This action might not be possible to undo. Are you sure you want to continue?

Lars Davidson Division of Fluid Dynamics Department of Applied Mechanics Chalmers University of Technology SE-412 96 G¨ teborg, Sweden o http://www.tfd.chalmers.se/˜lada, lada@chalmers.se July 7, 2012

Abstract This course material is used in two courses in the International Master’s programme Applied Mechanics at Chalmers. The two courses are TME225 Mechanics of ﬂuids, and MTF270 Turbulence Modeling. MSc students who follow these courses are supposed to have taken one basic course in ﬂuid mechanics. 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

1

Contents

1 Motion, ﬂow 1.1 Eulerian, Lagrangian, material derivative . . . . . . . . . 1.2 Viscous stress, pressure . . . . . . . . . . . . . . . . . . 1.3 Strain rate tensor, vorticity . . . . . . . . . . . . . . . . . 1.4 Product of a symmetric and antisymmetric tensor . . . . 1.5 Deformation, rotation . . . . . . . . . . . . . . . . . . . 1.6 Irrotational and rotational ﬂow . . . . . . . . . . . . . . 1.6.1 Ideal vortex line . . . . . . . . . . . . . . . . . . 1.6.2 Shear ﬂow . . . . . . . . . . . . . . . . . . . . . 1.7 Eigenvalues and and eigenvectors: physical interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 9 10 11 13 13 15 16 17 18 20 20 20 20 21 22 23 24 25 25 28 28 29 31 32 34 34 35 38 39 41 41 42 43 47 50 50 51 51 55 57 59

2

Governing ﬂow equations 2.1 The Navier-Stokes equation . . . . . . . . . . . . . . . . . . . . . 2.1.1 The continuity equation . . . . . . . . . . . . . . . . . . . 2.1.2 The momentum equation . . . . . . . . . . . . . . . . . . 2.2 The energy equation . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Transformation of energy . . . . . . . . . . . . . . . . . . . . . . 2.4 Left side of the transport equations . . . . . . . . . . . . . . . . . 2.5 Material particle vs. control volume (Reynolds Transport Theorem) Exact solutions to the Navier-Stokes equation: two examples 3.1 The Rayleigh problem . . . . . . . . . . . . . . . . . . . 3.2 Flow between two plates . . . . . . . . . . . . . . . . . . 3.2.1 Curved plates . . . . . . . . . . . . . . . . . . . 3.2.2 Flat plates . . . . . . . . . . . . . . . . . . . . . 3.2.3 Force balance . . . . . . . . . . . . . . . . . . . 3.2.4 Balance equation for the kinetic energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

3

4

Vorticity equation and potential ﬂow 4.1 Vorticity and rotation . . . . . . . . . . . . . . . . . . . . . 4.2 The vorticity transport equation in three dimensions . . . . . 4.3 The vorticity transport equation in two dimensions . . . . . . 4.3.1 Boundary layer thickness from the Rayleigh problem Turbulence 5.1 Introduction . . . . . . . . . . . . . . . 5.2 Turbulent scales . . . . . . . . . . . . . 5.3 Energy spectrum . . . . . . . . . . . . . 5.4 The cascade process created by vorticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

5

6

Turbulent mean ﬂow 6.1 Time averaged Navier-Stokes . . . . . . . . . . . 6.1.1 Boundary-layer approximation . . . . . . 6.2 Wall region in fully developed channel ﬂow . . . 6.3 Reynolds stresses in fully developed channel ﬂow 6.4 Boundary layer . . . . . . . . . . . . . . . . . . . Probability density functions

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

7

2

3

8

Transport equations for kinetic energy 8.1 The Exact k Equation . . . . . . . . . . . . 8.2 The Exact k Equation: 2D Boundary Layers 8.3 Spatial vs. spectral energy transfer . . . . . 8.4 The overall effect of the transport terms . . . 8.5 The transport equation for vi vi /2 . . . . . . ¯¯

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

. . . . .

62 62 65 67 67 68 71 75 77 77 79 81 81 81 82 82 83 85 86 86 87 88 88 90 90 93 99 101 102 102 103 104 106 106 107 110 111 112 113 115 118 121 122

9

Transport equations for Reynolds stresses 9.1 Reynolds shear stress vs. the velocity gradient . . . . . . . . . . . .

10 Correlations 10.1 Two-point correlations . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Auto correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Reynolds stress models and two-equation models 11.1 Mean ﬂow equations . . . . . . . . . . . . . . 11.1.1 Flow equations . . . . . . . . . . . . 11.1.2 Temperature equation . . . . . . . . . ′ ′ 11.2 The exact vi vj equation . . . . . . . . . . . . ′ 11.3 The exact vi θ′ equation . . . . . . . . . . . . 11.4 The k equation . . . . . . . . . . . . . . . . . 11.5 The ε equation . . . . . . . . . . . . . . . . . 11.6 The Boussinesq assumption . . . . . . . . . . 11.7 Modelling assumptions . . . . . . . . . . . . 11.7.1 Production terms . . . . . . . . . . . 11.7.2 Diffusion terms . . . . . . . . . . . . 11.7.3 Dissipation term, εij . . . . . . . . . 11.7.4 Slow pressure-strain term . . . . . . . 11.7.5 Rapid pressure-strain term . . . . . . 11.7.6 Wall model of the pressure-strain term 11.8 The k − ε model . . . . . . . . . . . . . . . . ′ ′ 11.9 The modelled vi vj equation with IP model . . 11.10 Algebraic Reynolds Stress Model (ASM) . . . 11.11 Explicit ASM (EASM or EARSM) . . . . . . 11.12 Boundary layer ﬂow . . . . . . . . . . . . . . 12 Reynolds stress models vs. eddy-viscosity models 12.1 Stable and unstable stratiﬁcation . . . . . . 12.2 Curvature effects . . . . . . . . . . . . . . . 12.3 Stagnation ﬂow . . . . . . . . . . . . . . . 12.4 RSM/ASM versus k − ε models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

13 Realizability 13.1 Two-component limit . . . . . . . . . . . . . . . . . . . . . . . . . 14 Non-linear Eddy-viscosity Models 15 The V2F Model 15.1 Modiﬁed V2F model . . . . . . . . . . . . . . . . . . . . . . . . . . 15.2 Realizable V2F model . . . . . . . . . . . . . . . . . . . . . . . . .

4 To ensure that v 2 ≤ 2k/3 [1] . . . . . . . . . . . . . . . . . . . . .

15.3

122 123 127 127 128 129 130 131 131 132 132 133 133 134 135 136 137 138 139 139 140 141 142 143 144 144 145 146 146 146 148 151 151 153 154 155 156 160 160 160 163 163 163 165

16 The SST Model 17 Large Eddy Simulations 17.1 Time averaging and ﬁltering . . . . . . . . . . . . . . . . . . . . . . 17.2 Differences between time-averaging (RANS) and space ﬁltering (LES) 17.3 Resolved & SGS scales . . . . . . . . . . . . . . . . . . . . . . . . 17.4 The box-ﬁlter and the cut-off ﬁlter . . . . . . . . . . . . . . . . . . 17.5 Highest resolved wavenumbers . . . . . . . . . . . . . . . . . . . . 17.6 Subgrid model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.7 Smagorinsky model vs. mixing-length model . . . . . . . . . . . . . 17.8 Energy path . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.9 SGS kinetic energy . . . . . . . . . . . . . . . . . . . . . . . . . . 17.10 LES vs. RANS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.11 The dynamic model . . . . . . . . . . . . . . . . . . . . . . . . . . 17.12 The test ﬁlter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.13 Stresses on grid, test and intermediate level . . . . . . . . . . . . . . 17.14 Numerical dissipation . . . . . . . . . . . . . . . . . . . . . . . . . 17.15 Scale-similarity Models . . . . . . . . . . . . . . . . . . . . . . . . 17.16 The Bardina Model . . . . . . . . . . . . . . . . . . . . . . . . . . 17.17 Redeﬁned terms in the Bardina Model . . . . . . . . . . . . . . . . 17.18 A dissipative scale-similarity model. . . . . . . . . . . . . . . . . . 17.19 Forcing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.20 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.20.1 RANS vs. LES . . . . . . . . . . . . . . . . . . . . . . . . 17.21 One-equation ksgs model . . . . . . . . . . . . . . . . . . . . . . . 17.22 Smagorinsky model derived from the ksgs equation . . . . . . . . . 17.23 A dynamic one-equation model . . . . . . . . . . . . . . . . . . . . 17.24 A Mixed Model Based on a One-Eq. Model . . . . . . . . . . . . . 17.25 Applied LES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17.26 Resolution requirements . . . . . . . . . . . . . . . . . . . . . . . . 18 Unsteady RANS 18.1 Turbulence Modelling . . . . . . . . . . . . . . . . . . . . . . . . . 18.2 Discretization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 DES 19.1 DES based on two-equation models . . . . . . . . . . . . . . . . . . 19.2 DES based on the k − ω SST model . . . . . . . . . . . . . . . . . 20 Hybrid LES-RANS 20.1 Momentum equations in hybrid LES-RANS . . . . . . . . . . . . . 20.2 The equation for turbulent kinetic energy in hybrid LES-RANS . . . 20.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 The SAS model 21.1 Resolved motions in unsteady . . . . . . . . . . . . . . . . . . . . 21.2 The von K´ rm´ n length scale . . . . . . . . . . . . . . . . . . . . . a a 21.3 The second derivative of the velocity . . . . . . . . . . . . . . . . .

5

21.4

Evaluation of the von K´ rm´ n length scale in channel ﬂow . . . . . a a

165 168

22 The PANS Model 23 Hybrid LES/RANS for Dummies 23.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 23.1.1 Reynolds-Averaging Navier-Stokes equations: RANS 23.1.2 Large Eddy Simulations: LES . . . . . . . . . . . . 23.1.3 Zonal LES/RANS . . . . . . . . . . . . . . . . . . . 23.2 The PANS k − ε turbulence model . . . . . . . . . . . . . . 23.3 Zonal LES/RANS: wall modeling . . . . . . . . . . . . . . . 23.3.1 The interface conditions . . . . . . . . . . . . . . . . 23.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 23.4 Zonal LES/RANS: embedded LES . . . . . . . . . . . . . . 23.4.1 The interface conditions . . . . . . . . . . . . . . . . 23.4.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . 24 Inlet boundary conditions 24.1 Synthesized turbulence . . . . 24.2 Random angles . . . . . . . . . 24.3 Highest wave number . . . . . 24.4 Smallest wave number . . . . . 24.5 Divide the wave number range 24.6 von K´ rm´ n spectrum . . . . . a a 24.7 Computing the ﬂuctuations . . 24.8 Introducing time correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

172 172 172 172 173 173 174 174 175 175 175 177 178 178 178 179 179 179 180 180 180 183 183 183 184 184 185

25 Best practice guidelines (BPG) 25.1 EU projects . . . . . . . . . . . . . . . . 25.2 Ercoftac workshops . . . . . . . . . . . . 25.3 Ercoftac Classical Database . . . . . . . . 25.4 ERCOFTAC QNET Knowledge Base Wiki A TME225: ε − δ identity

B TME225 Assignment 1: laminar ﬂow B.1 Fully developed region . . . . . . . . . . . . B.2 Wall shear stress . . . . . . . . . . . . . . . . B.3 Inlet region . . . . . . . . . . . . . . . . . . . B.4 Wall-normal velocity in the developing region B.5 Vorticity . . . . . . . . . . . . . . . . . . . . B.6 Deformation . . . . . . . . . . . . . . . . . . B.7 Dissipation . . . . . . . . . . . . . . . . . . . B.8 Eigenvalues . . . . . . . . . . . . . . . . . . B.9 Eigenvectors . . . . . . . . . . . . . . . . . . C TME225: Fourier series C.1 Orthogonal functions . . . . . . C.2 Trigonometric functions . . . . . C.3 Fourier series of a function . . . C.4 Derivation of Parseval’s formula

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

186 187 187 187 188 188 188 188 188 189 190 190 191 193 193

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . . . . . . . . . . . . . . . . . . .3 Compute derivatives on a curvi-linear mesh . D. . . . . . . . . . . K. . .2 Time averaging . . . . . . . . . E. .11 Do something fun! . . . . . L. . . . . . . . . . . . . . . . . . . . . . . . . . . .7 Fluctuating wall shear stress . .2. . . . . . . .3 . . . . . . . . . . . . L. . . . . . MTF270: 1D and 3D energy spectra J. . . . . . . . . . . . . . . .2 An example of using FFT . . . . . . .6 C. . . . . . . L. . . K. . . . . . . . . .9 Pressure-strain terms . E. . . . . . . . . . . . . . . J K MTF270. . . K. . .1 Geometrical quantities . . .5 Wall shear stress . .10 Dissipation . . . . . D. . . . .1 Introduction . . . .1 The wavenumber vector. L. . . . . .1 The momentum equations . . . . . .6 . . . . . . . . . . . . . . L. K. . . . . . . . F TME225 Learning outcomes G MTF270: Some properties of the pressure-strain term H MTF270: Galilean invariance I . L. . . . . E. . . E. . . Assignment 2: L. . . . . . LES . . . . . . . .3. . . . . . . .8 Production terms . . . . . . . . . . . . . . E. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Task 2. . . . . . . . . . . . . . . .6 Resolved stresses . . . . . . . . . . . . . . . . . . . . . . . . . . K. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . L MTF270. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Task 2. . . . . . κn .7 .2 Unit vector σi . . . . . . . .5 Task 2. . . . . . . .5 Complex Fourier series . . . . . . . . . . . . . . . . . . . . . . . . .2 . . . . . . .2. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . E TME225 Assignment 2: turbulent ﬂow E. . . . . . . . . . . . . . . . . . . . . . .3 The Reynolds stress equations . .4 The time-averaged momentum equation E. . MTF270: Computation of wavenumber vector and angles I. E. .4 Task 2. . . . . . . . . . . . . . . . .1 Two-dimensional ﬂow . . Assignment 1: Reynolds averaged Navier-Stokes K.1 Energy spectra from two-point correlations . . . . . . . . . . . . . . .3 Energy spectrum from the two-point correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Energy spectra from the autocorrelation . . . . . . . . . . . . . . . . . .2 The turbulent kinetic energy equation . E. . . . . . . .6 Task 2. . . . . . . . . . .5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . K. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . D. E. .3 Mean ﬂow . . . . . j n I. . . . . .2. . . . . . . . . . . . .4 . . . . . . . . . 195 196 196 196 198 200 201 201 201 202 202 203 203 203 203 203 204 204 205 216 217 219 219 220 221 222 224 224 224 225 226 226 228 229 230 230 232 232 233 234 234 235 D TME225: Compute energy spectra from LES/DNS data using Matlab D. . .1 . .3 Task 2. . . . . .1 Time history . . . . . . . . . . .2 Analysis . . . . . . . . . . . . . . . . . . . . . . . .7 Task 2. . . E. . . . . . . . . . .

. . . . .2 Mean velocity proﬁle . . .2 Green’s second formula . . . . . . . . . . . . . . . . M. . . . . . . . . .4 Turbulent kinetic energy . . . . . . .3 Resolved stresses . . . . . . . . M. . . . . .7 L. Assignment 5: Embedded LES with PANS N. O. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Assignment 4: Hybrid LES-RANS M. . . . . . . . . . . .11 . . . . . . . . . . . P. . . . . . . . . P. . .3 Turbulent viscosity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .8 Task 2. . . . . . . . . . . . . . . . . .1 Green’s ﬁrst formula . . . . . . . . . . . . . . . .10 Task 2.2 Transformation of a velocity gradient . . . . . .4 Analytical solution to Poisson’s equation Q MTF270: Learning outcomes for 2012 R References . . . . . . . . . . . . . . . . . . . . . . . . . .1 Time history . . N MTF270. . . . . . . . M. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Green’s third formula . . . . . . . . . . . . . . N. . . . . . 236 236 236 237 238 238 239 239 240 240 240 240 242 242 243 243 244 244 246 247 248 249 249 249 249 252 253 260 M MTF270. . . . . . . . . . . . . . . . . . . . .7 SAS turbulent length scales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N. . . . . . . . . . . . . .2 Resolved stresses . .4 Modelled stresses . . . . .10 L. . . . . . . M. . . . .11 Task 2. . . . . . . . . . . . . . .5 The modelled turbulent shear stress . . . . N. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . N. . . . . . . . . . . . . . . O MTF270: Transformation of a tensor O. . . . . . . . P. . . P MTF270: Green’s formulas P. . .8 L. . . . . . . . . . . . . .9 Task 2. . . .1 Time history . . . . M. . . . . . . . . . . . . . . . . . . .9 L. . . . . . .5 Turbulent SGS dissipation . . . . . . . . . . . . . . . . . . . . . . .6 Turbulent length scales . . .1 Rotation to principal directions . . . . . . . . . . . . M.

tfd.se/˜lada.8 TME225 Mechanics of ﬂuids L. Department of Applied Mechanics Chalmers University of Technology. Sweden o http://www.se This report can be downloaded at http://www. lada@chalmers. G¨ teborg.tfd.chalmers.se/˜lada/MoF/ . Davidson Division of Fluid Dynamics.chalmers.

T = T (xi ). dT /dt. see Fig. The temperature of the ﬂuid.1: The temperature of a ﬂuid particle described in Lagrangian. t).1. In the Lagrangian approach we keep track of its original position (Xi ) and follow its path which is described by xi (Xi . Chapt.1) Note that we have to use partial derivative on T since it is a function of more than one (independent) variable. In the Lagrangian approach we ﬁrst pick the particle (this gives its starting position. Assume a ﬂuid particle is moving along the line in Fig. is expressed as a function of the position. The speed of the particle is then expressed as a function of time and its position at time zero. T (t) and the temperature gradient can be written dT /dt.e. for example. t2 ). vi = dxi /dt.e. and temperature varies only with time.1. t2 ) i Figure 1. at time t1 the temperature of the particle is T (Xi .g. The second term on the right side is called the convective rate of change. It is however used when simulating movement of particles in ﬂuids (for example soot particles in gasoline-air mixtures in combustion applications).1 Eulerian. We can choose to study its motion in two ways: Lagrangian or Eulerian. and at time t2 its temperature is T (Xi . This approach is not used for ﬂuids because it is very tricky to deﬁne and follow a ﬂuid particle. Motion. It may be that the temperature at position xi . ﬂow T (Xi .2. t1 ) i T (Xi . we know that the velocity of a ﬂuid particle is the time derivative of its space location. see Fig. vi = vi (Xi . t2 ) 9 T (x2 .e. From classical mechanics. i. The chain-rule now gives ∂T ∂T ∂T dxj ∂T dT = = + + vj dt ∂t dt ∂xj ∂t ∂xj (1. e. t). i. ﬂow 1. Once we have chosen a particle its starting position is ﬁxed. by this we mean that it describes the variation of T in time at position xi . For example. In the Eulerian approach we pick a position. T (Xi . but since we are looking at ﬁxed points in space we need to express the temperature as a function of both time and space. 1. This approach is used for ﬂuids. material derivative See also [2]. The ﬁrst term on the right side is the local rate of change. We are looking for the temperature gradient. t. T (xi . i. 1. In the Eulerian approach it is a little bit more difﬁcult. and watch the particle pass i by. Now we want to express how the temperature of a ﬂuid particle varies. 1. x1 . 1 Motion. T . approach. t1 ) Xi T (x1 . t). varies in time. Xi ). Lagrangian. t). t). or Eulerian.1. which means that it describes local rate of change Conv.1. 3. rate of change . and then T = T (xi .e. for example. t1 ). i.

3) where the ﬁrst and the second term on the right side represents. Stress is force per unit area. 6. dT /dt. σ12 and σ13 . Chapts. Imagine that you’re a ﬂuid particle and that you ride on a bike. v. on a surface whose normal is ni = (1. The ﬁrst term includes the viscous stress tensor. p = −σkk /3 (1. 0) act the three stress components σ11 .4) .e.2 Viscous stress. For example. Material derivative 1. the net force due to surface and volume forces (σij denotes the stress tensor). As you have learnt earlier. v3 ) and the coordinate by x = xi = (x1 . u3 ). The left side in Eq.3 and 8.1. Viscous stress.1. x2 . u2 . 0. y. We have in Part I [3] derived the balance equation for linear momentum which reads ρvi − σji.2: Deﬁnition of stress components on a surface. pressure 10 x2 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 x1 σ12 σ11 σ13 Figure 1.1 can be illustrated as follows. x3 ). respectively. The diagonal components of σij represent normal stresses and the off-diagonal components of σij represent the shear stresses. the ﬁrst index relates to the surface at which the stress acts and the second index is related to the stress component.2.1 is called the material derivative and is in this text denoted by dT /dt. the variation of T in space when is passes the point xi . 1. In Part I [3] you learnt that the pressure is deﬁned as minus the sum of the normal stress. In the literature.2. 1. Equation 1.1. i. v2 . pressure See also [2]. Put your ﬁnger out in the blowing wind. The temperature gradient you experience is the material derivative. see Fig. In the present notation we denote the velocity vector by v = vi = (v1 . you may ﬁnd other notations of the velocity vector such as ui = (u1 . term-by-term. z).2) Switch notation for the material derivative and derivatives so that ρ dvi ∂σji + ρfi = dt ∂xj (1. w) and the coordinates as (x. If no tensor notation is used the velocity vector is usually denoted as (u. 1.j − ρfi = 0 ˙ (1. Exercise 1 Write out Eq. τij . The temperature gradient you’re ﬁnger experiences is ∂T /∂t.

e. for example. The minus-sign in front of p appears because the pressure acts into the surface. the viscous stresses are zero and then of course the normal stresses are the same as the pressure. p. vorticity See also [2]. In that case the thermodynamics pressure. σ32 . The expression for the viscous stress tensor is found in Eq. which can be obtained – for example – from the ideal gas law. p. from σij . i. Chapt. 1. vorticity 11 The pressure.8) Strain-rate tensor vorticity tensor The vorticity represents rotation of a ﬂuid particle. 1. The viscous stress tensor.5. 3. may not be the same but Eq. acts as a normal stress. 1.2.5) τij is the deviator of σij .6) Exercise 2 Consider Fig. 0). 3. 1).4 is nevertheless used. is obtained by subtracting the trace. pt .3 Strain rate tensor. Inserting Eq. 1. pressure is a thermodynamic property.8 gives ωi = ǫijk (Skj + Ωkj ) = ǫijk Ωkj (1. Show also how σ31 . Show how σ21 . ω = ∇ × v. the stress tensor can then be written as σij = −pδij + τij (1.9) ∂vk ∂xj (1. In general. σ33 = −p + τ33 . and the mechanical pressure.3. The velocity gradient tensor can be split into two parts as ∂vi 1 ∂vi ∂vi ∂vj ∂vj = ∂xj + ∂xj + ∂xi − ∂xi ∂xj 2 2∂vi /∂xj =0 (1. 0. or in tensor notation ωi = ǫijk If we set. i. σ33 act on a surface with normal vector ni = (0. 20.e.10) . 1.5 on matrix form.6. however. τij . σ23 act on a surface with normal vector ni = (0. σkk /3 = −p. σ22 . pt . 1. (1. Strain rate tensor. (1.3.7 into Eq. In general.1. 1.7) = Sij + Ωij = where 1 2 ∂vi ∂vj + ∂xj ∂xi + 1 2 ∂vi ∂vj − ∂xj ∂xi Sij is a symmetric tensor called the strain-rate tensor Ωij is a anti-symmetric tensor called the vorticity tensor The vorticity tensor is related to the familiar vorticity vector which is the curl of the velocity vector. σ22 = −p + τ22 .4 at p. Exercise 3 Write out Eq. 2. i = 3 we get ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 . When there’s no movement. the normal stresses are the sum of the pressure and the viscous stresses. σ11 = −p + τ11 .

10. see Table A.15) 2 A much easier way to go from Eq. vorticity 12 since ǫijk Skj = 0 because the product of a symmetric tensor (Skj ) and a anti-symmetric tensor (εijk ) is zero. 1.8 (i. 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 + 0 · S12 + 0 · S22 + 1 · S32 + 0 · S13 − 1 · S23 + 0 · S33 = S32 − S23 = 0 (1. Show that you get the same result as in Eq. .16 also for i = 2 and i = 3 and ﬁnd an expression for Ω12 and Ω13 (cf.11) Now les us invert Eq. Exercise 7 In Eq. Hence we get with Eq. ε113 = ε221 = . .8 Ωmℓ = or. What about the diagonal components of Ωij ? What do you get from Eq. 1. S13 = S31 .1) εiℓm ǫijk Ωkj = (δℓj δmk − δℓk δmj )Ωkj = Ωmℓ − Ωℓm = 2Ωmℓ (1. ε123 = −ε132 = ε231 . 1.11 for i = 2 and i = 3.12) This can easily be proved by writing all the components. 1.7? . A.15 is to write out the components of Eq.1. A. switching indices 1 1 1 εiℓm ωi = εℓmi ωi = − εmℓi ωi 2 2 2 (1.10 to Eq.15. Eq. We start by multiplying it with εiℓm so that εiℓm ωi = εiℓm ǫijk Ωkj The ε-δ-identity gives (see Table A. 1. Strain rate tensor. and we get (1. Let us show this for i = 1 by writing out the full equation. 1.1 at p. S23 = S32 ) and ǫijk = −ǫikj = ǫjki etc (i. 1. 1.17 we proved the relation between Ωij and ωi for the off-diagonal components. Exercise 5 Complete the proof of Eq. .14) 1 Ωij = − εijk ωk (1. 1.3. ω2 and ω3 ). S12 = S21 .10.1. 1. Recall that Sij = Sji (i. 1.15. Here we do it for i = 1 ω1 = ε123 Ω32 + ε132 Ω23 = Ω32 − Ω23 = −2Ω23 1 Ω23 = − ω1 2 which indeed is identical to Eq. .17). Exercise 6 Write out Eq.1 at p. .e.e.16) (1.13) (1.e. 1.17) Exercise 4 Write out the second and third component of the vorticity vector given in Eq. 1.

and then that bij = −bji (antisymmetric).8. 1. First. 1. The deformation due to shear is caused by the off-diagonal terms of Sij . . Chapt. for example. 1. Elongation caused by S11 = ∂v1 /∂x1 is illustrated in Fig. a ﬂuid particle experiences a combination of rotation. a pure shear deformation by S12 = (∂v1 /∂x2 + ∂v2 /∂x1 )/2 is shown. + a32 b32 + a33 b33 = 0 1. i. The (inner) product of a symmetric and antisymmetric tensor is always zero. It follows that for an antisymmetric tensor all diagonal components must be zero. In general. This can of course also be shown be writing out aij bij on component form.5. The vorticity ω3 should be interpreted as twice the average rotation of the horizontal edge (∂v2 /∂x1 ) and vertical edge (−∂v1 /∂x2 ). 1.4 Product of a symmetric and antisymmetric tensor In this ssction we show the proof that the product of a symmetric and antisymmetric tensor is zero. Exercise 8 From you course in linear algebra. This movement can be illustrated by Fig. It is assumed that the ﬂuid particle is rotated the angle α during the time ∆t.3.3. Use it to compute the vector product ˆ1 e ˆ2 e ˆ3 e ∂ ∂ ∂ ω = ∇ × v = ∂x1 ∂x2 ∂x3 v1 v2 v3 1. It can be divided into two parts.4.7. so that aij bij = −aji bij = −aij bij . We have shown that the latter is connected to rotation of a ﬂuid particle.e. you should remember how to compute a vector product using Sarrus’ rule. 3.1. rotation . and thus the product must be zero. namely shear and elongation (also called extension or dilatation). The vorticity during this rotation is ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 = −2Ω12 . 1. rotation See also [2]. In Fig. During rotation the ﬂuid particle is not deformed. The velocity gradient can. b11 = −b11 can only be satisﬁed if b11 = 0.5 Deformation. Product of a symmetric and antisymmetric tensor 13 Verify that this agrees with the expression in tensor notation in Eq.4. aij bij = a11 b11 + a12 b12 + a13 b13 + . be divided into two parts: Sij and Ωij . • A tensor bij is antisymmetric if bij = −bji . This can be shown as follows aij bij = aji bij = −aij bji . . Since the indices i and j are both dummy indices we can interchange them. deformation and elongation as indeed is given by Eq. where we ﬁrst used the fact that aij = aji (symmetric). as shown above. The deformation due to elongation is caused by the diagonal terms of Sij . we have the deﬁntions: • A tensor aij is symmetric if aij = aji . Next let us have a look at the deformation caused by Sij .

Here ∂v1 /∂x2 = −∂v2 /∂x1 so that −Ω12 = ω3 /2 = ∂v2 /∂x1 > 0. . Exercise 9 Consider Fig. ∆x1 ∂v1 ∆x2 ∆t ∂x2 α ∆x2 ∂v2 ∆x1 ∆t ∂x1 α x2 x1 Figure 1. Deformation. 1.3. 1. Show and formulate the deformation by S23 .3: Rotation of a ﬂuid particle during time ∆t.4. Exercise 10 Consider Fig.4: Deformation of a ﬂuid particle by shear during time ∆t. rotation ∆x1 ∂v1 ∆x2 ∆t ∂x2 14 α ∆x2 ∂v2 ∆x1 ∆t ∂x1 α x2 x1 Figure 1.5.1. Here ∂v1 /∂x2 = ∂v2 /∂x1 so that S12 = ∂v1 /∂x2 > 0. Show and formulate the rotation by ω1 .

In potential ﬂow. 1. ti dℓ x2 S x1 Figure 1.1. S. from which the velocity components can be obtained as ∂Φ (1. there exists a potential. The vector.5.6 Irrotational and rotational ﬂow In the previous subsection we introduced different types of movement of a ﬂuid particle. We’ll talk more about that later on.18) vk = ∂xk . 1.6: The surface. the latter type is also called inviscid ﬂow or potential ﬂow.3. Show and formulate the elongation by S22 . 1. One type of movement was rotation. see Fig. denotes the unit tangential vector of the enclosing line. Φ.5: Deformation of a ﬂuid particle by elongation during time ∆t. Flows are often classiﬁed based on rotation: they are rotational (ωi = 0) or irrotational (ωi = 0). Exercise 11 Consider Fig.6. Irrotational and rotational ﬂow 15 ∆x1 ∂v1 ∆x1 ∆t ∂x1 ∆x2 x2 x1 Figure 1. is enclosing by the line ℓ. ti . ℓ. In this subsection we will give examples of one irrotational and one rotational ﬂow.

6. an inviscid simulation method (based on the circulation and vorticity sources) was used to compute the aerodynamic loads for windturbines [4].20 reads in vector notation Γ= ℓ v · tdℓ = S (∇ × v) · ndS = ω3 dS S (1. see Fig.7.6. Its potential reads Φ= The velocity. The Cartesian velocity vectors are expressed as v1 = −vθ sin(θ) = −vθ x2 x2 = −vθ 2 r (x1 + x2 )1/2 2 x1 x1 = vθ 2 v2 = vθ cos(θ) = vθ r (x1 + x2 )1/2 2 (1. vθ .1 Ideal vortex line The ideal vortex line is an irrotational (potential) ﬂow where the ﬂuid moves along circular paths. When the velocity is integrated along this line and projected onto the line we obtain the circulation Γ = vm tm dℓ (1.23 into Cartesian velocity components. 1.19) Using Stokes’s theorem we can relate the circulation to the vorticity as Γ= ℓ vm tm dℓ = S εijk ∂vk ni dS = ∂xj ω3 dS S (1. 1. 1. 1. consider Fig. we need to introduce the concept circulation.6. The lift force is computed as L = ρV Γ (1. i. Multiply Eq. see Fig.21) The circulation is useful in aeronautics and windpower engineering where the lift of an airfoil or a rotorblade is expressed in the circulation for a 2D section.23) where Γ is the circulation.8.20) where ni = (0. 1.24) To transform Eq. Equation 1. Exercise 12 In potential ﬂow ωi = εijk ∂vk /∂xj = 0. In a recent MSc thesis project. εijk ∂ 2 Φ/(∂xk ∂xj ) = 0.22) where V is the velocity around the airfoil (for a rotorblade it is the relative velocity.25) . 2πr vr = 0 (1.e.18 by εijk and derivate with respect to xk (i.e. 1) is the unit normal vector of the surface S. 0. since the rotorblade is rotating). is then obtained as vθ = 1 ∂Φ r ∂θ Γθ 2π (1. Irrotational and rotational ﬂow 16 Before we talk about the ideal vortex line in the next section.1. The velocity ﬁeld in polar coordinates reads vθ = Γ . Consider a closed line on a surface in the x1 − x2 plane. take the curl of) and show that the right side becomes zero as it should. 1.

That the ﬂow has no vorticity (i.e. The ﬂuid particle (i. 2π(x2 + x2 ) 1 2 v2 = Γx1 . 2π(x2 + x2 ) 1 2 (1. This is a singular point as is seen from Eq. The velocity derivatives are obtained as Γ x2 − x2 ∂v1 1 2 =− . Since it is a two-dimensional ﬂow (v3 = ∂/∂x3 = 0).7.e. vorticity . 1. we need to show that the vorticity. we only need to compute ω3 = ∂v2 /∂x1 − ∂v1 /∂x2 .6. Note that the deformation is not zero.29) Hence a ﬂuid particle in an ideal vortex does deform but it does not rotate (i. Note that generally a vortex has vorticity. 1. its diagonal does not rotate. 2 v2 = 0 (1. Inserting Eq. S12 = 1 2 ∂v1 ∂v2 + ∂x2 ∂x1 = x2 Γ 2 2π (x2 + x2 )2 1 2 (1. 1. The ideal vortex is a very special ﬂow case. Irrotational and rotational ﬂow a 17 b Figure 1. 1. 1. its diagonal. irrotational (ωi ≡ 0). see Fig. ωi = εijk ∂vk /∂xj is zero.e. see Section 4. This is true for the whole ﬂow ﬁeld. i. except at the center where the ﬂuid particle does rotate.7). ∂x2 2π (x2 + x2 )2 1 2 and we get ω3 = 1 Γ (x2 − x2 + x2 − x2 ) = 0 1 1 2 2π (x2 + x2 )2 2 1 2 (1.2 Shear ﬂow Another example – which is rotational – is a shear ﬂow in which v1 = cx2 .26) To verify that this ﬂow is a potential ﬂow.e.2. i.3) does not rotate.25 into Eq. ω1 = ω2 = 0. Thus one must be very careful when using the words “vortex” and ”vorticity”. 1.28) Γ x2 − x2 ∂v2 2 1 = ∂x1 2π (x2 + x2 )2 1 2 (1. 1.e.23 we get v1 = − Γx2 .1.30) vortex vs. It may be little confusing that the ﬂow path forms a vortex but the ﬂow itself has no vorticity. By vortex we usually mean a recirculation region of the mean ﬂow. see Fig.27) which shows that the ﬂow is indeed a potential ﬂow.23 for which ω3 → ∞.6. no rotation) means that a ﬂuid particle moves as illustrated in Fig. As a ﬂuid particle moves from position a to b – on its counter-clockwise-rotating path – the particle itself is not rotating.7: Ideal vortex.

112. not imaginary). This is why the vorticity.5 or Eq. the eigenvalues are real (i. via b to position c it is indeed rotating. In the left ﬁgure it is oriented along the x1 − x2 coordinate system. 13. σij .1. x2 > 0.9.5 at p.5. The ﬂuid particle rotates. The vorticity vector for this ﬂow reads ω1 = ω2 = 0. Note that the positive rotating direction is deﬁned as the counter-clockwise direction. the eigenvectors can be computed. The eigenvalues are obtained from the characteristic equation.8: Transformation of vθ into Cartesian components.5. the ﬂuid element is rotated α degrees so that its edges are aligned with ˆ ˆ the eigenvectors.9. 1. Eigenvalues and and eigenvectors: physical interpretation vθ x2 r θ x1 18 θ Figure 1. Since σij is symmetric.10.7 Eigenvalues and and eigenvectors: physical interpretation See also [2]. Chapt. 2 with c. Note that the the ˆ ˆ . It is rotating in clockwise direction. see right part of Fig. is negative (= −2cx2 ).5. ω3 .31) When the ﬂuid particle is moving from position a.7.9: A shear ﬂow. v1 a b v1 c x2 a x3 x1 Figure 1.e. v1 = x1′ and v2 = x2′ . 1. Any tensor has eigenvectors and eigenvalues (also called principal vectors and principal values). On the surfaces act normal stresses (σ11 . σ22 ) and shear stresses (σ12 . ω3 = ∂v2 ∂v1 − = −2cx2 ∂x1 ∂x2 (1. indicated by α in Fig. Given the eigenvectors. see Fig. 1. 1. v1 = cx2 . see Fig. 2. σ21 ). 2. see [2].10. 1. Consider a two-dimensional ﬂuid (or solid) element. When the eigenvalues have been obtained. The stresses form a tensor. Chapt.

10). λ1 = σ1′ 1′ and λ2 = σ2′ 2′ . v1 and v2 denote unit ˆ ˆ eigenvectors. right: rotated to principal coordinate directions. There are only normal stresses. Eigenvalues and and eigenvectors: physical interpretation σ23 σ21 σ12 x2 x1 α σ11 v1 ˆ λ2 λ1 x2′ x1′ 19 v2 ˆ Figure 1. the eigenvalues are the normal stresses in the principal coordinates.e. which means that the eigenvectors can equally well be chosen as −ˆ1 and/or −ˆ2 . λ1 and λ2 denote eigenvalues. i. . This is the very deﬁnition of eigenvectors. In the principal coordinates x1′ − x2′ (right part v v of Fig.7.10: A two-dimensional ﬂuid element. Furthermore.PSfrag 1. sign of the eigenvectors is not deﬁned. there are no shear stresses on the surfaces of the ﬂuid element. Left: in original state. 1.

i = 0 ˙ (2. 2.1.5 can now – for constant µ and incompressible ﬂow – be written ρ ∂ 2 vi ∂p dvi +µ + ρfi =− dt ∂xi ∂xj ∂xj (2.7) In inviscid (potential) ﬂow. 5 and 8. µ. It is also called the transport equation for momentum. if the ﬂow is incompressible the second term in the parenthesis on the right side is zero because of the continuity equation. Chapts. Governing ﬂow equations 20 2 Governing ﬂow equations See also [2].4) σij = −pδij + 2µSij − µSkk δij 3 Inserting Eq. .2 The momentum equation The next equation is the momentum equation. In this case.5) (2.6) because of the continuity equation. 1.8) Euler equations Exercise 13 Formulate the Navier-Stokes equation for incompressible ﬂow but nonconstant viscosity. This is the Navier-Stokes equations (sometimes the continuity equation is also included in the name “Navier-Stokes”).1 The Navier-Stokes equation 2.1) Change of notation gives dρ ∂vi =0 +ρ dt ∂xi For incompressible ﬂow (ρ = const) we get ∂vi =0 ∂xi 2. 2. We have formulated the constitutive law for Newtonian viscous ﬂuids [3] 2 (2. the Navier-Stokes equation reduces to the Euler equations ρ ∂p dvi + ρfi =− dt ∂xi (2.1. there are no viscous (friction) forces.3.2) (2. Eq.2.1 The continuity equation The ﬁrst equation is the continuity equation (the balance equation for mass) which reads [3] ρ + ρvi.4 into the balance equations. Equation 2. Furthermore. If these two requirements are satisﬁed we can also re-write the ﬁrst term in the parenthesis as ∂ ∂ (2µSij ) = µ ∂xj ∂xj ∂vj ∂vi + ∂xj ∂xi =µ ∂ 2 vi ∂xj ∂xj (2. is constant it can be moved outside the derivative. If the viscosity.3) where µ denotes the dynamic viscosity. we get ρ ∂τji ∂p ∂ ∂p dvi + + ρfi = − + =− dt ∂xi ∂xj ∂xi ∂xj 2 ∂vk 2µSij − µ δij 3 ∂xk + ρfi (2.1.

e.e.12 is the transport equation for (internal) energy. transformation of kinetic energy into thermal energy).2. these terms are important at high-speed ﬂow1 (for example re-entry from outer space) and for highly viscous ﬂows (lubricants). Ωij .10 gives ρ du 2 ∂vi ∂ + 2µSij Sij − µSkk Sii + = −p dt ∂xi 3 ∂xi Φ k ∂T ∂xi (2.e.j σji + qi. is zero. Chapts.12) where we have used Sij ∂vi /∂xj = Sij (Sij + Ωij ) = Sij Sij because the product of a symmetric tensor. 6. (2. 1 High-speed ﬂows relevant for aeronautics will be treated in detail in the course “Compressible ﬂow” in the MSc programme.2 The energy equation See also [2]. the velocity should be smaller than approximately 1/3 of the speed of sound) for which du = cp dT (2. u. for simplicity.11) qi = −k ∂xi Inserting the constitutive laws.1.15) thermal diffusivity . into Eq. we neglect the radiation from here on.4 and 2. qi denotes the conductive heat ﬂux and z the net radiative heat source. not an lubricant oil).9) where u denotes internal energy.i = ρz ˙ (2. Equation 2.10) In Part I [3] we formulated the constitutive law for the heat ﬂux vector (Fourier’s law) ∂T (2. We have in Part I [3] derived the energy equation which reads ρu − vi. Sij . Now we assume that the ﬂow is incompressible (i.13) where cp is the heat capacity (see Part I) [3] so that Eq.rad . 2. Change of notation gives ρ du ∂vi ∂qi − = σji dt ∂xj ∂xi (2. and an anti-symmetric tensor. zi.11. The ﬁrst term on the right side represents reversible heating and cooling due to compression and expansion of the ﬂuid. we get ∂2T dT =α dt ∂xi ∂xi where α = k/(ρcp ) is the thermal diffusivity.2. The energy equation 21 2. The latter can also be seen as a vector.14) =Φ+ ρcp dt ∂xi ∂xi The dissipation term is simpliﬁed to Φ = 2µSij Sij because Sii = ∂vi /∂xi = 0. Eqs. 2. 2. If we furthermore assume that the heat conductivity coefﬁcient is constant and that the ﬂuid is a gas or a common liquid (i.12 gives (cp is assumed to be constant) ∂T ∂ dT k (2. Two of the viscous terms (denoted by Φ) represent irreversible viscous heating (i.4 and 8.

20 to Eq. Let us take a closer look at Eqs. The ﬂuid is a gas with low velocity (lower than 1/3 of the speed of sound). 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.3.3 with vi ∂σji dvi − vi ρfi = 0 (2.18) (2. 2.16) α is deﬁned where ν = µ/ρ is the kinematic viscosity. The dissipation term.3 Transformation of energy Now we will derive the equation for the kinetic energy.12. i. k = vi vi /2.17) − vi ρvi dt ∂xj The ﬁrst term on the left side can be re-written ρvi (vi vi /2 = k) so that ρ dvi 1 d(vi vi ) dk = ρ =ρ dt 2 dt dt dk ∂σji + ρvi fi = vi dt ∂xj (2. 2.e. The ﬂuid is a common liquid (i. 1. are you sure that Φ > 0? 2.e. the dissipation.5). σji ∂vi ∂vi ∂vi + τji = −p ∂xj ∂xi ∂xj a b=Φ (2.20) This is the transport equation for kinetic energy.2. This is the transport equation for total energy. . u + k. 2.20 into work related to the pressure and viscous stresses respectively (see Eq.10 and 2. 2. Transformation of energy 22 ν (2. Φ.21) This is an equation for the sum of internal and kinetic energy.20 and 2.19) Re-write the stress-velocity term so that ρ ∂vi dk ∂vi σji − σji + ρvi fi = dt ∂xj ∂xj (2. this assumption was made when we assumed that the ﬂuid is incompressible 2. First we separate the term σji ∂vi /∂xj in Eqs. In lubricant oils the viscous heating (i.10 gives ρ ∂qi ∂σji vi d(u + k) − + ρvi fi = dt ∂xj ∂xi (2.e. Adding Eq.22) The following things should be noted. The physical meaning of the Prandtl number is the ratio of how well the ﬂuid diffuses momentum to the how well it diffuses internal energy (i. 2. in Eq. Exercise 14 Write out and simplify the dissipation term. not an lubricant oil). is neglected in Eq. Multiply Eq.15 because one of two assumtions are valid: Pr = 1. Φ) is large.e. 2. 1. temperature). k.21. 2. u + k. Φ.10. The ﬁrst term is positive and the second term is negative.

22 – which include the pressure. At the same time we have the usual viscous dissipation from the mean ﬂow to thermal energy.4. k (Eq. Let Ψ denote a transported quantity (i. 2. 2. Then we have transformation of kinetic energy from turbulence kinetic energy to thermal energy. i. Φ. The physical process is called production of turbulent kinetic energy.4 Left side of the transport equations So far.e. the left side in transport equations have been formulated using the material derivative. d/dt.23) nonconservative This is often called the non-conservative form. Left side of the transport equations 23 • The physical meaning of the a-term in Eq. the left side of the equation for momentum.10).se/˜lada . . and the other three terms can be collected into a convective term. The transformation of kinetic energy into internal energy takes place through this source term. 2. see Eq.e. .). For more detail. Eq. First energy is transferred from the mean ﬂow to the turbulent ﬂuctuations. 2. temperature etc reads ρ dΨ ∂Ψ ∂Ψ =ρ + ρvj dt ∂t ∂xj (2.chalmers. appears as a sink term in the equation for the kinetic energy. ρ dΨ ∂ρΨ ∂ρvj Ψ = + dt ∂t ∂xj (2. Ψ = vi . it can be re-written as ρ ∂Ψ ∂Ψ dΨ +Ψ =ρ + ρvj dt ∂t ∂xj ∂vj dρ +ρ dt ∂xj =0 = (2. u (Eq. which means that kinetic energy is transformed to thermal energy.2. It is denoted Φ. τij – is a dissipation. T .20) and it appears a source term in the equation for the internal energy. Dissipation is very important in turbulence where transfer of energy takes place at several levels. see section 2.e. and is called viscous dissipation. total energy.tfd. 2. thermal energy. u + k. 2. • The dissipation. but this is much smaller than that from the turbulence kinetic energy. When solving the Navier-Stokes equations nu2 can be downloaded from http://www. • The physical meaning of the b-term in Eq. this is turbulence dissipation (or heating). this makes sense since Φ represents a energy transfer between u and k and does not affect their sum. Using the continuity equation. i. p – is heating/cooling by compression/expansion.2.25) conservative This is called the conservative form. no loss of energy but only transformation of energy.21).22 – which include the viscous stress tensor.24) ∂Ψ ∂Ψ + ρvj +Ψ ρ ∂t ∂xj ∂ρ ∂vj ∂ρ + vj +ρ ∂t ∂xj ∂xj The two underlined terms will form a time derivative term. This is a reversible process. 2. u. It is always positive and represents irreversible heating.12.4 in [5]2 . 2. • Φ does not appear in the equation for the total energy u+k (Eq.

2. Vpart . ρvi (momentum) or ρu (energy). 2. The Reynolds transport theorem reads d dt = V ΦdV = Vpart V ∂vi dΦ +Φ dt ∂xi dV (2.3.5.12 and 2. . d/dt ρdV = 0. must be moving and it may expand or contract )if the density is non-constant). in this way we ensure that the transported quantity is conserved. to being valid for a ﬁxed volume.7. Newton’s second law. 5. the latter is usually ﬁxed (this is not necessary). momentum. where Vpart is a volume that includes the same ﬂuid particles all the time. How come? The answer is the Reynolds transport theorem.20) are all given for a ﬁxed control volume. the left side in the transport equation is always written in the form of Eq.2. The last term on the last line represents the net ﬂow of Φ across the ﬁxed non-deformable volume. control volume (Reynolds Transport Theorem) See also [2].2. 2. Vpart . . Chapt. This equation applies to any volume at every instant and the restriction to a collection of a material particles is no longer necessary. momentum and energy) for a collection of material particles. but no mass. The conservation of mass. Hence. Material particle vs.5. Φ in the equation above can be ρ (mass).26) ∂Φ ∂Φ ∂vi +Φ + vi ∂t ∂xi ∂xi dV = V = V ∂Φ ∂vi Φ dV + ∂t ∂xi ∂Φ vi ni ΦdS dV + ∂t S where V denotes a ﬁxed non-deformable volume in space. The results may be inaccurate due to too coarse a numerical grid. V . in ﬂuid mechanics the transport equations (Eqs. ) are valid both for a material collection of particles as well as for a volume. This means that the volume. the Navier-Stokes equation 2. The divergence theorem was used to obtain the last line and S denotes the bounding surface of volume V . V . control volume (Reynolds Transport Theorem) 24 merically using ﬁnite volume methods.5 Material particle vs.25. 2. In Part I [3] we initially derived all balance equations (mass. which converts the equations from being valid for a moving volume with a collection. .10. otherwise particles would move across its boundaries. the energy equations 2.2. 2. The equations we have looked at so far (the continuity equation 2. energy etc is lost (provided a transport equation for the quantity is solved): “what comes in goes out”. d/dt ρvi = Fi etc were derived for a collection of particles in the volume Vpart . .

v1 = v1 (x2 .1 is a similarity solution. 3.7 gives (no body forces. µ. Because the plate is inﬁnitely long. At the lower boundary (x2 = 0) and at the upper boundary (x2 → ∞) the velocity component v2 = 0. is related to x2 and t as x2 η= √ 2 νt (3.1.1) We will ﬁnd that the diffusion process depends on the kinematic viscosity.2: The v1 velocity at three different times. t) and p = p(x2 .8 1 v1 /V0 Figure 3. For increasing time (t3 > t2 > t1 ). the moving plate affects the ﬂuid further and further away from the plate.2. Eq. t) = 0 (3. f1 = 0) for the v1 velocity component ρ ∂ 2 v1 ∂v1 =µ 2 ∂t ∂x2 (3. 3 Exact solutions to the Navier-Stokes equation: two examples 3.e. 3. η.1: The plate moves to the right with speed V0 for t > 0. The similarity variable. It turns out that the solution to Eq. rather than the dynamic one. v1 (x2 → ∞. ν = µ/ρ. v1 (x2 = 0.3) similarity solution .3. x2 t2 0. 3. Exact solutions to the Navier-Stokes equation: two examples x2 x1 V0 25 Figure 3.6 0.2 t3 t1 0 0. see Fig. ∂v1 /∂x1 = ∂v3 /∂x3 = 0 so that the continuity equation gives ∂v2 /∂x2 = 0. this means that the number of independent variables is reduced by one.4 0. t3 > t2 > t1 .2) The solution to Eq. So. in this case from two (x2 and t) to one (η). i.1 The Rayleigh problem Imagine the sudden motion of an inﬁnitely long ﬂat plate. 3. t = 0) = 0. t) = V0 . Furthermore. which means that v2 = 0 in the entire domain. 3. t).e. there is no x1 dependency.1 is shown in Fig. The boundary conditions for Eq. For time greater than zero the plate is moving with the speed V0 . i.1 are v1 (x2 . 2. Hence the ﬂow depends only on x2 and t.

1) and the boundary conditions (Eq. i.8) f = C1 0 exp(−η ′2 )dη ′ + C2 (3. Integration once gives df = C1 exp(−η 2 ) dη Integration a second time gives η (3.6) v1 V0 (3.e.1 depends only on η. we get v1 (x2 .7) v1 (x2 → ∞. We get ∂v1 dv1 ∂η x2 t−3/2 dv1 1 η dv1 = =− √ =− ∂t dη ∂t 4 ν dη 2 t dη dv1 ∂η 1 dv1 ∂v1 = = √ ∂x2 dη ∂x2 2 νt dη 1 ∂ ∂ ∂v1 1 dv1 ∂ 2 v1 ∂ √ = √ = = ∂x2 ∂x2 ∂x2 ∂x2 2 νt dη 2 νt ∂x2 2 He introduce a non-dimensional velocity f= Inserting Eqs. it means that the solution for a given ﬂuid will√ the same (“similar”) for many (inﬁnite) values of x2 and t as long as the ratio be x2 / νt is constant.7) we conclude that the transformation is suitable. 3.1 from ∂/∂t and ∂/∂x2 to d/dη so that it becomes a function of η only.1 gives df d2 f + 2η =0 dη 2 dη (3. 3.4) We have now successfully transformed Eq. 3.10) At the limits. 3.3. the error function takes the values 0 and 1.6.4 and 3. t) = V0 ⇒ f (η = 0) = 1 (3. 3. Now let us solve Eq. Now let us ﬁnd out if the boundary conditions. 3. t = 0) = 0 ⇒ f (η → ∞) = 0 v1 (x2 = 0.1. 3. Now we need to transform the derivatives in Eq.1 and reduced the number of independent variables from two to one. 3.5) dv1 dη = 1 d2 v1 4νt dη 2 (3. The Rayleigh problem 26 If the solution of Eq.9 is (with C2 = 1 and C1 = −2/ π) f (η) = 1 − erf(η) (3. Taking into account the boundary conditions. also can be transformed in a physically meaningful way. Eq.5 in Eq.7.2. Eq. erf(0) = 0 and erf(η → ∞) = 1. the ﬁnal solution to √ Eq. 3. t) = 0 ⇒ f (η → ∞) = 0 Since we managed to transform both the equation (Eq. 3. 3.9) The integral above is the error function 2 erf(η) ≡ √ π η 0 exp(−η 2 ) (3.11) .

2 were obtained.4 0. To compute the velocity. is computed from Eq. The Rayleigh problem 27 3 2. the thickness of the boundary layer. v1 . This is how the graphs in Fig. 3. 3. From the velocity proﬁle we can get the shear stress as τ21 = µ ∂v1 µV0 df µV0 exp −η 2 = √ = −√ ∂x2 2 νt dη πνt (3. 3. increases. the magnitude of the shear stress increases for de√ creasing η and it is largest at the wall.3.5 0 −6 −5 −4 −3 −2 −1 0 τ21 /(ρV0 ) Figure 3. 3. i. As can be seen from Fig. τ21 .11 and the velocity.12) where we used ν = µ/ρ. 3.12 and circles are obtained by evaluating the derivative.5 2 η 1.4: The shear stress for water (ν = 10−6 ) obtained from Eq. f = v1 /V0 .4.3. 3.8 1 f Figure 3.2 0.11. 3. df /dη. the moving plate is felt further and further out in the ﬂow.4 below presents the shear stress.2 at p.13) ∂x2 dη 2 νt πνt From Fig.12 at time t = 100 000. The solid line is obtained from Eq. Figure 3.6 0.5 2 η 1. Compare this ﬁgure with Fig.3. 3. numerically using central differences (fj+1 − fj−1 )/(ηj+1 − ηj−1 ). 25 it is seen that for large times.e. τw = −ρV0 / πt The vorticity. 3. we pick a time t and insert x2 and t in Eq. ω3 .2 at p.1. 3. 25.5. 3. v1 . across the boundary layer is computed from its deﬁnition (Eq.5 0 0 0. δ. 1. all graphs in that ﬁgure collapse into one graph in Fig.3: The velocity.5 1 0.31) V0 df V0 ∂v1 =− √ =√ exp(−η 2 ) (3. Often ω3 = − . 3 2. Then f is obtained from Eq. The solution is presented in Fig.3.5 1 0. given by Eq. 3.

see Fig.2.1.5. v1 (x2 ).15) diffusion length As mentioned in the beginning of this section. this corresponds to the point where v1 = 0. 3. The physical agent which accomplish this is the pressure gradient which forces the ﬂow to follow the wall as closely as possible .2. Equation 3. incompressible ﬂow in a two-dimensional channel. see Fig.14) It can be seen that the boundary layer thickness increases with t1/2 .6 νt (3.e. µ. Create this graph with Matlab.3 and Eq. τ21 . Exercise 16 Consider the graphs in Fig. are δair = 10. say. Create this graph with Matlab for both air and engine oil. t2 and t3 .4. 3. the velocity near the walls is larger than in the center.3. 3.2 Flow between two plates Consider steady. with constant physical properties (i. 10 minutes the diffusion length for air and water.11 we ﬁnd that this occurs at δ η = 1. 3. The diffusion length can also be used to estimate the thickness of a developing boundary layer.14 can also be used to estimate the diffusion length. Exercise 15 Consider the graphs in Fig. µ = const). reaches 99% of the freestream velocity. 3. After. 3. From Fig. h 28 the boundary layer thickness is deﬁned by the position where the local velocity.3.2.5: Flow in a horizontal channel.01V0 . Choose suitable values on t1 . t2 and t3 . 3. In our case.1 Curved plates Provided that the walls at the inlet are well curved. Exercise 17 Repeat the exercise above for the shear stress.8cm δwater = 2.8cm (3. ν = µ/ρ rather than by dynamic one. see Fig. 3. respectively. note that the diffusion length is determined by the kinematic viscosity. Flow between two plates V P1 P2 P1 V x2 x1 Figure 3. see Section 4. Note that no scale is used on the x2 axis and that no numbers are given for t1 .3.5. The reason is that the ﬂow (with velocity V ) following the curved wall must change its direction. 3.8 = √ 2 νt √ ⇒ δ = 3. The inlet part of the channel is shown.

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. The ﬂow V approaches the bend and the ﬂow feels that it is approaching a bend through an increased pressure.e. is higher than the pressure near the wall. for moderately disturbed inﬂow.e. For ﬂow between two plates we get Dh = 2h. and the entrance length can. a swirling motion in the x2 − x3 plane). It is thus easier (i. Since the ﬂow is two-dimensional. x3 (i. The ﬂow is accelerated in the center because the mass ﬂow at each x1 must be constant because of continuity. in order to force the ﬂow to turn. P1 .e. Hence. the region until this occurs is called the entrance region.16) where V denotes the bulk (i. the pressure difference P2 − P1 creates secondary ﬂow downstream the bend (i. i. x1 (i.016 Dh ν (3. ∂v1 /∂x1 = ∂v2 /∂x1 = 0).2 Flat plates The ﬂow in the inlet section (Fig.2. 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. P2 .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. A and Sp denote the hydraulic diameter.6: Flow in a channel bend. the cross-sectional area and the perimeter. ∂/∂x3 ). Flow between two plates 29 P2 P1 x2 x1 V Figure 3. and Dh = 4A/Sp where Dh . Let us ﬁnd the governing equations for the fully developed ﬂow region. In a threedimensional duct or in a pipe.2. and the velocity in this direction is zero. P1 .5) is two dimensional. The pressure near the outer wall. the mean) velocity. it does not depend on the third coordinate direction.016ReDh ≡ 0. if a separation occurs because of sharp corners at the inlet. respectively. 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. be estimated as [6] x1. (if the wall is not sufﬁciently curved a separation will take place). Hence the pressure in the center of the channel. For large x1 the ﬂow will be fully developed. in this region the ﬂow does not change with respect to the streamwise coordinate.e. P2 . v3 = 0. 3. Taking these restrictions into account the continuity equation can be .6.e.3.e V Dh = 0. 3. 3. see Fig. must be higher than that near the inner wall. the pressure loss will be even higher.

when you want the physical pressure.e. 3.25 gives h dP x2 v1 = − (3.24 twice and using Eq. −ρg. see Fig.22. v1 (0) = v1 (h) = 0 (3. Since v2 = ∂v1 /∂x1 = 0 the left side is zero so dP ∂ 2 v1 (3.5). This is the vertical direction (x2 is positive upwards. The momentum equation can be written (see Eq.22) Hence the pressure. in the inlet section v2 = 0. P . see Fig. at x2 = 0) as P we get p = −ρgx2 + P (x1 ) (3.7 at p. fi = (0. decreases with vertical height.26) x2 1 − 2µ dx1 h . it means that v2 = 0 (3. This pressure is zero when the ﬂow is static.3. i. the ρgx2 as well as the surrounding atmospheric pressure must be added. If we denote the pressure at the lower wall (i. Flow between two plates 30 simpliﬁed as (see Eq.3) ∂v2 =0 ∂x2 Integration gives v2 = C1 and since v2 = 0 at the walls. see Eq.21) where the integration “constant” C1 may be a function of x1 but not of x2 . 3. We can now formulate the momentum equation in the streamwise direction ρ ∂v1 ∂v1 dP ∂ 2 v1 dv1 + ρv2 =− +µ 2 ≡ ρv1 dt ∂x1 ∂x2 dx1 ∂x2 (3.24) µ 2 =− ∂x2 dx1 Since the left side is a function of x2 and the right side is a function of x1 .e. p. v1 . This agrees with our experience that the pressure decreases at high altitudes in the atmosphere and increases the deeper we dive into the sea.e. 20) dv2 ∂v2 ∂v2 ∂p ∂ 2 v2 ρ + ρv2 =− + µ 2 − ρg (3. Now let us turn our attention to the momentum equation for v2 . is used in incompressible ﬂow. However. is zero at the walls. we conclude that they both are equal to a constant. 3. i. The gravity acts in the negative x2 direction. The velocity.2.5. 2.17) across the entire channel (recall that we are dealing with the part of the channel where the ﬂow is fully developed. Integrating Eq.23) hydrostatic pressure where p was replaced by P using Eq.5). Usually the hydrostatic pressure. 3.20) Integration gives (3.25) where h denotes the height of the channel.18) (3.e. 3.19) ≡ ρv1 dt ∂x1 ∂x2 ∂x2 ∂x2 Since v2 = 0 we get ∂p = −ρg ∂x2 p = −ρgx2 + C1 (x1 ) (3. 3. 0). 2. i. when the velocity ﬁeld is zero.

28) The mean velocity (often called the bulk velocity) is obtained by integrating Eq.2.max h h 0 4x2 1 − 2 x2 dx2 = v1. Eq. 3. Hence we integrate Eq.max Figure 3.28 across the channel.31 over the volume of a slice (length .28.31) = ρ dt ∂xj The left hand side is zero since the ﬂow is fully developed.26 gives ∂v1 h dP h ∆P τw = µ =− = (3. The negative pressure gradient is constant (see Eq. Flow between two plates 31 1 0. this result could have been obtained by simply taking a force balance of a slice of the ﬂow far downstream. v1.3 which reads for i=1 ∂σj1 dv1 (3. we start with Eq.e.2 0 0 0.27) We often write Eq.6 0. 3. the pressure is driving the ﬂow.4 0.4 0. we can compute the wall shear stress.2 0.7: The velocity proﬁle in fully developed channel ﬂow. for x2 = h/2.7 Since we know the velocity proﬁle. The velocity takes its maximum in the center.6 x2 /h 0.3. 3. Forces act on a volume and its bounding surface.max = h ∆P h 2µ L 2 1− 1 2 = h2 ∆P 8µ L (3.26 on the form v1 v1. i.8 1 v1 /v1. 3.30) ∂x2 2 dx1 2 L Actually.max h 3 (3.2.3 Force balance To formulate a force balance in the x1 direction. 3. 3. 1.e. The minus sign on the right side appears because the pressure gradient is decreasing for increasing x1 .8 0.24) and can be written as −dP/dx1 = ∆P/L.mean = v1. i. and reads v1. 3. Equation 3.29) The velocity proﬁle is shown in Fig.max = x2 4x2 1− h h (3.

see Fig.lef t = (−1. 0. 2. 0).U 32 P1 walls x2 τw.20. 3.20 .33) The bounding surface consists in our case of four surfaces (lower.2. upper. Eq. ni.lower = (0.L x1 V P2 h Figure 3. 3. L).36) where W is the width (in x3 direction) of the two plates (for convenience we set W = 1). left and right) so that 0= Slef t σj1 nj dS + Sright σj1 nj dS + Slower σj1 nj dS + Supper σj1 nj dS (3. 2.2.32) Recall that this is the form on which we originally derived the the momentum balance (Newton’s second law) in Part I. upper. [3] Now use Gauss divergence theorem 0= V ∂σj1 dV = ∂xj σj1 nj dS S (3. 0). Flow between two plates L τw. 3. Let us integrate this equation in the same way as we did for the force balance. 1.5 give 0=− (−p + τ11 )dS + Slef t Sright (−p + τ11 )dS − τ21 dS + Slower Supper τ21 dS (3. Inserting the normal vectors and using Eq.8 0= V ∂σj1 dV ∂xj (3. ni. −1.34) The normal vector on the lower.3. ni.35) τ11 = 0 because ∂v1 /∂x1 = 0 (fully developed ﬂow). The left side of Eq. 0.4 Balance equation for the kinetic energy In this subsection we will use the equation for kinetic energy. 0). Using this and Eq. The shear stress at the upper and lower surfaces have opposite sign because τw = µ(∂v1 /∂x2 )lower = −µ(∂v1 /∂x2 )upper . 0).upper = (0. 1. 3. left and right are ni.8: Force balance of the ﬂow between two plates. With ∆P = P1 − P2 we get Eq.22 give (the gravitation term on the left and right surface cancels and P and τw are constants and can thus be taken out in front of the integration) 0 = P1 W h − P2 W h − 2τw LW (3.30.right = (1.

40) Exercise 18 For the fully developed ﬂow.39 are zero on the upper and the lower surfaces in Fig.mean W h because ρgx2 n1 v1 on the left and right surfaces cancels.22 (term b). it should be!).2. compute the dissipation.39 is zero also on the left and right surfaces Fig. Exercise 21 Using the exact solution. 0. loss due to friction). 1. Exercise 22 From the dissipation.8 it is nj = (1. 0) and on the lower surface it is nj = (0. Now we apply Eq. 3. −1. 3. for the fully developed ﬂow. Finally we get ∆P = 1 W hv1. Flow between two plates 33 is zero because we assume that the ﬂow is fully developed. compute the vorticity.8.mean ΦdV V (3.37) On the ﬁrst line vi fi = v1 f1 + v2 f2 = 0 because v2 = f1 = 0. The last term corresponds to the viscous dissipation term. Φ (i. . The third term on the second line pδij ∂vi /∂xj = p∂vi /∂xi = 0 because of continuity.5 gives 0= ∂vi ∂vi σji − σji + ρvi fi ∂xj ∂xj =0 ∂vi ∂vi ∂vj p ∂vi τji + + pδij − τji =− ∂xj ∂xj ∂xj ∂xj Φ (3. compute the pressure drop. ﬁnd the error. 2. 3. 3. 3.39 to the ﬂuid enclosed by the ﬂat plates in Fig.38) Gauss divergence theorem on the two ﬁrst terms gives 0= S (−pvj + τji vi )nj dS − ΦdV V (3. For example.8. Exercise 20 Show that the second term in Eq. Is it the same as that obtained from the force balance (if not. Exercise 19 Show that the ﬁrst and second terms in Eq. 3. see Eq. using Eq. ωi . P can be taken out of the integral as it does not depend on x2 .22 so that (−P v1 + ρgx2 v1 )n1 dS = −(P2 − P1 ) v1 n1 dS Slef t &Sright Slef t &Sright = ∆P v1.e. Φ.28). The second term is zero on all four surfaces and the ﬁrst term is zero on the lower and upper surfaces (see Exercises below). 3. 3. The unit normal vector is denoted by nj which points out from the volume. on the right surface in Fig. We replace the pressure p with P using Eq. using the exact solution (Eq. 0). 3. Now we integrate the equation over a volume 0= V − ∂pvj ∂τji vi + − Φ dV ∂xj ∂xj (3.3.8.39) where S is the surface bounding the volume.

2) (4. was introduced in Eq.1 it is obvious that only the shear stresses are able to rotate the ﬂuid particle. x2 ) p(x1 + 0.4. ωi . the pressure acts through the center of the ﬂuid particle and is thus not able to affect rotation of the ﬂuid particle. 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.5∆x1 ) τ12 (x1 − 0. v1 = cx2 and v2 = 0.1 shows the surface forces in the x1 momentum equation acting on a ﬂuid particle in a shear ﬂow. vorticity is connected to rotation of a ﬂuid particle. Looking at Fig.5∆x2 ) 34 τ12 (x1 + 0. ∂τ21 /∂x2 > 0. 2 4 Vorticity equation and potential ﬂow 4.3) ∂ 2 vi ∂ 2 vj − ∂xj ∂xi ∂xj ∂xj = ∂ 2 vj ∂ 2 vk − εinm εmjk ∂xj ∂xi ∂xj ∂xn (4. In incompressible ﬂow the viscous terms read (see Eqs.3 at p.8 at p. 2.5 and 2. 2.5∆x1 ) (x1 . ωi .6) ∂ 2 vi ∂τji =µ ∂xj ∂xj ∂xj The right side can be re-written using the tensor identity ∂ 2 vi ∂ 2 vj = − ∂xj ∂xj ∂xj ∂xi Let’s verify that ∂ 2 vi ∂ 2 vj − ∂xj ∂xi ∂xj ∂xj = εinm εmjk ∂ 2 vk ∂xj ∂xn (4. As shown in Fig. ∂τ12 /∂x1 = 0. 11. 1. 14.5∆x1 ) p(x1 − 0. 38 εinm εmjk ∂ 2 vk ∂ 2 vk ∂ 2 vi ∂ 2 vk = (δij δnk − δik δnj ) = − ∂xj ∂xn ∂xj ∂xn ∂xi ∂xk ∂xj ∂xj (4.1: Surface forces in the x1 direction acting on a ﬂuid particle (assuming τ11 = τ22 = 0).1 at p. Vorticity equation and potential ﬂow τ21 (x2 + 0. Figure 4.1) Use the ε − δ-identity (see Table A.5∆x2 ) x1 Figure 4.4. 1.5∆x1 ) x2 τ21 (x2 − 0.1 Vorticity and rotation Vorticity. 4.4) .

6) (4. this means that the vorticity can’t be created or destroyed in inviscid (friction-less) ﬂow 2.e. The vorticity transport equation in three dimensions 35 The ﬁrst term on the right side is zero because of continuity and hence we ﬁnd that Eq.e.e. considering Item 1 this was to be expected. 4. An imbalance in shear stresses (left side of Eq. but also in that case no vorticity is generated or destroyed: it stays constant.9) (4. Eq. 4.7. Hence.7) (4. so that ∂ωm ∂ 2 vi = −εinm ∂xj ∂xj ∂xn In vector notation the identity Eq.9 for i = 1 and verify that it is satisﬁed. Exercise 23 Prove the ﬁrst equality of Eq. so that ∂ 2 vi ∂ωm ∂ 2 vj = − εinm ∂xj ∂xj ∂xj ∂xi ∂xn where the ﬁrst on the right side is zero because of continuity. ωm = εmjk ∂vk /∂xj .8) (4.6 reads ∇2 v = ∇(∇ · v) − ∇ × ∇ × v = −∇ × ω Using Eq. potential 4.4. 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.7 using the ε-δ-identity. (The exception is when vorticity is transported into an inviscid region. irrotational and potential ﬂow all denote frictionless ﬂow which is equivalent to zero vorticity. 4.1 reads ∂ωm ∂τji = −µεinm ∂xj ∂xn (4. unaffected. 4. There is a small difference between the three terms because there may be vorticity in inviscid ﬂow that is convected into the ﬂow at the inlet(s). there is a one-to-one relation between the viscous term and vorticity: no viscous terms means no vorticity and vice versa. The vorticity is always created at boundaries.) Inviscid ﬂow is often called irrotational ﬂow (i. Exercise 24 Write out Eq. The viscous terms in the momentum equations can be expressed in ωi . but also in this case the vorticity is not affected once it has frictionless . The main points that we have learnt in this section are: 1. no rotation) or potential ﬂow. generates vorticity (right side of Eq.2. 4. 4. The viscous terms are responsible for creating vorticity. i.5) Thus.1. see Section 4. The terms inviscid.2 can indeed be written as ∂ 2 vk ∂ 2 vj ∂ 2 vi = − εinm εmjk ∂xj ∂xj ∂xj ∂xi ∂xj ∂xn At the right side we recognize the vorticity.9) causes a change in vorticity. friction-less ﬂow) has no rotation.3.2 The vorticity transport equation in three dimensions Up to now we have talked quite a lot about vorticity. 4. inviscid ﬂow (i. 4.9).

14) The volume source is in most engineering ﬂows represented by the gravity which is conservative meaning that it is uniquely determined by the position (in this case the vertical position).11 can now be written as vj ∂vi = ∂xj ∂k − εijk vj ωk ∂xi no rotation rotation (4. Since the vorticity vector is deﬁned by the cross product εpqi ∂vi /∂xq (∇ × v in vector notation. i. In this section we will derive the transport equation for vorticity in incompressible ﬂow. 1.7) as vj 1 ∂vi = vj (Sij + Ωij ) = vj Sij − εijk ωk ∂xj 2 (4. see Exercise 8). we write fi = gi = −g ∂h ∂xi (4.20. 4. As usual we start with the Navier-Stokes equation. 4. However. we start by applying the operator εpqi ∂/∂xq to the Navier-Stokes equation (Eq. h.13) and one term including rotation (second term on the right side). Equation 4. The vorticity transport equation in three dimensions 36 entered the inviscid ﬂow region.14) so that εpqi ∂ 2 vi ∂2k ∂vj ωk + εpqi − εpqi εijk ∂t∂xq ∂xi ∂xq ∂xq ∂ 3 vi ∂2h 1 ∂2p + νεpqi − εpqi = −εpqi ρ ∂xi ∂xq ∂xj ∂xj ∂xq ∂xq ∂xi (4. Inserting Sij = (∂vi /∂xj + ∂vj /∂xi )/2 and multiplying by two gives 2vj ∂vi = vj ∂xj ∂vi ∂vj + ∂xj ∂xi − εijk vj ωk (4. Inserting Eq.2.7) yields ∂vi + ∂t 1 ∂p ∂ 2 vi ∂k − εijk vj ωk = − +ν + fi ∂xi ρ ∂xi ∂xj ∂xj no rotation rotation (4.16) . v × ω. we re-write the convective term of the incompressible momentum equation (Eq.4. Hence it can be expressed as a potential. 4.e.13 into the incompressible momentum equation (Eq.15) The negative sign appears because height is deﬁned positive upwards and the direction of gravity is downwards. 2. mostly no distinction is made between the three terms.15 on p. First.13) The last term on the right side is the vector product of v and ω. 12 was used.11) The second term on the right side can be written as vj ∂vj 1 ∂(vj vj ) ∂k = = = ∂xi 2 ∂xi ∂xi (4.7 at p. 2. The trick we have achieved is to split the convective term into one term without rotation (ﬁrst term on the right side of Eq.10) where Eq.12) where k = vj vj /2. 2. Eq.

1. That makes sense. We know that εijk is anti-symmetric in all indices. 0.1 at p. which means that the radius of the cylinder will decrease. 4. it will act as a source and increase ω1 and it will stretch the cylinder. Eq.e.21 has a new term on the right-hand side which represents ampliﬁcation and rotation/tilting of the vorticity lines.2) so that ωi = (ω1 .15. 4. 4. Hence vortex stretching will either make a ﬂuid element longer and thinner (as in the example above) or shorter and thicker (when ∂v1 /∂x1 < 0). A.17 can be written εpqi εijk ∂vj ωk ∂vp ∂ωp = ωk − vk ∂xq ∂xk ∂xk (4. 4. we have got rid of the pressure gradient term.20) Inserting Eqs. the vorticity) of a ﬂuid particle since the pressure acts through its center.20 into Eq. If we write it term-by-term it reads ω ∂v1 +ω ∂v1 + ω ∂v1 .1) εpqi εijk ∂vj ωk ∂vp ωk ∂vq ωp ∂vj ωk = (δpj δqk − δpk δqj ) = − ∂xq ∂xq ∂xk ∂xq ∂ωk ∂vp ∂ωp ∂vq = vp + ωk − vq − ωp ∂xk ∂xk ∂xq ∂xq (4.17) Using the deﬁnition of ωi we ﬁnd that its divergence ∂ ∂ωi = ∂xi ∂xi εijk ∂vk ∂xj = εijk ∂ 2 vk =0 ∂xj ∂xi (4.21) We recognize the usual unsteady term.19 and 4. The last term on line 1 is re-written using the ε-δ identity (see Table A. cylindrical ﬂuid particle with vorticity ωi and introduce a cylindrical coordinate system with the x1 -axis as the cylinder axis and r2 as the radial coordinate (see Fig. Equation 4. Using the continuity equation (∂vq /∂xq = 0) and Eq. We assume that a positive ∂v1 /∂x1 is acting on the ﬂuid cylinder. Furthermore.16 can be written as νεpqi ∂2 ∂ 3 vi =ν ∂xj ∂xj ∂xq ∂xj ∂xj εpqi ∂vi ∂xq =ν ∂ 2 ωp ∂xj ∂xj (4. and hence the second term on line 1 and the ﬁrst and the last term on line 2 are all zero (product of a symmetric and an anti-symmetric tensor). 1 p=1 2 3 ∂x1 ∂x2 ∂x3 ∂vp ∂v2 ∂v2 ∂v2 (4. 4. because as mentioned in connection to Fig.22) = ωk +ω2 + ω3 .2.16 gives ﬁnally ∂vp ∂ωp ∂ 2 ωp ∂ωp dωp = ωk +ν ≡ + vk dt ∂t ∂xk ∂xk ∂xj ∂xj (4. the pressure cannot affect the rotation (i. 4.19) The second term on line 2 in Eq. 1 p=3 2 3 ∂x1 ∂x2 ∂x3 The diagonal terms in this matrix represent vortex stretching. 4. Vortex stretching . 0).4. Imagine a slender. The volume of the ﬂuid element must stay constant during the stretching (the incompressible continuity equation).18. p=2 ω 1 ∂x1 ∂xk ∂x2 ∂x3 ω ∂v3 +ω ∂v3 + ω ∂v3 . The vorticity transport equation in three dimensions 37 where the body force in Eq. the convective term and the diffusive term. 4.18) is zero (product of a symmetric and an anti-symmetric tensor).16 was re-written using Eq. 4.

**4.3. The vorticity transport equation in two dimensions
**

v1 ω1 x2 x1 v1 ω1

38

Figure 4.2: Vortex stretching. Dashed lines denote ﬂuid element before stretching. ∂v1 > 0. ∂x1 ω2

v1 (x2 )

x2 ω2 x1

Figure 4.3: Vortex tilting. The off-diagonal terms in Eq. 4.22 represent vortex tilting. Again, take a slender ﬂuid particle, but this time with its axis aligned with the x2 axis, see Fig. 4.3. 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. 4.22 gives a contribution to ω1 . This means that vorticity in the x2 direction 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 threedimensional, see Section 5.4. Vortex tilting

**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 orthogonal 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 ∂ 2 ω3 dω3 =ν dt ∂xα ∂xα (4.23)

(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. 2.15. This means that vorticity diffuses in the same way as temperature does. In

4.3. The vorticity transport equation in two dimensions

39

fully developed channel ﬂow, for example, the vorticity and the temperature equations reduce to ∂ 2 ω3 ∂x2 2 ∂ 2T 0=k 2 ∂x2 0=ν (4.24a) (4.24b)

For the temperature equation the heat ﬂux is given by q2 = −∂T /∂x2 ; with a hot lower wall and a cold upper wall (constant wall temperatures) the heat ﬂux is constant 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 = −∂ω3 /∂x2 , is constant across the channel. You have plotted this quantity in TME225 Assignment 1. If wall-normal temperature derivative ∂T /∂x2 = 0 at both walls (adiabatic walls), the heat ﬂux is zero at the walls 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 ∂ω3 /∂x2 = 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). 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. In a boundary layer the streamwise pressure gradient is zero. This means that µ ∂ 2 v1 ∂x2 2 =0

wall

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. 2.7 and 3.23). Hence, the ﬂux of vorticity, γ2 = −∂ω3 /∂x2 = 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 and 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 boundary layer it is zero. 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. At the end of the plate the boundary thickness 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 . During this time vorticity will be transported by diffusion in the x2 direction the length δ according 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.

**4.3. The vorticity transport equation in two dimensions
**

x2 V0 δ

40

111111111 000000000

1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1 0 1111111 0000000 1 0 11111111111111111111111111111111111111111 00000000000000000000000000000000000000000 11111111111111111111111111111111111111111 00000000000000000000000000000000000000000

x1

L Figure 4.4: Boundary layer. The boundary layer thickness, δ, increases for increasing streamwise distance from leading edge (x1 = 0). Exercise 25 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 = 0 and ∂v1 /∂x1 = 0). The proper way to solve the problem is to use Blasius solution (you have probably learnt about this in your ﬁrst ﬂuid mechanics course; if not, you should go and ﬁnd out). Blasius solution gives 5 δ = , 1/2 L ReL Compute what δ(L) you get from Eq. 4.25. Exercise 26 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. ReL = V0 L ν (4.25)

5. Turbulence

41

5 Turbulence

5.1 Introduction

Almost all ﬂuid ﬂow which we encounter in daily life is turbulent. Typical examples 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 buildings 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 compute ﬂuid ﬂow it will most likely be turbulent. In turbulent ﬂow we usually divide the velocities in one time-averaged part vi , ¯ which is independent of time (when the mean ﬂow is steady), and one ﬂuctuating part ′ ′ vi so that vi = vi + vi . ¯ There is no deﬁnition on turbulent ﬂow, but it has a number of characteristic features (see Pope [7] and Tennekes & Lumley [8]) such as: I. Irregularity. Turbulent ﬂow is irregular, random and chaotic. 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 time and that it is subsequently destroyed (by the cascade process or by dissipation, 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 spectra 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. VI. Continuum. Even though we have small turbulent scales in the ﬂow they are much larger than the molecular scale and we can treat the ﬂow as a continuum.

turbulent eddy

cascade process

5.2. Turbulent scales

ﬂow of kinetic energy

42

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. ∂¯1 v = O(t−1 ) = O(v0 /ℓ0 ) 0 ∂x2 (5.1)

The kinetic energy of the large scales is lost to slightly smaller scales with which the large scales interact. Through the cascade process the 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 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 ﬂuctuating velocity gradient up to the power of two (see Section 8.1). The friction forces exist of course at all scales, but they are largest at the smallest eddies. Thus it is not quite true that eddies, which receive their kinetic energy from slightly larger scales, give away all of that to the slightly smaller scales. This is an idealized picture; in reality a small fraction is dissipated. However it is assumed that most of the energy (say 90%) that goes into the large scales 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 argument is as follows.

The Fourier coefﬁents are given by an = bn = Parseval’s formula states that L −L 1 L 1 L L f (x) cos(κn x)dx −L L f (x) sin(κn x)dx −L f 2 (x)dx = L 2 a +L (a2 + b2 ) n n 2 0 n=1 ∞ (5. f (x) = 1 a0 + (an cos(κn x) + bn sin(κn x)) 2 n=1 ∞ (5.3.e. and one for seconds [s] − 1 = −a − 3b. f (x) = f (x + 2L)). the turbulence ﬂuctuations are composed of a wide range of scales. The more energy that is to be transformed from kinetic energy to thermal energy. ℓη = ν3 ε 1/4 . see Fig. 5.5. dissipation: The amount of energy that is to be dissipated is ε. ℓη and τη in ν and ε using dimensional analysis. Energy spectrum 43 viscosity: Since the kinetic energy is destroyed by viscous forces it is natural to assume that viscosity plays a part in determining these scales.5) 5. f . We get two equations. Variable κn is called the wavenumber.1. can be expressed as a Fourier series. τη = ν ε 1/2 (5.e. It turns out that it is often convenient to use Fourier series to analyze turbulence. In general.4) (5. (5. one for meters [m] 1 = 2a + 2b.3) which give a = b = 1/4. In the same way we obtain the expressions for ℓη and τη so that vη = (νε) 1/4 .2) where below each variable its dimensions are given.3 Energy spectrum As mentioned above. the larger the velocity gradients must be. any periodic function.7) . Having assumed that the dissipative scales are determined by viscosity and dissipation. The dimensions of the left and the right side must be the same.6) where x is a spatial coordinate and κn = nπ/L. We can think of them as eddies. the larger scales. i. We write vη = νa εb 2 2 3 [m/s] = [m /s] [m /s ] (5. the larger viscosity. we can express vη . with a period of 2L (i.

9) Think of this equation as a way to compute the kinetic energy by ﬁrst sorting them by size (i. I: Range for the large. ′2 In this case the left side of Eq. is probably more familiar to “frequency”. The energy.e. isotropic scales. wavenumber). The left side of Eq. k. The dimension of wavenumber is one over length.7 expresses v1 as a function of time and the right side ′2 expresses v1 as a function of frequency. thus we can think of wavenumber as proportional to the inverse of an eddy’s diameter. a brief introduc tion is given in Section C. III: Range for small.e. E(κ)dκ). . In wavenumber space the energy of eddies can be expressed as E(κ)dκ (5. i. The reader who is not familiar to the term “wavenumber”.e κ ∝ 1/d. say v1 .1. 5. It is now convenient to study the kinetic energy of each eddy in wavenumber space. express f in Eq.e.e. 5. In that case. 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. For readers not familiar to Fourier series.2: Spectrum for turbulent kinetic energy. Let ′ now f be a ﬂuctuating velocity component. II: the inertial subrange. κn ).5. 5. Now let us think about how the kinetic energy of the eddies varies with eddy size. and ﬁnally summing the kinetic energy of all eddy sizes (i. see Fig. The total turbulent kinetic energy is obtained by integrating over the whole wavenumber space i.6 as a series in time rather than in space. x) and the right side v1 in wavenumber space (vs. 5. ∞ E ) (κ ∝ κ − 3 5/ III κ k= 0 E(κ)dκ = L f 2 (κn ) (5.7. 5.8 expresses the contribution from the scales with wavenumber between κ and κ + dκ to the turbulent kinetic energy k.7 expresses ′2 ′2 v1 in physical space (vs. vi vi /2. corresponds to f 2 (κ) in Eq. then computing the kinetic energy of each eddy size (i.8) where Eq. 5. Energy spectrum Pk E(κ) I II εκ 44 ε Figure 5. E(κ). carrying out the integration). Intuitively we assume that large eddies have large ﬂuctuating velocities ′ ′ which implies large kinetic energy.3. energy containing eddies.

The “transport” in wavenumber space is called spectral transfer. see Eq. The eddies’ velocity and length scales are v0 and ℓ0 . i. The scales of the eddies are described by the Kolmogorov scales (see Eq. ε 3 κ− 3 This is a very important law (Kolmogorov spectrum law or the −5/3 law) which states that. P k = ε. 5. 5. 8.2. Energy spectrum 45 The kinetic energy is the sum of the kinetic energy of the three ﬂuctuating velocity components. This region is a “transport region” (in wavenumber space i.11 we get E(κ) = const.i vκ.5) II. is given by the concept of the cascade process).11) 3 2 2 3 [m /s ] = [1/m] [m /s ] We get two equations. and one for seconds [s] so that b = 2/3 and a = −5/3.e. Since the cascade concept assumes that all turbulent kinetic energy is transferred from large to small eddies. The eddies in this region are independent of both the large. . The energy of the largest eddies is transferred to slightly smaller scales. k= 1 ′ ′ 1 ′2 ′2 ′2 v + v2 + v3 = vi vi 2 1 2 (5. This energy transfer takes places via the production term. Energy per time unit. 5. see Fig. if the ﬂow is fully turbulent (high Reynolds number). see Eq. 5. P k . The energy transfer from turbulent kinetic energy to thermal energy (increased temperature) is governed by ε in the transport equation for turbulent kinetic energy. One can argue that the eddies in this region should be characterized by the spectral transfer of energy (ε) and the size of the eddies 1/κ. is coming from the large eddies at the lower part of this range and is given off to the dissipation range at the higher part (note that the relation P k = {dissipation at small scales}. Nκ kκ is con′ ′ stant (Nκ and kκ = vκ. Dissipation range. energy-containing eddies and the eddies in the dissipation range. one for meters [m] 3 = −a + 2b. respectively.) in the cascade process. i. The eddies in this region represent the mid-region. respectively). 5. Fig.14. 2 5 spectral transfer −2 = −3b.e. it means than the number of small eddies must be much larger than that of large eddies.3. II and III which correspond to: I.14. We ﬁnd region I.2. The existence of this region requires that the Reynolds number is high (fully turbulent ﬂow). The eddies are small and isotropic and it is here that the dissipation occurs.10) The spectrum of E is shown in Fig. These eddies interact with the mean ﬂow and extract energy from the mean ﬂow.i /2 denote the number of eddies and kinetic energy eddies of size 1/κ. Inserted in Eq.e. In this region we have the large eddies which carry most of the energy.5. Dimensional analysis gives E = κa εb (5.2). in the transport equation for turbulent kinetic energy. the energy spectra should exhibit a −5/3-decay in the inertial region (region II. Inertial subrange. 8. III.

the energy spectral transfer rate. i. the cascade process assumes that εκ = ε. Hence. length and time scales of the energy-containing eddies to the Kolmogorov eddies increases with increasing Reynolds number. Hence. Energy spectrum 46 Above we state that the small eddies are isotropic. Using only if v1 2 our knowledge in tensor notation. the Reynolds stress tensor for small scales can be written as vi vj = const. εκ .13 give v0 3 = (νε)−1/4 v0 = νv0 /ℓ0 vη ℓ0 = ℓη τo = τη ν3 ε νℓ0 3 v0 −1/4 −1/4 v0 = (v0 ℓ0 /ν)1/4 = Re1/4 ℓ0 = ν3 3 v0 ℓ3 0 v0 ℓ0 ν −1/4 ℓ0 = −1/2 ν 3 ℓ0 3 v0 3 v0 νℓ0 −1/4 = Re3/4 1/2 (5. Equations 5. As discussed on p. increases with increasing Reynolds number.δij . all shear stresses are zero in isotropic turbulence. the inertial region). nothing should be changed. Note that is not true instantaneously. the larger the wavenumber range of the intermediate range where the eddies are independent of both the large scales and the viscosity. We can now estimate the ratio between the large eddies (with v0 and ℓ0 ) to the Kolmogorov eddies (vη and ℓη ). (region II.12) ℓκ ℓκ vκ The kinetic energy is transferred to smaller and smaller eddies until it is dissipated at the dissipative scales.e. Furthermore. In the inertial subrange. i.5. 42. The kinetic energy of the eddy is 2 proportional to vκ and the time for one revolution is proportional to ℓκ /vκ . This means that – in average – the eddies have no preferred direction. we know that an isotropic tensor can be written as ′ ′ const.14) 1/2 τ0 = ℓ0 = v0 = Re1/2 where Re = v0 ℓ0 /ν.13) The dissipation at small scales (large wavenumbers) is determined by how much energy that enters the cascade process at the large scales (small wavenumbers).δij which. An eddy loses its kinetic energy during one revolution.e. in general ′ ′ ′ v1 = v2 = v3 . the larger the Reynolds number. again.12 for the large energy-containing eddies gives ε0 = O 2 v0 ℓ0 v0 isotropic turbulence =O 3 v0 ℓ0 = εκ = ε (5. shows us that the shear stresses are zero in isotropic turbulence. the concept of the cascade process assumes that the energy extracted by the large turbulent eddies is transferred by non-linear interactions through the inertial range to the dissipative range where the kinetic energy is transformed to thermal energy (increased temperature).5 and 5. This means that the eddy range (wavenumber range) of the intermediate region. We ﬁnd that the ratio of the velocity. Applying Eq. v1 v2 = −v1 v2 .2) 2 3 vκ vκ εκ = O =O (5. the ﬂuctuations in all directions are the ′2 ′2 ′2 same so that v1 = v2 = v3 . Hence. 5. For example if the x1 coordinate direction is ′ ′ ′ ′ ′ ′ rotated 180o the v1 v2 should remain the same.e. isotropic turbulence implies that if a coordinate direction is switched. . Hence. 5. i.3. The spectral transfer rate of kinetic energy from eddies of size 1/κ to slightly smaller eddies can be estimated as follows. for a an eddy of length scale 1/κ can be estimated as (see Fig. This is possible ′ v ′ = 0.

2 and from the requirement that the volume must not change (incompressible continuity equation) we ﬁnd that the radius of the cylinder will decrease. is 1st generation.4. The cascade process created by vorticity 1 47 2 2 3 3 3 3 4 4 4 4 4 4 4 4 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 Figure 5. 4. see Fig. Two idealized phenomena in this interaction process can be identiﬁed: vortex stretching and vortex tilting.e. with axis aligned in the x1 direction. ′ The equation for the instantaneous vorticity (ωi = qi + qi ) reads (see Eq. Five generations. the disturbances are turned into chaotic.3: Family tree of turbulent eddies (see also Table 5.2 this equation is not an ordinary convection-diffusion equation: it has an additional term on the right side which represents ampliﬁcation and rotation/tilting of the vorticity lines (the ﬁrst term on right side).4 The cascade process created by vorticity The interaction between vorticity and velocity gradients is an essential ingredient to create and maintain turbulence.15) As we learnt in Section 4. The large original eddy. into turbulence.21) ¯ ∂vi ∂ωi ∂ 2 ωi ∂ωi = ωj +ν + vj ∂t ∂xj ∂xj ∂xj ∂xj ∂vk ωi = ǫijk ∂xj (5.1). i. three-dimensional. The i = j components of this term represent (see Eq.5. 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. 42). Thus we can assume that there are no viscous stresses acting Vortex stretching .22) vortex stretching. 4. 4. Adapted from [9] 5. random ﬂuctuations. A positive ∂v1 /∂x1 will stretch the cylinder. Disturbances are ampliﬁed by interaction between the vorticity vector and the velocity gradients.

Figure 5.5. For each generation the eddies become more and more isotropic as they get smaller. The increased v2 and v3 velocity components will in next stage stretch smaller ﬂuid elements aligned in these two directions and so on. 5. now with its axis aligned with the x2 axis. Hence an extension of a ﬂuid element in one direction (x1 direction) decreases the length scales and increases the velocity scales in the other two coordinate directions ′ ′ (x2 and x3 ).22 gives a contribution to ω1 . the small eddies “don’t remember” the characteristics of their original ancestor. 4. This shows how vorticity in one direction is transferred to the other two directions through vortex tilting.3. because vorticity generated by vortex stretching and Vortex tilting .2). The cascade process created by vorticity generation 1st 2nd 3rd 4th 5th 6th 7th x1 1 0 2 2 6 10 22 x2 0 1 1 3 5 11 21 x3 0 1 1 3 5 11 21 48 Table 5. on the cylindrical ﬂuid element surface which means that the angular momentum r2 ω1 = const. The i = j components in the ﬁrst term on the right side in Eq. As was mentioned above. x2 or x3 direction. Once this process has started it continues.2. the length scale of the eddies – whose velocity scale are increased – decreases.3 and Table 5. Equation 5. 5. Fig.1: Number of eddies at each generation with their axis aligned in the x1 . The smaller the eddies. the second term ω2 ∂v1 /∂x2 in line one in Eq. Here a generation is related to a wavenumber in the energy spectrum (Fig. (5.4. 4. which in turn each create new eddies in the x1 and x3 (3rd generation) and so on.e. the continuity equation shows that stretching results in a decrease of the radius of a slender ﬂuid element and an increase of the vorticity component (i.16 shows that the vorticity increases if the radius decreases (and vice versa). take a slender ﬂuid element. It creates eddies in the x2 and x3 direction (2nd generation).16) remains constant as the radius of the ﬂuid element decreases. They are isotropic. Again. The velocity gradient ∂v1 /∂x2 will tilt the ﬂuid element so that it rotates in the clock-wise direction. The small eddies have no preferred direction. As a result. 4. see Fig. Γ – which is the integral of the tangential velocity round the perimeter – is constant. young generations correspond to high wavenumbers. the tangential velocity component) aligned with the element. 5. 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.22 represent vortex tilting. At each stage. Note that also the circulation.1 The large original eddy (1st generation) is aligned in the x1 direction. In other words.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. The creation of multiple eddies by vortex stretching from one original eddies is illustrated in Fig. the less the original orientation of the large eddy is recalled.

The only non-zero component of vorticity vector is ω3 because ∂v3 ∂v2 − ≡0 ∂x2 ∂x3 ∂v1 ∂v3 − ≡ 0. . The vorticity and velocity ﬁeld becomes chaotic and random: turbulence has been created. 5. The turbulence is also maintained by these processes. we get ωj ∂vi /∂xj = 0.4. The cascade process created by vorticity 49 vortex tilting interacts with the velocity ﬁeld and creates further vorticity and so on. 41). From the discussion above we can now understand why turbulence always must be three-dimensional (Item IV on p. ∂x3 ∂x1 ω1 ω2 = = Since v3 = 0. If the instantaneous ﬂow is two-dimensional (x1 − x2 plane) we ﬁnd that the vortex-stretching/tilting term on the right side of Eq.15 vanishes because the vorticity vector and the velocity vector are orthogonal.5.

4). vi (Eq.e.5 which is called the Reynolds stress tensor. Inserting Eq.1 into the continuity equation (6. 6.e. We need a model for vi vj to close the equation system in Eq. 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). ¯. pressure.3) we obtain the time averaged continuity equation and Navier-Stokes equation ∂¯i v ∂xi ∂¯i vj v¯ ρ ∂xj = = 0 − ∂p ¯ ∂ + ∂xi ∂xj µ ∂¯i v ′ ′ − ρvi vj ∂xj (6. The continuity equation and the Navier-Stokes equation for incompressible ﬂow with constant viscosity read ∂vi ∂xi ∂vi ∂vi vj ρ +ρ ∂t ∂xj = 0 = − ∂p ∂ 2 vi +µ ∂xi ∂xj ∂xj (6. Turbulent mean ﬂow 50 6 Turbulent mean ﬂow 6.e. The ′ ′ ′ ′ tensor is symmetric (for example v1 v2 = v2 v1 ). see p.3) The gravitation term.5. It represents correlations between ﬂuctuating velocities. vi (Eq.6) ∂xi Reynolds equations closure problem . It is an additional stress term due to turbulence (ﬂuctuating ve′ ′ locities) and it is unknown. hence it applies also for the ﬂuctuating ¯ ′ velocity.1) where the bar. 6. One reason why we decompose the · variables is that when we measure ﬂow quantities we are usually interested in their mean values rather than their time histories. This is called the closure problem: the number of unknowns (ten: three velocity components.1 Time averaged Navier-Stokes When the ﬂow is turbulent it is preferable to decompose the instantaneous variables (for example the velocity components and the pressure) into a mean value and a ﬂuctuating value. −ρgi . denotes the time averaged value.2) and the Navier-Stokes equation (6. six stresses) is larger than the number of equations (four: the continuity equation and three components of the Navier-Stokes equations). and for the time-averaged velocity. 6. ′ ∂vi =0 (6. i. then p ≡ 0. The continuity equation applies both for the instantaneous velocity. This equation is the time-averaged ′ ′ Navier-Stokes equation and it is often called the Reynolds equation. when vi ≡ 0. 6. i.2). 6. vi .4) (6. ′ vi = vi + vi ¯ p = p + p′ ¯ (6.2) (6. A new term ρvi vj appears on the right side of Eq. 30). has been omitted which means that the p is the hydrostatic pressure (i.6.5) It is assumed that the mean ﬂow is steady.

pipe ﬂow. channel ﬂow.e. etc.5 using the continuity equation ρ ∂¯i v ∂¯j v ∂¯i v ∂¯i vj v¯ = ρ¯j v + ρ¯i v = ρ¯j v ∂xj ∂xj ∂xj ∂xj =0 (6.11) w .7) First we re-write the left side of Eq. 6. Here the velocity gradient is largest as the velocity drops down to zero at the wall over a very short distance.1: Flow between two inﬁnite parallel plates. v v In addition to the viscous shear stress.5 can be written ρ¯1 v ∂¯1 v ∂¯1 v + ρ¯2 v ∂x1 ∂x2 = − ∂¯1 v ∂p ¯ ∂ ′ ′ µ + − ρv1 v2 ∂x1 ∂x2 ∂x2 τtot (6.e. boundary layers along a ﬂat plate. Eq. ¯ ¯ ∂¯1 v ∂¯1 v ≪ . length in the x3 direction) of the plates. 6. 6.1. Zmax . is much larger that the separation between the plates.8. two-dimensional (¯3 = ∂/∂x3 = 0) boundary-layer type of v ﬂow (i. µ∂¯1 /∂x2 . ∂x1 ∂x2 (6. an additional turbulent one – a v Reynolds shear stress – appears on the right side of Eq. i. The total shear stress is thus ∂¯1 v ′ ′ (6.e.1.8) Using Eq.9) x1 and x2 denote the streamwise and wall-normal coordinate. because they both include the product of one large (¯1 or ∂/∂x2 ) and one small (¯2 or ∂/∂x1 ) part.6. see Fig. Zmax ≫ δ. respectively.1 Boundary-layer approximation For steady (∂/∂t = 0). 6. 6.10) − ρv1 v2 τtot = µ ∂x2 shear stress 6. One important quantity is the wall shear stress which is deﬁned as τw = µ ∂¯1 v ∂x2 (6.9.2.) where v2 ≪ v1 . 6. jet and wake ﬂow. The width (i. Note that the two terms on the left side are of the same order.2 Wall region in fully developed channel ﬂow The region near the wall is very important. Wall region in fully developed channel ﬂow 51 x2 v1 (x2 ) ¯ xA 2 2δ x1 x3 Figure 6.

The continuity v equation gives now v2 = 0. 3.2. i. the non-linear convection) term. In the logarithmic layer 2 the viscous stress is negligible compared to the Reynolds stress. see Fig. let us. 6.15 and Fig.3.24 and text related to this equation). Wall region in fully developed channel ﬂow From the wall shear stress. ∂p ¯ = -constant ∂x1 ∂τtot = constant = ∂x2 − ∂ ∂x2 ∂¯1 v ′ ′ µ − ρv1 v2 ∂x2 (6. 6.9.1.11 and 6. the velocity gradient is directly related to the wall shear stress.e. 6.30 (note that h = 2δ) and Eq. hence they must be constant (see Eq. In the inner region. The streamwise momentum equation. because the turbulent stresses stem from the inertial (i. The inner region includes the viscous region (dominated by the viscous diffusion) and the logarithmic region (dominated by turbulent diffusion). is given by Eq. we can deﬁne a wall friction velocity. 30. 6. 6. 6. Note that the total shear stress is constant only close to the wall (Fig. the streamwise derivative of the streamwise velocity component is zero (this is the deﬁnition of fully developed ﬂow).3a).e. again.16) ∂x2 w µ µ τ ν τ . as τw = ρu2 ⇒ uτ = τ τw ρ 1/2 52 (6. the logarithmic region is sometimes called the inertial region. i. 6. Eq.15) At the last step we used the fact that the pressure gradient balances the wall shear stresses. The ﬁrst term on the left side of ¯ v Eq.14) where the total stress. As we go away from the wall the viscous stress decreases and the turbulent one increases and at x+ ≃ 11 they are approximately equal.9 is zero because we have fully developed ﬂow (∂¯1 /∂x1 = 0) and the last term is zero because v2 ≡ 0.18 at p. The Reynolds shear stress vanishes at the ′ ′ wall because v1 = v2 = 0. 6.e. 6. uτ . −∂ p/∂x1 = τw /δ. Integrating Eq.10.2. ¯ The wall region can be divided into one outer and one inner region. 6. τtot . consider fully developed channel ﬂow between two inﬁnite plates. 6. (see Eq. see Fig.31. further away from the wall it decreases (in fully developed channel ﬂow it decreases linearly with the distance from the wall.13) 0=− µ + − ρv1 v2 ∂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. see Eq. and the viscous shear stress attains its wall-stress value 2 τw = ρuτ . see Eq.12) ∂¯1 v τw ρ 1 = = u2 = u2 (6. the total shear stress is approximately constant and equal to the wall shear stress τw .3b).13 from x2 = 0 to x2 τtot (x2 ) − τw = ∂p ¯ x2 ∂p ¯ x2 ⇒ τtot = τw + x2 = τw 1 − ∂x1 ∂x1 δ (6. see Eq. can now be ¯ written ∂p ¯ ∂¯1 v ∂ ′ ′ (6. i. 3.12) wall friction velocity In order to take a closer look at the near-wall region. The buffer region acts as a transition region between these two regions where viscous diffusion of streamwise momentum is gradually replaced by turbulent diffusion. i. At the wall. ∂¯1 /∂x1 = 0.e.6. 6. 3. In fully developed channel ﬂow. see Fig.e.

since the ﬂow is dominated by viscosity).003 x2 /δ 0. In this region the ﬂow is assumed to be ′ ′ independent of viscosity. v1 . 6. is a suitable velocity scale in the inner logarithmic region. is in the region x+ 200 2 (i.17 can then be written v1 = x+ ¯+ 2 x+ 2 (6. 6.17) Equation 6.19) .2.3). the integration constant C1 = 0 so that ¯ uτ x2 v1 ¯ = uτ ν (6.2. x2 /δ 0.6.e. 6. see Fig. it is used in the entire region. Integration gives (recall that both ν and u2 are constant) τ v1 = ¯ 1 2 u x2 + C1 ν τ Since the velocity. Hence the friction velocity.20) (6.18) Further away from the wall at 30 3000 (or 0. The Reynolds shear stress.e. 6.1 and x+ = x2 uτ /ν denotes the normalized wall 2 distance. uτ .2: The wall region (adapted from Ch.1) fairly constant and approximately equal to τw . Hence it seems reasonable to take the wall distance as the characteristic length scale. ρv1 v2 . i. the friction velocity stemming from the wall shear stress and the viscosity (here we regard viscosity as a quantity related to the wall.3b. What about the length scale? Near the wall. a constant is added so that ℓ = κx2 The velocity gradient can be estimated as ∂¯1 v uτ = ∂x2 κx2 (6. Wall region in fully developed channel ﬂow 10−4 10−3 10−2 10−1 outer region overlap region Centerline log-law region Wall 1 inner region buffer region viscous region 1 53 x2 /δ x+ = y + 2 5 10 30 100 1000p 10000 Figure 6. see Fig. δ denotes half width of the channel. is zero at the wall. Often the plus-sign (‘ + ‘) is used to denote inner scaling and equation Eq. 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 eddy-size.17 is expressed in inner scaling (or wall scaling) which means that v1 and ¯ x2 are normalized with quantities related to the wall. see Fig. we encounter the log-law region.7 in [7]) for Reτ = 10 000.

11]. : v1 /uτ = (ln x+ )/0. i.2 a) 0 0 0.21 gives (inserting −v1 v2 = u2 ) τ (6.30) in which a turbulent Reynolds stress is assumed to be equal to the product between the turbulent viscosity and the velocity gradient as ′ ′ −v1 v2 = νt ∂¯1 v ∂x2 (6.20 and 6. and the length scale κx2 . Hence νt can be expressed as a product of a turbulent velocity scale and a turbulent length scale.22) u2 = κuτ x2 τ ∂¯1 v ∂¯1 v uτ ⇒ = ∂x2 ∂x2 κx2 (6.6.41 + 5. 11].2.4 500 100 0.8 x2 /δ 1000 0. represents the turbulence and has the same dimension as ν. : −ρv1 v2 /τw .6 0.e.4 0.3: Reynolds shear stress. Wall region in fully developed channel ﬂow 2000 1 200 54 0.2 0. ¯ : v1 /uτ = x+ .1 1500 0.8 0 1 Figure 6. 11.21) The turbulent viscosity. Another way of deriving the expression in Eq. and in the log-law region that gives νt = uτ κx2 ′ ′ so that Eq.23) In non-dimensional form Eqs. Reτ = 2000.8 0 1 0 b) 0. : µ(∂¯1 /∂x2 )/τw . data [10. νt . 6.23 read ∂¯1 v+ 1 = ∂x+ κx+ 2 2 (6. 6. DNS data [10.2 0.20 is to use the Boussinesq assumption (see Eq. ¯ 2 2 based on the velocity scale.6 0. 6.4 0. v 2000 24 20 16 12 8 500 4 0 0 1500 x+ 1000 2 a) 5 10 15 20 25 v1 /uτ ¯ v1 /uτ ¯ b) 1 10 100 1000 x+ 2 : DNS Figure 6.6 x+ 2 x+ 2 .05 0. [m2 /s].2.4: Velocity proﬁles in fully developed channel ﬂow. Reτ = 2000. a) lower half of the channel.24) x2 /δ 0. uτ . b) zoom ′ ′ near the wall.

c − v1 ¯ ¯ x2 = FD uτ δ where c denotes centerline. From Eq.3. 6. Since nothing changes in the x3 direction. As can be seen in Fig.25) where B is an integration constant.5: Symmetry plane of channel ﬂow.2 – show that the log-law ﬁt the DNS up to x+ 500 (x2 /δ 0.41 and a a B = 5. 2 In the outer region of the boundary layer. as ν (6. Consider the x2 − x3 plane.tot x2 x3 Figure 6.6.29) . 6. The constants in the log-law are usually set to κ = 0.26) x+ = x2 /ℓν ⇒ ℓν = 2 uτ Equation 6. is called a a the von K´ rm´ n constant. the viscous shear stress τ32 = µ ∂¯3 v ∂¯2 v + ∂x2 ∂x3 =0 (6. The constant. the relevant length scale is the boundary layer thickness. Reynolds stresses in fully developed channel ﬂow 55 τ32.28) ′ ′ because v3 = ∂¯2 /∂x3 = 0. 11.2 the log-law applies for x+ 2 Figure 6. The turbulent part shear stress.4 – where the Reynolds number is lower than in Fig. (6.2. ℓν . v see Fig.25 is the logarithmic law due to von K´ rm´ n [12]. Integration gives now 1 ln x+ + B or 2 κ v1 ¯ 1 x2 uτ +B = ln uτ κ ν v1 = ¯+ (6.25 we can deﬁne the viscous length scale. 6.25). can be expressed ¯ v using the Boussinesq assumption (see Eq.5.30) ′ ′ −ρv2 v3 = µt ∂¯3 v ∂¯2 v + ∂x2 ∂x3 =0 (6.27) log-law 6. 3000 (x2 /δ 0.3). κ. The resulting velocity law is the defect law v1.3 Reynolds stresses in fully developed channel ﬂow The ﬂow is two-dimensional (¯3 = 0 and ∂/∂x3 = 0). 6. v2 v3 .

the forces acting on the ﬂuid.2. : −ρ(∂v1 v2 /∂x2 )/τw .1252 + 0. Hence v3 = 0 and v3 = 0 so that v3 = 0. the instantaneous turbulent ﬂow is always three¯ ′ ′2 dimensional and unsteady. the viscous shear stress is negligible except very near the wall. : µ(∂ 2 v1 /∂x2 )/τw . ¯ 2 ′ ′ and it is also zero since v3 = ∂¯2 /∂x3 = 0. 0.3 presents the Reynolds and viscous shear stresses for fully developed ﬂow. the Reynolds shear stress is zero at the wall (because the ﬂuctuating velocities are zero at the wall) and increases for increasing wall distance. for example) is driven by a fan. 0.125. Figure 6. This opposing force has its origin at the walls due to the viscous wall force (viscous shear stress multiplied by area). 6.2). 6.25 + 0. v1 v3 = 0. ′ ′ but its gradient. v3 = 0). b) lower half of the channel ex′ ′ cluding the near-wall region. −0. On the other hand.13 takes a positive value which pushes the ﬂow in the x1 direction.03475 = 0.2)/5 = 0 ¯ but ′2 2 v3 = v3 = (−0. in ¯ Eq.02). Reτ = 2000. ¯ see Eq.2)2 + 0. 6. The force that balances the pressure gradient is the gradient of the Reynolds shear stress. 2 Looking at Eq. for ′ example. This is the force opposing the movement of the ﬂuid. −∂(ρv1 v2 )/∂x2 and µ∂ 2 v1 /∂x2 ¯ 2 represent.25.25)2 + 0. Here the forces are two orders of magnitude larger than in Fig. This gives v3 = (−0.125 − 0.3. Start by looking at Fig.6. Forces in the v1 equation. 11].22 /5 = 0. a) near the lower wall of the channel. 6.125. We can imagine that the ﬂuid (air.6b but they act over a very thin region (x+ ≤ 40 or x2 /δ < 0. The pressure must decrease in the streamwise direction so that the pressure gradient term.6a. Now let’s have a look at the forces in the wall region. 0. The pressure gradient is constant and equal to one: this is the force driving the ﬂow.e. The gradient of the shear stress. This agrees – fortunately – with our intuition.13. the time series v3 = v3 = (−0. With the same argument. The reason is that although the time-averaged ﬂow However note that v3 3 is two-dimensional (i.6: Fully developed channel ﬂow. As can be seen. The intersection of the two shear stresses takes place at x+ ≃ 11. ¯ : −(∂ p/∂x1 )/τw . Reynolds stresses in fully developed channel ﬂow 200 2000 1800 150 1600 1400 56 x+ 2 100 x+ 2 1200 1000 800 50 600 400 a) 0 −150 −100 −50 0 50 100 150 b) 200 −2 −1 0 1 2 Figure 6.6 presents the forces.13 we ﬁnd that it is not really the shear stress that is interesting. 6. Another way to describe the behaviour of the pressure is to say that there is a pressure drop.125 + 0. It is equal to one near the wall and decreases rapidly for increasing wall distance. DNS data [10. 6.2 + 0.6b which shows the forces in the region away from the wall. see Fig. −∂ p/∂x1 . ¯ v ′2 = v 2 = 0. −∂ p/∂x1 .1252 + (−0. In this region the shear stress gradient term is driving 2 . ¯ Figure 6. together with the pressure gradient. Consider.

i. What is the difference between that ﬂow and a boundary layer ﬂow? First. see Eq. 3.8 at p.8: the linear and the log-law regions .1) is principally the same as for the fully developed channel ﬂow. In turbulent ﬂow the velocity proﬁle in the center region is much ﬂatter than in laminar ﬂow (cf. is also included. ′2 ′2 ′2 Figure 6. δ. 11]. the left side of Eq. : ρv2 /τw . 6. the convective terms must also be taken into account. The ﬂow in a boundary layer is continuously developing. ′ ′ ′2 k = vi vi /2. The former is largest because the mean ﬂow is in this direction. in a boundary layer ﬂow the wall shear stress is not determined by the pressure drop. Third.4.e.7: Normal Reynolds stresses and turbulent kinetic energy. This makes the velocity gradient near the wall (and the wall shear stress. : ρv1 /τw .9 is not zero. 6. see Fig.4 and Fig.9. the inner region of a boundary layer (x2 /δ < 0.31) (6. in a boundary layer ﬂow the convective terms are not zero (or negligible). ◦: k/u2 . The turbulent kinetic energy. i. 3. 6. its thickness. However.4 Boundary layer Up to now we have mainly discussed fully developed channel ﬂow.31 shows why the pressure drop is larger in the former case compared to the latter. consisting of turbulent/non-turbulent motion. or — in other words – why a larger fan is required to push the ﬂow in turbulent ﬂow than in laminar ﬂow. 31). τw ) much larger in turbulent ﬂow than in laminar ﬂow: Eq. the latter is smallest because the turbulent ﬂuctuations are dampened by the wall. ρv2 and ρv3 . Fig. Note that this is smaller than v1 . as we did for laminar ﬂow. τ the ﬂow and the opposing force is the viscous force. 6. As can be seen.7 at p. 32 which reads 0 = p1 Zmax 2δ − p2 Zmax 2δ − 2τw LZmax ¯ ¯ where L is the length of the section. : ρv3 /τw . The ﬂow in a boundary layer is described by Eq.6. 32 and Fig. Reτ = 2000. 6. We get ∆¯ p τw ∂p ¯ = =− L ∂x1 δ (6. We can of course make a force balance for a section of the channel.e. the streamwise stress is largest and the wall-normal stress is smallest.30) As can be seen the pressure drop is directly related to the wall shear stress.36 at p. increases continuously for increasing x1 . Second. 6. DNS ′2 ′2 ′2 data [10.7 presents the normal Reynolds stresses. ρv1 . the outer part of the boundary layer is highly intermittent. 3. Boundary layer 2000 100 57 1500 80 60 x+ 2 1000 x+ 2 40 20 500 a) 0 0 2 4 6 8 b) 0 0 2 4 6 8 Figure 6.

1 (compared to approximately x2 /δ ≃ 0. Boundary layer 2000 58 24 20 16 12 8 1500 x+ 1000 2 500 v2 /uτ ¯ 5 10 15 20 25 4 0 0 a) v2 /uτ ¯ b) 1 10 100 1000 x+ 2 : DNS data [13]. However.3 in channel ﬂow) .8: Velocity proﬁles in a boundary layer along a ﬂat plate. : v2 /uτ = (ln x+ )/0.6.41 + 5. in boundary layer ﬂow the log-law is valid only up to approximately x2 /δ ≃ 0.4. ¯ 2 2 are very similar for the two ﬂows. ¯ : v2 /uτ = x+ .2. Figure 6.

vrms = 0.15.2) Normalize the probability functions. Consider the probability density functions of the ﬂuctuations.e. t 16 18 20 12 14 (b) Point 2.3) Here we integrate over v.1: Time history of v ′ . 7. probability density function is a useful statistical tool to analyze turbulence.5 0 −1 −5 10 −1.4. v ′2 is usually computed by integrating in time. In order to extract more information. the mean velocity is computed as ∞ v= ¯ −∞ vfv (v)dv (7.44. vrms = 1. With a probability density. ∞ v ′2 = −∞ v ′2 fv′ (v ′ )dv ′ As in Eq. see Eq. t 16 18 20 Figure 7. The root-mean-square (RMS) can be ¯ deﬁned from the second moment as vrms = v ′2 1/2 root-meansquare RMS standard deviation variance (7.5 0 5 −0. i. v ′2 = 1 2T T v ′2 (t)dt −T A probability density function is symmetric if positive values are as frequent and large as the negative values. 6. T 1 v= ¯ vdt (7.e.e. (see Eq. vrms = 2. fv . 7 Probability density functions Some statistical information is obtained by forming the mean and second moments. The second moment corresponds to the variance of the ﬂuctuations (or the square of the RMS. 50).23.1). 7. as we do when we deﬁne a time average.1 presents the time history of the v ′ history at . i. Probability density functions 20 15 10 1. for ′2 example v and v2 . of the v velocity. as was done in Section 6. i.5 1 0.1 at p.1) The RMS is the same as the standard deviation which is equal to the square-root of the variance. Figure 7. The mean velocity can of course also be computed by integrating over time.5 10 −4 −6 10 −2 6 4 2 0 59 v ′ 12 14 (a) Point 1. so that ∞ fv (v)dv = 1 −∞ (7.4) 2T −T where T is “sufﬁciently” large. From the velocity signals we can compute the probability densities (sometimes called histograms).7. t 16 18 20 12 14 (c) Point 3. Horizontal red lines show ±vrms .

15 0. the PDF is symmetric).2.05 0 −4 60 fv′ 0. S = 1. so that the skewness.2c conﬁrms that there are many value around zero.5 1 0.1b and 7.2b.5 0 −5 0.77. Point 3.5vrms . S. .09 and F = 7. In Fig. see Figs. The time history of the velocity ﬂuctuation (Fig. of v ′ is deﬁned as Sv ′ = 1 3 vrms ∞ v ′3 fv′ (v ′ )dv ′ = −∞ 1 3 2vrms T T v ′3 (t)dt −T Note that f must be normalized (see Eq. 7. The extreme values of v ′ are approximately ±1.7.2 we can judge whether the PDF is symmetric. The positive values are often larger than +vrms (the peak is actually close to 8vrms ) but the negative values are seldom smaller than −vrms . It is deﬁned as ∞ skewness v ′3 = −∞ v ′3 fv′ (v ′ )dv ′ 3 and is commonly normalized by vrms . are given for the three time histories. this means that the PDF should be symmetric which is conﬁrmed in Fig.68.2b. The PDF function in Fig. The resulting probability densities functions are shown in Fig. This indicates that the distribution of v ′ is skewed towards the positive side.5 2 1.05 0 −2 0 2 4 6 8 10 0 5 −2 0 2 4 v ′ /vrms v ′ /vrms v ′ /vrms (a) Point 1.1c) shows that the ﬂuctuations are clustered around zero and much values are within ±vrms .51 and F = (b) Point 2.2 0.1a) shows that there exists large positive values but no large negative values. The time history shows that the positive and the negative values have the same magnitude. 7. Vertical red lines show ±vrms . The red horizontal lines indicate the RMS value of v ′ . The skewness. Sv′ .2a.1. 2.25 0.1 0. and the ﬂatness.81.3). This is conﬁrmed in the PDF distribution. Probability density functions 3 0. At this point the time history (Fig. see Fig. 5. 7. S = −0.25 0. but instead of “looking” at the probability density functions. Figure 7. S = −0. 7. Point 2. Point 1.e. 7. three different points in a ﬂow (note that v ′ = 0). 7.15 0. The red vertical lines show plus and minus RMS of v ′ .27 and F = (c) Point 3.2: Probability density functions av time histories in Fig. 7. The ﬂuctuations at this point are much smaller and the positive values are as large the negative values. which is the skewness.3 0.1 0. 7. 7. that the extreme value are small and that positive and negative values are equally frequent (i. 7. Let us analyze the data at the three points. F . we should use a deﬁnition of the degree of symmetry.2 2.

The variance (the square of RMS) tells us how large the ﬂuctuations are in average.7.2c. Probability density functions 61 There is yet another statistical quantity which sometimes is used for describing turbulent ﬂuctuations. F = 1 4 vrms ∞ ﬂatness v ′4 fv′ (v)dv −∞ The ﬂuctuations at Point 1 (see Fig. see Fig.2a). 7.1c and the caption in Fig. 7.1a) includes some samples which are very large and hence its ﬂatness is large (see caption in Fig. 7.e. The ﬂatness gives 4 this information. 1 vrms exp − v ′ − vrms 2 2vrms . namely the ﬂatness. whereas the ﬂuctuation for Point 3 all mostly clustered within ±2vrms giving a small ﬂatness. and it is deﬁned computed from v ′4 and normalized by vrms . 7. but it does not tell us if the time history includes few very large ﬂuctuations or if all are rather close to vrms . For a Gaussian distribution f (v ′ ) = for which F = 3. i.

We subtract Eq. ¯ ′ ∂xj ∂xj (8. 8. vi vj = vi vj ¯ ′ ∂xj ∂xj (8. Again. 6. Eq.3) This trick is usually used backwards. Both tricks simply use the product rule for derivative backwards.8. 6. Trick 2: Using the product rule we get 1 1 ∂Ai Ai = 2 ∂xj 2 Ai ∂Ai ∂Ai + Ai ∂xj ∂xj = Ai ∂Ai ∂xj (8. Transport equations for kinetic energy 62 8 Transport equations for kinetic energy In this section and Section 9 we will derive various transport equations. i.4) 8.2) (8. multiply ′ by vi and time average which gives ′ vi ∂ [vi vj − vi vj ] = ¯¯ ∂xj ′ ′ ∂vi vj ′ 1 ′ ∂ ∂2 ′ − vi [p − p] + νvi ¯ [vi − vi ] + ¯ v ρ ∂xi ∂xj ∂xj ∂xj i ′ Using vj = vj + vj .6). we assume incompressible ﬂow (constant density) and constant viscosity (cf.8) .6 we start using ∂¯j /∂xj = 0 v ′ vi ∂ ∂v ′ ′¯ ¯ ′ (vi vj ) = vj vi i ∂xj ∂xj (8.6) ′ Using the continuity equation ∂vj /∂xj = 0 (see Eq.1 The Exact k Equation 1 ′ ′ The equation for turbulent kinetic energy.7) For the second term in Eq. k = 2 vi vi .1) and then we call it the “product rule backwards”.5) ∂ ′ ∂ ′ ′ ′ ′¯ ′ v ¯¯ (¯i + vi )(¯j + vj ) − vi vj = vi v vi vj + vi vj + vi vj . 6. the ﬁrst term is rewritten as ′ vi v ∂ ′ ′ ∂¯i .5 from Eq. 6. is derived from the Navier-Stokes equation.3). Trick 1: Using the product rule we get ∂Ai Bj ∂Bj ∂Ai = Ai + Bj ∂xk ∂xk ∂xk This expression can be re-written as Ai ∂Ai ∂Ai Bj ∂Bj = − Bj ∂xk ∂xk ∂xk (8. There are two tricks which often will be used. Ai ∂Ai 1 ∂Ai Ai = ∂xj 2 ∂xj (8.3 and divide by density.e. the left side can be rewritten as ¯ ′ vi (8.

8. 8. For the ﬁrst term in Eq.9) ′ ¯ The third term in Eq.14 have the following meaning. The large turbulent scales extract energy from the mean ﬂow. 8. II. Equations 8. 8.11. I. IV. see Fig.9 so that ν ∂ ∂xj ′ ∂vi ′ v ∂xj i =ν ∂ ∂xj 1 2 ′ ∂vi ′ ∂v ′ ′ vi + i vi ∂xj ∂xj = (8. ab′ = ab′ = 0.13 give v ∂¯j k v ′ ′ ∂¯i − ∂ = −vi vj ∂xj ∂xj ∂xj I II ′ ∂k 1 ′ ′ 1 ′ ′ ′ ∂v ′ ∂vi vj p + vj vi vi − ν −ν i ρ 2 ∂xj ∂xj ∂xj III IV (8. 8. 8.2.9) 1 ∂ v′ v′ v′ .5 is re-written using the continuity equation (8. i. The two ﬁrst terms represent turbulent diffusion by pressure-velocity ﬂuctuations. we use Trick 2 vj ¯ ′ vi ′ ∂vi ∂xj = vj ¯ ∂ ∂xj 1 ′ ′ vv 2 i i = vj ¯ ∂ ∂ (k) = (¯j k) v ∂xj ∂xj (8. 8. Now we can assemble the transport equation for the ¯ ¯¯ turbulent kinetic energy.12 and 8.e.2 .12 we get the left side and the second term on the right side). The Exact k Equation 63 Next. This term is responsible for transformation of kinetic energy at small scales to thermal energy. 8.6 can be written as (replace vj by vj and use the same technique as in Eq.5 can be written ′ νvi ′ ∂ 2 vi ∂ =ν ∂xj ∂xj ∂xj ′ ∂vi ′ v ∂xj i −ν ′ ′ ∂vi ∂vi ∂xj ∂xj (8. 8.11) The second term on the right side of Eq. Production.5 is zero because it is time averaging of a ﬂuctuation. 2 ∂xj j i i as ′ 1 ′ ∂p′ 1 ∂p′ vi − vi =− ρ ∂xi ρ ∂xi (8. 5. for small wavenumbers.8. and velocity ﬂuctuations. This term (including the minus sign) is almost always positive. respectively. Dissipation.10) The ﬁrst term on the right side of Eq. III.e. 8.14) The terms in Eq. Convection. It is largest for the energy-containing eddies. see Fig. 8. The last term is viscous diffusion. i. P k .12 we use the same trick as in Eq.9. It is largest for large wavenumbers.13) ∂ ν ∂xj 1 2 ′ ′ ∂vi vi ∂xj ′ ′ 1 ∂ 2 vi vi ∂ 2k =ν =ν 2 ∂xj ∂xj ∂xj ∂xj The last term on the right side of Eq.7.12) applying Trick 1 (if we apply the product rule on the ﬁrst term on the right side of Eq. The term (excluding the minus sign) is always positive (it consists of velocity gradients squared). ε. 5.1. 8.

E ∝ vκ /κ. P k . Now we know that the energy spectrum.17) . see Eq. The velocity gradient for an eddy characterized by velocity vκ and lengthscale ℓκ can be estimated as ∂v ∂x ∝ vκ 2 ∝ vκ ℓκ 1/2 κ (8. is equal to the viscous dissipation. ε. How do we know that the majority of the dissipation takes place at the smallest scales? First. How large is ε at wavenumber κ (denoted by ε(κ))? Recall that the viscous dissipation. i.1. 8. its magnitude. ∂¯i /∂xj . Actually. it is stated that the production takes place at the large energy-containing eddies.2. In order to extract energy from the mean ﬂow. time scale etc) In the cascade process (see Section 5. 8. the less they remember the characteristic of mean ﬂow gradient (i.8. ε. direction. must be of the same magnitude. is expressed as the viscosity times the square of the velocity gradient.16) κ 2 since ℓκ ∝ κ−1 . This rev quirement is best satisﬁed by the large scales. we assume that the large eddies contribute much more to the production term more than the small eddies.e.15) where C k . in the cascade process we argue that the smaller the eddies. takes places at the smallest scales. εκ . The Exact k Equation E(κ) ε(κ) εκ 64 εκ+dκ κ dκ κ + dκ κ Figure 8. Dk and ε correspond to terms I-IV in Eq. let us investigate how the time scale varies with eddy size. Consider the inertial subrange. see Fig. The energy that is transferred in spectral space. the time scale of the eddy and the mean velocity gradient. Above.3) we assume that the viscous dissipation. The Reynolds stresses (which appear in P k ) are larger for large eddies than for small eddies 2.14.e.14. ε. follows the −5/3 law in the inertial region which gives ∂v ∂x κ ∝ κ−2/3 1/2 κ ∝ κ−1/3 κ ∝ κ2/3 (8. 5. There are two arguments for this: 1. The transport equation for k can also be written in a simpliﬁed easy-to-read symbolic form as C k = P k + Dk − ε (8.1: Zoom of the energy spectrum for a wavenumber located in Region II or III.

8.1 which shows a zoom of the energy spectrum. τκ = ℓκ ℓ0 2/3 τ0 (8. 5. i. i. For a given spectrum. then ε(κ) is simply obtained through an energy balance. We assume that no mean ﬂow energy production occurs between κ and κ + dκ. then the usual assumption is that εκ ≃ εκ+dκ and that there is no viscous dissipation to internal energy.2. and the “true” viscous dissipation.e. see Fig. εκ = ε = u3 /ℓκ = κ 3 3 3 3 ℓ2 /τκ = ℓ2 /τ0 . the exact k ¯ ¯ equation reads v ∂¯1 k ∂¯2 k v v ′ ′ ∂¯1 + = −v1 v2 ∂x1 ∂x2 ∂x2 ′ ∂k 1 ′ ′ ′ ∂ 1 ′ ′ ∂v ′ ∂vi p v2 + v2 vi vi − ν − −ν i ∂x2 ρ 2 ∂x2 ∂xj ∂xj (8. .8. i. ε. i. the time scale.20) κ which is the same as Eq. also decreases.e. If κ and κ + dκ are located in the inertial region (i. see Eq. ε(κ) does indeed increase for increasing wavenumber. i.e. the −5/3 region).2.e.2 The Exact k Equation: 2D Boundary Layers In 2D boundary-layer ﬂow.e.12. the region may be in the −5/3 region or in the dissipation region.21) 8.e. the cascade process) can also be used for estimating the velocity gradient of an eddy. and the latter is the energy transformed per unit time to internal energy for the entire spectrum (occurring mainly at the small. ε(κ) = εκ+dκ − εκ (8.19) where τ0 and ℓ0 are constants (we have chosen a spectrum). it may be useful to look at the difference between the spectral transfer dissipation εκ . Hence ∂v ∂x −1 −2/3 = τκ ∝ ℓκ ∝ κ2/3 . 8. The cascade process assumes that this energy transfer is the same for each eddy.e. (8. As a ﬁnal note. Now consider Fig.17. The Exact k Equation: 2D Boundary Layers Thus the viscous dissipation at wavenumber κ can be estimated as ε(κ) ∝ ∂v ∂x 2 κ 65 ∝ κ4/3 . τκ . 5. (8. ε(κ) ≃ 0.e. dissipative scales). we ﬁnd from ℓ2 /τκ = ℓ2 /τ0 that κ 0 κ 0 for decreasing eddy size (decreasing ℓκ ). If there is viscous dissipation at wavenumber κ (which indeed is the case if the zoomed region is located in the dissipative region). 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. the former is the energy transferred from eddy to eddy per unit time. The energy transferred from eddy-to-eddy in spectral space (i. Turbulent kinetic energy enters at the left end of the ﬁgure at a rate of εκ and leaves at a rate of εκ+dεκ . This is because the dissipation term is at its largest for small.18) i.22) Note that the dissipation includes all derivatives. for which ∂/∂x2 ≫ ∂/∂x1 and v2 ≪ v1 .

3 0.2b) all terms are negligible except the production term and the dissipation term which balance each other. 11]. 8.4 0. τ Reτ = 2000.e. ▽: −∂v ′ p′ /∂x2 . k is transported in the x2 direction by viscous and turbulent diffusion and it is destroyed (i.8 1 20 0.3a) where the shear stress is zero (see Fig.3b)).02 −0.2a) the other terms do also play a role.1 −0. 8. Closer to the wall (Fig. ◦: ν∂ 2 k/∂x2 .2 0.3: Channel ﬂow at Reτ = 2000. The Exact k Equation: 2D Boundary Layers 0.8 0 1 ∂¯1 /∂x+ v+ 2 (a) Velocity gradient (b) Reynolds shear stress ′ ′ −v1 v2 /u2 τ Figure 8. 8. Since P k is largest here so is also k. : P k. b) Outer region. Figure 8.01 0 −0. 8. 2 40 35 30 25 40 0. Since the region near the wall is dominated by viscosity the turbulent diffusion terms due to pressure and velocity are also small.04 66 a) 0 x+ 2 b) 100 200 300 400 500 x+ 2 Figure 8. 6.04 0.6 0. P k = −v1 v2 ∂¯1 /∂x2 .2: Channel ﬂow at Reτ = 2000. DNS data [10. 8.02 x+ 20 2 15 10 5 0 0 0.2. the former decreases and the magnitude of the latter increases with wall distance and their product takes its maximum at x+ ≃ 2 11.02 0.4 0.2 0. At the wall the turbulent ﬂuctuations are zero which means that the production term is zero. The turbulence kinetic energy is produced by its main source term. a) Zoom near the wall. +: −∂v2 vi vi /2/∂x2 . dissipated) by ε. The left side is – since the ﬂow is fully developed – zero. ′ ′ ′ : −ε.03 0.1 0 −0. the production ′ ′ v term. 8.03 10 20 30 40 −0.22 for fully developed channel ﬂow. Terms in the k equation scaled by u4 /ν.01 0 0. 11].2 0. x2 /δ x+ 2 .2 presents the terms in Eq.6 0. 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.8. Note that the production and the dissipation terms close to the wall are two orders of magnitude larger than in the logarithmic region (Fig.7. see Fig. The velocity gradient is largest at the wall (see Fig.2b).2 0. In the outer region (Fig.01 −0. DNS data [10.

Spatial vs. kκ = vκ. if we integrate the production term over the entire domain. In summary. 0. In homogeneous turbulence the spatial transport terms (i.3 we discussed spectral transfer of turbulent kinetic energy from large eddies to small eddies (which also applies to transport of the Reynolds stresses). and the diffusion terms. In Section 8. the instantaneous ﬂow ﬁeld in turbulent ﬂow is – as we mentioned at the very beginning at p. spectral energy transfer 67 8.1. we get Pk dV > 0 V (8. The reason is that the transfer of turbulent kinetic energy. This should be kept in mind when thinking in terms of the cascade process. the time scale of the cascade 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.3. ′ ∂v1 /∂xi = 0. spectral energy transfer In Section 5.i /2. i.23) .e. 5. care should be taken in non-homogeneous turbulence. for example. from eddy-to-eddy occurs at a much faster rate than the spatial transport by convection and diffusion. How are the spectral transfer and the spatial transport related? The reason that we in Section 5.14) are zero. the convective term. the cascade process is not valid. kκ . We say that the turbulence at these scales is in local equilibrium. 41 – always three-dimensional and unsteady).14) corresponds to the process in which large energy-containing eddies extract energy from the mean ﬂow. V . i. there is spectral transfer of turbulent kinetic energy which takes place in wavenumber space. It may ﬁrst be transported in physical space towards the center. If the channel walls are very long and wide compared to the distance between the walls. turbulent eddy at a position xA (see Fig. term I. Large eddies which extract their energy from the mean ﬂow may not give their energy to the slightly smaller eddies as assumed in Figs. for example ∂v1 /∂xi = 0. Some ﬂows may have one or two homogeneous directions. the cascade process is still a good approximation even in non-homogeneous turbulence.e.e.e. (Note that the derivatives of the instantaneous turbulent ﬂuctuations are non-zero even in homogeneous turbulence.e. 6.8. but kκ may ﬁrst be transported in physical space and then transferred in spectral space. In the inertial range (Region II). The 2 ′ ′ instantaneous turbulent kinetic energy. ∂¯1 /∂x1 = v ′2 /∂x = 0 ∂v ′2 /∂x = 0 etc.1) in fully developed channel ﬂow. In other words. The production term (term II in Eq. the net effect) of the production term is to increase k. real ﬂows are hardly ever homogeneous. term III in Eq.2 and 5. Consider a large. ∂vi 1 3 i In non-homogeneous turbulence. in homogeneous turbulence there is no time-averaged spatial transport. from large eddies to small eddies. 8. ∂k/∂xi = 0 etc. 2δ.14) corresponds to transformation of the turbulent kinetic energy at the small eddies to thermal energy. 8. regarding the validity of the cascade process for the large scales (Region I). 8. Consider. The dissipation term (term IV in Eq.3 Spatial vs. However. of this eddy may either be transferred in wavenumber space or transported in physical (spatial) space.3 only talked about spectral transfer was that we assumed homogeneous turbulence in which the spatial derivatives of time-averaged turbulent quantities are ′2 zero. and there lose its kinetic energy to smaller eddies. However. fully developed channel turbulent ﬂow. i. homogeneous turbulence local equilibrium 8. or both. then the turbulence (and the ﬂow) is homogeneous in the streamwise direction and the spanwise direction.4 The overall effect of the transport terms The overall effect (i. Hence.1 we derived the equation for spatial transport of turbulent kinetic energy.i vκ. however.

27) divergence terms (8. (8.25) V where S is the bounding surface of V . convection and diffusion? Integration of the convection term over V gives. ¯ ∂xj ρ ∂xi ∂xj ∂xj ∂xj vi ¯ (8.8.29 can be written as ∂¯i p v ¯ ∂p ¯ =− . we get ∂ 1 ′ ∂k 1 ′ ′ ′ vj vk vk + p′ vj − ν V 2 ρ ∂xj V ∂xj 1 ′ ′ ′ 1 ′ ∂k nj dS =− vj vk vk + p′ vj − ν 2 ρ ∂xj S − (8.5. Eqs.5. Hence.26) The only net contribution occurs at the boundaries. −εdV < 0 68 (8.25 and 8. This shows that the net effect of the convection term occurs only at the boundaries. but in the remaining part of the domain it acts as an equally large sink term. 6.e. Mathematically these terms are called divergence terms. i. the convection merely transports k with out adding or subtracting anything to k. using Gauss divergence law ∂¯j k v dV = ∂xj vj knj dS ¯ S (8. i. the net effect of the dissipation term is a negative contribution.30) Using the continuity equation. 8. the ﬁrst term on the right side of Eq.29) Using the continuity equation and Trick 2 the term on the left side can be rewritten as vi ¯ ∂¯i vj v¯ ∂¯i v 1 ∂¯i vi v¯ ∂¯j K v = vj vi ¯ ¯ = vj = ¯ ∂xj ∂xj 2 ∂xj ∂xj (8.e.5 The transport equation for vi vi /2 ¯¯ ′ ′ The equation for K = vi vi /2 is derived in the same way as that for vi vi /2.31) −¯i v ∂xi ∂xi . i. they can both be written as the divergence of a vector Aj . Multiply ¯¯ the time-averaged Navier-Stokes equations. the convection acts as a source term in part of the domain. by vi so that ¯ ′ ′ ∂vi vj ∂¯i vj v¯ ∂ 2 vi ¯ 1 ∂p ¯ = − vi + ν¯i v − vi ¯ . Similarly for the diffusion term.24) V What about the transport terms. Eq. ∂Aj ∂xj where Aj for the convection and the diffusion term reads vj k ¯ convection term 1 ′ ′ ′ 1 ′ ′ ∂k Aj = diffusion term v v v + pv −ν − 2 j k k ρ j ∂xj (8.28) 8.26 show that the transport terms only – as the word implies – transports k without giving any net effect except at the boundaries.e. Inside the domain. 8. The transport equation for vi vi /2 ¯¯ Similarly.

DNS data [10.5. . 8. 11]. 8.4: Channel ﬂow at Reτ = 2000. 8. =− + vi vj ∂xj ∂xj ∂xj (8. 8.2 0 0 20 40 60 80 100 x+ 2 Figure 8.35 is the same as the ﬁrst term in Eq.32 inserted in Eq.e. v : ε.13.14 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. 8.6 0. Inserted in Eq. ν¯i v ∂K ∂¯i ∂¯i v v ∂ 2 vi ¯ =ν −ν . εmean Note that the ﬁrst term in Eq.33) The last term is rewritten using Trick 1 as −¯i v ′ ′ ∂¯i vi vj v ′ ′ ∂vi vj v ′ ′ ∂¯i . ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj (8.33 gives (cf.32) Equations 8. see Eqs. The viscous term in Eq. sink ∂K 1 vj p + vi vi vj − ν ¯ ¯ ¯ ′ ′ ρ ∂xj −ν ∂¯i ∂¯i v v ∂xj ∂xj (8.29 gives ′ ′ ∂vi vj v ¯ ∂2K 1 ∂¯i p ∂¯i ∂¯i v v ∂¯j K v =ν − −ν − vi ¯ .4 0. ∂xj ∂xj ∂xj ρ ∂xi ∂xj ∂xj ∂xj (8.31 and 8.35) εmean . 8. On the right side we ﬁnd: • loss of energy to k due to the production term • diffusion by pressure-velocity interaction • diffusion by velocity-stress interaction • viscous diffusion • viscous dissipation. i.1. 8.8.14) ¯ ∂¯j K v v ′ ′ ∂¯i − ∂ = vi vj ∂xj ∂xj ∂xj −P k .30.8 0. Both terms are normalized by u4 /ν. : τ ν(∂¯1 /∂x2 )2 . 8. Eq.34) Note that this term differs to the corresponding term in Eq.14 by a factor of two since ′ “Trick 2” cannot be used because vi = vi .29 is rewritten in the same way as the viscous term in Section 8.12 and 8. sink On the left side we have the usual convective term. 8. Comparison of mean and ﬂuctuating dissipation terms. The transport equation for vi vi /2 ¯¯ 69 1 0.

ν(∂¯1 /∂x2 )2 . In the K equation the dissipation term and the negative production term (representing loss of kinetic energy to the k ﬁeld) read −ν v ∂¯i ∂¯i v v ′ ′ ∂¯i . K = 1 1 ′ ′ ¯¯ 2 vi vi and k = 2 vi vi .5: Transfer of energy between mean kinetic energy (K). is much v larger than ε. ∆T ). The energy ﬂow from the mean ﬂow to internal energy is illustrated in Fig. is much smoother than that of the ¯ ′ ﬂuctuating velocity ﬁeld. τ . turbulent kinetic energy (k) and internal energy (denoted as an increase in temperature. + vi vj ∂xj ∂xj ∂xj (8.36). vi . the dissipation by the ﬂuctuations. In the viscous-dominated region. vi . 8.37) The gradient of the time-averaged velocity ﬁeld. The transport equation for vi vi /2 ¯¯ 70 ∆T εm n ea ε K Pk k Figure 8. In fully turbulent ﬂow. the mean dissipation takes the value ν = 1/2000 (normalized by u4 /ν).5. 8. is much larger than the dissipation by the mean ﬂow (left side of Eq. 8. ε. the mean dissipation.5. The major part of the energy ﬂow goes from K to k and then to dissipation. This is seen in Fig.8.4.36) and in the k equation the production and the dissipation terms read ′ ′ −vi vj ′ ∂v ′ ∂vi ∂¯i v −ν i ∂xj ∂xj ∂xj (8. At the wall.

1.3 and ′ divide by density. some simpliﬁcations are introduced.3) + ′ ′ ′ ′ ∂vj vk ′ ∂vi vk ′ vi + v ∂xk ∂xk j Note that each line in the equation is symmetric: if you switch indices i and j in any ′ ′ of the terms nothing changes. i. This means ′ ′ ′2 that k is equal to twice the sum of the diagonal components of vi vj . k.5 is written with the index i as free index. k = vi vi /2.2 together gives ′ vj ∂ ′ ∂ [v v − v v ] = [vi vk − vi vk ] + vi ¯¯ ¯j ¯k j k ∂xk ∂xk 1 ′ ∂p′ 1 ′ ∂p′ − vi − vj ρ ∂xj ρ ∂xi ′ +νvi ′ 2 ′ ∂ 2 vj ′ ∂ vi + νvj ∂xk ∂xk ∂xk ∂xk (9. 2 or 3 so that the equation is an equation for v1 . no transport equation ′ ′ for vi vj is solved. which must ′ ′ be known before Eq. . v1 2 2 1 i j is modelled by expressing it in a turbulent viscosity and a velocity gradient. This is computationally expensive since ′ ′ we then need to solve six additional transport equations (recall that vi vj is symmetric. We get ′ vi ∂ [vj vk − vj vk ] = ¯ ¯ ∂xk ′ ′ ′ ∂vj vk ′ ∂ 2 vj 1 ′ ∂ ′ ′ − vi p + νvi + v ρ ∂xj ∂xk ∂xk ∂xk i (9.1) Equation 6.2 is conveniently obtained from Eq.e. 6. in which v ′ v ′ i. Adding Eqs. This is an unknown in the time-averaged Navier-Stokes equations. to solve a transport equation for it.e.1 by simply switching indices i and j. Steady. The most accurate way to ﬁnd vi vj is. 9. Twoequations models are commonly used in these simpliﬁed models.) Often.9. This approach is very similar to that we used when deriving the k equation in Section 8. Transport equations for Reynolds stresses 71 9 Transport equations for Reynolds stresses In Section 8 we derived transport equations for kinetic turbulent energy. multiply by vj and time average and we obtain ′ vj ∂ [vi vk − vi vk ] = ¯¯ ∂xk 2 ′ ∂v ′ v ′ ′ 1 ′ ∂ ′ ′ ∂ vi p + νvj + i k vj − vj ρ ∂xi ∂xk ∂xk ∂xk (9. i = 1. v2 or v3 .e. Eq. which is the ′ ′ ′ ′ trace of the Reynolds stress tensor vi vj divided by two. This is important since the tensor vi vj is symmetric. i.2) It may be noted that Eq. of course. 9. ′ v ′ = v ′ v ′ etc. k = 0.5 from Eq. i. This is the subject of Turbulence Modelling which you will learn about in other courses in the MSc programme. incompressible ﬂow with constant density and viscosity is assumed.e. ′ ′ Now let’s start to derive the transport equation for vi vj .5(v1 + ′2 ′2 v2 +v3 ). 6. 6.5. Subtract Eq.5 can be solved. Here we will now derive the transport equation for the Reynolds stress tensor.5 as an equation for vj and ′ multiply this equation by vi .1 and 9. 6. 6. 9. Now write Eq.

3 is also re-written using Trick 1 − ′ ′ 1 ∂vj 1 ∂ ′ ′ 1 ∂ ′ ′ 1 ′ ∂vi + p′ vi p − vj p + p ρ ∂xj ρ ∂xi ρ ∂xj ρ ∂xi (9.5) We merge the second terms in the two groups in Eq. 9.4. Therefore we re-write the derivative in the two ﬁrst terms as ∂ ∂ ∂ ∂ = δjk and = δik ∂xj ∂xk ∂xi ∂xk so that −δjk ′ ′ 1 ∂ ′ ′ 1 ∂ ′ ′ 1 ′ ∂vi 1 ∂vj vi p − δik vj p + p + p′ ρ ∂xk ρ ∂xk ρ ∂xj ρ ∂xi (9. ′ Using vi = vi + vi . but usually we use the same dummy index in every term.7) The second line in Eq.8) It will later turn out that it is convenient to express all derivatives as ∂/∂xk . Indices i and j appear once in each term (not more and not less) and index k (the dummy index) appears exactly twice in each term (implying summation). 9. 9.4 in the same way.4) Using the continuity equation the ﬁrst terms in the two groups are rewritten as ′ ′ vj vk ∂¯i v v ′ ′ ∂¯j + vi vk ∂xk ∂xk (9. for example. replace k with m in the ﬁrst term and with q in the second term. 9. Re-writing also the third terms in the two groups in Eq.3 is also re-written using Trick 1 ν ′ ∂vj ∂ ′ vi ∂xk ∂xk +ν ′ ∂ ′ ∂v vj i ∂xk ∂xk − 2ν ′ ′ ∂vi ∂vj ∂xk ∂xk .9) (9. the ﬁrst line can be rewritten as ¯ ′ vj ∂ ′ ′ ′¯ ′ ′ ′¯ ′ ∂ [¯i vk + vi vk + vi vk ] + vi v ′ vj vk + vj vk + vj vk ¯ ′ ∂xk ∂xk (9.9. You could. it is permissible. ′ vj ′¯ ′ ′¯ ∂vj vk ∂vj ∂vi vk ∂v ′ ′ + vi = vk vj i + vk vi ¯ ′ ¯ ′ ∂xk ∂xk ∂xk ∂xk ′ ′ ′ ′¯ ∂vi vj ∂vi vj vk = vk ¯ = ∂xk ∂xk (9. Note that it is correct to use any other index than k in some terms (but you must not use i and j).4 can be written ′ ′¯ ′ ′ ′ ∂vi vj vk ∂vi vj vk + ∂xk ∂xk (9. the second and the third terms in Eq. Transport equations for Reynolds stresses 72 Furthermore.10) The third line in Eq. you can check that the equation is correct according to the tensor notation rules.6) The continuity equation was used twice (to get the right side on the ﬁrst line and to get the ﬁnal expression) and the the product rule was used backwards to get the second line. 9.

Put the ﬁrst term in Eq.11 so that ∂ v v ′ ′ ∂¯i − v ′ v ′ ∂¯j (¯k vi vj ) = −vj vk v ′ ′ i k ∂xk ∂xk ∂xk I II − ∂ ∂xk ′ ′ ∂vi vj 1 1 ′ ′ ′ ′ ′ vi vj vk + δjk vi p′ + δik vj p′ − ν ρ ρ ∂xk III (9. • Every term – or group of terms – should be symmetric in i and j. Since it is an equation for a second-order tensor. 9. An alternative – and maybe easier – way to derive Eq. a diffusion term Reynolds stress equations .3 are zero because ab′ = ab′ = 0.14.9 allows the diffusion (term III) to be written on a more compact form. 9. 8. vi vj . • Every term – or group of terms – should include the free indices i and j (only once). 9.12 is the (exact) transport equation of the Reynolds stress. It is called the the Reynolds stress equations.9.10 and 9.12) 1 + p′ ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi V − 2ν ′ ′ ∂vi ∂vj ∂xk ∂xk IV Note that the manipulation in Eq. 9.13) (9.3 reads ν ∂ =ν ∂xx ′ ′ ∂vi vj ∂xk ∂ ∂xk ′ vi ′ ′ ∂vj ′ ∂v + vj i ∂xk ∂xk − 2ν ′ ′ ∂vi ∂vj ∂xk ∂xk ′ ′ ′ ′ ∂ 2 vi vj ∂v ′ ∂vj ∂v ′ ∂vj − 2ν i =ν − 2ν i ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk (9.12 with the equation for turbulent kinetic energy. 9. 9. • A dummy index (in this case index k) must appear exactly twice (=summation) in every term Equation 9. it is always useful to check that the equation is correct according to the tensor notation rules. Eq.12 and then take the trace (setting i = j) and dividing by two.7 on the left side and the second term on the right side together with Eqs. After a derivation. Transport equations for Reynolds stresses 73 Trick 1 is used – again – to merge the two ﬁrst terms so that the third line in Eq. 8. Compare Eq. a production term (II). but since it is symmetric we only need to consider six of them.14 is to ﬁrst derive Eq.11) ¯ ¯¯ The terms on the fourth line in Eq.14) ′ ′ Equation 9.5. In both the k and the ′ ′ vi vj equations there is a convection term (I). it consists of nine equations.12 can also be written in a simpliﬁed easy-to-read symbolic form as Cij = Pij + Dij + Πij − εij where Πij denotes the pressure-strain term Πij = p′ ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi (9. 9. 9. We can now put everything together.

9.1: Channel ﬂow at Reτ = 2000. b) Outer region.15) because of the continuity equation and this is the reason why this term does not appear in the k equation.16 reads ′ ′ Pij = −vj v2 ∂¯i v v ′ ′ ∂¯j − vi v2 ∂x2 ∂x2 (9. Π11 . For a normal stress that is larger ′2 .1 x+ 2 b) 100 200 300 400 500 x+ 2 ′2 Figure 9. ′ ′ ′2 The average of the normal stresses is vav = vi vi /3.05 a) −0. 9.5 0 10 20 30 40 −0.e. It is often called the Robin than vav Hood term because it – as Robin Hood – “takes from the rich and gives to the poor”. : −ε11 .17) The second line shows that it is the streamwise and spanwise derivative that operate on time-averaged quantities that are zero. The physical meaning of this term is to redistribute energy between the normal stress components (if we transform Eq.16) − 2ν ′ ′ ∂vi ∂vj ∂xk ∂xk Now let’s look at this equation for fully developed channel ﬂow for which v2 = v3 = 0 ¯ ¯ ∂(·) ∂(·) = =0 ∂x1 ∂x3 (9. : ′ v ′2 )/∂x . : P11 . pressure strain Robin Hood Πii = 1 ′ p ρ ′ ∂vi ∂v ′ + i ∂xi ∂xi =0 (9. For 2D boundary layer ﬂow. The production term in Eq. 9. see Eq. i. Eq. only normal stresses). DNS data [10. 9. ◦: ν∂ 2 v ′2 /∂x2 .18) . Transport equations for Reynolds stresses 0.14. not those that operate on instantaneous quantities such as in εij and Πij .12 ′ ′ to the principal coordinates of vi vj there are no shear stresses.12 reads ∂ v v ∂ ′ ′ ∂¯i − v ′ v ′ ∂¯j (¯1 vi vj ) + v ′ ′ (¯2 vi vj ) = −vj v2 v ′ ′ i 2 ∂x1 ∂x2 ∂x2 ∂x2 ′ ′ ∂vi vj ∂ 1 1 ′ ′ ′ ′ ′ − vi vj v2 + δj2 vi p′ + δi2 vj p′ − ν ∂x2 ρ ρ ∂x2 1 + p′ ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi (9. τ a) Zoom near the wall. which is called the pressure strain term. 9. In the vi vj equation there is a ﬁfth term (V).5 0. 11].05 0 0 −0. the pressure-strain term is negative and vice-versa. Note that the trace of the pressure-strain term is zero. Terms in the v1 equation scaled by u4 /ν.1 74 0. +: −∂(v2 1 2 2 1 ′ ′ (III) and a dissipation term (IV).

17.20) ε12 = 2ν 1 2 = 0 ∂xk ∂xk However. 9. 9. Π22 . P12 = −v2 ∂¯1 /∂x2 .1b) the production term balances the pressure-strain term and the dissipation term.3. Note that in the upper half of the channel ∂¯1 /∂x2 < 0 and hence P12 v v ′ ′ and v1 v2 are positive. P11 . see Fig. the other way around: ′ ′ ′2 v v1 v2 is negative because its production term. for i = j.3 presents the terms in the v1 v2 equation. also their general.4. v2 . For channel ﬂow.1 Reynolds shear stress vs. the Reynolds shear stress and the velocity gradient ∂¯1 /∂x2 have nearly always opposite signs.19b) (9. is zero. ′ ′ Figure 9.19a) (9. This can also be shown by v physical argumentation.19d) using Eq. v1 v2 = 0 (in ′ v ′ in isotropic turbulence are zero). e. Reynolds shear stress vs.19 v ′ ′ shows that P12 is negative (and hence also v1 v2 ) in the lower half because ∂¯1 /∂x2 > 0 v and it is positive in the upper half because ∂¯1 /∂x2 < 0. Eq. Hence.se/˜lada/allpapers. Furthermore. see Fig.html 9.1. 9. 9. acts as the main source term. for example. 6. ε12 .tfd. or. very close to the wall. A ﬂuid . If you want to learn more how to derive transport equations of turbulent quantities. ε12 is positive since v1 v2 < 0. Π22 – the “Robin Hood” term – takes from the “rich” v1 ′2 ′2 ′2 equation and gives to the “poor” v2 equation energy because v1 is large and v2 is small. rather. the velocity gradient In boundary-layer type of ﬂow. and the dissipation term act as sink terms. The production term – which ′ ′ should be a source term – is here negative. note that the dissipation. 9. Indeed it should be. v3 (i = j = 3) and v1 v2 (i = 1. x+ ≤ 10. In the outer region (Fig. Recall that v1 v2 is here negative and hence its source must be negative. j = 2) equations we get ′ ′ P11 = −2v1 v2 P22 P33 P12 ∂¯1 v ∂x2 v ′ ′ ∂¯2 = 0 = −2v2 v2 ∂x2 v ′ ′ ∂¯3 = 0 = −2v3 v2 ∂x2 ∂¯1 v v ′ ′ v − v ′ v ′ ∂¯2 = −v ′2 ∂¯1 = −v2 v2 1 2 2 ∂x2 ∂x2 ∂x2 (9. ε12 = 0 because here the wall affects 2 ′ ′ the dissipative scales making them non-isotropic. v2 (i = j = 2). This is because dissipation takes place at the smallest scales and they are isotropic. ′2 Figure 9.7).16 with i = j = 1). see [14] which can be downloaded here http://www. As we saw for the k equation.1 presents the terms in the v1 equation (Eq. is negative since ∂¯1 /∂x2 > 0. Consider the ﬂow in a boundary layer. Π11 .chalmers.g. 9.2. That implies ′ ′ there is no correlation between two ﬂuctuating velocity components.19c) (9. the production term. reaches its maximum at x2 ≃ 11 ′2 where also v1 takes its maximum (Fig.9. the velocity gradient 75 ′2 ′2 ′2 ′ ′ For the v1 (i = j = 1). the stresses vi j gradients are zero so that ∂v ′ ∂v ′ (9. 9. The pressure-strain term. ′2 The terms in the wall-normal stress equation. ′2 As mentioned previously. are shown in Fig. Here we ﬁnd – as expected – that the pressure-strain term.

in a wall jet) the reason is that the other terms (usually the transport terms) are more important than the production term.B ).01 −0. v1 v2 < 0. v1 (x2.05 0 −0.e.A ) < v1 (x2.05 0 0. . This means that when the particle ¯ ¯ at x2. 9.005 0 0 −0. If we study this ﬂow for a long time and ′ ′ average over time we get v1 v2 < 0.A (where the streamwise velocity is v1 (x2. τ : P12 .15 0. b) Outer region.B )) comes down to x2.4) we reach the same con′ clusion. : −ε22 .05 0. i.3: Channel ﬂow at Reτ = 2000. Thus the streamwise ﬂuctuation is positive. 2 0.05 −0. 11]. ◦: ν∂ 2 v1 v2 /∂x2 . 2 particle is moving downwards (particle drawn with solid line) from x2. Reynolds shear stress vs.015 76 a) −0.02 x+ 2 b) 100 200 300 400 500 x+ 2 ′2 Figure 9. +: −∂(v1 v2 )/∂x2 . ◦: ν∂ 2 v2 /∂x2 . Thus.2: Channel ﬂow at Reτ = 2000. the velocity gradient 0. Terms in the v1 v2 equation scaled by u4 /ν. i. DNS data [10. If we change the sign of the velocity gradient so ′ ′ that ∂¯1 /∂x2 < 0 we will ﬁnd that the sign of v1 v2 also changes. again. At its new location the v1 velocity is in average smaller than at its old.05 a) 0 10 20 30 40 x+ 2 b) 100 200 300 400 500 x+ 2 ′ ′ Figure 9. DNS data [10. b) Outer region.02 0. Terms in the v2 equation scaled by u4 /ν.B to x2. and it is bringing a deﬁcit in v1 ′ ′ ′ so that v1 < 0.A )) it has an excess of streamwise velocity compared to ′ its new environment at x2.1 −0.9. : P22 . and the correlation between v1 and v2 is in average negative (v1 2 If we look at the other particle (dashed line in Fig.01 0. v1 > 0 ′ ′ ′ v ′ < 0). +: −∂(v2 v2 )/∂x2 . v In cases where the shear stress and the velocity gradient have the same sign (for example. The particle is moving upwards (v2 > 0). : Π22 . a) Zoom near the wall.005 −0.015 0. ′ ′ ′2 ′2 ▽ : −2∂v2 p′ /∂x2 .05 0 10 20 30 40 −0.1. : −ε12 .A with ′ (the turbulent ﬂuctuating) velocity v2 .e. 11]. ▽ : −∂v2 p′ /∂x2 .A .B (which has streamwise velocity v1 (x2.1 0. ′ ′ ′2 ′ ′ : Π12 . τ a) Zoom near the wall.

then B11 will go to zero. xC ) = v1 (xA )v1 (xC ) (10. then B11 = v ′2 (xA ). and sample the ﬂuctuating 1 1 ′ velocity.10. say xA and xC . x1 ) = v1 (xA )v1 (xA + x1 ) 1 ˆ 1 1 (10. B Figure 10.1: Two-point correlation. we expect that the two-point correlation function will be related to the largest eddies.1) 1 1 1 1 Often. the x1 direction. 1 on the other hand.B ′ v2 < 0 ′ v2 > 0 x2. in. |xA − xC | 1 1 10 Correlations 10.A v1 (x2 ) ¯ x2 x1 ′ ′ Figure 9. we move point C further and further away from point A. ˆ 1 1 It is obvious that if we move point A and C closer to each other. it is expressed as ′ ′ ˆ B11 (xA . for example. It is convenient to normalize B11 so that it varies between . Correlations 77 x2. B11 increases.2) where x1 = xC − xA is the separation distance between point A and C. Furthermore. We can then form the correlation of v1 at these two points as ′ ′ B11 (xA . Pick two points along the x1 axis.4: Sign of the Reynolds shear stress −ρv1 v2 in a boundary layer.1 Two-point correlations Two-point correlations are useful when describing some characteristics of the turbulence. when the two points are moved so close that they merge. If.

10.1. Two-point correlations

78

xA 1

Lint

B

x1

(a) Small integral length scale

xA 1 Lint

B

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

−1 and +1. The normalized two-point correlation reads

norm B11 (xA , x1 ) = 1 ˆ

79

1 v1,rms (xA )v1,rms (xA 1 1 + x1 ) ˆ

′ ′ v1 (xA )v1 (xA + x1 ) ˆ 1 1

(10.3)

′ where subscript rms denotes root-mean-square, which for v1 , for example, is deﬁned as ′2 v1,rms = v1 1/2

(10.4)

RMS is the same as standard deviation (Matlab command std) which is the squareroot of the variance (Matlab command var). Consider a ﬂow where the largest eddies have an eddy scale (length scale) of Lint , see Fig. 10.2. We expect that the two point correlation, B11 , approaches zero for separation distance, |xA −xC | > Lint because for separation distances larger than |xA −xB | 1 1 1 1 ′ ′ there is no correlation between v1 (xA ) and v1 (xC ). Hence, ﬂows with large eddies will 1 1 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 x1 . ˆ If the ﬂow is homogeneous (see p.67) in the x1 direction, the two-point correlation does not depend on the location of xA , i.e. it is only dependent on the separation of the 1 two points, x1 . ˆ From the two-point correlation, B11 , an integral length scale, Lint which is deﬁned as the integral of B11 over the separation distance, i.e.

∞

integral length scale

Lint (x1 ) =

0

B11 (x1 , x1 ) ˆ dˆ1 x A C v1,rms v1,rms

(10.5)

If the ﬂow is homogeneous in the x1 direction then Lint does not depend on x1 .

**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 ′ ′ ˆ ˆ B11 (tA , t ) = v1 (tA )v1 (tA + t ) (10.6) ˆ where t = 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 ˆ this case the auto-correlation depends only on time separation, t, i.e.

′ ′ ˆ ˆ B11 (t ) = v1 (t)v1 (t + t )

(10.7)

**where the right side is time-averaged over t. The normalized auto-correlation reads
**

norm ˆ B11 (t ) =

1 ′ ˆ v ′ (t)v1 (t + t ) 2 v1,rms 1

(10.8) integral time scale

In analogy to the integral length scale, Lint , the integral time scale, Tint , is deﬁned as (assuming steady ﬂow)

∞

Tint =

0

norm ˆ ˆ B11 (t)dt

(10.9)

10.2. Auto correlation

80

**MTF270 Turbulence Modelling
**

L. Davidson

Division of Fluid Dynamics, Department of Applied Mechanics Chalmers University of Technology, G¨ teborg, Sweden o http://www.tfd.chalmers.se/˜lada, lada@chalmers.se

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

81

**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 averaged) and ﬂuctuating part. For the velocity vector this means that vi is divided into ′ ′ a mean part vi and a ﬂuctuating part vi so that vi = vi + vi . Time average and we get ¯ ¯ (see Eq. 6.4 at. p. 50): ∂¯i v =0 ∂xj ∂p ¯ ∂ 2 vi ¯ ∂τij ∂ ∂ρ0 vi ¯ ¯ (ρ0 vi vj ) = − ¯¯ +µ − − βρ0 (θ − θ0 )gi + ∂t ∂xj ∂xi ∂xj ∂xj ∂xj (11.1) (11.2)

(note that θ denotes temperature) where ρ0 is a constant reference density and fi = ¯ −β(θ − θ0 )gi and the turbulent stress tensor (also called Reynolds stress tensor) is written as: ′ ′ τij = ρ0 vi vj The pressure, p, denotes the hydro-static pressure, see Eq. 3.22, which means that ¯ when the ﬂow is still (i.e. vi ≡ 0), then the pressure is zero (i.e. p ≡ 0). ¯ ¯ The body force fi – which was omitted for convenience in Eq. 6.4 – has here been 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 acceleration. Since the pressure, p, is deﬁned as the hydro-static pressure we must rewrite ¯ the buoyancy source as ρ0 fi = (ρ − ρ0 )gi (11.3)

Reynolds stress tensor

so that p ≡ 0 when vi ≡ 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 temperature, differentiation gives dρ = ∂ρ ∂θ dθ +

p

∂ρ ∂p

dp

θ

(11.4)

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. 11.4 is zero and we introduce the volumetric thermal expansion, β, so that β=− 1 ρ0 ∂ρ ∂θ ⇒ (11.5)

p

dρ = −ρ0 βdθ ⇒ ρ − ρ0 = −βρ0 (θ − θ0 ) where β is a physical property which is tabulated in physical handbooks. For a perfekt gas it is simply β = θ−1 (with θ in degrees Kelvin). Now we can re-write the buoyancy source as ¯ ρ0 fi = (ρ − ρ0 )gi = −ρ0 β(θ − θ0 )gi (11.6) 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. 11.6 results in a force vertically upwards, which agrees well with our intuition.

′ ′ 11.2. The exact vi vj equation

82

11.1.2 Temperature equation The instantaneous temperature, θ, is also decomposed into a mean and a ﬂuctuating ¯ component as θ = θ + θ′ . The transport equation for θ reads (see Eq. 2.15 where temperature was denoted by T ) ∂2θ ∂θ ∂vi θ =α + ∂t ∂xi ∂xi ∂xi ¯ where α = k/(ρcp ), see Eq.2.15 on p. 21. Introducing θ = θ + θ′ we get ¯ ¯ ∂ θ ∂¯i θ ∂2θ ∂v ′ θ′ v ¯ =α − i + ∂t ∂xi ∂xi ∂xi ∂xi (11.8) (11.7)

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. 11.8 reads (cf. Eq. 2.11) ¯ qi,tot qi qi,turb ∂θ ′ = + = −α − vi θ′ ρcp ρcp ρcp ∂xi

′ ′ 11.2 The exact vi vj equation

(11.9)

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 (¯i , p) plus six turbulent stresses, vi vj . We must close this equation system; it is v ¯ called the closure problem. We must ﬁnd some new equations for the turbulent stresses. We need a turbulence model. 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 = vi + vi → ¯ Eq. (A)

closure problem

**• Time average → equation for vi , 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. 71 these steps are given in some detail. More details can also be a found in [14] (set the SGS tensor to zero, i.e. τij = 0).

3 The exact vi θ′ equation If temperature variations occurs we must solve for the mean temperature ﬁeld. see Eq.10 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 ′ 11. The exact vi θ′ equation 83 ′ ′ The ﬁnal vi vj -equation (Reynolds Stress equation) reads (see Eq. 11. 9. subtract the equation for the mean velocity vi (Eq.12) ′ ′ ′ ′ ′ ∂vi vj ∂vi vj v v ′ ′ ∂¯j − v ′ v ′ ∂¯i + p = −vi vk + vk ¯ j k ∂t ∂xk ∂xk ∂xk ρ Cij Pij ′ ′ ∂vj ∂vi + ∂xj ∂xi Πij ′ ′ ∂ 2 vi vj − ′ ′ p′ vj p′ vi ∂ ′ ′ ′ vi vj vk + δik + δjk + ν ∂xk ρ ρ ∂xk ∂xk Dij. The total diffusion reads Dij = Dij.12) + ∂t ∂xk ρ ∂xi ∂xk ∂xk ∂xk .3. vi . 11. vi (Eq.8.t and Dij.ν .2) from the equation for the instantaneous velocity.3) so ¯ that ′ ′ ∂v ′ v ′ 1 ∂p′ ∂ 2 vi ∂ ∂vi ′ ′ ′ (vk vi + vk vi + vk vi ) = − ¯ ¯ ′ +ν + i k − gi βθ′ (11. 6.10) ′ ′ −gi βvj θ′ − gj βvi θ′ − 2ν Gij where Dij.ν denote turbulent and viscous diffusion.8 from Eq. Equation 11. subtract Eq.t ′ ′ ∂vi ∂vj ∂xk ∂xk εij (11.7 ∂v ′ θ′ ∂θ′ ∂ 2 θ′ ∂ ′ ¯ ′ (vk θ + vk θ′ + vk θ′ ) = α ¯ + k + ∂t ∂xk ∂xk ∂xk ∂xk (11.t + Dij. 11. To obtain the equation for the ﬂuctuating temperature. respectively.11) ′ To get the equation for the ﬂuctuating velocity.′ 11.ν D ij. This is analogous to the momentum equation where we have gradients of viscous and turbulent stresses which correspond to viscous and turbulent diffusion. 11.

′ ′ ′ ′ ∂vi vj ∂vi vj θ′ = 0 θ′ = ∂xk ∂xk The ﬁrst term in the two parentheses on line 1 in Eq. 11.13) The Reynolds stress term in Eq.13 are re-written using the continuity equation ′ vi ∂¯k vi v ′ ∂¯k θ′ v + θ′ = vk ¯ ∂xk ∂xk ′ vi ∂θ′ ∂v ′ + θ′ i ∂xk ∂xk (11.13 are combined into two ′ production terms (using the continuity equation.18 into Eq.18) Inserting Eqs.16. The last two terms in Eq.14.17) ∂xk The viscos diffusion terms on the right side are re-written using the product rule backwards (Trick 1.15 and 11. The exact vi θ′ equation 84 ′ Multiply Eq. i.ν ∂ ∂xk ∂v ′ θ′ i ∂xk − (ν + α) ′ ∂vi ∂θ′ (11.15) Now the two terms can be merged (product rule backwards) vk ¯ ′ ∂vi θ′ ∂¯k vi θ′ v ′ = ∂xk ∂xk (11.19) ∂xk ∂xk −gi βθ′2 Giθ εiθ .16 ′ ′ ∂vi vk θ′ (11. 11.16) where we used the continuity equation to obtain the right side.′ 11. 11.11 with vi and multiply Eq. 11. 11. 11. add them together and time average ′ ∂vi θ′ ′ ∂ (v ′ θ + v θ ′ + v ′ θ ′ ) + θ ′ ∂ (¯ v ′ + v v ′ + v ′ v ′ ) ¯ ¯k vi k ¯k i + vi i k k ∂t ∂xk k ∂xk ′ 2 ′ θ′ ∂p′ ∂ 2 vi ′ ∂ θ =− + αvi + νθ′ − gi βθ′ θ′ ρ ∂xi ∂xk ∂xk ∂xk ∂xk (11.17 and 11.14) The second term in the two parenthesis on the ﬁrst line of Eq. 11. ∂vk /∂xk = 0) ′ ′ vi vk ¯ ∂θ ∂¯ v ′ + vk θ′ ∂xk ∂xk (11. see p.12 multiplied by θ′ and time averaged is zero. 62) ′ αvi ∂ 2 θ′ ′ ∂ = αvi ∂xk ∂xk ∂xk ′ ∂ 2 vi ∂ = νθ′ ∂xk ∂xk ∂xk ∂θ′ ∂xk ′ ∂vi ∂xk =α ∂ ∂xk ′ vi ∂θ′ ∂xk −α ′ ∂θ′ ∂vk ∂xk ∂xk νθ′ ∂ =ν ∂xk θ′ ′ ∂vi ∂xk ′ ∂θ′ ∂vk −ν ∂xk ∂xk (11. 11.13 gives the transport ′ equation for the heat ﬂux vector vi θ′ ′ ′ ′ ¯ v ∂ ∂vi θ′ ′ ′ ∂ θ − v ′ θ ′ ∂¯i − θ ∂p − ∂ v ′ v ′ θ ′ vk vi θ′ = −vi vk ¯ ′ + k ∂t ∂xk ∂xk ∂xk ρ ∂xi ∂xk k i Piθ Πiθ Diθ.12 with θ′ .3. 11. 11. 11.e.t +α ∂ ∂xk ′ vi ∂θ′ ∂xk +ν Diθ.13 are re-cast into turbulent diffusion terms using the same procedure as in Eqs.

k= 1 ′ ′ 1 ′2 ′2 ′2 v1 + v2 + v3 ≡ 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 v ′ ′ ∂¯i − ν ∂vi ∂vi + vj ¯ = − vi vj ∂t ∂xj ∂xj ∂xj ∂xj ε k k C P (11. Diθ.4 The k equation The turbulent kinetic energy is the sum of all normal Reynolds stresses. The total diffusion reads Dk = Dt + Dν . Πiθ and Diθ. However. 11. The k equation 85 where Piθ .21 can symbolically be written: C k = P k + Dk + Gk − ε (11. It can be noted that there is no usual viscous diffusion term in Eq.e. εiθ . see Eq. dissipation and buoyancy term. 2. if ν ≃ α (which corresponds to a Prandtl number of unity. 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).t denote the production. i. this is out of the scope of the present course but the interested reader is referred to [15].20) ′ ∂vi ∂xk ∂ −ν ∂xk ′ ∂ 2 vi θ′ ∂xk ∂xk ∂θ′ ∂xk = 2ν ′ ∂ 2 vi θ′ ν = ν+ ∂xk ∂xk Pr Often the viscous diffusion is simpliﬁed in this way.22) . respectively.21) − ∂ ∂xj ′ vj p′ ρ k Dt 1 ′ ′ + vi vi 2 +ν ∂ k ′ −gi βvi θ′ ∂xj ∂xj k k Dν G 2 k k ′ ′ where – as in the vi vj equation – Dt and Dν denotes turbulent and viscous diffusion.4.19.e. 11. respectively.ν . P r = ν/α ≃ 1. k k respectively.19 assumes the familiar form α ∂ ∂xk ′ vi ∂θ′ ∂xk +ν ∂ ∂xk θ′ θ′ θ′ ′ ∂vi ∂xk =α ′ ∂ 2 vi θ′ ∂ −α ∂xk ∂xk ∂xk ′ ∂vi ∂xk +ν ′ ∂ 2 vi θ′ ∂ −ν ∂xk ∂xk ∂xk ′ vi ′ vi ∂θ′ ∂xk ′ ∂ ∂ 2 vi θ′ −ν ≃2ν ∂xk ∂xk ∂xk (11. The unknown terms – Πiθ . εiθ and Giθ denote viscous diffusion. The production term include one term with the mean velocity gradient and one term with the mean temperature gradient. Equation 11. Diθ . i. On the last line. 11.16). the diffusion term in Eq. Giθ – have to be modelled as usual. scramble and turbulent diffusion term.11.

Ψε . C ε = P ε + Dε + Gε − Ψε (11.33).5. Hence. a turbulent) viscosity is introduced to model the unknown Reynolds stresses in Eq. 2. Gε . An exact equation for the transport equation for the dissipation ε=ν ′ ′ ∂vi ∂vi ∂xj ∂xj can be derived (see. In the k − ε model. and a destruction term.5) ∂ ∂xj ν ∂¯j v ∂¯i v + ∂xj ∂xi ′ ′ − vi vj (11. One quantity with this dimension is the mean velocity gradient which might be relevant for the production term.23) The production and destruction terms.5 The ε equation Two quantities are usually used in eddy-viscosity model to express the turbulent viscosity. The transport equation should include a convective term.27) .. A better choice should be ε/k = [1/s].09. but it is very complicated and in the end many terms are found negligible.e. 11. but not for the destruction. σε (see Eq. C ε . 11. i. and to setup a similar equation for ε.24) where we have added new unknown coefﬁcients in front of each term. a production term.26) Note that this is a modelled equation since we have modelled the production. we must multiply P k and ε by a quantity which has the dimension [1/s]. Eq. It is much easier to look at the k equation.11. The turbulent diffusion term is expressed in the same way as that in the k equation (see Eq. Dε . The terms in the k equation have the dimension ∂k/∂t = [m2 /s3 ] whereas the terms in the ε equation have the dimension ∂ε/∂t = [m2 /s4 ]. destruction and turbulent diffusion terms. a production term due to buoyancy. P ε .2. Consider the diffusion terms in the incompressible momentum equation in the case of non-constant viscosity (see Eq. 11. The ε equation 86 11. k and ε are used. we get P ε + Gε − Ψε = ε cε1 P k + cε1 Gk − cε2 ε k (11. 11. in the k equation are used to formulate the corresponding terms in the ε equation.g. The turbulent viscosity is then computed as k2 νt = Cµ ε where Cµ = 0.e.25) The ﬁnal form of the ε transport equation reads ∂ε ε ∂ ∂ε = (cε1 P k + cε1 Gk − cε2 ε) + + vj ¯ ∂t ∂xj k ∂xj ν+ νt σε ∂ε ∂xj (11.36) but with its own turbulent Prandtl number. 11.22. a diffusion term. [16]).6 The Boussinesq assumption In the Boussinesq assumption an eddy (i. P k and ε.e. i. Hence. e. Dε = ∂ ∂xj ν+ νt σε ∂ε ∂xj (11.

11. Eq. ′ vi θ′ . σθ = νt αt (11.7 Modelling assumptions ′ ′ Now we will compare the modelling assumptions for the unknown terms in the vi vj . Modelling assumptions 87 ′ ′ Now we want to replace the Reynolds stress tensor.30) Now the equation valid also when it is contracted (i.16. One option is to solve its transport equation. the Prandtl number (P r).29) This equation is not valid upon contraction (the left side will be zero. but not the right side). 11. the turbulent viscosity (νt ). 11.e.g.g. water or air) and its conditions (e. If the mean temperature equation is solved for we need an equation for the heat ′ ﬂux vector. The Boussinesq assumption reads ¯ ∂θ ′ (11. This will give us the Reynolds Stress Model [RSM] (also . If an eddyviscosity model (i.11. νt . The physical meaning of the turbulent Prandtl number.30) is used for the Reynolds stresses. the thermal diffusivity (α) are physical parameters which depend on the ﬂuid (e.32) σθ where σθ is the turbulent Prandtl number. However. it deﬁnes how efﬁcient the turbulence transports (by diffusion) momentum compared to how efﬁcient it transports thermal energy. Hence we add the trace of the left side to the right side so that ′ ′ vi vj = −νt ∂¯j v ∂¯i v + ∂xj ∂xi 2 2 + δij k = −2νt sij + δij k ¯ 3 3 (11. i. is analogous to the physical meaning of the usual Prandtl number. Eq. vi vj .19. by a turbulent viscosity. It is usually obtained from the turbulent viscosity as νt αt = (11.31) vi θ′ = −αt ∂xi where αt denotes the turbulent thermal diffusivity. νt ). 11.7. This is of course a drastic simpliﬁcation. k and ε equations and formulate modelling assumptions for the remaining terms in the Reynolds stress equation. 2.g. temperature).e taking the trace).9.7 ≤ σθ ≤ 0.28 gives ′ ′ vi vj = −νt ∂¯i v ∂¯j v + ∂xj ∂xi (11.27 and 11. the turbulent thermal diffusivity (αt ) and the turbulent Prandtl number (σθ ) depend on the ﬂow (e. vi θ′ .33) It is important to recognize that the viscosity (ν). 11.28) Identiﬁcation of Eqs. so that the the diffusion terms can be written ∂ ∂xj (ν + νt ) ∂¯j v ∂¯i v + ∂xj ∂xi (11. see Eq. 11. it is an empirical constant which is usually set to 0. mean ﬂow gradients and turbulence). σθ . after contrac′ ′ tion both left and right side are equal (as they must be) and equal to vi vi = 2k.2 we replace six turbulent stresses with one new unknown (the turbulent viscosity. an eddy-viscosity model is commonly used also for the heat ﬂux vector.e. When Eq.30 is included in Eq.

vi θ′ and k due to buoyancy k ′ ′ ′ Dij. s 11. − vj vk ∂xk ∂xk ¯ v ′ ′ ∂ θ − v ′ θ ′ ∂¯i = −vi vk k ∂xk ∂xk Pk = 1 v ′ ′ ∂¯i Pii = −vi vj 2 ∂xj (11. The dissipation takes place at the small-scale turbulence.11.7. ε or ω) is always part of an RSM and that equation includes P k .2 Diffusion terms The diffusion terms in the k and ε-equations in the k − ε model are modelled using the standard gradient hypothesis which reads Dk = ∂ ∂xj ∂ Dε = ∂xj νt σk νt ν+ σε ν+ ∂k ∂xj ∂ε ∂xj (11.t .t . In the k − ε model. Piθ and P k are production terms of vi vj . Later on. vi θ′ and k ′ Πiθ is the pressure-scramble terms of vi θ′ Πij is the pressure-strain correlation term.1 Production terms In RSM and ASM the production terms are computed exactly ′ ′ Pij = −vi vk Piθ ∂¯j v v ′ ′ ∂¯i . εiθ and ε are dissipation of vi vj . we will introduce a simpliﬁed algebraic model. Ωij = + − ∂xj ∂xi 2 ∂xj ∂xi (11. ARSM) Summary of physical meaning: ′ ′ ′ Pij . Diθ. Modelling assumptions 88 called the Reynolds Stress underline Transport Model [RSTM]) where a (modelled) transport equation is solved for each stress.7.e. which is called the Algebraic Stress Model [ASM] (this model is also called Algebraic Reynolds Stress Model. vi θ′ and k ′ ′ ′ Gij . respectively. which promotes isotropy of the turbulence ′ ′ ′ εij . vi θ′ and k. which gives ′ ′ −vi vj = νt P k = νt sij = ¯ 1 2 ∂¯i v 2 ∂¯j v − δij k + ∂xj ∂xi 3 ∂¯i v ∂¯j ∂¯i v v + = νt 2¯ij (¯ij + Ωij ) = 2νt sij sij s s ¯ ¯ ∂xj ∂xi ∂xj ∂¯i v 1 ∂¯i v ∂¯j v ∂¯j v .36) .34) The k is usually not solved for in RSM but a length-scale equation (i. Dt are the turbulent diffusion terms of vi vj . Giθ and Gk are production terms of vi vj . the Reynolds stresses in the production term are computed using the Boussinesq assumption.35) where on the second line we used the fact that sij Ωij = 0 because the product between ¯ a symmetric tensor (¯ij ) and an asymmetric tensor (Ωij ) is zero.7. 11.

36 which for vi vj reads Dij.t.24 we take k/ε so that k ′ ′ ∂k (11.11. 11.t νt ∂k 1 ′ ′ ′ vj vi vi = − 2 σk ∂xj (11.t. is computed as 1. A quantity of dimension [s] must be added to get the correct dimension.21) gives the turbulent diffusion term in Eq. Equation 11. opens possibilities for a more advanced model of the turbulent diffusion terms. e.G ′ ′ dk 1. for example. Modelling assumptions 89 The gradient hypothesis simply assumes that turbulent diffusion acts as to even out k all inhomogeneities. 8.t ∂ = ∂xk (11.37 assumes that if the gradient is zero in xi direction. 11.36.t. Dt . transports k from regions where k is large to regions where k is small. 66).t. ′ ′ The corresponding diffusion terms for the ε and vi vj equations read ε Dt = ∂ ∂xj ′ ′ cε vj vk k ∂ε ε ∂xk ′ ′ k ∂vi vj ε ∂xm Dij.37 (including the minus sign in Eq. in the k equation is obtained by taking the divergence of this equation ∂dk k ′ ′ ∂k ∂ j. 11.G ∝ vj vk ∂k ∂xk (11.t. In GGDH the turbulent ﬂux dk . 11. it assumes that the turbulent diffusion term. Taking the divergence of Eq.2 at p. and as in Eq. ′ ′ dk j. A more stable alternative is to model ′ ′ the diffusion terms as in 11. then there is no diffusion ﬂux in that direction.41) Dt = = ∂xj ∂xj ε ∂xk This diffusion model may be used when the k equation is solved in an RSM or an ASM. even if ∂k/∂x1 = 0 the diffusion ﬂux dk 1. A more general gradient hypothesis can be formulated without this limitation.G = ck vj vk ε ∂xk k The diffusion term.38) which is called the general gradient diffusion hypothesis (GGDH). The turbulent diffusion ﬂux of k is expressed as dk = j.7.40) dk j.42 often causes numerical problems.37) Only the triple correlations are included since the pressure diffusion usually is negligible (see Fig.G may be non-zero.g. Dt . vi vj . ′ ′ Solving the equations for the Reynolds stresses.t. In other words.43) .t = ∂ ∂xm ′ ′ νt ∂vi vj σk ∂xm (11.G ∝ v1 v1 ∂k ′ ′ ∂k + v ′ v ′ ∂k + v1 v2 1 3 ∂x1 ∂x2 ∂x3 (11.G k ck vj vk (11.42) ′ ′ ck vk vm Equation 11.39) Hence. It was derived in ′ ′ ′ [17] from the transport equation of the triple correlation vj vi vi .

and spanwise. ′2 ′2 ′2 v1 .11.10 and 11. see Eqs.44 are conveniently expressed in tensor notation as εij = 2 εδij 3 (11. 46. 5.10) is active for the small-scale turbulence. moving ﬂuid particles against the pressure gradient. and positive for the wall-normal. v1 = v2 = v3 .7.4 Slow pressure-strain term ′ ′ The pressure-strain term. 11.45) where the factor 2/3 is included so that ε = 1 εii is satisﬁed. 11. >0 (11.e. All shear stresses are zero. i. What applies for the small-scale ﬂuctuations (Items 1 and 2. see Fig.e. it acts as a sink term for the shear stress equation. 11. i. above) must also apply to the gradients of the ﬂuctuations.e. makes a large contribution to the vi vj equation. 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. As′ sume that two ﬂuid particles with ﬂuctuating velocities v1 bounce into each other at O ′ so that ∂v1 /∂x1 < 0.44) The relations in Eq. εij 90 The dissipation term εij (see Eq. 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. ′ ′ ′ ′ ′ ′ ∂v1 ∂v1 ∂v2 ∂v2 ∂v3 ∂v3 = = ∂xk ∂xk ∂xk ∂xk ∂xk ∂xk ′ ∂v ′ ∂vi j = 0 if i = j ∂xk ∂xk (11. Because of the cascade process and vortex stretching (see Figs. In Section 9 it was shown that for channel ﬂow it is negative for the streamwise equation. ′ ′ ∂v3 ∂v2 > 0.e. it was shown that the term acts as to make the turbulence more isotropic.2 and 5. v2 . This gives: ′2 ′2 ′2 1. 11. Modelling assumptions 11.3 Dissipation term. ′ ′ vi vj = 0 if i=j because the ﬂuctuations in two different coordinate directions are not correlated.3) the smallscale turbulence is isotropic. 2. decreasing the large normal stresses and the magnitude of the shear stress and increasing the small normal stresses. see p.7. This means that the velocity ﬂuctuations of the smallscale turbulence have no preferred direction. 2 11. The pressure-strain term is often called the Robin Hood terms.21. The role of the pressure strain can be described in physical terms as follows. Furthermore. i. equations. v3 .1.7. i. Πij . because it “takes from the rich and gives to the poor”.46) ∂x2 ∂x3 . In summary.

respectively.49) the pressure-strain term in Eqs. should be proportional to the difference of their energy. and adding them gives ′ ′ ∂v2 ∂v ′ ∂v1∗ ∂v ′ + 1 = − 2∗ ∂x1 ∂x2 ∂x1∗ ∂x2∗ (11.11.e. (x1∗ . i.50) . in Eq.5 v1∗ − v2∗ (11.6c: ′ ′ replacing u12 and u21 by ∂v1 /∂x2 and ∂v2 /∂x1 . x2 ) by rotating it angle α = π/4. 11. the principal axis ′ ′ of vi vj .47) The expression in Eq. Modelling assumptions 91 ′ v1 ′ v1 O x2 x1 Figure 11.e. the velocity gradients. Assume that we express Eq. i.47 can be written p′ ′ ∂v2 ∂v ′ + 1 ∂x1 ∂x2 = p′ ′ ∂v1∗ ∂v ′ − 2∗ ∂x1∗ ∂x2∗ (11. O. The amount of kinetic energy transferred from the x1 direction to the x2 and x3 directions. the continuity equation gives ∂v2 /∂x2 + ∂v3 /∂x3 > 0. ′ 1 ′ ∂v1 1 p ∝− ρ ∂x1 2 ′2 ′2 ′2 ′2 v1 − v2 + v1 − v3 ′2 = − v1 − 1 ′2 ′2 v + v3 2 2 =− 3 ′2 1 ′2 ′2 ′2 v − v + v2 + v3 2 1 2 1 =− 3 ′2 v −k 2 1 (11. We use Eqs. Now we have transformed the right side of Eq. and then transform the equation to (x1 . ′ ′ ′ Indeed.47 in principal coordinates. i. the ﬂuctuation v1 would not be able to create an acceleration of both v2 and v3 . i. If this is to happen the kinetic energy in the x1 direction must be ′2 ′2 ′2 ′2 larger than that in the x2 and x3 direction.6b and O. However. v1 > v2 and v1 > v3 .48) ′ ′ ′ ′ since v1∗ v2 ∗ = v2∗ v1 ∗ . O.47.e. 11.e. if ∂v1 /∂x1 < 0. 11. x2∗ ). Next step is to transform the left side. By transforming to a coordinate system which is rotated π/4 it is shown ′ ′ ′ ′ that the sign of p′ (∂vi /∂xj + ∂vj /∂xi ) and vi vj are opposite.47 applies only to the normal stresses.1: Physical illustration of the pressure-strain term.6b by ′ ′ v1 v2 we get ′ ′ ′2 ′2 v1 v2 = 0. 11. If this were not ′ ′ ′ true.7. 11. see Appendix O. Replacing u12 in Eq.46 we assume that not only their sum is positive but also that they both are positive.10 and 11.

52) This shows that the pressure-strain term acts as a sink term in the shear stress equation. the concept “slow” and “rapid” is discussed at p.11.e.2: Decaying grid turbulence. see Fig.53 behaves for decaying grid turbulence.51) Inserting Eqs.7. From this point and further downstream the ﬂow represents homogeneous turbulence which is slowly approaching isotropic turbulence. 11. Eqs.48 and 11. We have introduced the turbulent time scale k/ε. Now we apply Eq. 11. The circles (a) and the thin rectangles (b) illustrates past of the grid which consists of a mesh of circular cylinders. Further downstream the velocity gradients are smoothed out and the mean ﬂow becomes constant.50 p′ ′ ∂v1∗ ∂v ′ − 2∗ ∂x1∗ ∂x2∗ ′2 ′2 ∝ − v1∗ − v2∗ (11. 11. Thus. The grid creates velocity gradi¯ ents behind the grid which generate turbulence. The exact vi vj equation for this ﬂow reads (no production or diffusion because of homogeneity) v1 ¯ ′ ′ ∂vi vj p′ = ∂x1 ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi − εij (11. furthermore the turbulence is slowly dying (i.2.47 and 11.47 using the right side of Eq. Let us investigate how Eq. Modelling assumptions 92 v1 ¯ x2 x2 a) x1 b) x3 Figure 11. 11. This pressure-strain model for the slow part was proposed by Rotta in 1951 [18].50 into Eq.1 ≡ p′ ′ ′ ∂vj ∂vi + ∂xj ∂xi = −c1 ρ ε k 2 ′ ′ vi vj − δij k 3 (11. decaying) due ′ ′ to dissipation.54) . 11.51 gives ﬁnally p′ ′ ∂v ′ ∂v2 + 1 ∂x1 ∂x2 ′ ′ = −ρv1 v2 (11.53) where Φ denotes the modelled pressure-strain term and subscript 1 means the slow part.52 lead as to write Φij. 11. Flow from left with velocity v1 passes through a grid. 11. 93.

56 and dividing by 2k we obtain v1 ¯ ε 1 ε ε 1 ε ε ∂bij = −c1 bij − δij + bij + δij = bij (1 − c1 ) ∂x1 k 3 k k 3 k k (11. It is assumed that the effect of the mean gradients is much larger than the effect of the turbulence.45) into Eq. We introduce bij (Eq. then bij = 0.11. 11.59) aij is the anisotropy tensor whose trace is zero.56) ∂x1 3 ∂x1 3 Analogously to Eq. 11.7. ∂¯i v (ε/k) → ∞ (11. see Eq. In isotropic ﬂow all its components are zero.54. To make it general it is enough to include ′ v ′ . 11. i. i. 11. The most general form of Φij.60) ∂xj .58) Provided that c1 > 1 Rotta’s model does indeed reduce non-isotropy as it should. ′ ′ this means that terms cubic in vi vj can be expressed in terms linear and quadratic in ′ ′ vi vj . Φii = 2p′ ∂vi /∂xi = 0.1 = −c1 ρ εaij + c′ aik akj − δij akℓ aℓk 1 3 ′ ′ vi vj 2 aij = − δij k 3 (11. since according to the Cayley-Hamilton theorem.53) and the model for the dissipation tensor (11.55) bij = 2k 3 Note that when the turbulence is isotropic. 11. Modelling assumptions 93 Rotta’s pressure-strain model is supposed to reduce anisotropy.5 Rapid pressure-strain term Above a model for the slow part of the pressure-strain term was developed using physical arguments. 11. the trace is zero). 11.57 in Eq.20 in [19]). where a very strong velocity gradient ∂¯i /∂xj is imposed so that v initially the second term (the slow term) can be neglected. This should be so ′ since the exact form of Φij is trace-less.54 expressed in the normalized anisotropy Reynolds stress tensor which is deﬁned as ′ ′ vi vj 1 − δij (11. Here we will carry out a mathematical derivation of a model for the rapid part of the pressure-strain term. a terms which are quadratic in vi j second-order tensor satisﬁes its own characteristic equation (see Section 1.57) Inserting Eq. 11. the k equation in decaying grid turbulence reads v1 ¯ ∂k = −ε ∂x1 (11.e.54 so that ∂(kbij ) 2 1 ∂k 2¯1 v = −2c1 εbij − δij ε + δij (11.e. The model of the slow pressure-strain term in Eq.55).61.e.53 can be extended by in′ ′ cluding terms which are non-linear in vi vj .7. Rotta’s model (Eq. 11. 11. Thus it should be interesting to re-write Eq.1 can be formulated as [20] 1 Φij. Note that the right side is trace-less (i. The notation “rapid” comes from a classical problem in turbulence called the rapid distortion problem.

62. .e. Now we want to derive an exact equation for the pressure-strain term. vj ¯ = . Subtracting of Equation B from Equation A gives a Poisson equation for the ﬂuctuating pressure p′ ′ ∂¯i ∂vj v ∂2 1 ∂ 2 p′ ′ ′ v ′ v ′ − vi vj = −2 − ρ ∂xj ∂xj ∂xj ∂xi ∂xi ∂xj i j rapid term slow term (11. 1.62) there exists an exact analytical solution given by Green’s formula. .61) For a Poisson equation ∂2ϕ =f ∂xj ∂xj (11. ∂xi ∂xj 2. Since it includes the ﬂuctuating pressure. in volume V . see Appendix P (it is derived from Gauss divergence law) ϕ(x) = − 1 4π f (y)dy1 dy2 dy3 |y − x| (11. Modelling assumptions 94 y V x2 . stant viscosity (see Eq. Take the divergence of incompressible time-averaged Navier-Stokes equation as∂ ∂¯i v suming constant viscosity (see Eq.61 which gives the most “rapid” response in p′ . Thus in this case it is the ﬁrst term in Eq. p′ . 11.7. ⇒ Equation A. Take the divergence of incompressible Navier-Stokes equation assuming con∂vi ∂ vj = .3: The exact solution to Eq. . 6. we start by deriving an exact equation for p′ starting from Navier-Stokes equations. y1 Figure 11.3) i. y. ⇒ Equation ∂xi ∂xj B. The integral is carried out for all points. Πij .e. y2 x x1 . . The second “slow” term becomes important ﬁrst at a later stage when turbulence has been generated.11. 6.5) i.63) V . 11.

Applying Eq. If we assume that g(ξ) varies slowly with ξ.e.65) + 1 4π V ′ ′ ∂vi (x) ∂vj (x) + ∂xj ∂xi Aij dy3 ∂2 ′ ′ (vk (y)vℓ (y)) ∂yk ∂yℓ |y − x| Note that the mean velocity gradient.7. 11.64: ′ dy3 ∂¯i (y) ∂vj (y) ∂2 ′ ′ ′ ′ 2 v vi (y)vj (y) − vi (y)vj (y) + |y − x| ∂yj ∂yi ∂yi ∂yj rapid term slow term (11. ii) the variation of ∂¯i /∂xj in space is small. 11. ′ ′ ∂ 2 vi vj =0 ∂yi ∂yj Assumption ii) means that the mean velocity gradient can be taken outside the integral. i. 11.63 on Eq. the spatial derivative of all time-averaged ﬂuctuating quantities is zero). In order to understand this better. 11.e. ∂vj (y)/∂yi Assumption i) means that the last term in the integral in Eq. Hence. then the denominator is large and the contribution to the integral is small.61 gives ρ p′ (x) = 4π V where dy3 = dy1 dy2 dy3 .66) Note that x and ξ are coordinates along the same axis (think of them as two different points along the x axis).64 is zero.11. Now make two assumptions in Eq. we only need to consider ξ points which are close to x. 11. x and ξ. ′ ′ Now multiply Eq. If the two points. is taken at point x because it has been v moved out of the integral. vi (y)vj (y. The same argument can be used as v above: the mean gradient ∂¯i /∂xj varies indeed much more slowly than the v ′ instantaneous velocity gradient. This term is indeed very small ′ ′ compared to the second derivative of the instantaneous ﬂuctuations. 11. Since this term is not a function of y it can be moved in under the integral. Modelling assumptions 95 where the integrals at the boundaries vanish because it is assumed that f → 0 at the boundaries.64 with ∂vi /∂xj + ∂vj /∂xi . see Fig. are far from each other.64) i) the turbulence is homogeneous (i. ∂¯/∂xℓ . We obtain after time averaging 1 ′ p (x) ρ = ′ ′ ∂vi (x) ∂vj (x) + ∂xj ∂xi ′ ′ ∂vi (x) ∂vj (x) + ∂xj ∂xi Mijkℓ ′ ∂vℓ (y) dy3 ∂yk |y − x| ∂¯k (x) 1 v ∂xℓ 2π V (11.3. This requirement is not as drastic as it may sound (very few turbulent ﬂows are homogeneous). consider the integral L f (x) = 0 g(ξ)dξ |x − ξ| (11. .

Now we will take a closer look at rapid part (i. ∂ ∂ =− ∂xi ∂ri ∂ ∂ = ∂yi ∂ri (11.2 ∂xℓ (11. 11.65 are two-point correlations. In other words. Note that the terms above as well as in Eq. Introduce the distance vector between the two points ri = yi − xi Differentiating Eq.11. Modelling assumptions 96 Going from Eq. and second term the rapid term. Equation 11.1 (see Eq.67) f (x) = g(x) |x − ξ| 0 = Aij + Mijkℓ ∂¯k v = Φij. the two-point correlations are independent of where in space the two points are located.72) .53).70 is a coordinate transformation where we replace xi and yi with I. 11. or II. 11.70) (11. For the same reason the last term on line 2 is zero. (11.68) where the ﬁrst term represents the slow term. 11.1 + Φij.69) ′ ′ ∂ 2 vj (x)/∂yk ∂xi on line 1 is zero because vj (x) is not a function of y. Φij. they are only dependent on the distance between the two points (i.7. xi and ri .67 corresponds to moving the mean velocity gradient out of the integral. 95 gives that ∂/∂xi = 0 (Item I) or ∂/∂yi = 0 (Item II). 11.65 can be written on shorter form as p′ ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi g(ξ) can be moved out of the integral and since x is close to ξ. the two points being x and y. ri ). Φij.e. 11. yi and ri .66 to Eq.e.e.2 (index 2 denotes the rapid part). Hence we can replace the spatial derivative by the distance derivative. the second term) of Mijkℓ .65 can be rewritten as ′ ′ ∂vj (x) ∂vℓ (y) ∂ = ∂xi ∂yk ∂yk ′ vℓ (y) ′ ∂vj (x) ∂xi ′ − vℓ (y) ′ ∂ 2 vj (x) ∂yk ∂xi ′ vj (x) ′ ∂vℓ (y) ∂xi = = ∂2 ∂ ′ v ′ (y)vj (x) − ∂yk ∂xi ℓ ∂yk ∂2 ′ v ′ (y)vj (x) ∂yk ∂xi ℓ (11. i. Eq.70 gives ∂ ∂ ∂ = − ∂ri ∂yi ∂xi Equation 11.71) Assumption i) at p. The second term of Mijkℓ in the integral in Eq. 11.66 can be written as L dξ (11.

looking at Eq. 11.e. 11.73 is independent of in which order the two derivatives are taken.76 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) + c4 δkℓ k + c5 (3δkℓ + δkℓ )k ′ ′ = vk vℓ (c1 + c2 + 5c3 ) + kδkℓ (c4 + 2c3 + 4c5 ) (11.69 and 11.72. Eq.e. 11. i.76) + c4 δjℓ δik k + c5 (δij δkℓ + δjk δiℓ )k Each line is symmetric in (j.73 we get aiikℓ = 0 Applying this condition to Eq. ℓ) and (i. 216. so that aijkℓ is symmetric with respect to i and k. 216) ′ ′ aijiℓ = 2vj vℓ (11.76 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 ′ ′ ′ ′ = 3c1 vj vℓ + 2c2 δjℓ k + 4c3 vj vℓ + (3c4 + 2c5 )δjℓ )k ′ ′ = vj vℓ (3c1 + 4c3 ) + δjℓ k(2c2 + 3c4 + 2c5 ) (11.80) . 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.11. k).75) ′ ′ Now let us formulate a general expression of aijkℓ which is linear in vi vj and symmetric in (j. 11. using Eqs. Modelling assumptions We can now write Mijkℓ in Eq.79) Using Eq. k). Furthermore. 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. 11. aijkℓ = akjiℓ (11. 11.7. Here it is seen that if i = j then Mijkℓ = 0 due to the continuity equation. i.65.77) (11. on line 3. aijkℓ = aiℓkj (11. 11. For example.78) Green’s third formula reads (see Appendix G on p. Consider Eq.74) see Appendix G on p.65. 11.79 in Eq. as Mijkℓ = − 1 2π ∂2 ∂2 ′ ′ vℓ vj + v′ v′ ∂rk ∂ri ∂rk ∂rj ℓ i dr3 |r| 97 V (11. ℓ) and (i.73) = 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 ).

7.83) +c3 We ﬁnd that the c1 term and the second and third part of the c3 term can be merged.86 is commonly called the IP model (IP=Isotropization by Production) . This pressure-strain model is called the LRR model and it was proposed in [21].77. 11.85 satisfy continuity and symmetry conditions. It might be possible to use a simpler pressure-strain model using one or any two terms. 11.84) Finally we re-write this equation so that it is expressed in trace-less tensors Φij. 11 3c2 + 2 . Note that Φii = 0 as we required in Eq. Since Eq. Since the ﬁrst term is the most important one.6 was (c2 = −1. 11. 11. a value of γ = 0.82) Inserting Eq. a simpler model has been proposed [21.4.78 and 11.76 and 11. 11.68 gives ∂¯k v ∂¯k v = (aijkℓ + ajikℓ ) ∂xℓ ∂xℓ v v v v ′ ′ ∂¯i + v ′ v ′ ∂¯j ′ ′ ∂¯k + v ′ v ′ ∂¯k + c2 vi vk = c1 vj vℓ i ℓ j k ∂xℓ ∂xℓ ∂xj ∂xi v v v v v ′ ′ ∂¯k + v ′ v ′ ∂¯j + v ′ v ′ ∂¯i + v ′ v ′ ∂¯k + v ′ v ′ ∂¯k 2δij vk vℓ i ℓ j ℓ k j k i ∂xℓ ∂xℓ ∂xℓ ∂xi ∂xj ∂¯i v ∂¯j v ∂¯i v ∂¯j v +c4 k + c5 k + + ∂xj ∂xi ∂xi ∂xj φij. 55 c5 = 20c2 + 6 55 (11. 11.2 = − c2 + 8 8c2 − 2 6c2 + 4 k 4 − 60c2 Pij − Dij + P + k¯ij s 11 11 11 55 v v ′ ′ ∂¯k − v ′ v ′ ∂¯k Dij = −vi vk j k ∂xj ∂xi (11. 11. 3c1 + 4c3 − 2 = 0.81 we get φij. using Eq. 22] 2 Φij.2 = −c2 ρ Pij − δij P k 3 (11.82 into Eq.81) for the ﬁve unknown constants. 4c2 + 10 . 11 c4 + 2c3 + 4c5 = 0 2c2 + 3c4 + 2c5 = 0 (11. the c2 term and the third and fourth part of the c3 term can be merged as well as the c4 and c5 terms.86: ′ ′ both models represent “return-to-isotropy”. the ﬁrst expressed in vi vj and the second in Pij .2 = −ρ c2 + 8 2 Pij − δij P k 11 3 2 60c2 − 4 8c2 − 2 Dij − δij P k − ρk¯ij s −ρ 11 3 55 (11. 11.11. Let us express all constants in c2 which gives c1 = c3 = − c4 = − 50c2 + 4 .85) where c2 = 0.80 give four equations c1 + c2 + 5c3 = 0.86 is a truncated form of . The model in Eq. Furthermore. All three terms in Eq. Since two terms are omitted we should expect that the best value of γ should be different than (c2 + 8)/11.86) It can be noted that there is a close similarity between the Rotta model and Eq. Modelling assumptions 98 Equations 11.4) found to give good agreement with experimental data.2 = Mijkℓ (11.

it is one of the most popular models of the rapid pressure-strain term. When a ﬂuid particle approaches a wall.7. 11.86 does satisfy symmetry condition and continuity but it does not satisfy the integral condition in Eq. 11. 11. The coefﬁcients for the slow and rapid terms in the LRR and LRR-IP models are summarized in Table 11.85 do. Modelling assumptions LRR model 1. 11. 1. Viscosity. In both cases it is the pressure that informs the ﬂuid particle of the presence of the wall. Pressure. There are two main effects whose underlying physics are entirely different.4 − LRR-IP model 1.7. Since the IP model is both simpler and seems to be more accurate than Eq. The effect of the wall is to dampen turbulence.85. the slow and the . 11. 11.5 − 0.53 c2 (Eq.6 Wall model of the pressure-strain term When we derived the rapid pressure-strain model using Green’s function in Eq. Equation 11. x2 x1 Figure 11.1 11.79.64 we neglected the inﬂuence of any boundaries. truncated version of Eq. 11. 11.4: Modelling of wall correction in pressure-strain terms. 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. Since the pressure-strain term includes the ﬂuctuating pressure. the presence of the wall is felt by the ﬂuid particle over a long distance.85. Both the rapid term in the LRR model and the IP model must be modiﬁed to include wall modelling. Eq.85) c2 (Eq. In wall-bounded domains it turns out that the effect of the walls must be taken into account.86) Table 11. Although Eq. Close to the wall the viscous processes (viscous diffusion and dissipation) dominate over the turbulent ones (production and turbulent diffusion).86 is a simpler. it is often found to give more accurate results [23]. 11.85 it does not satisfy all requirements that Eq.5 0.6 99 c1 (Eq. Up to now we have introduced two terms for modelling the pressure-strain term.1: Constants in the LRR and LRR-IP pressure-strain models. it is obviously the second of these two processes that we want to include in the wall model.11. 2. 11.

Φ22. The pressure ﬂuctuations dampens the wall-normal ﬂuctuations. Φ12.w and ni. is now satisﬁed since Φ11.1w = c1w v2 f k (11. Away from the wall.55|ni.1w = 0. The IP wall model for the wall-parallel ﬂuctuations reads ε ′2 Φ11.1w .89) The requirement that the sum of the pressure strain term should be zero.1w ] and taking the trace of Φ in the principal coordinates system (i.11. It has nothing to do with viscous damping.w nm. i. The wall model for the shear stress is set as ε ′ ′ 3 Φ12.1w = 0 is satisﬁed when the coordinate system is rotated.2 + Φij.w δij − Φki. Consider a wall. You can prove this by rotating the matrix [Φ11.w − Φkj. taking the sum of the eigenvalues).1w + Φ33.e.1w . Moreover.1w . in the fully turbulent region. As explained above.1w = 0.1w = − c1w v1 v2 f 2 k (11.1w + Φij.92) ε k .w (xi − xi. We need to scale the wall-normal distance with a relevant quantity and the turbulent length scale. Φij = Φij. 11. respectively. Φ21.2 nk. Furthermore. It is suitable to include a slow and a fast wall model term.1w = c1w 3 3 Φij.w nj.1w = Φ33.w nj. For the wall-normal ﬂuctuations.w f 2 2 (11. k 3/2 /ε.e. the IP wall model reads [24] ε ′2 Φ22.2 nk. see Fig.w denotes the distance vector to the wall and the unit wallnormal vector.w nm. the damping effect of the wall should decrease for increasing wall distance.w δij − vk vi nk.1w + Φ22. this damping is inviscid (due to pressure) and affects the turbulent ﬂuctuations well into the log-region. Modelling assumptions 100 fast term. the damping function goes to zero since the distance to the wall (|xi − xi.88) where xi − xi.w nk.4.91) An analogous wall model is used for the rapid part which reads Φij.87) where subscript w denotes wall modelling. In the viscous region the wall model term. The general formula for a wall that is not aligned with a Cartesian coordinate axis reads [24] 3 ′ ′ 3 ′ ′ ′ ′ vk vm nk.1w = −2c1w v2 f k 3 k2 f= 2. Φ22. k 3/2 /ε. is not relevant and should be zero since it should account only for inviscid damping. Φii.1 + Φij.w |)ε (11.2w = c2w Φkm.w − vk vj ni. seems to be a good candidate.1w . i.w f 2 2 (11.e.7.2w (11.w nk.2 ni.w |) increases faster than the turbulence length scale. function f should not exceed one.90) The factor 3/2 is needed to ensure that Φii.

21.09.35) and the buoyancy term (Eqs.11. 1. 11. 1. 1. 11.31 and 11. the production (Eq. By inserting the model assumptions for the turbulent diffusion (Eq.94) The turbulent viscosity is computed as νt = cµ The standard values for the coefﬁcients read (cµ .44.93) ν+ νt σε ∂ε ∂xj (11. 1. 11.3) (11.36).26 ∂ε ε ∂ε = cε1 νt + vj ¯ ∂t ∂xj k ∂¯i v ∂¯j ∂¯i v v + ∂xj ∂xi ∂xj ¯ ε2 ∂ ε νt ∂ θ − cε2 + + cε1 gi k σθ ∂xi k ∂xj (11. The k − ε model 101 11.8.92. σε ) = (0.95) . the modelled ε equation is obtained from Eq. 11. σk .8 The k − ε model The exact k equation is given by Eq. 11. cε2 . cε1 .32) we get the modelled equation for k ¯ ∂¯i v ∂k ∂¯j ∂¯i v νt ∂ θ v ∂k = νt + + gi β + vj ¯ ∂t ∂xj ∂xj ∂xi ∂xj σθ ∂xi ∂ νt ∂k −ε + ν+ ∂xj σk ∂xj In the same way.96) k2 ε (11.

11. ′ ′ vi vj Cij − Dij = C k − Dk k Inserting Eq.10 Algebraic Reynolds Stress Model (ASM) The Algebraic Reynolds Stress Model is a simpliﬁed Reynolds Stress Model. 83 & 85) as: RSM : Cij − Dij = Pij + Φij − εij k − ε : C k − Dk = P k − ε (11.9 The modelled vi vj equation with IP model With the models for diffusion.99) ′ ′ Thus the transport equation (PDE) for vi vj has been transformed into an algebraic equation based on the assumption in Eq. rapid part) (viscous diﬀusion) (turbulent diﬀusion) ′ ′ −gi βvj θ′ − gj βvi θ′ (buoyancy production) 2 − εδij (dissipation) 3 11. i. The modelled vi vj equation with IP model 102 ′ ′ 11.2 nk ni ]f 2 ′ ′ ∂ 2 vi vj +ν ∂xk ∂xk ′ ′ ∂vi vj ∂ ′ ′ k ck vk vm + ∂xk ε ∂xm vk ¯ (convection) (production) (slow part) (rapid part) (wall.2 nk nj 2 3 − Φjk.′ ′ 11.98) ′ ′ In ASM we assume that the transport (convective and diffusive) of vi vj is related to that of k. pressure-strain and dissipation we get ′ ′ ∂vi vj = ∂xk v v ′ ′ ∂¯j − v ′ v ′ ∂¯i −vi vk j k ∂xk ∂xk ε 2 ′ ′ −c1 vi vj − δij k k 3 2 −c2 Pij − δij P k 3 3 ′ ′ ε ′ ′ +c1w ρ [ vk vm nk nm δij − vi vk nk nj k 2 3 ′ ′ − vj vk nk ni ]f 2 3 +c2w [ Φkm. 11.98 into the equation above gives ′ ′ vi vj Pk − ε k Pij + Φij − εij = (11.98. slow part) (11.9. The RSM and k − ε models are written in symbolic form (see p.e.2 nk nm δij − Φik.97) (wall. .

11 Explicit ASM (EASM or EARSM) ′ ′ Equation 11. 11.100 reads ′ ′ −v1 v2 = As can be seen.1 (Eq.53) and Φij.1w + Φij.1w + Φij. It would of course be advantageous to be able to get an explicit expression for the Reynolds stresses.2w 3 δij k + 3 ε c1 + P k /ε − 1 v c1 − 1 + c2 P k /ε k 2 ∂¯ 2 (1 − c2 ) k /ε) ε ∂y 3 (c1 − 1 + P cµ (11. 11. Ωij .99 and multiply by k/ε so that k 2 k ′ ′ Pij − c1 vi vj − δij k − c2 ε 3 ε 2 Pij − δij P k 3 2 − δij k 3 ′ ′ vi vj k + (Φij. s ¯ ¯ ¯ s ¯ ¯ (k 2 /ε2 )Ωij Ωji . the Reynolds stresses appear both on the left and the right side of the equation.100 is an implicit equation for vi vj . Pope [25] managed to derive an explicit expression for ASM in two dimensions.45) in Eq. 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. 11.101) In two dimension the expression reads ′ ′ vi vj = k2 k3 2 ¯ ¯ s ¯ kδij + G(1) sij + +G(2) 2 (¯ik Ωkj − Ωik skj ) ¯ 3 ε ε (11. He assumed that the Reynolds stress tensor can expressed in the strain-rate tensor. G(n) .2 (Eq.2w ) = Pk − ε ε ε ′ ′ Collect all vi vj terms so that ′ ′ vi vj k 2 Pij − c2 Pij − δij P k ε 3 = 2 2 k Pij −δij P k − c2 Pij − δij P k ε 3 3 + Φij.86) and the isotropic model for εij (Eq.11. and the vorticity ¯ tensor. Explicit ASM (EASM or EARSM) 103 ′ ′ Now we want to re-write this equation as an equation for vi vj . 11. 11. Dividing both sides by P k /ε − 1 + c1 gives ﬁnally ′ ′ vi vj = 2 k (1 − c2 ) Pij − 2 δij P k + Φij.11. (k /ε )Ωij Ωjk skm smi (11.e. (k 3 /ε3 )¯ij sjk ski 4 4 ¯ ¯ 3 3 ¯ ¯ ¯ ¯ ¯ (k /ε )Ωij Ωjk ski . in that expression can be a function of not more than the following ﬁve invariants (k 2 /ε2 )¯ij sji .1w + Φij.100) In boundary layer ﬂow Eq.2w + 2 δij k(P k /ε − 1 + c1 ) 3 where (2/3)δij P k k/ε was added and subtracted at the last line (shown in boxes).1w + Φij. sij . i.102) . Furthermore.2w Pk − 1 + c1 = ε 2 + δij k(−1 + c1 ) 3 + Φij. Insert the IP models for Φij. he showed that the coefﬁcients.

11. ¯ 2 Tij ¯ ¯ = sik Ωkj − sjk Ωki .e.103 includes only linear and quadratic terms of sij ¯ ¯ and Ωij . It may be noted that Eq.5) where v2 = 0.11. vi vj − 2δij k/3.103 are symmetric and traceless as s ¯ ′ ′ required by the left side. i. 3 2 ¯ ¯ 6 ¯ ¯ ¯ ¯ ¯ ¯ ¯ Tij = Ωim Ωmk skj + sik Ωkm Ωmj − δij Ωpm Ωmk skp 3 8 7 ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ Tij = Ωim smk Ωkn Ωnj − Ωim Ωmk skn Ωnj .12.12 Boundary layer ﬂow ¯ ¯ ¯ Let us study boundary layer ﬂow (Fig. 11. ¯ ¯ 1 ¯ ¯ 4 ¯ ¯ Tij = Ωik Ωkj − δij Ωik Ωki . note that all terms in Eq. 11.20 in [19]). Any ASM may be written on the form of Eq. the Reynolds stress tensor depends on 10 tenn sors. Equation 11.103. hence cubic terms or higher can recursively be expressed in linear (¯ij ) and quadratic tensors s (¯ik skj ). Tij = sim Ωmk skn snj − sim smk Ωkn snj 2 ¯ ¯ 9 ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ Tij = Ωim Ωmk skn snj − sim smk Ωkn Ωnj − δij Ωpm Ωmk skn snp ¯ ¯ 3 ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ T 10 = Ωim smk skn Ωnp Ωpj − Ωim Ωmk skn snp Ωpj ij 1 3 Tij = sik skj − δij sik ski ¯ ¯ ¯ ¯ 3 5 ¯ ¯ ¯ Tij = Ωik skm smj − sim smk Ωkj ¯ ¯ ¯ (11.103 is a general form of a non-linear eddy-viscosity model. 11. 10 ′ ′ vi vj − kδij = 1 Tij n G(n) Tij n=1 = sij . That is because of Cayley-Hamilton theorem which states that a secondorder tensor satisﬁes its own characteristic equation (see Section 1. Boundary layer ﬂow 104 x2 n v1 (x2 ) ¯ Xx x1 s Xx Figure 11. Furthermore.103) n where Tij may depend on the ﬁve invariants in Eq. Tij [25]. In general three-dimensional ﬂow. 11. In general the production Pij has the form: ′ ′ Pij = −vi vk ∂¯j v v ′ ′ ∂¯i − vj vk ∂xk ∂xk . v1 = v1 (x2 ).5: Boundary layer ﬂow.101. 11.

1 and Φij.11.1 = c1 Φ22.1 and Φij. v1 ) and gives to the poor (i.e. but it takes the value 3 ε for the ′2 ′2 ′ ′ v1 and v2 equations (see p.2 ε 2 ′2 k − v2 > 0 k 3 1 v 2 ′ ′ ∂¯1 = c2 P11 = −c2 v1 v2 >0 3 3 ∂x2 2 ′ ′ Note also that the dissipation term for the v1 v2 is zero. the pressure strain term Φij. Boundary layer ﬂow 105 In this special case we get: ′ ′ P11 = −2v1 v2 ∂¯1 v ∂x2 ′2 P12 = −v2 ∂¯1 v ∂x2 P22 = 0 ′2 Is v2 zero because its production term P22 is zero? No! The sympathetic term Φij ′2 ′2 which takes from the rich (i. v2 ) saves the unfair situation! The IP model for Φij. the question arises: what is the main sink term in the v1 v2 equation? The answer is.e.2 .12. 90). . Since the modelled v1 v2 does not have any dissipation ′ ′ term.2 gives Φ22. again.

1) ′ since gi = (0. is moved down to level 1. The reason for the superiority of the former model is in all cases that the production term is treated exactly. hence it continues to move upwards. whereas it in eddy-viscosity models is modelled. If it is moved down to level 1 it is heavier than its new environments and it will then continue downwards. From the equation for the turbulent heat ﬂux. i. turbulent ﬂuctuations are enhanced. If the temperature increases ¯ upwards (i. This is illustrated in ¯ Fig. Hence.e. Cold ﬂuid is ¯ located on top of hot ﬂuid. the buoyancy forces the particle back to its original position 0. For the case of unstable stratiﬁcation.e.1 Stable and unstable stratiﬁcation In ﬂows where buoyancy is dominating. Hence. 12.e. the temperature has a large effect on the turbulence through the buoyancy term Gij . and hence buoyancy makes it to move make to its original level 0. see Eq. The production term due to buoyancy reads (see Eq. Similarly if a particle originating at level 0. Here it is lighter than its new environment. eddy-viscosity models ¯ ¯ θ2 > θ0 ρ 2 < ρ0 F 106 2 0 ¯ ∂ θ/∂x3 > 0 ∂ρ/∂x3 < 0 x3 F ¯ ¯ θ1 < θ0 ρ 1 > ρ0 1 ¯ Figure 12. 12. 11. it is at this location lighter than its new environment. see Fig.e. If a ﬂuid particle at level 0 is displaced upwards to level 2.19 . 12. ∂ θ/∂x3 > 0). 12 Reynolds stress models vs. −g). In Fig. ∂ θ/∂x3 < 0 and ∂ρ/∂x3 > 0. the situation is reversed. 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. 11. Consider ∂ θ/∂x3 > 0. v3 θ′ (i.10) ′ G33 = 2gβv3 θ′ (12.1. This means that the density decreases with increasing vertical height. then the ﬂow is stably stratiﬁed.1. In this way the vertical turbulent ﬂuctuations are dampened.10. 11.1: Stable stratiﬁcation due to positive temperature gradient ∂ θ/∂x3 > 0. 12. i. Reynolds stress models vs.12. 0. ∂ρ/∂x3 < 0. If a ﬂuid particle is displaced from its equilibrium level 0 up to level 2.1 we would then have ρ2 > ρ0 . Eq. it is heavier then the surrounding at this new level (ρ0 > ρ2 ). This ﬂow situation is called unstable stratiﬁcation.

A polar ˆ coordinate system r − θ with θ locally aligned with the streamline is introduced. it makes v3 θ′ < 0 so that G33 < 0 (see Eq. In [28] they carried out an experimental investigation on a conﬁguration simulating the ﬂow near a trailing edge of an airfoil. Gij . Thus. 11.2. transport equations are usually not solved ′ for vi θ′ . it reads Gk = gβv3 θ′ which with Eq. we ﬁnd that the production term. As vθ = vθ (r) (with ∂vθ /∂r > 0 and vr = 0). ¯ If the situation in Fig.2 Curvature effects When the streamlines in boundary layer ﬂow have a convex curvature. we ﬁnd the production term for v3 θ′ ′ ′ P3θ = −v3 vk 107 ¯ ∂θ ∂¯3 v ′ − vk θ′ ∂xk ∂xk (12.4) Hence it is seen that in stably stratiﬁed conditions.01 can have a signiﬁcant effect on the turbulence. 0. the production term due to temperature gradient ′2 ¯ reads P3θ = −v3 ∂ θ/∂x3 < 0 (recall that we assume that buoyancy dominates so that the ﬁrst term in Eq. Since the main source ′ ′ term in the v3 θ′ equation. −g). P3θ . g1 = g2 = 0).1 is reversed so that ∂ θ/∂x3 < 0 the vertical ﬂuctuations are instead augmented.12.1). When eddy-viscosity models are used. is that the former reduces k whereas the latter reduces only the vertical ﬂuctuations. see Eq. The ratio of boundary layer thickness δ to curvature radius R is a common parameter for quantifying the curvature effects on the turbulence. 27]. since G11 = G22 = 0 because the gravity is in the x3 direction (i.1.2. for the case illustrated in Fig. the turbulence is stabilized.10 (take the trace of Gij and divide by two) ′ Gk = 0. G33 . The ′2 ′ ′ reduction of ρv1 and −ρv1 v2 was also substantial. 12. see Fig.2) In the case illustrated in Fig.1. the radial inviscid momentum equation degenerates to 2 ∂p ρvθ − =0 r ∂r (12.e. 12. 12. Concave curvature destabilizes the turbulence.31 gives (12. 11. The buoyancy term. The work reviewed by Bradshaw [26] demonstrates that even such small amounts of convex curvature as δ/R = 0. An illustrative model case is curved boundary layer ﬂow. is negative.5) . They reported a 50 percent decrease ′2 of ρv2 (Reynolds stress in the normal direction to the wall) owing to curvature. 12. The difference between an eddy-viscosity model and a Reynolds stress model. especially the shear stress and the Reynolds stress normal to the wall. This is called unstably stratiﬁed conditions. Note that the horizontal turbulent ﬂuctuations are not affected by the buoyancy term. This dampens the turbulence [26.3) Gk = −gβ ¯ νt ∂ θ σθ ∂x3 (12.03.2 is much larger than the second one).5Gii = −gi βvi θ′ ′ For gi = (0. Curvature effects ′ with i = 3). Instead the heat ﬂux tensor is modelled with an eddy-viscosity assumption using the Boussinesq assumption. where they measured δ/R ≃ 0. 11. Gk < 0 as required. due to buoyancy yields a damping of the vertical ﬂuctuations as it should. in the k equation reads.31. 12. 12. 12. Gk . In addition they reported signiﬁcant damping of the turbulence in the shear layer in the outer part of the separation region. see Eq.

3: Streamline curvature occurring when the ﬂow approaches.2. r v1 ¯ x2 x1 streamline θ Figure 12. for example. . Curvature effects 108 A 0 B r B 0 A vθ (r) Figure 12.12. a separation region or an obstacle.2: Flow in a polar coordinate system illustrating streamline curvature. The streamline is aligned with the θ axis.

when ∂vθ /∂r > 0. v1 v2 − eq. and so on.12. The turbulence is destabilized owing to concave curvature of the streamlines. Similarly. there are more terms v resulting from ∂¯2 /∂x1 . which continuously increases the Reynolds stresses. Eq.5 gives (∂p/∂r)A > (∂p/∂r)0 . However as soon as the streamlines are deﬂected.5 shows that the pressure gradient increases with r. It is seen that there is a positive feedback. at least in the radial direction. How will this affect turbulence? Let us study the effect of concave streamline curvature. 12. At one x1 station. turbulent ﬂuctuation) outwards to level A. 11.2. The ratio of ′2 ′2 the normal stresses ρv1 to ρv2 is typically 5. : P22 = −2v1 v2 k k−ε P = νt ∂¯1 v ∂¯2 v + ∂x2 ∂x1 (12. Since we have assumed that ∂vθ /∂r > 0. v2 − eq. As long as the streamlines are parallel to the wall. Assume that there is a ﬂat-plate boundary layer ﬂow.3. Instead the centrifugal force drives it back to its original level. It is discussed below how the Reynolds stress model responds to streamline curvature. Here the variables are instantaneous or laminar.6b) (12. Thus the magnitude contribute non-negligibly to P12 as ρv1 2 of P12 will increase (P12 is negative) as ∂¯2 /∂x1 > 0. has a stabilizing effect on (turbulent) ﬂuctuations. if the ﬂuid is displaced inwards to level B. : P11 = −2v1 v2 ′ ′ ′2 RSM. : P12 = −v1 ∂¯1 v ∂x2 (12. Curvature effects ∂Vθ /∂r > 0 convex curvature concave curvature stabilizing destabilizing ∂Vθ /∂r < 0 destabilizing stabilizing 109 Table 12. which in turn will increase P11 and P22 .6a) (12. all production is a result of ∂¯1 /∂x2 . The centrifugal force exerts a force in the normal direction (outward) on a ﬂuid following the streamline. It is clear from the model problem above that convex curvature. see Fig. This means that ρv1 ′2 and ρv2 will be larger and the magnitude of P12 will be further increased. If the ﬂuid is displaced by some disturbance (e.6c) 2 ∂¯2 v v ′2 ∂¯1 − v2 ∂x1 ∂x2 ∂¯2 v ∂x1 ′2 ′ ′ RSM. Even if ∂¯2 /∂x1 is much smaller than ∂¯1 /∂x2 it will still v v v ′2 is much larger than ρv ′2 . An increase in the magnitude of v ′ ′ ′2 P12 will increase −v1 v2 . it encounters a pressure gradient larger than that to which it was accustomed at r = r0 . as (vθ )A > (vθ )0 . 12. the ﬂow is deﬂected upwards. 12. The production terms Pij owing to rotational strains (∂¯1 /∂x2 .g. v1 − eq. the pressure gradient is smaller here than at r = r0 and cannot keep the ﬂuid at level B. which from Eq.10): ′2 ′ ′ RSM. Hence the ﬂuid is forced back to r = r0 . ∂¯2 /∂x1 ) v v can be written as (see Eq. which is balanced by the pressure gradient.6d) The terms in boxes appear because of the streamline curvature.1: Effect of streamline curvature on turbulence. Note than eddy-viscosity models such as k − ε and k − ω models cannot account for .

12. 12. It should be mentioned that one part of the effect of curved streamlines in Eq. and whether there is an increase or decrease in momentum in the tangential direction with radial distance from its origin (i. 12. ∂¯1 /∂x2 and ∂¯2 /∂x1 . for example. The production term in the k equations for RSM/ASM and k − ε model in stagnation .1.12.6 ′ ′ is due to the transformation of the advective term of the vi vj -equation (cf. polar coordinates where additional terms appear both in the momentum equations and the transport ′ ′ equation for vi vj ). It should be noted that concave or convex depends on from which the streamline is viewed. If the streamline (and the wall) is deﬂected downwards. Stagnation ﬂow v1 (x2 ) ¯ 110 x2 x1 Figure 12.3. the sign of ∂Vθ /∂r). The streamline in Fig. whereas ASM/RSM do.4) the situation will be reversed: the turbulence will be stabilized. streamline curvature since the two rotational strains. If the ﬂow (concave curvature) is a wall jet ﬂow where ∂¯1 /∂x2 < 0 in the outer v part (see Fig.e. these cases are summarised in Table 12. In [29] they proposed a correction term to take this effect into account.3. For convenience.5: The ﬂow pattern for stagnation ﬂow. v v The stabilizing or destabilizing effect of streamline curvature is thus dependent on the type of curvature (convex or concave).3 Stagnation ﬂow The k − ε model does not model the normal stresses properly.4: The velocity proﬁle for a wall jet. x2 x2 x1 x1 x2 x1 Figure 12. and destabilizing for ∂¯1 /∂x2 < 0. in the v v production term are multiplied by the same coefﬁcient (the turbulent viscosity). is concave when viewed from the wall but convex when viewed from the orig of the circle with radius r. the situation will be as follows: the turbulence is stabilizing when ∂¯1 /∂x2 > 0. 12.

v3 ) 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 (stagnation ﬂow) iv) in boundary layers approaching separation. • Disadvantages with ASM/RSM: i) RSM is complex and difﬁcult to implement. 12. boundary layers approaching separation. especially implicit ASM ii) numerically unstable because small stabilizing second-order derivatives in the momentum equations (only laminar diffusion) iii) CPU time consuming . the two terms v v are added with sign. 12.5) due to ∂¯1 /∂x1 and ∂¯2 /∂x2 is: v v ′2 RSM : 0.12. and thus not good in predicting normal stresses (v1 . v2 . buoyancy etc. In the k − ε model. however.4.8) k − ε : P k = 2νt + ∂¯2 v ∂x2 2 where continuity ∂¯1 /∂x1 = −∂¯2 /∂x2 has been employed. the production will be large because the difference in sign of the two terms is not taken into account. RSM/ASM versus k − ε models ﬂow (see Fig. the production due to normal stresses is of the same magnitude as that due to shear stresses [30].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 equations iii) work reasonably well for a large number of engineering ﬂows • Disadvantages: ′2 ′2 ′2 i) isotropic.5 (P11 + P22 ) = −v1 111 v ∂¯1 v v ′2 ∂¯2 = − ∂¯1 (v ′2 − v ′2 ) − v2 2 ∂x1 ∂x2 ∂x1 1 ∂¯1 v ∂x1 2 (12. • Advantages with ASM/RSM: i) the production terms do not need to be modelled ii) thanks to i) it can selectively augment or damp the stresses due to curvature effects (RSM is better than ASM because the convective terms are accounted for).7) (12. In RSM.

[19] or [32]) |Cij − δij λ| = 0 which gives.e. i. ¯ Let’s now brieﬂy repeat the concept “invariants”. 13. Realizability 112 13 Realizability There are a number of realizability constraints.2). 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.4) 2D I1 = Cii 1 2D I2 = (Cii Cjj − Cij Cij ) = det(Cij ) 2 (13. it follows that its coefﬁcients I1 and I2 are invariants. 13. 13. This means something that is independent of the coordinate system.g. Here we mean independent of rotation of the coordinate system.5) Since the above equation is the same irrespectively of how the coordinate system is 2D 2D rotated. 12. If a tensor is symmetric. The equation for ﬁnding the eigenvalues of a tensor Cij is (see e. The Boussinesq ′2 assumption for the normal stress v1 reads (cf.6) . Eq. i = j (13.13. Assume that the ﬂow is in the x1 direction and that it approaches the wall (see Fig. satisfying the ﬁrst criteria is actually of importance for eddy-viscosity models in stagnation ﬂow [31].2) ′2 It is seen that if s11 gets too large then v1 < 0 which is unphysical.g..5).3) C12 C22 − λ =0 (13.7) ′2 v1 = 2 ∂¯1 v 2 = k − 2νt s11 k − 2νt ¯ 3 ∂x1 3 (13. [19]). non-realizable. i.3 gives C11 − λ C12 C21 C22 − λ C31 C32 C13 C23 C33 − λ =0 (13. Thus this is the coordinate system in ¯ ′2 which the danger of negative v1 from Eq. e.g. C11 − λ C21 The resulting equation is 2D 2D λ2 − I1 λ + I2 = 0 (13. However.1) These criteria are seldom used in RSMs. For the strain tensor this means that the off-diagonal components of sij vanish and this is the coordinate system where the diagonal com¯ ponents become largest (e. 12. s11 in Eq. then we know that it has real eigenvalues which means that we can rotate the coordinate system so that the off-diagonal components vanish (see. In 3D Eq.e. in 2D.2 is largest. ′2 vi ≥ 0 for all i 1/2 ′ ′ vi vj ′2 vi ′2 vj ≤ 1 no summation over i and j.

7 instead of Eq.9) (see Eq. 13. and Eq.14) . 13. should go to zero as vi ′2 vi → 0 ⇒ ′2 dvi →0 dt (13. s ¯ (13.2 and assume incompressible 2D ﬂow. One way to ensure this is to require that the derivative of vi ′2 goes to zero.11 νt ≤ k k = 3|λ1 | 3 2 sij sij ¯ ¯ 1/2 (13. we apply Eq. using Eq. The ﬁrst invariant reads (cf. 13. As discussed ¯ ¯ ¯ above. Two-component limit 113 which gives 3D 3D 3D λ3 − I1 λ2 + I2 λ − I3 = 0 3D I1 = Cii 1 3D I2 = (Cii Cjj − Cij Cij ) 2 1 3D I3 = (2Cij Cjk Cki − 3Cij Cji Ckk + Cii Cjj Ckk ) = det(Cij ) 6 (13. Eq. s11 in Eq.1 Two-component limit ′2 Another realizability constraint is to require that when vi approaches zero near walls. and it ensure that the normal stresses stay positive. is 2D λ1.10 is replaced by [31] |λk | = k 2¯ij sij s ¯ 3 1/2 (13. 13. 13.2 is replaced by the largest eigenvalue so that ′2 v1 = 2 k − 2νt λ1 3 (13.13) This is a simple modiﬁcation of an eddy-viscosity model. Hence. 13.2 in the principal coordinate directions of sij . ′2 it should do so smoothly.8) It is zero due to the continuity equation.8. The solution to Eq. 13.7) 3D 3D 3D The invariants are I1 . 13. Let’s go back to Eq.11) ′2 The requirement v1 ≥ 0 gives now together with Eq.12) In 3D. 13. I2 and I2 .1. 13.10) The eigenvalues of sij correspond to the strains in the principal axis.2 = ± −I2 1/2 =± sij sij ¯ ¯ 2 1/2 (13. Eq.13. The second invariant of sij reads ¯ 2D I2 = −¯ij sij /2. 13. i.e.5) 2D I1 = sii = s11 + s22 = λ1 + λ2 = 0 ¯ ¯ ¯ (13. 13.5.5) which is the same in all coordinate systems (hence the name ”invariant”).5 is used.

Since we are here concerned about the ′2 pressure-strain term. A convenient parameter proposed in [33] is A which is an expression of A2 and A3 (the second and third invariant of aij . 2 Φ22. we follow a ﬂuid particle as it approaches the wall). 11.17) The parameter A = 0 in the two-component limit and A = 1 in isotropic turbulence. 9 A2 = aij aji . Thus A is a suitable parameter to use when damping the constant c1 as the wall is approached.2 in Eq. gives 2ε Φ22. i. Two-component limit 114 where d/dt denotes the material derivative (think of Eq. the left side (and thus also the right side) of the trans′2 port equation of vi should also do so too. 11. for example. 11.15) Very complex forms of Φij.2 have been proposed [34] [CL96] which include terms cu′ ′ bic in vi vj . A = 1 − (A2 − A3 ) 8 (13. we’ll take a look at how it behaves near walls when vi → 0.2 = 0 when v2 = 0 [20]. 13.59 (non-linear model) do also not satisfy the two-component limit. Neither the form of Φij. The CL96 model does satisfy the two-component limit. for example.86. A3 = aij ajk aki .13. . The Rotta model.14 in Lagrangian coordinates. i. This is of some relevance 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 [33].2 → γ δij P k = 0 3 (13. 11.53 (linear model) and Eq. Equation 13. The models of the slow pressure-strain in Eq.14 requires ′2 that when vi approaches zero. In Eq.e.1. Another advantage of the CL96 model is that it does not need any wall distances.86 nor Eq. respectively).85 satisfy the requirement ′2 that Φ22. 11.1 → c1 ρ = 0 (13.e.16) 3 The only way to ensure this is to make c1 → 0 when the wall is approached. which is valuable in complex geometries.

e.3) where τ is a turbulent time scale. However.14. This is due to Cayley-Hamilton theorem which states that a ¯ tensor is only linearly independent up to quadratic terms.e. i. and for a nonlinear k − ω model τ = 1/ω. s3 = sik skℓ sℓj can be expressed as a linear combination of s2 = sik skj and ¯ij ¯ ¯ ¯ ¯ij ¯ ¯ sij . 14.1) where the anisotropy tensor is deﬁned as aij ≡ ′ ′ vi vj 2 − δij k 3 (14. 14.3 also has these properties. Non-linear Eddy-viscosity Models 115 14 Non-linear Eddy-viscosity Models ′ ′ In traditional eddy-viscosity models the turbulent stress vi vj is formulated from the Boussinesq assumption.2) ′ ′ The relation between the stress vi vj and the velocity gradient in Eq. One way to make eddy-viscosity models more general is to include non-linear terms of the strain-rate (i. for a non-linear k − ε model τ = k/ε. see p. 14. ¯ aij is symmetric and its trace is zero. A subset of the most general form reads [35] aij = −2cµ τ sij ¯ 1 ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ + c1 τ 2 sik skj − sℓk sℓk δij + c2 τ 2 Ωik skj − sik Ωkj 3 1¯ ¯ ¯ ¯ ¯ ¯ ¯ + c3 τ 2 Ωik Ωjk − Ωℓk Ωℓk δij + c4 τ 3 sik skℓ Ωℓj − Ωiℓ sℓk skj ¯ ¯ ¯ 3 2¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ + c5 τ 3 Ωiℓ Ωℓm smj + siℓ Ωℓm Ωmj − Ωmn Ωnℓ sℓm δij 3 ¯ ¯ ¯ + c6 τ 3 skℓ skℓ sij + c7 τ 3 Ωkℓ Ωkℓ sij ¯ ¯ ¯ 1 ¯ Ωij = 2 ∂¯i v ∂¯j v − ∂xj ∂xi (14. The tensor groups correspond to a subset of Eq. Examples of non-linear models (sometimes also . this means that. the velocity gradient) [25]. for example. 3 2 Line 2: Tij and Tij 4 5 Line 3: Tij and Tij 6 Line 4: Tij 1 ¯ ¯ Line 5: Tij multiplied by the invariants skℓ skℓ and Ωkℓ Ωkℓ ¯ ¯ The expression in Eq. note that it is only v ¯ quadratic in sij and Ωij . aij = −2νt sij = ¯ 1 2 sij ¯ k ∂¯i v ∂¯j v + ∂xj ∂xi (14. as can be seen.103: 1 Line 1: Tij . 93. it is easily veriﬁed that the right side of Eq. linear. 11.1 is.3 is cubic in ∂¯i /∂xj .

c3 = 0.5 and c7 = 0. 38].82 2 τ 12 −0.14. Non-linear Eddy-viscosity Models 116 called explicit algebraic Reynolds stress models. c4 = −0.4 and 14.4) (c1 + c3 ) ∂¯1 v ∂x2 3 a12 = −cµ τ 1 ∂¯1 v + τ3 ∂x2 4 (−c5 + c6 + c7 ) Using values on the constants as in [35]. v ¯ Let’s take a closer look on Eq. 37. EARSM) in the literature are the models presented in [36.5 we get a11 = a22 = a33 a12 0. c6 = −0.5 2 k− kτ 3 12 ∂¯1 v ∂x2 ∂¯1 v ∂x2 2 2 v −0. Let’s write down Eq. When v we discussed streamline curvature effects at p.82 2 k+ kτ 3 12 2 0. In Eqs. 35. c2 = 0.8 c5 = 0.5 we have assumed that the only strain is ∂¯1 /∂x2 . 110 we found that it is important to investigate the effect of secondary strains such as ∂¯2 /∂x1 . i.e c1 = −0.16 2 ∂¯1 τ = 12 ∂x2 k ∂¯1 v = −cµ ε ∂x2 2 0. more accurate than linear eddy-viscosity models and they don’t give rise to any numerical instabilities as in implicit solvers (like SIMPLE).21. They are computationally cheap.11.16 2 ′2 kτ ⇒ v3 = k − 3 12 ∂¯1 v ∂x2 2 (14. we obtain a11 = a22 = a33 1 2 τ 12 1 2 τ 12 ∂¯1 v ∂x2 ∂¯1 v ∂x2 ∂¯1 v ∂x2 2 (c1 + 6c2 + c3 ) 2 (c1 − 6c2 + c3 ) 2 1 = − τ2 6 (14.05.6) (c1 + c3 ) 1 a12 = − τ 3 4 3 (c5 + c6 + c7 ) . EARSMs are very popular — especially the model in [38] — in the aeronautical community where explicit time-marching solvers are used.5) We ﬁnd that indeed the non-linear model gives anisotropic normal Reynolds stresses.3 in fully developed channel ﬂow (¯2 = v3 = ∂/∂x1 = ∂/∂x3 ≡ 0). 14. 14.5 2 τ 12 ∂¯1 v ∂x2 ∂ u1 ¯ ∂x2 2 ′2 ⇒ v1 = 2 ′2 ⇒ v2 = 2 2 0. In implicit solvers a large turbulent viscosity in the diffusion term of the momentum equations is needed to stabilize the solution procedure.3 v for the strain ∂¯2 /∂x1 v a11 = a22 = a33 1 2 τ 12 1 2 τ 12 ∂¯2 v ∂x1 ∂¯2 v ∂x1 ∂¯2 v ∂x1 ∂¯2 v ∂x1 2 (c1 − 6c2 + c3 ) 2 (c1 + 6c2 + c3 ) 2 1 = − τ2 6 (14. 14.

and hence the model is slightly more sensitive to the secondary strain ∂¯2 /∂x1 than to the primary v one ∂¯1 /∂x2 . the non-linear models are able to account for streamline curvature.5.7) 2 a33 = − .82 2 τ 12 0. 14. Non-linear Eddy-viscosity Models 117 Inserting with values on the constants from [35] (see above) we obtain a11 = − a22 = 0. a12 = 0 As can be seen the coefﬁcient for a22 is larger than that in Eq. v but due to the choice of constants so that c5 + c6 + c7 = 0 this effect is weak.16 2 τ 12 2 (14.14. Thus. .5 2 τ 12 ∂¯2 v ∂x1 ∂¯2 v ∂x1 ∂¯2 v ∂x1 2 0.

11. 11. 2 ′2 In the V2F model the problem of accounting for the wall damping of v2 is simply ′2 equation in boundary-layer form resolved by solving its transport equation. v2 . The damping of v2 is in the RSM accounted for through the modelled pressure-strain terms Φ22. 15. 11. 40.1w and Φ22.2 can be written as ∂ρ¯1 v2 v ′2 ∂v ′2 ∂ρ¯2 v2 v ′2 ∂ v ′2 (µ + µt ) 2 + ρP − ρ 2 ε + = ∂x1 ∂x2 ∂x2 ∂x2 k (15. In usual eddy-viscosity models both 2 ′2 these effects are accounted for through damping functions. 31] two additional equations. 11.e. The V2F Model 118 15 The V2F Model In the V2F model of [39. They go to zero far away from the wall (x+ 10). are solved: the wall-normal stress v2 and a function f . the pressure-strain term) whereas the viscous damping takes place within the viscous and buffer layer (x+ 10).15.2w (see Eqs. The model for the dissipation ε22 is ¯ ¯ taken as in RSM (see Eq.43 at p.91 and Eq.45) εmodel = 22 ′2 v2 ε k Add and subtract εmodel on the right side of Eq.4) (15.3) ′2 P is the source term in the v2 -equation above. 89. Note that the production term P22 = 0 because in boundary-layer approximation v2 ≪ v1 and ∂/∂x1 ≪ ∂/∂x2 . This is a model which is aimed at improving modelling of wall effects on the turbulence. in two ways.1) in which the diffusion term has been modelled with an eddy-viscosity assumption. The v2 reads (see Eq.2) In the V2F model P is now deﬁned as 2 ′ ∂p′ v ′2 − ε22 + 2 ε P = − v2 ρ ∂x2 k so that Eq.92). 9. ′2 Walls affect the ﬂuctuations in the wall-normal direction.16 at p. apart from the k and ε′2 equations. and it includes the velocity-pressure gradient term and the difference between the exact and the modelled dissipation. 74) ′ ∂ρ¯1 v2 v ′2 ∂v ′2 ∂ρ¯2 v2 v ′2 ∂ ′ ∂p − ρε (µ + µt ) 2 − 2v2 + = 22 ∂x1 ∂x2 ∂x2 ∂x2 ∂x2 (15. 15.1 yields 22 ∂ρ¯1 v2 v ′2 ∂ρ¯2 v2 v ′2 + = ∂x1 ∂x2 ′2 ′ ∂v ′2 ∂ v2 v ′2 ′ ∂p − ρε (µ + µt ) 2 − 2v2 ε−ρ 2 ε 22 + ρ ∂x2 ∂x2 ∂x2 k k (15. Note that this term is commonly split into the pressure-strain term and a diffusion term as ′ v2 ∂p′ ∂v ′ ∂v ′ p′ = 2 − p′ 2 ∂x2 ∂x2 ∂x2 . see Eq. The wall ′2 is felt by the turbulence fairly far from the wall (x+ damping of v2 200) through the 2 pressure ﬁeld (i.

σε = 1. CT ε ′2 2 v2 − 3 k ′2 2 v2 − k 3 ν ε 1/2 νt + C2 k ν3 ε ∂¯1 v ∂x2 1/4 2 (15.5) L = CL max k 3/2 . ′2 ′2 The boundary condition for f is obtained from v2 equation. However. and the second the rapid term. The constants are given the following values: cµ = 0.6) 0=ν ∂x2 k 2 The ﬁrst and the last term behave as O(x2 ) as x2 → 0 because Taylor analysis gives 2 ′2 v2 = O(x4 ).23.9) In the V2F model a transport equation for the normal stress normal to walls is solved ′2 for. Cη ε where Φ22 is the IP model of the pressure-strain term. ′2 Above we have derived the v2 equation in boundary layer form assuming that x2 is the wall-normal coordinate. then v 2 = v2 .3.2.8) For more details. i. 11.44. ε = O(x0 ) and k = O(x2 ). three-dimensional ﬂow it reads ∂v 2 v2 ∂ ∂ρ¯j v 2 v (µ + µt ) + ρf k − ρ ε = ∂xj ∂xj ∂xj k (15. if a wall lies in . CT = 6. If the wall lies in the x1 − x3 plane. A = B = 0.9.3. CL = 0.53 and 11.6 gives 0= ′2 ′2 ∂ 2 v2 f εx2 2v2 2 2 + 2ν 2 − x2 ∂x2 2 (15. Near the wall. Cη = 90. σk = 0.9. ε = 2νk/x2 [5] . see Eqs.15. this equation turns into an ordinary second-order differential equation with the solution ′2 v2 = Ax2 + 2 B x4 − εf 2 2 x2 20ν ′2 Since v2 = O(x4 ) as x2 → 0. ce1 = 1.e. Furthermore.86. C1 = 1. A new variable f = P/k is deﬁned and a relaxation equation is formulated for f as L2 Φ22 1 ∂2f −f =− − ∂x2 k T 2 T = max Φ22 C1 = k T k . 2 2 2 2 using this expression to replace k in Eq.3. the v2 equation reads ′2 v ′2 ∂ 2 v2 + fk − 2 ε (15. In general. cε2 = 1. The V2F Model 119 Physically. 15. see example in Section 11. the ﬁrst term being the slow term. see [5].7) Assuming that f and ε are constant very close to the wall. see [41]. both constants must be zero.12. the main agent for generating wall-normal stress is indeed the pressurestrain term via re-distribution. C2 = 0. so we get 2 f =− ′2 20ν 2 v2 ε x4 2 (15.

The inﬂuence of the lengthscale L is nicely illustrated: the larger L.2. because far from the wall when ∂ 2 f /∂x2 ≃ 0.1 where the equation L2 ∂2f −f +S =0 ∂x2 2 (15. S=0.S=1 L=0.e. As can be seen. as required. 15. 15. Thus the f equation acts so as to let f go from the value of its source term to its (negative) wall value (see Eq.8 1 0 0 x2 Figure 15.2 0. In the V2F model the turbulent viscosity is computed from ′2 νt = Cµ v2 T (15.12 ′2 the x2 − x3 plane.6 0. The V2F Model 120 2 1.4 we get ∂ρ¯1 v2 v ′2 ∂v ′2 2 ∂ρ¯2 v2 v ′2 ∂ (µ + µt ) 2 + ρΦ22 − ρε + = ∂x1 ∂x2 ∂x2 ∂x2 3 (15.S=1 L=0. ′2 the source term in the v2 -equation simpliﬁes to Φ22 plus isotropic dissipation (see Eq.1).S=1 L=0. for example. 15. v 2 = v ′2 . This is what happens.8) over lengthscale L.5 yields (T = k/ε) ′2 kf ≡ P → Φ22 + ε(v2 /k − 2/3) (15. 15. 15.08.15. the further away from the wall does f go to its far-ﬁeld value.13) .04.1: Illustration of Eq.12. 15. This is done automatically since in the general into an equation for v1 1 formulation in Eq. 15. 15.4 0. ∂¯1 /∂x2 in the expression for Φ22 is replaced by P k . Furthermore. S=2 L=0. f is. This is how the reduction of the source term P in Eq.5) for different right sides is illustrated in the Fig.5 has the form it has? Far from the wall.4 is achieved as the wall is approached.5 0. this means that the transport equation for v2 is turned ′2 . reduced as the wall is approached.2.10) When this expression is inserted in Eq. v v Why does the right side of Eq. f approaches the value of the source term as x2 > L.5 f 1 0.S=1 L=0.11) ′2 which is the usual form of the modelled v2 -equation with isotropic dissipation.9. and 2 Eq.5 L=0. 15.12) has been solved with f = 0 at the wall and with different L and S. i. If the v wall lies in the x2 − x3 plane the largest velocity gradient will be ∂¯2 /∂x1 or ∂¯3 /∂x1 . 15. The behavior of the equation for f (Eq.2.

14) The boundary condition for f makes the equation system numerically unstable. They introduced a new variable f ∗ = f − 5εv 2 /k 2 and they neglected the term −5L2 ∂2 ∂xj ∂xj εv 2 k2 ′2 The resulting v2 and f ∗ -equation read [43] ∂¯j v 2 v ∂v 2 v2 ∂ (ν + νt ) + kf ∗ − 6 ε = ∂xj ∂xj ∂xj k v2 1 Pk 2 ∂2f ∗ (C1 − 6) − (C1 − 1) + C2 + f∗ = − ∂xj ∂xj T k 3 k ∂¯j ∂¯i v ∂¯i v v + P k = νt ∂xj ∂xi ∂xj ν 1/2 k . ε and v2 . 15.6 T = max ε ε L = CL max k 3/2 . namely f = 0.15. Modiﬁed V2F model 121 The k and ε-equations are also solved (without damping functions). the boundary conditions are given again ′2 k = 0. . v 2 = 0 ε = 2νk/x2 2 f∗ = 0 This modiﬁed model is numerically much more stable.1 Modiﬁed V2F model In [43] they proposed a modiﬁcation of the V2F model allowing the simple explicit boundary condition f = 0 at walls.1. v2 = 0 ε = 2νk/x2 2 ′2 20ν 2 v2 f =− 4 εx2 (15. Note that the modiﬁed model is identical to the original model far from the wall. f equations coupled [41]. An alternative is to use the ζ − f model [42] which is more stable. ′2 One way to get around that problem is to solve both the k. In ′2 ′2 this model they solve for the ratio v2 /k instead of for v2 which gives a simpler wall boundary condition for f .16) Boundary conditions at the walls are k = 0.15) −L2 (15. For convenience. Cη ε ν3 ε 1/4 (15.

it was found that it is important to impose the limitation on T in a consistent manner.√ 1/2 ε 6Cµ v 2 (2¯ij sij ) s ¯ (15. 15. This is not ensured in the V2F models presented above.16 as an upper bound on the source term kf in the v 2 -equation. 45.15.2 Realizable V2F model The realizable condition for stagnation ﬂow (see p. In the V2F model. i. − 15.6k k . It turns out that in the region far away from the wall. Since v 2 represents the wall-normal normal stress.3 To ensure that v 2 ≤ 2k/3 [1] ε 2k (C1 − 1) + C2 P k (C1 − 6)v 2 − k 3 (15.√ 2 (¯ s )1/2 ε 6Cµ v sij ¯ij k 3/2 k 3/2 L = min .e. 2 ′2 ′2 ′2 ′2 ′2 v2 ≤ v1 and v2 ≤ v3 . and it was found that the limitation on T is indeed important. i. 112) is used also for the V2F model. Then Eq. For instance. The source term kf in the v 2 -equation (Eq. 47]. A simple modiﬁcation 3 is to use the right side of Eq. 2 vsource = min kf. the Laplace term is not negligible. ∂ 2 f /∂xj ∂xj → 0. v 2 denotes the generic wall-normal stress. the Laplace term is assumed to be negligible i.15) includes the modelled velocitypressure gradient term which is dampened near walls as f goes to zero. if the limit is used in the f equation.16 reduces to f = right side. it should be the smallest normal stress. 15. Realizable V2F model 122 15. Below the simple modiﬁcation proposed by [1] is presented.e. and as a consequence v 2 gets too large so that v 2 > 2 k. . In the homogeneous region far away from the wall. see [1]. whereas that for L is not. Furthermore.2. and thus v2 should be smaller than or equal to 3 k.15.17) These realizable conditions have been further investigated by Sveningsson [41.18) This modiﬁcation ensures that v 2 ≤ 2k/3. For more details.e. it must for consistency also be used for ε/k in Eq. 44. 15. 46. and they read [43] T = min 0. Thus it should be the smallest one. 15.

1) .e. Red: high pressure. where β ∗ = cµ . dω/dt. it was suggested in [48] to combine the two models. 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). see Eq. 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 dissipation) and it requires near-wall modiﬁcation (i. blue: low pressure 16 The SST Model The SST (Shear Stress Transport) model of [48] is an eddy-viscosity model which includes two main novelties: 1. A limitation of the shear stress in adverse pressure gradient regions.16. 2. 2. Consider the upper surface (suction side).e. Let us express the lefthand side of the ω equation as a combination of the left-hand sides of the ε and the k equations by using the chain rule. i. The lefthand side of the ω equation will consist of the convection term. the disadvantage of the standard k − ω model is that it is dependent on the free-stream value of ω [50].1: Flow around an airfoil.23. The k − ω model is better than the k − ε model at predicting adverse pressure gradient ﬂow and the standard model of [49] does not use any damping functions. which denotes the material derivative assuming steady ﬂow. Pressure contours. However. In order to improve both the k − ε and the k − ω model. Before doing this. the pressure decreases because the velocity increases.15) the pressure reaches its minimum and increases further downstream as the velocity decreases. 16. dω ε d ε d(1/k) 1 dε = + ∗ = ∗ ∗k dt dt β β k dt β dt ε dk 1 dε ω dk 1 dε − ∗ 2 = ∗ − = ∗ β k dt β k dt β k dt k dt (16. low-Re number damping functions/terms) One example of adverse pressure gradient is the ﬂow along the surface of an airfoil.1. The SST Model 123 Figure 16. it is convenient to transform the k − ε model into a k − ω model using the relation ω = ε/(β ∗ k). Starting from the leading edge. This region is called the adverse pressure gradient (APG) region. At the crest (at x/c ≃ 0. see Fig.

2) and one from the k equation ω − P k (the second term at the right side in Eq.16.1) (16.tfd.1) (16.se/˜lada) T Dω = 2νt ∂k ∂ω ω + σε k ∂xj ∂xj k 1 1 − σε σk νt νt − σε σk ∂2k + ∂x2 j νt ∂ω σε ∂xj ω + k ∂ ∂νt ∂k + ∂xj ∂xj ∂xj (16.4) T Turbulent diffusion.6) • Viscous diffusion term ν Dω = ν ∂2ε νω ∂ 2 k ν ∂ 2 ωk νω ∂ 2 k − = − β ∗ k ∂x2 k ∂x2 k ∂x2 k ∂x2 j j j j ω ∂k ∂ω +k ∂xj ∂xj −ν ω ∂2k k ∂x2 j = ν ∂ k ∂xj ν = k = ∂ω ∂k ∂2k ∂k ∂ω ∂2ω ω ∂2k +ω 2 + +k 2 −ν ∂xj ∂xj ∂xj ∂xj ∂xj ∂xj k ∂x2 j ν ∂ω ∂xj (16. The right side should be transformed in the same manner.3) k In the same way we transform the entire right side inserting the modelled equations for k and ε so that Dω ω ω 1 1 = ∗ Pε − P k − ∗ Ψε − Ψk + Dt β k k β k k Production.5) • Destruction term Ψω = (16. one term from the ε equation 1 Pε β∗k (the ﬁrst term at the right side in Eq.8) . 16. 16. Dω (16.chalmers. The SST Model 124 Now we have transformed the left side of the ω equation.7) ∂ 2ν ∂ω ∂k + k ∂xj ∂xj ∂xj The turbulent diffusion term is obtained as (the derivation is found in [51] which can be downloaded from www. Dω • Production term Pω = ε ω ω 1 Pε − P k = Cε1 ∗ 2 P k − P k β∗k k β k k ω k = (Cε1 − 1) P k ε2 ω ω Ψk = Cε2 − ε k k k ∗ 2 = (Cε2 − 1)β ω 1 β∗k Ψε − (16. For example. Pω Destruction. the production of the ω equation will consist of two terms. Ψω 1 β∗k T Dε ω T − Dk k ν ∂2ε νω ∂ 2 k + ∗ − β k ∂x2 k ∂x2 j j ν Viscous diffusion.

Since the standard k − ω model does not include any cross-diffusion term.10) α = Cε1 − 1 = 0. ∗ ωy y 2 ω β CDω y 2 are used. ξ = min max . β = (Cε2 − 1)β ∗ = 0. (16.10) should only be active in the k − ε region. for example. the Boussinesq assumption can be cµ k 2 ∂¯1 cµ k 2 ∂¯1 v v = = c1/2 k µ ∂x2 ε ∂x2 ε2 νt . the last term in the ω equation (second line in Eq. the predicted shear stress is too large.14) 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 −v1 v2 ≃ cµ k.King model [JK]) which is based on transport of the main shear ′ ′ stress v1 v2 .0828 Since the k − ε model will be used for the outer part of the boundary layer. If we assume that σk = σε in the second and third term of the right-hand side. hence it is multiplied by (1 − F1 ).k−ε k ω νt . which . At p. F1 = 1 in the near-wall region and F1 = 0 in the outer region. 123 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. the v1 v2 transport equation is built on Bradshaw’s assumption [52] ′ ′ −v1 v2 = a1 k (16. is computed as βSST = F1 βk−ω + (1 − F1 )βk−ε (16. Functions of the form √ k 500ν 4σω2 k 4 . In boundary layer ﬂow. 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.12) where βk−ω = 0.11) F1 = tanh(ξ ).9) We can now ﬁnally write the ε equation formulated as an equation for ω ∂ ∂ ν+ (¯j ω) = v ∂xj ∂xj νt 2 ν+ + k σε νt ∂ω ω + α P k − βω 2 σε ∂xj k ∂k ∂ω ∂xi ∂xi (16. 16. predicts adverse pressure gradient ﬂows much better than the k − ω model.16.44. 123). the author in [48] noted that a model (the Johnson . The βcoefﬁcient.3.13) where a1 = cµ written as ′ ′ −v1 v2 = 1/2 = β ∗1/2 . This brings us to the second modiﬁcation (see p.075 and βk−ε = 0. we can considerably simplify the turbulence diffusion so that T Dω = ∂ 2νt ∂k ∂ω + σε k ∂xj ∂xj ∂xj νt ∂ω σε ∂xj (16. ′ ′ In the JK model. the viscous part of the cross-diffusion term (second line) is usually neglected (the viscous term are negligible in the outer region). When introducing this second modiﬁcation.0828.k−ω ∂¯1 v ∂x2 2 1/2 = c1/2 k µ Pk ε 1/2 (16. The SST Model 125 In the standard k − ε model we have σk = 1 and σε = 1. Still.

18) ¯ ¯ where sij Ωij = 0 because sij is symmetric and Ωij is anti-symmetric. . Eq. The SST Model 126 explains why Eq. Eq. in (a) the v Boussinesq assumption together with Eq.13. 16.17) (16.17 shows that (it is likely that) the second part is only used in APG regions and not elsewhere.15b) ¯ ¯ where Ω is the absolute vorticity (in boundary layer ﬂow Ω = ∂¯1 /∂x2 ). 11. 155. Then we take the minimum of (a) and (b) so that νt = cµ k 1/2 ¯ max(cµ ω.4 at p. To summarize the SST modiﬁcation: • the second part. i.King model.19) The ﬁnal form of the SST model is given in Eq. i. Equation 16. The second modiﬁcation s in the SST model is that the production term in the new SST model is limited by 10ε.13 were used and (b) is taken from the k − ω model.12. [48] proposed to re-deﬁne the turbulent eddy viscosity including the expression in Eq. 16.14 over-predicts the shear stress and works poorly in this type of ′ ′ ﬂow.16 reduces νt according to the Johnson . 16. i. 16. in regions where P k < ε. 16.16 returns to νt = k/ω as it should.14 in adverse pressure gradient ﬂow. The ﬁrst modiﬁcation is that the absolute vorticity Ω in Eq.15a) (16.16 (which corresponds to the usual Boussinesq model). when Ω is large). 16. To reduce |v1 v2 | in Eq. 10ε (16.35). We have two expressions for the turbulent viscosity νt = ′ ′ −v1 v2 cµ k ¯ = Ω ¯ Ω 1/2 cµ k k νt = = 1/2 ω cµ ω 1/2 (16. It is important to ensure that this limitation is not active in usual boundary layer ﬂows where P k ≃ ε. 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. Equation 16.e.e. |¯|2 = s ∂¯j ∂¯i v v ∂¯i v ¯ + = 2¯ij (¯ij + Ωij ) = 2¯ij sij s s s ¯ ∂xj ∂xi ∂xj v ∂¯j v 1 ∂¯i ¯ − Ωij = 2 ∂xj ∂xi 1/2 (16.16 (which mimics the Johnson-King model).16. should be used in APG ﬂow • the ﬁrst part.e.new = min P k .16 ¯ ¯ with |¯| limits νt in stagnation regions similar to Eq.e. Two modiﬁcations have ¯ been introduced [53].15a.16) ¯ When the production is large (i. 16. It can be seen that νt is 1/2 ¯ reduced only in regions where P k > ε.16 1/2 has been replaced by |¯| = (2¯ij sij ) s s ¯ which comes from the production term using the Boussinesq assumption (see Eq. the SST model has been slightly further developed. Pk. k/ω in Eq. 16. F2 Ω) 1/2 (16. should be in the remaining of the ﬂow. because if P k < ε then Ω < cµ ω since ω ωε 1 ¯ ¯ = cµ ω 2 Ω2 = νt Ω2 = P k < νt k k Hence. Eq. cµ k/Ω in Eq. 16. Today. 16. 13. 19.

1: Filtering the velocity.1 2 2. The equations for the ﬁltered variables have the same form as Navier-Stokes.35 0. the ﬁltered variables are functions of space and time.2) (17.4 127 v1 .3 0.5 4 Figure 17. Large Eddy Simulations 0.5∆x 1 ¯ Φ(x.1) (note that we use the notation .2 0.5 3 no filter one filter x1 3. This is called Reynolds time averaging: Φ = 1 2T T Φ(t)dt. In LES we ﬁlter (volume average) the equations. we have here ′′ vi = 0 (17.1) x+0.3) v i = vi ¯ ¯ . 17.25 0.15 0.5∆x ¯ Φ = Φ + Φ′′ Since in LES we do not average in time.5 0.45 0. for time averaging). ¯ 1 ∂p ∂ 2 vi ¯ ∂τij ∂ ∂¯i v (¯i vj ) = − v¯ +ν − + ∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj ∂ vi ¯ =0 ∂xi where the subgrid stresses are given by ¯¯ τij = vi vj − vi vj ′ Contrary to Reynolds time averaging where vi = 0. Φ = Φ + Φ′ −T (17.1 Time averaging and ﬁltering In CFD we time average our equations to get the equations in steady form. i. v1 ¯ 0. In 1D we get (see Fig. t) = Φ(ξ.e. 17 Large Eddy Simulations 17. t)dξ ∆x x−0.17.

i. i. For simplicity we do it in ¯ 1D.e. the size of the control volume varies in space. vi vj . if not otherwise stated.2. see Fig. i. see ¯ ¯ p. Let’s look at the ﬁltering of Eq.2. for example. ∂p ∂ = ∂xi ∂xi 1 V pdV V +O V2 = ∂p ¯ +O V2 ∂xi All linear terms are treated in the same way. it can be shown that if V is a function of xi . if V is constant. ∂ ∂vi vj = ∂xj ∂xj 1 V vi vj dV V +O V2 = ∂ (vi vj ) + O V 2 ∂xj There is still a problem with the formulation of this term: it includes an integral of a ¯¯ product. i.e. V . To achieve this we simple add the term we want (¯i vj ) and subtract the one we don’t want ( vi vj ) v¯ on both the right and left side. i. 17. if a variable is time averaged twice ( v ). This is because v is not dependent on time. reads 1 ∂p ∂p = dV ∂xi V V ∂xi Now we want to move the derivative out of the integral.2 in more detail. 17. Now we take a look at the non-linear term in Eq.e. Fortunately. 17.) vI = ¯ = 1 ∆x 1 ∆x ∆x/2 v (ξ)dξ = ¯ −∆x/2 1 ∆x 0 ∆x/2 v (ξ)dξ + ¯ −∆x/2 0 v (ξ)dξ ¯ = ∆x ∆x vA + ¯ vB . Now let’s move the derivative out of the integral. vi vj . the ﬁltering volume. ¯ ¯ ¯ In LES.e. The pressure gradient term. From Eq.e. it is an error of second order.e. i. This is how we end up with the convective term and the SGS term in Eq. In this course we use box ﬁlters. it is the same as time averaging once ( v ). In general. i. 130. in ﬁnite volume methods. 17. If the ﬂow is unsteady. First we ﬁlter the term and move the derivative out of the integral. 17. When is that allowed? The answer is “if the integration region is not a function of xi ”. we want it to appear like a product of integrals. Let’s ﬁlter vI once more (ﬁlter size ∆x. In practice this requirement is rarely satisﬁed. the convective term. i. (Note that subscript I denotes node number. Differences between time-averaging (RANS) and space ﬁltering (LES) 128 This is true for box ﬁlters. In ﬁnite volume methods. ¯ 2 2 . v = v (and since v = v + v ′′ we get v ′′ = 0). box ﬁlters are always used.2. is (almost always) identical to the control volume. However. Note that for the spectral cut-off ﬁlter vi = v i .1 we get v = 1 2T T v dt = −T 1 v 2T = v 2T This is obvious if the ﬂow is steady. Since this is the order of accuracy of our ﬁnite volume method anyway.2.e. we can accept this error. ∂ v /∂t = 0. the error we do when moving the derivative out of the integral is proportional to V 2 . ∂/∂t ≪ 1/T . i.e.2 Differences between time-averaging (RANS) and space ﬁltering (LES) In RANS. we must assume a separation in time scales so that the variation of v during the time interval T is negligible.17. 17.e.

17. which gives ¯ vI = ¯ 1 3 3 1 1 vI−1 + vI + ¯ ¯ vI + vI+1 ¯ ¯ 2 4 4 4 4 1 = (¯I−1 + 6¯I + vI+1 ) = vI v v ¯ ¯ 8 (17.3. which is second-order accurate. was used to estimate the integrals. and to model (small) subgrid-scales (SGS). E(κ) I cut-off II resolved scales III SGS κ Figure 17. The limit (cut-off) between GS and SGS is supposed to take place in the inertial subrange (II). v at locations A and B (see Fig.2) is estimated by linear interpolation. 17. see Fig. 17.3: Spectrum of velocity.3 Resolved & SGS scales The basic idea in LES is to resolve (large) grid scales (GS). Resolved & SGS scales ∆x 129 x I −1 A I x B I +1 x Figure 17.4) 17. . The trapezoidal rule.2: Box ﬁlter illustrated for a control volume.3.

Hence the cut-off ﬁlter ﬁlters out all scales with wavenumber larger than the cut-off wavenumber κc = π/∆.17. Note that it is physically meaningful to use Fourier transforms only in a homogeneous coordinate direction.8) √ where κ denotes the wavenumber and ı = −1. energy-containing scales inertial subrange (Kolmogorov −5/3-range) dissipation subrange 130 17. in nonhomogeneous directions the Fourier coefﬁcients – which are not a function of space – have no meaning.9) ˆ ˆ ˆ ˆ ˆv v = GC Gˆ = GC v = v ˆ (17.4 The box-ﬁlter and the cut-off ﬁlter The ﬁltering is formally deﬁned as (1D) ∞ v (x) = ¯ −∞ GB (r)v(x − r)dr 1/∆. The ﬁlter in spectral space is particular simple: we simply set the contribution from wavenumbers larger than cut-off to zero.4.6) The Fourier transform is deﬁned as (see Section C) v (κ) = ˆ and its inverse v(r) = 0 1 2π ∞ v(r) exp(−ıκr)dr 0 ∞ (17. Using the convolution theorem (saying that the integrated product of two functions is equal to the product of their Fourier transforms) the ﬁltering in Eq.5) GB (r)dr = 1 −∞ It is often convenient to study the ﬁltering process in the spectral space.9) = 0 ∞ exp(−ıκρ) exp(−ıκ(η − ρ))GC (ρ)v(η − ρ)dρdη = 0 0 ˆ exp(−ıκρ) exp(−ıκξ)GC (ρ)v(ξ)dξdρ = GC (κ)ˆ(κ) v If we ﬁlter twice with the cut-off ﬁlter we get (see Eq. 17.7) v (κ) exp(ıκr)dκ ˆ (17. It is deﬁned as ˆ GC (κ) = 1/∆ if κ ≤ κc 0 otherwise (17. if r > ∆ ∞ GB (r) = (17. if r ≤ ∆/2 0. 17.5 is conveniently written ∞ v (κ) = v (κ) = ˆ ¯ 0 ∞ ∞ 0 v (η) exp(−ıκη)dη ¯ = 0 ∞ ∞ 0 ∞ exp(−ıκη)GC (ρ)v(η − ρ)dρdη (17.10) . The box-ﬁlter and the cut-off ﬁlter I: II: III: large.

17. This means that the ﬁltering is the same as the discretization (=integration over the control volume which is equal to the ﬁlter volume. We require it to be homogeneous.5 0 −0.c = 2π/(2∆x1 ) = π/∆x1 (17.4a reads ′ v1 = 0.5 (a) One sinus period covering two cells x1 /L (b) One sinus period covering four cells Figure 17.5 3 x1 /L 4 1 5 1. in other words.4).13) 2¯ij sij ≡ (CS ∆) |¯| s ¯ s .c L = κ1. does not vary with x1 . v1 . v1. i. and let it be a function of x1 .9 and 17. 17. Let’s choose the ﬂuctuating ′ velocity in the x1 direction. what is the highest wavenumber that is resolved? Or. κ1 = 2π/L (17. Furthermore implicit ﬁltering is employed. for the cut-off ﬁlter it is vice versa.6. Sinus curves with different wavenumbers illustrated in physical space. what is the cut-off wavenumber? The wave shown in Fig. 17. 17.17. If we deﬁne this as the cut-off wavenumber we get κ1.rms .4b.2 0. In ﬁnite volume methods box ﬁltering is always used.4 ′ v1 0.c = 2π/(4∆x1 ) = π/(2∆x1 ).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 box ﬁlter is sharp in physical space but not in wavenumber space. The simplest model is the Smagorinsky model [54]: 1 τij − δij τkk = −νsgs 3 νsgs = (CS ∆) 2 ∂¯i v ∂¯j v + ∂xj ∂xi = −2νsgs sij ¯ 2 (17.8 sin(κ1 x1 )] . If we require that the highest resolved wavenumber should be covered by four cells (∆x1 /L = 0. Now we ask the question: on a given grid.4a really is resolved since the sinus wave covers only two cells. its RMS. However this is the usual deﬁnition of the cut-off wavenumber.11) and it covers two cells (∆x1 /L = 0.5 Highest resolved wavenumbers Any function can be expressed as a Fourier series such as Eq. as in Fig. Highest resolved wavenumbers 131 0. i. 17.5.25).4: Physical and wavenumber space. see Eq. 17. using Eqs.e.8 (see also Eq. then the cut-off wavenumber is given by κ1.5).e.12) ′ It is of course questionable if v1 in Fig.5 2 1 3 1. contrary to the box-ﬁlter (see Eq.27) provided that the coordinate direction is homogeneous.c 2∆x1 = 2π so that κ1. Thus. 17.14). 17.2 0 1 2 0.25 [1 + 0.2 0 node 1 0.2 0 −0. 17. nothing happens when we ﬁlter twice in spectral space.4 ′ v1 0. C.

but in reality dissipation (i. However. the “length”) of ∂¯i /∂xj + ∂¯j /∂xi in the Boussinesq s v v assumption. mixing-length model The eddy viscosity according to the mixing length theory reads in boundary-layer ﬂow [57.5. increase in temperature) takes place at all wave numbers.25 [56]. i. This is a good approximation.18.e.13 17. the SGS viscosity becomes quite large since the velocity gradient is very large at the wall. so must the SGS viscosity. and the dissipation increases for increasing wave number. SGS kinetic energy is dissipated εsgs = −τij sij = 2νsgs sij sij ¯ ¯ ¯ (17. It is found to vary in the range from CS = 0. . We assume that ALL dissipation takes place in the dissipation range. 58] ∂¯1 v . s In the Smagorinsky model the SGS turbulent length scale corresponds to ℓ = CS ∆ so that νsgs = (CS ∆)2 |¯| s which is the same as Eq.7. Disadvantage of Smagorinsky model: the “constant” CS is not constant. 17. transfer of energy from kinetic energy to internal energy. 17.8 Energy path The path of kinetic energy is illustrated in Fig.17) from the resolved turbulence. It should be mentioned that this process is an idealized one. Near the wall. i.e. A damping function fµ is added to ensure this fµ = 1 − exp(−x+ /26) 2 (17.e. mixing-length model 132 and the ﬁlter-width is taken as the local grid size ∆ = (∆VIJK ) 1/3 (17. Smagorinsky model vs.7 Smagorinsky model vs. 16. because the SGS turbulent ﬂuctuations near a wall go to zero. 17. but it is ﬂow-dependent. At cut-off.14) The scalar |¯| is the norm (i. we have νt = ℓ2 ∂¯i v ∂¯j v + ∂xj ∂xi ∂¯i v ∂xj 1/2 = ℓ2 (2¯ij sij ) s ¯ 1/2 ≡ ℓ2 |¯|. This energy is transferred to the SGS scales and act as production term (Pksgs ) in the ksgs equation.065 [55] to CS = 0. κn (17.e. see Eq.17.15) A more convenient way to dampen the SGS viscosity near the wall is simply to use the RANS length scale as an upper limit.16) where n is the distance to the nearest wall. ∆ = min (∆VIJK ) 1/3 . The SGS kinetic energy is then transferred to higher wave-numbers via the cascade effect and the kinetic energy is ﬁnally dissipated (ε=physical dissipation) in the dissipation range. νt = ℓ2 ∂x2 Generalized to three dimensions.

18) The Kolmogorov −5/3 law now gives ∞ ksgs = κc Cκ−5/3 ε2/3 dκ (Note that for these high wavenumbers. the Kolmogorov spectrum ought to be replaced by the Kolmogorov-Pau spectrum in which an exponential decaying function is added for high wavenumbers [57.10 LES vs.18. κ 17.e. RANS LES can handle many ﬂows which RANS (Reynolds Averaged Navier Stokes) cannot. SGS kinetic energy 133 E(κ) PK in e rti al gs εs ran P ge ε gs = E(κ) ∝ κ −5/3 ks dissipating range κc Figure 17.19) ksgs = C 2 π In the same way as ksgs can be computed from Eq. is obtained from κc kres = 0 E(κ)dκ 17. Examples are: .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 ∞ k= 0 E(κ)dκ Changing the lower integration limit to wavenumbers larger than cut-off (i. κc ) gives the SGS kinetic energy ∞ ksgs = κc E(κ)dκ (17. Carrying out the integration and replacing κc with π/∆ we get 2/3 ∆ε 3 (17. turbulent scales are resolved. 17.5: Energy spectrum. the reason is that in LES large.17.9. the resolved turbulent kinetic energy. kres . Chapter 3]).

but it is computed. If we apply two ﬁlters to Navier-Stokes [grid ﬁlter and a second.17.22) . isotropic turbulent scales are modelled ⇒ accurate • LES is very much more expensive than RANS. denoted by . )] where ∆ = 2∆ we get ∂vi ¯ ∂ + ∂t ∂xj vivj ¯ ¯ =− 1∂ p ¯ ¯ ∂2 v i ∂Tij +ν − ρ ∂xi ∂xj ∂xj ∂xj (17.11. coarser ﬁlter (test ﬁlter. o Flows with large separation o Bluff-body ﬂows (e.20) where the subgrid stresses on the test level now are given by ¯ ¯ Tij = vi vj − v i v j ∂vi ¯ ∂ + ∂t ∂xj ¯ ∂2 v i ¯ ∂ τ ij 1∂ p +ν − ρ ∂xi ∂xj ∂xj ∂xj ∂ − vi vj − v i v j ¯¯ ¯ ¯ ∂xj =− (17. turbulent structures o Transition • In RANS all turbulent scales are modelled ⇒ inaccurate • In LES only small. ﬂow around a car). The dynamic model test ﬁlter E(κ) SGS grid scales I SGS test ﬁlter scales II 134 cut-off. grid ﬁlter resolved on test ﬁlter resolved on grid ﬁlter III κ κ = π/ ∆ κc = π/∆ Figure 17. unsteady.11 The dynamic model In this model of [59] the constant C is not arbitrarily chosen (or optimized).g. 17.21) vivj ¯ ¯ (17. the wake often includes large.6: Energy spectrum with grid and test ﬁlter.

see Fig.20 and 17. Identiﬁcation of Eqs. 17. 17.17. The test ﬁlter test ﬁlter grid ﬁlter 135 W ∆x P ∆x E Figure 17.12 The test ﬁlter The test ﬁlter is twice the size of the grid ﬁlter. ﬁltering at the test level is carried out in the same way by integrating over . For ex¯ ample.26) 1 1 (¯w ∆x + ve ∆x) = v ¯ 2∆x 2 1 v v ¯ = (¯W + 2¯P + vE ) 4 vW + vP ¯ ¯ vP + vE ¯ ¯ + 2 2 For 3D. 17. 17.24 as 1 1 Lij − δij Lkk = Tij − δij Tkk − 3 3 1 τ ij − δij τ kk 3 (17. The test-ﬁltered variables are computed by integration over the test ﬁlter.25) (17.24) (17.7 v is computed as (∆x = 2∆x) v = ¯ = 1 2∆x E v dx = ¯ W 1 2∆x P E v dx + ¯ W P v dx ¯ (17.22 gives vi vj − v i v j + τ ij = Tij ¯¯ ¯ ¯ The dynamic Leonard stresses are now deﬁned as ¯¯ ¯ ¯ Lij ≡ vi vj − v i v j = Tij − τ ij The trace of this relation reads Lii ≡ Tii − τ ii With this expression we can re-formulate Eq. the 1D example in Fig. ∆ = 2∆. the test ﬁlter is located at lower wave number than the grid ﬁlter.7: Control volume for grid and test ﬁlter.12.e.23) In the energy spectrum.6. 17. i.

test and intermediate level The stresses on the grid level.28) 3 2 1 Tij − δij Tkk = −2C ∆ | s | s ij ¯ ¯ 3 (17.K−1/2 + vI+1/2. Stresses on grid.K+1/2 ) v ¯ (17.K = ¯ 1 (¯I−1/2. K I.17. If we use the Smagorinsky model we get 1 τij − δij τkk = −2C∆2 |¯|¯ij ss (17. 17.J−1/2.K+1/2 + vI+1/2.29) 1/2 where s ij ¯ ¯ ∂vj ¯ 1 ∂ v i + .J−1/2. J − 1.J−1/2.K+1/2 v ¯ +¯I−1/2. K Figure 17.J+1/2.K−1/2 v ¯ +¯I−1/2. ℓ. J + 1. test level and intermediate level (dynamic Leonard stresses) have the form ¯¯ τij = vi vj − vi vj stresses with ℓ < ∆ Tij = vi vj − v i v j stresses with ℓ < ∆ ¯ ¯ Lij = Tij − τ ij stresses with ∆ < ℓ < ∆ Thus the dynamic Leonard stresses represent the stresses with lengthscale. K 136 I − 1. i. 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 ). K x2 x1 I.K−1/2 v ¯ 8 +¯I−1/2.8) v I.K+1/2 + vI+1/2. K I + 1.J+1/2.J−1/2. J. | s |= ¯ = 2 ∂xj ∂xi 2 s ij s ij ¯ ¯ .J. the test cell assuming linear variation of the variables [60].27) 17. (see Fig.e. J. in the range between ∆ and ∆. test and intermediate level I. J.J+1/2.J+1/2.8: A 2D test ﬁlter control volume.K−1/2 + vI+1/2.13 Stresses on grid.13.

14 Numerical dissipation The main function of an SGS model is to dissipate (i. Applying the test ﬁlter to Eq. Use of one-equation models solve these numerical problems (see p.17.e.e. 17. 145).31b) ss ∆ | s | s ij − ∆2 |¯|¯ij ¯ ¯ The error. substituting this equation and Eq.29 into Eq. This is the reason why in LES we should use a central differencing scheme. 17. In real 3D ﬂows. Lilly [61] suggested to satisfy Eq. 17. i. Hence. Usually local averaging and clipping (i.e. 17. and we have ﬁve (¯ij is symmetric and traces less) equations for C. 17.30) ss Lij − δij Lkk = −2C ∆ | s | s ij − ∆2 |¯|¯ij ¯ ¯ 3 Note that the “constant” C really is a function of both space and time.13 at 2 p. All upwind schemes give numerical dissipation in addition to the modelled SGS dissipation. Let us deﬁne the error as the difference between the left-hand side and the right-hand side of Eq.30 in a least-square sense. 17. i. and it has been found necessary to average C in homogeneous direction(s).14. t).131). Q.28 (assuming that C varies slowly). 17.31a gives 1 ∂Q = 4Mij Lij − δij Lkk + 2CMij ∂C 3 =0 (17. Equation 17.30 is a tensor equation. C = C(xi . Carrying out the derivation of 17.33) C=− 2Mij Mij It turns out that the dynamic coefﬁcient C ﬂuctuates wildly both in space and time. MILES [62]).28 is not squared (cf. This causes numerical problems. C must be clipped to ensure that the total viscosity stays positive (ν + νsgs ≥ 0). Q= Mij = 1 Lij − δij Lkk + 2CMij 3 2 2 (17. has a minimum (or maximum) when ∂Q/∂C = 0. Equation 17. Indeed. to dampen) resolved turbulent ﬂuctuations. because this class of schemes does not give any numerical dissipation.g. here we focus on ensuring proper dissipation through an SGS model rather than via upwind differencing. there is no homogeneous direction.32) Since ∂ 2 Q/∂C 2 = 8Mij Mij > 0 it is a minimum. Numerical dissipation 137 Note that C in Eq. there are LES-methods in which upwind schemes are used to create dissipation and where no SGS model is used at all (e. Furthermore. Eq. However.30 raised to the power of two. C should be compared with CS . requiring that C stays within pre-deﬁned limits) of the dynamic coefﬁcient is used. The SGS model is – hopefully – designed to give a proper amount of dissipation. 17.31 is re-written so that Lij Mij (17. the Smagorinsky model.25 gives 2 1 (17.e. It can be shown using Neumann stability analysis that all upwind schemes are dissipative (see Further reading at .31a) (17.

9) vI ¯ ∂¯ v ∂x = vI ¯ vI − vI−1 ¯ ¯ + O (∆x) ∆x (17. The ﬁrst-derivative in the convective term is estimated by ﬁrst-order upwind differencing as (ﬁnite difference.5|¯I |∆x. the second term on the right-hand side will act as an additional (numerical) diffusion term.15. The total diffusion term will have the form diﬀusion term = ∂ ∂x (ν + νsgs + νnum ) ∂¯ v ∂x (17.9: Numerical dissipation.17) εsgs+num = 2(νsgs + νnum )¯ij sij s ¯ For more details on derivation of equations transport equations of turbulent kinetic energies. νnum ≃ 0. http://www. 17. see Fig. Scale-similarity Models vI ¯ I −1 I I +1 138 Figure 17.15 Scale-similarity Models In the models presented in the previous sections (the Smagorinsky and the dynamic ¯¯ models) the total SGS stress τij = vi vj − vi vj was modelled with an eddy-viscosity .tfd.chalmers. 17. Taylor expansion gives ¯ vI−1 = vI − ∆x ¯ ¯ so that vI − vI−1 ¯ ¯ = ∆x ∂¯ v ∂x ∂¯ v ∂x 1 + (∆x)2 2 1 − ∆x 2 ∂2v ¯ ∂x2 + O (∆x)3 I I I ∂2v ¯ ∂x2 I + O (∆x)2 Insert this into Eq. 17. Below it is shown that ﬁrst-order upwind schemes are dissipative. This means that the total v dissipation due to SGS viscosity and numerical viscosity is (cf. see [14].34.35) where the additional numerical viscosity. 17.17.34) where we have assumed vI > 0.34 ∂¯ v ∂x ∂¯ v = vI ¯ ∂x 1 − ∆x 2 ∂ v ¯ ∂x2 2 v ¯ I I O(∆x) +O (∆x)2 where the second term on the right side corresponds to the error term in Eq.se/˜lada/comp turb model/). Eq. When this expression is inserted into the LES momentum equations. 17.

36 it seems natural to assume that the cross term is responsible for the interaction between resolved scales (¯i ) and v ′′ modelled scales (vi ).38) 17.16. i. i. Thus τij = Lij + Cij + Rij ¯ ′′ ¯ ′′ Cij = vi vj + vj vi ′′ ′′ Rij = vi vj . 64] M Cij = cr (¯i vj − v i v j ) v¯ ¯¯ (17. M Lm = Lij + Cij ij . and the last term is denoted the Reynolds SGS stress.37). In scale-similarity models the total stress is split up as ′′ ′′ v ¯¯ τij = vi vj − vi vj = (¯i + vi )(¯j + vj ) − vi vj ¯¯ v ′′ ′′ = vi vj + vi vj + vj vi + vi vj − vi vj ¯¯ ¯¯ ¯ ′′ ¯ ′′ ′′ ′′ v¯ ¯¯ ¯ ′′ ¯ ′′ = (¯i vj − vi vj ) + vi vj + vj vi + vi vj where the term in brackets is denoted the Leonard stresses.e. It was found that this model was not sufﬁciently dissipative. they are exact and don’t need to be modelled.17 Redeﬁned terms in the Bardina Model The stresses in the Bardina model can be redeﬁned to make them Galilean invariant for any value cr (see Appendix H). 17. hence the word ”scale-similar”.e.16 The Bardina Model In the Bardina model the Leonard stresses Lij are computed explicitly. The Bardina Model 139 hypothesis.37) M and Rij = 0 (superscript M denotes Modelled). (smaller than ∆) are similar to the ones just below κc (larger than ∆).39) m ′′ ′′ Rij = Rij = vi vj Note that the modiﬁed Leonard stresses is the same as the “unmodiﬁed” one plus the modelled cross term Cij in the Bardina model with cr = 1 (right-hand side of Eq. In scale-similarity models the main idea is that the turbulent scales just above cutoff wavenumber. since Cij includes both scales. and thus a Smagorinsky model was added M Cij = cr (¯i vj − v i v j ) v¯ ¯¯ M 2 Rij = −2CS ∆2 |¯|¯ij ss (17. ¯¯ ¯¯ Lij = vi vj − vi vj (17. κc .36) Note that the Leonard stresses Lij are computable.17. and the sum of the cross term Cij and the Reynolds term is modelled as [63. Looking at Eq. A modiﬁed Leonard stress tensor Lm is deﬁned as [65] ij m m m τij = τij = Cij + Lm + Rij ij m Cij = 0 Lm = cr (¯i vj − v i v j ) v¯ ¯¯ ij (17. 17. 17. the term in square brackets is denoted cross terms.

**17.18. A dissipative scale-similarity model.
**

E(κ)

10

140

0

ε+ SGS ε− SGS

−10

−20

−30 0 50 100 150

x+ 2

κc

κ

Figure 17.10: Dissipation terms and production term from DNS data. 963 mesh data ﬁltered onto a 483 mesh. Reτ = 500. : −ε+ ; : −ε− ; +: −εSGS . SGS SGS In order to make the model sufﬁciently dissipative a Smagorinsky model is added, and the total SGS stress τij is modelled as τij = vi vj − v i v j − 2(CS ∆)2 |¯|¯ij ¯¯ ¯¯ ss Below we verify that the modiﬁed Leonard stress is Galilean invariant. 1 m∗ ∗ ∗ ¯∗ ¯∗ ¯ ¯ v v v v L = vi vj − v i v j = (¯i + Vi )(¯j + Vj ) − (¯i + Vi ) (¯j + Vj ) cr ij = vi vj + vi Vj + vj Vi − v i v j − v i Vj − Vi v j ¯¯ ¯ ¯ ¯¯ ¯ ¯ 1 ¯¯ ¯¯ = vi v j − v i v j = Lm cr ij (17.40)

(17.41)

**17.18 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 [63]. An eddy-viscosity model has to be added to make the model sufﬁciently dissipative; these models are called mixed models. [66] (can be downloaded from www.tfd.chalmers.se/˜lada) presents and evaluates a dissipative scale-similarity model. The ﬁltered Navier-Stokes read d¯i v ¯ ∂ 2 vi ¯ ∂τik 1 ∂p =ν − + dt ρ ∂xi ∂xk ∂xk ∂xk (17.42)

where d/dt and τik denote the material derivative and the SGS stress tensor, respectively. The SGS stress tensor is given by τik = vi vk − vi vk . ¯¯ (17.43)

When it is modelled with the standard scale-similarity model, it is not sufﬁciently dissipative. Let us take a closer look at the equation for the resolved, turbulent kinetic

17.19. Forcing

′ ′ energy, k = vi vi /2, which reads

141

¯′ ¯′ ¯′ ∂ p′ vi ¯ ¯′ 1 ∂ vk vi vi ∂ vi ¯ ∂ 2 vi ′ ¯′ dk + + =ν v − ¯ + vk vi ¯′ ¯′ dt ∂xk ∂xi 2 ∂xk ∂xk ∂xk i ∂τik ∂ 2 vi ′ ¯′ ∂τik ∂τik ′ vi = ν ¯′ − v − ¯ v = ¯ ∂xk ∂xk ∂xk ∂xk i ∂xk i ∂τik ′ ∂2k ∂¯i ∂¯i v′ v′ − ν −ν v ¯ ∂xk ∂xk ∂xk ∂xk ∂xk i

ε εSGS

(17.44)

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 17.10 presents SGS dissipation, εSGS in Eq. 17.44, computed from ﬁltered DNS data. The forward scatter, ε+ , and back scatter, ε− , SGS dissipation are SGS SGS 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 . 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 vi (see the last term in Eq. 17.44). 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 NavierStokes equations, Eq. 17.42. The second derivative on the right side is usually called a diffusion term because it acts like a diffusion transport term. When analyzing the stability properties of discretized equations to an imposed disturbance, v ′ , using Neumann ¯ analysis (see, for example, Chapter 8 in [67]), this term is referred to as a dissipation term. In stability analysis the concern is to dampen numerical oscillations; in connection 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. 17.44, that we want to dissipate, we must consider the ﬁltered Navier-Stokes equations for the ﬂuctuating velocity, vi . It is the diffusion term in this ¯′ equation which appears in the ﬁrst term on the right side (ﬁrst line) in Eq. 17.44. 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. 17.44). This is achieved by deﬁning a sign function; for details, see [66].

17.19 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 [68].

17.20. Numerical method

142

**17.20 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 vi ¯

n+1/2

= vi + ∆tH v n , vi ¯n ¯ ¯

n+1/2

−α∆t

∂ pn ¯ ∂ pn+1/2 ¯ − (1 − α)∆t ∂xi ∂xi

(17.45)

where H includes convective, viscous and SGS terms. In SIMPLE notation this equation reads ∂ pn+1/2 ¯ n+1/2 = aP vi ¯ anb v 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 + vi,j−1 ) (note that j denotes node number and i is a tensor ¯ vf,i ¯ = 0.5(¯i,j v index) do not satisfy continuity. Create an intermediate velocity ﬁeld by subtracting the implicit pressure gradient from Eq. 17.45, i.e. vi = vi + ∆tH v n , vi ¯∗ ¯n ¯ ¯ ⇒ vi = vi ¯∗ ¯

n+1/2 n+1/2

− (1 − α)∆t

∂ pn ¯ ∂xi

(17.46a) (17.46b)

+ α∆t

∂ pn+1/2 ¯ ∂xi

n+1/2

v Take the divergence of Eq. 17.46b and require that ∂¯f,i v∗ ∂ 2 pn+1 ¯ 1 ∂¯f,i = ∂xi ∂xi ∆tα ∂xi

/∂xi = 0 so that (17.47)

The Poisson equation for pn+1 is solved with an efﬁcient multigrid method [69]. In the ¯ 3D MG we use a plane-by-plane 2D MG. The face velocities are corrected as vf,i = vf,i − α∆t ¯n+1 ¯∗ ∂ pn+1 ¯ ∂xi (17.48)

A few iterations (typically two) solving the momentum equations and the Poisson pressure equation are required each time step to obtain convergence. More details can be found [70] ¯ ¯ 1. Solve the discretized ﬁltered Navier-Stokes equation, Eq. 17.46a, for v1 , v2 and v3 . ¯ 2. Create an intermediate velocity ﬁeld vi from Eq. 17.46b. ¯∗ 3. The Poisson equation (Eq. 17.47) is solved with an efﬁcient multigrid method [69]. 4. Compute the face velocities (which satisfy continuity) from the pressure and the intermediate face velocity from Eq. 17.48 5. Step 1 to 4 is performed till convergence (normally two or three iterations) is reached.

17.20. Numerical method

RANS 2D or 3D steady or unsteady 2nd order upwind 1st order more than two-equations

143 LES always 3D always unsteady central differencing 2nd order (e.g. C-N) zero- or one-equation

Domain Time domain Space discretization Time discretization Turbulence model

Table 17.1: Differences between a ﬁnite volume RANS and LES code. v1 ¯

t t1 : start Figure 17.11: Time averaging in LES. 6. The turbulent viscosity is computed. 7. Next time step. Since the Poisson solver in [69] 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. 17.20.1 RANS vs. LES Above a numerical procedure suitable for LES was described. However, in general, any numerical procedure used for RANS can also be used for LES; for example pressurecorrection methods such as SIMPLE [71, 72] 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 and RANS, see Table 17.1. It is important to use a non-dissipative discretization scheme which does not introduce any additional numerical dissipation, see Section 17.14; 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 modelled and, second, no equation for the turbulent length scale is required since the turbulent length scale can be taken as the ﬁlter width, ∆. t2 : end

17.21. One-equation ksgs model

144

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 our self whenever doing an experiment in, for example, a windtunnel. We start the windtunnel: when has the ﬂow (and turbulence) reached fully developed conditions so that we can start the 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 v1 velocity at ¯ a point may look like in Fig. 17.11. 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 nondimensional time scale from a velocity, V (free-stream or bulk velocity), and a length scale, L (width of a wake, width or 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.

**17.21 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 ∂ ∂ (ν + νsgs ) + Pksgs − ε (¯j ksgs ) = v + ∂t ∂xj ∂xj ∂xj

1/2 ck ∆ksgs ,

(17.49) 3/2 ksgs ¯ ¯ νsgs = Pksgs = 2νsgs sij sij , ε = Cε ∆ Note that the production term, Pksgs , is equivalent to the SGS dissipation in the equation for the resolved turbulent kinetic energy (look at the ﬂow of kinetic energy discussed at the end of [73]).

**17.22 Smagorinsky model derived from the ksgs equation
**

We can use the one-equation model to derive the Smagorinsky model, Eq. 17.13. The length scale in the Smagorinsky model is the ﬁlter width, ∆ ∝ κII , see Fig. 17.12. 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, sij ). Hence, production and dissipation in ¯ Eq. 17.49 are in balance so that ¯ ¯ Pksgs = 2νsgs sij sij = ε Let us replace ε by SGS viscosity and ∆. We can write the SGS viscosity as νsgs = εa (CS ∆)b Dimensional analysis yields a = 1/3, b = 4/3 so that νsgs = (CS ∆)4/3 ε1/3 . Eq. 17.50 substituted into Eq. 17.52 gives

3 νsgs = (CS ∆)4 ε = (CS ∆)4 νsgs (2¯ij sij ) s ¯

(17.50)

(17.51)

(17.52)

⇒ νsgs = (CS ∆)2 |¯| s |¯| = (2¯ij sij ) s s ¯ which is the Smagorinsky model.

(17.53)

1/2

77]) ∂ ∂ ∂ksgs (¯j ksgs ) = Pksgs + v + ∂t ∂xj ∂xj Pksgs = 1 2 with νef f = ν + 2Chom ∆ksgs . being a dynamic model.55) Now we want to ﬁnd a dynamic equation for C∗ . The C in the production term Pksgs is computed dynamically (cf. Chom is computed by requiring that Chom should yield the same total production of ksgs as C. independent of the large scales (ℓ) and the dissipative scales (ν). 75] (see also [76. it has the great advantage that the coefﬁcients are computed rather than being prescribed. ¯ a τij = −2C∆ksgs sij ¯ xyz 2 = 2Chom ∆ksgs sij sij ¯ ¯ xyz The dissipation term εksgs is estimated as: εksgs ≡ νTf (vi. dissipative scales. The equation for the subgrid kinetic energy reads [74.j ) = C∗ ksgs . a constant value (in space) of C (Chom ) is used in the diffusion term in Eq.12: Spectrum for k. 17.e. Usually this problem is ﬁxed by averaging the coefﬁcient in some homogeneous ﬂow direction. 1 2 ¯ ¯ 2C∆ksgs sij sij 1 νef f ∂ksgs ∂xj 1 2 − C∗ ksgs ∆ 3/2 (17. Furthermore. ∆ 3/2 (17.23 A dynamic one-equation model One of the drawbacks of the dynamic model of [59] (see p.54) a −τij vi. vi. isotropic.j . Eq. Below a dynamic one-equation model is presented. The equations for ksgs and K read in symbolic form T (ksgs ) ≡ Cksgs − Dksgs = Pksgs − C∗ T (K) ≡ CK − DK = P K − C∗ ksgs ∆ K 3/2 ∆ 3/2 (17. 134) is the numerical instability associated with the negative values and large variation of the C coefﬁcient.17. The main object when developing this model was that it should be applicable to real industrial ﬂows. In real applications ad-hoc local smoothing and clipping is used. III: Range for small.54 and in the momentum equations. II: the inertial subrange for isotropic scales.56) . energy containing eddies. i. 17. I: Range for the large.23.j . A dynamic one-equation model 145 E(κ) I II III κ Figure 17. To ensure numerical stability.33). 17.

comparing the resolution to the boundary layer thickness. 17. This enables resolution of the near-wall turbulent structures in the viscous sub-layer and the buffer layer consisting of high-speed inrushes and low-speed ejections [100]. 97. 99].e. the left-hand side of the two equations in Eq. trains and buses subjected to sidewinds and wind gusts [90. Then the spectra of the resolved turbulence will exhibit −5/3 range. the ﬂow and heat transfer in a square rotating duct [81. 80]. δ). 94. in the production and the dissipation terms of the ksgs equation (Eq. In this model a dynamic one-equation SGS model is solved. P ksgs − C∗ ksgs ∆ ∆ so that n+1 C∗ = P K − P ksgs + 1 n 3/2 C k ∆ ∗ sgs ∆ K2 3 . In [101. see ﬁgure on p. 137). These structures must be resolved in an LES in order to achieve accurate results. 92]. 17. the ﬂow around a simpliﬁed bus [83. The resolution requirements in wall-parallel planes for a wellresolved LES in the near-wall region expressed in wall units are approximately 100 (streamwise) and 30 (spanwise). 17. 96. detailed SUV [89]. 98. respectively. 88]. The . 17. In outer scaling (i. i. In this way the dynamic coefﬁcients C and C∗ don’t need to be clipped or averaged in any way.57) The idea is to put the local dynamic coefﬁcients in the source terms. An even better approximation should be to assume T (ksgs ) = T (K). 86] and the ﬂow around an airfoil [87.e. we can afford δ/∆x1 and δ/∆x3 in the region of 10 − 20 and 20 − 40. the spectra in the boundary layer will look something like that shown in Fig.40. (17. often called the streak process. This is similar to the requirement in RANS (Reynolds-Averaged Navier-Stokes) using low-Re number models.24 A Mixed Model Based on a One-Eq.17. Model Recently a new dynamic scale-similarity model was presented by [78]. This is a big advantage compared to the standard dynamic model of Germano (see discussion on p.13 [101]. Energy spectra are actually not very reliable to judge if a LES simulation is well resolved or not. 17. A Mixed Model Based on a One-Eq.e.54). 44. 95. i. 82].26 Resolution requirements The near-wall grid spacing should be about one wall unit in the wall-normal direction. 85. 3/2 K 3/2 1 = P K − C∗ .25 Applied LES At the Department we used LES for applied ﬂows such as ﬂow around a cube [79. Model 146 Since the turbulence on both the grid level and the test level should be in local equilibrium (in the inertial −5/3 region). In this case. At low to medium Reynolds numbers the streak process is responsible for the major part of the turbulence production. this kind of resolution can hardly ever be afforded.24. 17. In applied LES. a simpliﬁed car [84. 17. 80]. and the scale-similarity part is estimated in a similar way as in Eq. 91.56 should be close to zero. 102] different ways to estimate the resolution of an LES were investigated. We have also done some work on buoyancy-affected ﬂows [93.

33. 17. 20). 17. Nevertheless. Thus. Number of cells expressed as (δ/∆x1 . the requirement of near-wall grid resolution is the main reason why LES is too expensive for engineering ﬂows. the ﬂow in recirculation regions and shear layer can.44.14: Onera bump. ∆x3 /δin = 0. H . the turbulence in the recirculation re1 3 gion and in the shear layer downstream the bump turned out to be well resolved. 105]. which was one of the lessons learned in the LESFOIL project [104. 20). ◦: (5. ∆x+ = 1300 and ∆x+ = 1800. L1 V1. : channel width. Even if the turbulence in boundary layer seldom can be resolved. 40). Indeed.15. see Fig. The geometry is shown in Fig. the computational resource requirement of accurate LES is prohibitively large. Resolution requirements Thick dashed line −5/3 slope 147 10 −1 E33 (k3 ) 10 −2 10 −3 10 −4 10 0 10 1 10 2 κ3 = 2πk3 /x3.17. 20).14. δ/∆x3 ).max Figure 17. for wall-bounded ﬂows at high Reynolds numbers of engineering interest. (20. Computational domain (not to scale). δ denotes half : (10. 10).13: Energy spectra in fully developed channel ﬂow [101]. suggestion in these works was that two-point correlations is the best way to estimate if an LES is sufﬁciently resolved or not. +: (10. The turbulence in the boundary layer on the bump was very poorly resolved: ∆x1 /δin = 0.in y h x L2 L Figure 17.26. : (10. In [103] the ﬂow (Re ≃ 106 ) over a bump was computed.

v = v + v ′′ ¯ t−T (18. x . In URANS the usual Reynolds decomposition is employed.1) The URANS equations are the usual RANS equations. An alternative to LES for industrial ﬂows can then be unsteady RANS (ReynoldsAveraged Navier-Stokes). v .15: Energy spectra E33 (κ3 ) in the recirculation region and the shear layer : downstream the bump (x1 /H = 1.e. : x2 /H = 0. i. x2 . layer). 18 Unsteady RANS To perform an accurate LES. x2 .18. one is often interested only in the time-averaged ﬂow.34 (in the shear x2 /H = 0. for example. Thick dashed line shows −5/3 slope. which means ¯ that we can decompose the results from an URANS as a time-averaged part. i. but with the transient (unsteady) term retained. ′′ ′′ but also function of time.0035 (near the wall). t). However. if attached boundary layers are important. x3 . t) and vi vj = ¯ ¯ ¯ ¯ ′′ v ′′ (x . ¯ v = v + v ′′ = v + v ′ + v ′′ ¯ ¯ ¯ (18. near walls. p = p(x1 .2). a ¯ resolved ﬂuctuation. t). a very ﬁne mesh must be used. (on incompressible form) ′′ ′′ ∂vi vj ∂¯i v ¯ 1 ∂p ∂ 2 vi ¯ ∂ (¯i vj ) = − v¯ +ν − + ∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj ∂ vi ¯ =0 ∂xi (18. Unsteady RANS 148 10 −1 E33 (κ3 ) 10 −2 10 −3 10 100 200 κ3 = 2πk3 /x3. i. We denote here the time-averaged velocity as v . vi = vi (x1 . : x2 /H = 0. LES will probably give poor predictions in these regions. i.13. LES is very good for wake ﬂow.2) Note that the dependent variables are now not only function of the space coordinates. where the ﬂow is governed by large.3) . which can be captured by a fairly coarse mesh. This causes problems. vi j 1 2 3 Even if the results from URANS are unsteady.e. often denoted URANS (Unsteady RANS) or TRANS (Transient RANS). x . turbulent ﬂuctuation. v ′ . and the modelled.e. unless ﬁne grids are used. turbulent structures.e.max Figure 17. x3 . v (t) = ¯ 1 2T t+T v(t)dt. v ′′ .

νt . the turbulent viscosity. The common deﬁnition of URANS is that the turbulent length scale is not determined by the grid. 18. T . the URANS equations read µt = cµ ρ ∂ρ¯i v ¯ ∂¯i v 1 ∂p ∂ ∂ρ¯i vk v¯ (µ + µt ) =− + + ∂t ∂xk ρ ∂xi ∂xk ∂xk (18. How is this possible? The theoretical answer is that the time. in Eq. In URANS we do usually not care about scale v. If the ﬂow has strong vortex shedding. This is called scale separa¯ tion. On the other hand. v ′′ . v ′′ . and.7) So we are doing unsteady simulations. 18. whereas in LES it is. except that we denote the turbulent viscosity in the former case by νt and in the latter case by νsgs . The modelled turbulent ﬂuctuation. is much larger than the SGS viscosity. . but still we time average the equations.18. Unsteady RANS 149 see Fig. i. should have a much smaller time scale than the resolved ones. νsgs . is not shown in the ﬁgure. much more of the turbulence is modelled than in LES. In URANS. ¯ ¯ What type of turbulence model should be used in URANS? That depends on type of ﬂow. 18. In practice this requirement is often not satisﬁed [70].e.6) (18.7. the modelled turbulent ﬂuctuations. v ′ .4) µ+ ε ∂ε + c1ε P k − cε2 ρε ∂xj k (18. hence. know how they were time averaged? Or if they were volume ﬁltered? The answer is that they don’t.1 should be much smaller than the resolved time scale.1. i.1: Decomposition of velocities in URANS. the standard high-Re number k − ε model can be used. The URANS momentum equation and the LES momentum equation are exactly the same. Eq. how do the momentum equation. ∂ρk ∂ρ¯j k ∂ v = + ∂t ∂xj ∂xj ∂ v ∂ρε ∂ρ¯j ε = + ∂t ∂xj ∂xj µt σε µ+ µt σk ∂k + P k − ρε ∂xj (18.5) k2 ε With an eddy-viscosity. if this is added to v + v ′ we obtain v. v ¯ ¯ v ¯ v ¯ v′ ¯ t Figure 18.e.

18. 17.e. and this was discussed in connection to Fig. when should we start to average and for how long. they should not kill the resolved ﬂuctuations.2: Conﬁguration of the ﬂow past a triangular ﬂameholder.18. 18.4 and 18. separation. and the ﬂow actually has a scale separa¯ tion.e. i. . When we’re doing URANS. In Fig. This ﬂow has a very regular vortex shedding.3 the v2 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. 18. the question arises how the results should be time averaged. One cycle of the v2 velocity in a ¯ cell near the upper-right corner of the ﬂameholder. Flow from left to right Figure 18. i. v′ .2. see Fig. Unsteady RANS 150 Figure 18. What we care about is that the turbulence model and the discretization scheme should not be too dissipative. This issue is the same when doing LES.3: 2D URANS k − ε simulations [106]. ¯ The standard k − ε model (Eq.5) was used in [106] for URANS simulations computing the ﬂow around a triangular ﬂame-holder in a channel.11.

Thus. part of the turbulence is modelled (v ′′ ) and part of the turbulence is resolved (¯′ ). Turbulence Modelling 151 Figure 18.1H (x = 0 at the downstream vertical plane of the ﬂame-holder). v we must add these two parts together. Left ﬁgure: x = 0. both the discretization and the turbulence model have high dissipation.18.4: 2D URANS k − ε simulations compared with experiment [106]. The turbulence model that was used was the standard k − ε model.2 Discretization In LES it is well-known that non-dissipative discretization schemes should be used. Non-linear models like that of [109] was found to be less dissipative. because the unsteadiness is dampened out. and was successfully applied in URANS-simulations for these two ﬂows. It can be seen that here the resolved and the modelled turbulence are of the same magnitude. The reason is that we don’t want to dampen out resolved. right ﬁgure: x = 1.1 Turbulence Modelling In URANS.43H. In general a discretization scheme which has little numerical dissipation should be . It was found in [107. 18. 18.4. the reason for this is that the turbulence model is too dissipative. the hybrid discretization scheme for the convective terms was used together with fully implicit ﬁrst-order discretization in time. In the predictions on the ﬂame-holder presented above. turbulent ﬂuctuations. 18. The reason why the unsteadiness in these computations was not dampened out is that the vortex shedding in this ﬂow is very strong.1. The same is to some extent true also for URANS. Proﬁles downstream the ﬂameholder are shown in Fig. the ﬂow may not become unsteady at all. If the turbulence model in URANS generates ”too much” eddy viscosity. 108] when using URANS for the ﬂow around a surface-mounted cube and around a car. dashed lines: resolved turbulent kinetic energy: ∗: experimental data. this gives ﬁrst-order accuracy in both space and time. Solid lines: total turbulent kinetic energy. that the standard k − ε model was too dissipative. If we want to compare computed turbulence with experimental turbulence.

18.18.6H) using URANS and LES with central differencing are shown together with URANS and Mars scheme. whereas for industrially complex ﬂows maybe a bounded second-order scheme must be used.6 the time-averaged velocity proﬁle upstream of the cube (x1 = −0. For time discretization. Two different discretization schemes were used: the central scheme and the Mars scheme (a blend between central differencing and a bounded upwind scheme of second-order accuracy).e. the unphysical oscillations do usually not appear. How dissipative a scheme needs to be in order to be stable is ﬂow dependent. LES with the Mars scheme (in which some numerical dissipation is present) and URANS with the central scheme (where the modelling dissipation is larger than in LES) no such unphysical oscillations are present.5: URANS simulations of the ﬂow around a surface-mounted cube. It is seen that with LES and central differencing unphysical oscillations are present (this was also found by [79]). In this case the turbulent. resolved ﬂuctuations dominate over any numerical oscillations. 18. even if a central differencing scheme is used. used. Discretization 152 Figure 18. However. In Fig. In [107] LES and URANS simulations were carried out of the ﬂow around a surfacemounted cube (Fig. it may work with no dissipation at all (i. for some simple ﬂows. the second-order accurate Crank-Nicolson works in most cases. If turbulent unsteady inlet ﬂuctuations are used.5) with a coarse mesh using wall-functions.2. The main reason to the unphysical oscillations is that the predicted ﬂow in this region does not have any resolved ﬂuctuations. central differencing). .

Sect. η and ζ.65. ∆xζ ). The RANS model that was originally used was the one-equation model by [110]. In [111] the DES model was proposed in which d is taken as the minimum of the RANS turbulent length scale d and the cell length ∆ = max(∆xξ . It can be written [110. i. DES 153 Figure 18. Cdes ∆) (19. The constant Cdes is usually set to 0. The model was originally developed for wings at very high angles of attack. Velocity proﬁles upstream the cube [107].e. ∆xη .19.1) (19. Outside the turbulent boundary layer d > Cdes ∆ so that the model operates in LES .2) 1/2 Ψ = Cw1 ρfw d in the RANS SA model is equal to the distance to the nearest wall. In the boundary layer d < Cdes ∆ and thus the model operates in RANS mode. 4. νt ˜ νt ˜ d 2 ˜ µ + µt ∂ νt σνt ∂xj ˜ + ˜ ˜ Cb2 ρ ∂ νt ∂ νt +P −Ψ σνt ∂xj ∂xj ˜ (19.6] ∂ρ˜t ν ∂ ∂ρ¯j νt v ˜ = + ∂t ∂xj ∂xj νt = νt f1 ˜ The production term P and the destruction term Ψ have the form P = Cb1 ρ s + ¯ s = (2¯ij sij ) ¯ s ¯ νt ˜ f 2 d2 2 κ . The aim is to treat the boundary layer with RANS and capture the outer detached eddies with LES. 19 DES DES (Detached Eddy Simulation) is a mix of LES and URANS.6: URANS simulations of the ﬂow around a surface-mounted cube. 104. ˜ d = min(d. ∆xη and ∆xζ denote the cell length in the three grid directions ξ.3) ∆xξ .

say.1 DES based on two-equation models The model described above is a one-equation model. which in many situations is not a relevant turbulent length scale. If the switch occurs at location x1 . This poses a major problem with this type of models. εT is modiﬁed. When the model switches to LES mode. In regions where the turbulent length scales are taken from ∆ (LES mode) the ε-equation is still solved. This is exactly what happens at an inlet in an LES simulation if no real turbulence is given as inlet boundary conditions. However. A model based on the k − ε model can read ∂ ∂k (¯j k) = v + ∂t ∂xj ∂ε ∂ (¯j ε) = v + ∂t ∂xj ∂ ∂xj ∂ ∂xj νt σk νt ν+ σε ν+ ∂k + P k − εT ∂xj ε ∂ε + (C1 P k − C2 ε) ∂xj k P k = 2νt sij sij . ¯ ¯ νt = k 1/2 ℓt The turbulent length scale. A rather new approach is to reduce the destruction term in the ε equation as in PANS [117. 116] ℓt = min Cµ k 3/2 .2). A reduced νt ˜ ˜ gives a smaller production term P so that the turbulent viscosity is further reduced. which gives a reduction in the turbulent viscosity νt . In the RANS mode the major part of the turbulence is modelled. DES models based on two-equation models were proposed [113. and the turbulent dissipation. εT . A low-Reynolds number PANS was recently proposed [118] in which the near-wall modiﬁﬁcations were taken from the AKN model [120]. A third alternative is to modify only the turbulent length scale appearing in the turbulent viscosity [116]. but ε is not used. In RANS mode it takes its length scale from the wall distance. this would give rise to an increased production term P through the second term (see Eq. DES based on two-equation models 154 mode.g. Cε In other models [113. ε is needed as soon as the model switches to RANS model again. 114. In these models the turbulent length scale is either obtained from the two turbulent quantities (e. the length scale is taken as ∆. the turbulence is supposed to be represented by resolved turbulence. 19.1.19. However. it will take some distance L before the momentum equations start to resolve any turbulence. When the grid is sufﬁciently ﬁne. In these models ε increases because of its reduced destruction term which decreases both k and νt . Ck ∆ ε k 3/2 ∆ εT = max ε. 118] (Partially Averaged Navier-Stokes) and PITM [119] (Partially Integrated Transport Modelling). are computed as [115. The result is that the dissipation in the k equation increases so that k decreases which gives a reduced νt . This term is sometimes neglected [112] 19. The modelled length scale is reduced and the consequence is that the destruction term Ψ increases. 115]. Recently. this second term is a viscous term and is active only close to the wall. k 3/2 /ε or k 1/2 /ω) or the ﬁlter width ∆. At ﬁrst sight it may seem that as the model switches from RANS mode to LES mode thus reducing d. 53] only the dissipation term. ℓt . One way to get around this .

73. The forcing is added in the form of a source term (per unit volume) in the momentum equations.2. ∗ ωd d2 ω β νt = a1 k max(a1 ω. 122. 124.4) of the SST model. 123. 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 sufﬁciently ﬁne.4) F2 = tanh(η 2 ). ∆x2 . ∆x3 } .1 CDES ∆ k 1/2 β∗ω β ∗ kω → β ∗ kωFDES . 53] ∂k ∂ (¯j k) = v + ∂t ∂xj ∂ ∂ω (¯j ω) = v + ∂t ∂xj ∂ ∂xj ∂ ∂xj ν+ νt ∂k + Pk − β ∗ kω σk ∂xj ∂ω νt Pk − βω 2 ν+ +α σω ∂xj νt 1 ∂k ∂ω + 2(1 − F1 )σω2 ω ∂xi ∂xi √ k 500ν 4 F1 = tanh(ξ ).e the damping) is reduced and the ﬂow is induced to go unsteady. With a smaller turbulent viscosity in the momentum equations. β ∗ ωd d2 ω where d is the distance to the closest wall node. All coefﬁcients are blended between the k − ω and the k − ε model using the function F1 . DES based on the k − ω SST model 155 is to impose turbulence ﬂuctuations as forcing conditions [121. 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). |¯|F2 ) s η = max 4σω2 k CDω d2 (19. FDES = max Lt = ∆ = max {∆x1 . In some ﬂows it may occur that the FDES term switches to DES in the boundary layer because ∆z is too small (smaller than the boundary layer thickness. the dissipation term in the k equation increases which in turn decreases k and thereby also the turbulent viscosity. hopefully. 19. In DES the dissipation term in the k equation is modiﬁed as [53] Lt . 126] to protect the boundary layer from the LES mode FDES = max Lt (1 − FS ). 1 CDES ∆ where FS is taken as F1 or F2 (see Eq. the modelled dissipation (i. When FDES is larger than one.19. δ). Again. Different proposals have been made [125. The result is. ] at the location where the model switches from RANS mode to LES mode.2 DES based on the k − ω SST model The standard k − ω model SST reads [48. 2k 1/2 500ν . that a large part of the turbulence is resolved rather than being modelled. 19. ξ = min max . .

.3. 20.20. i. U . Hybrid LES-RANS 156 20 Hybrid LES-RANS When simulating bluff body ﬂows. the requirement of near-wall grid resolution is the main reason why LES is too expensive for engineering ﬂows.1). 20. 20. 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). The vorticity of the legs lift each other through self-induction which helps lifting the tip even more. often called the streak process. ′ v1 in the viscous wall region. The streamwise distance between the events is related to the boundary layer thickness (4δ. LES (Large Eddy Simulation) is the ideal method. see Fig. 20. On the other hand. An event of a high-speed in-rush is illustrated in Fig.1. The object of hybrid LES-RANS (and of DES) is to eliminate the requirement of high near-wall resolution in wall-parallel planes. Figure 20. 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 [100]. see Fig. As can be seen. respectively) are shown as the hairpin vortex is created.1). In the lower part of the ﬁgure the spanwise vortex line is shown. 20. The spanwise separation between sweeps and bursts is very small (approximately 100 viscous units. the turbulent structures very elongated in the streamwise direction.g. coarser grid spacing in wall-parallel planes can be used. Thus. The process by which the events are formed is similar to the later stage in the transition process from laminar to turbulent ﬂow. These structures must be resolved in an LES in order to achieve accurate results. Indeed. These events are called bursts or ejections. In the outer region (the LES region). As ¯ a result the mid-part is lifted up even more and a tip of a hairpin vortex is formed. The in-rush event is also called a sweep. a turbulent ﬂuctuation – the mid-part of the vortex line is somewhat lifted up away from the wall. In the LES region.2 presents the instantaneous ﬁeld of the streamwise velocity ﬂuctuation. There are also events which occurs in the other direction. It can be seen that an inﬂexion point is created in the instantaneous velocity proﬁle. 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. lowspeed ﬂuid is ejected away from the wall. it is a challenging task to make accurate predictions of wallbounded ﬂows with LES. the usual LES is used. but due to a disturbance – e. which was one of the lessons learned in the LESFOIL project [104. The grid resolution in this region is . a low-Re number RANS turbulence model (usually an eddy-viscosity model) is used. Initially it is a straight line. The mid-part of the vortex line experiences now a higher v1 velocity (denoted by U in the ﬁgure) than the remaining part of the vortex line. 105]. for wall-bounded ﬂows at high Reynolds numbers of engineering interest. and the momentum deﬁcit in the inner layer increases for increasing x1 . This is the main reason why the grid must be very ﬁne in the spanwise direction. Eventually the momentum deﬁcit becomes too large and the highspeed ﬂuid rushes in compensating for the momentum deﬁcit. the computational resource requirement of accurate LES is prohibitively large. see Fig. The idea is that the effect of the near-wall turbulent structures should be modelled by the RANS turbulence model rather than being resolved. In the x1 − x2 plane (upper part of Fig. 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. At low to medium Reynolds numbers the streak process is responsible for the major part of the turbulence production.e.1) the instantaneous and ¯ mean velocity proﬁles (denoted by U and U in the ﬁgure. In the near-wall region (the URANS region).

ml . DNS 2 of channel ﬂow [73].3: The LES and URANS region.20. x3 x1 Figure 20.2: Fluctuating streamwise velocity in a wall-parallel plane at x+ = 5. wall URANS region LES region x2 URANS region wall x1 Figure 20.1: Illustration of near-wall turbulence (taken from [57]). Hybrid LES-RANS 157 Figure 20. x+ 2.

δ/∆x3 ≃ 5. e.) [73]. 128] twoequation models were used in the URANS region and a one-equation SGS model was employed in the LES region. Markers in right ﬁgure indicate resolution.f . 70. In [127] they used a two-equation model in the URANS region and blended it into a one-equation model in the LES region.2 10 x+ 2 10 10 0 0 0. 130]. A one-equation model was used in both regions in [131].4 0.4 ln(y + ) + 5. presumably dictated by the requirement of resolving the largest turbulent scales in the ﬂow (which are related to the outer length scales. ◦: 0. 1 3 LES-RANS. The unsteady momentum equations are solved throughout the computational domain.g. v2. and the x+ 2.f LES region URANS region x2 wall x1 Figure 20.20.2. In some work [70.6 v1 15 ¯+ 10 5 0 0. In [129] it was determined by comparing the URANS and the LES turbulent length scales or was computed from turbulence/physics requirements. v3. The turbulent RANS viscosity is used in the URANS region.5 2 (a) Velocity proﬁles (b) Two-point correlation Figure 20. interface ′ ′ ′ v1.5: Using forcing at the interface between the LES and URANS region.ml . In [127.rms x 1 2 3 25 0.4: Comparison of standard hybrid LES-RANS in channel ﬂow on a very : hybrid coarse mesh (∆x+ = 2∆x+ = 785.5. The locations of the matching planes were determined in different ways. 128] it was chosen along a pre-selected grid plane. and the turbulent SGS viscosity is used in the LES region.f . Different partial differential equations for automatically ﬁnding the matching plane were investigated in [130].8 0. One-equation models were used in both regions in [129. Hybrid LES-RANS 1 30 158 20 2 B11 (ˆ1 )/v1. δ/∆x1 ≃ 2.5 x1 1 1. the boundary layer thickness) rather than the near-wall turbulent processes. Much work on hybrid LES-RANS has been carried out.

the streamwise lengthscale predicted with hybrid LESRANS is extremely large. 20. This results in too poorly resolved stresses in the interface region and thereby gives a ramp – also referred to as a shift – in the velocity proﬁle approximately at the location of the matching plane [70. Hybrid LES-RANS is similar to DES (Detached Eddy Simulations) [111. industrial ﬂows. 136. The overly small resolved stresses in the LES region are translated into too small a wall shear stress.1 . 122]. 126]. Although the results obtained with hybrid LES-RANS are better than those obtained with LES. Hybrid LES-RANS gives much improved results. 135.6: Added ﬂuctuations. The ﬂuctuations are either synthesized (subscript f = S) or taken from channel DNS (subscript f = DN S). The main difference is that the original DES aims at covering the whole attached boundary layer with URANS. The momentum equations are solved in the entire domain and the turbulent viscosity is in both regions obtained from a one-equations ksgs equation and an algebraic length scale (see Sections 20. On a ﬁne mesh the model switches smoothly to LES and in the limit cε1 = cε2 so that a pure DNS solution is obtained.f Control volume j LES region Interface j − 1/2 URANS region An ′ ′ ′ Figure 20. in this form. Hybrid LES-RANS 159 j + 1/2 ′ v2. in which the cε2 is made into a function of the ratio of the RANS and LES length scales. 128. DES is similar hybrid LES-RANS. The resolution in the wall-parallel plane is comparable to what can be afforded for boundary layer in real. In later work DES has been used as a wall model [135. The resolved turbulence supplied by the URANS region to the LES region has no reasonable turbulent characteristics and is not appropriate for triggering the LES equations to resolve turbulence. The normalized streamwise two-point correlation is shown in Fig.20. 131]. at least in terms of viscous units (∆x+ and ∆x+ ). In [132] they proposed a k − ε turbulence model. 129. vf. still not very good however.4b.4a presents comparison of LES and hybrid LES-RANS in channel ﬂow at Reτ = 2000 on a very coarse mesh. As can be seen. and.f ′ v1. 122]. Several modiﬁcations have been proposed to remove this deﬁciency. 134. In [136. vf. cµ coefﬁcient at the interface was computed dynamically to yield a smoother transition between the URANS and LES regions. It should be mentioned that standard hybrid LES-RANS does – of course – give better results on ﬁner grids [101]. Figure 20. it has been found that the treatment of the interface between the URANS region and the LES region is crucial for the success of the method. in a control volume (j = jml + 1) in the LES region adjacent to the interface. The LES 1 3 cannot resolve the ﬂow at all. 123.1 and 20. they suggested dampening the modelled stresses in the URANS region to .3 .2). later also used by [133]. whereas hybrid LES-RANS aims at covering only the inner part of the boundary layer with URANS. vf. but these ﬁner grids are rarely affordable in industrial ﬂows.2 .

3) and. which reads ∂kT ∂kT ∂ ∂ (ν + νT ) (¯j kT ) = v + ∂t ∂xj ∂xj ∂xj ¯ PkT = −τij sij . 20.ml (see Fig. 20. In attempts to improve the performance of LES in wall-bounded ﬂows. 73]. νT = νsgs .5 and 20.2) τij = −2νT sij ¯ In the inner region (x2 ≤ x2.e. This ﬂow may seem to be an easy test case.20. Numerical smoothing was used at the interface in [129]. kT = 0. but it is not.2 The equation for turbulent kinetic energy in hybrid LES-RANS A one-equation model is employed in both the URANS region and the LES region. k. see Figs.2. . in the outer region (x2 > x2. the resolved turbulence in the LES region near the URANS region will be too small. One way to improve hybrid LES-RANS is to add ﬂuctuations to the momentum equations at the interface [123.ml . see Eq. νT .1. 20. The object is to trigger the equations to resolve turbulence. for x2 > x2. At the walls. In [122] backscatter was introduced in the interface region with the object of generating resolved ﬂuctuations.1 Momentum equations in hybrid LES-RANS The incompressible Navier-Stokes equations with an added turbulent/SGS viscosity read ¯ ∂¯i v 1 ∂p ∂ ∂ ∂¯i v (ν + νT ) (20. the latter is obtained by solving its transport equation. 20.1) and kT . resolved plus modelled) shear stress in the URANS region and thereby reduce the jump in shear stress across the matching plane. This is also the case in hybrid LESRANS: if no triggering (forcing) is applied at the interface between the LES region and the URANS region.1. 20.1) (¯i vj ) = − v¯ + + ∂t ∂xj ρ ∂xi ∂xj ∂xj where νT = νt (νt denotes the turbulent RANS viscosity) for x2 ≤ x2. the resolved turbulence near the inlet will be too small and a large streamwise distance is required before the equations trigger themselves into describing turbulent ﬂow. is computed from an algebraic turbulent length scale (see Table 20. see Table 20. + PkT − Cε kT ℓ 3/2 (20.ml ) kT corresponds to the RANS turbulent kinetic energy. No special treatment is used in the equations at the matching plane except that the form of the turbulent viscosity and the turbulent length scale are different in the two regions. [128] proposed a modiﬁcation of the discretized streamwise equation at the interface in order to avoid ﬁltering out any resolved ﬂuctuations at the interface. the Achilles’ heel is the near-wall ﬂow region.3 Results Fully developed channel ﬂow at Reτ = uτ δ/ν = 2000 (δ denotes the channel half width) is used as a test case to evaluate the effect of different forcing conditions. Adding ﬂuctuations in order to trigger the equations to resolve turbulence is actually very similar to prescribing ﬂuctuating turbulent inlet boundary conditions for DNS or LES (or hybrid LES-RANS). 20. Momentum equations in hybrid LES-RANS 160 reduce the total (i.6. The turbulent viscosity.ml ) it corresponds to the subgrid-scale kinetic turbulent energy (ksgs ). If no triggering inlet boundary conditions are prescribed in DNS or LES.

20.3. Results

URANS region ℓ νT Cε

−3/4 κcµ n[1 − exp(−0.2k 1/2 n/ν)] 1/4 κcµ k 1/2 n[1 − exp(−0.014k 1/2n/ν)]

161 LES region ℓ=∆ 0.07k 1/2 ℓ 1.05

1.0

Table 20.1: Turbulent viscosity and turbulent length scales in the URANS and LES regions. n and κ denote the distance to the nearest wall and von K´ rm´ n constant a a (= 0.41), respectively. ∆ = (δV )1/3 .

30 25 30 25 20 15 10 5

0 2

v /u∗,w ¯

20 15 10 5 0

0

10

x+ 2

10

10

0

x+ 2

10

2

(a) Solid lines: isotropic forcing, MS = 0.25; dashdotted lines: isotropic forcing, MS = 1; dashed lines: no forcing; +: present 963 DNS.

(b) DNS forcing. Solid line: MDNS = 0.25; dashed line: MDNS = 0.5.

**¯ Figure 20.7: Streamwise velocities [73]. v proﬁles. ◦: 2.5 ln(x+ ) + 5.2. 2 τν + τ12 /u2 ∗,w kT /u2 ∗,w
**

′ ′ 0.5vi vi /u2 , ∗,w

1

7 6 5 4 3 2 1 0 0 0.2 0.4 0.6 0.8 1

0.8

0.6

′ ′ −v1 v2 /u2 , ∗,w

0.4

0.2

0 0

0.2

0.4

x2

0.6

0.8

1

x2

(a) Shear stresses. τν denotes viscous stress.

(b) Turbulent kinetic energy.

Figure 20.8: Shear stress and turbulent kinetic energy [73]. Solid lines: no forcing; dashed lines: forcing with isotropic ﬂuctuations with MS = 0.25; ◦: present 963 DNS. Thick lines: resolved; thin lines: modelled. The bulk velocity in fully developed channel ﬂow with periodic boundary conditions (see Eq. 20.1) is entirely determined by the wall shear stress; consequently the ﬂow is extremely sensitive to the turbulence in the near-wall region.

20.3. Results

162

The streamwise velocity proﬁles obtained with and without forcing are compared in Fig. 20.7 with the present DNS and the log-law. It can be seen that the centerline velocity is strongly over-predicted when no forcing is used, whereas forcing with MS = MDN S = 0.25 gives excellent agreement with the log-law (MS and MDN S denote forcing with synthetic and DNS ﬂuctuations, respectively). The reason for the overly large velocities without forcing is that the resolved shear is too small. It can be seen in Fig. 20.8a that it is the resolved shear stress that increases when forcing is introduced, indicating that the resolved shear stress without forcing is too small. This was also observed by [122]: when forcing is introduced, the resolved shear stress increases, which reduces the bulk and centerline velocity. Recently a novel way for generating ﬂuctuations to be used as forcing at the interface was presented [101]. In this work backscatter obtained from a scale-similarity model was used.

21. The SAS model

163

**21 The SAS model
**

21.1 Resolved motions in unsteady

When doing URANS or DES, the momentum equations are triggered through instabilities 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 modelled. 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

**21.2 The von K´ rm´ n length scale a a
**

The von K´ rm´ n length scale a a LvK,1D = κ ∂ v /∂x2 ¯ 2 v /∂x2 ∂ ¯ 2 (21.1)

which includes the second velocity gradient is a suitable length scale for detecting unsteadiness. The von K´ rm´ n length scale is smaller for an instantaneous velocity a a proﬁle than for a time averaged velocity, see Fig. 21.1. This is interesting because, as noted in [137], the von K´ rm´ n length scale decreases when the momentum equations a a resolve (part of) the turbulence spectrum. The ﬁrst and second derivatives in Eq. 21.1 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 |¯| = (2¯ij sij )1/2 . The second derivative can be generalized in a number of ways. In s s ¯ the SAS model it is taken as U ′′ = ∂ 2 vi ¯ ∂ 2 vi ¯ ∂xj ∂xj ∂xk ∂xk

0.5

(21.2)

Hence, the general three-dimensional expression for the von K´ rm´ n length scale reads a a LvK,3D = κ |¯| s |U ′′ | (21.3)

In [138] they derived a one-equation νt turbulence model where the von K´ rm´ n a a length scale was used. The model was called the SAS model. Later, based on the k − k 1/2 L model of Rotta [139], Menter & Egorov [137] derived a new k − kL model using the von K´ rm´ n length scale. Finally, in [140] they modiﬁed the k − ω-SST a a model to include the SAS features; they called this model the SST-SAS model. This model is described in more detail below.

**The SST-SAS model
**

The k − ω SST model is given in Eq. 19.4 at p. 155 (see also the section starting at p. 123) Now, Menter & Egorov [140] introduced a SAS-term in the ω equation. The

21.2. The von K´ rm´ n length scale a a

164

1

0.8

x2

0.6

LvK,1D

0.4 0.2

LvK,3D

0 0

5

10

v1 ¯

15

20

Figure 21.1: Velocity proﬁles from a DNS of channel ﬂow. Solid line: time-averaged velocity with length scale Lx,1D , Eq. 21.1 ; dashed line: instantaneous velocity with length scale LvK,3D , Eq. 21.3. object of this term is to decrease the turbulent viscosity when unsteadiness is detected, i.e. when the von K´ rm´ n length scale becomes small. The production term in the ω a a equation in the k−ω-SST model reads Pω = αP k /νt ∝ |¯|2 . To decrease the turbulent s 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 ˜ s (21.4) ζ2 κ|¯|2 , Lt = 1/4 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. 21.1 • This results in a decrease in the von K´ rm´ n length scale, LvK,3D a a • As a consequence the additional source, Eq. 21.4, 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 kinetic 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.

21.3. The second derivative of the velocity

165

When incorporating the additional production term (Eq. 21.4) in the k − ω-SST model, the last term in the ω equation is replaced by (for further details, see [140]) PSAS = FSAS max (T1 − T2 , 0) L ˜ T1 = ζ2 κS 2 LvK,3D 2k 1 ∂ω ∂ω 1 ∂k ∂k T2 = max , σΦ ω 2 ∂xj ∂xj k 2 ∂xj ∂xj L= k 1/2 ωcµ

1/4

(21.5)

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.

**21.3 The second derivative of the velocity
**

To compute U ′′ in Eq. 21.2, 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 ∂x2 and then ⇒ =

j+1/2

vj+1 − vj , ∆x2 =

j

∂v ∂x2

=

j−1/2

vj − vj−1 ∆x2

∂2v ∂x2 2

(∆x2 )2 ∂ 4 v vj+1 − 2vj + vj−1 + (∆x2 )2 12 ∂x4 2

**Option II: compute the ﬁrst derivatives at the center ∂v ∂x2 and then ⇒ =
**

j+1

vj+2 − vj , 2∆x2 =

j

∂v ∂x2

=

j−1

vj − vj−2 2∆x2

∂2v ∂x2 2

(∆x2 )2 ∂ 4 v vj+2 − 2vj + vj−2 + 2 4(∆x2 ) 3 ∂x4 2

In [141], Option I was used unless otherwise stated.

**21.4 Evaluation of the von K´ rm´ n length scale in channel ﬂow a a
**

In Fig. 21.2 the turbulent length scale, LvK,3D , is evaluated using DNS data of fully developed channel ﬂow. When using DNS data only viscous dissipation of resolved turbulence affects the equations. This implies that the smallest scales that can be resolved are related to the grid scale. The von K´ rm´ n length scale based on instantaneous vea a locities, LvK,3D , is presented in Fig. 21.2. For x2 > 0.2, its magnitude is close to ∆x2 which conﬁrms that the von K´ rm´ n length scale is related to the smallest resolva a able scales. Closer to the wall, LvK,3D increases slightly whereas ∆x2 continues to decrease. The von K´ rm´ n length scale, LvK,1D , based on the averaged velocity proﬁle a a v1 = v1 (x2 ) is also included in Fig. 21.2, and as can be seen it is much larger than ¯ ¯ LvK,3D . Near the wall LvK,1D increases because the time-average second derivative,

25 0.03 0. Reτ = 500.05 0.02 0.03 0.8 1 0 0 0. ∆x1 /δ = : LvK. ∂ 2 v1 /∂x2 .02 0. goes to zero as the wall is approached.4 0. Hybrid LESRANS.05 0 0 0.1 0.1D .1D . DNS.6 0.3D .3D . +: ℓk−ω = k 0. x2 -stretching of 17%.3D follows closely LvK.3: Turbulent length scales in fully developed channel ﬂow. : LvK. Reτ = 2000.2 0. Left: global view. : LvK.065.2 0.39.1D . 32 × 64 × 32 mesh. x2 -stretching of 9%.3D . Left: global view.02 0. −2 νT sij sij ¯ ¯ (νT denotes SGS or RANS turbulent viscosity). 21. LvK. part of the turbulence is resolved and part of the turbulence is modelled. 963 mesh. : LvK.2 0. In Fig. the length scale of the smallest resolved turbulence is larger in hybrid LES-RANS than in DNS. Evaluation of the von K´ rm´ n length scale in channel ﬂow a a x+ 2 x+ 2 166 0 0.4. and νT ≫ ν.016. 0 500 x+ 2 1500 2000 0 0. the resolved turbulence is dampened by the high turbulent viscosity. : 0. No such behavior is seen for the ¯ 2 three-dimensional formulation.05 x2 /δ x2 /δ Figure 21.05 0 0 0.04 0.2 0.6 0. 1/4 (∆x1 ∆x2 ∆x3 )1/3 .04 x2 /δ x2 /δ Figure 21.25 0. from a 1D RANS simulation at Reτ = .031δ). When using hybrid LES-RANS. right: zoom.01 0. : (∆x1 ∆x2 ∆x3 )1/3 .19.2: Turbulent length scales in fully developed channel ﬂow.04 0.1 5 10 15 20 25 0.15 0.15 0.8 1 0 0 0.06 0. The resolved turbulence is dissipated by a modelled dissipation.3.21.04 0.01 0. ∆x3 /δ = 0. right: zoom.5 /(cµ ω). ◦: ∆x2 . and as a results LvK.08 0. ∆x1 /δ = 0. Close to the wall in the URANS region (x2 < 0.06 0.08 0. ◦: ∆x2 .1 0. ∆x3 /δ = 0.3 100 200 300 400 500 0 0. The RANS turbulent length scale.1 20 40 x+ 2 60 80 100 0. data from hybrid LES-RANS are used (taken from [73]). As a result.02 0.4 0. ℓk−ω .

In the a a center region the RANS turbulent length scale continues to increase which is physically correct. In [141]. (∆x1 ∆x3 ∆x3 )1/3 > 2 LvK. . In the DNS-simulations.5δ). its behavior is close to that of the von K´ rm´ n length scale. the SST-SAS model has been evaluated in channel ﬂow.3D . However.3.3. Evaluation of the von K´ rm´ n length scale in channel ﬂow a a 167 2000 with the k − ω SST model is also included in Fig. In the hybrid simulations. a a i. In the inner region (x2 < 0. LvK.2 and 21.1D . ∆x2 < (∆x1 ∆x2 ∆x3 )1/3 near the wall.21. 21. goes to zero because the a a velocity derivative goes to zero.1D . the von K´ rm´ n length scale. LvK. Two ﬁlter scales are included in Figs. it can be noted that the three-dimensional ﬁlter width is more that twice as large as the three-dimensional formulation of the von K´ rm´ n length scale. ﬂow in an asymmetric diffuser and ﬂow over an axi-symmetric hill. 21.e.4. 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 .

namely f incompressible ﬂows. 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 spatial ﬁltering in LES. ku and its dissipation rate εu . F . .2) where τij is the central second moment resulting from the partial averaging for the nonlinear terms. Parameter fk deﬁnes the ratio of unresolved (partially-averaged) turbulent kinetic energy (ku ) to the total kinetic energy (k). a model is needed for τij .4) where Vi denotes the RANS velocity. that is τij = (P(vi vj ) − vi vj ). we also use the terminology of Reynolds stresses for the term τij in Eq. In [143.93) in the k − ε model by fk . 22. which corresponds to a ﬁltering operation for a portion of the ﬂuctuating scales [142]. Applying the partial averaging to the governing equations gives ∂¯i v =0 ∂xi ∂¯i v ¯ 1 ∂p ∂ ∂(¯i vj ) v¯ =− + + ∂t ∂xj ρ ∂xi ∂xj ν ∂¯i v − τij ∂xj (22. where P denotes the partial-averaging operator. they deﬁned in [142] another two quantities. as in RANS and LES. In the derivation of the transport equations for ku and εu . and fε is the ratio between the unresolved (εu ) and the total (ε) dissipation rates. fk and fε . 142] they employed the standard k − ε model as the base model. 11.5 is usually neglected.2. The left side can be re-written fk ∂k ¯ ∂k + Vj ∂t ∂xj = ∂ku ¯ ¯ ∂ku = ∂ku + vj ∂ku + (Vj − vj ) ∂ku ¯ + Vj ¯ ∂t ∂xj ∂t ∂xj ∂xj (22.e. (for simplicity we omit the buoyancy term) fk ∂k ¯ ∂k + Vj ∂t ∂xj = fk P k − ε + ∂ ∂xj ν+ νt σk ∂k ∂xj (22. where sij is the strain-rate tensor of the ¯ ¯ computed ﬂow and νu is the PANS eddy viscosity. For simplicity. ¯ For an instantaneous ﬂow variable. To close the system of the partially-averaged Navier-Stokes equations. the partially-averaged turbulent kinetic energy. The ku equation is derived by multiplying the RANS k equation (Eq. have been introduced.1) (22. 2 so that νu = Cµ ku /εu .22. In [142] they proposed using the conventional eddy viscosity concept so that τij = −2νu sij .3) The extent of the resolved part is now determined by fk and fε . two parameters. relating the unresolved to the resolved ﬂuctuating scales.5) The convective term must be expressed in vj (the PANS averaged velocity) rather than ¯ ¯ in Vj (the RANS averaged velocity). The last term on the right side in Eq. 22. because it is vj that transports ku because vj ¯ ¯ represents the PANS resolved part of vj . where vi indicates instantaneous ve¯¯ locity components. These give k= εu ku and ε = fk fε (22. i. we use f to denote the partially-averaged ¯ = P(F ). We consider part. In order to formulate the PANS eddy viscosity. The PANS Model 168 22 The PANS Model The PANS method uses the so-called “partial averaging” concept.

Pu = νu where ∂¯j v ∂¯i v + ∂xj ∂xi ∂¯i v ∂xj (22.13) The diffusion term is re-written using Eq.14) where σεku = σε k (22.8 the ﬁnal transport equation for ku can now be written as ∂ku ∂ ∂(ku vj ) ¯ = + ∂t ∂xj ∂xj ν+ νu σku ∂ku + Pu − εu ∂xj (22.15) . i.e. 22.9) Using Eqs.5.e.3 fε ∂ ∂xj ν+ νt σε ∂ε ∂xj ∂ ∂xj ∂ = ∂xj = 2 fk fε νt σε νu ν+ σεu ν+ ∂εu ∂xj ∂εu ∂xj (22. 22. is expressed in terms of the PANS eddy viscosity.4 must be equal to the sum of the source terms of the ku equation.8) fk P k − ε = Pu − εu This relation implies Pk = 1 εu (Pu − εu ) + fk fε (22.3 fk ∂ ∂xj ν+ νt σk ∂k ∂xj ∂ ∂xj ∂ = ∂xj = 2 fk fε νt σk νu ν+ σku ν+ ∂ku ∂xj ∂ku ∂xj (22. i.22. 22. Pu . 22. i. and the strain rate of PANS-resolved ﬂow ﬁeld.7) The sum of the source terms in Eq. 22.11) 2 ku ε The εu equation is derived by multiplying the RANS ε equation by fε .10) where the production term.6 and 22.e.12) ¯ ∂εu ∂ε ∂(εVj ) ∂(εu vj ) ¯ = fε + + ∂t ∂xj ∂t ∂xj νt ∂ ν+ = fε ∂xj σε ε ε2 ∂ε + Cε1 Pk − Cε2 ∂xj k k (22. The PANS Model 169 The diffusion term is re-written using Eq. νu .6) where σku = σk (22. νu = cµ (22. (22.

10 and 22. 22.18) ¯ v As in the ku equation.19(a) the total (i. 22. but not the diffusion term. the turbulent diffusion term was obtained as fk ∂ ∂xj νt ∂k σk ∂xj ∂ νt ∂ku ∂xj σk ∂xj νu ∂ku ∂ = ∂xj σku ∂xj = (22.3 and 22. 22.20) ∂xj σk ∂xj In Eq. ε or ω equation is modiﬁed. The PANS equation for ku . ku . 22. 22. σku . for example. the 2 modelled turbulent diffusion in the PANS formulation is a factor of σk /σku = fε /fk larger than in Eq.7. Equation 22. such as DES [144].17) The εu equation in the PANS model now takes the following form ∂εu ∂ ∂(εu vj ) ¯ = + ∂t ∂xj ∂xj ν+ νu σεu 2 εu ∂εu ∗ ε + Cε1 Pu − Cε2 u ∂xj ku ku (22. the sink term in the k.9) fε Cε1 Pk ε ε2 − Cε2 k k εu ε2 = Cε1 Pu − Cε1 u ku ku εu fk 1 εu (Pu − εu ) + ku fk fε 2 2 ε fk ε fk + Cε1 u − Cε2 u ku fε fε ku 2 εu ∗ ε = Cε1 Pu − Cε2 u ku ku = Cε1 fk (Cε2 − Cε1 ) fε − Cε2 ε2 fk u fε ku (22.e.6. the RANS) turbulent viscosity has been used for ku . for improved modelling of near-wall turbulence. is a unique property of the PANS model. the the additional term (Vj −¯j )∂εu /∂xj has been neglected.12 and 22. 22. In other models.20.19a) (22.10.22. X-LES [115] and PITM [119]. Eq. this factor is larger than six. A recently developed LRN PANS model is employed. when used as an SGS model.19(a) suggests that the turbulent transport for the PANS-modelled turbulent kinetic energy. The PANS Model 170 In the same way.4. A Low Reynolds number PANS model was presented in [145]. .19b) The expression on the right-hand side of Eq. which reads νsgs ∂ksgs ∂ (22. is actually formulated in terms of the RANS turbulent viscosity from the base model. whereas on the right-hand side of Eq. This is different from the turbulent diffusion in subgrid scale (SGS) modelling of LES with a one-equation ksgs model. see Eqs. the production and destruction terms are re-formulated as (using Eqs. was derived by multiplying the RANS equation for k by fk which was assumed to be constant in space and in time.19(a) suggests that.19(b). 22.16) where ∗ Cε2 = Cε1 + (22. By referring to Eqs. 22. With fε = 1 and fk = 0. The modiﬁcation of the diffusion coefﬁcient. 22.20 the SGS turbulent viscosity is invoked for the transport of ksgs .

9. The PANS Model 171 which reads [118] ∂ ∂(ku vj ) ¯ ∂ku = + ∂t ∂xj ∂xj ∂εu ∂ ∂(εu vj ) ¯ = + ∂t ∂xj ∂xj νu = Cµ fµ σku ≡ σk νu σku νu ν+ σεu ν+ ∂ku + (Pu − εu ) ∂xj 2 εu ∂εu ∗ ε + Cε1 Pu − Cε2 u ∂xj ku ku (22.21) 2 ku ∗ fk (Cε2 f2 − Cε1 ) .22.5. σεu ≡ σε k fε fε The modiﬁcation introduced by the PANS modelling as compared to its parent RANS model is highlighted by boxes.09 (22.4.e. σε = 1. Cµ = 0. Cε2 = 1. The model constants take the same values as in the LRN model [120].22) . Cε1 = 1.4. i. σk = 1. Cε2 = Cε1 + εu fε 2 fk f2 .

3) Note that the time dependence term (the ﬁrst term on the left side) has been retained. 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. i. 23. Vi = vi + vi . p is the pressure and ν and ρ are the viscosity and density of the ﬂuid.1) where vi denotes the velocity vector. 23. time dependent turbulent (i.2) The term in front of the second equal sign is called the Reynolds stress and it is unknown and must be modelled.1. because the large. 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.e.e. CFD (Computationally Fluid Dynamics) based on ﬁnite volume methods is used extensively to solve the RANS equations. 23. often called eddies. i. The term in front of ¯ ¯ (ν + νsgs ) ∂¯i v ¯ ¯ 1 ∂p ∂ 2 vi ¯ ∂τij 1 ∂p ∂ ∂¯i vj v¯ =− +ν − =− + + ∂t ∂xj ρ ∂xi ∂xj ∂xj ∂xj ρ ∂xi ∂xj . In turbulent ﬂow. The smallest scales are related to the eddies where dissipation takes place. the velocity and pressure are unsteady and vi and p include all turbulent motions. where the kinetic energy of the eddies is transformed into internal energy causing increased temperature.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 modelled with a turbulence model. the velocity vector and the pressure are split into a time-averaged part (Vi ′ ′ and P ) and a ﬂuctuating part (vi and p′ ). The ratio of νt to ν may be of the order of 1000 or larger. The LES equations read ∂¯i v ∂xj (23. in case of a boundary layer). 23.2.2 the unknown Reynolds stresses are expressed by a turbulence model in which a new unknown variable is introduced which is called the turbulent viscosity. Eq. The ratio of the largest to the smallest eddies increases with Reynolds number.2 are highly dependent on the accuracy of the turbulence model. δ. On the right side of Eq.1 Introduction Fluid ﬂow problems are governed by the Navier-Stokes equations ∂vi 1 ∂p ∂ 2 vi ∂vi vj =− +ν + ∂t ∂xj ρ ∂xi ∂xj ∂xj (23. In industry today. the resolved) ﬂuctuations are part of vi and p and are not modelled with the turbulence model.1. All turbulent ﬂuctuation are modelled with a turbulence model and the results when solving Eq. The resulting equation is called the RANS (Reynolds-Averaging Navier-Stokes) equations ′ ′ ∂vi vj 1 ∂P ∂ 2 Vi 1 ∂P ∂ ∂Vi Vj =− +ν − =− + ∂xj ρ ∂xi ∂xj ∂xj ∂xj ρ ∂xi ∂xj (ν + νt ) ∂Vi ∂xj (23. respectively. Hybrid LES/RANS for Dummies 172 23 Hybrid LES/RANS for Dummies 23. 23. p = P + p′ .1 Reynolds-Averaging Navier-Stokes equations: RANS In order to be able to solve the Navier-Stokes equations with a reasonable computational cost.e. νt .23.

νT = νt and away from the walls an LES turbulence model is employed. Hence. The ratio of νsgs to ν is of the order of 1 to 100.4 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.e. LES is much more accurate than RANS because only a small part of the turbulence is modelled with the turbulence SGS model whereas in RANS all turbulence is modelled. The turbulence model reads [117. 118]. The ﬂow around the vehicle can be computed with RANS. 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 dimensions (time and three spatial directions) whereas RANS can be solved in steady state (no time dependence).2 and 23.3 can be written in a same form as ∂¯i v ¯ 1 ∂p ∂ ∂¯i vj v¯ =− + + ∂t ∂xj ρ ∂xi ∂xj (ν + νT ) ∂¯i v ∂xj (23. However. In Section 23.1.2 The PANS k − ε turbulence model . we have describe how to use the zonal LES/RANS method for ﬂows near walls.3) is much v ′ ¯′ smaller than one. In the present work. their spatial scales get smaller for increasing Reynolds number.1. i. i. which are called SGS (sub-grid stresses).2. 23. 23. The PANS k − ε turbulence model 173 the second equal sign includes the Reynolds stresses of the small eddies. Much research has the last ten years been carried out to circumvent this problem.4. This term must also – as in Eq. Note that the time dependence term is now retained also in the RANS region: near the wall we are using an unsteady RANS. a RANS turbulence model is used for the turbulent viscosity. The focus of this report is zonal LES/RANS. 23. Above. When the ﬂow near walls is of importance. The difference of νsgs compared to νt in Eq. 23.e.e. and at the right side it has been modelled with a SGS turbulent viscosity. URANS. is turns out that LES is prohibitively expensive because very ﬁne cells must be used there. These methods are called Detached Eddy Simulation (DES). Furthermore. νT = νsgs . The reason is entirely due to physics: near the walls. the ratio of the resolved to the modelled turbulence. 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.23. Another form of zonal LES/RANS is embedded LES. |¯i vj |/|τij | (see Eqs. the PANS k − ε model is used to simulate wall-bounded ﬂow at high Reynolds number as well as embedded LES. 23. in which an LES mode is embedded in a RANS region.4) Near the walls. νsgs . see Eq. One example is prediction of aeroacoustic noise created by the turbulence around an external mirror on a vehicle [89]. 22.5) 23. the spatial scales of the “large” turbulent eddies which should be resolved by LES are in reality rather small.3 Zonal LES/RANS Equations 23.2 is that it includes only the effect of the small eddies. i.21 (here in a slightly simpliﬁed form to enhance readability) ∂k ∂k¯j ∂ v = + ∂t ∂xj ∂xj ν+ νT σk ∂k + Pk − ε ∂xj (23. see Fig. hybrid LES/RANS or zonal LES/RANS.2 and 23. see Fig. 23. All proposed methods combines RANS and LES where RANS is used near walls and LES is used some distance away from the walls.2 – be modelled.

When fk is ∗ decreased (to 0.7 decreases. 23. New ﬂuxes are introduced using smaller SGS values [146].8 decreases because k decreases and ε increase. • k decreases because ε (last term in Eq. Cµ = 0.5) is the main sink term in the k equation increases.09 (23. In the LES region fk = 0. fk = 1. Zonal LES/RANS: wall modeling 174 URANS.5–23. Cε2 = 1. 23.7 is equal to one. Hence νT must decrease rapidly when going from the URANS region to the LES region. Fully developed channel ﬂow. 23.4 and in the URANS region fk = 1. the turbulent viscosity νT should be an SGS viscosity and in the latter region it should be an RANS viscosity. fk = 1.0 interface LES. the model acts as a standard k − ε RANS model giving a large turbulent viscosity. Periodic boundary conditions are applied at the left and right boundaries. 23. Cε2 in Eq. Hence. .1: The LES and URANS regions. 23. 23. Cε1 = 1. In the former region.23.8 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. and • νT in Eq. ∂ε ∂ε¯j ∂ v = + ∂t ∂xj ∂xj ν+ νT σε 2 ε ∂ε ∗ ε + Cε1 Pk − Cε2 ∂xj k k (23.1) separates the URANS regions near the walls and the LES region in the core region.2 Figure 23.0 x1 3.9 k2 .6) (23. fk = 0. The key elements in the present use of the PANS k − ε model are highlighted in red.6 which is the main sink term) in the ε equation decreases. When fk in Eq.1 The interface conditions The interface plane (see Fig. 23. 23.4 in the present study).3 Zonal LES/RANS: wall modeling 23.4.7) ∗ Cε2 = Cε1 + fk (Cε2 − Cε1 ).3. As a result νT = Cµ • ε increases because the destruction term (last term in Eq.8) ε Note that k and ε are always positive.3. This is achieved by setting the usual convection and diffusion ﬂuxes of k at the interface to zero.5.4 interface 2 x2 URANS. the turbulence model in Eqs.

max = 3. Zonal LES/RANS: embedded LES 175 0 30 25 −0. 23.2b presents the resolved shear stresses.2 Results : Fully developed channel ﬂow is computed for Reynolds numbers Reτ = uτ δ/ν = 4 000. : Reτ = 16 000.3) is sharply reduced when going across the interface from the URANS region to the LES region and the resolved ﬂuctuations (the Reynolds shear stress in Fig. (Nx × Nz ) = (64 × 64) Reτ = 4 000.2b) increase. 16 000 and 32 000. 8 000.2.4.1.3.3 for three different resolutions in the x1 − x3 plane.8 −1 0 0. The interface is shown by thick dashed lines and it moves towards the wall for increasing Reynolds number since it is located at x+ ≃ 500 for all Reynolds numbers.1 0.4 Zonal LES/RANS: embedded LES 23. 23. The grid in the x2 direction varies between 80 and 128 cells depending on Reynolds number. 23. : Reτ = 8 000.4. ////: Reτ = 32 000. 23. Figure 23. The size of the domain is x1.max = 2 and x3. see Fig. respectively. More detailed results can be found in [146]. Hence. The interface is set to x+ ≃ 500 for all grids. 23. The interface conditions for k and ε are treated in the same way as in Section 23. the model yields grid independent results contrary to other LES/RANS models.4 −0.3. 23. As can be seen. the predicted velocity proﬁles are in good agreement with the log-law which represents experiments.23.2 U+ 15 10 5 0 1 100 1000 30000 ′ ′ v1 v2 + 20 −0. This shows that the model is switching from RANS mode to LES mode as it should.15 0.2 (b) Resolved shear stresses Figure 23.2: Velocities and resolved shear stresses.4.2. 148]. Anisotropic synthetic turbulent ﬂuctuations are used [147.6 −0. 23. The difference is now that “inlet” turbulent ﬂuctuations must be added to the LES vi equations (Eq.6 (δ = uτ = 1). x2. 2 The turbulent viscosity proﬁles are shown in Fig. It is interesting to note that the turbulent viscosity is not affected by the grid resolution. 2 The velocity proﬁles and the resolved shear stresses are presented in Fig.1 The interface conditions The interface plane is now vertical. 23. .05 x+ 2 (a) Velocities x2 0. The turbulent viscosity (Fig.max = 1. The baseline mesh has 64 × 64 cells in the streamwise (x1 ) and spanwise (x3 ) directions.3) to trigger the ¯ ﬂow into turbulence-resolving mode.

015 0.5 3 0 3.5 1 1.1 0.4 x2 0.4.0 LES. The left boundary is an inlet and the right boundary is an outlet. Interface 2 x2 x1 URANS. x1 = 0. Figure 23.015 νT /(uτ δ) 0. : Reτ = 8 000.5 x+ 2 (a) Velocities. : Reτ = 4 000. : (Nx ×Nz ) = (128 × 128) (b) (Nx ×Nz ) = (64×64).25 0.01 0. : (Nx ×Nz ) = (64×64).01 0.15 0.2 0.8 1 (a) Reτ = 4000. x1 = 3. νT /ν max . ////: Reτ = 32 000.3 0. (a) Velocities. Zonal LES/RANS: embedded LES 176 0.4 1 2.005 νT /(uτ δ) 0.5 (b) : Turbulent viscosity (right x2 axis). x1 2 2. The log-law is plotted in symbols. : Reτ = 16 000.05 0. : (Nx ×Nz ) = (32×32).5: Channel ﬂow with inlet and outlet.005 0 0 x2 0 0 0. (b) maximum resolved streamwise turbulent ﬂuctuations and turbulent viscosity versus x1 . x1 = 1. : maximum streamwise ﬂuctuations (left x2 axis) Figure 23.6 0.4: The LES and URANS regions. 4 100 30 25 20 (maxx2 u′ u′ ) 1 1 1/2 U+ 15 10 5 0 2 50 10 0 10 1 10 2 0 0 0.3: Turbulent viscosity.23. fk = 1.19.2 Figure 23.2 0.25. fk = 0.

The resolved streamwise velocity ﬂuctuations are zero in the RANS region.2 Results The Reynolds number for the channel ﬂow is Reτ = 950. Figure 23.5b). 23.25 and 3 (recall that the interface is located at x1 = 1). 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). it is seen that the present model successfully switches from RANS to LES across the interface. Inlet conditions at x = 0 are created by computing fully developed channel ﬂow with the PANS k − ε model in RANS mode (i.19. respectively. At x1 = 3.4. the wallnormal (x2 ) and the spanwise (x3 ) direction. The results will be presented in more detail in [146].4. 23. x1 = 0. a mesh with 64 × 80 × 64 cells is used in. With a 3.e.23. see Fig. with fk = 1). Zonal LES/RANS: embedded LES 177 23.5a presents the mean velocity and the resolved shear stresses at three streamwise locations. 1. .2 × 2 × 1. and the maximum resolved values increase sharply over the interface thanks to the imposed synthetic turbulent “inlet” ﬂuctuations.4.6 domain. as they should (Fig. the streamwise (x1 ). Hence. the predicted velocity agrees very well with the experimental log-law proﬁle.

1) where an = cn cos(α) .3 and Eq. The method is able to reproduce ﬁrst and second-order statistics as well as two-point correlations. DES .17. It is based on a superposition of coherent eddies where each eddy is described by a shape function that is localized in space. see Appendix I. One way is to use a pre-cursor DNS or well resolved LES of channel ﬂow. bn = cn sin(αn ).24. v′ . phase and direction of Fourier mode n.1 Synthesized turbulence The method described below was developed in [151. Inlet boundary conditions 178 24 Inlet boundary conditions In RANS it is sufﬁcient to supply proﬁles of the mean quantities such as velocity and temperature plus the turbulent quantities (e.2) A general form for a turbulent velocity ﬁeld can thus be written as N v′ (x) = 2 n=1 un cos(κn · x + ψ n )σ n ˆ (24. but in many ﬂows it seems to be sufﬁcient to prescribe constant (in time) proﬁles [103. resolved ﬂuctuations. ) the time history of the velocity and temperature need to be prescribed. 152. αn .2 Random angles The angles ϕn and θn determine the direction of the wavenumber vector κ. Let us re-write this formula as an cos(nx) + bn sin(nx) = cn cos(αn ) cos(nx) + cn sin(αn ) sin(nx) = cn cos(nx − αn ) (24. 147]. . see Eq. the time history corresponds to turbulent. Below we present a method of generating synthesized inlet ﬂuctuations. k and ε). ψ n and σi are the amplitude. .1. 24.3) n where un . cn . 154. αn denotes the direction of the velocity vector. In some ﬂows it is critical to prescribe reasonable turbulent ﬂuctuations. C. 149]. URANS. 24. The eddies are generated randomly in the inﬂow plane and then convected through it. 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. It was later further developed for inlet boundary conditions [153. The new coefﬁcient. The ˆ synthesized turbulence at one time step is generated as follows.g. 73] for creating turbulence for generating noise. in unsteady simulations (LES. 24. 24. . re-scale them and use them as inlet ﬂuctuations. and the phase angle. see Eq. However. For more details. There are different ways to create turbulent inlet boundary conditions. Another method based partly on synthesized ﬂuctuations is the vortex method [150]. A turbulent ﬂuctuating velocity ﬂuctuating ﬁeld (whose average is zero) can be expressed using a Fourier series. are related to an and bn as cn = a 2 + b 2 n n 1/2 αn = arctan bn an (24. A third method is to take resolved ﬂuctuations at a plane downstream of the inlet plane.

3. In [153.24. κmax − κ1 . of size ∆κ. 147] it was found that Lt ≃ 0.1δin is suitable. i. α = 1. equally large. 154. where κe = α9π/(55Lt). where η denotes the wall-normal direction and N2 and N3 denote the number of cells in the x2 and x3 direction.4 Smallest wave number Deﬁne the smallest wave number from κ1 = κe /p. ∆x3 = x3. where ∆ is the grid spacing. may be estimated in the same way as in RANS simulations. and the velocity unit vector. Lt . Factor p should be larger than one to make the largest scales larger than those corresponding to κe .e. 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).max /N2 . respectively. κn . 24. The turbulent length scale. σi .3 Highest wave number Deﬁne the highest wave number based on mesh resolution κmax = 2π/(2∆). The ﬂuctuations are generated on a grid with equidistant spacing (or on a weakly stretched mesh). .5 Divide the wave number range Divide the wavenumber space. 24. 24.max /N3 . Lt ∝ δ where δ denotes the inlet boundary layer thickness. are orthogi onal (in physical space) for each wave number n. into N modes.453. A value p = 2 is suitable.1: The wave-number vector. Highest wave number 179 n ξ2 n σi x3 n ξ3 n ξ1 αn n ξ3 n ξ1 θ n κn i 0 x2 x1 ϕn n Figure 24. ∆η = x2.

von K´ rm´ n spectrum a a 180 24. 24. .3 to be computed. 24. 24. ∞ k= 0 E(κ)dκ (24. i.tfd.4. 24. v2 .24.453 π Γ(1/3) where Γ(z) = 0 ∞ (24. This is unphysical.6 von K´ rm´ n spectrum a a A modiﬁed von K´ rm´ n spectrum is chosen. ˆ j N ′ v1 = 2 n=1 N ′ v2 = 2 n=1 N ′ v3 = 2 n=1 n n n β n = k1 x1 + k2 x2 + k3 x3 + ψ n un cos(β n )σ1 ˆ un cos(β n )σ2 ˆ un cos(β n )σ3 ˆ (24.7) 24.7 Computing the ﬂuctuations n Having un .6.ht 24. v3 ) are created at the inlet x2 −x3 plane.3 is then obtained from ˆ un = (E(κ)∆κ)1/2 ˆ E(κ) = cE 2 (κ/κe )4 u2 rms e[−2(κ/κη ) ] κe [1 + (κ/κe )2 ]17/6 (24. The code for generating the isotropic ﬂuctuations can be downloaded here http://www.5) which gives [57] 4 Γ(17/6) cE = √ ≃ 1.e. allows the expression in Eq.8 Introducing time correlation A ﬂuctuating velocity ﬁeld is generated each time step as described above. κn . 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. They are independent of each other and their time correlation will thus be zero. σi and ψ n .2. The amplitude a a un of each mode in Eq.4 and Fig.e.6) e−z xz−1 dz ′ ′ (24. ˆ ′ ′ ′ In this way inlet ﬂuctuating velocity ﬁelds (v1 .se/˜lada/projects/inlet-boundary-conditions/proright.8) where un is computed from Eq.chalmers.4) κ = (κi κi )1/2 . see Eq. 24.

24.in . new ﬂuctuating velocity ﬁelds.5 which ensures that V1 = v1 ( · denotes averaging). t) ¯ (24.11) v3 (0. t) ¯ 1. x3 .in = (V2 )m and v3.in (x2 ) + v2. V3 . x3 . x2 . Introducing time correlation 181 E(κ) κ1 E(κn ) κe E(κ) ∝ κ−5/3 ∆κn κn Figure 24.9 is a convenient way to prescribe the turbulent time scale of the ﬂuctuations. t) = V2.in . V2 . t) = V1.in as [155] + x+ ≤ 5 x2 2 + + (24.in = (V1 )m . v2. ′2 ′2 The second coefﬁcient is taken as b = (1 − a2 )0. V3. The time correlation of will be equal to exp(−τ /T ) (24.in (x2 . if V2.in = −3. x3 ) v1 (0. V2.in ′ v2 (0. x3 .in (x2 ) + u′ (x2 . x3 . are 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 ′ (V3 )m (24. t) ′ ′ ′ ′ ′ ′ where v1.in .05 + 5 ln(x2 ) 5 < x+ < 30 2 1 ln(x+ ) + B x+ ≥ 30 2 2 κ .in = V3. 24. x3 . V1. x3 . for example. a RANS solution or from the law of the wall.9) = ′ a(V3 )m−1 + ′ b(v3 )m where m denotes the time step number and a = exp(−∆t/T ).8. are either taken from experimental data. x2 . 24.in (x2 . The mean inlet proﬁles.in = 0 we can estimate V1. x2 . 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.9).10) where τ is the time separation and thus Eq.2: Modiﬁed von K´ rm´ n spectrum a a log(κ) ′ ′ ′ To create correlation in time.12) V1. t) = V3. V1 .in = (V3 )m (see Eq.in (x2 ) + ¯ ′ v3.

. (V1 )m . t).10. 24.2 0.9. 24. ′ Eq.8 0.6 B(τ ) 0.2 0 0 0. : computed from synthetic data.2.24.8. : where κ = 0.4 τ 0.3: Auto correlation.8 1 ′ ′ Figure 24.6 0. B(τ ) = v1 (t)v1 (t − τ )t (averaged over time.4 0.4 and B = 5. The method to prescribed ﬂuctuating inlet boundary conditions have been used for channel ﬂow [147]. Introducing time correlation 182 1 0. for diffuser ﬂow [149] as well as for the ﬂow over a bump and an axisymmetric hill [156]. see Eq.

uk/ATAAC/WebHome 25.mace. turbulence models etc are deﬁned and the participants in the workshops and EU projects carry out CFD simulations for these test cases.org/ﬁleadmin/user upload/bigﬁles/sig15/database/index.se/˜lada/projects/lesfoil/proright.chalmers. The Special Interest Group Sig15 is focused on evaluating turbulence models. 25.25. Turbulence And Combustion).uk/ﬂomania/ DESIDER: Detached Eddy Simulation for Industrial Aerodynamics http://cfd.2 Ercoftac workshops Workshops are organized by Ercoftac (European Research Community On Flow.tfd. The publication Industrial Computational Fluid Dynamics of Single-Phase Flows can be ordered on Ercoftac www page http://www.ac.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.ercoftac. Test cases with mandatory grids.org/publications/ercoftac best practice guidelines/single-phase ﬂows spf/ . or that the discretization scheme in one code was more diffusive than in the other.ac.html FLOMANIA: Flow Physics Modelling: An Integrated Approach http://cfd. Today the situation is much improved.manchester. different CFD codes used to give different results.html Ercoftac also organizes workshops and courses on Best Practice Guidelines. Then they compare and discuss their results. The main reason for this improved situation is because of workshops and EU projects where academics.ac. boundary conditions.manchester.mace.ercoftac. The reasons to these differences could be that the turbulence model was not implemented in exactly the same way in the two codes.uk/desider ATAAC: Advanced Turbulence Simulation for Aerodynamic Application Challenges http://cfd. Two different CFD codes usually give the same results on the same grid. The outcome from all workshop are presented here http://www. Best practice guidelines (BPG) 183 25 Best practice guidelines (BPG) In the early days of CFD.mace. Even if the same grid and the same turbulence model were used. There could be small differences in the implementation of the boundary conditions in the two codes. there could be substantial differences between the results. engineers from industry and CFD software vendors regularly meet and discuss different aspects of CFD.manchester.

External Aerodynamics. Turbomachinery. visit ERCOFTAC QNET Knowledge Base Wiki http://www. Underlying Flow Regimes.4 ERCOFTAC QNET Knowledge Base Wiki The QNET is also the responsibility of Ercoftac. Chemical and Process Engineering. Ercoftac Classical Database 184 25. which includes some 100 experimental investigations. Combustion and Heat Transfer etc.3. These are sector disciplines such as Built Environment. well-studied test cases capturing important elements of the key ﬂow physics encountered across the Application Areas. Here you ﬁnd descriptions of how CFD simulations of more than 60 different ﬂows were carried out. The ﬂows are divided into Application Areas.25.org/products and services/classic collection database 25. These are generic.3 Ercoftac Classical Database A Classical Database. These are realistic industrial test cases which can be used to judge the competency and limitations of CFD for a given Application Area.ercoftac.ercoftac. can be found at Ercoftac’s www page http://www. Each Application Area is comprised of Application Challenges.org/products and services/wiki/ . For more information.

The ε − δ identity reads i 1 2 1 1 3 1 2 3 2 n 2 1 2 3 1 3 3 2 3 j 1 1 2 1 1 3 2 2 3 k 2 2 1 3 3 1 3 3 2 εinm εmjk ε12m εm12 = ε123 ε312 = 1 · 1 = 1 ε21m εm12 = ε213 ε312 = −1 · 1 = −1 ε12m εm21 = ε123 ε321 = 1 · −1 = −1 ε13m εm13 = ε132 ε213 = −1 · −1 = 1 ε31m εm13 = ε312 ε213 = 1 · −1 = −1 ε13m εm31 = ε132 ε231 = −1 · 1 = −1 ε23m εm23 = ε231 ε123 = 1 · 1 = 1 ε32m εm23 = ε321 ε123 = −1 · 1 = −1 ε23m εm32 = ε231 ε132 = 1 · −1 = −1 δij δnk − δik δnj 1−0=1 0 − 1 = −1 0 − 1 = −1 1−0=1 0 − 1 = −1 0 − 1 = −1 1−0=1 0 − 1 = −1 0 − 1 = −1 Table A.A. .1: The components of the ε − δ identity which are non-zero.1 the components of the ε − δ identity are given. TME225: ε − δ identity 185 A TME225: ε − δ identity εinm εmjk = εmin εmjk = εnmi εmjk = δij δnk − δik δnj In Table A.

ﬂow between two parallel plates).se/˜lada/MoF/. Divide the report into sections corresponding to the sections B.:) are the v1 values at the outlet. • First. i. Open Matlab and execute channel flow. the corresponding Matlab arrays are v1 2d.:) are the v1 values at the inlet. At the end of this Assignment the group should write and submit a report (in English).m which reads the data and plot some results. Octave is a Matlab clone which can be downloaded for free. v1 2d(1. B. There are three ﬁeld variables.6385m. Hence in v1 2d(:.9.011m and L = 0. the ﬂuid is air of 20o C. .B:. you cannot assume that the ﬂow is fully developed). but we don’t recommend it).e. You will use Matlab to analyze the data.chalmers. see Fig. ni = 199 grid points in the x1 direction and nj = 22 grid points in the x2 direction. TME225 Assignment 1: laminar ﬂow 186 V x2 x1 L Figure B. v1 2d(:.m in an editor and make sure that you understand it. v1 2d(i.e.tfd. The ﬂow is steady and incompressible. The grid is 199 × 22.1) are the v1 values at the lower wall. v1 2d(ni.7. these should clearly be described and presented.B. h B TME225 Assignment 1: laminar ﬂow You will get results of a developing two-dimensional channel ﬂow (i. download the data ﬁle channel flow data.1. Present the results in each section with a ﬁgure (or a numerical value).1: Flow between two plates (not to scale). The work should be carried out in groups of two (you may also do it on your own.e.dat and the m-ﬁle channel flow. ﬁnd out and write down the governing equations (N.j). v1 . i. The results should also be discussed and – as far as you can – explained.nj) are the v1 values at the upper wall. We denote the ﬁrst index as i and the second index as j. The height of the channel is h = 0. The simulations have been done with Calc-BFC [157]. v2 2d and p 2d. v2 and p. You can also use Octave on Linux/Ubuntu. In some sections you need to make derivations. From the course www page http://www.1 – B. The inlet boundary condition (left boundary) is v1 = Vin = 0. Open channel flow. The ﬁeld variables are stored at these grid points.

1. 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. |∂v1 /∂x1 | < 0. How do they compare? In the fully developed region.2.5 and 2. x2 )dx2 what do you get? How does ξ(x1 ) vary in the x1 direction? How should it vary? . use Cauchy’s formula (see [2].1. Skk = 0 because of the continuity equation. The expression for τij can be found in Eqs. nj = (0.7) to the fully developed proﬁle somewhere downstream. is computed as τw.B.4 and 1. How well does this agree with your values? In the fully developed region.2 to compute the wall shear stress at the upper wall. On the top wall. 0)).1.L ≡ τ21. the wall shear stress.L versus x1 . If you use Eq. Instead. recall that the ﬂow in incompressible. compare the velocity proﬁle with the analytical proﬁle (see Section 3.e. ∂v1 /∂x1 = 0. If you.2)? What value does it take (at x2 = h/4. h ξ(x1 ) = 0 v1 (x1 .1 Fully developed region Fully developed conditions mean that the ﬂow does not change in the streamwise direction. What value should it take in the fully developed region (see Section 3. compare with the analytical value (see Eq.2.e. Use Eq. for a ﬁxed x1 .3 Inlet region In the inlet region the ﬂow is developing from its inlet proﬁle (v1 = V = 0. for example. The v1 velocity is decelerated in the near-wall regions.2 Wall shear stress ∂v1 ∂x2 On the lower wall.U .e. 2. Plot the two wall shear stresses in the same ﬁgure.2. Eq. −1. Plot also ∂v1 /∂x1 . Plot τw. and hence the v1 velocity in the center must increase due to continuity.2). 4.2) ti (ˆ ) n = τji nj (B. τw. Fully developed region 187 B. Plot v1 in the center and near the wall as a function of x1 .01. B. for example)? B. a distance taken from the literature is given. i. Look at the vertical velocity component.01. B. What x1 value do you get? In Section 3. i.4. τw.L (index L denotes Lower).2.e.3.5) but at the wall ∂v2 /∂x1 = 0. you get the incorrect sign.L = µ (B. Why does it behave as it does? Now we will compute the wall shear stress at the upper wall. Chapt. 3. 99% of its ﬁnal value. integrate v1 . i.30). B.w. 2. the normal vector points out from the surface (i. If we deﬁne “fully developed” as the location where the velocity gradient in the center becomes smaller than 0.1) L Recall that τ12 = µ(∂v1 /∂x2 + ∂v2 /∂x1 ) (see Eqs. v2 .2) which is a general way to compute the stress vector on a surface whose (outward pointˆ ing) normal vector is n = nj .

is the ﬂow rotational anywhere? Why? Why not? B. (see Eq. by dissipation). . what happens with the vorticity in the inlet region? Now. q1 . Now you can compute the increase in bulk temperature. 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.5 Vorticity Do you expect the ﬂow to be irrotational anywhere? Let’s ﬁnd out by computing the vorticity. Wall-normal velocity in the developing region 188 B. and a vorticity tensor. i. Since the ﬂow is two-dimensional. Assume that the upper and lower walls are adiabatic.6 Deformation In Section 1. x2 .4.8 Eigenvalues Compute and plot the eigenvalues of the viscous stress tensor. v1 near the walls decreases for increasing x1 . furthermore we can neglect the heat ﬂux by conduction. the temperature. 2. Where is it largest? Plot the vorticity also in the inlet and developing regions. τij .5. If you have computed the four elements of the τij matrix you can use the following commands: tau=[tau_11 tau 12. see Section 1.e.lambda]=eig(tau).e.5.4 Wall-normal velocity in the developing region In Section B.9.B. Use Eq. This means that dissipation increases the internal energy. we divided the velocity gradient into a strain-rate tensor.3. tau_21 tau_22]. The bulk temperature is deﬁned at h 0 v1 T dx2 Tb = h v dx2 0 1 B.7 Dissipation Compute and plot the dissipation. respectively? Compare Ωij with the vorticity you plotted in Section B. [n. Use the Matlab command eig. Sij . Where are they largest? Why? What is the physical meaning of Sij and Ωij . 2. in the developing region. Plot it in the fully developed region as ω3 vs. Ωij . Φ = τji ∂vi /∂xj . we have only two off-diagonal terms (which ones?). Plot and compare one of the off-diagonal term of Sij and Ωij . 22. re-write the left side on conservative form and then apply the Gauss divergence theorem. see Eq.3 we found that. Next. Start by integrating the dissipation over the entire computational domain. This is discussed in some detail at p. 2. What about v2 ? How do you explain the behaviour of v2 ? B. Tb . from inlet to outlet.11) at the inlet and outlet.14 to compute the temperature increase that is created by the ﬂow (i. Are they similar? Any comment? B.

9 Eigenvectors Compute and plot the eigenvectors of τij .9. both v1 and −ˆ1 in Fig. Eigenvectors 189 where n and lambda denote eigenvalues and eigenvectors. Plot one eigenvalue as a x− y graph (eigenvalue versus x2 ). Recall that the sign of the eigenvector is not deﬁned (for example. 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? B. versus x2 . Plot also the four stress components. τij . An eigenvector is. Try to analyze why the eigenvectors ˆ v behave as they do. perpendicular to each other. It is enough to plot one of them. Recall that at each point you will get two eigenvectors.7)? Pick an x1 location where the ﬂow is fully developed. Use the Matlab command quiver to plot the ﬁeld of the eigenvectors. 1.B.10 at p. respectively. a vector. What is the physical meaning of the eigenvalues (see Chapter 1. 19 are eigenvectors). of course. .

e. C.3) . C.2) 2 where |Vi | denotes the length of Vi . Any vector.1: Scalar product. as a · b = (a|b).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 ) = c1 |V1 | + c2 |V2 | + c3 |V3 | = ci |Vi | 2 2 2 (C. in physical space which form an orthogonal base in R3 (i. Hence the coordinates.1) see Fig. V1 . the second line follows because of the orthogonality of Vi . T. V3 . i.1 Orthogonal functions Consider three vectors. can be determined by making a scalar product of Eq. Let us call them basis functions. in R3 can now be expressed in these three vectors. their scalar products are zero). V2 . T = c1 V1 + c2 V2 + c3 V3 (C. V1 C TME225: Fourier series Here a brief introduction to Fourier series extracted from [158] is given.1.e. TME225: Fourier series V2 190 T (V2 |T) (V1 |T) Figure C. a and b. ci . ci . are determined by ci = (T|Vi )/|Vi |2 (C. Now deﬁne the scalar product of two vectors. The coordinates.C. C.

The question is now how to choose the orthogonal function system {g}∞ on the interval 1 [−π. ci+1 (gi+1 |gi ) . C. as in Eq.1.] (C. . The length of the basis function. .2 Trigonometric functions Here we choose gn as trigonometric functions which are periodic in [−π. over the interval [a. ∞ (f |gi ) = n=1 cn (gn |gi ) (C. with orthogonal basis functions {g}∞ . cn . C. is ||gn || 3. Eq. |Vi | which is the length of Vi in Eq. ||gn || = (gn |gn ) 1/2 b 1/2 = a gn (x)gn (x)dx (C. The coordinates of Vi are computed as ci = (T|Vi )/|Vi |2 1.9) Let us now summarize and compare the basis functions in physical space and the basis functions in functional space 1. The coordinates of gn are computed as cn = (f |gn )/||gn ||2 C. π]. Vi . Multiply. Any vector in R3 can be expressed in the orthogonal basis vectors Vi 2.e. C. ci . The “scalar product” of two functions.6 is non-zero only if i = n. be expressed as ∞ f= n=1 cn g n (C. . is |Vi | 3.2. . . in Eq. (f |gi ) = (c1 g1 |gi ) + (c2 g2 |gi ) .2.7) Similar to Eq. f and gn . Trigonometric functions 191 Now let us deﬁne an inﬁnite (∞-dimensional) functional space. . cos(nx). So let us “guess” that the trigonometric series [1.1). we usually start by doing an intelligent “guess”.e. the coordinates. in a similar way to Eq. ci (gi |gi ) . sin x. = = ci (gi |gi ) = ci ||gi ||2 (C.3.8) The “coordinates”.10) .4) Then. C.C. . . . C. π]. .5) As above. . sin(2x). The length of the basis vector.6) Since we know that all gn are orthogonal. . B.3). . i. In mathematics. i. we must now ﬁnd the “coordinates” (cf. f with the basis functions. Any function in [a. gi . cos x. any function can. and then we prove that it is correct. sin(nx). is deﬁned as 1 b (f |gn ) = f (x)gn (x)dx a (C. the “coordinates” can be found from (switch from index i to n) cn = (f |gn )/||gn ||2 (C.e. gn . b] can be expressed in the orthogonal basis functions gn 2. b]. i. C. are called the Fourier coefﬁcients to f in system {g}∞ and 1 ||gn || is the “length” of gn (cf.

4. . .). π] and we need to compute their “length” (i.13) Orthogonality of φn and ψk π (φn |ψk ) = − sin(nx) cos(kx)dx = −π 1 2 π −π [sin((n + k)x) + sin((n − k)x)] dx = π 1 1 1 cos((n + k)x) + cos((n − k)x) 2 n+k n−k =0 −π (C.) and ψk (x) = cos(kx) (k = 0. . (C. .12) “Length” of ψk (ψk |ψk ) = ||ψk ||2 = (ψ0 |ψ0 ) = ||ψ0 ||2 = π −π π −π cos2 (kx)dx = 1 · dx = 2π x 1 + sin(2x) 2 4 π = π for k > 0 −π (C. 1.e. . i.C. ψk (x).10 can be deﬁned as gn (x) = φk (x).14) Orthogonality of φn and φk π (φn |φk ) = = sin(nx) sin(kx)dx = −π 1 2 π −π [cos((n − k)x) − cos((n + k)x)] dx π 1 1 1 sin((n − k)x) − sin((n + k)x) 2 n−k n+k = 0 for k = n −π (C.2.11) where φk (x) = sin(kx) (k = 1.16) . for n = 2k = 2. The function system in Eq. their norm). 2. . .15) “Length” of φk (φk |φk ) = ||φk ||2 = π −π sin2 (kx)dx = x 1 − sin(2x) 2 4 π −π = π for k ≥ 1 (C. for n = 2k + 1 = 1. Now we need to show that they are orthogonal. Trigonometric functions 192 is an orthogonal system. 3.e. that the integral of the product of any two functions φk and ψk is zero on B[−π. . . C. Orthogonality of ψn and ψk π (ψn |ψk ) = = cos(nx) cos(kx)dx = −π 1 2 π −π [cos((n + k)x) + cos((n − k)x)] dx π 1 1 1 sin((n + k)x) + sin((n − k)x) 2 n+k n−k = 0 for k = n −π (C. . . .

e.e. Fourier series of a function 193 C. Taking the average of f (i.C. C.18c) f (x) cos(nx)dx. if f = 0 then a0 = 0.19a) ¯ f= ¯ Hence. we know that we can express any periodic function. i.4 Derivation of Parseval’s formula Parseval’s formula reads π (f (x))2 dx = −π π 2 a +π (a2 + b2 ) n n 2 0 n=1 ∞ We will try to prove this formula.21) Now we want to prove that the Fourier coefﬁcients are the best choice to minimize the difference N ||f − n=1 an gn || (C.18a) (C.11 forms an orthogonal system of func1 tions.20) C. C. π f (x)dx = πa0 −π (C.17) where x is a spatial coordinate. Assume that we want to approximate the function f as well as possible with an orthogonal series ∞ an g n n=1 (C. then a0 is obtained from Eq. f (x) = a0 (an cos(nx) + bn sin(nx)) + 2 n=1 π ∞ (C.18b.19a) (C.19c) 1 π 1 an = (f |ψn )/||ψn ||2 = π bn = (f |φn )/||φn ||2 = f (x) sin(nx)dx −π π f (x) cos(nx)dx −π Note that a0 corresponds to the average of f . integrating f from −π to π) gives (see Eq. f (with a period of 2π) in {g}∞ as 1 ∞ f (x) = c + (an cos(nx) + bn sin(nx)) n=1 (C.19b) (C. C.3 Fourier series of a function Now that we have proved that {g}∞ in Eq. n > 0 −π π f (x)dx −π If we set c = a0 /2. The Fourier coefﬁents are given by 1 π 1 an = (f |ψn )/||ψn ||2 = π 1 c = (f |ψ0 )/||ψ0 ||2 = 2π bn = (f |φn )/||φn ||2 = π f (x) sin(nx)dx −π π (C.3.22) .18b) (C.

C.5 and C. Appendix D describes in detail how to create energy spectra from two-point correlations.24) = ||f ||2 + N n=1 ||gn ||2 a2 − 2an cn n N 2 = ||f ||2 + n=1 ||gn ||2 (an − cn ) − n=1 ||gn ||2 c2 n The left side of Eq. C. i.23) because of the orthogonality of the function system. π ||f ||2 ≡ (f (x))2 dx = −π π 2 a +π (a2 + b2 ) n n 2 0 n=1 ∞ Note that π/2 and π on the right side correspond to the “length” of ||gn ||. the magnitude of the left side increases. {g}N .C. and its magnitude gets closer and closer to that of the right side. Derivation of Parseval’s formula 194 Later we will let N → ∞.25) The left side must always be positive and hence N n=1 ||gn ||2 c2 ≤ ||f ||2 = n π −π (f (x))2 dx for all N (C. ||ψ0 ||.e.8) gives N N (f |f ) − 2 n=1 an cn (gn |gn ) + N n=1 a2 (gn |gn ) n (C. Using the deﬁnition of the norm and the laws of scalar product we can write N N N ||f − = (f |f ) − = (f |f ) − 2 n=1 an gn ||2 = f− N n=1 an g n f − N ak g k k=1 N n=1 N n=1 N an (f |gn ) − an (f |gn ) + k=1 N n=1 ak (f |gk ) + a2 (gn |gn ) n n=1 k=1 an ak (gn |gk ) = (C.23 is thus minimized if the coefﬁcients an are chosen as the Fourier coefﬁcients. which gives Parseval’s formula. Using the Fourier coefﬁcients in Eq. cn so that N N ||f − n=1 an gn ||2 = ||f ||2 − n=1 ||gn ||2 c2 n (C.26) As N is made larger. ||ψn || and ||φn ||. respectively. but it will always stay smaller than ||f ||2 . Express f in the second 1 term using the Fourier coefﬁcients cn (see Eqs.4. This means the series on the left side is convergent.19 and letting N → ∞ it can be shown that we get equality of the left and right side. . C.

It is more convenient to express a Fourier series in complex variables even if the function f itself is real. cn . On complex form it reads ∞ f (x) = cn = 1 2π cn exp(ınx)) n=−∞ π (C. This means that the real coefﬁcients are multiplied by a factor 1 two except the ﬁrst coefﬁcient. cn . C.19. Complex Fourier series 195 C. C. One advantage of Eq. C.27b) f (x) exp(−ınx)dx −π where the Fourier coefﬁcients.31) which veriﬁes that the complex Fourier series for a real function f is indeed identical to the usual formulation in Eq. The trick in the formulation in Eq.27 is that the imaginary coefﬁcients for negative and positive n cancel whereas the real coefﬁcients add. are complex.28) where an and bn are given by Eq. For negative n in Eq. C.30 into Eq.19. are complex.28. C.19.27a) (C.5. 2 n > 0 (C.27 gives f (x) = = = 1 1 a0 + (an − ıbn ) exp(ınx) + (an + ıbn ) exp(−ınx) 2 2 n=1 1 1 a0 + (an − ıbn )(cos(nx) + ı sin(nx)) + (an + ıbn )(cos(nx) − ı sin(nx)) 2 2 n=1 1 1 a0 + an cos(nx) − ı2 bn sin(nx) = a0 + an cos(nx) + bn sin(nx) 2 2 n=1 n=1 (C.29 and C. C.5 Complex Fourier series Equation C. For n = 0.27 is equivalent to Eq. C. The Fourier coefﬁcients.30) see Eq.19 gives the Fourier series of a real function. C.C. then Eq. C.19. Below we verify that if f is real.27 we get c−n = c∗ = n 1 2π π f (x)(cos(nx) + ı sin(nx))dx = −π 1 (an + ıbn ).27 over the formulation in Eq. which makes up for the factor 2 in front of a0 in Eq. cn . Eq. a0 . 2 n>0 (C. Inserting Eqs.19 although the Fourier coefﬁcients. C. read – assuming that f is real – according to Eq. C. ∞ ∞ ∞ ∞ . a0 .19 is that we don’t need any special deﬁnition for the ﬁrst Fourier coefﬁcient. C. C.29) where c∗ denotes the complex conjugate.27 reads n c0 = 1 2π π f (x)dx = −π 1 a0 2 (C.27 cn = 1 2π π −π f (x)(cos(nx) − ı sin(nx))dx = 1 (an − ıbn ). C. C.

On discrete form the expression for B33 reads B33 (k∆z) = 1 M M m=1 ′ ′ v3 (x3 − k∆z)v3 (x3 ) (D.2 An example of using FFT Here we will present a simple example. Here we assume that x3 is an ˆ homogeneous direction so that B33 is independent of x3 . we are interested to look at the energy spectra. from a two-point correlation in time). Finally. −L κ = 2π/L (D. Let’s use this function as input vector for the discrete Fourier transform (DFT) using Matlab. see Fig. we Fourier transform N instantaneous signals in space and then time average the N Fourier transforms.e. B33 = B33 (ˆ3 ).1) where x3 is the separation between the two points.2) x ′ B(x3 . An alternative way is to Fourier transform of a (time-averaged) two-point correlation. so we expect the Fourier coefﬁcients to be real. x3 ) = v3 (x3 − x3 )v3 (x3 ) ˆ ˆ ′ (D.5) . Then we show how to derive the energy spectrum from a spatial two-point correlation. 10. 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. In the following section we give a simple example how to use Matlab to Fourier transform a signal where we know the answer. In Matlab the DFT of u is deﬁned as (type help fft at the Matlab prompt) N U (k) = n=1 un exp −ı2π(k − 1)(n − 1) N (D.e. at which wave numbers) the ﬂuctuating kinetic turbulent energy reside. The x two-point correlation for an inﬁnite channel ﬂow is shown in Fig.1 Introduction When analyzing DNS or LES data. The function u is symmetric. time plus spatial homogeneous directions). From these we can ﬁnd out in which turbulence scales (i. Consider the function u = 1 + cos(2πx/L) = 1 + cos(2π(n − 1)/N ) (D.2) where m denotes summation in homogeneous directions (i. in the the continuous FFT U c (κ) = 1 L L 1≤k≤N u(x) exp(−ıκx)dx.e. If we want to have the energy spectrum versus wavenumber. D. TME225: Compute energy spectra from LES/DNS data using Matlab 196 D TME225: Compute energy spectra from LES/DNS data using Matlab D. D.e. B33 (ˆ3 ).3) where L is the length of the domain and N = 16 is the number of discrete points.1.4) √ where k is the non-dimensional wavenumber and ı = −1.D. i. which is deﬁned as (see Eq.2. D. x. some comments are given on how to create an energy spectrum versus frequency from an autocorrelation (i. The ratio (n − 1)/N corresponds to the physical coordinate.

3 using the commands N=16.3 0. In Matlab.4 1. U (k)/N . 2π] is used whereas the formulation in Eq.5 16 u 1 0.2.5 x 2 2. n=1:1:N.4 must be divided by N . Furthermore.1 0.4 0.2 1 B33 (ˆ3 ) x 0.2 0. 16 nodes are used. .5 3 Figure D.5 0.6 0. D.5). D.4 0. it can be noted that in Eq. The u function is shown in Fig.5 9 0 0 0. of DNS data in x ˆ ′ channel ﬂow taken from [73]. An example of using FFT 197 Note that the discrete Fourier U (k) coefﬁcients in Eq. we generate the function u in Eq. D.2: The u function.2 0 0. D. D.e. in order to correspond to the Fourier coefﬁcients U c (N corresponds to L in Eq. π].4 the period [0.8 0. B(ˆ3 ) = v3 (x3 − x3 )v3 (x3 ) .2 0 −0.5 is based on the interval [−π.2.6 x3 ˆ ′ Figure D. D.D. 1. node 1 is located at x = 0 and node 16 is located at 15L/16. i.1: Two-point correlation.5 1 1. u=1+cos(2*pi*(n-1)/N). 2 2 1.

The resulting Fourier coefﬁcients are shown in Fig. 198 Instead of using the built-in fft command in Matlab we can program Eq.e. The remaining coefﬁcients are zero.7) In wavenumber space the energy is – according to Parseval’s formula. end end Note that since u is symmetric.3. U (16)/N = 0. D.4 directly in Matlab as U=zeros(1.5.3 Energy spectrum from the two-point correlation Now that we have learnt how to use the FFT command in Matlab. Eqs. Equation D.6) which corresponds to cos(2πx/L) in Eq. The energy.1 and Fig. In Fig. i. u2 .4.3. U (1)/N = u . 16] corresponds to the negative.4 reads N ˆ B33 (k) = n=1 B33 (n) exp −ı2π(k − 1)(n − 1) N (D.4 and D.1. Two of them. see Eq.3. D. k. for k=1:N for n=1:N arg1=2*pi*(k-1)*(n-1)/N. D. D. U/N is plotted versus non-dimensional wavenumber. U(k)=U(k)+u(n)*cos(-arg1). let’s use it on our two-point correlation in Eq. N = 9.9) . i. D. Energy spectrum from the two-point correlation Now we take the discrete Fourier transform of u. we have only used cos(−x) = cos(x) (the symmetric part of exp(−ıx)).5.2 can be computed as u2 = 1 L L N u2 (x)dx = 0 n=1 u2 /N = 1. Type U=fft(u).5). It can be noted that the interval [k = N/2 + 1.8) D. of the signal in Fig.4 by inserting k = 1. u2 = 1 L ∞ U 2 (κ)dκ = 0 1 N N 2 Un /N = 1. Since the function u includes only one cosine function and a mean (which is equal to one) only three Fourier coefﬁcient are non-zero. This is easily veriﬁed from Eq.4 – equal to the integral of the square of the Fourier coefﬁcients. C.5 n (D.N).D. and versus wavenumber κ = 2π(n − 1)/L. correspond to the cosine functions (there must be two since U is symmetric) cos((N − 1)2π(n − 1)/N ) = cos(−2π(n − 1)/N ) = cos(2π(n − 1)/N ) cos(2π(n − 1)/N ) (D. D. D.5 n=1 (D. symmetric part of the wavenumbers in the physical formulation (cf. U (2)/N = 0.e. D. D. The ﬁrst Fourier coefﬁcient corresponds – as always – to the mean of u.

6 0.5b shows the same energy spectra in log-log scale (only half of the spectrum is included).D. i. B33 (ˆ3 ) = v3 (x3 )v3 (x3 + x3 ) . ˆ B33 (1)/N is equal to the mean of B33 . i. Energy spectrum from the two-point correlation 199 1. it is sufﬁcient to x use the cosine part of Eq.2 0 −0. Figure D. Figure D.3: The U/N Fourier coefﬁcients.2 0 −0.5 1 1. where x3.10.4.2 1 1 96 B33 (ˆ3 ) x 0.4: Periodic two-point correlation. The simulations have been carried out with periodic boundary conditions in x3 direction (and x1 ). D. and hence B33 (ˆ3 ) is symmetric.4 0. Thus. As usual.e.e. The dashed line shows the −5/3 slope which indicates that the energy spectra from the DNS follows the Kolmogorov −5/3 decay.11) compare with Eq.6 0.8 U/N 0.8 0.5a the Fourier coefﬁcients B33 κ3 are presented versus wavenumber κ3 = 2π(n − 1)/x3. D.9.4. N ˆ B33 (k) = n=1 B33 (n) cos 2π(k − 1)(n − 1) N (D. Note that this is almost the same expression as that for the .e.4 1.2 0 5 10 15 20 0 0 5 10 15 20 k (a) Versus non-dimensional wavenumber. see Fig. the Fourier coefﬁcient for the ﬁrst non-dimensional wavenumber.8 0. κ (b) Versus wavenumber. D.2 0. D.3.10) ˆ In Fig. k. which is the common way to present energy spectra.2 1 1 0.max ≃ 1. of DNS x ˆ data in channel ﬂow taken from [73].55. B33 = 1 N N n=1 B33 (n) ≡ 1 ˆ B33 (1) N (D.2 0 0. κ.max . see Fig.4 0.5 x3 ˆ ′ ′ Figure D. D.4 0.6 U/N 0. 1. i.

10. We can use the same approach in time.12) Hence the integral length scale is related to the ﬁrst Fourier mode as Lint = ˆ B33 (1) ′2 N v3 (D. In the previous section we computed the energy spectrum versus wavenumber by Fourier transforming the two-point correlation. Fourier transform each subset and then average. Another way to obtain v3 ′2 v3 = 1 x3. x3 )dˆ3 = ˆ x B33 ′2 v3 (D. Then B33 (τ ) is Fourier transformed to get B33 (f ) in ˆ33 (τ ) is a function of the same way as in Section D. κ3 .52 n=1 (D. say.5) Lint (x3 ) = 1 ′2 v3 0 ∞ B33 (x3 .D.15 10 0.4 Energy spectra from the autocorrelation ′ When computing the energy spectra of the v3 velocity. i.05 10 −2 −3 0 0 −4 100 200 κ3 300 400 10 1 (a) Linear-linear scale (b) Log-log scale κ3 10 2 ′2 Figure D.1 10 0. Dashed line in b) show −5/3 slope. ′ ′ First we create the autocorrelation B33 (τ ) = v3 (t)v3 (t + τ ) (this can be seen as a ˆ two-point correlation in time).51. D.3. The only difference is that B ˆ frequency whereas B33 (κ3 ) is a function of wavenumber. .25 10 −1 200 0. i. In Matlab. versus frequency.4.e. integral length scale which reads (see Eq. We can also split the time signal into a number subsets.2 B33 (k)/N 0.5: The energy spectrum of v3 versus wavenumber.5.14) D. the command pwelch is a convenient command which does all this. f .13) ′2 ′2 The two-point correlation for zero separation is equal to v3 . 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.max 0 ∞ 1 ˆ B33 (κ3 )dκ3 = N N ˆ B33 (n) = 1. Energy spectra from the autocorrelation 0. B33 (0) = v3 = ′2 is to integrate the energy spectrum in Fig. 1.e.

’b-’.u2. The sampling time step 2 2 2 2 is ∆t = 0. E.0033 (every second time step). x2 /δ = 0. x+ = 53. A 96 × 96 × 96 mesh has been used. see Fig. Make a zoom between. This is conveniently done with the command axis([10 11 3 21]) In order to see the value at each sampling time step.chalmers.E.m which loads and plots the time history of v1 . 53.0176. x+ = 8.org) which is free to download.47. to look at the details of the time history. wall-normal and spanwise directions are denoted by x (x1 ). Recall that the velocities have been scaled with the friction velocity uτ . You can also use Octave on Linux/Ubuntu.dat with the time history of v1 and v2 . 6. In DNS the unsteady. TME225 Assignment 2: turbulent ﬂow 201 E TME225 Assignment 2: turbulent ﬂow In this exercise you will use data from a Direct Numerical Simulation (DNS) for fully developed channel ﬂow.se/˜lada/MoF/ you ﬁnd a ﬁle u v time 4nodes. three-dimensional Navier-Stokes equations are solved numerically. the buffer layer and in the logarithmic layer. Study the time history of the blue line (x2 /δ = 0. It is recommended A A (but the not required) that you use LTEX(an example of how to write in LTEXis available on the course www page).0654 and ∆z = 0.’bo’) Use this technique to zoom. The time history of v1 at x2 /δ = 0.5 and x+ = 235.2 Time averaging Compute the average of the v1 velocity at node 2. and thus what you see is really v1 /uτ .95.u2. You should write a report where you analyze the results following the heading H1–H13. t = 10 and t = 11 and v1. change the plot command to plot(t. On Windows you can use.0039. Periodic boundary conditions were applied in the x and z direction (homogeneous directions). All data have been made non-dimensional by uτ and ρ. The four points are located in the viscous sublayer.tfd.0164.0176) more in detail. you can use the zoom buttons above the ﬁgure. y (x2 ) and z (x3 ) respectively. Start Matlab and run the program pl time. With uτ = 1 and ν = 1/Reτ = 1/500 this correspond to x+ = 1.107 and x2 /δ = 0. MikTex (www.miktex. You can do the assignment on your own or in a group of two.2 at p.107 are shown. for example. The streamwise.1 Time history At the course home page http://www. 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 ). It is available on Linux.min = 3 and v1.t.min = 21. What is the differences between v1 and v2 ? E.8. Octave is a Matlab clone which can be downloaded for free. How does the time variation of v1 vary for different positions? Plot also v2 at the four different positions. Start the program pl time. The cell size in x and z directions are ∆x = 0. Alternatively. The ﬁle has eight columns of v1 and v2 at four nodes: x2 /δ = 0.0176 and x2 /δ = 0. Add the following code (before the plotting section) . Plot v1 for all four nodes. Use the Matlab. x2 /δ = 0. for example.

is computed as u1_mean=mean(u1). Recall that all terms in the equation above represent forces (per unit volume). + In the log-linear plot. v1 = κ ln x2 + B (κ = 0.b and centerline velocity. Do the same exercise for the other three nodes. u1_fluct=u1-u1_mean. Let us investigate how these forces (the pressure gradient. y. node 1 and nj are located at the lower and upper wall. v1 at node 1. E. Find out how many samples must be used to get a correct mean value. How far out from the wall does the velocity proﬁle follow the linear law? At what x+ does it start to follow the log-law? 2 Compute the bulk velocity V1.mat does not include p.m which reads the data ﬁles. V1. Download the ﬁle uvw inst small.4).b = 1 2h 2h v1 dx2 ¯ 0 (E. Since the ﬂow is fully developed and two dimensional we get 0=− ′ ′ 1 ∂p ¯ ∂ 2 v1 ¯ ∂v1 v2 +ν 2 − ∂x ρ ∂x1 ∂x2 2 (E.24. There are nj = 98 nodes in the x2 direction.1) (recall that h denote half the channel width) What is the Reynolds number based on V1. Plot it both as linearlinear plot and a log-linear plot (cf. Your data are instantaneous.E. Compute and plot the three terms. see Eq.dat and the Matlab ﬁle pl vel.2) This equation is very similar to fully developed laminar ﬂow which you studied in Assignment 1. Mean ﬂow 202 umean=mean(u2) Here the number of samples is n = 5000 (the entire u2 array). respectively.) ¯ . (the ﬁle uvw inst small.3. the viscous term and the Reynolds stress term) affect ﬂuid particles located at different x2 locations.mat. Fig.3 Mean ﬂow All data in the data ﬁles below have been stored every 10th time step. 2 2 + + −1 and the log law. The data ﬁle includes v1 . set ¯ ∂ p/∂x = −1. Compute the mean velocity. v1 = x+ . respectively? E. use x+ for the wall distance. ′ Compute and plot also the instantaneous ﬂuctuations.4 The time-averaged momentum equation Let us time average the streamwise momentum equation. 3. for example. but now you are given the time history of all x2 nodes at one chosen x1 and x3 node.2). 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. 6. v2 and v3 from the same DNS as above.c . the difference is that we now have an additional Reynolds stress term.41 is the von K´ rm´ n constant and a a B = 5. Include the linear law.

can be decomposed into ¯ a mean and ﬂuctuation as φ = φ + φ′ .8 Production terms In order to understand why a stress is large. The root-mean-square (RMS) is then deﬁned as φrms = φ′2 1/2 (E. Which ones are zero (or close to)? Does any production term change sign at the centerline? If so. Pij . 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. Usually.5) . Wall shear stress 203 If a term is positive it means that it pushes the ﬂuid particle in the positive x1 direction. are largest). what about the sign of the corresponding shear stress plotted in Section E. This is probably the most common form of ﬂuid-solid interaction. E. They should be exactly equal. you can compute the ﬂuctuations as ′ vi = vi − vi ¯ (E. If heat transfer is involved.5 Wall shear stress Compute the wall shear stress at both walls.6? E. is large (there may be exceptions when other terms.7 Fluctuating wall shear stress In the same way as the velocity.5. the wall shear stress can be decomposed into a mean value and a ﬂuctuation.12. This is a measure of the ﬂuctuating tangential force on the wall due to turbulence.4) Compute the RMS of the wall shear stress.3 you computed the mean velocities. any ﬂuctuating variable. i.3) ′ ′ Now you can easily compute all stresses vi vj . Which shear stresses are zero? E.e. the ﬂuctuating temperature at the wall inducing ﬂuctuating heat transfer may be damaging to the material of the walls causing material fatigue. see Eq. a stress is large when its production term. In general.6 Resolved stresses In Section E.E. it is useful to look at its transport equation. Are they? E. Plot the production terms for all non-zero stresses across the entire channel. such as the diffusion term. Plot the normal stresses in one ﬁgure and the shear stresses in one ﬁgure (plot the stresses over the entire channel. From the instantaneous and the mean velocity. φ. 9.14) Πij = p′ ρ ′ ′ ∂vj ∂vi + ∂xj ∂xi (E.9 Pressure-strain terms The pressure-strain term reads (see Eq. Where is the Reynolds shear stress positive and where is it negative? E. 9. from x2 = 0 to x2 = 2h).

E. respectively. Using the velocities stored in uvw inst small. diss inst. E. x3 ).mat (see Section E. The data cover only the lower half of the channel.E. The dissipation in the K equation reads ∂¯i ∂¯i v v (E.8) – is small except very close to the wall (see Figs. x3 ).3.6.m which reads the data ﬁle. in which the pressure may vary unphysically in time (∂p/∂t does not appear in the equations).6) ′2 see the second line in Eq. we prefer to compute the velocity-pressure gradient term Πp = − ij ′ ′ vj ∂p′ vi ∂p′ − .7 and E. Compute and plot ′ ∂v ′ ∂vi ε=ν i (E. 8. 8.dat. 9.m which reads it. (x1 ± ∆x1 . Hence. E.mat and the Matlab ﬁle pl diss. Download the data ﬁle p inst small. The time histories of the pressure along ﬁve x2 lines [(x1 . k = vi vi /2 (Eq.5 and E.14).8.11 Do something fun! You have been provided with a lot of data which you have analyzed in many ways. K = vi vi /2 (Eq. Download the ﬁles y half. is transformation of turbulent kinetic energy into internal energy. x2 . Plot the pressure strain. are deﬁned in Eqs. the difference between Πp ij and Πij is small. Hence.mat and the Matlab ﬁle pl press strain. (x1 ±∆x1 . 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 ? E. ρ ∂xj ρ ∂xi (E. The pressure diffusion term in the v2 equation – which is the difference between Eqs.6 (the two ﬁrst terms in Eq.3). x3 ± ∆x3 )] are stored in this ﬁle.8) εmean = ν ∂xk ∂xk The ﬂow of kinetic energy between K. Compute and plot also εmean and P k . respectively? ij E. for the three normal stresses and the shear stress across ij the channel.10. x3 ±∆x3 )] so that you can compute all spatial derivatives. This allows you to compute all the three spatial derivatives of p′ . Πp . 9. x3 ) and (x1 . x2 . x3 ) and (x1 . Where is it largest? In which equation does this quantity appear? Let us now consider the equations for the mean kinetic energy. 8. Dissipation 204 Our data are obtained from incompressible simulations. The data ﬁle includes the time history of the velocities along ﬁve x2 lines [(x1 . Now think of some other way to analyze the data. increased temperature.35) ¯¯ ′ ′ and turbulent kinetic energy. For which stresses is it negative and positive? Why? Which term Πp is the largest source and sink term. x2 . x2 . x2 . you can compute all the terms in Eq.14. There are many interesting things yet to be analyzed! . k and ∆T is illustrated in Fig.e. i. x2 .2 and 9.3).7) ∂xk ∂xk see Eq. ε. ε and εmean . 8.10 Dissipation The physical meaning of dissipation.5 The dissipations. 9.

What is the deﬁnition of irrotational ﬂow? 9. σij . Explain the physical meaning of diagonal and off-diagonal components of Sij 7.html i.chalmers. σij . ˆ 5. that act on the Cartesian surfaces of a quadˆ rant (two dimensions). what are the boundary conditions in time (t) and space (x2 ). What is the physical meaning of irrotational ﬂow? 10. and the stress vector. The formula is nicely explained in Part 2 + vi ∂t ∂xi 3. Starting from the Navier-Stokes equations (see Formula sheet). Explain the physical meaning of Ωij 8. Show also the stress vector. ∂T ∂T ii. Part 1 describes the difference between Lagrangian and Eulerian points and velocities. TME225 Learning outcomes 205 F TME225 Learning outcomes TME225 Learning outcomes: week 1 1. Watch the on-line lecture Eulerian and Lagrangian Description. tn . Show that the product of a symmetric and an antisymmetric tensor is zero. i 6. tn . Derive the relation between the vorticity vector and the vorticity tensor 11. part 1 – 3 at http://www. Show which stress components.F. derive the ﬂow equation governing the Rayleigh problem expressed in f and η.se/˜lada/flow viz. Show the relation between the stress tensor. how are they expressed in the similarity variable η? . i 4.tfd. Explain the difference between Lagrangian and Eulerian description 2.

14).20. Derive the transport equation for temperature in incompressible ﬂow. Show that the divergence of the vorticity vector. 3.22 and 3. δ ∝ ℓ ν = Uℓ 1 (see also Re . 2. Show how the left side of the transport equations can be written on conservative and non-conservative form 12. Explain the ﬂow physics in a channel bend 5. 2. Use the diffusion of vorticity to show that Eq. i. Explain the physical meaning of the eigenvectors and the eigenvalues of the stress tensor 2. What is the physical meaning of the different terms? 10. is zero 15.7 7. 2. ωi . fully developed channel ﬂow (Eqs.23 18. Eq. Derive the ﬂow equations for fully developed ﬂow between two parallel plates. Derive the 2D equation transport equation for the vorticity vector from the 3D transport equation.5 6. TME225 Learning outcomes 206 TME225 Learning outcomes: week 3 1. 3.e. Eq. Explain vortex stretching and vortex tilting 16. 3. vi vi /2. Derive the incompressible Navier-Stokes equation. Derive the transport equation for the inner energy. The Navier-Stokes equation can be re-written on the form ∂vi + ∂t ∂k ∂xi no rotation − εijk vj ωk = − rotation 1 ∂p ∂ 2 vi +ν + fi ρ ∂xi ∂xj ∂xj Derive the transport equation (3D) for the vorticity vector.12. Eq.15 9. 2.18. 3. Eq. Explain the ﬂow physics at the entrance (smooth curved walls) to a plane channel 4. 4.21 14. Show how the boundary layer thickness can be estimated from the Rayleigh problem using f and η (Fig. Eq. 2. Derive the transport equation for kinetic energy. Eq.3) 3. Eq. Show that the vortex stretching/tilting term is zero in two-dimensional ﬂow 17. Explain the energy transfer between kinetic energy and inner energy 11. u. 4.F. Show the similarities between the vorticity and temperature transport equations in fully developed ﬂow between two parallel plates 19.26) 13. What is the physical meaning of the different terms? 8. Derive the Navier-Stokes equation.

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. The second part of the movie deals with turbulent ﬂow: we’ll talk about that in the next lecture (and the remaining ones). wall but not in the latter.F. Watch the on-line lecture Boundary layers parts 1 at http://www.1) 21. the vorticity) created? Where is the vorticity 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 ∂x1 ∂x2 = 0. What do we mean by a “separated boundary layer”? How large is the wall shear stress at the separation point? viii.3]) vi.e. Where is the circulation (i. How does the boundary layer thickness change when we increase the velocity? Explain why? ii. this has an implication for γ2. Consider the ﬂow over the ﬂat plate. 1.3. How does the boundary layer thickness change when we move downstream? ii. 4. . Is the wall shear stress lower or higher after the contraction? Why? v.chalmers.tfd.se/˜lada/flow viz. with Eq.html i. How do they estimate the boundary layer thickness? (cf. Γ.chalmers. What happens when the angle of the diffuser increases? vii. deﬁned? (cf. How is the circulation.tfd. 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.html i. TME225 Learning outcomes 207 20.19) 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. How is the wall shear stress deﬁned? How does it change when we move downstream? (how does this compare with the channel ﬂow in TME075 Assignment 1? iv. Watch the on-line lecture Boundary layers part 2 at http://www. Section.se/˜lada/flow viz.wall [see Section 4. Consider the ﬂow in a divergent channel (a diffuser): what happens with the boundary layer thickness and the wall shear stress? vi.

tfd. τη . Write the vortex stretching/tilting term in tensor notation. Why? vi. Try to explain why separation is delayed.tfd.se/˜lada/flow viz. 9. Explain the cascade process. Describe the cascade process created by vorticity. When the ﬂow along the lower wall of the diffuser is tripped into turbulent ﬂow.chalmers. E(κ). computed? Use dimensional analysis to derive the −5/3 Kolmogorov law. ℓη and the time scale. How large are the largest scales? What is dissipation? 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? 4. What are the Kolmogorov scales? Use dimensional analysis to derive the expression for the velocity scale. What is the difference in the two velocity proﬁles? Explain the differences. How is the energy transfer from eddy-to-eddy. iv. estimated? Show how the ratio of the large eddies to the dissipative eddies depend on the Reynolds number. Consider the airfoil: when the boundary layer on the upper (suction) side is turbulent. TME225 Learning outcomes 208 TME225 Learning outcomes: week 4 1. What is its physical meaning? Describe the physical process of vortex stretching which creates smaller and smaller eddies. 5. Given the energy spectrum. vη .html i. Try to explain why. What does isotropic turbulence mean? 7. Watch the on-line lecture Turbulence part 1 at http://www. iii. Make a ﬁgure of the energy spectrum.chalmers. What is a turbulent eddy? 3.F. the separation region is suppressed. Show and discuss the family tree of turbulence eddies and their vorticity. 8. Two boundary layers – one on each side of the plate – are shown. The upper one is turbulent and the lower one is laminar. The energy spectrum consists of thee subregions: which? describe their characteristics. 2. Why does the irregular motion of wave on the sea not qualify as turbulence? .html i.se/˜lada/flow viz. What characterizes turbulence? Explain the characteristics. k. 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. Watch the on-line lecture Boundary layers parts 2 & 3 at http://www. Show that in 2D ﬂow the vortex stretching/tilting term vanishes. how is the turbulent kinetic energy. the length scale. εκ . Vortex generator are place on the suction side in order prevent or delay separation. stall occurs at a higher angle of incidence compared when the boundary layer is laminar. Show the ﬂow of turbulent kinetic energy in the energy spectrum. How? ii. 6. The ﬂow is “tripped” into turbulence.

The total shear stress consists of a viscous and turbulent shear stress: show how they vary across the channel (show also a zoom near the wall). Make a ﬁgure and show where these regions are valid (Fig.tfd. TME225 Learning outcomes 209 ii. 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 8 000.2) 13.e. 6. buffer layer or log-layer) does the viscous stress dominate? In which region is the turbulent . v2 ) ﬂuctuations. vii. buffer layer and log-layer. How? v. viii. 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. Use the decomposition vi = vi + vi to derive the time-averaged Navier-Stokes ¯ equation.F. A new terms appears: what is it called? Simplify the time-averaged Navier-Stokes equation for boundary layer.chalmers. Why? The viscosity is further decreased. Consider fully developed channel ﬂow. the dye does not mix with the water. 10. In turbulent ﬂow.se/˜lada/flow viz.e. They look very similar in one way and very different in another way. ¯ 2 12. and the pressure drop (i. ′ 11. The inner region is furthermore divided into a viscous sublayer. Draw a laminar and turbulent velocity proﬁle for pipe ﬂow. and the pressure drop increases. v. Watch the on-line lecture Turbulence part 2 at http://www. How is the turbulence syndrome deﬁned? iii. The two turbulent jet ﬂows have the same energy input and hence the same dissipation.html i. one at low Reynolds number and one at high Reynolds number. What is the ∂¯1 v main difference? In which ﬂow is the wall shear stress τw = µ largest. For laminar ﬂow. uτ . 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. the resistance) decreases. The movie laminar shows ﬂow in a pipe. iv. Discuss the connection between mixing and the cross-stream (i. Explain the analogy of a water wall and the cascade process. In which region (viscous sublayer. Try to explain the increased pressure drop in turbulent ﬂow with the increased mixing. deﬁned? Deﬁne x+ and v + . The wall region is divided into an inner and outer region. Two turbulent jet ﬂows are shown. try to explain how turbulence creates a Reynolds shear stress. Dye is introduced into the pipe. The viscosity is decreased. ∂x2 laminar to turbulent? ii. Why? How does the water-ﬂow coming out of the pipe change due to the second decrease of viscosity? iv. vi. What is the total shear stress? How is the friction velocity. the velocity near the wall is larger than in laminar ﬂow. in turbulent ﬂow it does. Why? ′ iii.

TME225 Learning outcomes 210 shear stress large? Integrate the boundary layer equations and show that the total shear stress varies as 1 − x2 /δ. What are the relevant velocity and length scales in the viscous-dominated region? Derive the linear velocity law in this region (Eq. What are the suitable velocity and length scales in the inertial region? Derive the log-law.17). 6. . 14.F.

65. the situation is reversed: hot ﬂuid (dark blue) is moving on top of light ﬂuid (yellow). Where is the production term. Even though the Reynolds number may be large. TME225 Learning outcomes 211 TME225 Learning outcomes: week 5 July 7. how does the ratio of the large scales to the small scales change when the Reynolds number increases (see Eq. Which terms do only transport k? 3. Study how the two ﬂuids mix downstream of the partition. The ﬁlm says that there is a similarity of the small scales in a channel ﬂow and in a jet ﬂow. Show this. the k equation is dominated by two terms. . In the cascade process. show how −v1 v2 and ∂¯1 /∂x2 vary near the wall.chalmers.e.html i. After that.F.14) iii. P k = −v1 v2 ∂¯1 /∂x2 . 5. In the last example.se/˜lada/flow viz.18 at p. there are a couple of physical phenomena which may inhibit turbulence and keep the ﬂow laminar: mention three. What do they mean? ii. In fully developed channel ﬂow. Watch the on-line lecture Turbulence part 3 at http://www. τκ = ℓκ ℓ0 eddies is the production largest? Why? −2/3 4. Given the exact k equation. Make a ﬁgure and show how the velocity and shear stress vary across the channel. see Eq. v. Discuss the physical meaning of the different terms in the k equation. iv. the time-averaged Navier-Stokes consists only of three terms. Derive the exact transport equation for turbulent kinetic energy. their gradients plus the pressure gradient) vary across the channel. the ﬂuid on the top is hot (yellow) and the one at the bottom (dark blue) is cold: how do the ﬂuids mix downstream of the partition.tfd. All spatial derivatives are kept in the dissipation term: why? In the turbulent region of the boundary layer. 8. largest? In order to explain ′ ′ v this. How is the mixing affected? This ﬂow situation is called unstable stratiﬁcation. better or worse than in the previous example? This ﬂow situation is called stable stratiﬁcation. What happens with the small scales when the Reynolds number is increased? What happens with the large scales? Hence. we assume that the dissipation is largest at the smallest −1 τ0 . 5. i. Consider ﬂow in a channel where the ﬂuid on the top (red) and the bottom (yellow) are separated by a horizontal partition. which scales dies ﬁrst? The scenes of the clouds show this in a nice way. In the next example. Which ones? Which terms are non-zero at the wall? ′ ′ v 6. Compare in meteorology where heating of the ground may cause unstable stratiﬁcation or when inversion causes stable stratiﬁcation. Which two terms balance each other in the outer region? Which terms drives (“pushes”) the ﬂow in the x1 direction? Which two terms are large in the inner region? Which term drives the ﬂow? 2. In decaying turbulence. 2012 1. For which scales. k. show how the three terms (i.e. give the equation for boundary-layer ﬂow.

F. How is homogeneous turbulence deﬁned? . Give an example of how they are combined in non-homogeneous turbulence. Discuss the difference of spatial transport of k and spectral transfer of k. TME225 Learning outcomes 212 7.

′ ′ 5. Consider the fully turbulent region in fully developed channel ﬂow: which are ′2 ′2 ′2 ′ ′ the main source and sink terms in the v1 . 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. what sign will the pressure-strain term have for ′2 normal stresses. 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? Why? Show where they appear in the energy spectrum. respectively? Why is there no pressurestrain term in the k equation? 8. larger and smaller than vav ? What role does Π12 has? What sign? Why do we call the pressure-strain term the Robin Hood term? 6. vi vj . ′ ′ 2. The mean normal stress ′ ′ ′2 can be deﬁned as vav = vi vi /3. Is this consistent with the sign of P12 ? . respectively. Consider fully developed channel ﬂow: how are the expressions for the production terms simpliﬁed? Which production terms are zero and non-zero. Discuss the physical meaning of the different terms in the vi vj equation. TME225 Learning outcomes 213 TME225 Learning outcomes: week 6 July 7. K. respec′ ′ tively? Consider the production term for v1 v2 : which sign does it have in the lower and upper part of the channel. 3. v3 and v1 v2 equations? Which terms are the largest terms at the wall? Which terms are zero at the wall? ′ ′ 9.F. ε12 . 2012 1. Consider the pressure-strain term in the vi vj equation. respectively. Show that the role of the convection and diffusion terms is purely to transport the quantiy (k for example) and that they give no net effect except at the boundaries (use the Gauss divergence theorem) ′ ′ 4. v2 . Derive the exact transport equation for turbulent Reynolds stress. Derive the exact transport equation for mean kinetic energy. Discuss the physical meaning of the different terms. Take the ′ ′ trace of the vi vj equation to obtain the k equation. for the shear stress: how large is it? 7. Consider the dissipation term.

.e. Explain why.html i. Since there is atmospheric pressure at the outlet. There is a pressure drop.2. i. The pressure difference in the contraction region and the outlet increases. Fluid particles become thinner and elongated in the contraction. Why? iii. What happens with the pressure drop when there is a separation in the diffuser? vii. Compare this relation with the three-dimensional form of sure gradient ds Navier-Stokes equations for incompressible ﬂow.se/˜lada/flow viz.chalmers.F. Vs ds dp .html i.1). The bulk velocities at the inlet and outlet are equal. dVs . Why? ii. What happens with the velocity and pressure. What is the static pressure? How can it be measured? What is the difference between the stagnation and the static pressures? v. Watch the on-line lecture Pressure ﬁeld and acceleration part 1 at http://www. What is the Coanda effect? ii. This is called cavitation (this causes large damages in water turbines). this means that the pressure in the contraction region must decrease as we increase the velocity of the water. They increase the speed in the venturi meter. TME225 Learning outcomes 214 TME225 Learning outcomes: just for fun! 1.se/˜lada/flow viz. What is the stagnation pressure? How large is the velocity at a stagnation point? iv. Watch the on-line lecture Pressure ﬁeld and acceleration part 3 at http://www.chalmers. They show that the acceleration along s.tfd.7 2. How large is the pressure at the surface of the pipe relative to the surrounding pressure. although the water temperature may be around 10o C. The bleeders (outlets) are opened. Why? vi. Eq. 2. is related to the presiii. iii.tfd. Watch the on-line lecture Pressure ﬁeld and acceleration part 2 at http://www. The pressure decreases slowly downstream. 3.tfd.e. Finally the water starts to boil.se/˜lada/flow viz. Water ﬂow in a manifold (a pipe with many outlets) is presented. The pressure now increases in the downstream direction. Explain how suction can be created by blowing in a pipe. viii.chalmers. a diffuser). The water from the tap hitting horizontal pipe follows attaches to the surface of the pipe because of the Coanda effect. Why? ii. Section 3. Explain the relation between streamline curvature and pressure (cf.html i. A venturi meter is a pipe that consists of a contraction and an expansion (i. but still the pressure at the outlet is lower than that at the inlet. The water ﬂow goes through the contraction.

. TME225 Learning outcomes 215 iv.F. there is an adverse pressure gradient (∂p/∂x > 0). At the end of the contraction. Explain why.

1) Since Eq. 11. aijkℓ is symmetric with respect to index j and ℓ.e. MTF270: Some properties of the pressure-strain term 216 G MTF270: Some properties of the pressure-strain term In this Appendix we will investigate some properties of aijkℓ in Eq.3) V ′ ′ where the boundary integrals have been omitted.2) Green’s third formula (it is derived from Gauss divergence law) reads ϕ(x) = − 1 4π ∇2 ϕ dy3 |y − x| (G.73. .1 in the deﬁnition of aijkℓ in Eq. 11. aijkℓ = aiℓkj (G.73 at p. 97. 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 ∂Bℓj (−r) ∂ 2 Bjℓ (r) ∂ 2 Bℓj (−r) ∂Bjℓ (r) =− ⇒ = ∂ri ∂ri ∂rk ∂ri ∂rk ∂ri (G.73 is integrated over r3 covering both ′ ′ r and −r (recall that vℓ and vj are separated by r). G. i.3 gives ′ ′ vj vℓ = − 1 4π V ′ ′ ∂ 2 vℓ vj dy3 1 = aijiℓ ∂xi ∂xi |y − x| 2 (G.4) where the last equality is given by Equation 11. Setting ϕ = vℓ vj in Eq. G.G.

e. if the cross term is neglected. the Leonard stresses must not be computed explicitly. cr = 1 (see Eq. i. MTF270: Galilean invariance 217 H MTF270: Galilean invariance In [64] he found that the Leonard term Lij and the cross term Cij are not Galilean invariant by themselves. 159].e. As a consequence. but only the sum Lij + Cij is. t) k = ∗ + ∂t ∂t∗ = Vk ∂x∗ + ∂t∗ . 17.1) (H. and thus the terms are not Galilean invariant. This is shown by transforming both the exact . x∗ )-coordinate system gives i ∂φ ∂φ ∂φ ∂φ ∂φ ∗ = ∗ + Vk ∗ + (vk − Vk ) ∗ + vk ∂t ∂xk ∂t ∂xk ∂xk ∂φ ∂φ ∗ = ∗ + vk ∗ . x∗ = xk + Vk t. Below we repeat some of the details of the derivation given in [64]. t) ∂t∗ ∂φ ∂φ j = + = ∗ ∗ ∂xk ∂xk ∂xj ∂xk ∂t ∂x∗ k ∂φ ∂x∗ ∂φ ∂t∗ ∂φ ∂φ ∂φ(xi . i. ∂t ∂xk It shows that the left hand side does not depend on whether the coordinate system moves or not. let’s look at the Leonard term and the cross term.38).3) ∗ ′′ ′′ ¯∗ ′′∗ ¯∗ ′′∗ v v Cij = vi vj + vj vi = (¯i + Vi )vj + (¯j + Vj )vi = ′′ ′′ ′′ ′′ = vi vj + vj Vi + vj vi + vi Vj = Cij + vj Vi + vi Vj ¯ ′′ ¯ ′′ From Eq. Let’s denote the moving coordinate system by ∗. Now. vk = vk + Vk ¯∗ ¯ k By differentiating a variable φ = φ(t∗ . H. (H. i. ∂t ∂t ∂xk k (H. we have vk = vk + Vk and consequently also vk = vk .2 is it easy to show that the Navier-Stokes (both with and without ﬁlter) is Galilean invariant [64. Galilean invariance means that the equations do not change if the coordinate system is moving with a constant speed Vk . However. note that the sum is.e. because then the modelled momentum equations do not satisfy Galilean invariance. t∗ = t.H. it is Galilean invariant. H. xi )coordinate system to the (t∗ . Since the ﬁltering operation ′′∗ ′′ ¯∗ ¯ is Galilean invariant [64]. Transforming the material derivative from the (t.4) ij The requirement for the Bardina model to be Galilean invariant is that the constant must be one.3 we ﬁnd that the Leonard term and the cross term are different in the two coordinate systems.2) From Eq. x∗ ) we get i ∂x∗ ∂φ ∂φ(xi . ∗ L∗ + Cij = Lij + Cij . For ¯ ¯ the Leonard and the cross term we get (note that since Vi is constant Vi = Vi = Vi ) ¯∗ ¯∗ ¯∗ ¯∗ v v v v L∗ = vi vj − vi vj = (¯i + Vi )(¯j + Vj ) − (¯i + Vi )(¯j + Vj ) ij ¯¯ ¯ ¯ ¯¯ ¯ ¯ = vi vj + vi Vj + vj Vi − vi vj − vi Vj − Vi vj ¯¯ ¯¯ v ¯ v ¯ = vi vj − vi vj + Vj (¯i − vi ) + Vi (¯j − vj ) ′′ ′′ = Lij − Vj vi − Vi vj (H.

4). The exact form of Cij transforms as in Eq. Cij (see Eq. but here this does not matter. .37). because provided cr = 1 the M modelled stress. Cij . (H. 17. Cij (i.e. MTF270: Galilean invariance 218 M Cij (Eq.36) and the modelled one. 17.3. Eq. H. we have Cij + L∗ = Cij + Lij . Note that in order ij to make the Bardina model Galilean invariant the Leonard stress must be computed explicitly. transforms in the same way as the exact one.H. Cij = Cij . Cij . The Bardina term transforms as ∗M Cij = cr (¯i vj − v i v j ) v ∗ ¯∗ ¯ ¯ ∗ ∗ v v = cr (¯i + Vi )(¯j + Vj ) − (¯i + Vi )(¯j + Vj ) v v M = Cij + cr v ′′ i Vj + v ′′ j Vi . H. as for the ∗M M exact stress.5) = cr [¯i vj − vi vj − (¯i − vi )Vj − (¯j − vj )Vi ] v¯ ¯¯ v ¯ v ¯ ∗M M As is seen. Thus.

1. I.1 and 24. κn . κn j x3 dAi θn ϕ x1 Figure I. αn and θn (see Figs. I.1.1 The wavenumber vector.I. Compute the wavenumber vector. They are chosen so as to give a uniform distribution over a spherical shell of the direction of the wavenumber vector.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. i. I.1) and random phase ψ n . MTF270: Computation of wavenumber vector and angles 219 I MTF270: Computation of wavenumber vector and angles For each mode n.1) κn i n x2 Table I. κn = sin(θn ) cos(ϕn ) 1 κn = sin(θn ) sin(ϕn ) 2 κn = cos(θn ) 3 p(ϕn ) = 1/(2π) p(ψ n ) = 1/(2π) p(θn ) = 1/2 sin(θ) p(αn ) = 1/(2π) 0 ≤ ϕn ≤ 2π 0 ≤ ψ n ≤ 2π 0 ≤ θn ≤ π 0 ≤ αn ≤ 2π (I. see Fig. The probability distributions are given in Table I. .1. create random angles ϕn .e. using the angles in Section I according to j Fig. I.1: Probability distributions of the random variables.

0. 1.1 and I. 0) (1. σi . 0) (−1. 0) n Table I. 0. σi will lie in a plane i n normal to the vector κi .3) = − sin(θ ) cos(α ) n n n n n The direction of σi in this plane (the ξ1 − ξ2 plane) is randomly chosen through αn . 0) (0. i n I. 0) (0. 0. σi and αn from Eqs. Table I. and κn are orthogonal.2: Examples of value of κn . 0) (0. 24. σi κn = 0 (superscript n denotes Fourier mode n). 24.2. −1) (−1. −1) (0. 0) (0. I. . see Fig. 1. 0. 1) (0. 0. Hence.n I. This can be seen j by taking the divergence of Eq. 0. 1) n σi (0.2 gives the direction of the two vectors in the case that κi is along one coordinate direction and α = 0 and α = 90o . 0.2) n n i. 1.e.2 Unit vector σi n Continuity requires that the unit vector.3. 1.3 which gives N ∇ · v′ = 2 n=1 un cos(κn · x + ψ n )σ n · κn ˆ (I. 0.1. 0) αn 0 90 0 90 0 90 (0. This gives n σ1 = cos(ϕn ) cos(θn ) cos(αn ) − sin(ϕn ) sin(αn ) n σ2 = sin(ϕn ) cos(θn ) cos(αn ) + cos(ϕn ) sin(αn ) n σ3 (I. Unit vector σi 220 κn i (1.

κ2 or κ3 ). They are formed by integrating over a wavenumber plane. The complex Fourier transform exp(ıκm xm ) is deﬁned in Appendix C.6) Bij (ˆ1 ) cos(κ1 x1 )dˆ1 x ˆ x 0 ′2 The Reynolds stress ρv1 . J. The one-dimensional two-point correlation. The v1 can be computed both from the three- .e. which are a function of scalar wavenumber.5) 2 2π Bij (ˆ1 ) exp(−ıκ1 x1 )dˆ1 x ˆ x −∞ where Eij is twice the Fourier transform of Bij because of the factor two in Eq. Hence. x2 . i. −∞ < κ2 < ∞ and −∞ < κ3 < +∞.4) (J.8 and J.J. The diagonal components of the two-point correlation tensor are real and symmetric and hence the antisymmetric part of exp(−ıκ1 x1 ) – i. J. ˆ respectively. 17. Bij (ˆ1 ). and the one-dimensional x spectrum. form a Fourier-transform pair.8) +∞ Bij (x1 . 10. Bij .3) 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. Note that the maximum magnitude of the wavenumber vector contributing to Eij (κ1 ) is very large since it includes all κ2 and κ3 . ˆ The two-point correlation. the sinus part – is zero and ˆ Eqs.10. xm .e.3. for example. see Eqs. Eq. MTF270: 1D and 3D energy spectra 221 J MTF270: 1D and 3D energy spectra ′ ′ The general two-point correlation Bij of vi and vj (see Eq. i. the energy spectrum for the wavenumber κ1 . for example.1) where xm and κm are the separation vector the two points and the wavenumber vector. x3 )) = −∞ Ψij (κ) exp(ıκm xm )dκ1 dκ2 dκ3 ˆ (J. and the energy spectrum tensor. form a Fouriertransform pair Ψij (κ) = 1 (2π)3 +∞ Bij (ˆ) exp(−ıκm rm )dˆ1 dˆ2 dˆ3 x x x x −∞ (J. J.2) can be expressed by the energy spectrum tensor as [57. κ (κ1 . reads Eij (κ1 ) = 1 2 +∞ Ψij (κ)dκ2 dκ3 −∞ (J. one-dimensional energy spectra.e. for example. Eij (κ1 ).5 are simpliﬁed as Bij (ˆ1 ) = x 1 2 +∞ ∞ Eij (κ1 ) cos(κ1 x1 )dκ1 = ˆ −∞ +∞ −∞ 0 Eij (κ1 ) cos(κ1 x1 )dκ1 ˆ +∞ 1 Eij (κ1 ) = π 2 Bij (ˆ1 ) cos(κ1 x1 )dˆ1 = x ˆ x π (J.4 and J. is equal to the two-point correlation tensor ′2 ρBij with with zero separation distance. Eij (κ). Chapter 3] (cf. Both in experiments and in LES it is usually sufﬁcient to study the twoˆ point correlation and the energy spectra along a line.2) The separation between the two points is described by a general three-dimensional vector. Bij (ˆ1 ) = x Eij (κ1 ) = 1 2 +∞ Eij (κ1 ) exp(ıκ1 x1 )dκ1 ˆ −∞ +∞ (J. Ψij . are often used.

Eq. J. (J. i. k= 1 2 ∞ 4πκ2 Ψii dκ 0 (J. E11 (κ1 ) = v1 (κ1 ) ˆ2 1 2 v (κ) + v3 (κ) + v3 (κ) ˆ ˆ2 ˆ2 E(κ) = 2 1 Below the properties of the three energy spectra are summarized.7) = B11 (0) = 0 E11 (κ1 )dκ1 ′ ′ Hence the turbulent kinetic energy. The proof is given in this section.6) +∞ ′2 v1 = B11 (x1 . E(κ) = 4πκ2 Ψii so that κ k= 0 E(κ)dκ (J. see Parseval’s formula. • The three-dimensional spectrum tensor.5 and J. 0) = ∞ ′2 v1 Ψii (κ)dκ1 dκ2 dκ3 −∞ (J. k.12) J. an be written as k= k= 1 2 ∞ 1 2 +∞ Ψii (κ)dκ1 dκ2 dκ3 −∞ (J. The energy spectra E11 (κ1 ) and E(κ). Ψij (κ). J. |κ| ≡ κ. Eq. E(κ). is a tensor which is a function of a scalar (one component of κm ).e.9) The integral in Eq.10) where 4πκ2 is the surface area of the shell. Instead of integrating over dκ1 dκ2 dκ3 we can integrate over a shell with radius κ and letting the radius go from zero to inﬁnity. Eij (κ1 ). J.6 we stated that the one-dimensional energy spectra and the two-point correlations form Fourier-transform pairs. correspond to the square of the Fourier coefﬁcient of the velocity ﬂuctuation (see Parseval’s formula. C.1 Energy spectra from two-point correlations In connection to Eqs.4. J. J. We now deﬁne an energy spectrum. for example. • The energy spectrum. i.4.8) ∞ E11 (κ1 )κ1 + 0 1 2 ∞ E22 (κ2 )κ2 + 0 1 2 E33 (κ3 )dκ3 0 (J.8 has no directional dependence: it results in a scalar.11) where E(κ) = 2πκ2 Ψii (κ).J. • The one-dimensional spectrum. C.4).1 ) and one-dimensional spectrum (Eq. k = vi vi /2.1. 0.e. Let u be the Fourier coefﬁcient of the velocity ﬂuctuation ˆ . is a scalar which is a function of the length of the wavenumber vector. Energy spectra from two-point correlations 222 dimensional spectrum (Eq. is a tensor which is a function of the wavenumber vector. The energy spectrum is given by the square of the Fourier coefﬁcients.

5. all averaged turbulence quantities are independent of x and the two-point correlation is not dependent on x (or x′ ) but only on the separation distance x − x′ .15 for κ = κ′ is evaluated as (see “length of of ψk ” in Appendix C. The remaining integral includes trigonometric function with wavelengths κ and κ′ .14) The second integral (in parenthesis) is the Fourier transform of the two-point correlation B11 . it is seen that the Fourier transform of a two-point correlation (in this example B11 (x1 ) ) indeed gives the corresponding one-dimensional energy spectrum (in this example E11 (κ1 ) = (ˆ(κ))2 ). J.15 can now be written 4πx L −L ˆ u(κ)ˆ(κ) = B11 (x) ˆ u (J.12) and since it does not depends ˆ on the spatial coordinate it has been moved out of the integral. Hence we replace x′ by y + x′′ so that u(κ)ˆ(κ′ ) = ˆ u = 1 L2 1 L L L −L L−x u(x)u(x + x′′ ) exp(−ı(κx + κ′ (x + x′′ ))dx′′ −L−x dx dx exp(−ı(κ + κ′ )x) −L 1 L L−x B11 (x′′ ) exp(−ıκ′ x′′ ))dx′′ −L−x (J. u . and use ψ1 = cos(2πx/L)) L (ψ1 |ψ1 ) = ||ψ1 ||2 = cos2 −L 2πx L L dx (J. i. Energy spectra from two-point correlations 223 u′ in the x direction which is periodic with period L. C.e.16) =L x L = + sin 2 8π Equation J.e. u(κ)ˆ(κ′ ) = ˆ u 1 ˆ B11 (x′′ ) L L exp(−ı(κ + κ′ )x))dx −L (J.15) ˆ where B11 denotes the Fourier transform of B11 (cf.4 except the factor of two.17) Hence.13. 10. J.J. J.1. Take the covariance of the Fourier coefﬁcients. this equation corresponds to Eq. Furthermore. see Appendix C. They are orthogonal functions.13) u(x)u(x ) exp(−ı(κx + κ x )dxdx ′ ′ ′ −L where · denotes averaging over time. i. Eq. u(κ′ ) and u(κ) where κ and κ′ denote two different wavenumbers and x ˆ ˆ and x′ denote two points separated in the x directions so that u(κ)ˆ(κ′ ) = ˆ u 1 L L u(x) exp(−ıκx)dx −L 1 L ′ L u(x′ ) exp(−ıκ′ x′ )dx′ −L 1 = 2 L L −L L (J. and the integral of these functions is zero unless κ = κ′ . Since we are performing a Fourier transform in x we must assume that this direction is homogeneous. This integral in Eq. B11 is real and symmetric since B11 is real and symmetric. see discussion in connection to Eq.

It is available on Linux. should be submitted electronically at the Student Portal www.m which computes the gradients.44 · 10−5 .miktex. Assignment 1: Reynolds averaged Navier-Stokes 224 K MTF270.e. for example. At the wwwpage you can download a M-ﬁle (pl vect. v2 .org) which are both free to download. ρ = 1) based on the bulk velocity bulk velocity in the channel and the height of the hill at the lower wall.2 Analysis Study the ﬂow.1 Two-dimensional ﬂow You can do the assignment on your own or in a group of two. The work should be carried out in groups of two (if you want to work on you own that is also possible) . Octave is a Matlab clone which can be downloaded for free.portal.tfd. v2 . The report.1) Recall that all the terms on the right-hand side represent x components of forces per unit volume. ρ = 1) based on the bulk velocity in the channel and the width of the the channel. The two-dimensional time-averaged Navier-Stokes for the x1 momentum reads (the density is set to one. Although some of the data are probably not fully accurate.lyx. Assignment 1: Reynolds averaged NavierStokes K. ρ = 1) based on the bulk velocity in the channel and the hill height. v ¯ ′2 ′2 ′2 ′ ′ p) and turbulent quantities.org) or MikTex (www. Use Matlab or Octave to read data ﬁles of the mean ﬂow (¯1 .student. i.chalmers. You’ll use data from a coarse DNS. On Windows you can use. . Case 3: Flow in a diverging/converging section Re = 18 000 (ν = 5. You must also download the function dphidx dy.se/˜lada/comp turb model. It is recommended (but A A the not required) that you use LTEX(an example of how to write in LTEXis available on the course www page). Re = 10 595 (ν = 9. the deadline can be found at the Student Portal.K. K. You can also use Octave on Linux/Ubuntu. Re = 10 595 (ν = 9. v1 v2 .m) which reads the data and plots the vector ﬁeld and the pressure contours. Case 2: Flow over two small hills.56 · 10−5 . Contact the teacher to get a Case No. ρ = 1) ′2 ′ ′ ∂¯1 v2 v ¯ ∂p ¯ ∂ 2 v1 ¯ ∂v1 ∂ 2 v1 ¯ ∂v1 v2 ∂¯1 v1 v ¯ + =− +ν 2 − ∂x + ν ∂x2 − ∂x ∂x1 ∂x2 ∂x1 ∂x1 1 2 2 (K. You will analyze one of the ¯ following ﬂows: Case 1: Flow over a 2D hill. Make sure you put this function in the directory where you execute pl vect. in this exercise we consider the data to be exact. Download the data from http://www. v3 . along with the Matlab ﬁles(s). Lyx (www. (v1 . You will use Matlab.m. In which regions do you expect the turbulence to be important? Now let’s ﬁnd out.44 · 10−5 . Periodic boundary conditions are imposed in streamwise (x1 ) and spanwise (x3 ) directions in all ﬂows.se. and ε). MTF270.

1 The momentum equations ′2 The ﬁle pl vect.[20]) ylabel(’y/H’. Make also a zoom near the walls. You will need to compute the derivatives of e.m loads the data ﬁle and plots the proﬁles of v1 at some x stations.1 0.2.y. Plot the stresses along vertical grid lines at these two locations using the Matlab command plot(x.1. K. 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 show that on paper).− 1 2 ∂x2 ∂x1 (K.3) .m is used to compute ∂¯1 /∂x1 and ∂¯1 /∂x2 . Assignment 1. Analysis 225 K. The database corresponds to ¯ a two-dimensional ﬂow.g.4 0 0. Now let’s think of the forces as vectors. The gradient of the normal stresses in the x1 − x2 plane represent the force vector FN = − ′2 ∂v1 ∂v ′2 .− 2 ∂x1 ∂x2 (K.4].[20]) title(’velocity’. you can use the command h1=gca. Plot also all terms in Eq. set(h1. v1 and p.01] and u=[-0.2.2.m the function dphidx dy. with ’surf’ and ’quiver’.2) and the corresponding force vector due to the shear stresses reads FS = − ′ ′ ∂v1 v2 ∂v ′ v ′ . Find two (or more) x1 ′2 locations (vertical grid lines) where the v1 stress is large and small.’fontsize’. the velocity vector ﬁeld and a contour plot of velocity gradient ∂¯1 /∂x2 .1.1. In pl vect. K.[20]) Assignment 1. Please make sure that in your report the numbering on the axis and the text in the legend is large enough. Compute v ¯ ¯ all terms in Eq. for a x − y plot plot(u. One way to ﬁnd these locations is to use the Matlab surf command.01]) The ’axis’ command can be used together with any plot.’fontsize’. For example.[20]) %the number ’20’ gives the fontsize The size of the labels and the title is similarly controlled by xlabel(’x/H’. v v Use this function to compute all derivatives that you need.K. respectively.g.y).’fontsize’.’fontsize’.’linew’. e. To enhance readability you may omit the small terms or use two plots per vertical grid line. So far we have looked at the v1 -momentum equation. this is achieved by axis([-0.1 0.2) % linewidth=2 you may want to zoom in on y=[0 0.

1))2 + (y(I. The production term consists of the sum of four terms. P k ≃ ε) anywhere? K. the lower wall is the closest wall to node (I. J) − y(I. f (see Eq. In the damping function. K. again.4) Assignment 1. 1.3). Plot the production term along the two grid lines. reads ∂2k ∂ (¯j k) = ν v + Pk + Dk − ε ∂xj ∂xj ∂xj v ′ ′ ∂¯i Pk = −vi vj ∂xj (K. Zoom up in interesting regions. J) and that the lower wall is horizontal.86. Zoom-in on interesting regions. that the lower wall is the closest wall to cell (I. J). Analysis 226 Identify the ﬁrst term in Eqs.w = (0.87.92. K. If we assume.3 The Reynolds stress equations The modelled transport equation for the Reynolds stresses can be written as ′ ′ ∂ 2 vi vj ∂ vk vi vj = ν ¯ ′ ′ + Pij + Φij + Dij − εij ∂xk ∂xk ∂xk v v ′ ′ ∂¯j − v ′ v ′ ∂¯i Pij = −vi vk j k ∂xk ∂xk (K.w for the general case. 11. the pressure gradient and the viscous terms.3.91 and 11. Plot the vector ﬁeld τ N to ﬁnd out some features.1.2. Assignment 1. and compute then ni. k.w from si. need to be modelled. Write out the momentume equation also for v2 and identify the other two terms in ¯ Eqs. two of which involve the shear stress while the other include the normal stresses.2 The turbulent kinetic energy equation The exact transport equation for for the turbulent kinetic energy. xn denotes the distance to the nearest wall. τ S (see Eq.11) .K. Dij . K. ni. 11. si. 11.88). Here we use the models in Eqs. 11.5) The pressure-strain term. 11. Eq.4. then ni. J) − x(I.6. Do you have local equilibrium (i. what happens with the vector ﬁeld? Zoom-in on interesting regions.2. and the diffusion term. Φij . If.2. Plot also vector ﬁelds of the shear stress. K. Assignment 1.5.w denotes the unit normal vector of the wall to which the distance xn is computed.6) 2.w . 1). then xn = (x(I.2 and K. Assignment 1.3 in the momentum equation for v1 .53. Note that FN and FS are forces per unit volume ([N/m3 ]).3. 1))2 1/2 (K.e. compute ﬁrst the vector which is parallel to the wall. for example. Explain why it is large at some locations and small at others. K.w (see Eq. Plot the dissipation and compare it with the production.2 and K. Compare the contributions due the shear stress and the normal stresses. K. The compute ni. Anything interesting? ′2 When v2 reaches a maximum or a minimum along a grid line normal to the wall.

e. Use the simple eddy viscosity model for the turbulent diffusion term.e.4) is usually positive. K. Zoom in on interesting regions. K. and the strain rate tensor. Compute the stresses using the Boussinesq assumption. ∂¯i /∂xj are parallel.09. To ﬁnd out to what degree the exact Reynolds stress and the strain rate are parallel. Make also a zoom-in near walls. explain why. 0. Plot the eigenvectors as a vector ﬁeld. Plot the different terms in the equations for one vertical grid line fairly close to the inlet (not too close!). the eigenvectors (π/4. Since the two eigenvectors are perpendicular to each other it is sufﬁcient to plot one of them ( for example. Usually. vi vj . The eigen¯ values correspond to the normal strain in the direction of the eigenvectors (see Section 13). i. The diffusion terms Dij and Dε can be modelled using the Generalized Gradient Diffusion Hypothesis GGDH of [160] Dij = ∂ ∂xm cuk um ′ ′ k ∂vi vj ε ∂xk (K. the off-diagonal components) dominate. you will get eigenvectors in the direction ±π/4 ± π/2 and if the normal strains (i.9 must be positive is of course that neither νt nor sij sij can go negative.9. c1 . −π/4). Compute the eigenvalues and eigenvectors of the strain tensor. The reason why the eddy-viscosity production in Eq. If the ﬁgure becomes too crowdy.4 in the entire domain to investigate if the production is negative anywhere. c2 .9) is always positive.10. (−π/4.6. sij . −π/4) and (π/4.e vi vj = −2νt sij + 2 (2k/3)δij where νt = cµ k /ε. 1. The exact production of turbulent kinetic energy (see Eq. Choose two stresses.3. c1w . K. Our ﬂow is 2D. 0.8. all represent the same principal coordinate system).2. Assignment 1.7) This diffusion model can cause numerical problems. σk ) = (0. use two plots per vertical grid line or simply omit terms that are negligible. Assignment 1. (−π/4. 1) Assignment 1. If so. π/4). Is that the case for you? ′ ′ ¯ Assignment 1.5. Another way to explain this fact is ¯ ¯ ′ ′ v that the modelled Reynolds stress.K. νt = Cµ k 2 /ε (K.7. It can however become negative. Compute the exact production in Eq. Analysis 227 3. one can compute the eigenvectors. c2w . a stress is large in locations where its production (or pressure-strain) term is large. Compare the eddy-viscosity stresses with two of the Reynolds stresses from the database. Try to explain why some terms are large and vice versa.8) The following constants should be used: (cµ . and the GGDH is then replaced by a simple eddy viscosity model Dij = ∂ ∂xm ′ ′ νt ∂vi vj σk ∂xm . 0. π/4). If the shear strains (i.5. thus we get two eigenvectors and two eigenvalues. the diagonal components) dominate the direction of the eigenvectors will be along the x1 and x2 axes (explain why!). When using the Boussinesq assumption the production of turbulent kinetic energy P k = 2νt sij sij ¯ ¯ (K. .

at the center of the cell.3 Compute derivatives on a curvi-linear mesh In this appendix we describe how the derivatives on a curvi-linear grid are computed in the provided Matlab function dphidx dy. ). The divergence theorem for a scalar. The eigenvalues correspond to the normal stresses in the direction of the eigenvectors.nw ) and (x1.K. compute it as v1. The velocity v1 is stored at the corners (ne. Compute derivatives on a curvi-linear mesh N ne nw se = (s1e .nw . Compute the eigenvalues and eigenvectors of the Reynolds stresses. including Cartesian ones. say v1 . K.).sw ).m. K. In order to do that we use Gauss’ law over a control volume centered at face e (dashed control volume in Fig.se . .1: Control volume.se + v1. Zoom in on interesting regions. nw.nw + v1. .. When you need a variable.3. x2 are given at the corners (ne. Coordinates x1 . The data you have been given.11. x2. at (x1. reads ∂φ φni dA dV = ∂xi A V To compute ∂v1 /∂x1 we set φ = v1 and i = 1 which gives V ∂v1 dV = ∂x1 v1 n1 dA A Assuming that ∂v1 /∂x1 is constant in the volume V we obtain 1 ∂v1 = ∂x1 V v1 n1 dA A . . nw. .10) Let’s compute ∂v1 /∂x1 .1). x2. In which regions are the eigenvectors of the Reynolds stress tensor and those of the strain tensor not parallel? This should indicate regions in which an eddy-viscosity model would perform poorly. but the approach used below works for all meshes.sw + v1. x1 and x2 and all variables are stored at the grid points. (x1. (x1.P = 1 (v1.sw . x2.. On a Cartesian grid it is more convenient to use the built-in Matlab function gradient. s2e ) P W sw se S ne = (n1e .e. φ.ne ). vi vj . Zoom in on interesting regions. i.ne ) 4 (K. n2e ) E 228 Figure K. x2.se ). ′ ′ Assignment 1.ne .

we get s1e = x1. (K.3.3. ˆ gives us the normal vector for the east face as n1e = s2e n2e = −s1e . For the east face.w.1) ∂v1 ∂x1 = 1 V (v1 n1 A)i = i=e.ne − x1.se de x2.ne − x1.ne − x2.11) . K.K.ne − x2. Compute derivatives on a curvi-linear mesh 229 In discrete form we can write (see Fig.se )2 + (x2. The relation between the normal vector n. s and the unit vector in the z-direction s·n =0 s × z = n.1 Geometrical quantities It is useful to ﬁrst compute the unit vectors s along the control volume.w 1 {(v1 An1 )e + (v1 An1 )n + (v1 An1 )w + (v1 An1 )s } V K.se )2 s2e = de = (note that the area of the east face Ae is equal to de since ∆z = 1).n. for example.se de (x1.

then use also data ﬁles ’2’ and ’3’ averaging by use of the three ﬁles.lyx. You can also use Octave on Linux/Ubuntu. Use Matlab to analyze the data.2) Wait with analysis of the results till you have done next part. When everything works.org) or MikTex (www.L. always denote time averaging.chalmers. The cell size in x and z directions are ∆x = 0. z. vi vj . for example ′ ′ v1 v2 = (v1 − v1 ) (v2 − v2 ) = v1 v2 − v1 v2 − v2 v1 + v1 v2 = v1 v2 − 2 v1 v2 + v1 v2 = v1 v2 − v1 v2 . v (v2 ).tfd. The streamwise. MTF270. It is only in this special academic test case where we have three homogeneous directions (x. Lyx (www. Use the deﬁnition to compute the stresses. denotes averaging in the homogeneous directions x and z. for example. Since the data ﬁles are rather large.0164. Octave is a Matlab clone which can be downloaded for free. wall-normal and spanwise directions are denoted by x (x1 ). . It is recommended (but the not A A required) that you use LTEX(an example of how to write in LTEXis available on the course www page). p = p + p′ The symbol . L.1 We decompose the instantaneous variables in time-averaged and ﬂuctuating quantities as ′ vi = vi + vi .org) which are both free to download.1 Task 2. The size od the domain is (L. It is available on Linux. y (x2 ) and z (x3 ) respectively. Assignment 2: LES 230 L MTF270. At the www-page (http://www.1) 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 ). w (v3 )and p (made nondimensional by uτ and ρ). it is recommended that you do all tasks using only data ﬁles ’1’.miktex. The equations that have been solved are ∂vi =0 ∂xi ∂vi ∂p 1 ∂ 2 vi ∂ (vi vj ) = δi1 − + + ∂t ∂xj ∂xi Reτ ∂xj ∂xj (L. Zmax) in (x.L. On Windows you can use.se/˜lada/comp turb model) you ﬁnd data ﬁles with three instantaneous ﬂow ﬁelds (statistically independent). Periodic boundary conditions were applied in the x and z direction (homogeneous directions).1. Assignment 2: LES You can do the assignment on your own or in a group of two. z). y. You will receive data from a DNS of fully developed ﬂow in a channel.0654 and ∆z = 0. You ﬁnd a Matlab/Octave program at the www-page which reads the data and computes the mean velocity. A 96 × 96 × 96 mesh has been used. The data ﬁles are Matlab binary ﬁles. (L. Note that in reality . The data ﬁles include the instantaneous variables u (v1 ). t) where we can – in addition to time averaging – also can use x and z averaging. Compute the six stresses of the stress ′ ′ tensor. see Fig. h.

1: Channel domain.max (b) x2 − x3 plane.c.L.1 231 periodic b. periodic b. Task 2.1. x2 j 2h x3 k X3. .c. Figure L. x2 j 2h v1 ¯ x1 i L (a) x1 − x2 plane.

. Repeat Item 1. Discuss the differences between no ﬁlter. ¯ ¯ 2. but now for a 2D ﬁlter (x1 and x3 direction). the formula for a 3D ﬁlter is given in Eq. see Eq.2 232 L.g. ∆ = 2∆x1 and ∆ = 4∆x1 . If we would use many more time steps – or. Do the same thing again but with a ﬁlter width of ∆ = 4∆x1 (now you must derive the expression on your own!). e. in Item 1 and 2 you have just carried out explicit ﬁltering.e.L. if we let T → ∞ when time averaging. Task 2. in general. Which stresses do you think are symmetric with respect to the centerline? or anti-symmetric? What’s the reason? When averaging. we use an implicit ﬁlter).26.3 Task 2. we use only three time steps (three ﬁles).3 Plot v1 and v2 along x1 at two different x2 values at x3 = x3. Compare v1 and v2 with v1 and v2 .2.2 ∂ v1 ∂y ∂ v1 ′ ′ = −2 v1 v2 ∂y ′ ′ ∂ v1 = − v2 v2 ∂y ′ ∂v = 2 p′ 1 ∂x ∂v ′ ∂v ′ = p′ 1 + p′ 2 ∂y ∂x ′ ∂v = 2 p′ 2 ∂y Compute the production term and the pressure-strain terms ′ ′ Pk = − v1 v2 P11 P12 Φ11 Φ12 Φ22 Do the production terms and the pressure-strain term appear as you had expected? (see the previous course MTF256) Now analyze the ﬂuctuations in the previous subsection. In LES we almost always assume that the ﬁlter width is equal to the control volume (i. Compare the results along the same lines as in Item 1 and 2. Filter v1 and v2 to get v and v2 using a 1D box-ﬁlter (in the x1 direction) with ¯ ¯ ﬁlter width ∆ = 2∆x1 (this corresponds to a test ﬁlter. 17.max /2.27. Above. φ = lim 1 T →∞ 2T +T φdt −T then some of the stresses would be zero: which ones? Why? L. 1. 17.2 Task 2.

4) where n is the distance to the nearest wall and κ = 0. Task 2.L. . ∆} (L. vi . Compare the SGS stress τ12 with the resolved stress u′ v ′ and compare the SGS viscosity with the physical one.4 2 Compute the SGS stress τ12 from the Smagorinsky model. νsgs = (Cs fµ ∆) ¯ sij = ¯ 1 2 ∂¯i v ∂¯j v + ∂xj ∂xi 2¯ij sij s ¯ (L.3) fµ = 1 − exp(−x+ /26) 2 The ﬁltered velocities. The constant Cs = 0.1. Plot them across the channel. ¯ we should really have used a 3D ﬁlter.4 (von K` rm` n constant).2: Spectrum with cut-off. compute the ﬁlter length as ∆ = min{κn. fµ . which reads τij = −2νsgs sij . Any thoughts? As an alternative to the damping function. we use the 2D ﬁlter. look in the Lecture Notes how a 3D ﬁlter is done.4 cut-off 233 E(κ) I II G III S κ Figure L. In this a a case you should set fµ = 1. but in order to keep it simple.4.4 Task 2. L.3 using the 2D ﬁlter (in x1 and x3 ). are taken from Task 2. Before doing the 2D ﬁlter.

7 appears as an ¯′ instantaneous production term in the equation for ksgs [74. Hence the SGS dissipation in Eq.6 Task 2. the production term in the equation for turbulent kinetic energy appears as a sink term in the equation for the mean kinetic energy. L.chalmers.tfd. and ¯ ¯′ ′′ SGS ﬂuctuation. When deriving the ksgs equation. i. vi . compare and discuss the four dissipations (see Fig.3 & 2.1.tfd.7) (L. resolved ﬂuctuation. what is the relation between εsgs and ε? Considering the cascade process. L. 8. Task 2. 20 in [73]) ′ ′ The transport equation for 1 vi vi is derived in [14]. vi . vi .5 234 L. In LES we introduce a ﬁlter which is assumed to cut off the spectrum at κc in the inertial region. Introduce a 2D ﬁlter (2∆x1 and 2∆x3 ) as in Tasks 2. L. Compute the SGS stresses from the deﬁnition ¯ ¯ ¯ ¯¯ τij = vi vj − vi vj and compute the SGS dissipation εsgs = − τij ∂¯i v ∂xj (L. see Fig. but now for the WALE model by [161]. kinetic energy is extracted from the resolved ﬂow by the SGS dissipation εsgs . The ﬂow of kinetic energy can be illustrated as in Fig. Since the cut-off is assumed to be located in the inertial sub-range (II).8) 1 ′ ′ 1 ¯ ¯ ¯ ¯ ¯¯ Now we have three equations for kinetic energy: K = 2 vi vi .35. L. the SGS dissipation is at high Re numbers equal to the dissipation.5 ∂¯i 2 v . is made. At cut-off. Plot (along x2 ).3) .325 which corresponds to Cs = 0. ′′ ′′ vi = vi + vi = vi + vi + vi ¯ ¯ ¯′ (L.4. no decomposition into time-averaged. but now we have a decomposition into time-averaged ﬁltered velocity. k = 2 vi vi and 1 ′′ ′′ ksgs = 2 vi vi .se/˜lada). ] (can be downloaded from www. v2 and v3 .e.L. (can be downloaded from 2¯ ¯ www.5. L. This is the case also in LES. which reads gij = (L. g = gik gkj ∂xj ij 1 1 2 2 2 g + gji − δij gkk sd = ¯ij 2 ij 3 3/2 sd sd ¯ij ¯ij 2 (Cm ∆) 5/4 5/2 (¯ij sij ) + sd sd s ¯ ¯ij ¯ij Repeat the Task 2.chalmers.se/˜lada). vi . vi . Fig.3 (cf.5) νsgs = with Cm = 0. what did you expect? Recall that when we do traditional Reynolds decomposition.6 ε=ν ′ ′ ∂vi ∂vi ∂xj ∂xj Compute the dissipation and plot ε across the channel.6) Now.4 and ﬁlter all velocities to obtain v1 .5 Task 2.2. 75. see Eq. and ¯ resolved ﬂuctuations.

Task 2. ν ∂¯i ∂¯i v′ v′ ¯ : viscous dissipation term in the k equation ∂xj ∂xj νsgs ∂¯1 v ∂x2 ∂ v1 ¯ ≃ νsgs ∂x2 ∂ v1 ¯ ∂x2 2 ¯ : SGS dissipation term in the K equation. ∆ = 2∆. i.3: Transfer of kinetic turbulent energy. L. This is the modelled SGS dissipation. K = 1 vi vi and k = 2 vi vi 2 ¯ denote time-averaged kinetic and resolved turbulent kinetic energy.7. The exact SGS dissipation is computed as (since ε′ is a product of two ﬂuctuating quantities. 3 can be downloaded from course home page .7 Above the ﬁltered velocities were computed using the ﬁlter width ∆ = 2∆x1 . L. In dynamic models.7 Task 2. ∆T denotes increase in internal energy.e. L.7.7 235 ksgs ≃ ¯ K 2 s ν sg ¯ s ij ¯ s ij ∂ vi ¯ ∂xj ε εs′ gs vi vj ¯′ ¯′ ¯ k ν ∂v ¯ ∂x i ∂ v ¯ j ∂x i j ¯ vi ′ ¯ ∂ ∂v i xj ∂ ν ∂xj ′ ∆T 1 ′ ′ ¯ ¯ ¯ ¯¯ Figure L.2) νsgs ε′ = sgs ′ −τij ′ ε′ = sgs ∂¯i v′ ∂xj = − τij ∂¯i v ∂xj + τij ∂ vi ¯ ∂xj (L. respectively. v1 v2 ¯′ ¯′ ∂ v1 ¯ : dissipation (which is equal to production with minus sign) by resolved ∂x2 ¯ turbulence in the K equation ∂¯′ ∂¯i v v′ ∂¯i ∂¯i v v ¯ ≃ νsgs i : SGS dissipation term in the k equa∂xj ∂xj ∂xj ∂xj tion. dissipation.L.e.9) Note that ε′ is deﬁned using the ﬂuctuating velocity gradient (see [14]3 ). consgs trary to εsgs = Pksgs in Eq. we compute it with sgs the same formula as in Eq. we often deﬁne the test ﬁlter as twice the usual ﬁlter. i.

10 Task 2. ζm ) = 1 96 × 96 (L. for example.1 and Fig. In order to estimate m.8. it is no point showing both symmetric parts. v1 and v2 (i. the sum of all coefﬁcients should give the energy. ζm = 2∆x3 .9 96 96 ′ ′ v1 (xI . What’s the difference between L1 and L3 ? L. The reason is that the Fourier coefﬁcients correspond to the energy spectrum. Take advantage of the 3 fact that the ﬂow is periodic. the two-point correlations in Section 10.10) and compare it (across the channel) with the resolved stress and the SGS stress τ12 deﬁned in Eq. If.rms = 1 ˆ B11 /N (L. what is m in v1 = m O(x2 )). 3 3 Plot the two-point correlation at a couple of x2 positions.L. for v1 . When plotting two-point correlations.12) Compute also L3 .10 The energy spectrum of any second moment can be obtained by taking the FFT of the corresponding two-point correlation. Compute and plot the integral length scale. L.e. i. When we take the FFT of Eq.1). we get ˆ B11 (κz ) = F F T (B11 ) and summation gives N 2 v1.11) I=1 K=1 where xK and ζm are the spanwise locations of the two points.e. Task 2.8 Task 2.8 ′2 ′2 What is the near-wall behavior of v1 . You can ﬁnd some details on how to use Matlab’s FFT in Appendix D. Do the quantities exhibit the near-wall behaviour that you expected? L.11. The energy spectrum of any second moment can be obtained by taking the FFT of the corresponding two-point correlation.9 Task 2. ∆ = 4∆x1 ) to compute the dynamic Leonard stress L12 from the deﬁnition ¯¯ ¯ ¯ Lij = vi vj − v i v j ′ ′ v1 v2 (L.6. We have 96 cells in the x3 direction. x2 . x2 . plot the quantities in log-log coordinates. show only half of it (cf.. and one of the points (x1 ) is at K = 1 then the other (x1 − 2∆) is at K = 95. ζ)dζ (L. D.13) . L1 .rms 0 ∞ B11 (x2 . If you have computed the Fourier coefﬁcients properly. which is deﬁned by L1 (x2 ) = 1 2 v1. for example. Do you expect the magnitude of stresses to be similar? L.8 236 Use this deﬁnition (1D ﬁlter. xK )v1 (xI . but be careful when integrating the correlation above in the x3 direction. xK − ζm ) 1 3 1 3 The two-point correlation for u′ B11 (x2 . and if we integrate the energy spectrum over all wave numbers we get the total energy. L.

Conﬁrm that Eq. show only half of it (cf. L.L. it is no point showing both symmetric parts.11 Task 2.11 237 see Appendix D Plot the energy spectra at a couple of x2 locations.13 is satisﬁed. L. When plotting two energy spectrum. the energy spectrum in Fig.11 Think of an interesting turbulent quantity and plot it and analyze it! .5 b).11. Task 2. D.

v. Above we See the maximum separation in time to 500 samples. use the axis-command in the same way as when you studied the DNS data. 2 A 64 × 96 × 64 mesh has been used. two nodes are located in the URANS region and two nodes in the LES region. The Re number based on the friction velocity and the half channel width is Reτ = uτ h/ν = 8000. Assignment 4: Hybrid LES-RANS 238 M MTF270. Use 2 the Matlab program pl time hybrid to load and plot the time history of v1 . All data have been made non-dimensional by uτ and ρ. You will also ﬁnd a ﬁle with time history of u. Assignment 4: Hybrid LES-RANS In this exercise you will use data from a Hybrid LES-RANS for fully developed channel ﬂow. Then compute the integral timescale . and thus what you see is really v1 /uτ .05 and ∆x3 = 0. w and kT (made non-dimensional by uτ and ρ).35 ¯ ¯ are shown.0028. Does the time history of v1 behave as you expect? ¯ ¯ What about v2 ? ¯ Compute the autocorrelation of the four points imax=500. 28 cells (29 nodes including the boundary) are located in the URANS region at each wall. How do you expect that the difference in νt affects the time history of v1 . it is recommended that you do all tasks using only data ﬁles ’1’. The time history of v1 at x2 /δ = 0.1 Time history At the www page you ﬁnd a ﬁle u v time 4nodes hybrid. averaging by use of the four ﬁles. The turbulence model is the same as in [73] (no forcing). The cell size in x1 and x3 directions are ∆x1 = 0. but the domain and Reynolds number is taken from [146]. In the URANS region the turbulent viscosity is much larger than in the LES region. M. x2 /δ = 0. Use one of these programs to analyze the data. x+ = 120. With uτ = 1 and ν = 1/Reτ = 1/8000. ¯ Recall that the velocities have been scaled with the friction velocity uτ .06. MTF270.dat with the time history of v1 and v2 . The matching line is located at x+ ≃ 500.M. Since the data ﬁles are rather large. ’3’ and ’4’.099 and x2 /δ = 0. this correspond to x+ = 22. Octave is a Matlab clone which can be downloaded for free. The data ﬁles are Matlab binary ﬁles.35. You ﬁnd a Matlab/Octave program at the www page which reads the data and computes the mean velocity.imax). x2 /δ = 0. Periodic boundary conditions were applied in the x1 and x3 direction (homogeneous directions).025. x2 /δ = 0.250E − 4 (every time step). The ﬁle has nine columns of v1 and v2 at four nodes (and time): ¯ ¯ ¯ ¯ x2 /δ = 0. At the course www page you ﬁnd data ﬁles with instantaneous ﬂow ﬁelds (statistically independent) of The data ﬁles include the instantaneous variables u. then use also data ﬁles ’2’. The sampling time step is 6. Hence. respectively. When everything works.015 and x2 /δ = 0. To study the proﬁles in closer detail. How does the time variation of v1 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. Plot v1 for all four nodes.015. x+ = 792 and 2 2 2 x+ = 2800. two_uu_1_mat=autocorr(u1. Use Matlab or Octave on Linux/Ubuntu.

two_uu_1_mat(1:imax).M.j.[20]) How does it compare to the integral timescale. Use more than one ﬁle ¯ to perform a better averaging.2 Mean velocity proﬁle After having performed a hybrid LES-RANS. but if you use only one ﬁle this may not be the case due to too few samples).k)-vmean(j).5 and plot it in a new ﬁgure. Compute the autocorrelation and integral timescale also for the other three points. int_T_1=trapz(two_uu_1_mat)*dt. Plot the autocorrelation. Now compute ′ ′ v1 v2 .’fontsi’.3 Resolved stresses We want to ﬁnd out how much of the turbulence that has been resolved and how much that has been modelled.j. Plot it in a new ﬁgure (a ﬁgure is created by the command figure(2)). Compute also the resolved turbulent kinetic energy kres = 0.k)-umean(j). Mean velocity proﬁle 239 dt=t(1). ′2 ′2 ′2 v1 + v2 + v3 . vfluct=v3d(i. The time averaged velocity proﬁle is compared with the log proﬁle (markers). for k=1:nk for j=1:nj for i=1:ni ufluct=u3d(i. plot(t(1:imax).m to look at the mean velocity proﬁles. Compute ﬁrst vmean (this quantity should be very small. uv(j)=uv(j)+ufluct*vfluct. end end end uv=uv/ni/nk.1). Here’s an example how to do: uv=zeros(nj. we want to look at the time-averaged results.2) xlabel(’t’) ylabel(’B_{uu}’) handle=gca set(handle.m reads the instantaneous v1 ﬁeld and performs an averaging in the homogeneous direc¯ tions x1 and x3 . Do you see any difference between the points in the URANS region and the LES region? M. pl uvw hybrid.2. M.’linew’. There are four ﬁles with instantaneous values of v1 . Use the ﬁle pl uvw hybrid.

of v1 . δV . Let’s ﬁnd the modelled shear stress.e. you need the volume. How does it behave (i.5 The modelled turbulent shear stress We have computed the resolved shear stress. Zoom ¯ ¯ up near the wall.7 SAS turbulent length scales Compute the 1D von K´ rm´ n length scale deﬁned as a a LvK. Is it smooth? Do you need more samples? If so. use more ﬁles.1D = κ ∂ v1 /∂x2 ¯ ∂ 2 v1 /∂x2 ¯ 2 (M. When computing ∆. Recall that ν = 1/8000. look at the beginning of the m-ﬁle.07kT ℓ 240 ℓ νT Table M. of the cells. i.2kT n/ν)] 1/2 1/2 0. M. Plot νT /ν.5n[1 − exp(−0.09 · 2. Are they smooth across the interface? (recall that forcing is used) Is the resolved shear stress large in the URANS region? Should it be large? Why/why not? M.M. Compute the turbulent viscosity according to Table M. ∆x1 and ∆x3 are constant and ∆x2 is stored in the array dy(j). Where is it large and where is it small? (Recall that the URANS region is located in the ﬁrst ten cells).5kT n[1 − exp(−0.1: Expressions for ℓ and νT in the LES and URANS regions.e.4. Load a ﬁle with the modelled turbulent kinetic energy.1. M. Plot the ℓSGS and ℓURAN S length scales in both regions. kT (ﬁle te1 hybrid.014kT n/ν)] LES region ℓ = ∆ = (δV )1/3 1/2 0.1) Note that you should take the derivatives of the averaged v1 velocity.4 Turbulent kinetic energy Now plot and compare the resolved and modelled turbulent kinetic energies.1 and do the usual averaging. One would expect that (∆x1 ∆x2 ∆x3 )1/3 < ℓURAN S everywhere. what is n in O(xn )? What should n be? 2 .6 Turbulent length scales Compute and plot the turbulent length scales given in Table M. Compute the modelled shear stress from the Boussinesq assumption τ12 = −2νT s12 = −νT ¯ ∂¯2 v ∂¯1 v + ∂x2 ∂x1 Plot it and compare with the resolved shear stress. Which is largest? Which is largest in the URANS region and in the LES region. Is this the case? M.mat). Turbulent kinetic energy URANS region 1/2 2. n denotes the distance from the wall. respectively? What about the sum? The magnitude of resolved and modelled turbulent kinetic energies is discussed in the last subsection in [73]. It is computed as δV = (∆x1 ∆x2 ∆x3 ). Which is largest? Any surprises? Compare them with ∆x2 and (∆x1 ∆x2 ∆x3 )1/3 .

M.2) How does it compare with LvK. T1 in Eq.1D ? When we’re doing real 3D simulations. 21. M. the ﬁrst and second derivative must be deﬁned in 3D.4) 2 Plot the von K´ rm´ n length scale using Eqs. Compare them with a a Eq.4. M.7.e.3) ∂ vi ¯ ∂ vi ¯ ∂xj ∂xj ∂xj ∂xj The second derivative is then computed as U ′′2 = + + ∂ 2 v1 ¯ ∂ 2 v1 ¯ ∂ 2 v1 ¯ 2 + ∂x2 + ∂x2 ∂x1 2 3 ∂ 2 v2 ¯ ∂ 2 v2 ¯ ∂ 2 v2 ¯ + + 2 2 ∂x1 ∂x2 ∂x2 3 ∂ 2 w3 ¯ ∂ 2 w3 ¯ ∂ 2 w3 ¯ 2 + ∂x2 + ∂x2 ∂x1 2 3 2 2 (M. a a ¯ LvK.inst = κ ∂¯1 /∂x2 v ∂ 2 v1 /∂x2 ¯ 2 (M.e.5) and what effect does it give to ω? Another way to compute the second derivative is U ′′2 = + + ∂2v ¯ ∂x2 ∂2v ¯ ∂x2 ∂2w ¯ 2 ∂x 2 + 2 ∂2v ¯ ∂y 2 ∂2v ¯ ∂y 2 ∂2w ¯ 2 ∂y 2 + 2 ∂2v ¯ ∂z 2 ∂2v ¯ ∂z 2 ∂ 2w ¯ 2 ∂z 2 2 + 2 + 2 (M.1D. .5 2 0.3 and M.inst = κ S U ′′ 0. SAS turbulent length scales 241 Compare that with the von K´ rm´ n length scale deﬁned from instantaneous v1 . M. One way of deﬁning the von K´ rm´ n length scale in 3D is [140. 141] a a LvK.5) 2 + + Plot and compare the von K´ rm´ n length scales using the second derivatives deﬁned in a a Eqs. What’s the difference? What effect do the different length scales give for PSAS (i.4 and M.3D. i.5.1.5 S = (2νt sij sij ) ¯ ¯ U ′′ = 2 (M.

4.10) and the εu (Eq. The ﬁle has eight columns of v2 along two lines: x2 = 0. The interface condition for εu is computed with the baseline value Cs = 0. Assignment 5: Embedded LES with PANS Interface 2 x2 x1 RANS.6 − 103) × 24 in viscous units.dat with the time history of v2 . In the present simulations.6 domain. respectively. the wall-normal (y) and the spanwise (z) direction.e. εinter . 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. see Fig. The Reynolds number for the channel ﬂow is Reτ = 950 based on the friction velocity. a mesh with 64 × 80 × 64 cells is used in. N. where kRAN S is taken at x = 0.2) LES with fk = 0. with fk = 1).N. which is estimated from the Smagorinsky model as ℓsgs = Cs ∆ and the interface condition for ku is computed as LES kinter = fk kRAN S (N.12. N. fk = 0. δ = 1 and uτ ≃ 1. 22.07. Assignment 5: Embedded LES with PANS 242 N MTF270. Anisotropic synthetic ﬂuctuations are added at the interface. see Fig. The turbulent viscosity is computed from Eq. is set from kinter and an SGS length scale. The modelled dissipation.5. The ku (Eq. 22. N. and half the channel width. The PANS model is a modiﬁed k − ε model which can operate both in RANS mode and LES mode.95.1. uτ . we have set ρ = 1.1. see Fig. The interface separates the RANS and the LES regions. N.1. Inlet conditions at x = 0 are created by computing fully developed channel ﬂow with the LRN PANS model in RANS mode (i.4).0 LES LES. 22.0139 (x+ ≃ 13) and ¯ ¯ 2 . fk = fk <1 0. δ. ℓsgs . With a 3. The resolution is approximately (the wall shear stress varies slightly along the wall) 48 × (0. In this exercise you will use data from an embedded PANS of channel ﬂow. the streamwise (x). fk = 1.e.25 Figure N.2 × 2 × 1.95 2. The data are taken from [148].1 Time history At the www-page you ﬁnd a ﬁle u time interior.1) (N.1: Channel ﬂow conﬁguration. downstream the equations operate in LES mode ((i.18) equations are solved. MTF270. The RANS part extends up to x1 = 0.

2.e. Where is it large and where is it small? Why? Now plot it also vs.954 and from Eq. i. Two x1 stations are shown in ﬁgure 1.e.m to look at the mean quantities such as velocity.95. both in the RANS region and in the Now plot the resolved shear stresses. You can also use Octave on Linux/Ubuntu. 1. the integral timescale T is used.925. Use the ﬁle pl uuvvww 2d. x1 (ﬁgure 2). The v1 proﬁles are very different in the RANS region (x1 < 0. for x1 < 0. 2. N. The resolved stresses v1 are plotted vs x2 (ﬁgure 1) and vs. x2 in both regions. The autocorrelation is deﬁned as ∞ B(τ ) = 0 v(t)v(t − τ )dt (N. Use Matlab.2 Resolved stresses Now we will look at the time-averaged results. 11 in [148] we can then compute the prescribed integral timescale. 1. right? . Plot the ′2 resolved stress also in the RANS region. Plot v2 for the other nodes and study the differences. Try to understand the coding. Normalize it with ν . turbulent viscosities etc. use the axis-command in the same way as when you studied the DNS data. The 2 sampling time step is 0. Something drastically happens at x1 = 0.3) Study the coding and try to understand it. i. resolved and modelled stresses. Octave is a Matlab clone which can be downloaded for free.0139 at x1 = 0.775.m reads the ﬁelds and transforms them into 2D arrays such as u 2d. When prescribing the time correlation of the synthetic ﬂuctuations.95).95. Why is there such a big difference in the ﬂuctuations? If you’re not interested in integral time scales. x1 ′ v ′ . The constant a is in [148] set to 0.N. Resolved stresses 243 x2 = 0. 11 in [148].675.000625 (every time step). pl uuvvww 2d. skip the rest of this section and proceed to Section N. ¯ To study the proﬁles in closer detail.675 are shown.175 and x1 = 2. In Matlab ﬁgure 2. don’t you? N.3 Turbulent viscosity Plot the turbulent viscosity vs. Use the ¯ Matlab/Octave program pl time pans to load and plot the time history of v2 . Compute the autocorrelations ¯ and the integral timescales.175. uu 2d. they are located at x1 = 0. The integral time scale is deﬁned as ∞ T = B norm (τ )dτ 0 (N.95) and in the LES (x1 > 0.175.775 and x1 = 1.. The time history of v2 at x2 = 0. aren’t they? ′2 Why? This can also be seen in ﬁgure 2 where v1 is plotted vs. plot νu /ν. see Eq. v1 2 ′2 LES region. ′2 Run pl uuvvww 2d.2.24 (x+ ≃ 230). In the Matlab ﬁle the integral timescale is computed from the autocorrelation. the autocorrelation is plotted. x1 . x1 = 1.4) where B norm = B(τ )/B(0) so that B norm (0) = 1. You ﬁnd the same difference between RANS and LES region as for v1 .

ku .5) vi vj mod = − νu ∂xj ∂xi 3 Compare the resolved and the modelled shear stress and streamwise normal stresses in the RANS region and in the LES region.N. Now investigate the LES region the relation between Pu = εsgs and the production. k. Computer the modelled Reynolds stresses from the Boussinesq assumption 2 ∂ vj ¯ ∂ vi ¯ ′ ′ + δij ku + (N.5 Turbulent SGS dissipation In an LES the resolved turbulent ﬂuctuations can be represented by a energy spectrum as in Fig. 8.4 Modelled stresses In Section N. which represents a source term in the k equation (Eq. L. due to the resolved turbulence ′ ′ P k = − vi vj ∂ vi ¯ ∂xj .2. εu .3 where the energy in ¯ ¯ K mostly goes to resolved turbulence. Now let’s look at the modelled stresses. there is no resolved turbulence. The resolved turbulence extracts kinetic energy via the production term. then to modelled turbulence. 8. the production term in the ku equation includes both mean and ﬂuctuating strain rates since Pu = εsgs = νu ∂¯j v ∂¯i v + ∂xj ∂xi ∂¯i v ∂xj which in the Matlab ﬁle is stored as pksgs 2d. however.35).4.14) and a sink term in the ¯ K equation (Eq.2: Energy spectrum. P k . In RANS mode. N. κ N. N. In the LES region. P k .2 you looked at the resolved Reynolds stresses. ksgs (or ku ) and ﬁnally to internal energy via dissipation. Modelled stresses 244 E(κ) Pk in e rti al gs ran = εs ge ε u E(κ) ∝ κ −5/3 P dissipating range κc Figure N. The energy ﬂow is visualized in Fig. Hence the kinetic energy ¯ goes directly from K to the modelled turbulence.

does the relation Pu = εu hold? . plot also this quantity. In both the RANS and the LES region the process of viscous dissipation takes place via εu . Is the turbulence in local equilibrium.e. i. Hence.N.5. Turbulent SGS dissipation 245 Compare also P k in the LES region and in the RANS region.

the transformation reads uij = bik bjm uk∗m∗ As an example. xj∗ ) In Fig. the bij is given by b11 = cos α. . set α = π/4. MTF270: Transformation of a tensor 246 O MTF270: Transformation of a tensor The rotation of a vector from the xi∗ coordinate system to xi reads (see. Equation O.O. b22 = cos α The relations bik bjk = bki bkj = δij are fulﬁlled as they should.g.1) where bij denotes the cosine between the axis bij = cos (xi . x2∗ ) and (x1 . b12 = −1/ 2. O.4 gives u11 = b11 b11 u1∗1∗ + b12 b11 u2∗1∗ + b11 b12 u1∗2∗ + b12 b12 u2∗2∗ 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∗ 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∗ 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∗ 1 = (u1∗1∗ + u2∗1∗ + u1∗2∗ + u2∗2∗ ) 2 x2 x2∗ x1∗ (O.6d) β α x1 Figure O.4) (O.1: Transformation between the coordinate systems (x1∗ . e. b21 = 1/ 2. Chapter 1 in [19]) ui = bij uj∗ (O. O.3) (O. For a second-order tensor.6b) (O. Inserting Eq.6a) √ b22 = 1/ 2 (O. O. x2 ).5 into Eq.3 gives √ √ √ b11 = 1/ 2. b12 = cos β = − sin α b21 = cos(π/2 − α) = sin α.5) (O.6c) (O..2) (O.1.

O. s1∗2∗ = 0.1 correspond to the streamwise and wall-normal directions. sij .7 are any two orthogonal vectors with angles ±π/4. Rotation to principal directions 247 O. s2∗2∗ = −s21 (O. The eigenvectors for Eq. xj ) (O.O. s2∗1∗ = 0.10d) Since the diagonal elements are zero it conﬁrms that the coordinate system x1∗ − x2∗ with α = π/4 is indeed a principal coordinate system. c12 = cos(π/2 − α) = sin α c22 = cos α (O. where c11 = cos α. s12 = 2 ∂x2 Assume that the x1 and x2 coordinates in Fig. It can be seen that the relation cji = bij is satisﬁed as it should. O. O.1.11.8 reads (α = π/4) s1∗1∗ = c11 c11 s11 + c12 c11 s21 + c11 c12 s12 + c12 c12 s22 1 = (s11 + s21 + s12 + s22 ) 2 s1∗2∗ = c11 c21 s11 + c12 c21 s21 + c11 c22 s12 + c12 c22 s22 1 = (−s11 − s21 + s12 + s22 ) 2 s2∗1∗ = c21 c11 s11 + c22 c11 s21 + c21 c12 s12 + c22 c12 s22 1 = (−s11 + s21 − s12 + s22 ) 2 s2∗2∗ = c21 c21 s11 + c22 c21 s21 + c21 c22 s12 + c22 c22 s22 1 = (−s11 − s21 − s12 + s22 ) 2 The fully developed channel ﬂow is obtained by inserting Eq. see Appendix B. O. c21 = cos β = − cos α.e.1 Rotation to principal directions Consider fully developed ﬂow in a channel.10b) (O. 2 ∂x2 λ(2) ≡ s2∗2∗ = −s12 = − 1 ∂v1 2 ∂x2 (O.7) s11 = 0. ±3π/4. respectively. The strain-rate tensor. The transformation from x1 − x2 to x1∗ − x2∗ reads si∗j∗ = cik skm . i.10c) (O. Let the x1∗ − x2∗ coordinate system denote the eigenvectors. s21 = s12 .7 s1∗1∗ = s12 . O. λ(k) . O.1.11) (O. s22 = 0 (O. of sij correspond to the diagonal elements in Eq. The eigenvalues. λ(1) ≡ s1∗1∗ = s12 = 1 ∂v1 . reads 1 ∂v1 .12) .8) see Fig.10a) (O.9) cij = cos (xi∗ . Let us choose π/4 and 3π/4 for which the transformation in Eq.

13) Insert Eq.10d with α = π/4 A1∗1∗ = A1∗2∗ A2∗1∗ A2∗2∗ 1 (A11 + A21 + A12 + A22 ) 2 1 = (−A11 − A21 + A12 + A22 ) 2 1 = (−A11 + A21 − A12 + A22 ) 2 1 = (−A11 − A21 − A12 + A22 ) 2 (O. . O.2.2 Transformation of a velocity gradient Consider the velocity gradient Aij = ∂vi /∂xj . Apply the transformation from the x1 − x2 system to the x1∗ − x2∗ in Eqs.O.12 and O. Transformation of a velocity gradient 248 O.14) It can be seen that ∂v1∗ /∂x1∗ = s1∗1∗ and ∂v1∗ /∂x2∗ + ∂v2∗ /∂x1∗ = 2s1∗2∗ = 0 (see Eqs.10a-O.9 with α = π/4 and replace Aij by the velocity gradient ∂v1∗ 1 ∂v1 ∂v1∗ = = .14) as it should. O. ∂x1∗ ∂x2∗ 2 ∂x2 ∂v2∗ ∂v2∗ 1 ∂v1 = =− . ∂x1∗ ∂x2∗ 2 ∂x2 (O. O.

Eq. P.7) As usual we are considering a volume V with bounding surface S and normal vector ni .2) The left side is re-written as ∂ ∂xi ϕ ∂ψ ∂xi =ϕ ∂2ψ ∂ψ ∂ϕ + ∂xi ∂xi ∂xi ∂xi (P. see .1) V where S is the bounding surface of the volume.5 from P.5) Subtract Eq. consider a small sphere in V .4) This is Green’s ﬁrst formula.4 gives ϕ V ∂2ψ ∂2ϕ −ψ ∂xi ∂xi ∂xi ∂xi dV = S ϕ ∂ψ ∂ϕ −ψ ∂xi ∂xi ni dS (P. Replacing Fi by ϕ gives ∂xi ∂ ∂xi ϕ ∂ψ ∂xi dV = S ϕ V ∂ψ ni dS ∂xi (P. P. P.3 Green’s third formula In Green’s second formula.4 gives ψ V ∂2ϕ dV + ∂xi ∂xi V ∂ϕ ∂ψ dV = ∂xi ∂xi ψ S ∂ϕ ni dS ∂xi (P. MTF270: Green’s formulas 249 P MTF270: Green’s formulas In this appendix we will derive Green’s three formulas from Gauss divergence law. P.2 Green’s second formula Switching ϕ and ψ in Eq. V .2 gives ϕ V ∂2ψ dV + ∂xi ∂xi V ∂ψ ∂ϕ dV = ∂xi ∂xi ϕ S ∂ψ ni dS ∂xi (P.P. P. and ni is the normal vector of S ∂ψ pointing out of V .3) which inserted in Eq. In the last section we will derive the analytical solution to the Poisson equation.6.6) This is Green’s second formula. The derivations below are partly taken from [162]. Since function ψ(r) is singular for r = rP . P. P. set ψ(r) = 1 |r − rP | (P.1 Green’s ﬁrst formula Gauss divergence law reads ∂Fi dV = ∂xi Fi ni dS S (P.

3.e. Green’s third formula ni S 250 V r Sε rP nε i x2 x1 Figure P. Now we replace ri = r by r − rP = ri − rP. respectively.i |r − rP |3 (P. P.P.1. i.1. In Eq. Apply Green’s second formula for this volume which has the bounding surfaces S and S ε with normal vectors ni (outwards) and nε (inwards). P.1: Green’s third formula.9) =0 = −3 To get the right side on the second line we used the fact that ri ri = r2 .9 which gives ∂ ∂xi ∂2 ∂xi ∂xi 1 |r − rP | 1 |r − rP | =− =0 ri − rP.10) P for ri = ri .8 and P.e. A volume V with bounding surface S with normal vector ni . see Fig. The ﬁrst derivative of 1/ri is computed as ∂ ∂xi 1 r =− ∂r/∂xi ri =− 3 2 r r (P. i. In V there is a small sphere S ε located at rP with radius ε and normal vector nε . P. We get i − V −S ε ∂2ϕ 1 dV = |r − rP | ∂xi ∂xi + Sε S −ϕ P ri − ri ∂ϕ 1 −ϕ − 3 |r − rP | |r − rP | ∂xi P ri − ri ∂ϕ 1 − |r − rP |3 |r − rP | ∂xi ni dS (P.11) (−nε )dS i . The second derivative is obtained as ∂2 ∂xi ∂xi 1 r =− ∂ ∂xi 1 r3 ri r3 =− ri + r ∂ri ∂xi 3ri r4 ∂r 3ri ∂xi r4 3 r2 3 =− 3 + r r r4 1 r3 + (P. P. i Fig.8) since the derivative of a distance X is a vector along the increment of the distance.6 we need the ﬁrst and the second derivative of ψ. ∂X/∂xi = Xi /X where X = |Xi |.i in Eqs. for V excluding the sphere S ε .

P.i | (P.i is the radius of sphere S ε . the radius. i. The length of the vector ri − rP.e. where xQ denotes a point on [a.13 and P. is the direction from point rP . over the interval [a. of sphere S ε goes to zero so that the integral in Eq.i r − rP = |r − rP | |ri − rP. |r − rP | = ε The surface area.3.19) This is Green’s third formula.17) where rQ ≡ rQ. there exists (at least) on point for which the the relation b a g(x)dx = (a − b)g(xQ ) (P. ε.i denotes a point on S ε .12. this is because the normal vector must point out of the volume V − S ε .20) Hence it is seen that the volume element dV goes to zero faster than the singularity 1/|r − rP |.18 gives ∂2ϕ 1 dV V |r − rP | ∂xi ∂xi P 1 ∂ϕ 1 ri − ri 1 + ni dS + ni dS ϕ 4π S |r − rP |3 4π S |r − rP | ∂xi ϕ(rP ) = − 1 4π (P.e.15 gives IS ε = ϕ(rQ ) Sε dΩ + rQ. Consider a small sphere with radius r1 = |r − rP | centered at point P . As we let Q → P . into the sphere. In one dimension this theorem simply states that for the integral of a function. V − S ε .15) To re-write this integral we will use the mean value theorem for integrals.i ∂ϕ + ε2 ε2 ∂xi ε2 dΩ = Sε ϕ + (ri − rP. P. Inserting Eqs. .i in order to make its length equal to one. g(x). for sphere S ε can be expressed in spherical coordinates as dS = ε2 Ω = ε2 sin θdθdα (P. i. P. i.i − rP. i −nε = i ri − rP.13) where Ω is the solid angle.12) where we have normalized the vector ri − rP. P. nε . Green’s third formula 251 where the volume integral is taken over the volume V but excluding the sphere S ε .e.P.19 is not a problem. In spherical coordinates the volume element can then be expressed as 2 2 dV = r1 sin θdr1 dθdα = r1 dr1 dΩ (P.i ) ∂ϕ (rQ ) ∂xi dΩ Sε (P. Inserted in Eq. Note the minus sign in front of the normal vector in the S ε integral. dS. P.18) since Sε dΩ = 4π. In the sphere the normal vector.14 in the last integral in Eq.17 reads ε→0 lim IS ε = 4πϕ(rQ ) (P. b]. S ε .i ) ∂ϕ ∂xi dΩ (P. P. The singularity 1/|r − rP | in the volume integral in Eq. i. b].14) (P. Applying this theorem to the integral in Eq.e. P.11 gives IS ε = Sε ϕ ri − rP.16) holds.

25) V This is the analytical solution to Poisson’s equation. This integral goes to zero since ϕ → 0 as R → ∞.14.22 reads ϕ(rP ) = − 1 4π f (r) dV |r − rP | (P. Analytical solution to Poisson’s equation 252 P. the ﬁrst surface integral can be written as 1 4π ϕ S P 1 ri − ri ni dS = |r − rP |3 4πR2 ϕni ni dS = S 1 4π ϕdΩ S (P.13. P. P. Eq.21) where we assume that ϕ goes to zero at inﬁnity and that the right side is limited.13 and P.23) using ni ni = 1.4 Analytical solution to Poisson’s equation Poisson’s equation reads ∂2ϕ =f ∂xj ∂xj (P. The second integral in Eq.21.21 as 1 4π ∂ϕ 1 1 ni dS = |r − rP | ∂xi 4πR S 1 ∂2ϕ 1 dV = = 4πR S ∂xi ∂xi 4πR ∂ϕ ni dS ∂xi f dV V S (P.19. Gauss divergence law and Eq. P. P. Using Eqs. gives ϕ(rP ) = − f (r) dV |r − rP | V P 1 ri − ri 1 ni dS + ϕ + 4π S |r − rP |3 4π 1 4π S ∂ϕ 1 ni dS |r − rP | ∂xi (P.P. P. Hence the ﬁnal form of Eq. Eq. P. P.4. .24) This integral also goes to zero for large R since we have assumed that f is limited. P.22) We choose the volume as a large sphere with radius R. Green’s third formula.22 can be re-written using Eq.12.

6. Discuss and show how the dissipation term. Given the modelled k equation. the term including the triple correlation) in the k equation is modelled. Show how the turbulent diffusion (i. re-written in incompressible ﬂow? 2. Which terms in the vi vj equation need to be modelled? Explain the physical ′ ′ meaning of the different terms in the vi vj equation. 9. 5. Derive the Boussinesq assumption. How is the buoyancy term. What is the expression for the total heat ﬂux that appear in the θ equation? ′ ′ 4. εij . Discuss the physical meaning of the different terms. θ. derive the modelled ε equation.Q.e. ρgi .19). 7. How is the production term modelled in the k − ε model (Boussinesq)? Show how it can be expressed in sij ¯ 8. 11. MTF270: Learning outcomes for 2012 253 Q MTF270: Learning outcomes for 2012 Week 1 1. and the transport equation ′ ¯ for θ. is modelled. . Which terms need to be modelled? ¯ 3. Given the transport equation for the temperature. Derive the transport equation for vi θ′ (Eq.

vi vj . 10.2 = −c2 Pij − 2 δij P k . modelled in the Boussinesq approach? 2. . MTF270: Learning outcomes for 2012 254 Week 2 ′ ′ ′ 1. ). Show that the turbulence is dampened if ∂vθ /∂r > 0 and that it is enhanced if the sign of ∂vθ /∂r is negative. Show also that k in the k − ε model is affected in the same way. Give the expression for the pro3 duction terms. Derive the exact Poisson equation for the pressure ﬂuctuation. 11. Describe the physical effect of stable stratiﬁcation and unstable stratiﬁcation on turbulence. .1) V Use Eqs.1w have? 2 ε ′ ′ 7. What sign must hence Φ22.g. respectively. 4. In some stress equations there is no production terms nor any dissipation term.Q.1 to derive the exact analytical solution for the pressurestrain term. For a Poisson equation ∂2ϕ =f ∂xj ∂xj there exists an exact analytical solution ϕ(x) = − 1 4π f (y)dy1 dy2 dy3 |y − x| (Q. modelled pressure-strain terms and modelled dissipation terms for a simple shear ﬂow (e. Show that the Reynolds stress model gives an enhanced turbulence production (as it should) because of positive feedback between the production terms. Describe the physical effect of the pressure-strain term in the near-wall region. boundary layer. How are the Reynolds stress. vi θ′ . . The modelled slow and rapid pressure strain term read Φij. ′ ′ Piθ = −vi vk ¯ ∂θ ∂xk respectively. respectively.61 and Q. Derive the algebraic stress model (ASM). channel ﬂow. Streamline curvature: now consider a boundary layer where the streamlines are curved away from the wall (concave curvature). Use physical reasoning to derive a model for the diagonal components of the pressure-strain term (slow part). and the turbulent heat ﬂux. 11. Eq. Show that the effect of streamline curvature in the k − ε model is much smaller. What main assumption is made? 6. Consider buoyancy-dominated ﬂow with x3 vertically upwards. as it should. The production ′ ′ ′ term for the vi vj and the vi θ′ equations read ′ ′ Gij = −gi βvj θ′ − gj βvi θ′ . Show that the Reynolds stress model dampens and increases the vertical ﬂuctuation in stable and unstable stratiﬁcation. jet ﬂow .61. 11. How come? Which is the main source term (or sink term) in these equations? 8. Consider streamline curvature for a streamline formed as a circular arc (convex curvature). What are the “slow” and “rapid” terms? 5. 3. 9.1 = −c1 ρ k vi vj − 3 δij k and Φij.

e. Show that the Boussinesq assumption may give negative normal stresses. 13. 13. 15. non-dimensional. the anisotropy tensor aij = −2νt sij /k is often used. Consider stagnation ﬂow. What is a non-linear eddy-viscosity model? When formulating a non-linear model. traceless and symmetric. SΩ) in the non-linear model in the lecture notes. . What is the two-component limit? What requirement does it put on the pressurestrain models? Show that the standard IP model and the Rotta model do not satisfy this requirement. 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 turbulent viscosity (Eq. Show the three ¯ ﬁrst terms (S 2 .Q. 16. Show that in the Reynolds stress model.12). there is only a small production of turbulence whereas eddy-viscosity models (such as the k − ε model) give a large production of turbulence. MTF270: Learning outcomes for 2012 255 12. i. Ω2 . What is a realizability constraint? There are two main realizability constraints on the normal and the shear stresses: which ones? 14. Show that each term has the same properties as aij .

Which equations are solved in the V2F model? ′2 2. T ∝ k . the f and the v2 equation) returns to the v2 equation in the Reynolds stress model. 4. MTF270: Learning outcomes for 2012 256 Week 3 1. The transport equation for v2 reads (the turbulent diffusion terms are modelled) ∂ρ¯1 v2 v ′2 ∂v ′2 ∂ρ¯v2 v ′2 ∂ ′ (µ + µt ) 2 −2v2 ∂p′ /∂x2 −ρε22 + = ∂x1 ∂x2 ∂x2 ∂x2 Φ22 Show how this equation is re-written in the V2F model. 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 being used and why? How is ω expressed in k and ε? . the V2F model (i.Q. 3. How does f enter into ′2 the v2 equation? What is the physical meaning of f ? Show that far from the ′2 ′2 walls. ε L∝ k 3/2 ε Show how the magnitude of the right side and L affect f .e. The f equation in the V2F model reads L2 Φ22 1 ∂2f −f =− − ∂x2 k T 2 ′2 v2 2 − k 3 .

Derive a transport equation for ω from the k and ε transport equations. grid (i. κc are located in the spectrum. Derive the Smagorinsky model in two different ways (Sections 17. 8. 11.e resolved) scales and the cut-off. Consider the energy spectrum. 5. The ﬁltered non-linear term has the form ∂vi vj ∂xj Show that it can be re-written as ∂¯i vj v¯ ∂xj giving an additional term − on the right side. 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 additional second-order term).6 and 17. Consider the energy spectrum and discuss the physical meaning of Pksgs and εsgs . 13. what does this function do? In which region is each model being used and why? 4. the −5/3 range and the dissipating scales). a blending function F1 is used. derive the relation between κc and ∆. Show the three different regions (the large energycontaining scales. Consider a 1D ﬁnite volume grid. 6. In the SST model. Derive the one-equation ksgs equation 12. 16. 3. Where should the cut-off be located? What does cut-off mean? Show where the SGS scales. Carry out a second ﬁltering of v at node I and ¯ ¯ ¯ show that v I = vI .14). What is the purpose of the shear stress limiter in the SST model? Show that the eddy-viscosity assumption gives too high shear stress in APG since P k /ε ≫ 1 (Eq. respectively. 7. How is κc related to the grid size ∆x for these cases? Using the sin wave. Show how a sinus wave sin(κc x) corresponding to cut-off is represented on a grid with two and four nodes. 10. 2. Show the difference between volume averaging (ﬁltering) in LES and timeaveraging in RANS. 9.22) ∂ ∂ ∂τij (vi vj ) + (¯i vj ) = − v¯ ∂xj ∂xj ∂xj . Describe the k − ω SST model.Q. MTF270: Learning outcomes for 2012 257 Week 4 1.

3. τij .1 and 18.33. What is DES? The length scale in the RANS S-A model reads computed in the corresponding DES model? 9. 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 (see “Summary of lectures”) 10.Q. Use Eq. MTF270: Learning outcomes for 2012 258 Week 5 ¯ ¯ 1.test cover. Eq. how is it . ksgs. The modiﬁed (reduced) length scale in two-equation DES models can be introduced in different equations. see Fig. Show that when a ﬁrst-order upwind schemes is used for the convection term. L. What is a test ﬁlter? Grid and test ﬁlter Naiver-Stokes equation and derive the relation (Q.2 and derive the relation 1 Lij − δij Lkk = −2C 3 2 ss ∆ | s | s ij − ∆2 |¯|¯ij ¯ ¯ This equation is a tensor equation for C. Describe URANS. How are k energy spectrum? 2. 17.11) 7. how ﬁne does the mesh need to be in the wall region be? Why does it need to be that ﬁne? 8. C. ksgs .2) νt ˜ d 2 . When doing LES. and the test ﬁlter SGS stress. Discuss the energy path in connection to the source and sink terms in the k. 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. Describe hybrid LES-RANS based on a one-equation model. Derive these three terms. ′ ′ 5. 3. In scale-similarity ¯¯ models τij is written as three different terms. K ¯ and ksgs computed from the and the ksgs equations. 17. Which equations and which term? What is the effect on the modelled. turbulent quantities? 11. 4. Discuss the choice of discretization scheme and turbulence model in URANS (see Sections 18. How is the instantaneous velocity decomposed? What turbulence models are used? What is scale separation? 13. Formulate the Smagorinsky model for the grid ﬁlter SGS stress. What are they called? What does the work “scale-similar” mean? 6. 12.2) vi vj − v i v j + τ ij = Tij ¯¯ ¯ ¯ ¯ Draw an energy spectrum and show which wavenumber range k. an additional diffusion term and dissipation terms appear because of a numerical SGS viscosity. Use this relation and derive the ﬁnal expression for the dynamic coefﬁcient. Tij . Q. We usually deﬁne the SGS stress tensor as τij = vi vj − vi vj .

MTF270: Learning outcomes for 2012 259 Week 6 1. the generated shear stress is zero: why? How is the correlation in time achieved? . κmax . Give a short description of the method to generate synthetic turbulent inlet ﬂuctuations. κmin . Describe the SAS model. With this method. 3. What form on the spectrum is assumed? How are the maximum and minimum wavelengths. Describe the PANS model. What is the main modiﬁcation compared to the standard k − ε model? What is the physical meaning of fk ? Describe what happens to the equation system when fk is reduced.Q. How is the von K´ rm´ n length scale deﬁned? An a a additional source term is introduced in the ω equation: what is the form of this term? What is the object of this term? When is it large and small. respectively? 2. determined.

P. Turbulence Heat and Mass Transfer 4. Physics of Fluids A. Mechanische ahnlichkeit und Turbulenz. [3] S. Div.O.dmt. G¨ teborg. of Applied Mechanics. [8] H. Turbulent Flow. DCW Industries. Nagano. [12] Th. o [5] L. editors. o 5:58–76. The MIT Press. and A. Cambridge. 2011. and c M. Bradshaw. Hanjali´ . of Fluid Dynamics. Springer-Verlag. 2006. Wojtkowiak and C. Aerodynamic loads on rotor blades. Davidson.R. http://torroja. Gothenburg. Mechanics of solids & ﬂuids. Davidson. 1972. Chalmers University of Technology. [14] L. Hanjali´ . o [4] H. Inherently linear annular-duct-type laminar ﬂowmeter. of Material and Computational Mechanics. Reddy. 15:178–203. Department of Applied Mechanics. Tennekes and J. Toll and M. 2001. La Ca˜ ada. Transport equations in incompressible URANS and LES. Cambridge University Press. of Applied Mechanics. Chalmers University of Technology.upm. 1998. California 91011. A First Course in Turbulence. Modiﬁcations of the v 2 − f model for computing the ﬂow in a 3D wall jet. 1994. Dept. Inc. [7] S. Popiel. 5354 Palm Drive. New York. Report 2006/01.N. UK. Dept. Nachrichten von a a ¨ der Gesellschaft der Wissenchaften zu G¨ ttingen. G¨ teborg. [16] D. Int. An introduction to turbulence models. Wilcox. Technical Report 97/2. Turbulence. of Thermo and Fluid Dynamics. Part I: Fundamentals. Y. Journal of Fluids Engineering. inc. Nielsen. [10] S. Heat and Fluid Flow. 2003. An Introduction to Continuum Mechanics. Report. In K. Tummers. New York. Lumley. 1930. Sweden.L. e 2006. Ekh. J. o [15] K. Spalart.R. Dept. Turbulence Modeling for CFD. Davidson. MSc Thesis 2011:18. Sveningsson. 1997. Div. [11] S. 18(011702). 128(1):196–198. Jim´ nez. Scaling of the velocity ﬂuctuations in turbulent chane nels up to Reτ = 2003.J. Massachusetts. pages 577–584. n . von K´ rm´ n. Fachgruppe 1 (Mathematik).C. References 260 R References [1] L. [2] J. 1976. Journal of Fluid Mechanics.. [9] P.V. 187:61–98. begell house. Sweden. Wallingford (UK). Advanced turbulence closure models: A view of current status and c future prospects. 2006. 2011. Pope. 1988. Berlin. Direct simulation of a turbulent boundary layer up to Rθ = 1410. 2006. Jim´ nez. 2 edition. Hoyas and J. Cambridge University Press. Chalmers University of Technology. Hoyas and J. [6] J. Chalmers University of Technology. Division of Fluid Dynamics. Cambridge.es/ftp/channels/data/statistics/. Abedi. G¨ teborg. 2008. [13] P. Sweden.B.

[21] B. Technical Report TFD/87/5. Mase. E. Reece. Johansson.E. Wolfshtein. Computational modeling of turbulent ﬂows.E. [26] P. 23(5):721–730. New York. Continuum Mechanics. Journal of Fluids Engineering. [25] S. Launder.E. 1975..M. Whitelaw. 1995. [33] J. References 261 [17] K. [29] S.L. UMIST. Agardograph progress no. [23] B. editors. 86:491–511. Effects of streamline curvature on turbulent ﬂow. 26:1422–1435. 1985. a 1972. 1972. Naot.A. Thompson and J. . Lumpley. and D. [31] P. and J. B. In T. Rodi and G. [27] W. [18] J. G. Lumley. 17:89–90. and M. A. 1996. Journal of Fluid Mechanics. [30] L. Launder.V. 1970. Oxford University Press. Physics of Fluids. A more general effective-viscosity hypothesis. 1970. [22] D. Launder. Accomodating the effects of high strain rates in modelling the pressure-starin correlation. Characteristics of a trailing-edge ﬂow with turbulent boundary-layer separation. Rodi. 117:50–57. An introduction to single-point closure methodology. 472:331–340. Bradshaw. Journal of Fluid Mechanics.J. [28] B. Mechanical Engineering Dept. Prediction of the ﬂow around an airfoil using a Reynolds stress transport model.L. Scheuerer. 157:305–326. Hanjali´ and B.H. Journal of Fluid Mechanics. 68(3):537– 566. 1978. Almqvist & Wiksell. [20] S. 18:123–176. [32] J. Durbin. 52:609–638. 1978. Israel J.B. Hult.C. Simulation and Modeling of Turbulent Flows.Y. Tselepidakes.E. A Reynolds stress model of turbulence and its c application to thin shear ﬂows. u 129:547–572. Rotta. AGARD. International Journal of Heat and Fluid Flow. Journal of Fluid Mechanics. 1987. Calculation of curved shear layers with two-equation turbulence models. Statistische theorie nichthomogener turbulenz. 1975. 1973. Techn.B.E. Schaum’s Outline Series. and W.E. [19] G. Davidson.. Hussaini. Fu.R. 1951. Pope. Shavit. Wallin and A. McGraw-Hill. Modelling streamline curvature effects in explicit algebraic reynolds stress turbulence models. Journal of Fluid Mechanics. International Journal of Heat and Fluid Flow. Zeitschrift f¨ r Physik. [24] M.P. Advances in Applied Mechanics. Gatski. 169. Gibson and B. 8:259. M. 1983. Launder. Ground effects on pressure ﬂuctuations in the atmospheric boundary layer. Cartesiska tensorer: Ett praktiskt ﬁkonspr˚ k. Launder. 2002. 1996. Progress in the development of a Reynolds-stress turbulence closure. Interactions between components of the turbulent velocity correlation tensor. On the k − 3 stagnation point anomaly.

Journal of Fluid Mechanics.B. Computations of transonic ﬂow with the v2f turbulence model. Davidson. Vienna. 17(3):245–254. of Thermo and Fluid Dynamics. and K. 1991. 1994. Launder. of Applied Mechanics. [36] T. [44] A. Hanjalic.R. A Reynolds stress closure designated for complex geometries. Suga. [40] P. 2000. Menter. A new explicit algebraic Reynolds stress model for incompressible and compressible turbulent ﬂows. 18:15–28. Craft. PhD thesis. Analysis of the performance of different v 2 − f turbulence models in a stator vane passage ﬂow. 3:1–13. [47] A. Engng. Turbulence Transport Modelling in Gas Turbine Related Applications. G¨ teborg.A. J.L. 154:59–78. Johansson. 2004. 127(3):627–634. Chalmers University of Technology. and J. Two-equation eddy-viscosity turbulence models for engineering applications.R. Habziabdic. G¨ teborg. International Journal of Heat and Fluid Flow. 1993. Near-wall turbulence closure modeling without damping functions. Sveningsson and L. [38] S. Sveningsson and L.-H. 403:89–132. A new Reynolds stress algebraic equation model. o Sweden. Journal of Fluid Mechanics. Div. Craft and B. ASME TURBO EXPO 2004. 2006. Theoretical and Computational Fluid Dynamics.E. 2004. [39] P. . Comput.J. and M. [41] A. References 262 [34] T. Shih. 2001. Dept. Launder. 1997. Thesis for licentiate of engineering. 2003.A. Davidson. Durbin. 1993. Prediction of turbulent transitional phenomena with a nonlinear eddy-viscosity model. International Journal of Heat and Fluid Flow. Sveningsson. Methods Appl. [42] K. o [48] F. Application of a near-wall turbulence model to boundary layers and heat transfer. Lumley. International Journal of Heat and Fluid Flow. On explicit algebraic stress models for complex turbulent ﬂows. Speziale. 1996. International Journal of Heat and Fluid Flow. of Fluid Dynamics. 25(5):785– 794. Journal of Turbomachinery. [45] A.G. [46] A. Paper GT2004-53586. Gatski and C. Chalmers University of Technology. Sveningsson. Zhu. [37] T. 2005. B. Dept. Wallin and A. AIAA Journal. Sweden. [35] T. 22(1):53–61. Sveningsson and L. Computations of ﬂow ﬁeld and heat transfer in a stator vane passage using the v 2 − f turbulence model. Computations of ﬂow ﬁeld and heat transfer in a stator vane passage using the v 2 − f turbulence model. Davidson. 14(4):316– 323.. A robust near-wall elliptic relaxation eddy-viscosity turbulence model for CFD. Kalitzin. 1995.E. 2004. Assessment of realizability constraints in v 2 − f turbulence models.J. Popovacx. International Journal of Heat and Fluid Flow. Durbin.V. M. 125:287–302. 32:1598–1605. International Journal of Heat and Fluid Flow. [43] F-S Lien and G.

F. P. Calculation of boundary-layer development using the turbulent energy equation. and R. 10:199–228. New York. Ten years of industrial experience of the SST turbulence model. Wallingford (UK). Monthly Weather Review. of Thermo and Fluid Dynamics. 2001. Grinstein. Nagano. Fluids A. [53] F. [58] H. Koseff.html. Hinze.P. Bardina. 1995. Kolbe. [54] J. Cabot.chalmers. Germano.R. Germano. A dynamic subgrid-scale eddy viscosity model. Zang. 4:633–635. Snomass. Bradshaw. 3:1760–1765. Ferziger. Y. 1963.S. pages 624–632. [63] J. 1993. 5:3186– 3196. [61] D. 156:55–62.P. Menter.K. A proposed modiﬁcation of the Germano subgrid-scale closure method.H. and M.L. E. A dynamic mixed subgrid-scale model and its application to turbulent recirculating ﬂows. On two-equation eddy-viscosity models. Journal of Fluid Mechanics. and W.R. Oran. Phys. and N. Smagorinsky. 118:341–377. Wille. Physics of Fluids A. New York. Physics of Fluids A.H. [56] W. Reynolds. Galilean invariance of subgrid-scale stress models in the largeeddy simulation of turbulence. 7 edition. References 263 [49] D. c Turbulence Heat and Mass Transfer 4. 28:593–616. 26(11):1299–1310. 1991. Boundary-Layer Theory. and W. Hanjali´ .F. A proposal for a redeﬁnition of the turbulent stresses in the ﬁltered Navier-Stokes equations. M. Large eddy simulation of a jet in a cross-ﬂow. [51] J. Street. 1992.G. 1985. Physics of Fluids A. Bredberg. R. AIAA Journal. [50] F. Jones and M. McGraw-Hill.R. Lilly. AIAA Journal. Piomelli. Moin. Journal of Fluid Mechanics. begell house. General circulation experiments with the primitive equations.J. 7:2323–2324. D. 1979. Atwell. 2nd edition.O.C. 1986. 2003. Improved subgrid scale models for large eddy simulation.tfd. Turbulence. Numerical investigation of turbulent channel ﬂow. The Pennsylvania State University. .P. 1982. [57] J. [59] M. 30(6):1657–1659. New York. Schlichting. [62] J. inc. Moin and J.R.H. and R. [65] M.se/˜lada/allpaper. Dept. Chalmers University of Technology. Menter. and J. AIAA 80-1357. Fluid Dynamic Research. pages 4:1 – 4:6. 1992. 1992. Inﬂuence of freestream values on k − ω turbulence model predictions. McGraw-Hill. Wilcox. [52] P.L. Speziale. Journal of Fluid Mechanics. Report 01/8. U. Kim. can be downloaded from www. 91:99–165. 1980. [60] Y. Boris. Ferriss. Colorado. [55] P. In 10th Symp. In K.C. on Turbulent Shear Flows. Tummers. Langtry. 1975. New insights into large eddy simulation. editors. 1988. Kuntz. J. Reassessment of the scale-determining equation. [64] C. 1967.

Longman Scientiﬁc & Technical. McGraw-Hill. International Journal of Heat and Fluid Flow. Trans. Harlow. 2002. [73] L. Armenio. 2006. Hybrid LES-RANS: A one-equation SGS model combined with a k − ω model for predicting recirculating ﬂows. A dissipative scale-similarity model. Chalmers University of Technology. on Turbulent Shear Flow. Journal of Fluids Engineering. Patankar. International Journal for Numerical Methods in Fluids. [70] L. 43:1003–1018. Dept. Norberg. In V. Large eddy simulation: A dynamic one-equation subgrid model for three-dimensional recirculating ﬂow. L. [77] A. Hybrid LES/RANS using synthesized turbulence for forcing at the interface. Sohankar. volume 3. of the Royal Society A. 2003. 2000. Davidson.-H. [72] H. New York. International Journal of Heat and Fluid Flow. 122(1):39–47. 2009. Billson. Numerical Computation of Internal and External Flows: Fundamentals of Numerical Discretization. In 11th Int. Davidson. Krajnovi´ and L. [78] S. Dept. Malalasekera. 2002. Fr¨ hlich. 1995. A mixed one-equation subgrid model for largec eddy simulation. of Thermo and Fluid Dynamics. Large eddy simulation of ﬂow past a square cylinder: Comparison of different subgrid scale models. o 1997. Emvin. Davidson. Versteegh and W. Grenoble. PhD thesis. G¨ teborg.R. Journal of Fluids Engineering. Hirsch. volume 13 o of ERCOFTAC series. Davidson. editors. 40(5):927–936. 122(3):643. Chalmers University of Technology. Krajnovi´ and L. Geurts. 1980. [69] P. Phil. [68] L. [74] L. Davidson.V.The Finite Volume Method.K. Chichester. 2000. [76] A. The Full Multigrid Method Applied to Turbulent Flow in Ventilated Enclosures Using Structured and Unstructured Grids. Technical Report 97/11. Large eddy simulations: A note on derivation of the equations for the subgrid turbulent kinetic energies. B. 367(1899):2905–2915. 2010. . L. Numerical Heat Transfer and Fluid Flow. 27(6):1028–1042. Norberg. [67] C. Davidson. Hybrid LES-RANS: back scatter from a scale-similarity model used as forcing. and J. Erratum. and C. References 264 [66] L. Gothenburg. pages 26.1–26. pages 261–266. Davidson and M. UK. [75] L. AIAA Journal. England. 23(4):413–425. Sohankar. 1988. DLES7: Direct and Large-Eddy Simulations 7. Davidson. [79] S. An Introduction to Computational Fluid Dynamics . 1997. volume 1. Large eddy simulation of the ﬂow around a bluff c body. Symp. John Wiley & Sons. [71] S. Davidson. Peng.6. of Thermo and Fluid Dynamics. 1997. and C. Davidson and S.

In International Conference on Railway Engineering 2008 (IET ICRE2008). [91] H. Numerical simulation of the ﬂow around an ICE2 train under the c inﬂuence of a wind gust. o [89] J.R. 127(5):919–928. Inﬂuence of ﬂoor motions in wind tunnels on c the aerodynamics of road vehicles. Pallares and L. 2000. Ghent. [90] H. Large Eddy Simulation of the Flow Around a High-Lift Airfoil. Krajnovi´ . o [81] J. [92] S. [82] J. Hong Kong. 25-28 March. International Journal of Heat and Mass Transfer. 2003. Davidson. part II: Understandc ing the ﬂow. 14(8):2804–2816. Davidson. 2008. Dept. Hemida and S. Krajnovi´ . In ECCOMAS CFD 2006. Davidson. Davidson. Peng and L. 2003. Journal of Fluids Engineering. On a subgrid-scale heat ﬂux model for large eddy simulation of turbulent thermal ﬂow. 2006. Journal of Fluids Engineering. In Fourth International Conference on Computational Fluid Dynamics (ICCFD4). Krajnovi´ . Egmond aan Zee. 2005. AIAA paper 2003-0776. Chalmers University of Technology. Large eddy simulation for turbulent buoyant ﬂow in a conﬁned cavity. Dept. Davidson. 2002. 2005. 2002. [94] S. References 265 [80] S. 10-14 July. Dahlstr¨ m and L. Journal of Fluids Engineering. Krajnovi´ and L. Physics of Fluids. International Journal of Heat and Fluid Flow. Peng and L. Flow around a simpliﬁed car.-H. G¨ teborg. Journal of Wind Engineering and Industrial Aerodynamics. Ask and L. G¨ teborg. 12(11):2878–2894.-H. Flow around a simpliﬁed car. Krajnovi´ and L. Large-eddy simulations of turbulent ﬂow in a rotating square duct. Krajnovi´ . [83] S. PhD thesis. Large Eddy Simulations for Computing the Flow Around Vehic cles. [93] S. of Thermo and Fluid Dynamics. 92(5):677–696. Numerical study of the ﬂow around the busc shaped body. Davidson. 2001. 125:500–509. Sweden. of Thermo and Fluid Dynamics. Hemida and S. [88] S. Davidson. Flow and dipole source evaluation of a generic SUV. Large-eddy simulations of turbulent heat transfer in stationary and rotating square duct. 45:1393–1405. China. Krajnovi´ and L. Davidson. . 2006. 127(5):907–918. o PhD thesis. 2003. The Netherlands. Belgium. Numerical study of the unsteady ﬂow structures c around train-shaped body subjected to side winds. 132(051111). Pallares and L. Davidson. Sweden. [87] S. Journal of Fluids Engineering. Krajnovi´ and L. Physics of Fluids. Large eddy simulation applied to a higho Reynolds ﬂow around an airfoil close to stall. part I: Large eddy c simulations. 2010. 2005. 2002. Davidson. 22:323–331. LES study of the impact of the wake structures on c the aerodynamics of a simpliﬁed ICE2 train subjected to a side wind. [84] S. Chalmers University of Technology. [85] S. Dahlstr¨ m. [86] S. 5-8 September.

1998. Barhaghi and L. Davidson. Rodi. 2006. Davidson. Davidson. Detroit. Karlsson. 1993. Shih. 16:859–878.R. 2010. L. Davidson. o [96] D. In SAE Technical Paper 2000-01-0873. 10-13 September. Cokljat. J. Wu. 2003. 2003. Rodi. In J. NASA. o [98] D.-P. References 266 [95] S. J. 2007. Les of mixed convection boundary layer between radiating parallel plates. Numerical simulation of vortex shedding past triangular cylinders at high reynolds numbers using a k − ε turbulence model. editors. Barhaghi.G. pages 269– 286. 19(125106). C. Fr¨ hlich. Technical memo 105993. 2000. Modeling of Turbulent ﬂow and Heat Transfer for Building Ventilation.P. Olsson. [106] S. G¨ teborg. 27(5):811–820. of Thermo and Fluid Dynamics. Wen. Mellen. ERCOFTAC. G¨ teborg. LESFOIL: Large Eddy Simulation of Flow Around a High Lift Airfoil. [99] D. Zhu. Fr¨ hlich. In M. 41(4):573–581. Coherent motions in the turbulent boundary layer. Perzon and L. Davidson. 23:601–639. Large eddy simulations: how to evaluate resolution. Sagaut. Perzon and L. Large-eddy simulation of natural convection boundary layer on a vertical cylinder. Croatia. [104] L. 2009. and P. Heat and Mass Transfer. Lumley. Chalmers University of Technology. X. [108] S. volume 16 of Quality and Reliability of Large-Eddy Simulations II. Meyers.-H. International Journal of Heat and Fluid Flow. J. Davidson. PhD thesis. 2007. o editors. and E. 2007. M. Davidson.-B. and R. Barhaghi and L. and W. ACFD 2000. of Fluid Dynamics.-H. volume 83 of Notes on Numerical Fluid Mechanics. L. Leschziner. of Applied Mechanics. Annual Review of Fluid Mechanics.R. V. Robinson. and W. . [101] L. 1993. How to estimate the resolution of an LES of recirculating ﬂow.-J. PhD thesis. 30(5):1016–1025.G. Chalmers University of Technology. Sweden. Lessons from LESFOIL project on large o eddy simulation of ﬂow around an airfoil. AIAA Journal. Z. [103] L. Dubrovnik. Dept. Mellen. Jia. International Journal of Heat and Fluid Flow. 2006. Johansson. Zhu. In First CEAS European Air and Space Conference. Davidson. [107] S. Hu. 2000. Physics of Fluids. Bejing. [109] T. editors. [105] C.G.H. D. On transient modeling of the ﬂow around vehicles using the Reynolds equation. Peng. [97] D. Salvetti. [102] L. Davidson. A realizable Reynolds stress algebraic equation model. L. and W. Barhaghi. In 5th International Symposium on Turbulence. Div. J. B. International Journal for Numerical Methods in Fluids. 2006. A Study of Turbulent Natural Convection Boundary Layers Using Large-Eddy Simulation.A. Inlet boundary conditions for embedded LES. [100] S. Natural convection boundary layer in a 5:1 cavity. Springer Verlag. Davidson. Springer. and J. Dept. On CFD and transient ﬂow in vehicle aerodynamics. Berlin. September 25-29. Geurts. 1991. F.

[121] P. M. Abe. 32:652–669.Friederich and W. B. K. Wallingford (UK). 2004. Spalart and S. Davidson. Spalart. L. Allmaras. NV. M. Reno. H. Strelets. editors. Chakravarthy. Liu. Kuntz. A one equation turbulence model for aerodynamic ﬂows. A new turbulence model for predicting ﬂuid ﬂow and heat transfer in separating and reattaching ﬂows . Dordrecht. and S.R. Lodeﬁer. and P. Wang. Yan. 2006. Travin. Flow. Turbulence and Combustion. Merci.K. Turbulence Heat and Mass Transfer 4. and S. Dol. Menter. Interfacing statistical turbulence closures with large-eddy simulation. editors. and Y. 2003. Shur. [116] J. International Journal of Heat and Fluid Flow. 1994. International Journal of Heat and Fluid Flow. Fluid Mechanics and its Applications. Dejoan. Reno. K. [120] K. Advances in LES/DNS. Piomelli. A scale-adaptive simulation model for turbulent ﬂow prediction. In K. Batten. Journal of Fluids Engineering. [111] P. 2004. van der Ven. Louisiana Tech University. volume 65 of Euromech Colloquium 412. Fluid Mechanics and its Applications. Goldberg. [114] C. . Balaras.R. Greyden Press. First Int. 1997. M. 2011. References 267 [110] P. Heat Mass Transfer. Reno. Girimaji.R. Peng. E. 42(3):485–492. and F. H. Physical and numerical upgrades in the detached-eddy simulations of complex turbulent ﬂows. In C. NV. [119] R. Hybrid RANS-LES modelling with the renormalization group. Advances in LES of Complex Flows. Nagano. 1992.C. Pasinato. B. Jou. Nagano. [112] F. Theoretical and Computational Fluid Dynamics. 2002. 37:139–151. The innerouter layer interface in large-eddy simulations with wall-layer models. Schiestel and A. AIAA paper 2004-264.1. U. AIAA Journal. Investigation of alternative length scale substitutions in detached-eddy simulation. 24:538–550. and H. [113] A. 2003.Rodi. Academic Publishers. Kondoh. Allmaras. Advances in LES of Complex Flows. conf. and E. Int. Towards a new partially integrated transport model for coarse grid and unsteady turbulent ﬂow simulations. NV. Strelets. Comments on the feasability of LES for wings and on a hybrid RANS/LES approach. and R. Squire. on DNS/LES. 73(2):413–421. and M. [122] U. Partially-Averaged Navier-Stokes model for turbulence: A Reynolds-averaged Navier-Stokes to direct numerical simulation bridging method. T. Kok. inc. 18:443–468. and F. M. J. [118] J. c Tummers. Spalart. Extra-large eddy simulation of massively separated ﬂows. New York. pages 239–254.J.-H. Thiele. [117] S. Ma. In R. Y. A low Reynolds number variant of Partially-Averaged Navier-Stokes model for turbulence.D. Bender. 2003. AIAA paper 92-0439. S. 2005. Oskam.S. AIAA paper 2003–0767.S. and P. De Langhe. Mocket. C. pages 697–704. Dick.R. .R. begell house. 2005. W. Liu and Z.-H. Hanjali´ . Spalart. editors.R. [115] J. 74:85– 102. Flow ﬁeld calculations.

[132] A. Lindborg. N. In W. 2001. AIAA paper 2003–0079. A. Large-eddy simulation of non-equilibrium pulsed turbulent ﬂow using transport equations subgrid model. J.V. Eaton. M. Hassan. [134] P. and Trains.R. Hanjali´ . 2000. M. International Journal of Computational Fluid Dynamics. Stockholm. 2004. and K. Inﬂow boundary conditions for LES/RANS simulations with applications to shock wave boundary layer interactions. Reno. Kasagi. In E. In W.A. Engineering Turbulence Modelling and Measurements 6. [126] M. Schiestel. 28(6):1443–1459. and James Ross. and P. [133] S. A hybrid z c two-layer URANS-LES approach for large eddy simulation at high Reynolds numbers. Computers & Fluids. Edwards. Rodi and N. Hamba. Reno. editors. In Rose McCallen. Engineering Turbulence Modelling and Experiments 5. Squires. 2003. volume 19 of Lecture Notes in Applied and Computational Mechanics. Davidson and S. 21:252–263. [129] P. [124] J.R. Mulas.R. 19(6):415–427. 2000. Xiao. 2003.A. Johansson. pages 297–305. Hybrid LES-RANS: An approach to make LES o applicable at high Reynolds number. Tucker. Humphrey. An approach to hybrid RANS/LES calculation of channel ﬂow. Detached eddy simulation of massively separated ﬂows. pages 341–346. 2005. and M. [135] N. Buses. J. 26:173–190. Elsevier. and H. 2005. Physics of Fluids. Adaption of eddy-viscosity turbulence models to unsteady separated ﬂows behind vehicles. Nicoud. 16th AIAA CFD Conference. Lien. Hadˇ iadbic. Wasistho. on Turbulence and Shear Flow Phenomena. International Journal of Heat and Fluid Flow. 12(7):1629–1632. Fred Browand. International Journal of Heat and Fluid Flow. Rodi and M. Leschziner. International Journal of Heat and Fluid Flow. Yee. The Second International Symp. Nikitin. . F. Differential equation based length scales to improve DES and RANS simulations. The artiﬁcial buffer layer and the effects of forcing in hybrid LES/RANS. Leschziner. [128] F. Spalart. References 268 [123] L. AIAA paper 2003-3968. Larsson. pages 369–378. NV. 2004. and E. Fueyo. Menter and M. F. 33(2):267–287.S. Strelets. NV. Dahlstr¨ m. Davidson. [130] P. Strategies for turbulence modelling and simulations. K. Dejoan and R. Elsevier. Spalart. [127] X. [131] L. T-RANS and hybrid simulations of thermal s c convection at high RA numbers. M. The Aerodynamics of Heavy Vehicles: Trucks. Springer Verlag. Tucker and L. AIAA paper 2001–0879. 2007.R. editors. 2001. B.D. [125] F. LES. volume 2. Kenjereˇ and K. J. An approach to wall modeling in large-eddy simulations. Zonal k-l based large eddy simulation. Temmerman. Kuntz. editors. 2005. Hanjali´ . editors. 2003. Sommerfeld.

M. Modelling and Measurements. L. M. France. A low Reynolds number partiallyaveraged Navier-Stokes model for turbulence. and R. Advances and Applications in Fluid Mechanics. A scale-adaptive simulation model using twoequation models.A. Bender. 29:1638–1649. Using isotropic synthetic ﬂuctuations as inlet boundary conditions for unsteady simulations. [149] L. A new approach of zonal hybrid RANS-LES based on a twoequation k − ε model (to be presented). Davidson. In 8th International ERCOFTAC Symposium on Engineering Turbulence.R. Kuntz. [148] L. In W. 2007. Hanjali´ . Rodi and N. Shur. Evaluation of the SST-SAS model: Channel ﬂow. Partially-Averaged Navier-Stokes model for turbulence: Implementation and Validation. Munich. Egorov. Menter and Y. Reno. Turbulente Str¨ mungen. NV. Greece. 2010. [141] L. Girimaji. L. and K. 2002. S. Davidson. AIAA paper 2011-3108. A hybrid RANS-LES approach with delayed-DES and wall-modelled LES capabilities. September 5-8. [146] L. 73(2):413–421. ASME Journal of Applied Mechanics. 1972. International Journal of Heat and Fluid Flow. Ma. AIAA paper 2005–1095. In ECCOMAS CFD 2006. Peng. pages 689–694. 2007.S. G¨ ttingen. Davidson. Girimaji. Hawaii. 2012. [139] J. and F. 2011. M. AIAA-2003-0767. Peng. References 269 [136] L. Honolulu. Fueyo. 2008. Teubner Verlag. Engineering Turbulence Modelling and Experiments 5. editors. [143] S. In ETMM9: International ERCOFTAC Symposium on Turbulence Modelling and Measurements. A-priori studies of near-wall c RANS model within a hybrid LES/RANS scheme. In 6th AIAA Theoretical Fluid Mechanics Conference.K. 2006. Davidson. . Menter. Rotta. 2005. Temmerman. [145] J.R. Leschziner. Davidson and S. 1(1):1–35. Emdedded LES with PANS. Marseille.-H.R. In 5th International Symposium on Turbulence and Shear Flow Phenomena. Hybrid LES-RANS: Inlet boundary conditions for ﬂows including recirculation. Egmond aan Zee. Kh. [144] M.C. [138] F. 2006.-H. Spalart. Reno.R. Stuttgart.R. Egorov. 27-29 August. o [140] F. asymmetric diffuser and axi-symmetric hill. In 43rd AIAA Aerospace Sciences Meeting and Exhibt. Revisiting the turbulent length scale equation. 2006. [137] F. 9-11 June. [142] S. Travin. 27-30 June. Davidson. 2003. Reno. pages 317–326. A scale-adaptive simulation model for turbulent ﬂow predictions. Partially-averaged Navier-Stokes model for turbulence: A Reynolds-averaged Navier-Stokes to direct numerical simulation bridging method. 10-13 January. Wang. Thessaloniki. Menter and Y. o 2004. P. Elsevier. Strelets. Germany. volume 2. In IUTAM Symposium: One Hundred Years of Boundary Layer Research. Nevada. [147] L. 2005. The Netherlands. and A.S.

Nicoud and F. 13:2634–2649. Andersson. 2004. International Journal of Heat and Fluid Flow. [155] J. [156] L. CALC-BFC: A ﬁnite-volume code employing collocated variable arrangement and cartesian velocity components for computation of ﬂuid ﬂow and heat transfer in complex three-dimensional geometries. and L.H. In Second Symposium on Hybrid RANS-LES Methods. 2006. [158] F. C.L. of Thermo and Fluid Dynamics. John Wiley & Sons. Benhamadouche. Chalmers University of Technology. A synthetic-eddymethod for generating inﬂow conditions for large-eddy simulations. 2005. Jarrin. Billson.R. 3rd National Conference on Computational Mechanics – MekIT’05 (invited paper). Davidson. HYBRID LES-RANS: Inlet boundary conditions for ﬂows with recirculation. S. Billson.E.-E. Corfu island. AIAA paper 2003-3282. Flerdimensionell analys (in Swedish). Turbulence and Combustion. [153] L.R. Prosser. Davidson. [160] B. 62:183–200. Eriksson. Daly and F. Lund. pages 7–22. In Advances in Hybrid RANS-LES Modelling.J. Greece. Physics of Fluids. Lund. PhD thesis. Chalmers University of Technology. Studentlitteratur. Brander. Vektoranalys (in Swedish).I.E. [159] R. volume 97 of Notes on Numerical Fluid Mechanics and Multidisciplinary Design. [157] L. 2003. Springer Verlag. Flow. Transport equation in turbulence. and Mass Transfer. 2008. and R. and R. G¨ teborg. [151] M. Harlow. Davidson. Rept. New York. Fundamentals of Momentum. Davidson. References 270 [150] N. Studentlitteratur. Sweden. 9th AIAA/CEAS Aeroacoustics Conference. [154] L. John Wiley & Sons. [161] F. 2007. D. Laurence. Sweden. Norway. Computational Techniques for Turbulence Generated Noise. Hybrid LES-RANS: Inlet boundary conditions. 1995. Jet noise prediction using stochastic turbulence modeling. In B. Skallerud and H. 1984. Heat. Incompressible Flow. of Thermo and Fluid Dynamics.Wicks. 1984. 4 edition. Farhanieh. Gothenburg. Trondheim. 1970. Panton. L. 95/11. Wilson. pages 55–66. 3 edition. Welty. [162] O. Eriksson. o [152] M. Dept. Hybrid LES-RANS: Inlet boundary conditions for ﬂows with recirculation. 2007. Davidson and B. editors. Subgrid-scale stress modelling based on the square of the velocity gradient tensor. 27(4):585–593. Ducros. . Dept. 1977. Sweden. 1999. New York.

- Coelho
- Termometria e Calorimetria
- Espelhos - Esféricos e Planos
- Lista de Exercícios - Análise Dimensional
- Lista de Exercícios Circuitos Elétricos
- Lista de Exercícios - Análise Dimensional
- PAS UFLA 2010
- DALE_Os Diferentes Papeis e Resultados Dos Modelos Nacionais e Regionais de Educação (1)
- lista de exercícios física
- Lista de Exercícios - Termologia, Termometria e Dilatação
- Lista - Força e Aceleração
- PII EMBRAPA Barraginhas
- Lista Biofísica
- Lista FHE
- A Ciência Nos Meios de Comnicação de Massa
- EDITAL COMPLETO.doc
- Modelo computacional.pdf
- Vunesp 2013 Cetesb Estatistico Prova
- Código V2013 - A4
- Lista ENEM - Física
- unifenas
- Ajuste Shimano
- Port.232
- IPEM1301_306_005365
- port. 163

- Compact Heat Exchangers - Guidance for Engineers
- David Shonnard Green Engineering
- Enzymes Kinetics
- Lecture Notes on Turbulence
- DNV-RP-C203-10
- Medical Physics and Biomedical Engineering
- Stand Alone Photovoltaic Systems
- Principles of Measurement Systems
- 38465940-L-T-Piping-Engineering-3-Day-Progremme
- Adsorption Analysis
- Conduction Heat Transfer Arpaci.
- Heat Transfer Solutions - James Van Sent
- How to Build a House
- Mathematical Modeling of Transport Processes
- DuPont Refrigerant Piping Handbook(2)
- Drilling Engineering - Dipl & Prassl
- Vector Calculus
- Engineering Properties of Food
- Pipeline Design for Water Engineers
- Instrumentation and Control Course
- Physics Lab Manual for engineering students
- Fluid Mechanics
- Desing Formulas for Plastics Engineers
- Electrical Applications for Air Conditioning Refrigeration Systems
- Advances in Thermal Design of Heat Ex Changers
- Fundamentals of POLYMER ENGINEERING
- Introduction to Rotating Equipment Maintenance
- HOW TO BUILD A SHIPPING CONTAINER HOUSE
- Solids and Fluids Turbulent Flow Turbulence Modelling

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd