BestMasters

Springer awards BestMasters to the best masters theses which have been completed at renowned universities in Germany, Austria, and Switzerland.
The studies received highest marks and were recommended for publication by
supervisors. They address current issues from various fields of research in natural
sciences, psychology, technology, and economics.
The series addresses practitioners as well as scientists and, in particular, offers guidance for early stage researchers.

David Roos Launchbury

Unsteady Turbulent
Flow Modelling and
Applications

David Roos Launchbury

Horw, Switzerland

BestMasters
ISBN 978-3-658-11911-9
ISBN 978-3-658-11912-6 (eBook)
DOI 10.1007/978-3-658-11912-6
Library of Congress Control Number: 2015954647
Springer Vieweg
Springer Fachmedien Wiesbaden 2016
This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or
part of the material is concerned, specically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microlms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by
similar or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specic statement, that such names are
exempt from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this
book are believed to be true and accurate at the date of publication. Neither the publisher nor the
authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper
Springer Vieweg is a brand of Springer Fachmedien Wiesbaden
Springer Fachmedien Wiesbaden is part of Springer Science+Business Media
(www.springer.com)

Acknowledgments

My rst thanks go to my supervisors Dr. Luca Mangani and Dr. Ernesto

Casartelli for supporting me throughout this thesis and for giving me carte
blanche on themes and procedure. I would also like to thank the many other
professors who gave me the knowledge that I have today. Most notably,
Id like to thank Dr. Thomas Staubli and Dr. Andreas Haselbacher for
their inspiration and their ability to motivate a passion for the subject
matter, as well as Thomas Tresch for sparking my interest in uid dynamics
in the rst place and for being a good friend. Further thanks go to Dr.
Giulio Romanelli for the many fruitful discussions and his patience when
answering my endless array of questions.
I also thank my colleagues Oliver Ryan and Simon Roth for their company during my studies and for making the long evening hours much more
interesting.
I would like to express my gratitude to my parents Robert and Eva
for the moral and nancial support over the course of my entire education
and for the patience they showed when it came to me nding my way. Im
also very grateful for the support of my parents-in-law Bryan and Maggie,
especially for their endless hours of babysitting.
Finally, and most importantly, I would like to thank my wife Susan and
my boys Andreas and William for putting up with me during my studies
and for being the most important things in my life.

Abstract

The present work deals with the improvement of a previously developed

explicit third-order time-accurate solver for transient ow problems implemented in OpenFOAM. As such, the solver is enabled to solve the spatially
ltered Navier-Stokes equations applied in large eddy simulations.
An optimised pressure-velocity coupling algorithm is implemented to
reduce pressure oscillations at small time steps. The support for a temperature transport equation is added with the aim of solving problems
involving heat transfer in incompressible ows. Momentum and temperature source terms are added to allow for periodic boundary conditions in
such cases.
The solver is validated on a series of test cases involving the ow between
parallel plates and around a square cylinder. The ow over a turbulator
geometry involving heated walls is investigated, as well as a jet-in-crossow
setup of a lm cooling case. In all these cases, the performance of the
static Smagorinsky, dynamic Lagrangian and dynamic one-equation turbulence models available in OpenFOAM are assessed. Additional turbulence
models (dynamic Smagorinsky and WALE), implemented by OpenFOAM
community members, are adapted for incompressible ows and tested as
well. In addition to this, the previously unavailable sigma-model was implemented in this work. Simulations without using any turbulence models,
ie. under-resolved DNS (UDNS) simulations, were performed for comparison. Very good results were obtained in all cases with variations among
the individual models. The no-model simulations performed surprisingly
well and occasionally better than some of the models.
The parallel performance of the solver is tested on a computational
cluster and the experiences gained during the course of this work are summarised as recommendations for future use.

Contents

Nomenclature

XI

List of Figures

XIII

List of Tables

XV

1 Introduction

2 Large Eddy Simulation

3 Subgrid Models
3.1 Smagorinsky Model . . . . . . .
3.2 Dynamic Smagorinsky Model .
3.3 Dynamic Lagrangian Model . .
3.4 Dynamic One-Equation Model
3.5 WALE Model . . . . . . . . . .
3.6 Sigma Model . . . . . . . . . .

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

7
8
9
10
10
11
12

4 Solver
4.1 Basics . . . . . . . . . . . . . . . . . . . .
4.2 Improvements . . . . . . . . . . . . . . . .
4.2.1 Pressure-Velocity Coupling . . . .
4.2.2 Turbulence Models . . . . . . . . .
4.2.3 Momentum Source Term . . . . . .
4.2.4 Energy Equation and Source Term

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

15
15
16
16
18
19
20

5 Validation
5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 Periodic Channel Flow . . . . . . . . . . . . . . . . . . . . .

23
23
24

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

.
.
.
.
.
.

5.3

5.4

5.5

5.6

5.2.1 Description . . . .
5.2.2 Setup . . . . . . .
5.2.3 Mesh . . . . . . . .
5.2.4 Results . . . . . .
Square Cylinder . . . . . .
5.3.1 Description . . . .
5.3.2 Setup . . . . . . .
5.3.3 Mesh . . . . . . . .
5.3.4 Results . . . . . .
Ribbed Channel . . . . .
5.4.1 Description . . . .
5.4.2 Setup . . . . . . .
5.4.3 Mesh . . . . . . . .
5.4.4 Results . . . . . .
Film Cooling . . . . . . .
5.5.1 Description . . . .
5.5.2 Setup . . . . . . .
5.5.3 Mesh . . . . . . . .
5.5.4 Results . . . . . .
Comments on the Results
5.6.1 No-Model Results
5.6.2 Grid Resolution . .

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

6 Parallel Performance
7 Recommendations for LES Simulations
7.1 Discretisation Schemes . . . . . . . . . .
7.2 Linear Solvers . . . . . . . . . . . . . . .
7.3 Initial Values . . . . . . . . . . . . . . .
7.4 Grid Resolution . . . . . . . . . . . . . .
7.5 Mesh Quality . . . . . . . . . . . . . . .

24
25
26
28
37
37
37
39
40
43
43
43
45
46
50
50
50
52
54
57
57
57
59

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

63
63
64
64
65
66

8 Conclusions

69

9 Unresolved Issues

71

10 Outlook

73

A Appendix: Recent Developments

83

Nomenclature

Symbol
A
c
C
CF L
D
e
E
f, F
h, H
I
IQ
k
l
L
M
Nu
OP
p
phi
Pr
q
R
Re
S

Description
Area
Heat capacity
Generic coecient
Courant-Friedrichs-Lewy stability number
Diameter
Reference length
Eciency
Force or frequency
Reference length
Momentum ux ratio or invariant
Quality index
Thermal conductivity or kinetic energy
Reference length
Operator in subgrid models
Blowing Rate, operator in subgrid models
or modelling amount
Nusselt number
Generic operator
Pressure
Face ux
Prandtl number
Heat ux
Generic ratio
Reynolds number
Rate of strain tensor or parallel speedup

St
t
T
u
x, y, z

Strouhal number
Time
Temperature or elapsed time
Velocity
Spatial coordinates

Greeks

Angle or thermal diusivity

Pressure drop per unit length
Temperature source
Kronecker delta symbol
Filter width
Adiabatic eectiveness
Kinematic viscosity
Density
Singular value
Shear stress, stress tensor

Subscripts
,b
c
cva
d
l
t

v
w

Bulk or fareld quantity

Coolant
Cell volume weighted average
Drag
Lift
Turbulent
Shear stress
Viscous
Wall

Superscripts

XII

Fluctuation
Averaged or ltered quantity
Normalised quantity

List of Figures

4.2.1
4.2.2

Temporal Discretisation Error . . . . . . . . . . . . . . . . .

Reduction of Pressure Oscillations . . . . . . . . . . . . . .

18
18

5.2.1
5.2.2
5.2.3
5.2.4
5.2.5
5.2.6
5.2.7
5.2.8
5.2.9
5.2.10
5.2.11
5.2.12
5.2.13
5.2.14
5.2.15
5.2.16
5.2.17
5.2.18
5.2.19
5.2.20
5.2.21
5.2.22
5.2.23
5.2.24

Channel Geometry Overview . . . . . . . . . . .

Channel Mesh Fine . . . . . . . . . . . . . . . . .
Channel Mesh Coarse . . . . . . . . . . . . . . .
Mean Velocity Proles, Smagorinsky Models . . .
Mean Velocity Proles, n-Equation Models . . . .
Mean Velocity Proles, Static Models . . . . . . .
Mean Velocity Proles, No Model . . . . . . . . .
Mean Velocity Fluctuations, Smagorinsky Models
Mean Velocity Fluctuations, n-Equation Models .
Mean Velocity Fluctuations, Static Models . . . .
Mean Velocity Fluctuations, No Model . . . . . .
Mean Shear Stress, Smagorinsky Models . . . . .
Mean Shear Stress, n-Equation Models . . . . . .
Mean Shear Stress, Static Models . . . . . . . . .
Mean Shear Stress, No Model . . . . . . . . . . .
Mean Velocity Proles, Smagorinsky Models . . .
Mean Velocity Proles, n-Equation Models . . . .
Mean Velocity Proles, Static Models . . . . . . .
Mean Velocity Proles, No Model . . . . . . . . .
Mean Velocity Fluctuations, Smagorinsky Models
Mean Velocity Fluctuations, n-Equation Models .
Mean Velocity Fluctuations, Static Models . . . .
Mean Velocity Fluctuations, No Model . . . . . .
Mean Shear Stress, Smagorinsky Models . . . . .

24
27
27
30
30
30
30
31
31
31
31
32
32
32
32
34
34
34
34
35
35
35
35
36

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

5.2.25
5.2.26
5.2.27
5.3.1
5.3.2
5.3.3
5.3.4
5.3.5
5.3.6
5.4.1
5.4.2
5.4.3
5.4.4
5.4.5
5.4.6
5.4.7
5.5.1
5.5.2
5.5.3
5.5.4
5.5.5
5.5.6
5.5.7
5.5.8
5.5.9

Mean Shear Stress, n-Equation Models . . . . . . .

Mean Shear Stress, Static Models . . . . . . . . . .
Mean Shear Stress, No Model . . . . . . . . . . . .
Channel Geometry Overview . . . . . . . . . . . .
Square Cylinder Mesh . . . . . . . . . . . . . . . .
Averaged Velocity Prole, Smagorinsky Models . .
Averaged Velocity Prole, n-Equation Models . . .
Averaged Velocity Prole, Static Models . . . . . .
Averaged Velocity Prole, No Model . . . . . . . . .
Ribbed Channel Geometry Overview . . . . . . . .
Ribbed Channel Mesh . . . . . . . . . . . . . . . .
Nusselt Number Distribution, Smagorinsky Models
Nusselt Number Distribution, n-Equation Models .
Nusselt Number Distribution, Static Models . . . .
Nusselt Number Distribution, No Model . . . . . .
Averaged Streamlines . . . . . . . . . . . . . . . . .
Film Cooling Geometry Overview . . . . . . . . . .
Film Cooling Mesh, Coarse, Regular . . . . . . . .
Film Cooling Mesh, Medium, Regular . . . . . . .
Cooling Hole Mesh, Fine, Smoothed . . . . . . . .
Centerline Adiabatic Eectiveness . . . . . . . . . .
Laterally Averaged Adiabatic Eectiveness . . . . .
Adiabatic Eectiveness at x/D = 6 . . . . . . . . .
Adiabatic Eectiveness at x/D = 15 . . . . . . . . .
Horseshoe Vortex Around Cooling Hole Exit . . .

6.0.1
6.0.2
6.0.3

Speedup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Parallel Eciency . . . . . . . . . . . . . . . . . . . . . . . . 61
Time Reduction Factor When Doubling the Number of Cores 62

10.0.1
10.0.2
10.0.3
10.0.4

Mean Velocity Proles, IDDES . . .

Mean Velocity Fluctuations, IDDES
Mean Shear Stress, IDDES . . . . .
Nusselt Number, IDDES . . . . . .

.
.
.
.

74
74
75
75

A.1
A.2

Adiabatic Eectiveness at x/D = 6 . . . . . . . . . . . . . .

Adiabatic Eectiveness at x/D = 15 . . . . . . . . . . . . . .

84
84

XIV

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

36
36
36
37
39
41
41
41
41
43
46
47
47
47
47
48
51
53
53
54
55
55
55
55
56

List of Tables

5.2.1
5.2.2
5.2.3
5.2.4
5.2.5
5.3.1
5.3.2
5.4.1
5.4.2
5.4.3
5.5.1
5.5.2

Dimensionless Mesh Sizes of DNS Simulation . . . . . . . . .

Boundary Conditions for Periodic Channel Flow . . . . . . .
Dimensionless Mesh Sizes of LES Simulation . . . . . . . . .
Friction Velocity, Coarse Mesh . . . . . . . . . . . . . . . . .
Friction Velocity, Fine Mesh . . . . . . . . . . . . . . . . . .
Square Cylinder Boundary Conditions . . . . . . . . . . . .
Comparison of Aerodynamic Data . . . . . . . . . . . . . . .
Ribbed Channel Boundary Conditions . . . . . . . . . . . .
Physical and Geometrical Data used in Ribbed Channel Case
Reattachment Lengths . . . . . . . . . . . . . . . . . . . . .
Film Cooling Setup Parameters . . . . . . . . . . . . . . . .
Film Cooling Boundary Conditions . . . . . . . . . . . . . .

25
26
26
29
33
38
42
44
45
49
51
52

6.0.1

Execution Times of Parallel Runs . . . . . . . . . . . . . . .

60

7.1.1

Discretisation Schemes . . . . . . . . . . . . . . . . . . . . .

63

1. Introduction

One of the major limitations when performing uid dynamics simulation

has been, and will always be, the available resources to do such calculations.
In recent years the computational power of computers and the availability
of large parallel clusters have drastically increased and with that, simulations have become more and more complex. This allows for more detailed
studies of physical phenomena in challenging environments and geometries, but only if the software tools are enhanced along with the hardware
improvements. The focus of this work lies on improving a previously developed solver (see [19]) to be used for the simulation of highly turbulent
ows using a method known as large eddy simulation (LES).
Large eddy simulation has proved to perform very well in cases where
turbulence is dominant, especially when the turbulent structures are anisotropic. The solver is therefore validated using cases where other, cheaper
solution approaches for the Navier-Stokes equations, such as RANS (see
below), have failed to produce accurate results. The simulation of cases
involving heat transfer in incompressible ow will also be a part of this investigation, both for heat-transfer-enhancing geometries (turbulators) and
for applications involving lm cooling.
Since all these calculations require considerable computational power,
the performance on a parallel cluster will also be investigated. Furthermore,
quality aspects of computational grids and applicable numerical schemes
will be treated as part of this work.
In the next chapters, the LES method itself and the models used in this
study will be presented. The improvements implemented in the solver are
explained in the chapters following the theory part.

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_1, Springer Fachmedien Wiesbaden 2016

2. Large Eddy Simulation

The behaviour of uids can be described by the well-known mathematical

model known as the Navier-Stokes equations. The original equations include formulations for the conservation of momentum, energy and mass,
therefore leading to three momentum equations, one energy equation and
one continuity equation. The form presented below is a simplication of
these equations for incompressible ows and a constant viscosity. For a
detailed derivation of the Navier-Stokes equations as well as the simplications applied for incompressibility, many textbooks on uid dynamics are
available, eg. [12], [22] or [5].
ui
=0
xi

(2.1)

ui
(ui uj )
1 p
2 ui
+
= fi
+
t
xj
xi
xj xj

(2.2)

Equation 2.1 shows the continuity equation and equation 2.2 shows the
vectorial equation for momentum conservation. Here ui is the velocity in
direction i (xi would be the corresponding spatial direction), fi are body
forces (eg. gravity), is the uid density, p the pressure and the laminar
kinematic viscosity.
In turbulent ow simulations, the viscosity is often modied by turbulence models in the way that ef f = + t where the laminar viscosity
acting in the momentum equations will be replaced ef f (t is a quantity
calculated by the turbulence model and is called turbulent or eddy viscosity). In consequence the value of ef f will vary in space, and this requires
that a part of the diusion term of the momentum equations that was zero
in the above equations now has to be retained (see [12] for a more detailed

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_2, Springer Fachmedien Wiesbaden 2016

explanation). The momentum equation now takes the following form:




ui
ui
(ui uj )
1 p

uj
ef f
(2.3)
+
= fi
+
+
t
xj
xi
xj
xj
xi
One of the main diculties when simulating ows is the presence of turbulence as it appears in a very broad spectrum of time and length scales. A
proper simulation where all scales are fully resolved is known as DNS (direct
numerical simulation). The computional eort to perform such simulations
scales approximately with the Reynolds number cubed (see eg. [12]) and
is currently only applicable to academic test cases at comparatively low
Reynolds numbers and simple geometries.
A lot of eort has been put into developing formulations that model
the behaviour of turbulence, therefore reducing the spatial and temporal
resolution required to obtain a solution. The above equations are valid for
incompressible ows with a spatially varying viscosity and form the basis
for further simplications of the Navier-Stokes Equations that allow the
modelling of turbulence. One of these simplications is the decomposition
of the solution variables into an average and a uctuating quantity, a procedure rst described by Reynolds [36] leading to the Reynolds-Averaged
Navier-Stokes (RANS) equations, shown below in equations 2.4 and 2.5.
Their derivation and further explanations are omitted here.
u
i
=0
xi

j )
(
ui u
u
i
1 p

+
= fi
+
t
xj
xi
xj

(2.4)



 
 u u
i j
u
i
u
j
+
(2.5)

xj
xi
xj

Here, the additional term ui uj , called the Reynolds stress tensor, needs
to be modeled. This procedure has found widespread acceptance, especially in the industry, as it allows the calculation of stationary and simplied (eg. symmetric) solutions of the Navier-Stokes equations at greatly
reduced computational costs compared to DNS. The main drawback of this
method is that the uctuating quantities of the solution variables are entirely calculated by turbulence models. Simulations where turbulent eects
are dominant can often not be accurately represented by RANS turbulence
4

models. Or it can be the case that a model performs very well in some
cases but fails in others.
Another approach of dealing with turbulence is the so-called Large Eddy
Simulation (LES), which is the main topic of this work. In contrast to the
RANS method, the Navier-Stokes equations are not averaged, but ltered.
A generic lter function (shown here in one-dimensional notation)

u
i (x) = G(x, x )ui (x )dx
(2.6)
is applied to the solution variables, where G(x, x ) is the lter kernel (see
[12]). The ltered Navier-Stokes equation then take the following form:
u
i
=0
xi
u
i
j )
(
ui u
1 p

+
= fi
+
t
xj
xi
xj

 

r
ij
u
i
u
j
+

xj
xi
xj

(2.7)

(2.8)

Here u
is the ltered velocity, ie. the new solution variable. It can be
seen that the continuity equation, due to its linearity, does not change.
Similar however, to the Reynolds stresses in the RANS equations above,
r
, can be found on the
an additional term, the residual stress tensor ij
right hand side. This term appears due to the fact that when ltering the
equations, the nonlinear convection term is actually
(ui uj )
.
xj

(2.9)

Since term ui uj is not easily calculated, Leonard [20] suggested the split
r
+u
i u
j , therefore resulting in the nal ltered forms given
into ui uj = ij
r
above in equations 2.7 and 2.8. The residual stress tensor ij
is unknown
and needs to be modeled by so-called subgrid or LES turbulence models.

3. Subgrid Models

The following sections describe the individual models used in the context of
this work and how they approximate these subgrid stresses. First and foremost, all models shown here are based on the eddy-viscosity assumption,
ie. it is assumed that the eects of the subgrid stresses cause increased
transport and dissipation and can therefore be approximated (Boussinesqapproximation, see [12]) by increasing the laminar viscosity by a turbulent
counterpart. The following approximation is used:


u
i
1 r
u
j
r
+
(3.1)
= 2t Sij
ij kk ij = t
3
xj
xi
Here the left hand terms form the deviatoric part of the subgrid stress
r
, t is the turbulent viscosity and Sij is once again the strain rate
tensor ij
tensor of the resolved eld. The task of the turbulence model is then to
approximate the value of this turbulent viscosity or eddy viscosity.
The models presented below all have the following form when calculating the turbulent viscosity:
t = (C)2 OP

(3.2)

where C is a model constant (or a combination of constants), is the lter

width and OP is a generic operator.
All of the models can be classied into dierent groups. The rst distinction can be made based on how the constant C is treated. If the constant is
an actual constant and remains the same over the course of the simulation
then the models are called static models. In contrast to this, dynamic models locally calculate the constant, usually by re-ltering the resolved eld
using a lter width of 2 and comparing the dierently resolved elds. A

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_3, Springer Fachmedien Wiesbaden 2016

detailed explanation of this technique is omitted here, but can be found in

eg. [8]. These new values for C can then be locally or globally averaged or
clipped to avoid unphysical results causing instabilities (such as negative
eective viscosities).
The models can be further distinguished by looking at the operator
OP . Algebraic models simply solve an algebraic equation based on known
quantities to calculate t . N-equation models on the other hand solve
additional transport equations, eg. for the turbulent kinetic energy, inside
the operator. The lter width mentioned above is a measure for the grid
size and in most cases calculated as the cube root of the cell volume. All
models presented below are in incompressible form therefore the density
was generally dropped.
3.1

Smagorinsky Model

The Smagorinsky model was among the rst models created for the purpose
of calculating subgrid eddy viscosities, created in 1963 by Joseph Smagorinsky [41]. It is the most well-known and widely-used model and still applied today. It belongs to the category of static models where the C is an
actual constant. Very good results can be obtained using this model when
treating external detached ows and uniform grid turbulence.
The operator OP is simply dened as:

OP = |S|

(3.3)

= (2Sij Sij ) 12 .
where |S|
The main issue with the Smagorinsky model is that the constant C
is not a universal constant but highly problem-dependent and needs to
be calibrated to the case. This is especially problematic in the near-wall
regions where the contribution of the turbulent viscosity should tend to
zero. In these regions the Smagorinsky model drastically overestimates the
turbulent viscosity.

One way around this problem is to apply a damping function to the

constant C such as one suggested by van Driest [10] (originally used for
RANS models):


+
+ 2
C = C0 1 ey /A
yu

u = w

y+ =

(3.4)
(3.5)
(3.6)

Here C0 is the desired coecient in the main ow, y + is the dimensionless wall distance and u is the friction velocity dened as the square
root of the wall shear stress w . The static, algebraic Smagorinsky model
is provided as a built-in model in OpenFOAM.
3.2

Dynamic Smagorinsky Model

The dynamic version of the Smagorinsky model uses the same operator
OP as above, but the coecient C is calculated dynamically in each cell
by re-ltering the resolved velocity eld and comparing the results. More
on the exact procedure can be found in [8].
The coecient is given by
C=

1 Lik Mik
2 Mik Mik

(3.7)

 2  

 
k and Mik is given as
i u
ui u
k + u
where Lik is equal to
S  Sik

 
S Sik . The double overline here indicates re-ltering using the test lter

width .
The value for C can reach very large negative values and in order to
stabilise the solution the values are usually averaged, either over the domain volume or planes parallel to the ow. In complex geometries the
plane averaging is not very meaningful and the averaging over the entire
domain only properly works for homogenous turbulence. Another method
is to average the values in a locally dened volume around the cell. The
implementation used in this work does not use any averaging but clips the
value for t at , therefore producing zero viscosity at minimum.
9

This model was implemented by Alberto Passalacqua, an OpenFOAM

community member [33], for compressible ow and was adapted here to an
incompressible version.
3.3

Dynamic Lagrangian Model

This dynamic model was developed by Meneveau et al. [23] to improve

the dynamic Smagorinsky models for ows in complex geometries. The
averaging procedure of the dynamic Smagorinsky model was considered
unsuitable for non-homogenous turbulence as the methods described in the
previous section all require either homogenous directions (planes) or an
arbitrary averaging volume. The idea was then to use lagrangian averaging
along the streamlines of the ow, therefore being able to incorporate a uid
particles past trajectory. A short description of the method can be found
in [44].
The operator OP is again the same as in the Smagorinsky model, but
the coecient C is calculated by solving two additional transport equations
(not given here). The quantities calculated there have units of [m4 /s4 ]
and require boundary and initial conditions just like other solution variables. Therein lies the diculty when applying this turbulence model to
non-periodic ow domains: the estimation of the inlet conditions for these
variables is relatively arbitrary.
3.4

Dynamic One-Equation Model

The dynamic one-equation model in OpenFOAM solves a transport equation for the subgrid kinetic energy. The Smagorinsky models assumed that
a local equilibrium exists between the transferred energy from the subgrid
scales and the dissipated kinetic energy. If this is not the case, solving an
additional transport equation for the subgrid kinetic energy can improve
the results. Unfortunately the exact model used in OpenFOAM is not referenced and due to the large number of one-equation models it is unclear
which one was actually implemented. For this reason, the equations shown
here and some of the explanations are taken from the Fluent manual [2],
which describes the model developed by Kim and Menon [17], as well as
the OpenFOAM source code itself.
10

The operator OP for the calculation of the turbulent viscosity t is

given by

(3.8)
OP = k
Here k is the subgrid kinetic energy dened as:

1 2
(3.9)
uk u
2k
k=
2
where u is the resolved velocity and the overbar denotes ltering. The constant C is calculated using a dynamic procedure. The transport equation
solved to evaluate k is given as follows:


3
k u
u
j
k( 2 )
k
j k

t
(3.10)
+
+
= ij
C
t
xj
xj

xj
xj
The additional constant C contained in the k-equation is also calculated
dynamically in OpenFOAM. Just as the dynamic Lagrangian model, the
one-equation model is available built-in to the standard distribution of
OpenFOAM.
3.5

WALE Model

The wall-adapting local eddy-viscosity (WALE) model was developed by

Nicoud and Ducros [29] as a static, algebraic subgrid model that is able
to reproduce the correct near-wall behaviour without the use of articial
damping functions (such as the van Driest damping described in the previous sections).
The model constant C is considered a true constant and a valueof 0.5 is
mostly used (the original paper suggests a value of C calculated as 10.6Cs
where Cs is the Smagorinsky constant).
The operator OP is based on the traceless symmetric part of the square
d
. The velocity gradient tensor is given as
of the velocity gradient tensor Sij
u
i
(3.11)
xj
The overbar indicates that the resolved velocity is used for the gradient
d
mentioned above is then calculated as
construction. The quantity Sij
gij =

d
=
Sij

1
1 2
2
2
g + gji
ij gkk
2 ij
3

(3.12)
11

The operator OP of the WALE model is then constructed as

 d d 3/2
Sij Sij
OP = 
5/2  d d 5/4
Sij Sij
+ Sij Sij

(3.13)

d
where Sij
is given above and Sij is the rate-of-strain tensor.
This model is widely used in commercial software and known for its
good performance in transitional turbulent ows (an example is given in
the original paper [29]). The OpenFOAM formulation of the operator was
given by Cosimo Bianchini, an OpenFOAM community member, [4] and
was completed here to a fully functional turbulence model.

3.6

Sigma Model

The sigma model was originally developed by Nicoud et al. [30] and is
not available in OpenFOAM. It was implemented during the course of this
work. The model is the spiritual successor of the previously described
WALE model and has the same positive properties with some new ones
added. The list of desirable properties for the operator OP that were used
in the derivation in [30] is the following:
1. OP needs to be a positive quantity which involves only locally dened
velocity gradients.
2. OP needs to follow cubic behaviour near solid walls (ie. tend cubically
to zero).
3. OP needs to be zero for two-component or two-dimensional ows
4. OP needs to be zero for axisymmetric or isotropic expansion/contraction
The full derivation is omitted here, but as a nal result the resulting
operator was chosen to be
OP =

3 (1 2 )(2 3 )
12

(3.14)

The constant C used in this model is considered to be a true constant

and the suggested value for it is 1.5.
12

In equation 3.14 above, the -values are the singular values of the velocity gradient tensor. They are ordered in such a way that 1 2 3 0.
This denition ensures that the rst property mentioned above is met in
the sense that OP will always be positive. This has the eect that the
turbulent viscosity can never be negative and can therefore never lead to
unphysical solutions and instabilities.
The second property is fullled as well since the second invariant of the
velocity gradient tensor is quadratic in y, the wall normal direction. This
fact is simply stated here without further explanation. A full discussion on
the derivation of the values can be found in the original paper [30] and
the numerical calculation of the values are given below.
The third property is given, as in the case of two-dimensional ows
the smallest singular value 3 will always be zero. The fourth property is
fullled as well, but is not as obvious as the previous one.
The implementation of this turbulence model is straightforward with
the only diculty being the calculation of the singular values of the velocity gradient tensor. OpenFOAM already oers a method to calculate the
eigenvalues of tensors, from which the singular values can be obtained by
taking the square root. During this work this built-in method has however
failed several times causing oating point exceptions and stopping the simulation from running. This is most likely due to incomplete boundedness
checks inside the algorithm. Therefore the following method, also suggested by [30], was implemented and boundedness of certain critical operations
was enforced.
First, the velocity gradient tensor
 g is constructed
for every cell. Then
the invariants I1 = tr(g), I2 = 12 tr(g)2 tr(g 2 ) and I3 = det(g) are
calculated, where tr denotes the trace and det the determinant of the tensor.
I2
I3
2
Then the angles 1 = 91 I32 , 2 = 271 I16I2 + I23 and 3 = 13 arccos 3/2
1

are built. During this step it was ensured that the value of 1 was positive
and non-zero and that the argument of arccos remained bounded between
-1 and 1.
13

From that the singular values can be computed as follows:




I1
1 =
+ 2 1 cos3
3





I1
2 =
2 1 cos
+ 3
3
3





I1
2 1 cos
3
3 =
3
3

(3.15)

(3.16)

(3.17)

Here the arguments of the square root were forced to be non-negative

but allowed to have a zero value.

14

4. Solver

4.1

Basics

In the previous work [19] a new solver for time-dependent incompressible

ows was developed. It uses an explicit, third-order accurate time integration method based on the third-order Runge-Kutta scheme. It also oers
the possibility of treating the viscous terms implicitly by applying a hybrid
time integration scheme consisting of the Runge-Kutta and the CrankNicolson method (see [14]). It was shown that the solver is able to reach
the predicted orders of accuracy by performing an order-verication study
using the method of manufactured solution [38]. The full details of the
numerical schemes as well as the verication procedure and results can be
found in [19].
At the time the solver was limited in the sense that it did not incorporate any kind of turbulence modelling, restricting its use for practical
ows. For this work the solver was considerably improved by adding a new
procedure for pressure-velocity coupling as well as including the support
for turbulent ows. In order to allow for the simulation of periodic geometries (ie. channel ows, ribs, bumps etc.) a source term for the momentum
equations was included. An energy equation was also added for simulations
involving the transport of a passive temperature scalar (eg. incompressible lm cooling). Again, a source term is added to this equation to allow
periodic simulations. The following sections provide the details of these
additions.

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_4, Springer Fachmedien Wiesbaden 2016

4.2

Improvements

4.2.1. Pressure-Velocity Coupling

The new solver adds a selectable option to treat the pressure-velocity coupling in the manner proposed by Rhie and Chow [37].
The algorithm used at every substep in the previous version can be
summarised as follows (note that in OpenFOAM the convection term is
split into the velocity variable u and a ux eld phi, see eg. [15]):
1. Setup momentum equation, collecting all the terms (convection, diffusion etc.) multiplied by the Runge-Kutta weights. Since the solver
is explicit, everything but the temporal term goes to the right hand
side RHS.
2. Create a vector eld from the timestep, the right hand side and the
current solution: HbyA = A1p RHS +U . Ap is the diagonal coecient
of the matrix and in case of an explicit solver simply equal to t1 .
3. Interpolate this vector eld onto the cell faces and calculate the dot
product with the face normal Sf to create a ux eld phiHbyA.
4. Solve pressure equation lap( A1p p) = div(phiHbyA), where lap is the
discrete laplacian operator and div the discrete divergence.
5. Update the ux eld phi with the interpolated ux eld from above
and subtract the face ux eld from the pressure matrix.
6. Update the existing velocity eld by adding A1p (RHS grad(p)),
where grad(p) is the discrete gradient of the new pressure eld.
7. Update boundary conditions and proceed to the next substep.
The new algorithm uses the same Runge-Kutta framework as the old
one, but the steps performed in the ux calculation are dierent. The
algorithm is formulated in such a way that it solves for a pressure correction
as opposed to the actual pressure as before. The steps are as follows:
1. Setup momentum equation. This step is identical to the one before.
2. The diagonal coecient
16

1
Ap

is interpolated to the faces to create

1
Ap .

3. A unit-length vector eld ed , pointing from the owner cell center to

the neighbour cell center, is created.
4. The discrete gradient of the current pressure eld is interpolated to
the cell faces to create grad(p).
5. A new pressure gradient at the face is created: gradpf = grad(p)
grad(p)proj + gradpsn . Here grad(p)proj is the interpolated gradient
projected onto ed and gradpsn is the uncorrected (central dierencing) gradient normal to the cell face.
6. The current velocity eld is interpolated
  onto the faces (u)
 and the
1
ux eld is then: phi = u Sf
Sf +
Ap gradpf grad(p)

(phi0 u0 ). Here Sf is the cell face normal with length of the face
area and phi0 and u0 are values from previous substeps. This last
term is only present in the second and third substep.

7. Solve pressure correction equation lap( A1p pcorr ) = div(phi).

8. Update phi with the face ux eld from the pressure correction equation.
9. Update pressure with pressure correction and update boundary conditions.
10. Update velocity with the gradient of the pressure correction.
The order verication performed in [19] was repeated using the new
method and it can be shown that the method performs very well and
achieves the projected third-order accuracy in time. Figure 4.2.1 shows
the evolution of the error with decreasing time steps. The error plateau
seen at lower timesteps is due to the spatial discretisation.
The following gures show the improvement in reducing pressure oscillations when using the new formulation compared to the old one. The
gures were taken during the initial stages of the ow around a sharp edge.
The case on the left in gure 4.2.2 uses the old formulation and the one on
the right uses the new one.
17

Maximum Error L [m/s]

105

Old Algorithm
New Algorithm
Third-Order Slope

106
107
108
109

104

103
t [s]

Figure 4.2.1: Temporal Discretisation Error

Figure 4.2.2: Reduction of Pressure Oscillations

4.2.2. Turbulence Models

General support for turbulence models was added in this step. This means
that the solver works with any of the selectable turbulence models provided
in OpenFOAM. The solver updates the turbulence elds between every
substep of the Runge-Kutta scheme. This has one drawback: Since the
18

third-order time scheme is hardcoded in the solver, non-algebraic turbulence models can only be solved in second-order temporal accuracy. Only
algebraic models can fully utilise the temporal accuracy of the solver. The
error in the n-equation turbulence models will however be lower than when
used in the default OpenFOAM solvers such as pisoFoam. This is due to
the fact that the turbulence equations are solved multiple times in between
every full timestep. It will be shown in the validation section (section 5)
that n-equation models do not perform any worse compared to algebraic
models. They do, however, require more computational eort.
4.2.3. Momentum Source Term
When simulating ows through channels and ducts periodic boundary conditions are usually applied in order to reduce the size of the computational
domain. The periodicity condition for the velocity is easily imposed by
simply treating the rst and last cells in the domain as if they were direct
neighbours. Periodic velocity is also physically valid in incompressible cases
since the mass ow through the domain is always conserved. The pressure
however drops from inlet to outlet due to wall friction and other energy
losses and cannot simply be imposed on both sides of the periodic interface. To circumvent this problem a source term is added to the momentum
equation that compensates for the pressure losses and corrects the velocity
accordingly. Acharya [1] describes this procedure. The pressure is divided
into a periodic part and a correction term:
p(x, y, z) = x + pp (x, y, z)

(4.1)

Here pp (x, y, z) is the periodic part of the pressure and is the channel
pressure drop per unit length, ie. a pressure gradient. Acharya calculated
the value of using the current and the desired ow rate and applied overrelaxation factors to increase convergence for stationary simulations. The
method used in this work is slightly dierent and common practice in uid
solvers.
First, the following term
u

gp
(4.2)
|
u|
is added to the right hand side of the momentum equation, where u
is the
desired bulk ow velocity and gp is the articial pressure gradient in ow
19

direction. The latter can usually be set to zero for the rst iteration.
After solving the momentum and pressure equations, the cell volume
weighted average of the velocity eld in direction of u
is calculated:
|
u| = (u

)cva
|
u|

(4.3)

Next, the dierence of the current ow eld and the target bulk velocity
is calculated:
|
u| |
u|
(4.4)
gp+ =
1
( Ap )cva
where Ap is the diagonal coecient of the momentum matrix and the subscript cva again denotes the cell volume weighted average.
The velocity eld is then corrected using the pressure gradient corrector
gp+
u
1 +
u=u+
gp
(4.5)
|
u| Ap
and the pressure gradient used in the momentum source term is also updated.
(4.6)
gp = gp + gp+
These corrections are performed after every substep and therefore do
not inuence the temporal accuracy of the solver.
4.2.4. Energy Equation and Source Term
In order to be able to simulate ows including heat transfer, an additional
energy equation was added to the solver. This is not an energy equation as
it appears in the fully compressible Navier-Stokes equations but a simple
scalar transport equation for the temperature. The temperature is transported as a passive scalar, meaning that it has no inuence on the ow
eld. The procedure followed here is again based on the work by Acharya
[1]. The equation is given below:
T
(ui T )

t
xi
xi
20

T
ef f
xi


=0

(4.7)

Here T is the temperature, u is again the velocity and ef f is the

eective thermal diusivity. This diusivity is given by:
ef f =

+
P r P rt

(4.8)

where is the kinematic laminar viscosity and t is its turbulent counterpart. P r and P rt are the respective laminar and turbulent Prandtl numbers
which have to be given as simulation parameters.
When simulating periodic ows with heat transfer, a source term (or
sink term depending on the conditions) has to be added to equation 4.7
above. This source term can be derived from an energy balance (again see
[1]) and has the following form:
u

|
u|

(4.9)

q
ref cp,ref Href |
u|

(4.10)

u
where
=

Here q is the total heat ux in or out of the domain, ref , cp,ref and
Href are reference quantities for the density, the specic heat capacity and
the length respectively. These quantities need to be dened when running
a simulation. The reference length Href can be calculated as the domain
volume divided by the heated surface.

21

5. Validation

5.1

Overview

The solver along with the previously described turbulence models were
validated by performing a series of dierent test cases. These cases were
selected in such a manner that dierent physical phenomena can be investigated. They are also cases where stationary RANS models usually give
unsatisfactory results due to the inadequacy of the turbulence models.
The rst case shown is a widely-used validation case for turbulence
models and wall functions. It involves the simulation of fully wall-bounded,
periodic ow between two at parallel plates at a given Reynolds number.
Velocity proles, uctuations and stresses can be investigated and compared to widely available DNS data. Next, the external ow around a
square cylinder is investigated. Here the ow is fully detached and statistical values such as the Strouhal number can be compared to measurements.
A mixture of attached and separated ow including heat transfer is shown
in a ribbed channel. Nusselt numbers and reattachment lenghts are the
primary focus in these simulations. The last case shown also involves energy transport and investigates the eciency of lm-cooling in inclined
cooling holes at a low blowing rate. All the details and results on theses
cases are shown in the following sections.

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_5, Springer Fachmedien Wiesbaden 2016

5.2

Periodic Channel Flow

5.2.1. Description
One of the oldest and simplest cases for numerical investigations is the ow
between two parallel plates. It has been extensively researched and has
the advantage that, due to its simple geometry, fully resolved DNS using
spectral codes can be performed at reasonably high Reynolds numbers.
One of the main contributors to the numerical data of such simulations is
the work by Moser et al. [25]. The data from this publication is available
to the public under [26] and provides the basis for all comparisons made in
this work.
Figure 5.2.1 below shows a sketch of the geometry.

Figure 5.2.1: Channel Geometry Overview

Flows between parallel plates are usually characterised by a Reynolds

number dened as:
hu
.
(5.1)
Re =

This Reynolds number is based on the friction velocity u which, for

incompressible ows, is given by w , where w is the wall shear stress.

The length scale h is taken as the half-height of the channel and is the
laminar kinematic viscosity. The specic case investigated in this work was
performed at a friction Reynolds number of 395.
24

The original authors performed the simulation in a domain width edge

lengths of 2h 2h h. The size of this domain was chosen in such
a manner that the largest occuring turbulent structures can be captured.
The mesh resolution in the corresponding directions were 256 193 192,
resulting in a mesh size of approximately 9.5 million cells. The important
criteria for such cases are the dimensionless distances in wall normal and
wall parallel-directions (x is assumed to be the ow direction):
xu
(5.2)

yu
(5.3)
y+ =

zu
(5.4)
z+ =

The -values are the mesh sizes in the corresponding directions. The
values for these dimensionless distances used in the original paper are given
here for comparison with the values used in the present work:
x+ =

x+
10.0

y+
0.0295 - 4.81

z+
6.5

Table 5.2.1: Dimensionless Mesh Sizes of DNS Simulation

The values for x+ and z + were taken from the original paper, whereas
the values for y + were extracted from the numerical data. The lower value
of y + corresponds to the rst cell near the wall and the higher value represents the mesh size at the centerline of the channel.
5.2.2. Setup
The simulation domain size was reduced to 3.5 2 1.3 h as suggested by
other authors (eg. [30]). This reduction directly leads to decreased mesh
size while not inuencing the results too drastically.
As mentioned above, the case was run at a friction Reynolds number of
395, which leads to a target value for u of 0.0079. A bulk velocity of 0.138
[m/s] was enforced using the source term described in section 4.2.3. The
value for the bulk velocity was obtained by integrating the velocity prole
of the DNS data.
25

The following table 5.2.2 shows the boundary conditions for the case.
The wall boundary condition implies a zero xed velocity value and zero
gradient value for all scalars except for the subgrid viscosity, which was
also set to zero. The boundary condition for this turbulent viscosity is not
critically important when the boundary layer is resolved, as the values at
the wall are nearly zero when using the zero gradient condition.
Boundary
inlet & outlet
front & back
top & bottom

Type
translational periodic
translational periodic
no slip wall boundary

Table 5.2.2: Boundary Conditions for Periodic Channel Flow

5.2.3. Mesh
Two meshes were constructed for this case to observe the inuence of the
mesh resolution: a coarse mesh consisting of 51000 cells and a ne mesh
with 414000 cells. The dimensionless mesh sizes for both meshes (based on
the u -value reported above) were as follows:
Mesh
Coarse
Fine

x+
46.22
23.03

y+
1.47 - 29.35
0.73 - 14.58

z+
20.54
10.27

Table 5.2.3: Dimensionless Mesh Sizes of LES Simulation

26

Figure 5.2.2: Channel Mesh Fine

Figure 5.2.3: Channel Mesh Coarse

27

5.2.4. Results
The mean velocity and stress proles were evaluated for all cases. Mean
values of the velocity and the velocity uctuations in all spatial directions
were recorded during the simulation and averaged onto a single line. The
the mean wall shear stress was recorded as well and averaged over all wall
surfaces to provide a single value for the friction velocity u .
The simulations were started from an articial turbulence eld in order to accelerate the transition to properly turbulent ow. Details of the
initialisation procedure can be found in [8] If constant values are used for
the initialisation, transition to turbulence is only triggered by numerical
eects such as round-o errors and will take a very long time to occur. The
simulation was allowed about 25 ow-through times to initialise and was
then averaged over another 240. The time used here for the initialisation
was comparatively high. In [8] it was shown that a lower initialisation time
can be adequate as well.
All velocity values and their uctuations, as well as the wall distance,
were normalised using the value of u in the following manner:
y+ =
u+ =
u+ =
v + =
w+ =
+ =

y
u

u

u

v
u

w

u

u
 v

=
2
u

(5.5)
(5.6)
(5.7)
(5.8)
(5.9)
(5.10)

Here y is the wall normal distance from wall to cell center, is the
molecular kinematic viscosity, u
is the mean value of the x-component of
the velocity, all primed values are averaged uctuations and w is the wall
shear stress.
28

Coarse Mesh
First the results obtained using the coarse mesh are presented. Figures 5.2.4
to 5.2.7 show the mean dimensionless velocity proles. In gures 5.2.8 to
5.2.11 the uctuations and in 5.2.12 to 5.2.15 the shear stress prole can
be seen.
The diagrams are ordered in such a fashion that the dierent models are
put into meaningful groups. The values of the predicted friction velocity are
provided in table 5.2.4. The results show that the dynamic models perform
signicantly better than any of the static models, even the wall-adapting
WALE and sigma models. The static Smagorinsky model using the van
Driest damping is not able to capture any of the meaningful quantities,
especially in the case of the mean uctuations. The best model for this
grid is no model at all. The reason behind this is given in section 5.6.
Case
Smagorinsky & VanDriest
Dynamic Smagorinsky
Dynamic Lagrangian
Dynamic 1-Equation Model
WALE
sigma
No Model
DNS

Friction Velocity u [m/s]

0.00774
0.00783
0.00776
0.00773
0.00741
0.00739
0.00781
0.00785

Table 5.2.4: Friction Velocity, Coarse Mesh

29

20

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

20

15
u+ [-]

u+ [-]

15

10

100

101
y + [-]

0
101

102

Figure 5.2.4: Mean Velocity Proles,

Smagorinsky Models

20

DNS Data
WALE
sigma

20

u+ [-]

u+ [-]

101
y + [-]

102

DNS Data
No Model

15

10

10

100

101
y + [-]

102

Figure 5.2.6: Mean Velocity Proles,

Static Models

30

100

Figure 5.2.5: Mean Velocity Proles,

n-Equation Models

15

0
101

10

0
101

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

0
101

100

101
y + [-]

102

Figure 5.2.7: Mean Velocity Proles,

No Model

3.5

3.5

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

2.5
u+ , v + , w+ [-]

u+ , v + , w+ [-]

2.5
2
1.5

2
1.5

0.5

0.5

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

20

40

60

80

100

20

40

Figure 5.2.8: Mean Velocity Fluctuations, Smagorinsky Models

3.5

100

DNS Data
No Model

3
2.5

2.5
u+ , v + , w+ [-]

u+ , v + , w+ [-]

80

Figure 5.2.9: Mean Velocity Fluctuations, n-Equation Models

DNS Data
WALE
sigma

3.5

2
1.5

2
1.5

0.5

0.5

60
y + [-]

y + [-]

20

40

60

80

100

y + [-]

Figure 5.2.10: Mean Velocity Fluctuations, Static Models

20

40

60

80

100

y + [-]

Figure 5.2.11: Mean Velocity Fluctuations, No Model

31

0.2

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

0.2

+ [-]

0.4

+ [-]

0.4

0.6

0.6

0.8

0.8

50

100

150

200
y + [-]

250

300

350

0.2

400

Figure 5.2.12: Mean Shear Stress,

Smagorinsky Models

50

100

150

200
y + [-]

250

300

350

400

Figure 5.2.13: Mean Shear Stress, nEquation Models

DNS Data
WALE
Sigma

DNS Data
No Model

0.2

+ [-]

0.4

+ [-]

0.4

0.6

0.6

0.8

0.8

50

100

Figure 5.2.14:
Static Models

32

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

150

200
y + [-]

250

300

350

400

Mean Shear Stress,

50

100

150

200
y + [-]

250

300

350

400

Figure 5.2.15: Mean Shear Stress, No

Model

Fine Mesh
The same evaluations have been performed on the ne mesh and the diagrams are grouped in the same way as before. All the results are much closer
to the reference data provided by the DNS solution. The Smagorinsky
model using the vanDriest damping function still shows some deciencies
which are due to the fact that the vanDriest damping is an articial modier to the subgrid viscosity and not physically meaningful. The velocity
proles of the dynamic Smagorinsky and the dynamic n-Equation models
(gures 5.2.16 and 5.2.17) all agree very well with the data whilst the static
models seem to slightly overestimate the subgrid viscosity, leading to higher
velocities in the logarithmic region of the prole. The no-model simulation
again shows surprisingly good results while very slightly underestimating
the prole. This indicates simply that for a DNS simulation the grid is still
too coarse.
The velocity uctuations and shear stresses are better represented here
compared with the results from the coarse grid. The agreement is very
good close to the wall but quickly deteriorates further towards the channel
centerline. This eect is observed in most papers doing such comparisons
(eg. [8] [30]).
The values of the friction velocity on the ne mesh are shown below
(table 5.2.5).
Case
Smagorinsky & VanDriest
Dynamic Smagorinsky
Dynamic Lagrangian
Dynamic 1-Equation Model
WALE
sigma
No Model
DNS

Friction Velocity u [m/s]

0.00787
0.00793
0.00791
0.00791
0.00779
0.00776
0.00797
0.00785

Table 5.2.5: Friction Velocity, Fine Mesh

33

20

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

20

15
u+ [-]

u+ [-]

15

10

0
101

100

101
y + [-]

0
101

102

DNS Data
WALE
sigma

20

u+ [-]

u+ [-]

101
y + [-]

102

DNS Data
No Model

15

10

10

100

101
y + [-]

102

Figure 5.2.18: Mean Velocity Proles,

Static Models

34

100

Figure 5.2.17: Mean Velocity Proles,

n-Equation Models

15

0
101

10

Figure 5.2.16: Mean Velocity Proles,

Smagorinsky Models

20

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

0
101

100

101
y + [-]

102

Figure 5.2.19: Mean Velocity Proles,

No Model

2.5

1.5

0.5

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

2.5

u+ , v + , w+ [-]

u+ , v + , w+ [-]

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

1.5

0.5

20

40

60

80

100

20

40

y + [-]

Figure 5.2.20: Mean Velocity Fluctuations, Smagorinsky Models

100

2
1.5
1
0.5

DNS Data
No Model

2.5

u+ , v + , w+ [-]

u+ , v + , w+ [-]

80

Figure 5.2.21: Mean Velocity Fluctuations, n-Equation Models

DNS Data
WALE
sigma

2.5

60
y + [-]

1.5

0.5

20

40

60

80

100

y + [-]

Figure 5.2.22: Mean Velocity Fluctuations, Static Models

20

40

60

80

100

y + [-]

Figure 5.2.23: Mean Velocity Fluctuations, No Model

35

0.2

DNS Data
Smagorinsky + VanDriest
dyn. Smagorinsky

0.2

+ [-]

0.4

+ [-]

0.4

0.6

0.6

0.8

0.8

50

100

150

200
y + [-]

250

300

350

0.2

400

Figure 5.2.24: Mean Shear Stress,

Smagorinsky Models

50

100

150

200
y + [-]

250

300

350

400

Figure 5.2.25: Mean Shear Stress, nEquation Models

DNS Data
WALE
Sigma

DNS Data
No Model

0.2

+ [-]

0.4

+ [-]

0.4

0.6

0.6

0.8

0.8

50

100

Figure 5.2.26:
Static Models

36

DNS Data
dyn. Lagrangian
dyn. One Eq. Eddy

150

200
y + [-]

250

300

350

400

Mean Shear Stress,

50

100

150

200
y + [-]

250

300

350

400

Figure 5.2.27: Mean Shear Stress, No

Model

5.3

Square Cylinder

5.3.1. Description
The external ow around blu bodies is the focus of many research investigations. The case shown here deals with a square cylinder of dimensions
e e 4e (where e = 0.04 [m]) in free ow that was experimentally investigated by Lyn et al. [21],[7]. The cylinder is placed in a three-dimensional
channel of dimensions 0.82 0.56 0.16 [m]. Figure 5.3.1 shows a sketch of
the simulation domain as well as the location of the origin of the coordinate
system used for all the evaluations below.

Figure 5.3.1: Channel Geometry Overview

The main interest here is to be able to accurately predict aerodynamic

properties of the body such as Strouhal number and drag coecient. The
geometry is simple and oers well-dened separation points for the ow at
the sharp edges of the cylinder.
5.3.2. Setup
The following table 5.3.1 provides an overview of the boundary conditions
used. The inlet velocity was given by the original paper [21] and had a
value of 0.535 [m/s]. The kinematic viscosity was set to 1e-5 [m2 /s]
and the timestep used was 0.001 [s]. The simulation was run over a total
duration 50 [s], ie. 50000 timesteps.
37

Boundary
inlet & outlet
front & back
cylinder
top & bottom

Type
xed value/zero gradient
translational periodic
no slip wall boundary
symmetry plane

Table 5.3.1: Square Cylinder Boundary Conditions

Note that the simulation is fully three-dimensional, 2D-simulations were

tested but did not give any meaningful results. This is obvious, as in large
eddy simulations, a large part of the turbulent spectrum is resolved, and
turbulent structures are always three-dimensional. In fact, in the sigma
model it is even the case that in locally two-dimensional or two-component
ows, the contribution from the turbulent eddy viscosity is zero.
Constant initial values were used to start the case as, due to the enforced
separation, the turbulent ow eld develops rather quickly. A slightly distorted initial eld for the velocity is benecial as the separation can remain
symmetrical around the x-axis for several thousand steps and only then
develop into a proper vortex street.

38

5.3.3. Mesh
Due to the simplicity of the geometry, a fully orthogonal, hex-cell mesh
was created. The resolutions in the three spatial coordinates were 160
198 32, resulting in an overall mesh size of 962560 cells. A frontal
view of the mesh can be seen in gure 5.3.2. The near wall resolution of
the mesh was rather coarse, resulting in average values of approximately
x+ = y + = 5 and z + = 25. As the ow is mostly detached and the
separation points are given by the geometry, increasing the resolution in
wall normal direction will not signicantly improve the results but will
cause issues for the allowable timestep limit.

Figure 5.3.2: Square Cylinder Mesh

39

5.3.4. Results
In this case the Strouhal number St, the drag and lift coecients Cd and
Cl as well as the averaged velocity were evaluated. These quantities are
dened as follows:
f lref
u
Fx
Cd = 1 2
u
2 Aref
Fy
Cl = 1 2
2 u Aref
St =

(5.11)
(5.12)
(5.13)

Here f is the shedding frequency of the ow, lref is a reference length

(in our case the edge length of the cylinder, ie. 0.04 [m]) and u is the
freestream velocity, set to 0.535 [m/s]. In the drag coecients F is the force
in the respective coordinate direction and Aref is a reference area which
here is equal to the cross-sectional area of 0.0064 [m2 ]. The total forces F
are evaluated by summing up the pressure and friction forces. These forces
were used not only in the calculation of the drag coecients but also in the
determination of the shedding frequency f . A frequency analysis via FFT
(Fast Fourier Transform) was performed on the signal of the forces in the
y-direction. This is also the reason why the simulation was run for a rather
long time as the resolution of the frequency spectrum depends on the total
number of available samples.
The rst 5000 timesteps were not used in the evaluation in order to
allow the ow to fully develop rst.
In gures 5.3.3 to 5.3.6 below the results of the averaged centerline
velocity in x-direction are shown and compared to the measurements by
[7]. Compared to the original paper, the origin of the x-axis is shifted by
0.5h, ie. the origin is not in the center of the cylinder but just at the end
where the wake begins.
As the original paper does not provide any information on the uctuation of the drag coecients and only provides mean values, simulation
results by other authors are listed here as well. Ochoa [32] provides a
helpful overview of many such comparisons.
This overview is listed in table 5.3.2 below.
40

0.8

0.8

0.6

0.6

0.4

0.4

u/u [-]

u/u [-]

0.2
0

0
Measurement
Smagorinsky + VanDriest
dyn. Smagorinsky

0.2
0.4

0.2

Measurement
dyn. Lagrangian
dyn. One Eq. Eddy

0.2
0.4

10

x/e [-]

10

Figure 5.3.4: Averaged Velocity Prole, n-Equation Models

0.8

0.8

0.6

0.6

0.4

0.4

u/u [-]

u/u [-]

Figure 5.3.3: Averaged Velocity Prole, Smagorinsky Models

0.2
0

0.2
0

Measurement
WALE
sigma

0.2
0.4

6
x/e [-]

0.2

10

x/e [-]

Figure 5.3.5: Averaged Velocity Prole, Static Models

0.4

Measurement
No Model
0

10

x/e [-]

Figure 5.3.6: Averaged Velocity Prole, No Model

It can clearly be seen in gures 5.3.3 to 5.3.6 that none of the models
are able to predict the nal recovery value of the mean velocity. This is
mostly due to the mesh resolution. The fact that constant boundary values
were used, and therefore no inlet turbulence level was present, could also
have inuenced this result. Since the no-model simulation shows the same
behaviour, the issue cannot be attributed to the models themselves. The
dynamic lagrangian and the sigma model are the only ones that are able
to capture the location and magnitude of the steep gradient accurately
and agree very well with the measured values. As this region in the ow
41

Case
Smagorinsky & vanDriest
Dynamic Smagorinsky
Dynamic Lagrangian
Dynamic One-Eq. Eddy
WALE
sigma
No Model
Verstappen and Veldman [43]
Porquie et al. [35]
Murakami et al. [27]
Wang and Vanka [45]
Nozawa and Tamura [31]
Ochoa and Fueyo [32]
Exp.: Lyn et al. [7] [21]

St
0.128
0.138
0.139
0.138
0.135
0.140
0.139
0.133
0.13
0.131
0.13
0.131
0.139
0.132

Cl
-0.017
-0.010
-0.001
0.01
0.008
0.001
-0.008
0.005
-0.02
-0.05
0.04
0.009
0.03
-

Cl
1.442
0.521
1.011
0.646
1.429
1.033
0.615
1.45
1.01
1.39
1.29
1.39
1.4
-

Cd
2.344
1.769
1.993
1.826
2.239
2.004
1.794
2.09
2.2
2.05
2.03
2.62
2.01
2.1

Cd
0.220
0.062
0.081
0.058
0.217
0.080
0.059
0.178
0.14
0.12
0.18
0.23
0.22
-

Table 5.3.2: Comparison of Aerodynamic Data

is the most important for the aerodynamical drag around blu bodies it
comes as no surprise that these two models are also able to predict the drag
coecient well.
The Strouhal number was comparably close to the measured value in
all of the cases and in the same range as obtained by Ochoa and Fueyo [32].
However, other authors have been able to reach values that are closer to
the measurements.
In order to improve the simulation results, the mesh would have to be
rened, especially in the wake region. It would also improve the prediction
of the Strouhal number. This point is explained further in chapter 5.6
below.

42

5.4

Ribbed Channel

5.4.1. Description
In ows where heat transfer is important, eg. for cooling certain machine
parts, high heat transfer coecients are always desirable. One way to
achieve such increased heat transfer coecients is by adding turbulators.
Turbulator is a generic term for any geometric modication of a surface that
increases turbulence levels, thereby improving mixing and energy transfer
into the core ow. The case investigated here involves such a geometry of
a rectangular turbulator in a periodic channel. The overall domain size is
0.127 0.061 0.060325 [m] and height and width of the turbulator rib,
which lies in the center of the bottom wall, is e = 0.00635 [m]. The case
was originally investigated by [1] whose measurements served as the basis
for comparisons in this work. A sketch of the geometry and the location of
the coordinate system origin can be seen in gure 5.4.1.

Figure 5.4.1: Ribbed Channel Geometry Overview

5.4.2. Setup
In this case, the temperature transport equation described in section 4.2.4
was used to simulate heat transfer. Both the momentum and energy source
terms had to be activated in order to allow for periodic boundary conditions. Table 5.4.1 below shows the boundary conditions used in this case.
43

Boundary
inlet & outlet
front & back
bottom wall & rib
top wall

Type
translational periodic
translational periodic
no slip wall boundary
no slip wall boundary

Table 5.4.1: Ribbed Channel Boundary Conditions

Additionally, the bottom wall, excluding the rib, was a heated wall
and required special boundary conditions for the temperature equation.
One way to achieve a heating wall would be to simply set a xed wall
temperature value. The drawback of this method is that the total heat ux
into the domain would depend on the ow eld, ie. the term in equation
4.9 would have to be recalculated for every timestep. For this reason the
total heat ux was xed to a value of 280 [W/m2 ], therefore keeping a
constant value. In order to achieve this, a xed gradient boundary condition
was imposed on the bottom wall, using a gradient value of 10415.7135. This
value was calculated through the following relation (see [2]):


T
q = kf
(5.14)
n wall
Here q is the wall heat ux, kf is the thermal conductivity of the uid
and the last term is the wall normal gradient of the temperature we are
looking for. The conductivity can be calculated as
kf = cp

(5.15)

where is the thermal diusivity, is the density and cp is the specic

heat capacity at constant pressure.
In our solver, is calculated as
=

+
P r P rt

(5.16)

where and t are the laminar and turbulent viscosities and P r and P rt
the respective Prandtl numbers. In our case, t , the subgrid viscosity, is

44

set to zero at the wall and no wall functions are used. For this reason the
turbulent contribution to the thermal diusivity at the wall is zero and
the gradient can be imposed as a constant. If this were not the case, the
wall heat ux would have to be constantly updated to maintain the total
required heat ux.
Care has to be taken that the wall gradient is calculated using the same
quantities as used in the temperature equation source term. The bulk
ow velocity was set to 3.6 [m/s], leading to a Reynolds number of 28341
(based on Dh ). The rest of the quantities used are summarised in table
5.4.2 below.
Quantity

cp
H
Pr
P rt
q
Dh

Value
1.55e-5 [m2 /s]
1.208 [kg/m3 ]
1005.2 [J/(kgK)]
0.06388 [m]
0.7 []
0.5 []
280 [W/m2 ]
0.122 [m]

Table 5.4.2: Physical and Geometrical Data used in Ribbed Channel Case

The reference length H was calculated as the total domain volume divided by the heated surface area. The turbulent Prandtl number was set
following a suggestion by Moin et al. [24] and twice the channel height was
used for the hydraulic diameter, as is the case for parallel plates.
A timestep of 1e-5 [s] was used and the simulation was run over the
duration of 1 [s] with the averaging procedure starting after 0.1 [s]. All
quantities were initialised as constant values. Due to the presence of the
rib and the resulting ow separation fully turbulent state is quickly reached
and no addition of articially turbulent elds was necessary.
5.4.3. Mesh
The mesh resolutions in the three coordinate directions were 136 104
33, amounting to a total cell number count of 493152. The y + value was
45

kept below 1 on the heated walls and varies over the rib. The values for x+
and z + vary strongly due to the dierent mesh densities but never surpass
a value of approximately 27 y + . Figure 5.4.2 gives an overview of the
mesh.

Figure 5.4.2: Ribbed Channel Mesh

5.4.4. Results
Nusselt Number Distribution
In order to assess the performance of the turbulator geometry, the Nusselt
number N u is evaluated. It is dened as the ratio of the convective and
the conductive heat transfer, therefore providing insight on the eectivity
of turbulator. It is dened as follows:
Nu =

qDh
k(Tw Tb )

(5.17)

Here q is the wall heat ux, Dh the hydraulic diameter, k the thermal
conductivity, Tw the temperature at the wall and Tb is the bulk temperature. It is evident that when the properties and the heat ux are constant, a
higher Nusselt number indicates a smaller temperature dierence between
the wall and the bulk ow, meaning better heat transport. The bulk temperature Tb is not a constant but depends on the simulation properties
and the turbulence model used. It was evaluated by taking the mass ow
average at the inlet periodic surface based on the time-averaged quantities.
The following gures 5.4.3 to 5.4.6 show the Nusselt number distribution over the position in ow direction. The position in x-direction was
46

normalised using the height of the rib and the rib is located at a value
between 9.5 and 10.5. The measurement values used for the comparison
were taken from [1]. The wall temperatures evaluated were averaged over
time and then averaged over the lateral direction to provide a single line
data source.
200

160

160

140

140

120

100

80

80

60

60
0

10 12
x/e [-]

14

16

18

200

40

20

Figure 5.4.3: Nusselt Number Distribution, Smagorinsky Models

200

140
N u [-]

160

140
120

100

80

80

60

60
2

10 12
x/e [-]

14

16

18

20

Figure 5.4.5: Nusselt Number Distribution, Static Models

10 12
x/e [-]

14

16

18

20

120

100

Measurement
No model

180

160

40

Figure 5.4.4: Nusselt Number Distribution, n-Equation Models

Measurement
WALE
sigma

180

N u [-]

120

100

40

Measurement
dyn. Lagrangian
dyn. One Eq. Eddy

180

N u [-]

N u [-]

200

Measurement
Smagorinsky + VanDriest
dyn. Smagorinsky

180

40

10 12
x/e [-]

14

16

18

20

Figure 5.4.6: Nusselt Number Distribution, No Model

All of the models tested are able to capture the overall shape of the
Nusselt number distribution. The dynamic Smagorinsky model agrees very
well with the measurements whilst its overall prole is slightly compressed,
47

underestimating the gradient before the rib. The static WALE and sigma
models predict the prole very well but the results are oset compared to
the measurements. This indicates a general over-estimation of the turbulent
subgrid viscosity and therefore the eective thermal diusivity. The same
can be said for the dynamic one-equation model. According to [1] the
maximum location of the Nusselt number after the rib lies at a position
of 15.6 (with an measuring uncertainty of 0.4). This is predicted well
by all models, however the exact location is dicult to determine in some
cases (eg. Smagorinsky + van Driest) as the solution contains wiggles in
this region. These could be evened out by averaging over a longer period
of time.
Again the no model simulation shows very good results while slightly
underestimating absolute values.
Just after the rib, in between locations 10.5 and approximately 11, a
strong increase in the Nusselt number can be observed that is not present in
the measured results. This is due to a local vortex in the vicinity of the rib
which increases the energy transport. Figure 5.4.7 shows the streamlines of
the averaged velocity in the wake region of the rib. It is possible that due
to the way the measurements were taken ([1]) this local peak was smoothed
out.

Figure 5.4.7: Averaged Streamlines

Reattachment Length
Also investigated in this case was the reattachment length behind the rib.
For this reason, the wall shear stress was calculated at every time step and
averaged along with the other quantities. The reattachment point was then
evaluated by looking at the change of sign of the wall shear stress in ow
direction. The reattachment location of separated ows are notoriously
48

dicult to accurately represent when using Reynolds averaged (RANS)

turbulence models.
Table 5.4.3 below shows the predicted reattachment lengths of all cases.
It is evident that all models are well within the accuracy of the measurements and signicantly better than RANS predictions conducted in an
ongoing study at the University of Applied Sciences & Arts Lucerne. Due
to the wide error margin of the measurements it is dicult to assess which
of the models actually performed best.
Case
Smagorinsky & VanDriest
Dynamic Smagorinsky
Dynamic Lagrangian
Dynamic 1-Equation Model
WALE
sigma
No Model
Measurement

Reattachment length x/e []

6.015
6.045
5.935
5.8
5.775
5.945
5.985
6 0.7

Table 5.4.3: Reattachment Lengths

49

5.5

Film Cooling

5.5.1. Description
The nal case treated in this work deals with lm cooling at low blowing
rates. Film cooling is a practice often used in turbomachines, specically
gas turbines. The rst stages of such turbines are exposed to very high
temperature gases, often beyond the allowable limit of the turbine blade
materials. For this reason, complex cooling geometries are built into the
blade casings. One of the methods used in such situations is lm cooling.
Here gas is extracted from the machine at lower temperatures and blown
into the critical parts of the machine. This cool gas then forms a protective
layer around the geometry, thus avoiding the destruction of the materials.
A quantity often used for classifying such cooling systems is the blowing
rate, dened as:
c uc
M=
(5.18)
u
where c and are the densities and uc and u are the velocities of the
cooling ow and the main ow respectively. The values for the cooling ow
part are evaluated inside the hole.
The prediction of such cooling layers is a dicult task, especially at low
blowing rates where the main ow strongly interacts with the jet. Turbulence in such regions is highly anisotropic making it virtually impossible for
isotropic RANS models to accurately predict the covered surface. RANS
turbulence models are usually modied to account for the anisotropy, such
as in [3]. An attempt was made in this work to employ a LES procedure to
such a test case. Figure 5.5.1 below shows the geometry and the position
of the coordinate systems origin.
5.5.2. Setup
The case investigated here follows the measurements conducted by Sinha
et al. [40]. The case and geometry details can be found in the following
table 5.5.1.
The rst thing to note is that in this work an incompressible solver
was used, therefore no density is available. The Mach numbers are low
enough for this to be valid but it poses the question on how to replicate the
ow conditions accurately. Johnson et al. [16] discuss the importance of
50

Figure 5.5.1: Film Cooling Geometry Overview

Quantity
Hole Diameter
Hole inclination
Diameter to Length Ratio
Channel Height
Bulk Flow Velocity
Cooling Flow Velocity
Blowing Rate
Bulk Flow Temperature
Cooling Flow Temperature
Bulk Flow Density
Cooling Flow Density

Value
0.0127 [m]
35 [ ]
1.75 []
0.127 [m]
20 [m/s]
5 [m/s]
0.5 [-]
302 [K]
153 [K]
1 [kg/m3 ]
2 [kg/m3 ]

Table 5.5.1: Film Cooling Setup Parameters

dierent similarity criteria. This leads to the following possible quantities

to keep constant:
1. The velocity ratio Ru =
2. The blowing rate M =

uc
u

c uc
u

3. The ratio of momentum uxes I =

M2
R ,

where R is the density ratio

In the rst case, the velocities remain the same therefore no adaptation needs to be performed. In the second case, keeping the blowing rate
constant, the velocity in the cooling hole needs to be doubled and when
51

keeping the momentum

uxes constant, the cooling hole velocity has to be
increased by a factor of 2. It is a priori not clear which of these methods
best represents the compressible case. Here the rst and third method were
chosen as they represent the minimum and the maximum.
Only one subgrid model, the sigma model (see section 3.6), was used
here as the computational eort for such a simulation is considerable.
The boundary conditions used are given in the following table 5.5.2:
Boundary
inlet & outlet
front & back
bottom wall & cooling hole wall
top wall

Type
xed value / zero gradient
translational periodic
no slip wall boundary
symmetry plane

Table 5.5.2: Film Cooling Boundary Conditions

The inlet boundary condition was chosen to be a constant xed value

for the velocity, which is also used in RANS simulations of the present
case. Using a periodic mapped boundary condition, simulating an innite
channel at the inlet, would have been the preferred method. Numerical
instabilities prevented this, however. More on this topic can be found in
chapter 9 on page 71.
The timestep used in this simulation was adapted automatically based
on a xed maximum CFL number of 1, which lead to time step sizes of
around 2.3e-6. The simulation was started from constant elds and was
run for 0.1 seconds, corresponding to about 40000 timesteps, before the
averaging procedure was started. The averaging was then performed over
two ow-through times of the bulk ow.
5.5.3. Mesh
Three meshes were generated for this case but only one was used in the
end. The coarsest mesh, consisting of only 266728 cells, was in fact too
coarse and did not deliver any meaningful results. The large variation
in cell densities occasionally even lead to numerical instability. A ner
and more even mesh was created using the same topology, resulting in a
mesh size of just under 1.4 million cells. The nest mesh generated used a
52

dierent topology and was smoothed to improve the overall orthogonality

of the cells. This mesh consisted of about 2.5 million cells. Due to the very
low timesteps required and the overall mesh size, the nest grid was not
used and the results shown below were obtained using the medium mesh.
Figures 5.5.2 and 5.5.3 show the coarse and the medium mesh respectively
and on gure 5.5.4 a close-up of the ne mesh region around the cooling
hole exit can be seen.

Figure 5.5.2: Film Cooling Mesh, Coarse, Regular

Figure 5.5.3: Film Cooling Mesh, Medium, Regular

53

Figure 5.5.4: Cooling Hole Mesh, Fine, Smoothed

5.5.4. Results
The main interest when doing lm cooling simulations is the surface coverage and the lateral spreading of the cool gas. In order to quantify this,
the adiabatic eectiveness is introduced, which is dened as follows:
=

T T
Tc T

(5.19)

Here T is the temporally averaged temperature of the simulation and

T and Tc are the bulk and cooling hole ow temperature respectively.
The value of ranges between 0 and 1, where 0 means no coverage and
1 indicates full coverage of the surface by the cool gas. For all diagrams
below, the coordinate directions have been normalised by the hole diameter.
Figures 5.5.5 and 5.5.6 below show the centerline eectiveness as well as the
laterally averaged centerline eectiveness over the normalised streamwise
direction.
The following two gures 5.5.7 and 5.5.8 show the lateral distribution
of at a normalised streamwise position of 6 and 15, respectively.
The results show that neither of the setups used accurately describes
the centerline and laterally averaged eectiveness. Keeping Ru constant
has lead to an overprediction of , while keeping M constant leads to ow
separation just after the cooling hole exit and therefore a locally reduced
eectiveness. This leads to the conclusion that keeping the momentum ux
ratio I constant would have been a better choice.
54

Measurements
Ru constant
M constant

0.8

Measurements
Ru constant
M constant

0.8

[-]

0.6

[-]

0.6

0.4

0.4

0.2

0.2

10

15
x/D [-]

20

25

Figure 5.5.5: Centerline Adiabatic Effectiveness

1

30

8
10
x/D [-]

12

14

16

Figure 5.5.6: Laterally Averaged Adiabatic Eectiveness

1

Measurements
Ru constant
M constant

0.8

Measurements
Ru constant
M constant

0.8

[-]

0.6

[-]

0.6

0.4

0.4

0.2

0.2

0.2

0.4

0.6

0.8
z/D [-]

1.2

1.4

1.6

Figure 5.5.7: Adiabatic Eectiveness

at x/D = 6

0.2

0.4

0.6

0.8
z/D [-]

1.2

1.4

1.6

Figure 5.5.8: Adiabatic Eectiveness

at x/D = 15

In gure 5.5.7, where the lateral distribution of is shown, a dip in

the eectiveness can be seen between z/D = 0.6 and 1.2. This dip is due
to a horseshoe-vortex forming around the jet entering the main ow and
transporting hot gas into the region past the cooling hole (see gure 5.5.9).
The eect can be seen in both cases investigated here and is therefore
not due to the inlet velocity of the cooling gas. The reason for the strong
eect of this vortex is most likely the lack of turbulent mixing, allowing the
vortex to fully develop. This is a problem related to the boundary condi55

Figure 5.5.9: Horseshoe Vortex Around Cooling Hole Exit

tions used in the main ow, which were set to uniformly constant values.
Due to the relatively short length of the domain inlet, at-plate transition
does not occur (which the sigma turbulence model is able to reproduce)
and the ow arriving at the cooling hole is fully laminar. The reason why
constant values were used at the inlet is given in section 9. In a future
simulation, turbulent boundary conditions and a constant momentum ux
ratio I will be applied and the results are expected to improve further.

56

5.6

Comments on the Results

5.6.1. No-Model Results

A few general comments on the above results need to be made. First we
have the surprisingly good results of the no-model simulations. This is a
phenomenon that has been observed by other authors as well (see eg. [6])
and is, after all, not that surprising. In this work we considered subgrid
models that are all based on the eddy viscosity concept by Boussinesq (see
eg. [12]). This means that all those models increase the eective viscosity
of the uid considered and therefore increase diusion. An exception to
this would be the localised dynamic Smagorinsky model where the turbulent viscosity can become negative and is clipped at the negative laminar
viscosty (ie. eliminating the diusion term fully). But even that model has
an overall positive value of the subgrid viscosity and therefore increases
diusion. Numerical diusion has exactly the same eect. We are using a
diusion-free central dierencing scheme for the convection term, but since
the turbulent ow eld in large eddy simulations is hardly ever aligned with
the computational grid, diusion occurs nevertheless in nite volume methods. Looking at the results from our simulations we can therefore deduce
that the numerical diusion is more or less of the same order of magnitude
as the contribution from the subgrid models. This also explains why all
the dynamic models provide much better results than the static ones in the
channel simulations: The dynamic models adapt their coecients taking
into account the numerical diusion as well, since they perform an additional ltering of the resolved velocity eld.
5.6.2. Grid Resolution
The second point to make is the lack of high-order spatial schemes in OpenFOAM. It has shown in this work that using only second-order schemes for
large eddy simulations has several major drawbacks, all related through the
grid resolution.
The rst point to mention is the grid resolution required for accuracy:
When using low-order spatial schemes, the grid resolution has to be very
ne in order to adequately reduce the discretisation error. This directly
leads to smaller time steps required to stay in the order of a CFL number of
around one. This means that using low-order schemes increase calculation
57

times signicantly.
Secondly, the grid density required can reach a level where the inuence
of the subgrid models practically vanishes, therefore almost invalidating the
use of any subgrid models. Again the example of the channel ow shall be
mentioned here, especially in the case of the sigma model. The original
paper [30] is able to obtain much better results with the model on a grid
with a resolution that lies between the coarse and the ne version used
here. This is mainly due to the fact that they were able to use fourth and
fth order schemes in their spatial discretisation (note however that other
issues such as implementation and clipping will certainly also play a role
here).
The last point to mention is that if a coarser grid is used, the spatial
discretisation error might be at a level where the third order time integration scheme will not be able to fully deliver its potential. Figure 4.2.1 on
page 18 shows the error evolution with respect to the time step obtained
in the order verication study performed in [19]. The error value of the
plateau on the left hand side is given by the spatial discretisation error, ie.
the discretisation scheme and the grid resolution. If the grid is coarsened,
the level of this plateau rises and may make the use of a (slower) third order
time scheme pointless. It is possible that this is the reason why hardly any
of the simulations performed in this work were able to accurately capture
the Strouhal number of the square cylinder (reduction of the timestep did
not improve the results in that respect).

58

6. Parallel Performance

Large eddy simulations are computationally expensive. Apart from a ne

mesh resolution a large number of timesteps is required to get signicant
results. This is due to the fact that not only does the simulation have to
reach a statistically steady state (eg. fully developed ow), it then needs to
run long enough to provide meaningful average values. This is even more
the case when transient statistics such as shedding frequencies are needed
(eg. around a square cylinder, see section 5.3) as the resolution of the
frequency spectrum depends on the total number of samples available. For
these reasons parallel computing is a necessary tool in order to deal with
large eddy simulations.
One of the most important qualities of an LES solver is parallel eciency. The code developed in this work was tested using the same mesh
on various amounts of cpu cores to obtain a measure for the speed gained
when using more cores. The most common performance metric for this is
the speedup S, dened as:
T1
S=
(6.1)
Tp
Here T1 is the execution time required using a single core and correspondingly Tp is the time required using p cores.
Also used in gure 6.0.2 below is the parallel eciency Ep dened as:
Ep =

T1
pTp

(6.2)

The test run was done using the mesh and setup from the square cylinder
case (see section 5.3), performing 1000 time steps and recording the time
required. The mesh consists of 962560 cells and uses all the boundary
types from the original case, including period ones (cyclicAMI). The old

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_6, Springer Fachmedien Wiesbaden 2016

and the new pressure velocity scheme were investigated, as well as two types
of linear solvers for the pressure equation: a multigrid solver (GAMG) and
Krylov-based solver (PCG). Figure 6.0.1 shows the parallel speedup of all
the cases. Table 6.0.1 provides the numerical values that were used in the
gure.
120
100

Ideal
Old Algorithm (PCG)
New Algorithm (PCG)
Old Algorithm (GAMG)

S [-]

80
60
40
20

20

40

60
n [-]

80

100

120

Figure 6.0.1: Speedup

No. of Cores
1
4
8
16
32
64
128

Cells/Core
962560
240640
120320
60160
30080
15040
7520

ToP CG [s]
30326
8749
5506
4385
1892
835
458

ToGAM G [s]
34807
9768
5991
3784
2034
1472
2047

TnP CG [s]
36831
10680
6987
4774
2294
1034
550

Table 6.0.1: Execution Times of Parallel Runs

It is evident that the GAMG solver is not suitable for massively parallel
calculations. The speedup even starts to decrease below a certain number
60

0.8

Old Algorithm (PCG)

New Algorithm (PCG)
Old Algorithm (GAMG)

Ep [-]

0.6

0.4

0.2

20

40

60
n [-]

80

100

120

Figure 6.0.2: Parallel Eciency

of cells per core. It could be observed that the number of iterations needed
to solve the pressure equation increased with the amount of cores used,
therefore slowing down the whole process. On the other hand, the PCG
solver scales relatively well. Even super-linear scaling can be observed
between 16 and 64 cores, ie. using twice as many cores reduces the time
required by more than a factor of two (see gure 6.0.3 below). One thing to
note is that the linear scaling (denoted as "ideal" in gure 6.0.1 above) can
be attained by fully explicit algorithms where no equation systems need to
be solved, ie. no matrix multiplications need to be performed. In our case,
the pressure equation needs to be solved implicitly, therefore fully linear
scaling (compared to the single-cpu eort) is out of reach.
The eciency of the PCG solver decreases at the very last step, which
leads to the conclusion that the optimal cell number per core is approximately 10000 (when pure simulation speed is the goal). This value most
likely depends on the setup of the computing cluster, however.
When comparing the old and new pressure-velocity coupling algorithms
it can be seen that the new version is slightly slower, but scales just as well.
This is due to an increase in molecule size in the pressure matrix, therefore
61

Reduction Factor []

Old Algorithm (PCG)

New Algorithm (PCG)
Old Algorithm (GAMG)

0
4-8

8-16
16-32
32-64
Number of Cores

64-128

Figure 6.0.3: Time Reduction Factor When Doubling the Number of Cores

making the pressure equation more dicult to solve. The experiment using
the GAMG solver with the new method leads to the same behaviour as
above and is therefore not shown here.

62

7. Recommendations for LES Simulations

This chapter provides a few guidelines when dealing with large eddy simulations in general as well as specically in OpenFOAM.
7.1

Discretisation Schemes

The discretisation practices successfully used during this work are given
in the following table 7.1.1. The scheme names correspond to the names
OpenFOAM uses. Other schemes might be possible for certain terms (such
as the least squares practice for gradients) but have not been tested.
Term
Time derivative
Gradients
Convection term
Diusion term
Laplacian terms
Surface normal gradients

Type
Euler or backward
Gauss linear
Gauss linear or Gauss limitedLinear 1.0
Gauss linear
Gauss linear corrected
corrected

Table 7.1.1: Discretisation Schemes

When using the solver developed in this work, the discretisation of the
time derivative must be set to the Euler scheme. The third-order Runge
Kutta scheme is hardcoded inside the solver and only works if this setting is
chosen. The additional temperature scalar is also solved using this scheme
and requires Euler as well. All other scalar transport equations, eg. used
by the turbulence models, should be solved using the backward scheme
(which is second order in time).

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_7, Springer Fachmedien Wiesbaden 2016

For the convection term of the momentum equations, only the Gauss
linear scheme, which corresponds to central dierencing, should be used
for LES simulations. The central dierencing scheme is non-dissipative
and therefore allows an accurate representation of the rapidly changing
ow eld. Oscillations may occur as the scheme is unbounded, but these
are not as critical. Any schemes that add articial diusion to keep the
solution bounded are not recommended. The added diusion may easily be
of the same order as the turbulence model contribution and will strongly
aect the solution. For additional scalars the limitedLinear 1.0 scheme has
proven to be eective. Other schemes might be more suitable but this has
not been tested.

7.2

Linear Solvers

The only linear system that needs to be solved when using the explicit solver
is the pressure equation. As seen in the section on parallel performance on
page 59 the PCG solver strongly outperforms the GAMG solver at high
numbers of processor cores. In this work the PCG solver, in conjunction
the FDIC preconditioner, has been used exclusively and performed very
well. The pressure equation was generally solved to an absolute residual
tolerance of 1e-6, which usually resulted in continuity errors lying in the
range of 1e-12. The solver tolerance might be relaxed to speed up the
simulation but the eect on the solutions has not been studied.

7.3

Initial Values

As far as initial values goes, it is generally advisable to start from a RANS

solution. This solution does not have to be fully converged but should
approximately represent the main ow eld structures. If a RANS solution
is not available and cannot be obtained it is advisable to use the perturbU
utility in OpenFOAM to generate articial, unstable turbulence structures.
While this utility was developed primarily with fully developed channel
ows in mind, it is also partially applicable to general cases. It is usually
not feasible to start from constant solution elds as the fully turbulent state
may take a very long time to develop.
64

7.4

Grid Resolution

The topic of grid resolution is widely discussed and apart from some general
guidelines, no clear consensus is apparent in the literature as to what a
proper LES-grid is. The preliminary work by Gaitonde [13] tries to address
this issue and oers a methodology whereby the grid size required can
be estimated a priori by evaluating the turbulent quantities of a RANS
solution. The work mentioned above was published in 2008 and is part
of a supposedly ongoing doctorate thesis, as the nal results are not yet
available.
One region where there seems to exist a general agreement on the mesh
resolution requirements is near the wall. The dimensionless wall spacings
(explained in section 5.2.4 on page 28) in the three coordinate directions
should lie in the following regions (again see [13]):
x+ 35 55
+

y 1
+

z 15

(7.1)
(7.2)
(7.3)

The paper by di Mare et al. [9] provides an overview of quality criteria

that can be applied to large eddy simulations. According to the authors
these quality indices should only be used as estimation tools and not as
denite criteria. Two of these quality indices from this source are presented here as they are the most easily applicable in general cases (the original
paper focuses strongly on the ow in internal combustion engines). These
indices give an estimation for the amount of modelling present in the simulation rather than an actual quality value.
The rst quality index IQ is based on turbulent viscosity t and is
dened as follows:
t
(7.4)
IQ =
+ t
Here a value of 0.5 would indicate that the turbulent viscosity is of
the same magnitude as the laminar viscosity. To give an idea about the
magnitude of these values IQ has been calculated and averaged (based on
the cell volumes) for the coarse and the ne channel case using the sigma
subgrid model at the last timestep. For the coarse grid a value of 0.31 has
been obtained while the ne grid gave a value of 0.18.
65

The second index shown here calculates a ratio between a characteristic

mesh length scale and an eective Kolmogorov scale. This ratio determines
how close the simulation is to a fully resolved DNS simulation.
IQ =

ef f

(7.5)

Here is the mesh length scale and simply equal to the lter width
used in the model. The value for ef f can be calculated using estimates
for turbulent and numerical dissipation. The procedure is explained in [9].
A third and very promising approach to obtain a measure for the adequacy of the grid was suggested by Pope [34]. This approach was also
used in an automatic grid renement procedure in [28]. Pope introduces
the following quantity:
k
M=
(7.6)
K +k
where
k = 0.5r T

 2
u2T
K = 0.5 
u T 

(7.7)
(7.8)

is the resolved velocity and the angled

Here r is the subgrid stress tensor, u
brackets denote temporal averaging over a period of T . The quantity M
gives an estimate of the ratio between modeled and resolved turbulence. A
value of one would indicate full modelling (i.e. RANS) and a value of zero
corresponds to a DNS simulation. A commonly accepted value for M in a
LES simulation is 0.2 (see [18]), which results in about 80% resolution of
the turbulent spectrum.
7.5

Mesh Quality

The following list is a selection from the Ercoftac Best Practice Guidelines
[11] and is taken verbatim from [13]. It provides a good overview of general
mesh quality aspects to be aware of. While anyone dealing with computational uid dynamics should be aware of these points, they are given
here as they are especially true in the context of large eddy simulations.
Bad quality meshes, such as aspect ratios over 50 in the ow eld, which
are acceptable in stationary RANS simulations, can actually lead to solver
failure.
66

1. Selection of a mesh global topology to help satisfy the specic codes

requirements with regard to skewness, aspect ratio and expansion
ratios.
2. Extent of the computational domain to be chosen to capture relevant ow and geometrical features. If required, the sensitivity of the
calculation to the choice of computational domain to be examined.
3. Assessment of the geometrical features that can be omitted. In areas
requiring ne local detail, considering local grid renement. When
employed, additional grid points should lie on the original geometry
and not simply be a linear interpolation of more grid points on the
original coarse grid.
4. Highly obstructed zones or ne detail of obstructions, if any, to be
accounted for by the use of distributed losses or porosity.
5. Highly skewed cells to be avoided. For hexahedral cells or prisms the
grid lines should be optimised in such a way that the included angles
are approximately 90 degrees. Tetrahedra should tend to have their
four angles equal.
6. Highly warped cells, that is cells with large deviations from co-planar
faces to be avoided. Warp angles (measured between the surface
normals of triangular parts of the faces) greater than 75 degrees can
lead to serious convergence problems and deterioration in the results.
7. Non-orthogonal cells near boundaries to be avoided. This requirement is stronger than the requirement for the non-orthogonal cells
away from the boundaries.
8. Tetrahedral elements to be avoided in boundary layers. Prismatic or
hexahedral cells to be preferred because of their regular shape and
ability to adjust in accordance with the near-wall turbulence model
requirements.
9. High aspect ratios to be avoided in important regions of the ow
domain but may be large in non-critical regions. This restriction
may be relaxed near walls.
67

10. Specic code requirements for cell mesh stretching or expansion ratios
to be observed. The change in mesh spacing should be continuous
and mesh size discontinuities are to be avoided, particularly in regions
of large changes.
11. Automatic grid adaptation techniques oered by some codes to be
noted, as they might not always improve the grid quality (skewness,
aspect ratio).
12. Critical regions with high ow gradients or with large changes (such
as shocks, high shear, signicant changes in geometry or where suggested by error estimators) to have a ner and a more regular mesh
in comparison with non critical regions. Local mesh renement to be
employed in these regions in accordance with the selected turbulence
wall modelling.
13. When using periodic boundary conditions, high geometric precision
of the periodic grid interface to be ensured.
14. Arbitrary mesh coupling, non-matching cell faces, grid renement
interfaces or extended changes of element types to be avoided in the
critical regions of high ow gradients.
15. Assumptions made when setting up the grid with regard to critical
regions of high ow gradients and large changes to be checked with
the result of the computation and grid points to be rearranged if
found to be necessary.
16. Grid dependency study to be employed to analyse the suitability of
the mesh and to give an estimate of the numerical error in the simulation.

68

8. Conclusions

In this work, the basic third-order time-accurate explicit solver for OpenFOAM previously developed in [19] was improved in such a way that it is
now possible to simulate complex cases using the technique of large eddy
simulation (LES). As a rst step, the theoretical background of the LES
procedure was explained. Afterwards, the turbulence models used in the
context of this work are briey described. More detailed information is
given on the sigma subgrid model developed by [30], which was implemented from scratch as a part of the present study. A pressure-velocity
coupling algorithm based on the works by Rhie and Chow [37] was implemented. It could be shown that the new algorithm eectively reduced
pressure-oscillations at low timesteps and still achieved the predicted thirdorder accuracy in time.
The support for eddy-viscosity based LES turbulence models was added,
with the limitation that models using additional scalar equations can only
be treated second-order accurate in time. This has, however, not shown to
be problematic for the quality of the results.
For the simulation of periodic cases, a switchable source term for the
momentum equations was added that allows the prescription of a desired
bulk velocity which is held upright by continuous correction of the ow
eld.
A temperature-transport equation was added for simulations involving
heat transfer in incompressible ows. This additional scalar transport equation is solved in the same framework as the pressure and momentum equations, therefore allowing for a third-order accurate solution in time. Again,
an articial source term is available to deal with periodic ows involving
heat ow in or out of the domain.
The solver was then validated on a series of test cases using measured

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_8, Springer Fachmedien Wiesbaden 2016

data or data obtained by direct numerical simulations (DNS). The test cases
were performed using a variety of dierent turbulence models described
earlier.
The rst case involved the simulation of the ow between two parallel
plates. The results of averaged velocity and stress proles were compared
to DNS data by [25] and good agreement could be obtained, especially in
the case of the ne grid.
Secondly, the ow around a square cylinder was investigated and aerodynamic data such as drag coecients and Strouhal numbers were compared against measurements by [7] and simulation results by other authors.
The sigma model produced good results in terms of averaged velocity proles and drag coecient, but all models predicted a Strouhal number that
was approximately 4% higher than the measured value.
In the third case a periodic ribbed channel with heated walls was investigated and compared against measurements by [1]. Good agreement
could be obtained for the Nusselt number proles shapes, while the absolute values diered between the turbulence models. Very good results
were achieved concerning the location of the reattachment point of the
ow behind the rib.
The last case involved the simulation of a lm cooling geometry where
the results were compared against measurements conducted by [40]. The
results were acceptable but the case needs further work. Good results are
expected in the near future.
Additionally, a test case to assess the parallel performance of the solver
was run. It was shown that the new pressure-velocity coupling algorithm
was slightly slower than the old one but scales just as well. It could also be
observed that the PCG linear solver using the FDIC preconditioner performed signicantly better than the geometric-algebraic multigrid (GAMG)
solver at high number of cpu cores.
Finally, based on the experiences gained in this work, recommendations
for future simulations regarding numerical schemes, solvers, grid quality
and grid resolution are given.

70

9. Unresolved Issues

During the course of this work a few issues have surfaced concerning solver
stability. The implementation of the new pressure-velocity coupling scheme
has delivered excellent results in the channel, square cylinder and ribbed
channel cases but has lead to solver divergence in the lm cooling simulation. The velocity values inside the cooling hole were increasing beyond the
stability limit given by the CFL condition, therefore causing the simulation
to abort. This phenomenon could be observed on dierent grids and even
at very low timesteps which lead to the conclusion that it is not just a matter of stability but a bug in the code. The fact that it only occurred in the
lm cooling case suggests that one of the terms on the non-orthogonality
correction of the pressure-velocity coupling is not working as intended. The
lm cooling case was the rst and only case where the geometry did not
allow for a fully orthogonal mesh, ie. the erroneous term was most likely
zero in all the previous simulations. This will be investigated further.
Another source for instability was the use of time-varying boundary conditions, again a feature that was only used in the lm cooling case. Here
the problem was much less obvious and a simulation could run for several
ten thousand timesteps and only crash later. The issue occurred when using the mapped boundary condition which copies values from inside the
domain back onto the boundary, therefore creating a quasi-innite channel
at the inlet providing fully developed and physically meaningful turbulent
boundary conditions. The divergence was observed using the new as well
as the old pressure-velocity coupling method. As a rst step the temporal
scheme was reduced to a rst-order Euler scheme and the problems remained. Another boundary condition creating randomised values around a
prescribed mean was tested as well and also lead to instabilities. The discretisation scheme was then changed to rst-order Euler implicit and the

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_9, Springer Fachmedien Wiesbaden 2016

oscillations leading to the eventual failure disappeared completely. It is assumed that at the moment the boundary conditions are not updated to the
proper values either before or after the solution of the pressure equation.
At the moment it is unclear which solution variable causes the instability,
but since the boundary condition for the pressure is a simple zero-gradient
condition it is evident that either the velocity itself or the boundary ux
eld is wrong and needs to be treated specially.
Both these issues require attention, as both the treatment of nonorthogonal grids as well as time-varying boundary conditions are important
for industrial ow problems.

72

10. Outlook

As a rst step it is imperative that the problems described in the previous

section are addressed.
Further, the ongoing discussion on mesh resolution will be followed, especially the results of the work by [13] as, if successful, they could provide a
powerful tool to estimate the required LES grid size based on a RANS simulation. In that context, the meshing procedure described in the same document should be implemented in OpenFOAM. Similar meshes can already
be produced by the SnappyHexMesh utility but the cells would not be able
to have the staggered arrangement shown in [13], therefore only allowing
coarsening factors of two from layer to layer. The BlockMesh utility could
be rewritten but it would be limited to simple geometries. Ideally, a utility
would be created that takes an existing hexahedral mesh and renes the
near wall region in the appropriate fashion.
One topic that has been entirely omitted in this work is the use of hybrid
RANS-LES models, due to the fact that all simulations here used meshes
that fully resolved the wall region. For complex geometries and especially
at high Reynolds numbers such methods could reduce the computational
eort of time dependent simulations considerably. For accurate simulation
results, the near wall resolution in wall normal direction would still have to
be of the same order as in LES simulations but the resolution in streamwise
and transverse direction could be relaxed dramatically. One of the most
promising models of this type is the improved detached eddy simulation
(IDDES) developed by Shur et al. [39]. It provides a framework for the
blending of RANS and LES models and is implemented in OpenFOAM
using the Spalart Allmaras (see [42]) formulations. The framework can be
used with any existing RANS and LES models and could, eg. be coupled
with the sigma LES model implemented in this work. The existing model

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6_10, Springer Fachmedien Wiesbaden 2016

of OpenFOAM has ben applied to the coarse channel mesh and the ribbed
channel case. The results (see below) look promising but need further
investigation. The Nusselt number prole is good, even if the absolute
value is rather far away from the measurements. This will also have to do
with the fact that the turbulent Prandtl number needs to be adjusted for
hybrid models.
20

3.5

DNS Data
IDDES

DNS Data
IDDES

3
2.5
u+ , v + , w+ [-]

u+ [-]

15

10

2
1.5
1

0.5
0
101

100

101
y + [-]

102

Figure 10.0.1: Mean Velocity Proles,

IDDES

20

40

60

80

100

y + [-]

Figure 10.0.2: Mean Velocity Fluctuations, IDDES

As a nal note, the implementation of stable spatial third-order procedure for unstructured grids would provide enormous benets and should
be the focus of future investigations. It would also give OpenFOAM an
edge over commercial products and would probably increase its acceptance
further.

74

DNS Data
IDDES

0.2

+ [-]

0.4

0.6

0.8

50

100

150

200
y + [-]

250

300

350

400

Figure 10.0.3: Mean Shear Stress, IDDES

200

Measurement
IDDES

180
160

N u [-]

140
120
100
80
60
40

10 12
x/e [-]

14

16

18

20

Figure 10.0.4: Nusselt Number, IDDES

75

Bibliography

[1] S. Acharya, S. Duttta, T. A. Myrum, and R. S. Baker. Periodically developed ow and heat transfer in a ribbed duct. International Journal
of Heat and Mass Transfer, 36(8):20692082, 1992.
[2] Ansys Inc. Fluent Users Guide, 2006.
[3] A. Azzi and B.A. Jubran. Numerical modeling of lm cooling from
short length stream-wise injection holes. Heat and Mass Transfer,
39:345353, 2003.
[4] Cosimo Bianchini.
http: // www. cfd-online. com/ Forums/
openfoam/ 96885-compressible-wale-model. html# post377749 .
University of Florence, 2012.
[5] J. Blazek. Computational Fluid Dynamics: Principles and Applications. Elsevier, England, 2001.
[6] F. Cadieux, G. Castiglioni, J. A. Domaradzki, T. Sayadi, S. Bose,
M. Grilli, and S. Hickel. LES of separated ows at moderate reynolds
numbers appropriate for turbine blades and unmanned aero vehicles.
In International Symposium on Turbulence and Shear Flow Phenomena, Poitiers, France, 2013.
[7] S. Einav W. Rodi D. Lyn and J. Park. Laser doppler velocimetry
study of ensemble-averaged characteristics of the turbulent near wake
of a square cylinder. Journal of Fluid Mechanics, 304:285319, 1995.
[8] Eugene de Villiers. The Potential of Large Eddy Simulation for the
Modeling of Wall Bounded Flows. PhD thesis, Imperial College of
Science, Technology and Medicine, 2006.
77

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6, Springer Fachmedien Wiesbaden 2016

[9] Francesca di Mare, Robert Knappstein, and Michael Baumann. Application of LES-quality criteria to internal combustion engine ows.
Computers & Fluids, 89:2002013, 2014.
[10] E. R. Van Driest. On turbulent ow near a wall. Journal of the
Aeronautical Sciences, 23(11):10071011, 1956.
[11] ERCOFTAC. Best practice guidelines. Technical report, Special Interest Group on Quality and Trust in Industrial CFD, 2000.
[12] J. H. Ferziger and M. Peric. Computational Methods for Fluid Dynamics. Springer, Germany, 2002.
[13] Ulka Gaitonde. Quality criteria for large eddy simulation, rst year
transfer report. Technical report, School of MACE University of
Manchester, 2008.
[14] R. Henniger, D. Obrist, and L. Kleiser. High-order accurate solution
of the incompressible Navier-Stokes equations on massively parallel
computers. Journal of Computational Physics, 229:35433572, 2010.
[15] H. Jasak. Error Analysis and Estimation for the Finite Volume Method
with Applications to Fluid Flows. PhD thesis, Imperial College of
Science, Technology and Medicine, 1996.
[16] Blake Johnsson, Khai Zhang, Wei Tian, and Hui Hu. An experimental
study of lm cooling eectiveness by using PIV and PSP techniques.
In 51st AIAA Aerospace Sciences Meeting including the New Horizons
Forum and Aerospace Exposition, Grapevine, Texas, 2013.
[17] W. Kim and S. Menon. Application of the localized dynamic subgridscale model to turbulent wall-bounded ows. Technical report, American Institute of Aeronautics and Astronautics, 1997.
[18] M. Klein. An attempt to assess the quality of large eddy simulations
in the context of implicit ltering. Flow, Turbulence and Combustion,
75(1):131147, 2005.
[19] David Roos Launchbury. Development of an explicit solver for incompressible ows in OpenFOAM. Semester work, University of Applied
Sciences and Arts, Lucerne, 2014.
78

[20] A. Leonard. Energy cascade in large-eddy simulations of turbulent

uid ows. Advances in Geophysics, 18:237248, 1974.
[21] D. Lyn and W. Rodi. The apping shear layer formed by ow separation from the forward corner of a square cylinder. Journal of Fluid
Mechanics, 267:353376, 1994.
[22] W. Malalasekera and H. K. Versteeg. An Introduction to Computational Fluid Dynamics: the Finite Volume Method. Pearson Education,
England, 2007.
[23] Charles Meneveau, Thomas S. Lund, and William H. Cabot. A lagrangian dynamic subgrid-scale model of turbulence. Journal of Fluid
Mechanics, 319:353385, 1996.
[24] P. Moin, K. Squires, W. Cabot, , and S. Lee. A dynamic subgrid-scale
model for compressible turbulence and scalar transport. Physics of
Fluids A, 3(11):27462757, 1991.
[25] Robert D. Moser, John Kim, and Nagi N. Mansour. Direct numerical
simulation of turbulent channel ow up to Re = 590. Physics of
Fluids, 11(4):943945, 1999.
[26] Robert D. Moser, John Kim, and Nagi N. Mansour.
turbulence. ices. utexas. edu/ MKM_ 1999. html , 2007.

http: //

[27] S. Murakami and A. Mochida. On turbulent vortex shedding ow past

a 2d square cylinder predicted by cfd. Journal of Wind Engineering
and Industrial Aerodynamics, 54-55:191211, 1995.
[28] A. Naudin, L. Vervisch, and P. Domingo. A turbulent-energy based
mesh renement procedure for large eddy simulation. Technical report,
INSA - Rouen, 2007.
[29] F. Nicoud and F. Ducros. Subgrid-scale stress modelling based on the
square of the velocity gradient tensor. Flow, Turbulence, and Combustion, 62(3):183200, 1999.
[30] Franck Nicoud, Hubert Baya Toda, Olivier Cabrit, Sanjeeb Bose, and
Jungil Lee. Using singular values to build a subgrid-scale model for
large eddy simulations. Physics of Fluids, 23, 2011.
79

[31] K. Nozawa and T. Tamura. Les of ow past a square cylinder using

embedded meshes. Technical report, Izumi Research Institute and
Tokio Institute of Technology, Japan.
[32] J.S. Ochoa and N. Fueyo. Large eddy simulation of the ow past
a square cylinder. Technical report, Area de Mecanica de Fluidos,
Centro Politecnico Superior, Spain.
[33] Alberto Passalacqua.
Iowa, USA, 2014.

https: // github. com/ AlbertoPa .

Ames,

[34] S. B. Pope. Ten questions concerning the large-eddy simulation of

turbulent ows. New Journal of Physics, 6(35), 2004.
[35] M. Porquie, M. Breuer, and W. Rodi. Computed test case: square
cylinder. Technical report, Institute for hydromechanics, University of
Karlsruhe, Germany.
[36] Osborne Reynolds. On the dynamical theory of incompressible viscous
uids and the determination of the criterion. Philosophical Transactions of the Royal Society of London, 186:123164, 1895.
[37] C. M. Rhie and W. L. Chow. Numerical study of the turbulent ow
past an airfoil with trailing edge separation. AIAA Journal, 21:1525
1532, 1983.
[38] P. J. Roache. Spectral methods for the navierstokes equations with
one innite and two periodic directions. Journal of Fluids Engineering,
124(1):410, 2001.
[39] Mikhail L. Shur, Philippe R. Spalart, Mikhail Kh. Strelets, and Andrey K. Travin. A hybrid rans-les approach with delayed-des and
wall-modelled les capabilities. International Journal of Heat and Fluid
Flow, 29:16381649, 2008.
[40] A. K. Sinha, D. G. Bogard, and M. E. Crawford. Film-cooling eectiveness downstream of a single row of holes with variable density ratio.
ASME Journal of Turbomachinery, 113:442449, 1991.
[41] J. Smagorinsky. General circulation experiments with the primitive
equations. Mon. Weather Rev., 91:99, 1963.
80

[42] P. R Spalart and S. R. Allmaras. A one-equation turbulence model

for aerodynamic ows. AIAA Paper, (92-0439), 1992.
[43] R. Verstappen and A. Veldman. Fourth-order DNS of ow past a
square cylinder: First results. Technical report, Department of Mathematics, University of Groeningen, The Netherlands.
[44] Feng Wan, Fernando Porte-Agel, and Rob Stoll. Evaluation of dynamic subgrid-scale models in large-eddy simulations of neutral turbulent ow over a two-dimensional sinusoidal hill. Technical report, Saint
Anthony Falls Laboratory, Department of Civil Engineering, University of Minnesota Twin Cities, Minneapolis, MN, USA, 2006.
[45] G. Wang and S.P. Vanka. LES of ow over a square cylinder. Technical report, Department of Mechanical and Industrial Engineering,
University of Illinois at Urbana-Champaign, USA.

81

A. Appendix: Recent Developments

This section gives a brief overview of the activities performed after the
initial report was handed in. These points were not part of the original
thesis and are simply mentioned here for completeness.
The instability issue on non-orthogonal grids was xed. The issue was
that on such grids the normal boundary uxes at wall boundaries was nonzero, therefore leading to conservation issues and ultimately simulation
failure.
The second problem concerned time-varying boundary conditions causing instabilities. As suspected the old time values were the cause of the
problem. The boundary values of the old time elds inside the RungeKutta cycle were never updated. This was not a problem in the case of
cyclic, zero gradient or uniform xed values, however, as soon as the xed
boundary values changed over time, the algorithm would fail.
The lm cooling case was investigated further as well. The issues concerning the blowing rate could not be fully resolved due to the incompressibility assumption being inadequate. However, as the gures A.1 and A.2
below show, the prediction of the spreading of the cooling lm downstream
could be improved dramatically. For further case details please see chapter
5.5.

D. R. Launchbury, Unsteady Turbulent Flow Modelling and Applications, BestMasters,

DOI 10.1007/978-3-658-11912-6, Springer Fachmedien Wiesbaden 2016

0.6

Measurements
Simulation

[-]

0.4
0.2
0

0.2

0.4

0.6

0.8
z/D [-]

1.2

1.4

1.6

Figure A.1: Adiabatic Eectiveness at x/D = 6

0.6

Measurements
Simulation

[-]

0.4
0.2
0

0.2

0.4

0.6

0.8
z/D [-]

1.2

1.4

Figure A.2: Adiabatic Eectiveness at x/D = 15

84

1.6