolds shear stress.
vi. Two turbulent jet flows are shown, one at low Reynolds number and one at
high Reynolds number. They look very similar in one way and very different
in another way. Which scales are similar and which are different?
G. TME225 Learning outcomes 2016 282

vii. The two turbulent jet flows have the same energy input and hence the same
dissipation. Use this fact to explain why the smallest scales in the high
Reynolds number jet must be smaller that those in the low Reynolds number
jet.
viii. At the end of the presentation of the jet flow, they explain the cascade pro-
cess.
ix. Explain the analogy of a water fall (cascade of water, the water passes down
the cascade) and the turbulent cascade process.
5. Use the decomposition vi = v̄i + vi′ to derive the time-averaged Navier-Stokes
equation. A new terms appears: what is it called? Simplify the time-averaged
Navier-Stokes equation for boundary layer flow. What is the total shear stress?
6. How is the friction velocity, uτ , defined? Define x+ +
2 and v̄1 .

7. The wall region is divided into an inner and outer region. The inner region is
furthermore divided into a viscous sublayer, buffer layer and log-layer. Make a
figure and show where these regions are valid (Fig. 6.2)
8. What are the relevant velocity and length scales in the viscous-dominated region
(x+
2 . 5)? Derive the linear velocity law in this region (Eq. 6.22). What are
the suitable velocity and length scales in the inertial region (the fully turbulent
region)? Derive the log-law for this region.
9. Consider fully developed turbulent channel flow. In which region (viscous sub-
layer, buffer layer or log-layer) does the viscous stress dominate? In which re-
gion is the turbulent shear stress large? Integrate the boundary layer equations
and show that the total shear stress varies as 1 − x2 /δ (Eq. 6.20).
10. In fully developed turbulent channel flow, the time-averaged Navier-Stokes con-
sists only of three terms. Make a figure and show how the velocity and Reynolds
shear stress vary across the channel. After that, show how the three terms vary
across the channel (Fig. 6.6). Which two terms balance each other in the outer
region? Which term drives (“pushes”) the flow in the x1 direction? Which two
terms are large in the inner region? Which term drives the flow?
11. Derive the exact transport equation for turbulent kinetic energy, k. Discuss the
physical meaning of the different terms in the k equation. Which terms do only
transport k? Which is the main source term? Main sink (i.e. negative source)
term?
G. TME225 Learning outcomes 2016 283

1. In the cascade process, we assume that the dissipation is largest at the smallest
scales, i.e. ε(κ) ∝ κ4/3 , see Eq. 8.19 at p. 95. Show this. For which eddies is
the production largest? Why?
2. Watch the on-line lecture Turbulence part 3 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. The movie says that there is a similarity of the small scales in a channel flow
and in a jet flow. What do they mean?
ii. What happens with the small scales when the Reynolds number is increased?
What happens with the large scales? Hence, how does the ratio of the large
scales to the small scales change when the Reynolds number increases (see
Eq. 5.16)
iii. In decaying turbulence, which scales die first? The scenes of the clouds
show this in a nice way.
iv. Even though the Reynolds number may be large, there are a couple of phys-
ical phenomena which may inhibit turbulence and keep the flow laminar:
mention three.
v. Consider the flow in the channel where the fluid on the top (red) and the
bottom (yellow) are separated by a horizontal partition. The two fluids are
identical. Study how the two fluids mix downstream of the partition. In the
next example, the fluid on the top is hot (yellow) and light, and the one at
the bottom (dark blue) is cold (heavy); how do the fluids mix downstream
of the partition, better or worse than in the previous example? This flow
situation is called stable stratification. In the last example, the situation is
reversed: cold, heavy fluid (dark blue) is moving on top of hot, light fluid
(yellow). How is the mixing affected? This flow situation is called unstable
stratification. Compare in meteorology where heating of the ground may
cause unstable stratification or when inversion causes stable stratification.
You can read about stable/unstable stratification in Section 12.1 at p. 147.
3. Given the exact k equation, give the equation for boundary-layer flow (Eq. 8.23).
All spatial derivatives are kept in the dissipation term: why? In the turbulent
region of the boundary layer, the k equation is dominated by two terms. Which
ones? Which terms are non-zero at the wall?
4. Where is the production term, P k = −v1′ v2′ ∂v̄1 /∂x2 , largest? In order to explain
this, show how −v1′ v2′ and ∂v̄1 /∂x2 vary near the wall.
5. Discuss the difference of spatial transport of k and spectral transfer of k. Give
an example of how they are combined in non-homogeneous turbulence. How is
homogeneous turbulence defined?
6. Derive the exact transport equation for mean kinetic energy, K. Discuss the
physical meaning of the different terms. One term appears in both the k and the
K equations: which one? Consider the dissipation terms in the k and the K
equations: which is largest near the wall and away from the wall, respectively?
Show where they appear in the energy spectrum.
7. Derive the exact transport equation for turbulent Reynolds stress, vi′ vj′ . Take the
trace of the vi′ vj′ equation to derive the k equation.
G. TME225 Learning outcomes 2016 284

8. Show that the role of the convection and diffusion terms is purely to transport the
quantity (k for example) and that they give no net effect except at the boundaries
(use the Gauss divergence theorem), see Eqs. 8.26 and 8.27.
9. Discuss the physical meaning of the different terms in the vi′ vj′ equation.

10. Consider the pressure-strain term in the vi′ vj′ equation. The mean normal stress
′2 = v ′ v ′ /3; what sign will the pressure-strain term have for
can be defined as vav i i
the normal stresses which are, respectively, larger and smaller than vav ′2 ? What

role does Π12 has? What sign? Why do we call the pressure-strain term the
Robin Hood term?
11. Consider the dissipation term, ε12 , for the shear stress: how large is it?
12. Consider fully developed channel flow: how are the expressions for the produc-
tion terms in the vi′ vj′ equations simplified? Which production terms are zero and
non-zero, respectively? Consider the production term for v1′ v2′ : what sign does
it have in the lower and upper part of the channel, respectively? Why is there no
pressure-strain term in the k equation?
13. Consider the fully turbulent region in fully developed channel flow: which are
the main source and sink terms in the v1′2 , v2′2 , v3′2 and v1′ v2′ equations? Which
are the largest terms at the wall? Which terms are zero at the wall?
14. Consider channel flow and use physical reasoning to show that v1′ v2′ must be
negative and positive in the lower and upper half of the channel, respectively. Is
this consistent with the sign of P12 ?
15. Explain why v1′ v2′ and ∂v̄1 /∂x2 must have different sign in fully-developed
channel flow by looking at the P12 .
16. Define the two-point correlation. How is it normalized? What is the physical
meaning of the two-point correlation? How is it related to the largest eddies?
How is the integral length scale defined?
17. Define the auto correlation. How is it normalized? What physical meaning does
it have? The integral time scale is defined in analogy to the integral length scale:
show how it is defined.
18. What is Taylor’s hypothesis? Shown that the integral length scale can be esti-
mated from the integral time scale
G. TME225 Learning outcomes 2016 285

TME225 Learning outcomes 2016: just for fun!

1. Watch the on-line lecture Pressure field and acceleration part 1 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. The water flow goes through the contraction. What happens with the veloc-
ity and pressure. Try to explain.
ii. Fluid particles become thinner and elongated in the contraction. Explain
why.
dVs
iii. In the movie they show that the acceleration along s, i.e. Vs , is related to
ds
dp
the pressure gradient . Compare this relation with the three-dimensional
ds
form of Navier-Stokes equations for incompressible flow, Eq. 2.8
2. Watch the on-line lecture Pressure field and acceleration part 2 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. Water flow in a manifold (a pipe with many outlets) is presented. The pres-
sure decreases slowly downstream. Why?
ii. The bleeders (outlets) are opened. The pressure now increases in the down-
stream direction. Why?
iii. What is the stagnation pressure? How large is the velocity at a stagnation
point?
iv. What is the static pressure? How can it be measured? What is the difference
between the stagnation and the static pressures?
v. A venturi meter is a pipe that consists of a contraction and an expansion (i.e.
a diffusor). The bulk velocities at the inlet and outlet are equal, but still the
pressure at the outlet is lower than that at the inlet. There is a pressure drop.
Why?
vi. What happens with the pressure drop when there is a separation in the dif-
fusor?
vii. They increase the speed in the venturi meter. The pressure difference in
the contraction region and the outlet increases. Since there is atmospheric
pressure at the outlet, this means that the pressure in the contraction region
must decrease as we increase the velocity of the water. Finally the water
starts to boil, although the water temperature may be around 10o C. This is
called cavitation (this causes large damages in water turbines).
viii. Explain how suction can be created by blowing in a pipe.
3. Watch the on-line lecture Pressure field and acceleration part 3 at
http://www.tfd.chalmers.se/˜lada/flow viz.html
i. What is the Coanda effect?
ii. The water from the tap which impinges on the horizontal pipe attaches to the
surface of the pipe because of the Coanda effect. How large is the pressure
at the surface of the pipe relative to the surrounding pressure?
iii. Explain the relation between streamline curvature and pressure (cf. Sec-
tion 3.2.1).
G. TME225 Learning outcomes 2016 286

iv. At the end of the contraction, there is an adverse pressure gradient (∂p/∂x >
0). Explain why.
H. MTF270: Some properties of the pressure-strain term 287

H MTF270: Some properties of the pressure-strain term

N this Appendix we will investigate some properties of aijkℓ in Eq. 11.77 at p. 133.
I Introduce the two-point correlation function

Bjℓ (r) = vj′ (x)vℓ′ (x + r)

Define the point x′ = x + r so that

Bjℓ (r) = vj′ (x′ − r)vℓ′ (x′ ) = vℓ′ (x′ )vj′ (x′ − r) = Bℓj (−r)

We get
∂Bjℓ (r) ∂Bℓj (−r) ∂ 2 Bjℓ (r) ∂ 2 Bℓj (−r)
=− ⇒ = (H.1)
∂ri ∂ri ∂rk ∂ri ∂rk ∂ri
Since Eq. H.1 in the definition of aijkℓ in Eq. 11.77 is integrated over r3 covering both
r and −r (recall that vℓ′ and vj′ are separated by r), aijkℓ is symmetric with respect to
index j and ℓ, i.e.
aijkℓ = aiℓkj (H.2)
Green’s third formula (it is derived from Gauss divergence law) reads
Z
1 ∇2 ϕ
ϕ(x) = − dy3 (H.3)
4π V |y − x|

where the boundary integrals have been omitted. Setting ϕ = vℓ′ vj′ in Eq. H.3 gives
Z
1 ∂ 2 vℓ′ vj′ dy3 1
vj′ vℓ′ = − = aijiℓ (H.4)
4π V ∂xi ∂xi |y − x| 2

where the last equality is given by Equation 11.77.

I. MTF270: Galilean invariance 288

I MTF270: Galilean invariance

ELOW we repeat some of the details of the derivation given in [80]. Galilean invari-
B ance means that the equations do not change if the coordinate system is moving
with a constant speed Vk . Let’s denote the moving coordinate system by ∗, i.e.
x∗k = xk + Vk t, t∗ = t, v̄k∗ = v̄k + Vk (I.1)
∗
By differentiating a variable φ = φ(t , x∗i ) we get
∂φ(xi , t) ∂x∗j
∂φ ∂t∗ ∂φ ∂φ
= ∗ + =
∂xk ∂xk ∂xj ∂xk ∂t∗ ∂x∗k
(I.2)
∂φ(xi , t) ∂x∗k ∂φ ∂t∗ ∂φ ∂φ ∂φ
= ∗ + = Vk ∗ + ∗ .
∂t ∂t ∂xk ∂t ∂t∗ ∂xk ∂t
From Eq. I.2 is it easy to show that the Navier-Stokes (both with and without filter)
is Galilean invariant [80, 184]. Transforming the material derivative from the (t, xi )-
coordinate system to the (t∗ , x∗i )-coordinate system gives
∂φ ∂φ ∂φ ∂φ ∂φ
+ vk = ∗ + Vk ∗ + (vk∗ − Vk ) ∗
∂t ∂xk ∂t ∂xk ∂xk
∂φ ∂φ
= ∗ + vk∗ ∗ ,
∂t ∂xk
It shows that the left hand side does not depend on whether the coordinate system
moves or not, i.e. it is Galilean invariant.
Now, let’s look at the Leonard term and the cross term. Since the filtering operation
is Galilean invariant [80], we have v̄k∗ = v̄k + Vk and consequently also vk′′∗ = vk′′ . For
the Leonard and the cross term we get (note that since Vi is constant Vi = V̄i = V̄i )
L∗ij = v̄i∗ v̄j∗ − v̄i∗ v̄j∗ = (v̄i + Vi )(v̄j + Vj ) − (v̄i + Vi )(v̄j + Vj )
= v̄i v̄j + v̄i Vj + v̄j Vi − v̄i v̄j − v̄i Vj − Vi v̄j
= v̄i v̄j − v̄i v̄j + Vj (v̄i − v̄i ) + Vi (v̄j − v̄j )
= Lij − Vj vi′′ − Vi vj′′ (I.3)

∗
Cij = v̄i∗ vj′′∗ + v̄j∗ vi′′∗ = (v̄i + Vi )vj′′ + (v̄j + Vj )vi′′ =
= v̄i vj′′ + vj′′ Vi + v̄j vi′′ + vi′′ Vj = Cij + vj′′ Vi + vi′′ Vj
From Eq. I.3 we find that the Leonard term and the cross term are different in the two
coordinate systems, and thus the terms are not Galilean invariant. However, note that
the sum is, i.e.
L∗ij + Cij
∗
= Lij + Cij . (I.4)
The requirement for the Bardina model to be Galilean invariant is that the constant
must be one, cr = 1 (see Eq. 18.44). This is shown by transforming both the exact
M
Cij (Eq. 18.42) and the modelled one, Cij (i.e. Eq. 18.43). The exact form of Cij
transforms as in Eq. I.3. The Bardina term transforms as
∗M ∗ ∗
Cij = cr (v̄i∗ v̄j∗ − v̄ i v̄ j )
h i
= cr (v̄i + Vi )(v̄j + Vj ) − (v̄i + Vi )(v̄j + Vj )
(I.5)
= cr [v̄i v̄j − v̄i v̄j − (v̄ i − v̄i )Vj − (v̄ j − v̄j )Vi ]
M
 
= Cij + cr v ′′ i Vj + v ′′ j Vi .
I. MTF270: Galilean invariance 289

∗M M
As is seen, Cij 6= Cij , but here this does not matter, because provided cr = 1 the
M
modelled stress, Cij , transforms in the same way as the exact one, Cij . Thus, as for
∗M
the exact stress, Cij (see Eq. I.4), we have Cij + L∗ij = CijM
+ Lij . Note that in order
to make the Bardina model Galilean invariant the Leonard stress must be computed
explicitly.
J. MTF270: Computation of wavenumber vector and angles 290

J MTF270: Computation of wavenumber vector and

angles
OReach mode n, create random angles ϕn , αn and θn (see Figs. J.1 and 27.1) and
F random phase ψ n . The probability distributions are given in Table J.1. They are
chosen so as to give a uniform distribution over a spherical shell of the direction of the
wavenumber vector, see Fig. J.1.

J.1 The wavenumber vector, κnj

x3

dAi
θn κni
x2
n
ϕ

x1
Figure J.1: The probability of a randomly selected direction of a wave in wave-space
is the same for all dAi on the shell of a sphere.

Compute the wavenumber vector, κnj , using the angles in Section J according to
Fig. J.1, i.e.

κn1 = sin(θn ) cos(ϕn )

κn2 = sin(θn ) sin(ϕn ) (J.1)
κn3 = cos(θn )

p(ϕn ) = 1/(2π) 0 ≤ ϕn ≤ 2π
p(ψ n ) = 1/(2π) 0 ≤ ψ n ≤ 2π
p(θn ) = 1/2 sin(θ) 0 ≤ θn ≤ π
p(αn ) = 1/(2π) 0 ≤ αn ≤ 2π

Table J.1: Probability distributions of the random variables.

J.2. Unit vector σin 291

κn i σin αn
(1, 0, 0) (0, 0, −1) 0
(1, 0, 0) (0, 1, 0) 90
(0, 1, 0) (0, 0, −1) 0
(0, 1, 0) (−1, 0, 0) 90
(0, 0, 1) (0, 1, 0) 0
(0, 0, 1) (−1, 0, 0) 90

Table J.2: Examples of value of κni , σin and αn from Eqs. J.1 and J.4.

J.2 Unit vector σin

Continuity requires that the unit vector, σin , and κnj are orthogonal. This can be seen
by taking the divergence of Eq. 27.3 which gives
N
X
∇ · v′ = 2 ûn cos(κn · x + ψ n )σ n · κn (J.2)
n=1

i.e.
σin κni = 0 (J.3)
(superscript n denotes Fourier mode n). Hence, σin will lie in a plane normal to the
vector κni , see Fig. 27.1. This gives

σ1n = cos(ϕn ) cos(θn ) cos(αn ) − sin(ϕn ) sin(αn )

σ2n = sin(ϕn ) cos(θn ) cos(αn ) + cos(ϕn ) sin(αn ) (J.4)
σ3n = − sin(θn ) cos(αn )

The direction of σin in this plane (the ξ1n − ξ2n plane) is randomly chosen through
αn . Table J.2 gives the direction of the two vectors in the case that κi is along one
coordinate direction and α = 0 and α = 90o .
K. MTF270: 1D and 3D energy spectra 292

K MTF270: 1D and 3D energy spectra

general two-point correlation Bij of vi′ and vj′ (see Eq. 10.2) can be expressed
T HE
by the energy spectrum tensor as [73, Chapter 3] (see Eq. D.28a)
Z +∞
Bij (x1 , x2 , x3 )) = Ψij (κ) exp(ıκm x̂m )dκ1 dκ2 dκ3 (K.1)
−∞

where x̂m and κm are the separation vector the two points and the wavenumber vector,
respectively. The complex Fourier transform exp(ıκm x̂m ) is defined in Appendix D
(see Eq. 18.8). The two-point correlation, Bij , and the energy spectrum tensor, Ψij ,
form a Fourier-transform pair (see Eq. D.28b)
Z +∞
1
Ψij (κ) = Bij (x̂) exp(−ıκm rm )dx̂1 dx̂2 dx̂3 (K.2)
(2π)3 −∞
The separation between the two points is described by a general three-dimensional
vector, x̂m . Both in experiments and in LES it is usually sufficient to study the two-
point correlation and the energy spectra along a line. Hence, one-dimensional energy
spectra, Eij (κ), which are a function of scalar wavenumber, κ (κ1 , κ2 or κ3 ), are often
used. They are formed by integrating over a wavenumber plane; the energy spectrum
for the wavenumber κ1 , for example, reads
Z
1 +∞
Eij (κ1 ) = Ψij (κ)dκ2 dκ3 (K.3)
2 −∞

A factor of two is included because E ∝ Ψii /2 is used to define a energy spectrum

for the turbulent kinetic energy k = vi′ vi′ /2, see Eqs. K.8 and K.10. Note that the
maximum magnitude of the wavenumber vector contributing to Eij (κ1 ) is very large
since it includes all κ2 and κ3 , i.e. −∞ < κ2 < ∞ and −∞ < κ3 < ∞. The
one-dimensional two-point correlation, Bij (x̂1 ), for example, and the one-dimensional
spectrum, Eij (κ1 ), form a Fourier-transform pair, i.e.
Z
1 +∞
Bij (x̂1 ) = Eij (κ1 ) exp(ıκ1 x̂1 )dκ1 (K.4)
2 −∞
Z +∞
2
Eij (κ1 ) = Bij (x̂1 ) exp(−ıκ1 x̂1 )dx̂1 (K.5)
2π −∞
where Eij is twice the Fourier transform of Bij because of the factor two in Eq. K.3.
The diagonal components of the two-point correlation tensor are real and symmetric
and hence the antisymmetric part of exp(−ıκ1 x̂1 ) – i.e. the sinus part – is zero and
Eqs. K.4 and K.5 are simplified as
Z Z ∞
1 +∞
Bij (x̂1 ) = Eij (κ1 ) cos(κ1 x̂1 )dκ1 = Eij (κ1 ) cos(κ1 x̂1 )dκ1
2 −∞ 0
Z Z
1 +∞ 2 +∞
Eij (κ1 ) = Bij (x̂1 ) cos(κ1 x̂1 )dx̂1 = Bij (x̂1 ) cos(κ1 x̂1 )dx̂1
π −∞ π 0
(K.6)

for i = j. The Reynolds stress ρv1′2 , for example, is equal to the two-point correlation
tensor ρBij with with zero separation distance. The v1′2 can be computed both from the
K.1. Energy spectra from two-point correlations 293

three-dimensional spectrum (Eq. K.1 ) and one-dimensional spectrum (Eq. K.6)

Z +∞
v1′2 = B11 (x1 , 0, 0) = Ψ11 (κ)dκ1 dκ2 dκ3
−∞
Z ∞ (K.7)
v1′2 = B11 (0) = E11 (κ1 )dκ1
0

Hence the turbulent kinetic energy, k = vi′ vi′ /2, an be written as

Z +∞
1
k= Ψii (κ)dκ1 dκ2 dκ3 (K.8)
2 −∞
Z ∞ Z ∞ Z
1 1 1 ∞
k= E11 (κ1 )κ1 + E22 (κ2 )κ2 + E33 (κ3 )dκ3 (K.9)
2 0 2 0 2 0

The integral in Eq. K.8 has no directional dependence: it results in a scalar, k. Instead
of integrating over dκ1 dκ2 dκ3 we can integrate over a shell with radius κ = (κi κi )1/2
and letting the radius go from zero to infinity, i.e.
Z
1 ∞
k= 4πκ2 Ψii dκ (K.10)
2 0

where 4πκ2 is the surface area of the shell. We now define an energy spectrum, E(κ) =
2πκ2 Ψii so that Z κ
k= E(κ)dκ. (K.11)
0

The energy spectrum E11 (κ1 ), for example, corresponds to the square of the Fourier
coefficient of the velocity fluctuation (see Parseval’s formula, Eq. D.4), i.e.

E11 (κ1 ) = v̂12 (κ1 ) (K.12)

Below the properties of the three energy spectra are summarized.

• The three-dimensional spectrum tensor, Ψij (κ), is a tensor which is a function

of the wavenumber vector.
• The one-dimensional spectrum, Eij (κ1 ), is a tensor which is a function of a
scalar (one component of κm ).
• The energy spectrum, E(κ), is a scalar which is a function of the length of the
wavenumber vector, |κ| ≡ κ.

K.1 Energy spectra from two-point correlations

In connection to Eqs. K.4, K.5 and K.6 we stated that the one-dimensional energy spec-
tra and the two-point correlations form Fourier-transform pairs. The proof is given in
this section. The energy spectrum is given by the square of the Fourier coefficients, see
Parseval’s formula, Eq. D.4. Let û be the Fourier coefficient of the velocity fluctua-
tion u′ in the x direction which is periodic with period 2L. Take the covariance of the
K.1. Energy spectra from two-point correlations 294

Fourier coefficients, û(κ′ ) and û(κ) where κ and κ′ denote two different wavenumbers
and x and x′ denote two points separated in the x directions so that
* Z L Z L +
′ 1 1 ′ ′ ′ ′
hû(κ)û(κ )i = u(x) exp(−ıκx)dx u(x ) exp(−ıκ x )dx
2L −L 2L −L
* Z LZ L +
1 ′ ′ ′ ′
= u(x)u(x ) exp(−ı(κx + κ x )dxdx
4L2 −L −L
(K.13)

where h·i denotes averaging over time; this equation corresponds to Eq. K.4 except the
factor of two. Since we are performing a Fourier transform in x we must assume that
this direction is homogeneous, i.e. all averaged turbulence quantities are independent
of x and the two-point correlation is not dependent of x (or x′ ) but only on the sepa-
ration distance x − x′ , see discussion in connection to Eq. 10.6 on p. 111. Hence we
replace x′ by x + x′′ so that
* Z L Z L−x ! +
′ 1 ′′ ′ ′′ ′′
hû(κ)û(κ )i = u(x)u(x + x ) exp(−ı(κx + κ (x + x ))dx dx
4L2 −L −L−x
* Z L Z L−x ! +
1 ′ 1 ′′ ′ ′′ ′′
= exp(−ı(κ + κ )x) B11 (x ) exp(−ıκ x ))dx dx
2L −L 2L −L−x
(K.14)

The second integral (in parenthesis) is the Fourier transform (which corresponds to the
Fourier coefficient in discrete space, see Eq. D.28b) of the two-point correlation B11 ,
i.e. * +
Z L
′ ′′ 1 ′
hû(κ)û(κ )i = B̂11 (x ) exp(−ı(κ + κ )x))dx (K.15)
2L −L

where B̂11 denotes the Fourier transform of B11 (cf. K.5) and since it does not depend
on the spatial coordinate it has been moved out of the integral. Furthermore, B̂11 is real
and symmetric since B11 is real and symmetric. The remaining integral in Eq. K.15
includes trigonometric function with wavelengths κ and κ′ . They are orthogonal func-
tions, see Appendix D, and the integral of these functions is zero unless κ = κ′ . The
integral in Eq. K.15 for κ = κ′ is evaluated as (see “length of of ψk ” in Appendix D,
Eq. D.13, and use ψ1 = cos(2πx/L))
Z L  
2πx
(ψ1 |ψ1 ) = ||ψ1 ||2 = cos2 dx
−L L
  L (K.16)
x L 4πx
= + sin =L
2 8π L −L

In the same way, the “length of of φk ” in Eq. D.16 is also L. Equation K.15 can now
be written
1
hû(κ)û(κ)i = (L + L)hB̂11 (x)i = hB̂11 (x)i (K.17)
2L
Hence, it is seen that the Fourier transform of a two-point correlation (in this example
hB11 (x1 )i) indeed gives the corresponding one-dimensional energy spectrum (in this
example E11 (κ1 ) = h(û(κ))2 i).
 L. MTF270, Assignment 1: Reynolds averaged Navier-Stokes 295

L MTF270, Assignment 1: Reynolds averaged Navier-

Stokes
L.1 Part I: Data of Two-dimensional flow
OU can do the assignment on your own or in a group of two. It is recommended
Y (but the not required) that you use LATEX(an example of how to write in LATEXis
available on the course www page). It is available on Linux. On Windows you can use,
for example, ShareLatex (https://www.sharelatex.com) Lyx (www.lyx.org)
or MikTex (www.miktex.org) which are both free to download.
You’ll use data from a coarse DNS. Although some of the data are probably not
fully accurate, in this exercise we consider the data to be exact. You will use Matlab.
You can also use Octave. Octave is a Matlab clone which can be downloaded for free.
Use Matlab or Octave to read data files of the mean flow (v̄1 , v̄2 , p̄) and turbulent
quantities. (v1′2 , v2′2 , v3′2 and v1′ v2′ ). You will analyze one of the following flows:
Case 1: Flow over a wavy wall (small wave) [185, 185]. Re = 10 000 (ν = 1 · 10−4 ,
ρ = 1) based on the bulk velocity in the channel at x = 0 and the channel height.
Case 2: Flow over a wavy wall (large wave) [185, 185]. Re = 10 000 (ν = 1 · 10−4 ,
ρ = 1) based on the bulk velocity in the channel at x = 0 and the channel height.
Case 3: Flow over a hill Re = 10 595 (ν = 1/10595, ρ = 1) based on the bulk velocity
in the channel at x = 0 and the channel height.
Case 4: Flow over two hills Re = 10 595 (ν = 1/10595, ρ = 1) based on the bulk
velocity in the channel at x = 0 and the channel height.
Periodic boundary conditions are imposed in streamwise (x1 ) and spanwise (x3 )
directions in all flows.
Get the case the case number by register in a group at Ping-pong. Download the
data from
http://www.tfd.chalmers.se/˜lada/comp turb model. At the www-
page you can download a M-file (pl vect.m) which reads the data and plots the vec-
tor field and the pressure contours. You must also download the function dphidx dy.m
which computes the gradients, see Section L.2. Make sure you put this function in the
directory where you execute pl vect.m.
The report, along with the Matlab files(s), should be submitted electronically at
Ping-pong; the deadline can be found in the time table.
The number of A4 pages in the report should not exceed 15. Font size must be 11pt
or larger. The minium size of each figure must not be smaller than 7cm.

L.1.1 Analysis
Study the flow. In which regions do you expect the turbulence to be important? Let’s
find out. The two-dimensional time-averaged Navier-Stokes for the x1 momentum
reads (the density is set to one, i.e. ρ = 1)
∂v̄1 v̄1 ∂v̄1 v̄2 ∂ p̄ ∂ 2 v̄1 ∂v ′2 ∂ 2 v̄1 ∂v ′ v ′
+ =− +ν 2 − 1 +ν 2 − 1 2
∂x1 ∂x2 ∂x1 ∂x1 ∂x1 ∂x2 ∂x2
(L.1)
∂v̄1 v̄2 ∂v̄2 v̄2 ∂ p̄ ∂ 2 v̄2 ∂v1′ v2′ ∂ 2 v̄2 ∂v2′2
+ =− +ν − + ν −
∂x1 ∂x2 ∂x2 ∂x21 ∂x1 ∂x22 ∂x2
 L.1. Part I: Data of Two-dimensional flow 296

Recall that all the terms on the right-hand side represent x1 components of forces per
unit volume.

L.1.2 The momentum equations

The file pl vect.m loads the data file and plots the profiles of v1′2 at some x stations,
the velocity vector field and a contour plot of velocity gradient ∂v̄1 /∂x2 . Find two
x1 locations (vertical grid lines) where the v1′2 stress have its minimum and maximum,
respectively. One way to find these locations is to use the Matlab surf command.

Assignment 1.1. Plot the stresses along vertical grid lines at these two locations using the Matlab
command plot(x,y). Please make sure that in your report the numbering on
the axis and the text in the legend is large enough; you can use the command

h1=gca;
set(h1,’fontsize’,[20]) %the number ’20’ gives the fontsize

The size of the labels and the title is similarly controlled by

xlabel(’x/H’,’fontsize’,[20])
ylabel(’y/H’,’fontsize’,[20])
title(’velocity’,’fontsize’,[20])

Assignment 1.2. Compute and plot all terms in Eq. L.1 (see Fig. 6.6). You will need to compute
the derivatives of e.g. v̄1 and p̄. In pl vect.m the function dphidx dy.m
is used to compute ∂v̄1 /∂x1 and ∂v̄1 /∂x2 . Use this function to compute all
derivatives that you need. To enhance readability you may omit the small terms
or use two plots per vertical grid line. Make also a zoom near the walls. For
example, for a x − y plot

plot(u,y,’linew’,2) % linewidth=2

you may want to zoom in on y=[0 0.01] and u=[-0.1 0.4]; this is achieved
by

axis([-0.1 0.4 0 0.01])

The ’axis’ command can be used together with any plot, e.g. with ’surf’ and
’quiver’.
Which terms are negligible? Can you explain why they are negligible?

What about the viscous terms: where do they play an important role? Which
terms are non-zero at the wall? (you can easily show that with paper and pen).
So far we have looked at the v̄1 -momentum equation. The database corresponds
to a two-dimensional flow. Now let’s think of the forces as vectors. The gradient
of the normal stresses in the x1 −x2 plane represent the force vector (see Eq. L.1)
!
∂v1′2 ∂v2′2
FN = − ,− (L.2)
∂x1 ∂x2
 L.1. Part I: Data of Two-dimensional flow 297

and the corresponding force vector due to the shear stresses reads (see Eq. L.1)
!
∂v1′ v2′ ∂v1′ v2′
FS = − ,− (L.3)
∂x2 ∂x1

Find the first term in Eqs. L.2 and L.3 in the v̄1 momentum equation, Eq. L.1.
Consider the second line in Eq. L.1 which is the v̄2 equation and find the other
two terms in Eqs. L.2 and L.3. Note that FN and FS are forces per unit volume
([N/m3 ]).

Assignment 1.3. Plot the vector field FN to learn something about its properties. When v2′2
reaches a maximum or a minimum along a grid line normal to the wall, what
happens with the vector field FN ? Zoom-in on interesting regions.
Assignment 1.4. Plot also vector fields of the shear stress, FS (see Eq. L.3), the pressure gradient
and the viscous terms. Zoom up in interesting regions. Anything interesting?

L.1.3 The turbulent kinetic energy equation

The exact transport equation for the turbulent kinetic energy, k, reads

∂ ∂2k
(v̄j k) = ν + P k + Dk − ε
∂xj ∂xj ∂xj
(L.4)
∂v̄i
P k = −vi′ vj′
∂xj

Assignment 1.5. Plot the production term along the two grid lines (see Fig. 8.3). Explain why
it is large at some locations and small at others. The production term consists
of the sum of four terms, two of which involve the shear stress while the other
include the normal stresses. Compare the contributions due the shear stress and
the normal stresses.
Assignment 1.6. Plot the dissipation and compare it with the production. Do you have local equi-
librium (i.e. P k ≃ ε) anywhere? Take ε from 2D RANS simulation.

L.1.4 The Reynolds stress equations

The modeled transport equation for the Reynolds stresses can be written as

∂  ′ ′ ∂ 2 vi′ vj′
v̄k vi vj = ν + Pij + Φij + Dij − εij
∂xk ∂xk ∂xk (L.5)
∂v̄j ∂v̄i
Pij = −vi′ vk′ − vj′ vk′
∂xk ∂xk
The pressure-strain term, Φij , use the models in Eqs. 11.91, 11.57, 11.90, 11.95 and
11.96.

1. In the damping function, f (see Eq. 11.92), |nk,w (xk − xk,w )| denotes the dis-
tance to the nearest wall. If, for example, a lower wall is the closest wall to node
(I, J), then
 1/2
|nk,w (xk − xk,n )| = (x(I, J) − x(I, 1))2 + (y(I, J) − y(I, 1))2 (L.6)
 L.1. Part I: Data of Two-dimensional flow 298

Note that you have to search through all wall nodes to find which wall node gives
the smallest value. The damping function, f , involves k and ε. These should be
taken from the 2D RANS simulation (they are loaded in pl vect.m).
2. If we assume, again, that the lower wall is the closest wall to cell (I, J) and that
the lower wall is horizontal, then ni,w = (0, 1). To compute ni,w for the general
case (see Eqs. 11.95 and 11.96), compute first the vector which is parallel to the
wall, si,w , and compute then ni,w from si,w (see Eq. L.11)
3. The diffusion terms Dij and Dε can be modeled using the Generalized Gradient
Diffusion Hypothesis GGDH of [186]
!
∂ k ∂vi′ vj′
Dij = cuk um (L.7)
∂xm ε ∂xk

This diffusion model can cause numerical problems, and the GGDH is then re-
placed by a simple eddy viscosity model
!
∂ νt ∂vi′ vj′
Dij = , νt = Cµ k 2 /ε (L.8)
∂xm σk ∂xm

The following constants should be used:

(cµ , c1 , c2 , c1w , c2w , σk ) = (0.09, 1.5, 0.6, 0.5, 0.3, 1)

4. The dissipation is modelled as (see Section 11.7.3)

2
εij = εδij
3
Assignment 1.7. Choose two stresses. Plot the different terms in the equations for one vertical
grid line fairly close to the inlet (not too close!), compare Fig. 9.1. Use the
simple eddy viscosity model for the turbulent diffusion term. Use k and ε from
the 2D RANS simulation. If the figure becomes too crowdy, use two plots per
vertical grid line or simply omit terms that are negligible. Try to explain why
some terms are large and vice versa. Usually, a stress is large in locations where
its production (or pressure-strain) term is large. Is that the case for you?
Assignment 1.8. Compute the stresses using the Boussinesq assumption, i.e vi′ vj′ = −2νt s̄ij +
(2k/3)δij where νt = cµ k 2 /ε. Use k and ε from the 2D RANS simulation.
Compare the eddy-viscosity stresses with two of the Reynolds stresses from the
database. Make also a zoom-in near walls.

When using the Boussinesq assumption the production of turbulent kinetic energy
(use k and ε from the 2D RANS simulation)

P k = 2νt s̄ij s̄ij (L.9)

is always positive. The exact production of turbulent kinetic energy (see Eq. L.4) is
usually positive. It can however become negative.

Assignment 1.9. Compute the exact production in Eq. L.4 in the entire domain to investigate if the
production is negative anywhere. If so, explain why (for physical explanation,
see Item II at p. 93).
 L.2. Compute derivatives on a curvi-linear mesh 299

The reason why the eddy-viscosity production in Eq. L.9 must be positive is that
neither νt nor s̄ij s̄ij can go negative. Another way to explain this fact is that the
modeled Reynolds stress, vi′ vj′ , and the strain rate tensor, s̄ij are parallel. To find out to
what degree the exact Reynolds stress and the strain rate are parallel, one can compute
the eigenvectors.

Assignment 1.10. The realizability concept is discussed in Section 13. It is shown that the Boussi-
nesq assumption can give unphysical negative normal Reynolds stresses (if ∂v̄1 /∂x1
in Eq. 13.2 is very large then v1′2 goes negative). A remedy is presented in Sec-
tion 13 in which the turbulent viscosity is limited using the eigenvaues of the
strain-rate tensor, s̄ij , see Eq. 13.12. Compute the eigenvalues s̄ij in the entire
domain. Check if the limiter in Eq. 13.12 reduces νt along the two vertical lines
you previously chose. Find out where in the entire domain the limiter has largest
effect. Is this region a stagnation region? (the object of Eq. 13.12 is to limit νt
in stagnation regions, see Section 12.3).
Assignment 1.11. In the course we have discussed a number of ways how to obtain the turbulent
Reynolds stresses. The most drastic simplification is the Boussinesq assumption.
A slightly more elaborate way is to use the algebraic stress model or a non-linear
turbulence model.
Compute the Reynolds stresses using the explicit Reynolds stress model (see
Section 11.12) and the non-linear k − ε model (see Section 14) and plot them
along the the two vertical lines you previously chose. Which model agrees best
with DNS? Is there any improvement compared to the Boussinesq assumption?

L.2 Compute derivatives on a curvi-linear mesh

In this section we describe how the derivatives on a curvi-linear grid are computed
in the provided Matlab function dphidx dy.m. On a Cartesian grid it is more con-
venient to use the built-in Matlab function gradient, but the approach used below
works for all meshes, including Cartesian ones.
The data you have been given, x1 and x2 and all variables, are stored at the grid
points, i.e. at (x1,sw , x2,sw ), (x1,se , x2,se ), (x1,nw , x2,nw ) and (x1,ne , x2,ne ). When
you need a variable, say v1 , at the center of the cell, compute it as
1
v1,P = (v1,sw + v1,se + v1,nw + v1,ne ) (L.10)
4
Let’s compute ∂v1 /∂x1 . In order to do that we use Gauss’ law over a control
volume centered at face e (dashed control volume in Fig. L.1). The divergence theorem
for a scalar, φ, reads Z Z
∂φ
dV = φni dA
V ∂xi A
To compute ∂v1 /∂x1 we set φ = v1 and i = 1 which gives
Z Z
∂v1
dV = v1 n1 dA
V ∂x1 A

Assuming that ∂v1 /∂x1 is constant in the volume V we obtain

Z
∂v1 1
= v1 n1 dA
∂x1 V A
L.2. Compute derivatives on a curvi-linear mesh 300

ne
nw
se = (s1e , s2e )

P
ne = (n1e , n2e )
E
W

sw
se

S
Figure L.1: Control volume. The velocity v1 is stored at the corners (ne, nw, . . . ).
Coordinates x1 , x2 are given at the corners (ne, nw, ...).

In discrete form we can write (see Fig. L.1)

  X
∂v1 1
= (v1 n1 A)i =
∂x1 V i=e,n,w,w

1
{(v1 An1 )e + (v1 An1 )n + (v1 An1 )w + (v1 An1 )s }
V

L.2.1 Geometrical quantities

It is useful to first compute the unit vectors s along the control volume. For the east
face, for example, we get
x1,ne − x1,se
s1e =
de

x2,ne − x2,se
s2e =
de

q
de = (x1,ne − x1,se )2 + (x2,ne − x2,se )2

(note that the area of the east face Ae is equal to de since ∆z = 1). The relation
between the normal vector n, s and the unit vector in the z-direction
s·n =0
s × ẑ = n,
gives us the normal vector for the east face as
n1e = s2e
(L.11)
n2e = −s1e .
 L.3. Part II: StarCCM+ 301

L.3 Part II: StarCCM+

In this task, a commercial CFD software (StarCCM+ 11.02.009) will be used. The
task is to do simulation of an asymmetric diffuser. Several turbulence models will be
used and the results will be compared with experimental data. Before doing the task,
it is recommended to first do one of the tutorials in the StarCCM+. The tutorial is the
steady backward facing step tutorial. The tutorials can be found in the StarCCM+ user
guide.

L.3.1 Backward Facing Step Tutorial

This tutorial is a good bridge before doing the asymmetric diffuser case. Here are some
steps to access the tutorial:
1. Open a terminal window. In the terminal window, type starccm+
2. To start a new simulation, click File → new simulation

3. Tick the Power-On-Demand box and fill the license box with thePOD code. The
POD code can be found on Ping Pong
4. Download the tutorial instruction and data from the course homepage

Assignment 1.13. Continue to work on the backward facing step flow. Choose a new turbulence
model (right click on Continua/Fluid and select Select models ....
Untick All y+ Wall treatment and then untick the turbulence model.
Now you can choose a new turbulence model. Run a couple different turbulence
model and compare the results. A short recirculation region is usually connected
to high turbulence in the shear layer bounding the recirculation region.
Assignment 1.14. Compare the turbulence levels in the shear layer for different turbulence models.
How much larger is the turbulence in the shear layer compared to the flow in the
boundary layer upstream of the step? How large is the turbulence intensities in
the two regions. Plot the turbulence production, P k . Do this by going to the
solvers folder, segregated flow and tick Temporary Storage Retained
Run one iteration to create the date and then open a Scalar Scene and plot the
turbulence production. Make other interesting comparisons!

L.3.2 Asymmetric diffuser case

In this case 2D Asymmetric diffuser will be studied. The Reynolds number (Re =
20000) is based on the bulk velocity and the inlet channel height. The case is built
based on these references. It is highly recommended to read these references before
doing the case :

1. Ercoftac case 8.2: Flow through an asymmetric plane diffuser

Ercoftac database.
2. El-Behery & Hamed [187]
3. Buice & Eaton [188]
4. Davidson [150].
L.3. Part II: StarCCM+ 302

L.3.3 Brief instruction to begin the asymmetric diffuser case

1. Start the StarCCM+
• Open a terminal window → type StarCCM+
2. Create a new simulation
• Click the new simulation icon → tick the power on demand box
• Fill the license box with the POD license found on Ping Pong
3. Draw the geometry
• Follow and modify the geometry in the ercoftac case 8.2
• Create a new 3D-CAD Models
• Create a new sketch on XY plane
• Start by drawing the asymmetric diffuser without the radius
• Right click on each lines and apply vertical / horizontal constraint accord-
ingly
• Set the diffuser angle to 10 degree
• Set the length of each lines. Use H = 1 m. Length of the inlet = 60H.
Length of the diffuser = 21H. Length of the outlet = 22H. Do not need to
set the height of the outlet.
• Create a fillet on the beginning and end of the diffuser line. (R = 9.7H)
• Set your base point (which is the point where the channel start to bend) at
X = 0 and Y = 3.7 m
• Extrude the sketch 1 m for easier renaming
• Rename the faces accordingly (inlet, outlet, top, bottom, leftside, rightside)
4. Create a new geometry part. (The same way as in backward facing step tutorial)
5. Assign parts to regions. (The same way as in backward facing step tutorial)
6. Set the boundary condition accordingly. Inlet = velocity inlet; Outlet = pressure
outlet; leftside = symmetry plane; rightside = symmetry plane.
7. Create directed mesh
• Right click on the operation → new → directed mesh → body1 → OK
• Expand operations → Right click on the directed mesh → edit
• Add rightSide as a source surfaces and leftSide as target surfaces
• Right click on source meshes → new source mesh → patch mesh → body1
→ OK
• Click auto populate feature edges icon
• Change the mode from patch topology into patch mesh
• Click the inlet line
– Number of divisions = 80
– Type = two sided hyperbolic
L.3. Part II: StarCCM+ 303

– Spacing start and end = 3e-4 m

• Click the top line
– Number of divisions = 421
– Type = constant
• Right click on the mesh distributions → create new volume distribution.
Set the number of layers = 2
• Close the directed mesh and execute the directed mesh
8. Create 2D Mesh
• Go to Menu bar → Mesh → Convert to 2D → Delete 3D regions after
conversion → OK
9. Select turbulence model
• Right click on Physics 1 2D → Select model
• In the new window select:
(a) Untick the auto-select recommended models
(b) Two dimensional
(c) Steady
(d) Gas
(e) Segregated flow
(f) Constant Density
(g) Turbulent
(h) Reynolds-Averaged Navier-Stokes
(i) Choose turbulence model that you want to use.
10. Set the Gas properties
• Expand Physics 1 folder → Gas → Air → Material Properties → Dynamic
Viscosity → Constant → Value = 6E − 5P a − s
11. Set the boundary condition
• Go to Regions.
• In the properties window, check that mesh continuum is parts meshes.
Check that physics continuum is physics 1 2D
• Expand Body 1 2D → Boundaries → Inlet
(a) Physics Conditions → Velocity Specification → Method → Compo-
nents
(b) Physics Values → Velocity → value → [1,0,0] m/s
12. Extract more data from the simulation (for example: Turbulent kinetic energy
production)
• Expand Solvers folder → go to every subfolder and tick temporary storage
retained if you find one
13. Set the stopping criteria
L.3. Part II: StarCCM+ 304

• For Reynolds Stress Model, use the default setting of the stopping criteria
• For the rest of the model use instruction below
• Disabled the maximum iteration criteria
• Create new stopping criteria
(a) Right click on the Report folder → go to New Report → click Pressure
Drop
(b) In the Pressure Drop 1 properties window. High Pressure = Outlet.
Low Pressure = Inlet.
• Right click on the Pressure Drop 1 → click on create monitor and plot from
report
• Right click on the Stopping criteria folder →choose create new criterion →
create from monitor → click on Pressure Drop 1 monitor
• In the Pressure Drop 1 Monitor criterion properties window
(a) Criterion option = Asymptotic
(b) Click on Asymptotic limit
• set [max-min] value = 1E − 4 → Number of samples = 100
14. Create a scalar scene to show the result
• Go to Scenes folder → New Scene → Scalar
• Right click on the select function → select velocity [i]
15. Create planes to extract the results at the same location as the experimental data
• Right click on Derived Parts folder → New Part → Section → Plane
• Input Parts → Select → All the region
• Set normal → x = 1, y = 0, z = 0
• Display → New Geometry Displayer
• Specify the x-coordinate according to the experimental data (see below)
• Click Apply for each of the x-coordinate
(a) X − 6 = −5.87m
(b) X03 = 2.59m
(c) X06 = 5.98m
(d) X14 = 13.56m
(e) X17 = 16.93m
(f) X20 = 20.32m
(g) X24 = 23.71m
(h) X27 = 27.09m
(i) X30 = 30.48m
(j) X34 = 33.87m
16. Create the X-Y plot to monitor the result
• Right click on Plots folder → choose New Plot → click on X-Y Plot
 L.3. Part II: StarCCM+ 305

• On the new X-Y Plot properties window, go to Parts → select all the derived
parts
• Expand the X-Y Plot 1
(a) X-type → Type = Scalar → Scalar Function → Field Function = Ve-
locity[i]
(b) Y-types → Y Type 1
– Type = Direction; smooth values
– Vector Quantity → Value = [0,1,0]
– Go to each Plane Section and change the X Offset accordingly
17. Insert experimental data
• Download the experimental data from the course homepage
• Expand Tools folder
• Right click on Tables folder → choose Create a New Table → click File
Table → Load all the experimental data
• Go back to X-Y plot 1 monitor → right click on Data Series folder → Add
Data → Select all the experimental data
• Expand Data series folder → select all the experimental data by click the
first data hold shift button and click the last data → right click on the high-
lighted area → click swap column
18. Run the simulation
• Go to Solution → Run
• Other method: click icon which show a running man.
19. Collecting the simulation data
• After the simulation has reached its stopping criteria → Right click on XY
plot 1 → Export → Save with appropriate name
• Add the simulation data to other simulation as you add experimental data

Assignment 1.14. Look at the results. How large is the pressure recovery (i.e. how much does the
pressure increase from inlet to outlet).
Assignment 1.15. Choose a new turbulence model Run the same turbulence models as for the back-
step case and compare the results. Are the results different? Where do the differ-
ence appear? In the plane channel upstream of the diffuser? Compare turbulent
quantities such a k and νt (or if you chose to use a Reynolds stress model, com-
pare the shear stresses v1′ v2′ . If the recirculation region is different, the reason
is maybe connected to the turbulence level in the shear layer above the recir-
culation region. Or maybe the reason is that the incoming boundary layers are
different. In the backstep flow, you found that one turbulence model was better
than the other one(s). Is the same turbulence model best for this flow? Make
other interesting comparisons!
M. MTF270, Assignment 2: LES 306

M MTF270, Assignment 2: LES

OU can do the assignment on your own or in a group of two. You will receive
Y data from a DNS of fully developed flow in a channel. It is recommended (but
not required) that you use LATEX(an example of how to write in LATEXis available on
the course www page). It is available on Linux. On Windows you can use, for exam-
ple, Lyx (www.lyx.org) or MikTex (www.miktex.org) which are both free to
download.
The report, along with the Matlab files(s), should be submitted electronically.
The number of A4 pages in the report should not exceed 15. Font size must be 11pt
or larger. The minium size of each figure must not be smaller than 7cm.
The equations that have been solved are

∂vi
=0
∂xi
(M.1)
∂vi ∂ ∂p 1 ∂ 2 vi
+ (vi vj ) = δi1 − +
∂t ∂xj ∂xi Reτ ∂xj ∂xj

The Re number based on the friction velocity and the half channel width is Reτ =
uτ h/ν = 500 (h = ρ = uτ = 1 so that ν = 1/Reτ ).
A 96 × 96 × 96 mesh has been used. The streamwise, wall-normal and spanwise
directions are denoted by x (x1 ), y (x2 ) and z (x3 ) respectively. The cell size in x and
z directions are ∆x = 0.0654 and ∆z = 0.0164. Periodic boundary conditions were
applied in the x and z direction (homogeneous directions). The size of the domain is
(L, h, Zmax) in (x, y, z), see Fig .M.1.

M.1 Time history

At the www-page
http://www.tfd.chalmers.se/˜lada/comp turb model/
you find a file u v time 4nodes.dat with the time history of u and v. The file has
eight columns of u and v at four nodes: y/δ = 0.0039, y/δ = 0.0176, y/δ = 0.107
and y/δ = 0.47. With uτ = 1 and ν = 1/Reτ = 1/500 this correspond to y + = 1.95,
y + = 8.8, y + = 53.5 and y + = 235. The sampling time step is 0.0033.
Use Matlab or Octave. Octave is a Matlab clone which can be downloaded for free.
Download the Matlab/Octave program pl time.m which loads and plots the time
history of u. Run the program pl time.m. Recall that the velocities have been scaled
with the friction velocity uτ , and thus what you see is really u/uτ . The time history
of u at y/δ = 0.0176 and y/δ = 0.107 are shown. Study the time history of the blue
line (y/δ = 0.0176) in greater detail. Make a zoom between, for example, t = 10 and
t = 11 and umin = 3 and umin = 21. This is conveniently done with the command

axis([10 11 3 21])

In order to see the value at each sampling time step, change the plot command to

plot(t,u2,’b-o’)

Use this technique to zoom and to look at the details of the time history. Alterna-
tively, you can use the zoom buttons above the figure.
M.2. Time averaging 307

Plot u for all four nodes. How does the time variation of u vary for different posi-
tions? Why? Plot also v at the four different positions. What is the difference between
u and v?

M.2 Time averaging

Compute the average of the u velocity at node 2. Add the following code (before the
plotting section)

umean=mean(u2)

Here the number of samples is n = 5000 (the entire u2 array). Find out how many
samples must be used to get a correct mean value. Start by trying with 100 samples as

umean_100=mean(u2(1:100))

Do the same exercise for the other three nodes.

M.3 Auto correlation

Auto correlation is defined in Section 10.2. Compute the autocorrelation for u′ at Node
1 using the Matlab command

imax=500;
two_uu_1_mat=autocorr(u1,imax);

where we set the maximum separation in time to imax = 500 (i.e. we carry out the
integration in Eq. 10.11 not to infinity, but to imax · ∆t). Note that the autocorr
norm
command returns the normalized autocorrelation, i.e. B11 , see Eq. 10.10. Plot the
autocorrelation as

plot(t(1:imax),two_uu_1_mat(1:imax),’linew’,2)
xlabel(’t’)
ylabel(’$B_{11}ˆ{norm}$’,’Interpreter’,’latex’)
handle=gca
set(handle,’fontsi’,[20])

Compute the integral time scale Tint as

dt=t(1);
int_T_1=trapz(two_uu_1_mat)*dt;

Plot the normalized autocorrelation and compute the integral time scales also for
the other three points. Where is Tint large and small, respectively? Try to explain why.
In Section M.2 you time averaged the velocities to get the mean value. You in-
vestigated how few samples you could use. In reality it is not only the number of
samples that is relevant, but also that they are independent. To find out if two samples
are independent, it is convenient to use the integral time scale, Tint . If the samples
are separated by Tint seconds they are independent. Hence, re-do the averaging you
did in Section M.2 but use samples every Tint second. The theoretical estimate of the
statistical error varies with number of independent samples, N , as
1 v1,rms
error = √ (M.2)
N hv1 i
M.4. Probability density/histogram 308

We see that the statistical error decreases as N −1/2 , provided that the samples are
independent.
Let us use Taylor’s frozen turbulence hypothesis to compute the integral length
scale. This hypothesis assumes that – if the turbulence level is not too strong – the
velocity fluctuation at point x and time t is the same as that at time (t − τ ) at point
(x−ξ) where τ = (x−ξ)/hui (it takes time τ for the particle to travel from point (x−ξ)
to point x with a velocity hui). The hypothesis assumes that the turbulence is frozen
between point (x − ξ) and x. When we want to find the velocity fluctuation at point
(x − ξ) at time (t − τ ) we can instead take it at point x at time t. The Taylor hypothesis
makes it possible to compute the integral lengthscale from the integral timescale as (see
Eq. 10.12) Z ∞
norm
Lint = hui B11 (t̂)dt̂ = huiTint (M.3)
0
Compute the integral lengthscale.

M.4 Probability density/histogram

Histogram (also called probability density function, PDF) can give additional useful
information, see Section 7. With a probability density, fv , of the v velocity, the mean
velocity is computed as Z ∞
hvi = vfv (v)dv (M.4)
−∞
Normalize the probability density function so that
Z ∞
fv (v)dv = 1 (M.5)
−∞

Here we integrate over v. The mean velocity can of course also be computed by
integrating over time, as we do when we define a time average. Compute the PDF as

u3_fluct=u3-mean(u3);
[pdf3 u3_pdf]= hist(u3_fluct,20)

Here we have divided u3 into 20 bins which span the variation of u3 , i.e. [min(u3 ), max(u3 )].
The variable pdf 3 is a vector of length 20 whose elements gives the number of samples
in each bin. Plot the histogram as

norm3=sum(pdf3)*(u3_pdf(2)-u3_pdf(1));
plot(u3_pdf,pdf3/norm3,’linew’,2)
xlabel(’u3’)
ylabel(’PDF’)
handle=gca
set(handle,’fontsi’,[20])

where norm3 is the integral in Eq. 7.3. You find that the PDF is rather symmetric.
Compute and plot the PDFs of the points close to the wall and you will find that they
are more skewed. Skewness, S, is a variable that quantify the skewness and it is defined
as Z ∞
1
Sv ′ = 3 v ′3 fv′ (v ′ )dv ′
vrms −∞
Compute it as
M.5. Frequency spectrum 309

urms3=std(u3_fluct);
S=mean(u3_fluct.ˆ3)/urms3ˆ3;
where urms3 = hu′2 i1/2 at Node 3. Verify that the magnitude of S is large close to
the wall. What does a negative skewness mean physically? (cf. Fig. 7.2) Do you have
any explanation why the skewness increases as we go closer to the wall?

M.5 Frequency spectrum

One way to verify that the LES you have performed resolves the turbulence properly,
is to look at the spectra of the resolved turbulence. One can analyze the time history of
a variable at a point. You do a FFT of that signal to get the Fourier coefficients ai and
then plot a2i . Then you get the frequency spectrum, i.e. how much energy resides in
each frequency. The other way to do it is to look at the instantaneous velocity along a
grid line, and do a FFT of that signal. Then you get the energy spectrum as a function
of the wave number, i.e. the inverse of the wave length. In this case you must average
over many instants to get a reasonably smooth spectrum.
The Matlab file pl spectrum.m does a FFT of the time history of fluctuating
velocity, u′ , (u = hui + u′ ) and plots a2i . The same data file as in the previous exercise
is used (u v time 4nodes.dat). It uses the Matlab function pwelch. Run the
program by typing
pl spectrum
As we mentioned in the Lecture Notes, in a well-resolved LES we want to have the
cut-off in the inertial subrange where the kinetic energy decays as the wave number (or
frequency) up to the power of −5/3. Thus we want the resolved turbulence to have
a behaviour like this for high frequencies. As you can see from the plot, this is the
case. Actually, you find that for even higher frequencies the kinetic energy decays even
faster. The decay at high frequencies occurs because in the LES which produced the
data we over-resolve in time compared to the resolution in space. This is usually the
case when the maximum CFL number (Courant-Friedrichs-Lewy condition) is set to
one, since CFL in many points is much smaller than one (CFL=1 in one cell means that
a fluid particle is transported across that cell during one time step). Investigate this by
using only, say, every 4th of the samples, i.e.

% 4*dt
m=4;
i=1:m:n;
dt=m*t(1);
[px_L,f_L]=pwelch(u4(i),nmax,[],[],1/dt);
plot(f_L,px_L,’k-’,’linew’,2)

Now you find that the spectrum became much more irrgular. The reason is that
pwelch chops the time history into segments each consisting of nmax samples, makes
an FFT and then averages the FFT:s. Since we are now only using every 4th sample, the
time signal is only 5000/4 = 1250 samples, which means that no averaging is done.
Decrease nmax to 256 and do it again. Has the region in which the energy spectrum
decays fast vanished? If not, use even coarser sampling (set m=8). You may note that
the low region of the frequency spectrum is also modified. The reason is that when we
increase nmax the time series on which pwelch makes FFT on has a larger total time,
T , and the lowest frequence is proportional to 1/T .
Plot the spectra for the three other points (make changes in pl spectrum.m).
M.6. Resolved stresses 310

periodic b.c.

x2 j
hv̄1 i
2h

x1 i L
(a) x1 − x2 plane.
periodic b.c.

x2 j
2h

x3 k X3,max

(b) x2 − x3 plane.

Figure M.1: Channel domain.

It should be mentioned that spectra may not be a reliable measure of resolution [117,
118]. Two-point correlations are usually better.

M.6 Resolved stresses

At the www-page (http://www.tfd.chalmers.se/˜lada/comp turb model)
you find data files with three instantaneous flow fields (statistically independent). The
data files include the instantaneous variables u (v1 ), v (v2 ), w (v3 )and p (made non-
dimensional by uτ and ρ). You find a Matlab/Octave program at the www-page which
reads the data and computes the mean velocity. The data files are Matlab binary files.
Since the data files are rather large, it is recommended that you do all tasks using only
data files ’1’. When everything works, then use also data files ’2’ and ’3’ averaging by
use of the three files.
We decompose the instantaneous variables in time-averaged and fluctuating quan-
tities as
vi = hvi i + vi′ , p = hpi + p′
The symbol h.i denotes averaging in the homogeneous directions x and z. Note that
in reality h.i always denote time averaging. It is only in this special academic test case
M.7. Resolved production and pressure-strain 311

where we have three homogeneous directions (x, z, t) where we can – in addition to

time averaging – also can use x and z averaging. Compute the six stresses of the stress
tensor, hvi′ vj′ i. Use the definition to compute the stresses, for example

hv1′ v2′ i = h(v1 − hv1 i) (v2 − hv2 i)i

= hv1 v2 i − hv1 hv2 ii − hv2 hv1 ii + hhv1 ihv2 ii (M.6)
= hv1 v2 i − 2hv1 ihv2 i + hv1 ihv2 i = hv1 v2 i − hv1 ihv2 i.
Wait with analysis of the results until you have done next part.

M.7 Resolved production and pressure-strain

Compute the production term and the pressure-strain terms
∂hv1 i
P k = −hv1′ v2′ i
∂y
∂hv1 i
P11 = −2hv1′ v2′ i
∂y
∂hv 1i
P12 = −hv2′ v2′ i
∂y
 ′

∂v
Φ11 = 2 p′ 1
∂x
 ′
  ′

′ ∂v1 ′ ∂v2
Φ12 = p + p
∂y ∂x
 ′

∂v
Φ22 = 2 p′ 2
∂y
When computing the gradient of, for example, v1 , use the Matlab command
[dudy dudx dudz]=gradient(u3d,y,dx,dz);
Do the production terms and the pressure-strain term appear as you had expected? (see
the previous course TME225).
Now analyze the fluctuations in the previous subsection. Which stresses do you
think are symmetric with respect to the centerline? or anti-symmetric? What’s the
reason?
When averaging, we use only three time steps (three files). If we would use many
more time steps – or, in general, if we let T → ∞ when time averaging, e.g.
Z +T
1
hφi = lim φdt
T →∞ 2T −T

then some of the stresses would be identical zero (now they are small): which ones?
Why?

M.8 Filtering
Plot v1 and v2 along x1 at two different x2 values at x3 = x3,max /2.
1. Filter v1 and v2 to get v̄1 and v̄2 using a 1D box-filter (in the x1 direction) with
filter width ∆ = 2∆x1 (this corresponds to a test filter, see Eq. 18.28). Compare
v̄1 and v̄2 with v1 and v2 .
M.9. SGS stresses: Smagorinsky model 312

I
E
II

εsgs

III

resolved cut-off SGS

scales κ
κc

Figure M.2: Spectrum with cut-off.

2. Do the same thing again but with a filter width of ∆ = 4∆x1 (now you must
derive the expression on your own!). Discuss the differences between no filter,
∆ = 2∆x1 and ∆ = 4∆x1 .

In LES we almost always assume that the filter width is equal to the control volume
(i.e. we use an implicit filter). Above, in Item 1 and 2 you have just carried out explicit
filtering. This type of filtering is used in the scale-similarity model, see Section 18.15.
Repeat Item 1, but now for a 2D filter (x1 and x3 direction); the formula for a 3D
filter is given in Eq. 18.32. Compare the results along the same lines as in Item 1 and
2.

M.9 SGS stresses: Smagorinsky model

Compute the SGS stress τ12 from the Smagorinsky model, which reads
2p
τij = −2νsgs s̄ij , νsgs = (Cs fµ ∆) 2s̄ij s̄ij
 
1 ∂v̄i ∂v̄j
s̄ij = + (M.7)
2 ∂xj ∂xi
fµ = 1 − exp(−x+
2 /26)

The filtered velocities, v̄i , are taken from Section M.8 using the 2D filter (in x1 and
x3 ); we should really have used a 3D filter, but in order to keep it simple, we use the
2D filter. The constant Cs = 0.1.
As an alternative to the damping function, fµ , one can compute the filter length as

∆ = min{κn, ∆} (M.8)

where n is the distance to the nearest wall and κ = 0.4 (von Kàrmàn constant). In this
case you should set fµ = 1.
M.10. SGS stresses: WALE model 313

Compare the SGS stress hτ12 i with the resolved stress hu′ v ′ i and compare the SGS
viscosity with the physical one. Use wall damping both according to Eqs. M.7 and
M.8. Plot them across the channel. Any thoughts?

M.10 SGS stresses: WALE model

Repeat the Task in Section M.9, but now for the WALE model by [189], which reads

∂v̄i 2
gij = , g = gik gkj
∂xj ij
1 2
1
s̄dij = 2
gij + gji − δij gkk 2
(M.9)
2 3

3/2
2 s̄dij s̄dij
νsgs = (Cm ∆)
5/4
5/2
(s̄ij s̄ij ) + s̄dij s̄dij

with Cm = 0.325 which corresponds to Cs = 0.1.

M.11 Dissipations
Compute the dissipation  
∂vi′ ∂vi′
ε=ν
∂xj ∂xj
and plot ε across the channel.
In LES we introduce a filter which is assumed to cut off the spectrum at κc in the
inertial region, see Fig. M.2. At cut-off, kinetic energy is extracted from the resolved
flow by the SGS dissipation εsgs . Since the cut-off is assumed to be located in the iner-
tial sub-range (II), the SGS dissipation is at high Re numbers equal to the dissipation.
Introduce a 2D filter (2∆x1 and 2∆x3 ) as in Sections M.8 and M.9 and filter all
velocities to obtain v̄1 , v̄2 and v̄3 . Compute the SGS stresses from the definition

τij = vi vj − v̄i v̄j (M.10)

and compute the SGS dissipation

 
∂v̄i
εsgs = − τij (M.11)
∂xj

Now, what is the relation between εsgs and ε? Considering the cascade process, what
did you expect?
Recall that when we do traditional Reynolds decomposition, the production term
in the equation for turbulent kinetic energy appears as a sink term in the equation for
the mean kinetic energy, see Eq. 8.35. This is the case also in LES, but now we have
a decomposition into time-averaged filtered velocity, hv̄i i, resolved fluctuation, v̄i′ , and
SGS fluctuation, vi′′ , i.e.

vi = v̄i + vi′′ = hv̄i i + v̄i′ + vi′′ (M.12)

Now we have three equations for kinetic energy: K̄ = 21 hv̄i ihv̄i i, k̄ = 21 hv̄i′ v̄i′ i and
ksgs = 12 hvi′′ vi′′ i. The flow of kinetic energy can be illustrated as in Figs. M.3 and M.4
(cf. Fig. 20 in [89])
 M.11. Dissipations 314

hksgs i

i
hs̄ ij
i
hs̄ ij εs′
i
s gs
2 hν sg
≃
 ∂hv̄i i
− v̄i′ v̄j′
∂xj

ε
K̄ k̄


ν ∂ hv̄ ∂
′
v̄ i
∂ x i i ∂ hv̄
′
 ∂ v̄ i
∂xj
∂x i i
j
j
ν ∂xj

∆T

Figure M.3: Transfer of kinetic turbulent energy. K̄ = 12 hv̄i ihv̄i i and k̄ = 21 hv̄i′ v̄i′ i
denote time-averaged kinetic and resolved turbulent kinetic energy, respectively. ∆T
denotes increase in internal energy, i.e. dissipation. The cascade process assumes that
the term in red is negligible (see also Fig. 8.2). The terms in blue show viscous and the
SGS dissipation of the mean flow which is not included in the cascad process.

The transport equation for h 12 v̄i′ v̄i′ i is derived in [23]. (can be downloaded from
www.tfd.chalmers.se/˜lada).
When deriving the ksgs equation, no decomposition into time-averaged, hv̄i i, and
resolved fluctuations, v̄i′ , is made. Hence the SGS dissipation in Eq. M.11 appears as
an instantaneous production term in the equation for ksgs [90, 91] (can be downloaded
from www.tfd.chalmers.se/˜lada).
Plot (along x2 ), compare and discuss the five dissipations (see Figs. M.3 and M.4)

∂hv̄1 i
hv̄1′ v̄2′ i : dissipation (which is equal to production with minus sign) by resolved
∂x2
turbulence in the K̄ equation
* ′ +  
∂v̄i ∂v̄i ∂v̄i′ ∂v̄i′
ε′sgs = νsgs ≃ νsgs : SGS dissipation term in the k̄ equa-
∂xj ∂xj ∂xj ∂xj
tion. This is the modelled SGS dissipation.
The exact SGS dissipation is computed as (since ε′sgs is a product of two fluctu-
ating quantities, we compute it with the same formula as in Eq. M.6)
   
∂v̄i′ ∂v̄i ∂hv̄i i ∂hv̄i i
ε′sgs = −τij′ = − τij + hτij i = εsgs + hτij i
∂xj ∂xj ∂xj ∂xj
(M.13)
see Eq. M.11. Note that ε′sgs is defined using the fluctuating velocity gradient
(see [23]5 ), contrary to εsgs = Pksgs in Eq. M.11. The difference between εsgs
5 can be downloaded from course home page
M.11. Dissipations 315

∂hv̄i i ∂hv̄i i
ν
∂xj ∂xj

≃
2h
∂hv̄i i
−hv̄i′ v̄j′ i

ν sg
∂xj  

s
ihs̄
∂v̄i′ ∂v̄i′
ν

ij
∂xj ∂xj

ihs̄
ij
E

i
ε′sgs

κ
κc

Figure M.4: Energy spectrum. Transfer of kinetic energy. The cascade process assumes
that the term in red is negligible (see also Fig. 8.2). The terms in blue show viscous
and the SGS dissipation of the mean flow which is not included in the cascad process.

and ε′sgs is that the former includes the SGS dissipation from the mean kinetic
energy; using the Boussinesq assumption, the last term in Eq. M.13 reads

∂hv̄i i
hτij i = h−2νsgs s̄ij ihs̄ij i ≃ −2hνsgs ihs̄ij ihs̄ij i (M.14)
∂xj

see Figs. M.3 and M.4.

 
∂v̄i′ ∂v̄i′
ν : viscous dissipation term in the k̄ equation
∂xj ∂xj
D E  2
∂v̄1 ∂hv̄1 i ∂hv̄1 i
νsgs ∂x 2 ∂x2 ≃ hν sgs i ∂x2 : SGS dissipation term in the K̄ equation. This
corresponds to 2hνsgs s̄ij ihs̄ij i in Figs. M.3 and M.4. This is seen by re-writing
the production term
    
∂v̄i ∂v̄j ∂v̄i
νsgs + = 2hνsgs s̄ij ihs̄ij i (M.15)
∂xj ∂xi ∂xj
M.12. Test filtering 316

see Eq. 11.39. In fully-developed channel flow, the left side reads
 
∂v̄1 ∂hv̄1 i
νsgs (M.16)
∂x2 ∂x2

M.12 Test filtering

Above the filtered velocities were computed using the filter width ∆ = 2∆x1 . In
z{
dynamic models, we often define the test filter as twice the usual filter, i.e. ∆ = 2∆.
z{
Use this definition (1D filter, i.e. ∆ = 4∆x1 ) to compute the dynamic Leonard stress
hL12 i from the definition
z { z{ z{
Lij = v̄i v̄j − v̄ i v̄ j (M.17)
and compare it (across the channel) with the resolved stress hv̄1′ v̄2′ i and the SGS stress
hτ12 i defined in Eq. M.10. Do you expect the magnitude of stresses to be similar?

M.13 Near-wall behavior

What is the near-wall behavior of hv1 i, hv1′2 i and hv2′2 i (i.e., for v1 , what is m in hv1 i =
O(xm2 )). In order to estimate m, plot the quantities in log-log coordinates. Do the
quantities exhibit the near-wall behaviour that you expected?

M.14 Two-point correlations

The two-point correlation for u′
96 X
X 96
1
B11 (x2 , ζm ) = v1′ (xI1 , x2 , xK ′ I K
3 )v1 (x1 , x2 , x3 − ζm ) (M.18)
96 × 96
I=1 K=1

where xK 3 and ζm are the spanwise locations of the two points. Take advantage of the
fact that the flow is periodic, but be careful when integrating the correlation above in
the x3 direction. We have 96 cells in the x3 direction. If, for example, ζm = 2∆x3 ,
and one of the points (x13 ) is at K = 1 then the other (x13 − 2∆) is at K = 95.
Plot the two-point correlation at a couple of x2 positions. When plotting two-point
correlations, it is no point showing both symmetric parts; show only half of it (cf. the
two-point correlations in Section 10.1 and Fig. N.1).
Compute and plot the integral length scale, L1 , which is defined by
Z ∞
1
L1 (x2 ) = 2 B11 (x2 , ζ1 )dζ1 . (M.19)
v1,rms 0

Compute also L3 , which is defined as

Z ∞
1
L3 (x2 ) = 2 B33 (x2 , ζ3 )dζ3 . (M.20)
v3,rms 0

What’s the difference between L1 and L3 ? Did you expect this difference?
M.15. Energy spectra, optional 317

M.15 Energy spectra, optional

The energy spectrum of any second moment can be obtained by taking the FFT of the
corresponding two-point correlation. You can find some details on how to use Matlab’s
FFT in Appendix N.
If you have computed the Fourier coefficients properly, the sum of all coefficients
should give the energy. The reason is that the Fourier coefficients correspond to the
energy spectrum, and if we integrate the energy spectrum over all wave numbers we
get the total energy. When we take the FFT of Eq. M.18, for example, we get

B̂11 (κz ) = F F T (B11 )

and summation gives

N
X
2
v1,rms = B̂11 /N (M.21)
1

see Appendix N.
Plot the energy spectra at a couple of x2 locations. When plotting two energy
spectrum, it is no point showing both symmetric parts; show only half of it (cf. the
energy spectrum in Fig. N.5 b). Confirm that Eq. M.21 is satisfied.

M.16 Something fun

Think of an interesting turbulent quantity and plot it and analyze it!
N. MTF270: Compute energy spectra from LES/DNS data using Matlab 318

N MTF270: Compute energy spectra from LES/DNS

data using Matlab
N.1 Introduction
HEN analyzing DNS or LES data, we are interested to look at the energy spectra.
W From these we can find out in which turbulence scales (i.e. at which wave num-
bers) the fluctuating kinetic turbulent energy reside. By taking the Fourier transform
of the time signal (a fluctuating turbulent velocity) and then taking the square of the
Fourier coefficients we obtain the energy spectrum versus frequency.
If we want to have the energy spectrum versus wavenumber there are two options;
either we
1. Fourier transform N instantaneous signals in space and then time average the N
Fourier transforms, or
2. we Fourier transform a (time-averaged) two-point correlation, B33 (x̂3 ).
Option 1 and 2 is really only a question when we time average. In Option 1, we take
first FFT and then time average. In 2, we first time average and then take FFT.
Here we will use Option 2. The two-point correlation, B33 (x̂3 ) is defined as (see
Eq. 10.2)
B(x3 , x̂3 ) = hv3′ (x3 − x̂3 )v3′ (x3 )i (N.1)
where x̂3 is the separation between the two points. Here we assume that x3 is an
homogeneous direction so that B33 is independent of x3 , i.e. B33 = B33 (x̂3 ). The
two-point correlation for an infinite channel flow is shown in Fig. N.1. On discrete
form the expression for B33 reads
M
1 X ′
B33 (k∆z) = v (x3 − k∆z)v3′ (x3 ) (N.2)
M m=1 3

where m denotes summation in homogeneous directions (i.e. time plus spatial homo-
geneous directions).
In the following section we give a simple example how to use Matlab to Fourier
transform a signal where we know the answer. Then we show how to derive the energy
spectrum from a spatial two-point correlation. Finally, some comments are given on
how to create an energy spectrum versus frequency from an autocorrelation (i.e. from
a two-point correlation in time).

N.2 An example of using FFT

Here we will present a simple example. Consider the function

u = 1 + cos(2πx/L) = 1 + cos(2π(n − 1)/N ) (N.3)

where L = 3 is the length of the domain and N = 16 is the number of discrete points,
see Fig. N.2. Let’s use this function as input vector for the discrete Fourier transform
(DFT) using Matlab. The function u is symmetric, so we expect the Fourier coefficients
N.2. An example of using FFT 319

to be real. In Matlab the DFT of u is defined as (type help fft at the Matlab prompt)
N
X  
−ı2π(k − 1)(n − 1)
U (k) = un exp
n=1
N (N.4)
1≤k≤N
√
where k is the non-dimensional wavenumber and ı = −1. The ratio (n − 1)/N
corresponds to the physical coordinate, x, in the the continuous FFT
Z L
c 1
U (κ) = u(x) exp(−ıκx)dx, κ = 2π/L (N.5)
L −L

Note that the discrete Fourier U (k) coefficients in Eq. N.4 must be divided by N , i.e.
U (k)/N , in order to correspond to the Fourier coefficients U c (N corresponds to L in
Eq. N.5). Furthermore, it can be noted that in Eq. N.4 the period [0, 2π] is used whereas
the formulation in Eq. N.5 is based on the interval [−π, π].
In Matlab, we generate the function u in Eq. N.3 using the commands

N=16;
n=1:1:N;
u=1+cos(2*pi*(n-1)/N);

The u function is shown in Fig. N.2. 16 nodes are used; node 1 is located at x = 0
and node 16 is located at 15L/16.
Now we take the discrete Fourier transform of u. Type

U=fft(u);

Instead of using the built-in fft command in Matlab we can program Eq. N.4

1.4
1.2
1
0.8
B33 (x̂3 )
0.6
0.4
0.2
0
−0.2
0 0.1 0.2 0.3 0.4 0.5 0.6
x̂3
Figure N.1: Two-point correlation, B(x̂3 ) = hv3′ (x3 − x̂3 )v3′ (x3 )i, of DNS data in
channel flow taken from [89].
N.2. An example of using FFT 320

2
2 16

1.5

u 1

0.5

9
0
0 0.5 1 1.5 2 2.5 3
x

Figure N.2: The u function.

directly in Matlab as

U=zeros(1,N);
for k=1:N
for n=1:N
arg1=2*pi*(k-1)*(n-1)/N;
U(k)=U(k)+u(n)*cos(-arg1);
end
end

Note that since u is symmetric, we have only used cos(−x) = cos(x) (the sym-
metric part of exp(−ıx)).
The resulting Fourier coefficients are shown in Fig. N.3. Since the function u in-
cludes only one cosine function and a mean (which is equal to one) only three Fourier
coefficient are non-zero. Two of them, U (2)/N = 0.5, U (16)/N = 0.5, correspond to
the cosine functions (there must be two since U is symmetric). For k = 2 we have
cos(2π(n − 1)/N )
and for k = N
cos((N − 1)2π(n − 1)/N ) = cos(−2π(n − 1)/N ) = cos(2π(n − 1)/N )
which corresponds to cos(2πx/L) in Eq. N.3. It can be noted that the interval [k =
N/2 + 1, N = 9, 16] corresponds to the negative, symmetric part of the wavenumbers
in the physical formulation (cf. Eqs. N.4 and N.5). The first Fourier coefficient corre-
sponds – as always – to the mean of u, i.e. U (1)/N = hui. This is easily verified from
Eq. N.4 by inserting k = 1. The remaining coefficients are zero.
In Fig. N.3, U/N is plotted versus non-dimensional wavenumber, k, and versus
wavenumber κ = 2π(n − 1)/L.
The energy, hu2 i, of the signal in Fig. N.2 can be computed as
Z XN
1 L 2
hu2 i = u (x)dx = u2n /N = 1.5 (N.6)
L 0 n=1
 N.3. Energy spectrum from the two-point correlation 321

1 1

0.8 0.8

0.6 0.6
U/N U/N
0.4 0.4

0.2 0.2

0 0
0 5 10 15 0 5 10 15 20 25 30
k κ
(a) Versus non-dimensional wavenumber, k. (b) Versus wavenumber, κ.

Figure N.3: The U/N Fourier coefficients.

In wavenumber space the energy is – according to Parseval’s formula, see Eq. D.4 –
equal to the integral of the square of the Fourier coefficients, i.e.
Z N
1 ∞
1 X 2
hu2 i = U 2 (κ)dκ = U /N = 1.5 (N.7)
L 0 N n=1 n

see Fig. N.3.

N.3 Energy spectrum from the two-point correlation

Now that we have learnt how to use the FFT command in Matlab, let’s use it on our
two-point correlation in Eq. N.1 and Fig. N.1. Equation N.4 reads
N
X  
−ı2π(k − 1)(n − 1)
B̂33 (k) = B33 (n) exp (N.8)
n=1
N

The simulations in [89] have been carried out with periodic boundary conditions in x3
direction (and x1 ), and hence B33 (x̂3 ) is symmetric, see Fig. N.4. Thus, it is sufficient
to use the cosine part of Eq. N.8, i.e.
N
X  
2π(k − 1)(n − 1)
B̂33 (k) = B33 (n) cos (N.9)
n=1
N

In Fig. N.5a the Fourier coefficients B̂33 κ3 are presented versus wavenumber κ3 =
2π(n − 1)/x3,max , where x3,max ≃ 1.55, see Fig. N.4. Figure N.5b shows the same
energy spectra in log-log scale (only half of the spectrum is included), which is the
common way to present energy spectra. The dashed line shows the −5/3 slope which
indicates that the energy spectra from the DNS follows the Kolmogorov −5/3 decay.
As usual, the Fourier coefficient for the first non-dimensional wavenumber, i.e.
B̂33 (1)/N is equal to the mean of B33 , i.e.
N
1 X 1
hB33 i = B33 (n) ≡ B̂33 (1) (N.10)
N n=1 N
 N.3. Energy spectrum from the two-point correlation 322

1.4 1 96
1.2
1
0.8
B33 (x̂3 )
0.6
0.4
0.2
0
−0.2
0 0.5 1 1.5
x̂3
Figure N.4: Periodic two-point correlation, B33 (x̂3 ) = hv3′ (x3 )v3′ (x3 + x̂3 )i, of DNS
data in channel flow taken from [89].

0.25

−1
0.2 10
B33 (k)/N

0.15 −2
10

0.1
−3
10
0.05

−4
0 10 1 2
0 100 200 300 400 10 10
κ3 κ3
(a) Linear-linear scale (b) Log-log scale

Figure N.5: The energy spectrum of v3′2 versus wavenumber, κ3 . Dashed line in b)
show −5/3 slope. Taken from [89].

compare with Eq. N.9. Note that this is almost the same expression as that for the
integral length scale which reads (see Eq. 10.6)
Z ∞
1 hB33 i
Lint (x3 ) = B33 (x3 , x̂3 )dx̂3 = (N.11)
′2
v3 0 v3′2

Hence the integral length scale is related to the first Fourier mode as

B̂33 (1)
Lint = (N.12)
N v3′2

The two-point correlation for zero separation is equal to v3′2 , i.e. B33 (0) = v3′2 =
1.51. Another way to obtain v3′2 is to integrate the energy spectrum in Fig. N.5, i.e.
Z N
∞
1 X
v3′2 = B̂33 (κ3 )dκ3 = B̂33 (n) = 1.51 (N.13)
0 N n=1
N.4. Energy spectra from the autocorrelation 323

N.4 Energy spectra from the autocorrelation

When computing the energy spectra of the v3′ velocity, say, versus frequency, the time
series of v3′ (t) is commonly Fourier transformed and the energy spectrum is obtained
by plotting the square of the Fourier coefficients versus frequency, f . We can also split
the time signal into a number subsets, Fourier transform each subset and then average.
In Matlab, the command pwelch is a convenient command which does all this.
In the previous section we computed the energy spectrum versus wavenumber by
Fourier transforming the two-point correlation. We can use the same approach in time.
First we create the autocorrelation B33 (τ ) = hv3′ (t)v3′ (t + τ )i (this can be seen as a
two-point correlation in time). Then B33 (τ ) is Fourier transformed to get B̂33 (f ) in
the same way as in Section N.3.
O. MTF270, Assignment 3: Zonal PANS, DES, DDES and SAS 324

O MTF270, Assignment 3: Zonal PANS, DES, DDES

and SAS
N this exercise you will use data from a Zonal PANS for fully developed channel
I flow. The data are taken from [155] 6 . The Re number based on the friction velocity
and the half channel width is Reτ = uτ h/ν = 8000. 28 cells (29 nodes including the
boundary) are located in the URANS region at each wall. The matching line is located
at gridline number 29 at which x+ 2 ≃ 500, x2 /δ = 0.06.
A 64 × 96 × 64 mesh has been used. The cell size in x1 and x3 directions are
∆x1 = 0.05 and ∆x3 = 0.025. Periodic boundary conditions were applied in the x1
and x3 direction (homogeneous directions). All data have been made non-dimensional
by uτ and ρ.
At the course www page you find data files with instantaneous flow fields (sta-
tistically independent). The data files include the instantaneous variables u (v1 ), v
(v2 ), w (v3 ) and kT (made non-dimensional by uτ and ρ). Use Matlab or Octave on
Linux/Ubuntu. Octave is a Matlab clone which can be downloaded for free. Use one
of these programs to analyze the data. You find a Matlab/Octave program at the www
page which reads the data and computes the mean velocity. The data files are Matlab
binary files. Since the data files are rather large, it is recommended that you do all tasks
using only data files ’1’. When everything works, then use also data files ’2’, ’3’ and
’4’, and average by use of all four files.
You will also find a file with time history of u.

O.1 Time history

At the www page you find a file u v time 5nodes zonal pans.dat with the
time history of v̄1 and v̄2 . The file has nine columns of v̄1 and v̄3 at five nodes (and
time): x2 /δ = 0.0028, x2 /δ = 0.0203, x2 /δ = 0.0364, x2 /δ = 0.0645 and x2 /δ =
0.20. Hence, three nodes are located in the URANS region and two nodes in the LES
region. With uτ = 1 and ν = 1/Reτ = 1/8000, this correspond to x+ 2 = 22,
x+ + + +
2 = 162, x2 = 291, x2 = 516 and x2 = 1600, respectively. The sampling time
step is 6.250E −4 (every time step). Use the Matlab program pl time zonal pans
to load and plot the time history of v̄1 .
Recall that the velocities have been scaled with the friction velocity uτ , and thus
what you see is really v̄1 /uτ . The time history of v̄1 at x2 /δ = 0.0203 and x2 /δ =
0.0645 are shown. To study the profiles in closer detail, use the axis-command in the
same way as when you studied the DNS data.
Plot v̄1 for all five nodes. How does the time variation of v̄1 differ for different
positions? Recall that the two points closest the wall are located in the URANS region
and the other two are located in the LES region. In the URANS region the turbulent
viscosity is much larger than in the LES region. How do you expect that the difference
in νt affects the time history of v̄1 . Does the time history of v̄1 behave as you expect?
What about v̄3 ?
Compute the autocorrelation of the five points

imax=500;
two_uu_1_mat=autocorr(u1,imax);
6 this paper can be downloaded from http://www.tfd.chalmers.se/˜lada/allpaper.html
O.2. Mean velocity profile 325

Above we set the maximum separation in time to 500 samples. Compute the inte-
gral timescale

dt=t(1);
int_T_1=trapz(two_uu_1_mat)*dt;

Plot the autocorrelation.

plot(t(1:imax),two_uu_1_mat(1:imax),’linew’,2)
xlabel(’t’)
ylabel(’B_{uu}’)
handle=gca
set(handle,’fontsi’,[20])

How does it compare to the integral timescale? Compute the autocorrelation and
integral timescale also for the other three points. Do you see any difference between
the points in the URANS region and the LES region?

O.2 Mean velocity profile

After having performed a Zonal PANS, we want to look at the time-averaged re-
sults. Use the file pl uvw zonal pans.m to look at the mean velocity profiles.
pl uvw zonal pans.m reads the instantaneous v̄1 field and performs an averaging
in the homogeneous directions x1 and x3 . The time averaged velocity profile is com-
pared with the log profile (markers). There are four files with instantaneous values of
v̄1 . Use more than one file to perform a better averaging.

O.3 Resolved stresses

We want to find out how much of the turbulence that has been resolved and how much
that has been modelled. Compute first vmean (this quantity should be very small, but
if you use only one file this may not be the case due to too few samples). Now compute
hv1′ v2′ i. Here’s an example how to do:

uv=zeros(nj,1);
for k=1:nk
for j=1:nj
for i=1:ni
ufluct=u3d(i,j,k)-umean(j);
vfluct=v3d(i,j,k)-vmean(j);
uv(j)=uv(j)+ufluct*vfluct;
end
end
end
uv=uv/ni/nk;

Plot it in a new figure (a figure is created by the command figure(2)).

Compute also the resolved turbulent kinetic energy


kres = 0.5 hv1′2 i + hv2′2 i + hv3′2 i

and plot it in a new figure.

O.4. Turbulent kinetic energy 326

O.4 Turbulent kinetic energy

Now plot and compare the resolved and modelled turbulent kinetic energies. Note that
the modelled turbulent kinetic energy, kT (te1 zonal pans.mat, te2 zonal pans.mat,
. . . ), can be downloaded from the www page and loaded at the beginning of pl uvw zonal pans.m.
Which is largest? Which is largest in the URANS region and in the LES region, re-
spectively? What about the sum? The magnitude of resolved and modelled turbulent
kinetic energies is discussed on p. 115 in relation to Fig. 6 in [155].

O.5 The modelled turbulent shear stress

We have computed the resolved shear stress. Let’s find the modelled shear stress.
The modelled turbulent kinetic energy, kT (file te1 zonal pans.mat, . . . ), and
the dissipation, ε (file ed1 zonal pans.mat, . . . ) will be used. Recall that ν =
1/8000. Compute the turbulent viscosity as 0.09k 2 /ε (we neglect the damping function
which is active only in the viscous sublayer; if you’re ambitious you may include it;
it’s given in the paper).
Compute the modelled shear stress from the Boussinesq assumption
 
∂v̄1 ∂v̄2
τ12 = −2νT s̄12 = −νT +
∂x2 ∂x1

Plot it and compare with the resolved shear stress (see Section O.3). Are they smooth
across the interface? Is the resolved shear stress large in the URANS region? Should it
be large? Why/why not?

O.6 Location of interface in DES and DDES

As mentioned above, the interface in the present simulations are prescribed along a
fixed grid line (No 29). Let’s compare that with DES and DDES.
In SA-DES, the interface is defined as the location where the wall distance is equal
to CDES ∆ where ∆ = max{∆x , ∆y , ∆z }, see Eq. 20.3. How does this compare with
gridline number 29?
In SST-DES, the location of the interface is computed using k and ω. Compute ω
from ε/(β ∗ k) and compute the location using Eq. 20.7. How does the location compare
with gridline 29 and SA-DES?
In DDES, the boundary layer is shielded with a damping function. In SST-DES,
the shielding function (see Eq. 20.8) may be one of the blending functions, F1 or F2
(see Eq. 20.4). Let’s use F2 as the shielding function as in [68]. Does DDES work, i.e.
does it make the model to be in RANS mode in the entire boundary layer?

O.7 SAS turbulent length scales

Compute the 1D von Kármán length scale defined as
 
 ∂hv̄1 i/∂x2 
LvK,1D = κ  2  (O.1)
∂ hv̄1 i/∂x22 

Note that you should take the derivatives of the averaged v̄1 velocity, i.e. of hv̄1 i. Zoom
up near the wall. How does it behave (i.e. what is n in O(xn2 )? What should n be?
O.7. SAS turbulent length scales 327

Compare that with the von Kármán length scale defined from instantaneous v̄1 , i.e.
 
 ∂v̄1 /∂x2 
LvK,1D,inst = κ  2  (O.2)
∂ v̄1 /∂x22 
How does it compare with LvK,1D ?
When we’re doing real 3D simulations, the first and second derivative must be
defined in 3D. One way of defining the von Kármán length scale in 3D is [149, 150]
 
 S 

LvK,3D,inst = κ  ′′ 
U
0.5
S = (2νt s̄ij s̄ij ) (O.3)
 2 0.5
∂ v̄i ∂ 2 v̄i
U ′′ =
∂xj ∂xj ∂xj ∂xj
The second derivative is then computed as
 2  2  2
∂ 2 v̄1 ∂ 2 v̄1 ∂ 2 v̄1
U ′′2 = + +
∂x21 ∂x22 ∂x23
 2  2  2
∂ 2 v̄2 ∂ 2 v̄2 ∂ 2 v̄2
+ + (O.4)
∂x21 ∂x22 ∂x23
 2  2  2
∂ 2 v̄3 ∂ 2 v̄3 ∂ 2 v̄3
+ +
∂x21 ∂x22 ∂x23
Plot the von Kármán length scale using Eqs. O.3 and O.4. Compare them with Eq. O.1.
What’s the difference? What effect do the different length scales give for PSAS (i.e.
T1 in Eq. 22.5) and what effect does it give to ω?
Another way to compute the second derivative is [190]
 0.5
′′ ∂ 2 v̄i ∂ 2 v̄i
U = (O.5)
∂xj ∂xk ∂xj ∂xk
 2  2 2 
′′2 ∂ 2 v̄1 ∂ 2 v̄1
∂ 2 v̄1
U = +2 +2
∂x21 ∂x1 ∂x2
∂x1 ∂x3
 2 2  2 2  2 2
∂ v̄1 ∂ v̄1 ∂ v̄1
+2 +
∂x22 ∂x2 ∂x3 ∂x23
 2 2  2 2  2 2
∂ v̄2 ∂ v̄2 ∂ v̄2
+2 +2
∂x21 ∂x1 ∂x2 ∂x1 ∂x3
 2 2  2 2  2 2 (O.6)
∂ v̄2 ∂ v̄2 ∂ v̄2
+2 +
∂x22 ∂x2 ∂x3 ∂x23
 2 2  2 2  2 2
∂ v̄3 ∂ v̄3 ∂ v̄3
+ 2 + 2
∂x21 ∂x1 ∂x2 ∂x1 ∂x3
 2 2  2 2  2 2
∂ v̄3 ∂ v̄3 ∂ v̄3
+2 +
∂x22 ∂x2 ∂x3 ∂x23
Plot and compare the von Kármán length scales using the second derivatives defined
in Eqs. O.4 and O.6.
P. MTF270, Assignment 4: Embedded LES with PANS 328

P MTF270, Assignment 4: Embedded LES with PANS

Interface

RANS, fk = 1.0 LES, fk = fkLES < 1

2
x2

x1
0.95 2.25

Figure P.1: Channel flow configuration. The interface separates the RANS and the LES
regions.

N this exercise you will use data from an embedded PANS of channel flow. The
I data are taken from [159]. The ku (Eq. 23.11) and the εu (Eq. 23.18) equations are
solved. The turbulent viscosity is computed from Eq. 23.8. The PANS model is a
modified k − ε model which can operate both in RANS mode and LES mode.
The Reynolds number for the channel flow is Reτ = 950 based on the friction
velocity, uτ , and half the channel width, δ. In the present simulations, we have set
ρ = 1, δ = 1 and uτ ≃ 1, see Fig. P.1. With a 3.2 × 2 × 1.6 domain, a mesh with
64 × 80 × 64 cells is used in, respectively, the streamwise (x), the wall-normal (y)
and the spanwise (z) direction, see Fig. P.1. The resolution is approximately (the wall
shear stress varies slightly along the wall) 48 × (0.6 − 103) × 24 in viscous units. Inlet
conditions at x = 0 are created by computing fully developed channel flow with the
LRN PANS model in RANS mode (i.e. with fk = 1). The RANS part extends up to
x1 = 0.95; downstream the equations operate in LES mode ((i.e. fk = 0.4).
Anisotropic synthetic fluctuations are added at the interface. The interface condi-
tion for εu is computed with the baseline value Cs = 0.07, where kRAN S is taken at
x = 0.5, see Fig. P.1. The modeled dissipation, εinter , is set from kinter and an SGS
length scale, ℓsgs , which is estimated from the Smagorinsky model as

ℓsgs = Cs ∆ (P.1)

and the interface condition for ku is computed as

kinter = fkLES kRAN S (P.2)

with fkLES = 0.4. The interface conditions on ku and εu will make the turbulent
viscosity steeply decrease from its large values in the RANS region to much smaller
values appropriate for the LES region.

P.1 Time history

At the www-page you find a file u time interior.dat with the time history of
v̄2 . The file has eight columns of v̄2 along two lines: x2 = 0.0139 (x+
2 ≃ 13) and
P.2. Resolved stresses 329

x2 = 0.24 (x+ 2 ≃ 230); they are located at x1 = 0.775, 1.175, 1.675, 2.175. The
sampling time step is 0.000625 (every time step). Use Matlab. You can also use Octave
on Linux/Ubuntu. Octave is a Matlab clone which can be downloaded for free. Use the
Matlab/Octave program pl time pans to load and plot the time history of v̄2 .
The time history of v̄2 at x2 = 0.0139 at x1 = 0.775 and x1 = 1.675 are shown.
To study the profiles in closer detail, use the axis-command in the same way as when
you studied the DNS data. Why is there such a big difference in the fluctuations?
If you’re not interested in integral time scales, skip the rest of this section and
proceed to Section P.2.
In Matlab figure 2, the autocorrelation is plotted. The autocorrelation is defined as
Z ∞
B(τ ) = v(t)v(t − τ )dt (P.3)
0

Study the coding and try to understand it. When prescribing the time correlation of
the synthetic fluctuations, the integral timescale T is used. The integral time scale is
defined as Z ∞
T = B norm (τ )dτ (P.4)
0

where B norm = B(τ )/B(0) so that B norm (0) = 1. The constant a is in [159] set to
0.954 and from Eq. 14 in [191] we can then compute the prescribed integral timescale.
In the Matlab file the integral timescale is computed from the autocorrelation. Try to
understand the coding.
Plot v̄2 for the other nodes and study the differences. Compute the autocorrelations
and the integral timescales.

P.2 Resolved stresses

Now we will look at the time-averaged results. Use the file pl uuvvww 2d.m to
look at the mean quantities such as velocity, resolved and modeled stresses, turbulent
viscosities etc. pl uuvvww 2d.m reads the fields and transforms them into 2D arrays
such as u 2d, uu 2d.
Run pl uuvvww 2d. The resolved stresses hv1′2 i are plotted vs x2 (figure 1) and
vs. x1 (figure 2). Two x1 stations are shown in figure 1, x1 = 1.175 and x1 = 2.925.
Plot the resolved stress also in the RANS region, i.e. for x1 < 0.95. The hv1′2 i profiles
are very different in the RANS region (x1 < 0.95) and in the LES (x1 > 0.95), aren’t
they? Why? This can also be seen in figure 2 where hv1′2 i is plotted vs. x1
Now plot the resolved shear stresses, hv1′ v2′ i, both in the RANS region and in the
LES region. You find the same difference between RANS and LES region as for hv1′2 i,
don’t you?

P.3 Turbulent viscosity

Plot the turbulent viscosity vs. x2 in both regions. Normalize it with hνi, i.e. plot
hνu i/ν. Where is it large and where is it small? Why? Now plot it also vs. x1 .
Something drastically happens at x1 = 0.95, right?
P.4. Modeled stresses 330

E(κ)
Pk in
e rti
al
ra
ng
e
εs
gs
= ε
P
−5/3 u
E(κ) ∝ κ

dissipating range

κ
κc
Figure P.2: Energy spectrum.

P.4 Modeled stresses

In Section P.2 you looked at the resolved Reynolds stresses. Now let’s look at the mod-
eled stresses. Compute the modeled Reynolds stresses from the Boussinesq assumption
 
∂hv̄i i ∂hv̄j i 2
hvi′ vj′ mod i = −hνu i + + δij hku i (P.5)
∂xj ∂xi 3

Compare the resolved and the modeled shear stress as well as the streamwise and the
normal stresses in the RANS region and in the LES region.

P.5 Turbulent SGS dissipation

In an LES the resolved turbulent fluctuations can be represented by a energy spectrum
as in Fig. P.2. The resolved turbulence extracts kinetic energy via the production term,
P k , which represents a source term in the k equation (Eq. 8.14) and a sink term in the
K̄ equation (Eq. 8.35). The energy flow is visualized in Fig. M.3 where the energy in
K̄ mostly goes to resolved turbulence, k̄, then to modeled turbulence, ksgs (or ku ) and
finally to internal energy via dissipation, εu .
In RANS mode, however, there is no resolved turbulence. Hence the kinetic energy
goes directly from K̄ to the modeled turbulence, ku .
In the LES region, the production term in the ku equation includes both mean and
fluctuating strain rates since
   
∂v̄i ∂v̄j ∂v̄i
Pu = εsgs = νu +
∂xj ∂xi ∂xj

which in the Matlab file is stored as pksgs 2d. Now consider the LES region. Investi-
gate the relation between Pu = εsgs and the production due to the resolved turbulence

∂hv̄i i
P k = −hvi′ vj′ i
∂xj

Compare also P k in the LES region and in the RANS region.

P.5. Turbulent SGS dissipation 331

In both the RANS and the LES region the process of viscous dissipation takes place
via εu . Hence, plot also this quantity. Is the turbulence in local equilibrium, i.e. does
the relation Pu = εu hold?
Q. MTF270, Assignment 5, recirculating flow: PANS, DES and DDES 332

Q MTF270, Assignment 5, recirculating flow: PANS,

DES and DDES
N this exercise you will use data from PANS and Zonal PANS for for the flow over a
I hump. The results are presented in detail in [155, 159, 191] 7

0.8

0.6
x2

0.4

0.2

0.6 1 2
x1

Figure Q.1: Hump flow. Grid. Every 4th grid line is shown.

The Reynolds number of the hump flow is Rec = 936 000, based on the hump
length, c, and the inlet mean velocity at the centerline, Uin,c . The grid is shown in
Fig. Q.1. Experiments were conducted by Greenblatt [192,193]; they are also described
in detail at http://cfdval2004.larc.nasa.gov/case3.html. The maximum height of the
hump, h, and the channel height, H, are given by H/c = 0.91 and h/c = 0.128,
respectively. The baseline mesh has 312 × 120 × 64 cells with Zmax = 0.2.
There are side-wall effects (3D flow) near the side plates in the experiment. Hence,
in order to compensate for the blockage effect of the side plates in the computation,
the surface shape of the upper wall (above the hump) is modified and the upper wall is
moved slightly downwards, see Fig. Q.1. The ratio of the local cross-sectional area of
the side plates (facing the flow) to the cross sectional area of the tunnel enclosed by the
side plates was computed. This ratio was used to scale the local height of the channel
and thus modifying the contour shape of the upper wall.
Neumann conditions are used at the outflow section located at x1 = 4.2. Slip
conditions are used at the upper wall and symmetric boundary conditions are used on
the spanwise boundaries. Inflow boundary (at x1 = 0.6) conditions are taken from
2D RANS SST k − ω simulations carried out by the group of Prof. Strelets in St.
Petersburg. The distributions of V1 and V2 at x1 = 0.6 from the RANS simulation are
used together with V3 = 0 as mean inlet velocities to which the fluctuating velocity V1′ ,
V2′ and V2′ are superimposed. The computed integral length scale for the synthetic inlet
fluctuations is L ≃ 0.040 and the integral time scale T ≃ 0.038. It is noted that the
prescribed inlet integral length scale is rather large [159] (approximately equal to the
7 these papers can be downloaded from http://www.tfd.chalmers.se/˜lada/allpaper.html
Q.1. Discretization schemes 333

inflow boundary layer thickness). The reason is that synthetic fluctuations with a large
integral length scale are efficient in generating resolved turbulent fluctuations [172].
At the course www page you find data files with time and spanwise averaged flow
fields. Use Matlab or Octave on Linux/Ubuntu. Octave is a Matlab clone which can be
downloaded for free. Use one of these programs to analyze the data.
Start by downloading pl vect hump.m, vectz aiaa paper.dat, xy hump.dat
and x065 off.dat. The m-file reads the data files. The velocities, the square of the
velocities and a number of other quantities are read. The resolved stresses are com-
puted as in Eq. M.6. Furthermore, the m file plots contours of hv1′ v1′ i, the grid, a vector
plot as well as hv̄1 i and hv2′ v2′ i at x = 0.65 (gridline 10 is chosen which corresponds
to x 2d(10,1)= 0.652). The experimental values are also included in the plots at
x = 0.65.

Q.1 Discretization schemes

You will analyze results from two publications [159, 191]. The main difference is that
in [191] pure central differencing was used whereas in [159] a blend of 95% central
differencing and 5% bounded second-order upwind scheme (van Leer scheme [194])
was employed.
Consider the plot of hv̄1 i and hv2′ v2′ i at x = 0.65. The velocity profile shows that
the boundary layer extends up to x2 ≃ 0.17. However, the hv2′ v2′ i exhibits large values
outside the boundary layer; the peak value at x2 ≃ 0.24 is almost 5 times larger than
in the boundary layer. These fluctuations are nonphysical and stem from the use of
central differencing.
Now plot hv2′ v2′ i at locations further downstream; do the nonphysical fluctuations
disappear (they do, don’t they?). Compare with hv1′ v2′ i in Fig. 12 in [191]. The non-
physical fluctuations appear in the outer region where no physical resolved turbulence
is present. Pure central differences sometimes give problems is flow regions like this.
It is tempting to draw the conclusion that these nonphysical fluctuations are caused by
the inlet synthetic fluctuations. However, when increasing the magnitude of the inlet
synthetic fluctuations the magnitude of the nonphysical fluctuation diminishes, see Fig.
13a in [191].
Similar (much worse) problems with central differences were encountered in LES
of the flow around an airfoil [104, 195]. Large nonphysical fluctuations were found in
regions far from the airfoil; in these studies the problems were solved by using 100%
van Leer scheme in the far-field regions.
In [159] the same simulations were carried out but now using a blend of central
differencing (95%) and van Leer scheme (5%). Analyze these results by loading file
vectz aiaa journal.dat in pl vect hump.m.
Consider the resolved fluctuations: note how strongly they increase when going
from the attached boundary layer (x1 = 0.65) to the recirculating region, 0.8 . x .
1.2. The peak of hv2′ v2′ i, for example, increases from 0.0017 (x1 = 0.65, gridline
i=10) to 0.046 (x1 = 0.800, gridline i=45). This is the main reason why the non-
physical fluctuations at x1 = 0.65 vanish quickly further downstream.

Q.2 The modeled turbulent shear stress

As usual, when analysing results of DES, PANS etc, we want to find out how much of
the turbulence is modeled and how much is resolved. This gives an indication how well
the turbulence has been resolved; if much turbulence is resolved it may be either good
Q.3. The turbulent viscosity 334

or bad. In recirculating regions it is usually good but it may be disastrous in boundary

layer regions close to the wall if the mesh is not fine enough (this is the very reason
why DDES was proposed, see Eq. 20.8). A suitable quantity to look at is the shear
stress which is much more relevant than the turbulent kinetic energy; the shear stress
is the largest diffusion term in the momentum equations, much larger than the normal
stresses (at least in boundary layer flows).
Now plot the resolved (uv 2d) and the modelled shear stress (uv model 2d)
in the boundary layer region (e.g. gridline i=10) and the recirculating region (e.g.
gridline i=45). Where is the modeled shear stress large and where is it small? Look
at other x1 locations.

Q.3 The turbulent viscosity

It may be interesting to look at the turbulent viscosity. Usually it is plotted scaled with
ν, i.e. νt /ν. Plot the viscosity, νt /ν (vis 2d/viscos) in the boundary layer and in
the recirculating region. You find that the turbulent viscosity is large in the recirculation
region, right? That might lead you to conclude that the turbulence is better resolved
in the boundary layer than in the recirculating region. But that would be an incorrect
conclusion. When estimating the influence of the modeled turbulence, we should not
look at the turbulent viscosity because it does not appear as a term in the momentum
equations; we should look at the ratio of modeled stresses to the resolved one, and,
as mentioned in Section Q.2, preferably the ratio of the modeled to the resolved shear
stress [117, 118]. It would actually be even more relevant to look at the divergence of
these terms, i.e.  
∂ ∂v̄i ∂hvi′ vj′ i
νt and − (Q.1)
∂xj ∂xj ∂xj
because these are the terms that indeed appear in the momentum equations.
You have noted that the turbulent viscosities are rather large; in the recirculation re-
gion, νt /ν is larger than 100. Go back and load the results of [191] (vectz aiaa paper.dat).
Plot the turbulent viscosity. You find that it is much smaller than for the data of [159].
It is true that the discretization schemes used in the two papers are slightly different,
but that is not the reason. The difference stems from the fact that the turbulent Prandtl
numbers in the k and ε equations were in the two papers set as
σk = 1.4, σε = 1.4 [191] (Q.2)
1.4 1.4
σk = , σε = [159] (Q.3)
fk2 fk2
Equation Q.3 corresponds to the original PANS model. Recall that the turbulent diffu-
sion in, for example, the k equation reads
 
∂ νt ∂k
(Q.4)
∂xj σk ∂xj
Since fk = 0.4, it means that the turbulent diffusion in the k and ε equations are
1/0.42 ≃ 6 times larger in [159] than in [191]. The consequence is that peaks in k and
ε (and also νt ) are reduced in the former case compared to the latter (this is the physical
role played by diffusion: it transports from regions of high k to regions of low k). This
explains why the peaks of k are much larger in [191] compared to [159].
Hence, in the original PANS model (Eq Q.3), the RANS turbulent viscosity appears
in the turbulent diffusion of k (and ε), because the turbulent diffusion term reads (recall
Q.4. Location of interface in DES and DDES 335

that fk = k/ktotal = k/kRAN S where kRAN S denotes the turbulent kinetic energy in
a RANS simulation)
   
∂ νt fk2 ∂k ∂ cµ k 2 ∂k
=
∂xj 1.4 ∂xj ∂xj εfk2 1.4 ∂xj
 2
   (Q.5)
∂ cµ kRAN S ∂k ∂ νt,RAN S ∂k
= =
∂xj ε1.4 ∂xj ∂xj 1.4 ∂xj

cf. Eqs. 18 and 19 in [134]. Thus the total (i.e. RANS) viscosity is responsible for the
transport of the modeled turbulent kinetic energy.

Q.4 Location of interface in DES and DDES

The results analyzed above were from LES simulations [159,191] (i.e. the PANS model
was used in LES mode). Now we will analyze results from Zonal PANS [155] where
the interface is prescribed along a fixed grid line (No 32). Let’s compare that with DES
and DDES. Load the file vectz zonal pans.dat in pl vect hump.m. Recall
that ∆z = 0.2/32 (note that in [159, 191] ∆z = 0.2/64)
In SA-DES, the interface is defined as the location where the wall distance is equal
to CDES ∆ where ∆ = max{∆x , ∆y , ∆z }, see Eq. 20.3. How does this compare with
gridline number 32?
In SST-DES, the location of the interface is computed using k and ω. Compute ω
from ε/(β ∗ k) and compute the location using Eq. 20.7. How does the location compare
with gridline 32 and SA-DES?
In DDES, the boundary layer is shielded with a damping function. In SST-DES,
the shielding function (see Eq. 20.8) may be one of the blending functions, F1 or F2
(see Eq. 20.4). Let’s use F2 as the shielding function as in [68]. Does DDES work, i.e.
does it make the model to be in RANS mode in the entire boundary layer? What about
the separation region?

Q.5 Compute fk
In [159, 191], fk = 0.4 in the entire domain. In the Zonal PANS simulations [155],
fk = 0.4 in the LES region or it is computed from Eq. Q.6; in the RANS region near
the lower wall fk = 1 (interface at grid line No 32). The expression for computing fk
reads [196]
3/2
∆ k
fk = cµ−2/3 , Lt = total (Q.6)
Lt hεi
where ∆ = (∆V )1/3 where ∆V denotes the cell volume and ktotal denotes resolved
plus modeled turbulent kinetic energy, i.e.
1

ktotal = hv1′2 i + hv2′2 i + hv3′2 i + hki (Q.7)
2
Compute fk from Eq. Q.6 and plot it at a couple of x1 locations (cf. Fig. 26
in [155]). It is bigger or smaller than the prescribed values (0.4 and 1)? Compare
it also with the definition of fk = k/ktotal (cf. Fig. 26 in [155]). How is k/ktotal
related to Eq. Q.6? (see the discussion in [155] leading to Eq. 14)
R. MTF270, Assignment 6: Hybrid LES-RANS 336

R MTF270, Assignment 6: Hybrid LES-RANS

N this exercise you will use data from a Hybrid LES-RANS for fully developed
I channel flow. The turbulence model is the same as in [89] (no forcing), but the
domain and Reynolds number is taken from [197]. The Re number based on the friction
velocity and the half channel width is Reτ = uτ h/ν = 8000. 28 cells (29 nodes
including the boundary) are located in the URANS region at each wall. The matching
line is located at gridline number 29 at which x+ 2 ≃ 500, x2 /δ = 0.06.
A 64 × 96 × 64 mesh has been used. The cell size in x1 and x3 directions are
∆x1 = 0.05 and ∆x3 = 0.025. Periodic boundary conditions were applied in the x1
and x3 direction (homogeneous directions). All data have been made non-dimensional
by uτ and ρ.
At the course www page you find data files with instantaneous flow fields (sta-
tistically independent). The data files include the instantaneous variables u (v1 ), v
(v2 ), w (v3 ) and kT (made non-dimensional by uτ and ρ). Use Matlab or Octave on
Linux/Ubuntu. Octave is a Matlab clone which can be downloaded for free. Use one
of these programs to analyze the data. You find a Matlab/Octave program at the www
page which reads the data and computes the mean velocity. The data files are Matlab
binary files. Since the data files are rather large, it is recommended that you do all tasks
using only data files ’1’. When everything works, then use also data files ’2’, ’3’ and
’4’, and average by use of all four files.
You will also find a file with time history of u.

R.1 Time history

At the www page you find a file u v time 4nodes hybrid.dat with the time
history of v̄1 and v̄3 . The file has nine columns of v̄1 and v̄3 at four nodes (and time):
x2 /δ = 0.0028, x2 /δ = 0.015, x2 /δ = 0.099 and x2 /δ = 0.35. Hence, two nodes
are located in the URANS region and two nodes in the LES region. With uτ = 1
and ν = 1/Reτ = 1/8000, this correspond to x+ + +
2 = 22, x2 = 120, x2 = 792 and
+
x2 = 2800, respectively. The sampling time step is 6.250E − 4 (every time step). Use
the Matlab program pl time hybrid to load and plot the time history of v̄1 .
Recall that the velocities have been scaled with the friction velocity uτ , and thus
what you see is really v̄1 /uτ . The time history of v̄1 at x2 /δ = 0.015 and x2 /δ = 0.35
are shown. To study the profiles in closer detail, use the axis-command in the same
way as when you studied the DNS data.
Plot v̄1 for all four nodes. How does the time variation of v̄1 differ for different
positions? Recall that the two points closest the wall are located in the URANS region
and the other two are located in the LES region. In the URANS region the turbulent
viscosity is much larger than in the LES region. How do you expect that the difference
in νt affects the time history of v̄1 . Does the time history of v̄1 behave as you expect?
What about v̄2 ?
Compute the autocorrelation of the four points

imax=500;
two_uu_1_mat=autocorr(u1,imax);

Above we set the maximum separation in time to 500 samples. Compute the inte-
gral timescale
R.2. Mean velocity profile 337

dt=t(1);
int_T_1=trapz(two_uu_1_mat)*dt;

Plot the autocorrelation.

plot(t(1:imax),two_uu_1_mat(1:imax),’linew’,2)
xlabel(’t’)
ylabel(’B_{uu}’)
handle=gca
set(handle,’fontsi’,[20])

How does it compare to the integral timescale? Compute the autocorrelation and
integral timescale also for the other three points. Do you see any difference between
the points in the URANS region and the LES region?

R.2 Mean velocity profile

After having performed a hybrid LES-RANS, we want to look at the time-averaged re-
sults. Use the file pl uvw hybrid.m to look at the mean velocity profiles. pl uvw hybrid.m
reads the instantaneous v̄1 field and performs an averaging in the homogeneous direc-
tions x1 and x3 . The time averaged velocity profile is compared with the log profile
(markers). There are four files with instantaneous values of v̄1 . Use more than one file
to perform a better averaging.

R.3 Resolved stresses

We want to find out how much of the turbulence that has been resolved and how much
that has been modelled. Compute first vmean (this quantity should be very small, but
if you use only one file this may not be the case due to too few samples). Now compute
hv1′ v2′ i. Here’s an example how to do:

uv=zeros(nj,1);
for k=1:nk
for j=1:nj
for i=1:ni
ufluct=u3d(i,j,k)-umean(j);
vfluct=v3d(i,j,k)-vmean(j);
uv(j)=uv(j)+ufluct*vfluct;
end
end
end
uv=uv/ni/nk;

Plot it in a new figure (a figure is created by the command figure(2)).

Compute also the resolved turbulent kinetic energy


kres = 0.5 hv1′2 i + hv2′2 i + hv3′2 i

and plot it in a new figure.

R.4. Turbulent kinetic energy 338

URANS region LES region

1/2
ℓ 2.5n[1 − exp(−0.2kT n/ν)] ℓ = ∆ = (δV )1/3
1/2 1/2 1/2
νT 0.09 · 2.5kT n[1 − exp(−0.014kT n/ν)] 0.07kT ℓ

Table R.1: Expressions for ℓ and νT in the LES and URANS regions. n denotes the
distance from the wall.

R.4 Turbulent kinetic energy

Now plot and compare the resolved and modelled turbulent kinetic energies. Note that
the modelled turbulent kinetic energy, kT (te1 hybrid.mat, te2 hybrid.mat,
. . . ), can be downloaded from the www page and loaded at the beginning of pl uvw hybrid.m.
Which is largest? Which is largest in the URANS region and in the LES region, re-
spectively? What about the sum? The magnitude of resolved and modelled turbulent
kinetic energies is discussed in the last subsection in [89].

R.5 The modelled turbulent shear stress

We have computed the resolved shear stress. Let’s find the modelled shear stress.
The modelled turbulent kinetic energy, kT (file te1 hybrid.mat, . . . ), will be
used. Recall that ν = 1/8000. Compute the turbulent viscosity according to Table R.1
and do the usual averaging. When computing ∆, you need the volume, δV , of the cells.
It is computed as δV = (∆x1 ∆x2 ∆x3 ); ∆x1 and ∆x3 are constant and ∆x2 is stored
in the array dy(j), look at the beginning of the m-file. Plot hνT i/ν. Where is it large
and where is it small? (Recall that the URANS region is located in the first 28 cells).
Is it smooth? Do you need more samples? If so, use more files.
Compute the modelled shear stress from the Boussinesq assumption
 
∂v̄1 ∂v̄2
τ12 = −2νT s̄12 = −νT +
∂x2 ∂x1
Plot it and compare with the resolved shear stress (see Section R.3). Are they smooth
across the interface? Is the resolved shear stress large in the URANS region? Should it
be large? Why/why not?

R.6 Location of interface in DES and DDES

As mentioned above, the interface in the present simulations is prescribed along a fixed
grid line (No 29). Let’s compare that with DES and DDES.
In SA-DES, the interface is defined as the location where the wall distance is equal
to CDES ∆ where ∆ = max ∆x , ∆y , ∆z , see Eq. 20.3. How does this compare with
gridline number 29?
In SST-DES, the location of the interface is computed using k and ω. Compute ω
from νT and k as ω = νT /k) and compute the location using Eq. 20.7.

R.7 Turbulent length scales

Compute and plot the turbulent length scales given in Table R.1. Plot the ℓSGS and
ℓURAN S length scales in both regions. Which is largest? Any surprises? Compare
them with ∆x2 and (∆x1 ∆x2 ∆x3 )1/3 . One would expect that (∆x1 ∆x2 ∆x3 )1/3 <
ℓURAN S everywhere. Is this the case?
S. MTF270: Transformation of tensors 339

x2

x∗2
x∗1

β α

x1

Figure S.1: Transformation between the coordinate systems (x1∗ , x2∗ ) and (x1 , x2 ).

S MTF270: Transformation of tensors

S.1 Rotation from x1∗ − x2∗ to x1 − x2
HE rotation of a vector from the xi∗ coordinate system to xi reads (see, e.g., Chapter
T 1 in [28])
ui = bij uj∗ (S.1)
where bij denotes the cosine between the axis
bij = cos (xi , xj∗ ) (S.2)
In Fig. S.1, the bij is given by
b11 = cos α, b12 = cos β = − cos(π/2 − α) = − sin(α)
(S.3)
b21 = cos(π/2 − α) = sin α, b22 = cos α
The relations bik bjk = bki bkj = δij are fulfilled as they should.
For a second-order tensor, the transformation reads
uij = bik bjm uk∗m∗ (S.4)
As an example, set α = π/4. Equation S.3 gives
√ √ √ √
b11 = 1/ 2, b12 = −1/ 2, b21 = 1/ 2, b22 = 1/ 2 (S.5)
Inserting Eq. S.5 into Eq. S.4 gives
u11 = b11 b11 u1∗1∗ + b12 b11 u2∗1∗ + b11 b12 u1∗2∗ + b12 b12 u2∗2∗ (S.6a)
1
= (u1∗1∗ − u2∗1∗ − u1∗2∗ + u2∗2∗ )
2
u12 = b11 b21 u1∗1∗ + b12 b21 u2∗1∗ + b11 b22 u1∗2∗ + b12 b22 u2∗2∗ (S.6b)
1
= (u1∗1∗ − u2∗1∗ + u1∗2∗ − u2∗2∗ )
2
u21 = b21 b11 u1∗1∗ + b22 b11 u2∗1∗ + b21 b12 u1∗2∗ + b22 b12 u2∗2∗ (S.6c)
1
= (u1∗1∗ + u2∗1∗ − u1∗2∗ − u2∗2∗ )
2
u22 = b21 b21 u1∗1∗ + b22 b21 u2∗1∗ + b21 b22 u1∗2∗ + b22 b22 u2∗2∗ (S.6d)
1
= (u1∗1∗ + u2∗1∗ + u1∗2∗ + u2∗2∗ )
2
S.2. Rotation from x1 − x2 to x1∗ − x2∗ 340

S.2 Rotation from x1 − x2 to x1∗ − x2∗

Now we do the same transformation as in Section S.1 but backwards, i.e. from x1 − x2
to x1∗ − x2∗ . Consider fully developed flow in a channel, see Appendix B, for which

∂v1
= v2 = 0 (S.7)
∂x1
so that the strain-rate tensor, sij , reads

1 ∂v1
s11 = 0, s12 = , s21 = s12 , s22 = 0 (S.8)
2 ∂x2
The x1 and x2 coordinates in Fig. S.1 correspond now to the streamwise and wall-
normal directions in the channel, respectively. Let the x1∗ − x2∗ coordinate system
denote the eigenvectors. The transformation from x1 − x2 to x1∗ − x2∗ reads

si∗j∗ = cik cjm skm , cij = cos (xi∗ , xj ) (S.9)

where
c11 = cos α, c12 = cos(π/2 − α) = sin α
(S.10)
c21 = cos β = − sin α, c22 = cos α

see Fig. S.1. It can be seen that the relation cji = bij is satisfied as it should. The
eigenvectors for Eq. S.8 are any two orthogonal vectors with angles ±π/4, ±3π/4.
Let us choose π/4 and 3π/4 for which the transformation in Eq. S.9 reads (α = π/4)

s1∗1∗ = c11 c11 s11 + c12 c11 s21 + c11 c12 s12 + c12 c12 s22 (S.11a)
1
= (s11 + s21 + s12 + s22 )
2
s1∗2∗ = c11 c21 s11 + c12 c21 s21 + c11 c22 s12 + c12 c22 s22 (S.11b)
1
= (−s11 − s21 + s12 + s22 )
2
s2∗1∗ = c21 c11 s11 + c22 c11 s21 + c21 c12 s12 + c22 c12 s22 (S.11c)
1
= (−s11 + s21 − s12 + s22 )
2
s2∗2∗ = c21 c21 s11 + c22 c21 s21 + c21 c22 s12 + c22 c22 s22 (S.11d)
1
= (s11 − s21 − s12 + s22 )
2
The fully developed channel flow is obtained by inserting Eq. S.8

s1∗1∗ = s12 , s1∗2∗ = 0, s2∗1∗ = 0, s2∗2∗ = −s21 (S.12)

Since the off-diagonal elements are zero it confirms that the coordinate system x1∗ −
x2∗ with α = π/4 is indeed a principal coordinate system. The eigenvalues, λ(k) , of
sij correspond to the diagonal elements in Eq. S.12, i.e.

1 ∂v1 1 ∂v1
λ(1) ≡ s1∗1∗ = s12 = , λ(2) ≡ s2∗2∗ = −s12 = − (S.13)
2 ∂x2 2 ∂x2
S.3. Transformation of a velocity gradient 341

S.3 Transformation of a velocity gradient

Consider the velocity gradient Aij = ∂vi /∂xj . Apply the transformation from the
x1 − x2 system to the principal coordinate system x1∗ − x2∗ in Eqs. S.11a-S.11d with
α = π/4
1
A1∗1∗ = (A11 + A21 + A12 + A22 )
2
1
A1∗2∗ = (−A11 − A21 + A12 + A22 )
2 (S.14)
1
A2∗1∗ = (−A11 + A21 − A12 + A22 )
2
1
A2∗2∗ = (A11 − A21 − A12 + A22 )
2
Insert Eq. S.10 with α = π/4 and replace Aij by the velocity gradient use Eq. S.7 (i.e.
fully-developed channel flow)

∂v1∗ ∂v1∗ 1 ∂v1 ∂v2∗ ∂v2∗ 1 ∂v1

= = , = =− , (S.15)
∂x1∗ ∂x2∗ 2 ∂x2 ∂x1∗ ∂x2∗ 2 ∂x2
It can be seen that ∂v1∗ /∂x1∗ = s1∗1∗ and ∂v1∗ /∂x2∗ + ∂v2∗ /∂x1∗ = 2s1∗2∗ = 0
(see Eqs. S.13 and S.15) as it should.
T. MTF270: Green’s formulas 342

T MTF270: Green’s formulas

N this appendix we will derive Green’s three formulas from Gauss divergence law.
I In the last section we will derive the analytical solution to the Poisson equation. The
derivations below are partly taken from [198].

T.1 Green’s first formula

Gauss divergence law reads
Z Z
∂Fi
dV = Fi ni dS (T.1)
V ∂xi S

where S is the bounding surface of the volume, V , and ni is the normal vector of S
∂ψ
pointing out of V . Replacing Fi by ϕ gives
∂xi
Z   Z
∂ ∂ψ ∂ψ
ϕ dV = ϕ ni dS (T.2)
V ∂x i ∂xi S ∂xi

The left side is re-written as

 
∂ ∂ψ ∂2ψ ∂ψ ∂ϕ
ϕ =ϕ + (T.3)
∂xi ∂xi ∂xi ∂xi ∂xi ∂xi
which inserted in Eq. T.2 gives
Z Z Z
∂2ψ ∂ψ ∂ϕ ∂ψ
ϕ dV + dV = ϕ ni dS (T.4)
V ∂xi ∂xi V ∂xi ∂xi S ∂xi

This is Green’s first formula.

T.2 Green’s second formula

Switching ϕ and ψ in Eq. T.4 gives
Z Z Z
∂2ϕ ∂ϕ ∂ψ ∂ϕ
ψ dV + dV = ψ ni dS (T.5)
V ∂xi ∂xi V ∂xi ∂xi S ∂xi
Subtract Eq. T.5 from T.4 gives
Z   Z  
∂2ψ ∂2ϕ ∂ψ ∂ϕ
ϕ −ψ dV = ϕ −ψ ni dS (T.6)
V ∂xi ∂xi ∂xi ∂xi S ∂xi ∂xi
This is Green’s second formula.

T.3 Green’s third formula

In Green’s second formula, Eq. T.6, set
1
ψ(r) = (T.7)
|r − rP |
As usual we are considering a volume V with bounding surface S and normal vector
ni . Since function ψ(r) is singular for r = rP , consider a small sphere in V , see
T.3. Green’s third formula 343

ni
S

V
r nεi
Sε

rP
x2

x1

Figure T.1: Green’s third formula. A volume V with bounding surface S with normal
vector ni . In V there is a small sphere S ε located at rP with radius ε and normal vector
nεi .

Fig. T.1. In Eq. T.6 we need the first and the second derivative of ψ. The first derivative
of 1/ri is computed as
 
∂ 1 ∂r/∂xi ri
=− =− 3 (T.8)
∂xi r r2 r
since the derivative of a distance X is a vector along the increment of the distance, i.e.
∂X/∂xi = Xi /X where X = |Xi |. The second derivative is obtained as
     
∂2 1 ∂  ri  ∂ri 1 ∂r 3ri
=− = − +
∂xi ∂xi r ∂xi r3 ∂xi r3 ∂xi r4
      (T.9)
1 ri 3ri 3 r2 3
= −3 3 + = − + = 0
r r r4 r3 r r4

To get the right side on the second line we used the fact that ri ri = r2 . Now we replace
ri = r by r − rP = ri − rP,i in Eqs. T.8 and T.9 which gives
 
∂ 1 ri − rP,i
=−
∂xi |r − rP | |r − rP |3
2
  (T.10)
∂ 1
=0
∂xi ∂xi |r − rP |

for ri 6= riP , i.e. for V excluding the sphere S ε , see Fig. T.1. Apply Green’s second
formula for this volume which has the bounding surfaces S and S ε with normal vectors
ni (outwards) and nεi (inwards), respectively. We get
Z Z  
1 ∂2ϕ ri − riP 1 ∂ϕ
− dV = −ϕ − ni dS
V −S ε |r − rP | ∂xi ∂xi S |r − rP |3 |r − rP | ∂xi
Z  
ri − riP 1 ∂ϕ (T.11)
+ −ϕ 3
− (−nεi )dS
S ε |r − rP | |r − rP | ∂xi
IS∗
T.3. Green’s third formula 344

where the volume integral is taken over the volume V but excluding the sphere S ε , i.e.
V − S ε . Note the minus sign in front of the normal vector in the S ε integral; this is
because the normal vector must point out of the volume V − S ε , i.e. into the sphere,
S ε . In the sphere the normal vector, nεi , correspond to the direction from point rP , i.e.
r − rP ri − rP,i
nεi = = (T.12)
|r − rP | |ri − rP,i |
where we have normalized the vector ri − rP,i in order to make its length equal to one.
The length of the vector ri − rP,i is the radius of sphere S ε , i.e.
|r − rP | = ε (T.13)
The surface area, dS, for sphere S ε can be expressed in spherical coordinates as
dS = ε2 Ω = ε2 sin θdθdα (T.14)
where Ω is the solid angle. Inserting Eqs. T.12, T.13 and T.14 in the last integral in
Eq. T.11 gives
Z   Z  
ϕ ri − rP,i ∂ϕ 2 ∂ϕ
IS ε = + ε dΩ = ϕ + (ri − rP,i ) dΩ (T.15)
Sε ε2 ε2 ∂xi Sε ∂xi
To re-write this integral we will use the mean value theorem for integrals. In one
dimension this theorem simply states that for the integral of a function, g(x), over an
interval [a, b], there exists (at least) on point for which the the relation
Z b
g(x)dx = (a − b)g(xQ ) (T.16)
a

holds, where xQ denotes a point on [a, b]. Applying this theorem to the integral in
Eq. T.15 gives
Z  Z
∂ϕ
IS ε = ϕ(rQ ) dΩ + rQ,i − rP,i ) (rQ ) dΩ (T.17)
Sε ∂xi Sε

where rQ ≡ rQ,i denotes a point on S ε . As we let Q → P , the radius, ε, of sphere S ε

goes to zero so that the integral in Eq. T.17 reads
lim IS ε = 4πϕ(rQ ) (T.18)
ε→0
R
since Sε
dΩ = 4π. Inserted in Eq. T.11 gives
Z
1 1 ∂2ϕ
ϕ(rP ) = − dV
4π
V |r − rP | ∂xi ∂xi
Z Z (T.19)
1 ri − riP 1 1 ∂ϕ
+ ϕ n i dS + ni dS
4π S |r − rP |3 4π S |r − rP | ∂xi
This is Green’s third formula.
The singularity 1/|r − rP | in the volume integral in Eq. T.19 is not a problem.
Consider a small sphere with radius r = |ra − rP | centered at point P . In spherical
coordinates the volume element can then be expressed as
dV = r2 sin θdrdθdα = r2 drdΩ (T.20)
Hence it is seen that volume element dV goes to zero faster than the singularity 1/|r −
rP |.
T.4. Analytical solution to Poisson’s equation 345

T.4 Analytical solution to Poisson’s equation

Poisson’s equation reads
∂2ϕ
=f (T.21)
∂xj ∂xj
where we assume that ϕ goes to zero at infinity and that the right side is limited.
Green’s third formula, Eq. T.19, gives
Z
1 f (r)
ϕ(rP ) = − dV
4π V |r − rP |
Z Z (T.22)
1 ri − riP 1 1 ∂ϕ
+ ϕ n i dS + n i dS
4π S |r − rP |3 4π S |r − rP | ∂xi

We choose the volume as a large sphere with radius R. Using Eqs. T.12, T.13 and T.14,
the first surface integral can be written as
Z Z Z
1 ri − riP 1 1
ϕ n i dS = ϕn n
i i dS = ϕdΩ (T.23)
4π S |r − rP |3 4πR2 S 4π S

using ni ni = 1. This integral goes to zero for large R since ϕ → 0 as R → ∞.

The second surface integral in Eq. T.22 can be re-written using Eq. T.13, Gauss
divergence law and Eq. T.21 as
Z Z
1 1 ∂ϕ 1 ∂ϕ
ni dS = ni dS
4π S |r − rP | ∂xi 4πR S ∂xi
Z Z (T.24)
1 ∂2ϕ 1
= dV = f dV
4πR S ∂xi ∂xi 4πR V

This integral also goes to zero for large R since we have assumed that f is limited.
Hence the final form of Eq. T.22 reads
Z
1 f (r)
ϕ(rP ) = − dV (T.25)
4π V |r − rP |

This is the analytical solution to Poisson’s equation, Eq. T.21.

U. MTF270: Learning outcomes for 2017 346

U MTF270: Learning outcomes for 2017

Week 1
1. How is the buoyancy term, ρgi , re-written in incompressible flow?
2. Given the transport equation for the temperature, θ, and the transport equation
for θ̄, show the principles (in the same way as is done for the vi′ vj′ equation on
Section 11.2 on p. 117) how to derive the transport equation for vi′ θ′ , Eq. 11.22.
Discuss the physical meaning of the different terms. Which terms need to be
modeled?
3. What is the expression for the total heat flux that appears in the θ̄ equation?
4. Show the principles how to derive the transport equation for vi′ vj′ , Eq. 11.11 (see
Section 11.2 on p. 117)
5. Which terms in the vi′ vj′ equation need to be modeled? Explain the physical
meaning of the different terms in the vi′ vj′ equation.
6. Show how the turbulent diffusion (i.e. the term which includes the triple corre-
lation) in the k equation is modeled.
7. How is the production term modeled in the k − ε model? Show how it can be
expressed in s̄ij
8. Given the modeled k equation, derive the modeled ε equation.
9. How are the Reynolds stress, vi′ vj′ , and the turbulent heat flux, vi′ θ′ , modeled in
the Boussinesq approach?
10. Discuss and show how the dissipation term, εij , is modeled.
11. Use physical reasoning to derive a model for the diagonal components of the
pressure-strain term (slow part).

12. Using physical reasoning, the model for the pressure-strain term above is for-
mulated only for the normal streses. Show that if the model is expressed in the
principal directins, then a model for the shear stress is also obtained.
U. MTF270: Learning outcomes for 2017 347

Week 2
 
1. The slow pressure-strain model reads Φij,1 = −c1 ρ kε vi′ vj′ − 32 δij k . The
v′ v′
anisotropy tensor is defined as aij = ik j − 32 δij . Show that for decaying grid
turbulence, the model for the slow pressure-strain model indeed acts as to make
the turbulence more isotropic if c1 > 1.
2. Derive the exact Poisson equation for the pressure fluctuation, Eq. 11.65.
3. For a Poisson equation
∂2ϕ
=f
∂xj ∂xj
there exists an exact analytical solution
Z
1 f (y)dy1 dy2 dy3
ϕ(x) = − (U.1)
4π V |y − x|

Use Eqs. 11.65 and U.1 to derive the exact analytical solution (Eq. 11.68) for
the pressure-strain term. Which are the “slow” and “rapid” terms? Why are they
called “slow” and “rapid”?
4. Derive the algebraic stress model (ASM). What main assumption is made?
5. Show the physical reasoning leading to the modeled slow pressure strain term,
Φ22,1w , for wall effects. What sign does it have? Give also the expressions for
Φ11,1w and Φ33,1w
U. MTF270: Learning outcomes for 2017 348

Week 3
 
1. The modeled slow and rapid pressure strain term read Φij,1 = −c1 ρ kε vi′ vj′ − 32 δij k


and Φij,2 = −c2 Pij − 23 δij P k , respectively. Give the expression for the pro-
duction terms, modeled pressure-strain terms and modeled dissipation terms for
a simple shear flow (e.g. boundary layer, channel flow, jet flow . . . ). In some
stress equations there is no production terms nor any dissipation term. How
come? Which is the main source term (or sink term) in these equations?
2. Describe the physical effect of stable stratification and unstable stratification on
turbulence.
3. Consider buoyancy-dominated flow with x3 vertically upwards. The production
term for the vi′ vj′ and the vi′ θ′ equations read

∂ θ̄
Gij = −gi βvj′ θ′ − gj βvi′ θ′ , Piθ = −vi′ vk′
∂xk
respectively (we assume that the velocity gradient is negligible). Show that the
Reynolds stress model dampens and increases the vertical fluctuation in stable
and unstable stratification, respectively, as it should. Show also that k in the
k − ε model is affected in the same way.
4. Consider streamline curvature for a streamline formed as a circular arc (convex
curvature). Show that the turbulence is dampened if ∂vθ /∂r > 0 and that it is
enhanced if the sign of ∂vθ /∂r is negative.
5. Streamline curvature: now consider a boundary layer where the streamlines are
curved away from the wall (concave curvature). Show that the Reynolds stress
model gives an enhanced turbulence production (as it should) because of positive
feedback between the production terms. Why is the effect of streamline curvature
in the k − ε model is much smaller?

6. Consider stagnation flow. Show that in the Reynolds stress model, there is only a
small production of turbulence whereas eddy-viscosity models (such as the k − ε
model) give a large production of turbulence.
7. What is a realizability constraint? There are two main realizability constraints,
one on the normal and one on the shear stresses: give the form of these con-
straints.
8. Show that the Boussinesq assumption may give negative normal stresses. In
which coordinate system is the risk largest for negative normal stresses? Derive
an expression (2D) how to avoid negative normal stresses by reducing the turbu-
lent viscosity (Eq. 13.12).
Hint: the eigenvalues, λ1 , λ2 , are obtained from |Sij − δij λ| = 0.
9. What is a non-linear eddy-viscosity model? When formulating a non-linear
model, the anisotropy tensor aij = −2νt s̄ij /k is often used. The three terms
U. MTF270: Learning outcomes for 2017 349

read
 
2 1
c1 τ s̄ik s̄kj − s̄ℓk s̄ℓk δij
3


+c2 τ 2 Ω̄ik s̄kj − s̄ik Ω̄kj
 
2 1
+c3 τ Ω̄ik Ω̄jk − Ω̄ℓk Ω̄ℓk δij
3

Show that each term has the same properties as aij , i.e. non-dimensional, trace-
less and symmetric.
U. MTF270: Learning outcomes for 2017 350

Week 4
1. Which equations are solved in the V2F model?
2. The transport equation for v2′2 reads (the turbulent diffusion term is modeled)
" #
∂ρv̄1 v2′2 ∂ρv̄v2′2 ∂ ∂v2′2
+ = (µ + µt ) −2v ′ ∂p′ /∂x2 −ρε22
∂x1 ∂x2 ∂x2 ∂x2 | 2 {z }
ρΦ22

Show how this equation is re-written in the V2F model.

3. The f equation in the V2F model reads
!
2
2∂f Φ22 1 v2′2 2 k k 3/2
L −f =− − − , T ∝ , L∝
∂x22 k T k 3 ε ε
Explain how the magnitude of the right side and L affect f (Fig 15.1). How
does f enter into the v2′2 equation? Show that far from the walls, the V2F model
(i.e. the f and the v2′2 equation) returns to the v2′2 equation in the Reynolds stress
model. In the V2F model, the v 2 equation is solved: what is the difference
between v2′2 and v 2 (see the discussion in connection to Eq. 15.9)?
4. What does the acronym SST mean? The SST model is a combination of the
k − ε and the k − ω model. In which region is each model used and why? How
is ω expressed in k and ε?
5. Derive a transport equation for ω from the k and ε transport equations; you only
need to do the production, the destruction and the viscous diffusion terms.
6. In the SST model, a blending function F1 is used; what does this function do?
7. What is the purpose of the shear stress limiter in the SST model? Show that the
eddy-viscosity assumption gives too high a shear stress in APG since P k /ε ≫ 1
(Eq. 16.15).
8. Show the difference between volume averaging (filtering) in LES and time-
averaging in RANS.
9. Consider the spatial derivative of the pressure in the filtered Navier-Stokes: show
that the derivative can be moved outside the filtering integral (it gives an addi-
tional second-order term).
10. The filtered non-linear term has the form
∂vi vj
∂xj
Show that it can be re-written as
∂v̄i v̄j
∂xj
giving an additional term
∂ ∂ ∂τij
− (vi vj ) + (v̄i v̄j ) = −
∂xj ∂xj ∂xj
on the right side.
U. MTF270: Learning outcomes for 2017 351

11. Consider a 1D finite volume grid. Carry out a second filtering of v̄ at node I and
show that v̄ I 6= v̄I .
12. Consider the energy spectrum. Show the three different regions (the large energy-
containing scales, the −5/3 range and the dissipating scales). Where should the
cut-off be located? Show where the SGS scales, grid (i.e resolved) scales and the
cut-off, κc are located in the spectrum.
13. Show how a sinus wave sin(κc x) corresponding to cut-off is represented on a
grid with two and four cells, respectively. How is κc related to the grid size ∆x
for these cases?
14. Taking guidance from the RANS k equation, formulate the one-equation ksgs
equation
15. Consider the energy spectrum and discuss the physical meaning of Pksgs and
εsgs .
16. Discuss the energy path in connection to the source and sink terms in the k̄, K̄
and the ksgs equations, see Figs. M.3 and M.4. How are k̄ and ksgs computed
from the energy spectrum?
U. MTF270: Learning outcomes for 2017 352

Week 5
1. What is a test filter? Grid and test filter Navier-Stokes equation and derive the
relation
z { z{ z{ z{ z{
v̄i v̄j − v̄ i v̄ j + τ ij = Lij + τ ij = Tij (U.2)
Draw an energy spectrum and show which wavenumber range k̄, ksgs , ksgs,test
cover.
2. Formulate the Smagorinsky model for the grid filter SGS stress, τij , and the test
filter SGS stress, Tij . Use Eq. U.2 and derive the relation
!
1 z{2 z{ z{ z {
Lij − δij Lkk = −2C ∆ | s̄ | s̄ ij − ∆2 |s̄|s̄ij
3

3. The equation you dervived above is a tensor equation for C. Use this relation
and derive the final expression for the dynamic coefficient, C, Eq. 18.38.

4. What are the five main differences between a RANS finite volume CFD code
and a LES finite volume CFD code? What do you need to consider in LES when
you want to compute time-averaged quantities? (see Fig. 18.11). How can the
integral time scale, Tint , be used?(see Section M.3)
U. MTF270: Learning outcomes for 2017 353

Week 6
1. Derive the Smagorinsky model in two different ways (Sections 18.6 and 18.22)
2. Show that when a first-order upwind schemes is used for the convection term,
an additional diffusion term and dissipation terms appear because of a numerical
SGS viscosity
3. When doing LES, how fine does the mesh need to be in the wall region be? Why
does it need to be that fine?
4. Describe URANS. How is the instantaneous velocity decomposed? What turbu-
lence models are used? What is scale separation?
 2
ν̃t
5. What is DES? The length scale in the RANS S-A model reads ; how is it
d
computed in the corresponding DES model?
6. How is the length scale computed in a k − ε two-equation DES model? Where
in a boundary layer does the DES model switch from RANS to LES?
7. The modified (reduced) length scale in two-equation DES models can be intro-
duced in different ways. It it usually introduced in one transport equation. Which
one and which term? Apart from this transport equation, it is sometimes used in
a another equation. Which one? What is the effect on the modeled, turbulent
quantities?
8. What is DDES? Why was it invented?
9. Describe hybrid LES-RANS based on a one-equation model.
U. MTF270: Learning outcomes for 2017 354

Week 7
1. What is the physical meaning of fk in PANS?
2. The PANS equations are given in Eqs. 23.21. Assume that fε = 1. Consider the
∗
destruction term in the ε equation and the coefficient Cε2 . Explain what happens
if fk is reduced from 1 (RANS mode) to fk = 0.4 (LES mode).
3. Describe the SAS model. How is the von Kármán length scale defined? An
additional source term is introduced in the ω equation: what is the form of this
term? Describe how this source reduces the turbulent viscosity. When is the term
large and small, respectively?
4. Give a short description of the method to generate synthetic turbulent inlet fluc-
tuations; see Chapter 27 and first 13 slides in
http://www.tfd.chalmers.se/˜lada/slides/slides inlet.pdf
What form on the spectrum is assumed? How is the wavenumber, κe , for the
energy-containing eddies, determined? How are the maximum and minimum
wavelengths, κmax , κmin , determined? With this method, the generated shear
stress is zero: why? How is the correlation in time achieved?
V. References 355

V References
[1] J. N. Reddy. An Introduction to Continuum Mechanics. Cambridge University
Press, New York, 2008.
[2] S. Toll and M. Ekh. Mechanics of solids & fluids. Part I: Fundamentals. Report,
Div. of Material and Computational Mechanics, Dept. of Applied Mechanics,
Chalmers University of Technology, Göteborg, Sweden, 2016.

[3] H. Abedi. Development of Vortex Filament Method for Wind Power Aerodynam-
ics. PhD thesis, Div. of Fluid Dynamics, Dept. of Applied Mechanics, Chalmers
University of Technology, Göteborg, Sweden, 2016.
[4] J. Wojtkowiak and C.O. Popiel. Inherently linear annular-duct-type laminar
flowmeter. Journal of Fluids Engineering, 128(1):196–198, 2006.
[5] H. Blasius. Grenzschichten in Flüssigkeiten mit kleiner Reibung. PhD thesis,
University of Göttingen, Göttingen, Germany, 1907.
[6] W. H. Hager. Blasius: A life in research and education. Experiments in Fluids,
34:566–571, 2003.
[7] S. G. Krantz. A Guide to Complex Variables. Cambridge University Press, 2007.
[8] L. Qu, Ch. Norberg, L. Davidson, S.-H. Peng, and F. Wang. Quantitative nu-
merical analysis of flow past a circular cylinder at Reynolds number between 50
and 200. Journal of Fluids and Structures, 39:347–370, 2013.
[9] S. Hong and T. Asai. Effect of panel shape of soccer ball on its flight character-
istics. Scientific Reports, 4(5068), 2014.
[10] W. Zhang, R. Bensow, M. Golubev, and V Chernoray. Flow past a rotating finite
length cylinder: numerical and experimental study. In 51st AIAA Aerospace
Sciences Meeting including the New Horizons Forum and Aerospace Exposition,
AIAA 2013-0987, 2012.
[11] H. Abedi. Development of vortex filament method for aerodynamic loads on
rotor blades. Thesis for licentiate of engineering, Div. of Fluid Dynamics, Dept.
of Applied Mechanics, Chalmers University of Technology, Göteborg, Sweden,
2013.
[12] S. B. Pope. Turbulent Flow. Cambridge University Press, Cambridge, UK,
2001.
[13] H. Tennekes and J. L. Lumley. A First Course in Turbulence. The MIT Press,
Cambridge, Massachusetts, 1972.
[14] P. Bradshaw. Turbulence. Springer-Verlag, Berlin, 1976.
[15] S. Hoyas and J. Jiménez. Scaling of the velocity fluctuations in turbulent chan-
nels up to Reτ = 2003. Physics of Fluids A, 18(011702), 2006.
[16] S. Hoyas and J. Jiménez. http://torroja.dmt.upm.es/ftp/channels/data/statistics/.
2006.
V. References 356

[17] Th. von Kármán. Mechanische ähnlichkeit und Turbulenz. Nachrichten von
der Gesellschaft der Wissenchaften zu Göttingen, Fachgruppe 1 (Mathematik),
5:58–76, 1930.
[18] J.M. Österlund. Experimental Studies of Zero Pressure-Gradient Turbulent
Boundary-Layer Flow. PhD thesis, Department of Mechanics, Royal Institute
of Technology, Stockholm, Sweden, 1999.
[19] M. Österlund, A.V. Johansson, H.M. Nagib, and M.H. Hites. Mean-flow charac-
teristics of high Reynolds number turbulent boundary layers from two facilities.
In 30th Fluid Dynamics Conference, 28 June-1 July, Norfolk, Virginia, 1999.
[20] M. Österlund, A.V. Johansson, H.M. Nagib, and M.H. Hites. A note on the
overlap region in turbulent boundary layers. Physics of Fluids A, 12(1):1–4,
2000.

[21] P. R. Spalart. Direct simulation of a turbulent boundary layer up to Rθ = 1410.

Journal of Fluid Mechanics, 187:61–98, 1988.
[22] A. A. Townsend. Equilibrium layers and wall turbulence. Journal of Fluid
Mechanics, 11:97–120, 1961.
[23] L. Davidson. Transport equations in incompressible URANS and LES. Re-
port 2006/01, Div. of Fluid Dynamics, Dept. of Applied Mechanics, Chalmers
University of Technology, Göteborg, Sweden, 2006.
[24] K. Hanjalić. Advanced turbulence closure models: A view of current status and
future prospects. Int. J. Heat and Fluid Flow, 15:178–203, 1994.
[25] D. C. Wilcox. Turbulence Modeling for CFD. DCW Industries, Inc., 5354 Palm
Drive, La Cañada, California 91011, 2 edition, 1998.

[26] K. Hanjalić and B. E. Launder. A Reynolds stress model of turbulence and its
application to thin shear flows. Journal of Fluid Mechanics, 52:609–638, 1972.
[27] J. C. Rotta. Statistische theorie nichthomogener turbulenz. Zeitschrift für Physik,
129:547–572, 1951.
[28] G. E. Mase. Continuum Mechanics. Schaum’s Outline Series. McGraw-Hill,
1970.

[29] S. Fu, B. E. Launder, and D. P. Tselepidakes. Accomodating the effects of

high strain rates in modelling the pressure-starin correlation. Technical Report
TFD/87/5, UMIST, Mechanical Engineering Dept., 1987.
[30] B. E. Launder, G. J. Reece, and W. Rodi. Progress in the development of a
Reynolds-stress turbulence closure. Journal of Fluid Mechanics, 68(3):537–
566, 1975.
[31] D. Naot, A. Shavit, and M. Wolfshtein. Interactions between components of the
turbulent velocity correlation tensor. Israel J. Techn., 8:259, 1970.
[32] B. E. Launder. An introduction to single-point closure methodology. In T. B.
Gatski, M.Y. Hussaini, and J. L. Lumpley, editors, Simulation and Modeling of
Turbulent Flows, New York, 1996. Oxford University Press.
V. References 357

[33] M. M. Gibson and B. E. Launder. Ground effects on pressure fluctuations in the

atmospheric boundary layer. Journal of Fluid Mechanics, 86:491–511, 1978.
[34] S. B. Pope. A more general effective-viscosity hypothesis. Journal of Fluid
Mechanics, 472:331–340, 1975.
[35] S. Wallin and A. V. Johansson. A new explicit algebraic Reynolds stress model
for incompressible and compressible turbulent flows. Journal of Fluid Mechan-
ics, 403:89–132, 2000.
[36] A. J. Johansson and S. Wallin. A new explicit algebraic Reynolds stress model.
In Advances in Turbulence VI, pages 31–34, Netherlands, 1996. Kluwer Aca-
demic Publishers.
[37] S. S. Girimaji. Fully-explicit and self-consistent algebraic Reynolds stress
model. Theoretical and Computational Fluid Dynamics, 8:387–402, 1996.
[38] S. S. Girimaji. Improved algebraic Reynolds stress model for engineering flows.
In W. Rodi and G. Bergeles, editors, Engineering Turbulence Modelling and
Measurements – ETMM3, pages 121–129. Elsevier, 1996.
[39] S. Wallin. private communication. Dept. of Mechanics, KTH, 2017.
[40] P. Bradshaw. Effects of streamline curvature on turbulent flow. Agardograph
progress no. 169, AGARD, 1973.
[41] W. Rodi and G. Scheuerer. Calculation of curved shear layers with two-equation
turbulence models. Physics of Fluids, 26:1422–1435, 1983.
[42] B. E. Thompson and J. H. Whitelaw. Characteristics of a trailing-edge flow with
turbulent boundary-layer separation. Journal of Fluid Mechanics, 157:305–326,
1985.
[43] L. Davidson. Prediction of the flow around an airfoil using a Reynolds stress
transport model. Journal of Fluids Engineering, 117:50–57, 1995.
[44] P. A. Durbin. On the k − 3 stagnation point anomaly. International Journal of
Heat and Fluid Flow, 17:89–90, 1996.
[45] J. Hult. Cartesiska tensorer: Ett praktiskt fikonspråk. Almqvist & Wiksell,
1972.
[46] J. L. Lumley. Computational modeling of turbulent flows. Advances in Applied
Mechanics, 18:123–176, 1978.
[47] T. J. Craft and B. E. Launder. A Reynolds stress closure designated for com-
plex geometries. International Journal of Heat and Fluid Flow, 17(3):245–254,
1996.
[48] T. J. Craft, B. E. Launder, and K. Suga. Prediction of turbulent transitional
phenomena with a nonlinear eddy-viscosity model. International Journal of
Heat and Fluid Flow, 18:15–28, 1997.
[49] T. B. Gatski and C. G. Speziale. On explicit algebraic stress models for complex
turbulent flows. Journal of Fluid Mechanics, 154:59–78, 1993.
V. References 358

[50] T.-H. Shih, J. Zhu, and J. L. Lumley. A new Reynolds stress algebraic equation
model. Comput. Methods Appl. Engng., 125:287–302, 1995.
[51] P. A. Durbin. Near-wall turbulence closure modeling without damping func-
tions. Theoretical and Computational Fluid Dynamics, 3:1–13, 1991.
[52] P. A. Durbin. Application of a near-wall turbulence model to boundary layers
and heat transfer. International Journal of Heat and Fluid Flow, 14(4):316–323,
1993.
[53] L. Davidson. An introduction to turbulence models. Technical Report 97/2,
Dept. of Thermo and Fluid Dynamics, Chalmers University of Technology,
Gothenburg, 1997.
[54] A. Sveningsson. Analysis of the performance of different v 2 − f turbulence
models in a stator vane passage flow. Thesis for licentiate of engineering, Dept.
of Thermo and Fluid Dynamics, Chalmers University of Technology, Göteborg,
Sweden, 2003.
[55] K. Hanjalic, M. Popovacx, and M. Habziabdic. A robust near-wall elliptic relax-
ation eddy-viscosity turbulence model for CFD. International Journal of Heat
and Fluid Flow, 2004.
[56] F-S Lien and G. Kalitzin. Computations of transonic flow with the v2f turbu-
lence model. International Journal of Heat and Fluid Flow, 22(1):53–61, 2001.
[57] A. Sveningsson and L. Davidson. Assessment of realizability constraints in v 2 −
f turbulence models. International Journal of Heat and Fluid Flow, 25(5):785–
794, 2004.
[58] A. Sveningsson and L. Davidson. Computations of flow field and heat transfer in
a stator vane passage using the v 2 − f turbulence model. Paper GT2004-53586,
ASME TURBO EXPO 2004, Vienna, 2004.
[59] A. Sveningsson and L. Davidson. Computations of flow field and heat transfer
in a stator vane passage using the v 2 − f turbulence model. Journal of Turbo-
machinery, 127(3):627–634, 2005.
[60] A. Sveningsson. Turbulence Transport Modelling in Gas Turbine Related Ap-
plications. PhD thesis, Div. of Fluid Dynamics, Dept. of Applied Mechanics,
Chalmers University of Technology, Göteborg, Sweden, 2006.
[61] L. Davidson, P. V. Nielsen, and A. Sveningsson. Modifications of the v 2 − f
model for computing the flow in a 3D wall jet. In K. Hanjalić, Y. Nagano, and
M. J. Tummers, editors, Turbulence Heat and Mass Transfer 4, pages 577–584,
New York, Wallingford (UK), 2003. begell house, inc.
[62] F. R. Menter. Two-equation eddy-viscosity turbulence models for engineering
applications. AIAA Journal, 32:1598–1605, 1994.
[63] D. C. Wilcox. Reassessment of the scale-determining equation. AIAA Journal,
26(11):1299–1310, 1988.
[64] F. R. Menter. Influence of freestream values on k − ω turbulence model predic-
tions. AIAA Journal, 30(6):1657–1659, 1992.
V. References 359

[65] A. Hellsten. New advanced k − ω turbulence model for high-lift aerodynamics.

AIAA Journal, 43(9):1857–1869, 2005.
[66] J. Bredberg. On two-equation eddy-viscosity models. Report 01/8, can be
downloaded from www.tfd.chalmers.se/˜lada/allpaper.html, Dept. of Thermo
and Fluid Dynamics, Chalmers University of Technology, 2001.
[67] P. Bradshaw, D. H. Ferriss, and N. P. Atwell. Calculation of boundary-layer
development using the turbulent energy equation. Journal of Fluid Mechanics,
28:593–616, 1967.
[68] F. R. Menter, M. Kuntz, and R. Langtry. Ten years of industrial experience of the
SST turbulence model. In K. Hanjalić, Y. Nagano, and M. J. Tummers, editors,
Turbulence Heat and Mass Transfer 4, pages 624–632, New York, Wallingford
(UK), 2003. begell house, inc.

[69] S. Ghosal and P. Moin. The basic equations for the large eddy simulation of
turbulent flows in complex geometry. J. Comp. Phys., 118:24–37, 1995.
[70] J. Smagorinsky. General circulation experiments with the primitive equations.
Monthly Weather Review, 91:99–165, 1963.
[71] P. Moin and J. Kim. Numerical investigation of turbulent channel flow. Journal
of Fluid Mechanics, 118:341–377, 1982.

[72] W. P. Jones and M. Wille. Large eddy simulation of a jet in a cross-flow. In

10th Symp. on Turbulent Shear Flows, pages 4:1 – 4:6, The Pennsylvania State
University, 1995.
[73] J. O. Hinze. Turbulence. McGraw-Hill, New York, 2nd edition, 1975.
[74] H. Schlichting. Boundary-Layer Theory. McGraw-Hill, New York, 7 edition,
1979.
[75] M. Germano, U. Piomelli, P. Moin, and W. H. Cabot. A dynamic subgrid-scale
eddy viscosity model. Physics of Fluids A, 3:1760–1765, 1991.
[76] Y. Zang, R. L. Street, and J. R. Koseff. A dynamic mixed subgrid-scale model
and its application to turbulent recirculating flows. Physics of Fluids A, 5:3186–
3196, 1993.

[77] D. K. Lilly. A proposed modification of the Germano subgrid-scale closure

method. Physics of Fluids A, 4:633–635, 1992.
[78] J. P. Boris, F. F. Grinstein, E. S. Oran, and R. L. Kolbe. New insights into large
eddy simulation. Fluid Dynamic Research, 10:199–228, 1992.
[79] J. Bardina, J. H. Ferziger, and W. C. Reynolds. Improved subgrid scale models
for large eddy simulation. AIAA 80-1357, Snomass, Colorado, 1980.

[80] C. G. Speziale. Galilean invariance of subgrid-scale stress models in the large-

eddy simulation of turbulence. Journal of Fluid Mechanics, 156:55–62, 1985.
[81] M. Germano. A proposal for a redefinition of the turbulent stresses in the filtered
Navier-Stokes equations. Phys. Fluids A, 7:2323–2324, 1986.
V. References 360

[82] L. Davidson. A dissipative scale-similarity model. In V. Armenio, B. Geurts, and

J. Fröhlich, editors, DLES7: Direct and Large-Eddy Simulations 7, volume 13
of ERCOFTAC series, pages 261–266, 2010.
[83] C. Hirsch. Numerical Computation of Internal and External Flows: Fundamen-
tals of Numerical Discretization, volume 1. John Wiley & Sons, Chichester, UK,
1988.
[84] L. Davidson. Hybrid LES-RANS: back scatter from a scale-similarity model
used as forcing. Phil. Trans. of the Royal Society A, 367(1899):2905–2915,
2009.
[85] P. Emvin. The Full Multigrid Method Applied to Turbulent Flow in Ventilated
Enclosures Using Structured and Unstructured Grids. PhD thesis, Dept. of
Thermo and Fluid Dynamics, Chalmers University of Technology, Göteborg,
1997.
[86] L. Davidson and S.-H. Peng. Hybrid LES-RANS: A one-equation SGS model
combined with a k − ω model for predicting recirculating flows. International
Journal for Numerical Methods in Fluids, 43:1003–1018, 2003.
[87] S. V. Patankar. Numerical Heat Transfer and Fluid Flow. McGraw-Hill, New
York, 1980.
[88] H. K. Versteegh and W. Malalasekera. An Introduction to Computational Fluid
Dynamics - The Finite Volume Method. Longman Scientific & Technical, Har-
low, England, 1995.
[89] L. Davidson and M. Billson. Hybrid LES/RANS using synthesized turbulence
for forcing at the interface. International Journal of Heat and Fluid Flow,
27(6):1028–1042, 2006.
[90] L. Davidson. Large eddy simulation: A dynamic one-equation subgrid model
for three-dimensional recirculating flow. In 11th Int. Symp. on Turbulent Shear
Flow, volume 3, pages 26.1–26.6, Grenoble, 1997.
[91] L. Davidson. Large eddy simulations: A note on derivation of the equations for
the subgrid turbulent kinetic energies. Technical Report 97/11, Dept. of Thermo
and Fluid Dynamics, Chalmers University of Technology, Gothenburg, 1997.
[92] 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(1):39–47, 2000.
[93] A. Sohankar, L. Davidson, and C. Norberg. Erratum. Journal of Fluids Engi-
neering, 122(3):643, 2000.
[94] S. Krajnović and L. Davidson. A mixed one-equation subgrid model for large-
eddy simulation. International Journal of Heat and Fluid Flow, 23(4):413–425,
2002.
[95] S. Krajnović and L. Davidson. Large eddy simulation of the flow around a bluff
body. AIAA Journal, 40(5):927–936, 2002.
V. References 361

[96] S. Krajnović. Large Eddy Simulations for Computing the Flow Around Vehi-
cles. PhD thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of
Technology, Göteborg, Sweden, 2002.
[97] J. Pallares and L. Davidson. Large-eddy simulations of turbulent flow in a rotat-
ing square duct. Physics of Fluids, 12(11):2878–2894, 2000.
[98] J. Pallares and L. Davidson. Large-eddy simulations of turbulent heat transfer in
stationary and rotating square duct. Physics of Fluids, 14(8):2804–2816, 2002.
[99] S. Krajnović and L. Davidson. Numerical study of the flow around the bus-
shaped body. Journal of Fluids Engineering, 125:500–509, 2003.
[100] S. Krajnović and L. Davidson. Flow around a simplified car. part I: Large eddy
simulations. Journal of Fluids Engineering, 127(5):907–918, 2005.
[101] S. Krajnović and L. Davidson. Flow around a simplified car. part II: Understand-
ing the flow. Journal of Fluids Engineering, 127(5):919–928, 2005.
[102] S. Krajnović and L. Davidson. Influence of floor motions in wind tunnels on
the aerodynamics of road vehicles. Journal of Wind Engineering and Industrial
Aerodynamics, 92(5):677–696, 2005.
[103] S. Dahlström and L. Davidson. Large eddy simulation applied to a high-
Reynolds flow around an airfoil close to stall. AIAA paper 2003-0776, 2003.
[104] S. Dahlström. Large Eddy Simulation of the Flow Around a High-Lift Airfoil.
PhD thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of Tech-
nology, Göteborg, Sweden, 2003.
[105] J. Ask and L. Davidson. Flow and dipole source evaluation of a generic SUV.
Journal of Fluids Engineering, 132(051111), 2010.
[106] H. Hemida and S. Krajnović. LES study of the impact of the wake structures on
the aerodynamics of a simplified ICE2 train subjected to a side wind. In Fourth
International Conference on Computational Fluid Dynamics (ICCFD4), 10-14
July, Ghent, Belgium, 2006.
[107] H. Hemida and S. Krajnović. Numerical study of the unsteady flow structures
around train-shaped body subjected to side winds. In ECCOMAS CFD 2006,
5-8 September, Egmond aan Zee, The Netherlands, 2006.
[108] S. Krajnović. Numerical simulation of the flow around an ICE2 train under the
influence of a wind gust. In International Conference on Railway Engineering
2008 (IET ICRE2008), 25-28 March, Hong Kong, China, 2008.
[109] S.-H. Peng and L. Davidson. On a subgrid-scale heat flux model for large eddy
simulation of turbulent thermal flow. International Journal of Heat and Mass
Transfer, 45:1393–1405, 2002.
[110] S.-H. Peng and L. Davidson. Large eddy simulation for turbulent buoyant flow
in a confined cavity. International Journal of Heat and Fluid Flow, 22:323–331,
2001.
V. References 362

[111] S.-H. Peng. Modeling of Turbulent flow and Heat Transfer for Building Ventila-
tion. PhD thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of
Technology, Göteborg, 1998.
[112] D. G. Barhaghi and L. Davidson. Natural convection boundary layer in a 5:1
cavity. Physics of Fluids, 19(125106), 2007.
[113] D. Barhaghi. A Study of Turbulent Natural Convection Boundary Layers Using
Large-Eddy Simulation. PhD thesis, Div. of Fluid Dynamics, Dept. of Applied
Mechanics, Chalmers University of Technology, Göteborg, Sweden, 2007.
[114] D. G. Barhaghi, L. Davidson, and R. Karlsson. Large-eddy simulation of natural
convection boundary layer on a vertical cylinder. International Journal of Heat
and Fluid Flow, 27(5):811–820, 2006.
[115] D. G. Barhaghi and L. Davidson. Les of mixed convection boundary layer be-
tween radiating parallel plates. In 5th International Symposium on Turbulence,
Heat and Mass Transfer, September 25-29, 2006, Dubrovnik, Croatia, 2006.
[116] S. R. Robinson. Coherent motions in the turbulent boundary layer. Annual
Review of Fluid Mechanics, 23:601–639, 1991.
[117] L. Davidson. Large eddy simulations: how to evaluate resolution. International
Journal of Heat and Fluid Flow, 30(5):1016–1025, 2009.

[118] L. Davidson. How to estimate the resolution of an LES of recirculating flow.

In M. V. Salvetti, B. Geurts, J. Meyers, and P. Sagaut, editors, ERCOFTAC,
volume 16 of Quality and Reliability of Large-Eddy Simulations II, pages 269–
286. Springer, 2010.
[119] L. Davidson. Inlet boundary conditions for embedded LES. In First CEAS
European Air and Space Conference, 10-13 September, Berlin, 2007.
[120] L. Davidson, D. Cokljat, J. Fröhlich, M. A. Leschziner, C. Mellen, and W. Rodi,
editors. LESFOIL: Large Eddy Simulation of Flow Around a High Lift Airfoil,
volume 83 of Notes on Numerical Fluid Mechanics. Springer Verlag, 2003.
[121] C. P. Mellen, J. Fröhlich, and W. Rodi. Lessons from LESFOIL project on large
eddy simulation of flow around an airfoil. AIAA Journal, 41(4):573–581, 2003.
[122] S. Johansson, L. Davidson, and E. Olsson. Numerical simulation of vortex shed-
ding past triangular cylinders at high reynolds numbers using a k − ε turbulence
model. International Journal for Numerical Methods in Fluids, 16:859–878,
1993.
[123] S. Perzon and L. Davidson. On CFD and transient flow in vehicle aerodynamics.
In SAE Technical Paper 2000-01-0873, Detroit, 2000.
[124] S. Perzon and L. Davidson. On transient modeling of the flow around vehicles
using the Reynolds equation. In J.-H. Wu, Z.-J. Zhu, F.-P. Jia, X.-B. Wen, and
W. Hu, editors, ACFD 2000, Bejing, 2000.
[125] T. H. Shih, J. Zhu, and J. L. Lumley. A realizable Reynolds stress algebraic
equation model. Technical memo 105993, NASA, 1993.
V. References 363

[126] P. Spalart and S. Allmaras. A one equation turbulence model for aerodynamic
flows. AIAA paper 92-0439, Reno, NV, 1992.
[127] P. R. Spalart, W.-H. Jou, M. Strelets, and S. R. Allmaras. Comments on the
feasability of LES for wings and on a hybrid RANS/LES approach. In C. Liu and
Z. Liu, editors, Advances in LES/DNS, First Int. conf. on DNS/LES, Louisiana
Tech University, 1997. Greyden Press.
[128] F. R. Menter, M. Kuntz, and R. Bender. A scale-adaptive simulation model for
turbulent flow prediction. AIAA paper 2003–0767, Reno, NV, 2003.
[129] A. K. Travin, M. Shur, , M. Strelets, and P. R. Spalart. Physical and numeri-
cal upgrades in the detached-eddy simulations of complex turbulent flows. In
R. Friederich and W.Rodi, editors, Fluid Mechanics and its Applications. Ad-
vances in LES of Complex Flows, volume 65 of Euromech Colloquium 412.
Fluid Mechanics and its Applications. Advances in LES of Complex Flows,
pages 239–254. Academic Publishers, Dordrecht, 2002.
[130] C. De Langhe, B. Merci, K. Lodefier, and E. Dick. Hybrid RANS-LES mod-
elling with the renormalization group. In K. Hanjalić, Y. Nagano, and M. J.
Tummers, editors, Turbulence Heat and Mass Transfer 4, pages 697–704, New
York, Wallingford (UK), 2003. begell house, inc.
[131] J. C. Kok, H. S. Dol, B. Oskam, and H. van der Ven. Extra-large eddy simulation
of massively separated flows. AIAA paper 2004-264, Reno, NV, 2004.
[132] J. Yan, C. Mocket, and F. Thiele. Investigation of alternative length scale substi-
tutions in detached-eddy simulation. Flow, Turbulence and Combustion, 74:85–
102, 2005.
[133] S. S. Girimaji. Partially-averaged Navier-Stokes model for turbulence: A
Reynolds-averaged Navier-Stokes to direct numerical simulation bridging
method. ASME Journal of Applied Mechanics, 73(2):413–421, 2006.
[134] J. Ma, S.-H. Peng, L. Davidson, and F. Wang. A low Reynolds
number variant of Partially-Averaged Navier-Stokes model for turbulence.
International Journal of Heat and Fluid Flow, 32(3):652–669, 2011.
10.1016/j.ijheatfluidflow.2011.02.001.
[135] R. Schiestel and A. Dejoan. Towards a new partially integrated transport model
for coarse grid and unsteady turbulent flow simulations. Theoretical and Com-
putational Fluid Dynamics, 18(6):443–468, 2005.
[136] K. Abe, T. Kondoh, and Y. Nagano. A new turbulence model for predicting
fluid flow and heat transfer in separating and reattaching flows - 1. Flow field
calculations. Int. J. Heat Mass Transfer, 37(1):139–151, 1994.
[137] L. Davidson. Zonal PANS: evaluation of different treatments of the RANS-LES
interface. Journal of Turbulence, 17(3):274–307, 2016.
[138] P. Batten, U. Goldberg, and S. Chakravarthy. Interfacing statistical turbulence
closures with large-eddy simulation. AIAA Journal, 42(3):485–492, 2004.
V. References 364

[139] U. Piomelli, E. Balaras, H. Pasinato, K. D. Squire, and P. R. Spalart. The inner-

outer layer interface in large-eddy simulations with wall-layer models. Interna-
tional Journal of Heat and Fluid Flow, 24:538–550, 2003.
[140] L. Davidson and S. Dahlström. Hybrid LES-RANS: An approach to make LES
applicable at high Reynolds number. International Journal of Computational
Fluid Dynamics, 19(6):415–427, 2005.
[141] J. Larsson, F. S. Lien, and E. Yee. The artificial buffer layer and the effects of
forcing in hybrid LES/RANS. International Journal of Heat and Fluid Flow,
28(6):1443–1459, 2007.
[142] F. R. Menter and M. Kuntz. Adaption of eddy-viscosity turbulence models to un-
steady separated flows behind vehicles. In Rose McCallen, Fred Browand, and
James Ross, editors, The Aerodynamics of Heavy Vehicles: Trucks, Buses, and
Trains, volume 19 of Lecture Notes in Applied and Computational Mechanics.
Springer Verlag, 2004.
[143] M. Strelets. Detached eddy simulation of massively separated flows. AIAA
paper 2001–0879, Reno, NV, 2001.
[144] P. R. Spalart. Strategies for turbulence modelling and simulations. International
Journal of Heat and Fluid Flow, 21:252–263, 2000.
[145] N. V. Nikitin, F. Nicoud, B. Wasistho, K. D. Squires, and P. Spalart. An approach
to wall modeling in large-eddy simulations. Physics of Fluids, 12(7):1629–1632,
2000.
[146] F. R. Menter and Y. Egorov. Revisiting the turbulent length scale equation. In IU-
TAM Symposium: One Hundred Years of Boundary Layer Research, Göttingen,
2004.
[147] F. R. Menter, M. Kuntz, and R. Bender. A scale-adaptive simulation model for
turbulent flow predictions. AIAA-2003-0767, Reno, 2003.
[148] J. C. Rotta. Turbulente Strömungen. Teubner Verlag, Stuttgart, 1972.
[149] F. R. Menter and Y. Egorov. A scale-adaptive simulation model using two-
equation models. AIAA paper 2005–1095, Reno, NV, 2005.
[150] L. Davidson. Evaluation of the SST-SAS model: Channel flow, asymmetric
diffuser and axi-symmetric hill. In ECCOMAS CFD 2006, September 5-8, 2006,
Egmond aan Zee, The Netherlands, 2006.
[151] S. S. Girimaji. Partially-averaged Navier-Stokes model for turbulence: A
Reynolds-averaged Navier-Stokes to direct numerical simulation bridging
method. ASME Journal of Applied Mechanics, 73(2):413–421, 2006.
[152] S. S. Girimaji and K. S. Abdol-Hamid. Partially-Averaged Navier-Stokes model
for turbulence: Implementation and Validation. AIAA paper 2005-0502, Reno,
N.V., 2005.
[153] M. L. Shur, P. R. Spalart, M. Kh. Strelets, and A. K. Travin. A hybrid RANS-
LES approach with delayed-DES and wall-modelled LES capabilities. Interna-
tional Journal of Heat and Fluid Flow, 29:1638–1649, 2008.
V. References 365

[154] J. Ma, S.-H. Peng, L. Davidson, and F. Wang. A low Reynolds number partially-
averaged Navier-Stokes model for turbulence. In 8th International ERCOFTAC
Symposium on Engineering Turbulence, Modelling and Measurements, Mar-
seille, France, 9-11 June, 2010.
[155] L. Davidson. The PANS k − ε model in a zonal hybrid RANS-LES formulation.
International Journal of Heat and Fluid Flow, 46:112–126, 2014.
[156] S. S. Girimaji and S. Wallin. Closure modeling in bridging regions of variable-
resolution (VR) turbulence computations. Journal of Turbulence, 14(1):72 – 98,
2013.
[157] B. Chaouat and R. Schiestel. A new partially integrated transport model
for subgrid-scale stresses and dissipation rate for turbulent developing flows.
Physics of Fluids, 17(065106), 2005.

[158] L. Davidson. Using isotropic synthetic fluctuations as inlet boundary condi-

tions for unsteady simulations. Advances and Applications in Fluid Mechanics,
1(1):1–35, 2007.
[159] L. Davidson and S.-H. Peng. Embedded large-eddy simulation using the par-
tially averaged Navier-Stokes model. AIAA Journal, 51(5):1066–1079, 2013.
[160] A. Travin, M. Shur, M., Strelets, and P. Spalart. Detached-eddy simulations past
a circular cylinder. Flow, Turbulence and Combustion, 63:293–313, 2000.
[161] F. Hamba. Analysis of filtered Navier-Stokes equation for hybrid RANS/LES
simulation. Physics of Fluids A, 23(015108), 2011.
[162] S. Arvidson, L. Davidson, and S.-H. Peng. Hybrid RANS-LES modeling based
on a low-Reynolds-number k − ω model. AIAA Journal, 54(12):4032–4037,
2016.
[163] L. Davidson. Two-equation hybrid RANS-LES models: A novel way to treat k
and ω at inlets and at embedded interfaces. Journal of Turbulence, 18(4):291–
315, 2017.
[164] S. Arvidson. Methodologies for RANS-LES interfaces in turbulence-resolving
simulations. PhD thesis, Div. of Fluid Dynamics, Dept. of Mechanics and Mar-
itime Sciences, Chalmers University of Technology, Göteborg, Sweden, 2017.
[165] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions for flows including
recirculation. In 5th International Symposium on Turbulence and Shear Flow
Phenomena, volume 2, pages 689–694, 27-29 August, Munich, Germany, 2007.
[166] N. Jarrin, S. Benhamadouche, D. Laurence, and R. Prosser. A synthetic-eddy-
method for generating inflow conditions for large-eddy simulations. Interna-
tional Journal of Heat and Fluid Flow, 27(4):585–593, 2006.
[167] M. Billson. Computational Techniques for Turbulence Generated Noise. PhD
thesis, Dept. of Thermo and Fluid Dynamics, Chalmers University of Technol-
ogy, Göteborg, Sweden, 2004.
V. References 366

[168] M. Billson, L.-E. Eriksson, and L. Davidson. Jet noise prediction using stochas-
tic turbulence modeling. AIAA paper 2003-3282, 9th AIAA/CEAS Aeroacous-
tics Conference, 2003.
[169] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions. In B. Skallerud
and H. I. Andersson, editors, 3rd National Conference on Computational Me-
chanics – MekIT’05 (invited paper), pages 7–22, Trondheim, Norway, 2005.
[170] L. Davidson. Hybrid LES-RANS: Inlet boundary conditions for flows with re-
circulation. In Second Symposium on Hybrid RANS-LES Methods, Corfu island,
Greece, 2007.
[171] J. R. Welty, C. E. Wicks, and R. E. Wilson. Fundamentals of Momentum, Heat,
and Mass Transfer. John Wiley & Sons, New York, 3 edition, 1984.
[172] L. Davidson. HYBRID LES-RANS: Inlet boundary conditions for flows with
recirculation. In Advances in Hybrid RANS-LES Modelling, volume 97 of Notes
on Numerical Fluid Mechanics and Multidisciplinary Design, pages 55–66.
Springer Verlag, 2008.
[173] M. Billson, L.-E. Eriksson, and L. Davidson. Modeling of synthetic anisotropic
turbulence and its sound emission. The 10th AIAA/CEAS Aeroacoustics Con-
ference, AIAA 2004-2857, Manchester, United Kindom, 2004.
[174] R. D. Moser, J. D. Kim, and N. N. Mansour. Direct numerical simulation of
turbulent channel flow up to Reτ = 590. Physics of Fluids, 11(4):943–945,
1999.
[175] L. Davidson. Matlab scriptt for synthetic fluctuations.
[176] Lars Davidson and Shia-Hui Peng. Hybrid LES-RANS: A one-equation SGS
model combined with a k − ω for predicting recirculating flows. International
Journal for Numerical Methods in Fluids, 43(9):1003–1018, 2003.
[177] F. R. Menter and Y. Egorov. The scale adaptive simulation method for unsteady
turbulent flow predictions. Part 1: Theory and description. Flow, Turbulence
and Combustion, 85:113–138, 2010.
[178] Y. Egorov, F. R. Menter, R. Lechner, and D. Cokljat. The scale adaptive sim-
ulation method for unsteady flow predictions. Part 2: Application to complex
flows. Flow, Turbulence and Combustion, 85:139–165, 2010.
[179] J. Fröhlich and D. von Terzi. Hybrid LES/RANS methods for the simulation of
turbulent flows. Progress in Aerospace, 44(5):349–377, 2008.
[180] A. Travin, M. Shur, M. Strelets, and P. Spalart. Detached-eddy simulations past
a circular cylinder. Flow Turbulence and Combustion, 63(1/4):293–313, 2000.
[181] L. Davidson and B. Farhanieh. CALC-BFC: A finite-volume code employing
collocated variable arrangement and cartesian velocity components for com-
putation of fluid flow and heat transfer in complex three-dimensional geome-
tries. Rept. 95/11, Dept. of Thermo and Fluid Dynamics, Chalmers University
of Technology, Gothenburg, 1995.
V. References 367

[182] B. P. Leonard. A stable and accurate convective modelling based on quadratic

upstream interpolation. Computational Methods in Applied Mechanical Engi-
neering, 19:59–98, 1979.
[183] F. Eriksson. Flerdimensionell analys (in Swedish). Studentlitteratur, Lund, Swe-
den, 4 edition, 1977.
[184] R. L. Panton. Incompressible Flow. John Wiley & Sons, New York, 1984.
[185] M. Mirzaei, L. Davidson, A. Sohankar, and F. Innings. The effect of corruga-
tion on heat transfer and pressure drop in channel flow with different Prandtl
numbers. International Journal of Heat and Mass Transfer, 66:164–176, 2013.
[186] B. J. Daly and F. H. Harlow. Transport equation in turbulence. Physics of Fluids,
13:2634–2649, 1970.
[187] S. M. El-Behery and M. H. Hamed. A comparative study of turbulence models
performance for separating flow in planar asymmetric diffuser. Computers &
Fluids, 44(1):248–257, 2011.
[188] C .U. Buice and J. K. Eaton. Experimental investigation of flow through an
asymmetric plane diffuser. Journal of Fluids Engineering, 122:433–435, 2000.
[189] F. Nicoud and F. Ducros. Subgrid-scale stress modelling based on the square of
the velocity gradient tensor. Flow, Turbulence and Combustion, 62(3):183–200,
1999.
[190] A. Mehdizadeh and A. Sadiki. A zonal hybrid turbulence model based on Rotta
integral length scale transport equation. In Turbulence, Heat and Mass Transfer,
THMT-12, Palermo, Sicily/Italy, 2012.
[191] L. Davidson and S.-H. Peng. Embedded LES with PANS. In 6th AIAA Theoreti-
cal Fluid Mechanics Conference, AIAA paper 2011-3108, 27-30 June, Honolulu,
Hawaii, 2011.
[192] D. Greenblatt, K. B. Paschal, C.-S. Yao, J. Harris, N. W. Schaeffler, and A. E.
Washburn. A separation control CFD validation test case. Part 1: Baseline &
steady suction. AIAA-2004-2220, 2004.
[193] D. Greenblatt, K. B. Paschal, C.-S. Yao, and J. Harris. A separation control CFD
validation test case Part 1: Zero efflux oscillatory blowing. AIAA-2005-0485,
2005.
[194] B. van Leer. Towards the ultimate conservative difference scheme. Monotonicity
and conservation combined in a second order scheme. Journal of Computational
Physics, 14(4):361–370, 1974.
[195] S. Dahlström and L. Davidson. Large eddy simulation of the flow around an
Aerospatiale A-aerofoil. In ECCOMAS 2000, European Congress on Com-
putational Methods in Applied Sciences and Engineering, 11-14 September,
Barcelona, Spain, 2000.
[196] B. Basara, S. Krajnović, S. Girimaji, and Z. Pavlović. Near-wall formula-
tion of the Partially Averaged Navier Stokes turbulence model. AIAA Journal,
49(12):2627–2636, 2011.
V. References 368

[197] L. Davidson. A new approach of zonal hybrid RANS-LES based on a two-

equation k − ε model. In ETMM9: 9th International ERCOFTAC Symposium
on Turbulence Modelling and Measurements, Thessaloniki, Greece, 2012.
[198] O. Brander. Vektoranalys (in Swedish). Studentlitteratur, Lund, Sweden, 2007.