IT 14 072

Examensarbete 15 hp
December 2014

Evaluation of RANS turbulence

models for the simulation of channel
flow

André Hedlund

Institutionen för teknikvetenskaper

Department of Engineering Sciences
 Abstract
Evaluation of RANS turbulence models for the
simulation of channel flow
André Hedlund

Teknisk- naturvetenskaplig fakultet

UTH-enheten The objective of this report is to investigate how
RANS models perform on fully developed channel
Besöksadress: flow, for Re = 13 350, and the simulations are made
Ångströmlaboratoriet
Lägerhyddsvägen 1 with the open source software OpenFOAM. The
Hus 4, Plan 0 velocity and turbulent kinetic energy profiles are
compared with previously published DNS results. A
Postadress: short introduction to turbulence modelling is
Box 536
751 21 Uppsala presented with focus on channel flow and the
boundary layer. In total eleven models are
Telefon: evaluated, and the results are of varying quality. A
018 – 471 30 03 convergence study is presented for two models,
Telefax: and reveals that the expected second order
018 – 471 30 00 convergence is fulfilled for one of them, whereas the
study for the other model is more ambiguous
Hemsida: without a clear conclusion. The OpenFOAM case
http://www.teknat.uu.se/student
setups for each model and the results gathered
from the simulations are publicly available.

Handledare: Mattias Liefvendahl

Ämnesgranskare: Gunilla Kreiss
Examinator: Jarmo Rantakokko
IT 14 072
Tryckt av: Reprocentralen ITC
Contents

1 Introduction 2

2 Theory of turbulence modelling 3

2.1 The turbulent viscosity hypothesis . . . . . . . . . . . . . . . 4
2.2 Fully developed channel flow . . . . . . . . . . . . . . . . . . 4
2.3 The boundary layer . . . . . . . . . . . . . . . . . . . . . . . 6
2.4 Turbulent viscosity models . . . . . . . . . . . . . . . . . . . . 7
2.4.1 k " model . . . . . . . . . . . . . . . . . . . . . . . . 7
2.4.2 k ! model . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4.3 Wall functions . . . . . . . . . . . . . . . . . . . . . . 9

3 Simulation 11
3.1 Test case setup . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 The mesh . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.1.2 Boundary and initial conditions . . . . . . . . . . . . . 12
3.1.3 Low Re-model setup . . . . . . . . . . . . . . . . . . . 14
3.1.4 High Re-model setup . . . . . . . . . . . . . . . . . . . 14
3.2 Convergence of the residuals . . . . . . . . . . . . . . . . . . . 15

4 Results 16
4.1 RANS model investigation . . . . . . . . . . . . . . . . . . . . 16
4.1.1 Low Re-models . . . . . . . . . . . . . . . . . . . . . . 16
4.1.2 High Re-models . . . . . . . . . . . . . . . . . . . . . . 18
4.2 Grid convergence study . . . . . . . . . . . . . . . . . . . . . 20

5 Discussion and conclusions 23

Acknowledgment 25

1
1 Introduction
Turbulence occur in many natural flows and is of paramount importance in
many engineering applications. The Navier-Stokes equations describe the
dynamics of fluids and can rarely be solved analytically. For this reason, we
are to a large extent dependent on numerical simulation for making fluid
predictions. Flow at a sufficiently low Reynolds (Re) number is laminar and
the Navier-Stokes equations can then be numerically solved directly. For
high Re-numbers, which typically occur in applications, the flow is turbulent
and hence consists of fluid motion with a wide range of spatial and temporal
scales. To numerically solve a turbulent flow it is important resolve the
flow such that the smallest scales can be represented, therefore a direct
numerical simulation (DNS) of turbulent flow is very costly and in most
cases unfeasible, see [1]. In the Reynolds-averaged Navier-Stokes (RANS)
formulation, the velocity is decomposed into the mean velocity distribution,
and the fluctuations around this mean. The equations are then solved for the
mean velocity, and the e↵ect of the fluctuations is modelled, see [2]. With
this approach it is not necessary to resolve the smallest turbulent scales –
and the computational cost can therefore be reduced significantly.
Channel flow is the case when a fluid flows between two parallel plates,
such as a rectangular duct. Turbulent channel flow is a well investigated
case both experimentally and numerically. Early DNS results for channel
flow were presented by Kim, Moin and Moser [3], and a continuation of this
work, which includes higher Re-numbers, where presented by Moser, Kim
and Mansour [4].
The objective of this report is to evaluate the performance of several
RANS models for incompressible turbulent fully developed channel flow.
The focus is to evaluate the models with respect to the velocity profile,
turbulent kinetic energy profile and wall shear stress, and if it is possible
rank the models based on how they predict the mentioned quantities. The
results will be compared with theoretical approximations (law of the wall)
and the DNS results of Moser, Kim and Mansour [4].
The simulations are carried out with the OpenFOAM toolbox, an open
source software for solving continuum mechanics problems [5]. There are
several RANS models implemented in OpenFOAM and each model has been
developed with a di↵erent objective in mind. This report will not try to
explain in depth why a certain model fails or succeeds in approximating the
quantities investigated, it will simply state that a model failed or succeeded.
The results from the simulation and the OpenFOAM case setups are
publicly available at https://github.com/AndreHed/channelFlow.git.

2
2 Theory of turbulence modelling
This section describes the theory of turbulent flows necessary for describing
turbulence modelling for fully developed channel flow.
The continuity equation is given by

r·U=0 (1)

and the Navier-Stokes equation is given by

DU @U 1
⌘ + U · rU = rp + ⌫r2 U (2)
Dt @t ⇢
where ⌫ = µ/⇢ is the kinematic viscosity. Here we have four unknowns: three
components of the velocity and the pressure; and we have four equations:
the continuity equation (eq. (1)) and the three components of Navier-Stokes
equations (eq. (2)). Hence, the system is closed and we can solve for the
unknowns. [1]
As the Re-number increases the flow enters the turbulent regime, and
smaller and smaller eddies will form. The eddies correspond to small random
fluctuations in the variables describing the flow. Therefore, in a turbulent
flow the velocity field, U(x, t), can be expressed by a mean and a fluctuating
part, viz.
U(x, t) = hU(x, t)i + u(x, t), (3)
this will be referred to as the Reynolds decomposition. The mean is given
by time averaging
Z
1 T
hU(x, t)i = lim U(x, t) dt, (4)
T !1 T 0

which is an applicable averaging for the stationary flows considered in this

report.
We can take the mean of equation (1) to obtain

@ hUi i
= 0. (5)
@xi
Taking the mean of equation (2) yields

D̄ hUj i @ hUj i @ hUj i @ hui uj i 1 @ hpi

⌘ + hUi i = ⌫r2 hUj i , (6)
D̄t @t @xi @xi ⇢ @xj
which will be referred to as the Reynolds equations. The di↵erence between
the Navier-Stokes equations (eq. (2)) and the Reynolds equations (eq. (6))
is that the term ⌧ij = hui uj i is introduced, which is called the Reynolds
stresses and describe the stress arising from the fluctuating velocity field.
It is important to note that we have introduced six new unknowns (⌧ij is a

3
symmetric tensor) without introducing additional equations, thus we have
ten unknowns and only four equations – the system is therefore not closed. [2]
Half of the trace of the Reynolds stress is called the turbulent kinetic
energy
1
k ⌘ hui ui i . (7)
2
The turbulent kinetic energy can be described as the kinetic energy per
unit mass in the oscillating velocity field. It will be used frequently in the
remaining of the report and is one of the quantities that is investigated.

2.1 The turbulent viscosity hypothesis

The turbulent viscosity hypothesis resolves the unclosed problem that is
created by the introduction of the Reynolds stresses. It states
✓ ◆
2 @ hUi i @ hUj i
hui uj i = k ij ⌫T + (8)
3 @xj @xi

where ⌫T is called the turbulent viscosity. The hypothesis inserted in equa-

tion (6) yields
 ✓ ◆ ✓ ◆
D̄ hUj i @ @ hUi i @ hUj i 1 @ 2
= ⌫e↵ + hpi + ⇢k (9)
D̄t @xi @xj @xi ⇢ @xj 3

where ⌫e↵ = ⌫+⌫T . Comparing this result with equation (2) we can conclude
that they are of the same form, therefore, if ⌫T is specified the modified
Reynolds equation (eq. (9)) is now in closed form.
There are two concerns with the hypothesis that should be discussed,

1. the accuracy of the hypothesis, and

2. the specification of the turbulent viscosity.

It has been proved that, unfortunately, the accuracy of the turbulent vis-
cosity hypothesis is poor for many flows, however it still serves a purpose
since it is reasonable for some simple flows. The turbulent viscosity can be
written as
⌫ T = l ⇤ u⇤ , (10)
where di↵erent approaches for specifying the length l⇤ and the velocity u⇤
are presented in the following sections. [1]

2.2 Fully developed channel flow

Channel flow (see figure 1) is a simple flow bounded by two parallel plates
separated by a distance of 2 and the flow direction is along the x-axis. For
fully developed flow the velocity is dependent of only y and it is statistically

4
 y

Flow h=2
x
z

Figure 1: Sketch of channel flow.

stationary and one-dimensional. Using the notation (U, V, W ) for the ve-
locity field, it is easy to conclude that hW i = 0. The Re-number for this
flow is given by
Ū · 2
Re = , (11)
⌫
where Ū is the bulk velocity defined as
Z
1
Ū ⌘ hU i dy. (12)
0

Because of the simplicity of the flow the mean continuity equation (eq. (5))
and the mean momentum equations (eq. (6)) can be greatly simplified. Using
the fact that hW i = 0 and that hU i is independent of x the mean continuity
equation reduces to
@Ui d hV i
= = 0. (13)
@xi dy
The boundary condition at the walls are hV iy=0 = hV iy=2 = 0, therefore
we obtain that hV i is zero for all y.
The lateral mean momentum equation is deduced to
⌦ ↵
1 d hpi d v 2
+ =0 (14)
⇢ dy dy
by noting that the nonlinear terms are neglectable due to the boundary-layer
approximation and that the axial derivatives of the Reynolds stresses are
small
⌦ 2 ↵ compared with the lateral derivatives. With the boundary condition
v y=0 = 0 we obtain

1 ⌦ ↵ 1
hpi + v 2 = pw (x) (15)
⇢ ⇢
where pw (x) = hp(x, 0, 0)i. Di↵erentiating equation (15) with respect to x
yields
d hpi dpw
= . (16)
dx dx

5
 The axial mean momentum equation for fully developed channel flow,
obtained in the same way as the lateral analog, is

@ @ hU i @ hpi
⇢⌫ ⇢ huvi = . (17)
@y @y @x
The total shear stress consists of the viscous stress and the Reynolds stress,
viz.
d hU i
⌧ (y) = ⇢⌫ ⇢ huvi . (18)
dy
Using equation (16)-(18) we finally obtain
@⌧ @pw
= . (19)
@y @x
Introducing ⌧w = ⌧ (0) as the wall shear stress, the boundary conditions for
equation (18) are ⌧ (0) = ⌧w and ⌧ (2 ) = ⌧w , which yields [1]
d⌧ ⌧w
= . (20)
dy

2.3 The boundary layer

The total shear stress (equation (18)) consists of two terms: the viscous
stress ⇢⌫d hU i /dy and the Reynolds stress ⇢ huvi. Due to the no slip
condition the Reynolds stress is zero at the wall and the shear stress is
dominated by the viscous stress in the vicinity of the wall. The wall shear
stress is therefore described by
✓ ◆
d hU i
⌧w = ⇢⌫ . (21)
dy y=0

Away from the wall the viscous stress is negligible compared with the Reynolds
stress and the total shear stress is therefore described by the Reynolds stress.
In the analysis of the flow close to the wall it is useful to define several
viscous scales from the viscosity ⌫, the wall shear stress ⌧w and the density
⇢, these are summarised in table 1.
Due to the reciprocal action between the viscous stress and the Reynolds
stress, di↵erent layers can be identified in the vicinity to the wall. For y + < 5
the viscous sublayer is apparent and the velocity adheres to a linear relation

u+ = y + . (22)

As the y + -values increases, the viscous stress becomes negligible compared

with the Reynolds stress. For y + > 30 there is a log-law region which obeys
the logarithmic law
1
u+ = ln y + + B, (23)


6
 Table 1: Viscous scales.

r
⌧w
Friction velocity u⌧ =
⇢
r
⇢ ⌫
Viscous length scale ⌫ =⌫ =
⌧w u⌧
u⌧
Friction Reynolds number Re⌧ = =
⌫ ⌫
+ y u⌧ y
Wall units y = =
⌫ ⌫
hU i
Viscous velocity u+ =
u⌧
k
Turbulent kinetic energy k+ = 2
u⌧

where  (von Kármán’s constant) and B are given by

 = 0.41, B = 5.2. (24)

The region between the two layers (i.e. 5 < y + < 30) is called the bu↵er
layer; in this region neither viscous nor turbulent processes dominate. These
layers are presented in figure 2, and the theoretical laws are compared with
DNS data. [1]

2.4 Turbulent viscosity models

The Reynolds decomposition (eq. (3)) separated the mean and the fluctua-
tions. RANS models solve only for the mean quantities, and are therefore
computationally much cheaper since there is no need to resolve the small
fluctuations of a turbulent flow. The Reynolds decomposition did however
introduce the Reynolds stresses, six new unknowns without additional equa-
tions, which resulted in an unclosed system. This section will describe how
to close the system by using the turbulent viscosity hypothesis and using
model transport equations to specify the turbulent viscosity ⌫T . There are
various other ways to close the problem that are not considered in this report
(e.g. Reynolds stress models and mixing-length model).

2.4.1 k " model

The k " model is a two-equation model - one transport equation is solved for
the turbulent kinetic energy, k, and one for the dissipation rate of turbulent
kinetic energy, ". When these model equations are solved, the turbulent

7
 y+ = 5 y + = 30
u+ = y +
u+ = 1 ln x + B
101 DNS

u+

100

viscous sublayer bu↵er layer log-law region

100 101 102

y+

Figure 2: The linear relation of the viscous sublayer and the logarithmic relation
of the log-law region compared with the DNS data of Moser, Kim and Mansour [4].

viscosity can be computed by

k2
⌫ T = Cµ , (25)
"
where Cµ is a model constant.
The exact equation for k is given by
D̄k
= r · T0 + P ", (26)
D̄t
where T0 is the flux of Reynolds stress and P is the rate of production of
turbulent kinetic energy, given by
@ hUi i
P= hui uj i . (27)
@xj
The mean derivative and P are in closed form and " will be modelled by a
separate equation, therefore only T0 need to be modelled, which can be done
with a gradient-di↵usion hypothesis. This results in the model transport
equation for k ✓ ◆
D̄k ⌫T
=r· rk + P ", (28)
D̄t k

8
where k is a model constant.
The dissipation rate is given by
⌧
@ui @ui
"=⌫ (29)
@xj @xj

and the model transport equation for " is given by

✓ ◆
D̄" ⌫T P" "2
=r· r" + C"1 C"2 . (30)
D̄t " k k
All the model constants introduced are given by

Cµ = 0.09, C"1 = 1.44, C"2 = 1.92, k = 1.0, " = 1.3. (31)

The inaccuracies of the model originates from the turbulent viscosity

hypothesis and the " equation. [1]

2.4.2 k ! model
The k ! model use a similar approach as the k " model, but instead of "
it uses the rate of dissipation of energy in unit volume and time, or merely
the specific dissipation rate, defined as
"
!= . (32)
k
The model equation for k is slightly modified

D̄k
= r · [(⌫ + ⌫T )rk] + P ", (33)
D̄t
and the model transport equation for ! is given by

D̄! P!
= r · [(⌫ + ⌫T )r!] + ↵ !2, (34)
D̄t k
where the constants are given by

= 12 , ↵ = 0.52, = 0.072. (35)

This is the classic Wilcox k ! model described in detail in [2].

2.4.3 Wall functions

For the models described, boundary conditions needs to be specified for the
velocity and the turbulent parameters. The boundary condition for the wall
can be problematic due to the fact that for y + < 30 the gradients of the
velocity and " are large, therefore the mesh needs to be very fine close to
the wall to accurately resolve the gradients – which can be very costly. If

9
the mesh is not fine enough the models described above deficiently predict
the flow behaviour close to the wall.
Wall functions are introduced to mitigate the computational cost, and
in the same time enable the model to predict accurate results. The wall
function is utilised between the wall and the mesh point closest to the wall
(matching point), and it will try to match the value at the matching point
with the law of the wall described in section 2.3. Each turbulent parameter
has its own wall function which is derived from the law of the wall.
It is important to emphasise that if the mesh is very refined (i.e. the
matching point lies before y + < 1) wall functions are not necessary. [2]

10
3 Simulation
As described in section 2.2 fully developed channel flow is a one-dimensional
problem. In OpenFOAM a case setup consists of: specifying the mesh, set-
ting the boundary and initial conditions, and specifying the physical prop-
erties. The di↵erent RANS models that are investigated in this report are
listed in table 2. The OpenFOAM solver that is used is boundaryFoam,
which is a steady-state solver for one-dimensional turbulent flows.

Table 2: RANS models implemented in OpenFOAM that are investigated.

OpenFOAM Type Description

Low Re LaunderSharmaKE 2-eq. Modified k " model
LamBremhorstKE 2-eq. Modified k " model
LienCubicKELowRe 3-eq. Modified k " model
LienLeschzinerLowRe 2-eq. Modified k " model
High Re kEpsilon 2-eq. See sec. 2.4.1
kOmega 2-eq. See sec. 2.4.2
kOmegaSST 2-eq. Combination of k " and k !
RNGkEpsilon 2-eq. Modified k " model
realizableKE 2-eq. Modified k " model
SpalartAllmaras 1-eq. Solves a transport eq. for ⌫˜
v2f 4-eq. Solves for k, ", hvi2 and f

3.1 Test case setup

Fully developed channel flow is obtained by specifying the bulk velocity of
the flow. The solver will add a pressure gradient to the momentum equation,
and after each iteration it will correct the pressure gradient such that the
specified Ū is maintained. The bulk velocity together with the kinematic
viscosity and half-channel width are the physical properties that completely
describe the flow (eq. (11)), and in OpenFOAM these are specified in the
file transportProperties. The values that are used in the simulations are
given in table 3.

Table 3: Physical properties for the simulation campaign.

1m
⌫ 2 ⇥ 10 5 m2 /s
Ū 0.1335 m/s
Re 13 350

11
 wall

cyclic 2 cyclic

z x wall

Figure 3: OpenFOAM setup for the one-dimensional fully developed channel flow.

3.1.1 The mesh

OpenFOAM solves the equations with the finite volume method, therefore,
the mesh is always three-dimensional. The fully developed channel flow is
simulated as a one-dimensional problem (see fig. 3), with only one cell in
the x- and z-direction. In y-direction the number of cells Ny is variable and
the cells are distributed geometrically such that there is a higher density of
cells near the wall than in the centre of the channel. In OpenFOAM a ratio
between the largest cell and the smallest cell is specified.
The half-channel width ( ), Ny and the grading are specified in the file
blockMeshProperties, and the mesh is created with the blockMesh utility.

3.1.2 Boundary and initial conditions

Since the case is one-dimensional, the sides facing the z-direction are set to
empty, which symbolises that no solution is required in that dimension. The
sides facing the flow direction, which could be seen as inlet and outlet, are
given a cyclic boundary condition. The remaining sides correspond to the
walls (see fig.3).
The boundary conditions at the wall is specified for each variable. For the
velocity the no slip condition, (U, V, W ) = (0, 0, 0), is applied. Section 2.4.3
described that wall functions may be utilised to reduce the computational
cost and improve the results. Wall functions are selected in the 0/ directory
for each variable that is needed; the various wall functions implemented in
OpenFOAM are presented in table 4.

12
 Table 4: Wall functions

⌫T nutUWallFunction Provides a boundary condition for ⌫T

based on velocity.
nutUSpaldingWallFunction Uses Spalding’s law to give a continu-
ous ⌫T profile, based on velocity.
nutLowReWallFunction For low Reynolds number models. It
sets ⌫T = 0, and provides a function to
calculate y + .
nutkWallFunction Based on turbulent kinetic energy.
k kqRWallFunction For k, q and R for high Reynolds num-
bers. Based on the zero-gradient con-
dition.
kLowReWallFunction Provides a wall function condition for
both low and high Reynolds numbers.
Operates in two modes based on y + .
✏ epsilonWallFunction For high Reynolds numbers.
epsilonLowReWallFunction Works for both low and high Reynolds
numbers. The model operates in two
modes based on an approximated y +
value.
! omegaWallFunction Works for both low and high Reynolds
numbers.
hvi2 v2WallFunction Works for both low and high Reynolds
numbers.

Initial conditions are required for the velocity field, k, " and !. The tur-
bulent kinetic energy is given by equation (7) which for isentropic turbulence
results in
3
k = u2 (36)
2
where the fluctuating velocity can be approximated by the product of the
bulk velocity and the turbulent intensity, viz.
u = ti Ū . (37)
For this case the turbulent intensity is approximately 5 %. The initial con-
dition for " is given by
3/4
Cµ k 3/2
"= , (38)
and for ! equations (25) and (32) yield
"
!= . (39)
Cµ k

13
 In OpenFOAM the boundary and initial conditions are specified in the
0/ directory for the variables that are required. The variables required are
determined by which solver and model that is used. [6]

3.1.3 Low Re-model setup

The low Reynolds number models are developed for flows with moderate
Re-numbers where the mesh is refined all the way to the wall, therefore no
wall functions are used. In table 4 there are, however, wall functions that
are specifically made for low Re-models; these are usually just a placeholder
for a regular boundary condition, such as zero gradient or no slip conditions.
In this case the wall function for ⌫T will produce the same results as setting
⌫T to be calculated from k and ".
Table 5 lists the settings for the low Re-models.

Table 5: Boundary conditions and mesh settings for low Re-models.

⌫T nutLowReWallFunction
k fixedValue = 0
" fixedValue = 0
U fixedValue = 0
Ny 400
grading 100

3.1.4 High Re-model setup

The high Reynolds number models are developed for high Re-numbers where
the utilisation of wall functions are necessary to limit the computational cost.
The Re-number that is investigated is relatively low and the logarithmic
boundary layer is limited to a couple of hundred y + units, therefore wall
functions are not necessary per se. In OpenFOAM the wall functions are
implemented such that they work e↵ectively for flows with low Reynolds
numbers as well, therefore they are used as boundary conditions for the
simulations. In table 6 the settings for the high Re-models are displayed.

14
 Table 6: Boundary conditions and mesh settings for high Re-models.

⌫T nutUWallFunction
k kqRWallFunction
" epsilonWallFunction
! omegaWallFunction
U fixedValue = 0
Ny 400
grading 100

3.2 Convergence of the residuals

Figure 4 shows the convergence of hU i, k and " with respect to the residuals
for one of the models that are investigated. The variables converge almost
monotonically until the residuals are reduced to a size where the machine
epsilon prevents further reduction.

1
hU i
10 k
"
10 4
Residual

10 7

10 10

10 13

10 16
0 10000 20000 30000 40000 50000
Iteration

Figure 4: The convergence of the initial residuals for the Launder-Sharma k "
model with Ny = 100.

15
4 Results
In this section the RANS models are evaluated with respect to how accurate
they predict fully developed channel flow compared with the DNS data of
Moser, Kim and Mansour [4]. A grid convergence study with respect to
the wall shear stress and velocity is also presented. The friction Reynolds
number of the investigated flow is Re⌧ = 395, which correspond to the
Reynolds number, Re ⇡ 13 350. The OpenFOAM test setups for each model
and the results gathered from the simulations are available at https://
github.com/AndreHed/channelFlow.git.

4.1 RANS model investigation

The RANS models that are investigated in this report are given in table 2.
The results presented are divided into low and high Re-models, this is mainly
because the problem setup di↵ers between the two groups.
In table 7 the friction velocity, friction Reynolds number and smallest
y + value are presented for all models that have been investigated. Note that
while the desirable friction Reynolds number is Re⌧ = 395, the resulting Re⌧
varies a lot. Note also that the models produce reasonably small y + values,
therefore the mesh is sufficiently refined at the walls.

Table 7: The obtained friction velocity, friction Reynolds number and first y +
value for the low Reynolds number models.

Model u⌧ /mm s 1 Re⌧ y + /10 2

Low Re Launder-Sharma k " 7.19 359 4.15

Lam-Bremhorst k " 7.66 383 4.42
Lien cubic k " 6.44 322 3.72
Lien-Leschziner 6.42 321 3.71
High Re k ! 7.87 393 4.54
k ! SST 7.71 386 4.45
Spalart-Allmaras 7.55 377 4.36
k " 10.7 535 6.18
RNG k " 10.5 525 6.06
Realizable k " 5.84 292 3.37
hvi2 f 4.45 223 2.57

4.1.1 Low Re-models

In figure 5 the velocity and turbulent kinetic energy profile for low-Re models
are compared with the DNS results. The results are normalised with both
the friction velocity and the bulk velocity (see table 1). The velocity profile

16
 20 4

u+ k+
10 2

0 0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/

101 1

hU i
u+
Ū
0.5

100
0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/
DNS; Launder-Sh.; Lam-Br.; Lien cubic; Lien-Le.

Figure 5: Simulating channel flow with low Reynolds number RANS models,
where Re = 13 350.

in plus units is accurate for all models for y + < 6, but for y + > 10 the
results vary significantly. Both the Lien cubic and Lien-Leschziner model
increase too much and at the channel centreline u+ is up to 20 percent o↵.
The Launder-Sharma and Lam-Bremhorst model is much better and deviate
approximately five percent or less at the centreline.
When the velocity is normalised with the bulk velocity the profiles are
instead inaccurate in the bu↵er layer, and at the centreline the di↵erence
is significantly less than when the results are normalised with the friction
velocity.
The results for the turbulent kinetic energy are also not accurate, espe-
cially where the maximum of the profile is. Note that an accurate result
for the velocity does not imply that the result for k is accurate as well. It
is also important to note that while the Lam-Bremhorst model produced

17
the best results for both quantities, it is the model that is the most difficult
to achieve convergence for. The other models converged with initial condi-
tions as described in section 3.1.2, whereas the Lam-Bremhorst model only
converged if the solution of one of the other models where used as initial
conditions for U , k and ", and even then the convergence of " is very slow.

4.1.2 High Re-models

Figure 6 plots the results for the k !, k ! SST and Spalart-Allmaras
models. These high Re-models predicted the quantities much better than
the other high Re-models which are presented later. The profiles for u+
shows the same behaviour as for the low Reynolds models in figure 5: for
y + < 6 the profiles agree very well with DNS data, whereas for y + > 10
the profiles are dispersed. Even though the profiles disperse, it is not to the
same extent as for some of the low Reynolds models, the velocity does not
vary significantly in comparison with the DNS data.
The results for k + are similar to the results of the low Reynolds number
models in the sense that they fail to reach the same peak as the DNS data.
Note that k is not required for the Spalart-Allmaras model, and therefore
no data is available.
When the velocity is normalised with the bulk velocity the profiles match
very well with the DNS data. The models in figure 6 are successful in
predicting the velocity in both the viscous sublayer as well as in the log-law
region.
In figure 7 the results for the k ", RNG k ", realizable k " and hvi2 f
models are presented. These models generally mispredicted the investigated
quantities. Realisable k " and hvi2 f were particularly bad since they
both had convergence issues and are dependent on the initial conditions.
Note the correlation between how the profile look in figures 6 and 7, and
the size of Re⌧ and u⌧ in table 7. For example, for k " and RNG k ",
both produce a large Re⌧ , the u+ profiles are very similar to each other; and
for RNG k " and hvi2 f , both produce a small Re⌧ , the u+ profiles have
the same characteristics.

18
 20 4

u+ k+
10 2

0 0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/

101 1

hU i
u+
Ū
0.5

100
0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/
DNS; k!; k!SST; Spalart-Al.

Figure 6: Simulating channel flow with high Reynolds number RANS models,
where Re = 13 350.

19
 20 4

u+ k+
10 2

0 0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/

101 1

hU i
u+
Ū
0.5

100
0
100 101 102 0 0.2 0.4 0.6 0.8 1
y+ y/

DNS; k"; RNG k"; Realizable k"; hvi2 f

Figure 7: Simulating channel flow with high Reynolds number RANS models,
where Re = 13 350.

4.2 Grid convergence study

The grid convergence was investigated with respect to the wall shear stress
⌧w and the velocity. When OpenFOAM computes the wall shear stress it
computes ⌧w /⇢, the values presented for the wall shear stress are therefore
divided by the density. Generally, the convergence rate is calculated by
|f2 f1 |
p = ln ln r (40)
|f3 f2 |
where r is the mesh refinement ratio, fi is the variable that is investigated
and i = 3 is the finest and i = 1 is the coarsest mesh. The number of
cells that are used across the channel are Ny = 200, 600 and 1800. Each
quantity is calculated at the cell centre, therefore in order to get the cell
centres to overlap for the di↵erent meshes the mesh refinement ratio is r =

20
3. No grading is imposed on the mesh since that would result in non-
overlapping cell centres for the di↵erent meshes. The solution is considered
to be converged when the residuals are below 1 ⇥ 10 9 .
The convergence study is performed for the Launder-Sharma and Spalart-
Allmaras models. Both of these models did produce accurate results (see
fig. 5 and 6), and are therefore reasonable models to investigate. In table 8
the results from the convergence study are presented, and equation 40 yields
for the wall shear stress

pLS = 1.81 pSA = 2.01. (41)

Note that the values for the wall shear stress are divided by density The
convergence rate for the velocity vary considerably across the channel it is
therefore more reasonable to plot the convergence rate, see figure 8. For
both models extrema occur, this is explained by that the velocity profiles of
the di↵erent meshes intersect each other. At the wall the no slip boundary
condition is imposed and the expected convergence rate in the vicinity of the
wall is expected to be p = 2. It is troubling that for Launder-Sharma the
convergence rate is predominantly negative, which indicates divergence. The
result for Spalart-Allmaras is approximately two across the entire channel,
which is what is to be expected for the numerical schemes that are utilised.

Table 8: Convergence study.

Ny y+ ⌧w /mm2 s 2 iterations
Launder-Sharma 200 2.07 55.5 50 000
600 0.608 56.1 300 000
1800 0.204 56.1 1 000 000
Spalart-Allmaras 200 1.92 58.8 50 000
600 0.630 57.2 500 000
1800 0.210 57.0 2 000 000

21
 4 Launder-Sharma k "
Spalart-Allmaras

2
p

4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
y

Figure 8: The convergence rate for the velocity across the channel.

22
5 Discussion and conclusions
The previous section presented the results of a simulation campaign made
with OpenFOAM for several RANS models, and the results were compared
with the DNS results of Moser, Kim and Mansour [4]. The results di↵er
significantly between the models, and while some models produce reason-
ably good results, others fail to accurately predict the investigated quanti-
ties. Apart from the flow quantities, the results of a grid convergence study
for two models were also presented. The study verified that the Spalart-
Allmaras model had second order convergence, whereas no convergence for
the Launder-Sharma k " model could be established. Now follows a discus-
sion of the results presented and statements about each models suitability
for simulating fully developed channel flow.
The results from the low Re-models agree well with the DNS data, but
there are significant di↵erences between the models. The Launder-Sharma
and Lam-Bremhorst models predicted the wall shear stress better than the
Lien cubic and Lien-Leschziner models, while for the turbulent kinetic energy
profile the Lam-Bremhorst and Lien cubic performed better than the other
two. This could be the basis of concluding that the Lam-Bremhorst model
is the superior model for this case, but as mentioned in section 4.1.1, the
model has severe convergence problems, to the point that it only converges
if the initial conditions for hU i, k and " are mapped from the solution of
another model such as Launder-Sharma.
The results from some of the high Re-models predicted the investigated
quantities well, whereas others mispredicted the profiles to the extent that
they are not acceptable. The first distinction that can be made between
the models, is that the models based on the transport equations for k and
" did not predict hU i, k and ⌧w very well, whereas the models based on the
transport equations for k and ! performed much better. The implementa-
tion of the k ! model in OpenFOAM is based on Wilcox’s (1988) k !
model (see OpenFOAM source code available at [5]). This implementation
is known to accurately predict the flow properties for wall bounded flow [2,
p. 128]. Since channel flow is a wall bounded flow, it is reasonable that the
results, for the models based on transport equations for k and !, agrees well
with the DNS results. The k " model is known to poorly predict boundary
layers with strong pressure gradients (see [1, p. 461]), which agrees with the
fact that the models based on the transport equations for k and " did not
predict the flow properties very well.
A surprising result is that the Spalart-Allmaras model performed very
well. Of the models that are investigated the Spalart-Allmaras model is
the only one-equation model and therefore the model with the simplest
turbulence description. The model predicted the velocity profile just as
accurately as the k ! and k ! SST.
The results of the “hvi2 f ” model are not good, which could be the

23
result of an incorrect setup for the model, therefore the results presented in
the report may not be representative for the model. A peculiar notion is that
the model is sensitive to the initial conditions, where a small perturbation
can produce a di↵erent solution, this has been noted before, see [7, p. 173].
The results of the models are ranked in table 9. The ranking is based
on the results presented in figures 5-7 and table 7, and only the top five
models are ranked. Based on the ranking in table 9, the RANS models that
are suitable for the simulation of fully developed channel flow are (in no
particular order): Launder-Sharma k ", k !, k ! SST and Spalart-
Allmaras.
Table 9: Ranking of the models for four di↵erent properties, the ranking is only
made for the five best models for each property. Note that ’⇤’ signifies that the
model did rank among the top five and ’-’ means that no data is available.

Model Re⌧ hU i k Convergence

Launder-Sharma k " 5 5 3 2
Lam-Bremhorst k " 3 2 1 *
Lien cubic k " * * 2 5
Lien-Leschziner * * * *
k ! 1 4 4 3
k ! SST 2 3 5 4
Spalart-Allmaras 4 1 - 1
k " * * * *
RNG k " * * * *
Realizable k " * * * *
hvi2 f * * * *

This investigation is for a constant Re-number, Re = 13 350, and further

investigations for higher Re-numbers would be worthwhile. The OpenFOAM
test setups for each model and the results gathered from the simulations
are available at https://github.com/AndreHed/channelFlow.git, read-
ers are encouraged to try the cases by themselves and any contribution or
comment is welcome.

24
Acknowledgment
I would like to thank my supervisor Mattias Liefvendahl for providing me
with the subject of this thesis, and for his guidance in during the process of
writing this thesis. I would also like to thank Timofey Mukha for taking his
time to answer my questions and providing his insights in the subject.

25
References
[1] S. Pope, Turbulent Flows. Cambridge University Press, 2000.

[2] D. Wilcox, Turbulence Modeling for CFD. DCW Industries, 2006.

[3] J. Kim, P. Moin, and R. Moser, “Turbulence statistics in fully developed

channel flow at low reynolds number,” J. Fluid Mech, 1987.

[4] R. D. Moser, J. Kim, and N. N. Mansour, “Direct numerical simulation

of turbulent channel flow up to re= 590,” Phys. Fluids, vol. 11, no. 4,
pp. 943–945, 1999.

[5] http://www.openfoam.com, October 2014.

[6] OpenFOAM Foundation, OpenFOAM - User Guide, 2.3.0 ed., 2014.

[7] D. Laurence, J. Uribe, and S. Utyuzhnikov, “A robust formulation of

the v2- f model,” Flow, Turbulence and Combustion, vol. 73, no. 3-4,
pp. 169–185, 2005.

26