AbstractThis paper presents a computational study of the

separated flow in a planer asymmetric diffuser. The steady RANS

equations for turbulent incompressible fluid flow and six turbulence
closures are used in the present study. The commercial software
code, FLUENT 6.3.26, was used for solving the set of governing
equations using various turbulence models. Five of the used
turbulence models are available directly in the code while the v2-f
turbulence model was implemented via User Defined Scalars (UDS)
and User Defined Functions (UDF). A series of computational
analysis is performed to assess the performance of turbulence models
at different grid density. The results show that the standard k-, SST
k- and v2-f models clearly performed better than other models
when an adverse pressure gradient was present. The RSM model
shows an acceptable agreement with the velocity and turbulent
kinetic energy profiles but it failed to predict the location of
separation and attachment points. The standard k- and the low-Re k-
delivered very poor results.

KeywordsTurbulence models, turbulent flow, wall functions,
separation, reattachment, diffuser.
I. INTRODUCTION
HE equations governing single-phase flows, the so-called
Navier-Stokes equations have been known for more than
a century. Despite intensive researches have been made; direct
analytical solutions to these equations are not available, except
for a very limited number of special cases. Since digital
computers became available in the 1960's, the analysis of
single-phase flows have increasingly been carried out
numerically using a range of techniques which together form
the field known as Computational Fluid Dynamics, CFD, [1].
In the present time, Computational Fluid Dynamics tools are
becoming standard in many fields of engineering involving
flow of gases and liquids; numerical simulations are used both
in the design phase to select between different concepts and in
the production phase to analyze performance.
Turbulence has a decisive influence on heat transfer,
species transport, drag, vorticity distribution, separation and
swirl flow. Separation and reattachment of turbulent shear
layers in the presence of adverse pressure gradient can be seen

Samy M. El-Behery is with Faculty of Engineering, Menoufiya University,
Shebin El-kom, Egypt (e-mail: [email protected]).
Mofreh H. Hamed is with Faculty of Engineering, Kafrelsheikh University,
Kafrelsheikh, Egypt (e-mail: [email protected]).

in many practical industrial and engineering applications,
either in internal flow systems such as diffusers, combustors
and channels with sudden expansion, or in external flows like
those past bluff structures and buildings. Turbulent flows are
characterized by fluctuating velocity fields. Therefore, one
way is to use a numerical mesh finer than the smallest length
scales and time step smaller than the fastest fluctuations of the
flow. This method is usually called Direct Numerical
Simulations (DNS). The exact governing equations can be
time-averaged, ensemble-averaged, or manipulated to remove
the small scales, resulting in equations that are
computationally less extensive to solve. However, the
modified equations contain additional unknown variables, and
thus the turbulence models are needed to determine these
variables in terms of known quantities. Generally, both the
Reynoldsaveraged Navier-Stokes approach (RANS) and the
Large Eddy Simulation (LES) approach can be employed to
transform the Navier-Stokes equations in such a way that the
small scale turbulent fluctuations do not have to be directly
simulated. However, here, only the Reynoldsaveraged
approach is considered for the simulation of flow through
asymmetric diffuser based on the following three factors: (1)
the mean flow in the diffuser is steady, (2) large computer
resources are required to resolve the energycontaining
turbulent eddies if LES approach is used, (3) the Reynolds-
averaged approach has been proven to be suitable for
industrial fluid simulations [1 and 2]. The RANS equations
need a turbulence model for computation of Reynolds stresses
that stems from averaging the non-linear convective terms, see
for more details [3]. A large family of turbulence models
exists in the literature which is far too extensive to be
reviewed here. The models are ranged from simple algebraic
expressions for the eddy viscosity to more elaborate
formulations which introduce a separate transport equation for
each component of the Reynolds stresses, see for more details
[4 and 5].
The most popular turbulence models are the standard k-
model [6], low-Re k- model [7], RNG k- model [8],
standard k- model [9], SST k- model [10]. These are based
on Boussinesq assumption that relates the apparent turbulent
shearing stresses to the rate of mean strain through an
apparent scalar turbulent or eddy viscosity. Consequently,
the relation between the Reynolds stresses and the velocity
A Comparative Study of Turbulence Models
Performance for Turbulent Flow in a Planar
Asymmetric Diffuser
Samy M. El-Behery, and Mofreh H. Hamed
T
World Academy of Science, Engineering and Technology 53 2009
769


gradient is linear. The v2-f turbulence model is a recently
developed four equation turbulence model [11]. This model is
similar to the standard k- model, and incorporates also some
near-wall turbulence anisotropy as well as non-local pressure-
strain effects. A more general model than those based on the
Boussinesq assumption is the Reynolds stress model [12]. In
this model, a separate transport equation for each component
of the Reynolds stresses is solved. Turbulence modeling is
commonly faulted as the case of deviations from measured
data in the predictions of flow through the tested case.
Therefore, turbulence modeling is a key issue in most CFD
simulations. Virtually all engineering applications are
turbulent and hence required turbulence model. Nevertheless,
no pretence has been made that any of these models can be
applied to all turbulent flows: such as universal model may
not exist. Each Model has its advantages, disadvantages,
limitations and appropriate flow regimes. Industry has many
pressing flow problems to solve that will not wait for the
conception of a universal turbulence model. Fortunately, many
sectors of industry are specifically interested in a limited class
of flows only; e.g. pipe flows for the oil transportation sectors
and diffusers for jet engines and compressors. The large
majority of turbulence research consists of case-by-case
examination and validation of existing turbulence models for
such specific problems. Performance of a propulsion system
as a whole is dependent on the efficiency of diffusers.
Therefore, identification of separation within diffusers is
important because separation increases drag and causes inflow
distortion to engine fans and compressors. Diffuser flow
computations are a particularly challenging task for
Computation Fluid Dynamics (CFD) simulations due to
adverse pressure gradients created by the decelerating flow,
frequently resulting in separation. These separations are
highly dependent on local turbulence level, viscous wall
effects, and diffuser pressure ratio, which are functions of the
velocity gradients and the physical geometry. Thus, turbulence
modeling and geometry modeling become dominant factors
that affect the ability of CFD to accurately predict flow
through diffusers.
Therefore, the goal of this study is to examine six
turbulence models in terms of their accuracy, convergence and
computational cost. In addition, comparisons are made
between FLUENT code and WIND code predictions for some
of these turbulence models.
II. MATHEMATICAL MODEL AND COMPUTATIONAL METHOD
The commercial FLUENT software package, FLUENT
6.3.26, was used for solving the set of governing equations.
The numerical method employed is based on the finite volume
approach. Fluent provides flexibility in choosing
discretization schemes for each governing equation. The
discretized equations, along with the initial condition and
boundary conditions, were solved using the segregated
solution method. Using the segregated solver, the
conservation of mass and momentum were solved sequentially
and a pressure-correction equation was used to ensure the
conservation of momentum and the conservation of mass
(continuity equation). Several turbulence models, such as, the
standard k model, the low-Re k model, the standard k
model, the shear-stress transport k model, the Reynolds
stress model (RSM) and the v2-f model. The first five models
are available directly in FLUENT while the last one (v2-f
model) was implemented using user defined functions (UDF)
and user defined scalars (UDS).
A. Governing Equations
In the present study steady RANS equations for turbulent
incompressible fluid flow with constant properties are used.
The governing flow field equations are the continuity and the
Reynolds averaged Navier-Stokes equations, which are given
by:

0.0
i
j
u
x
c
=
c
(1)
( )
1
i j
ij i j
j i j
u u
p
S u u
x x x
v

c
c c
' ' = +
c c c
(2)
Where, S
ij
is the main strain rate and calculated by:
1
2
j
i
ij
j i
u
u
S
x x
| | c
c
= + |
|
c c
\ .
(3)
and
i j ij
u u t ' ' = is the unknown turbulent or Reynolds-stress
tensor and
i
u' represents the velocity fluctuation in i-
direction. These equations are not a closed set and turbulence
models are required to model the turbulent or Reynolds-stress
tensor.
B. Turbulence Modeling
Several turbulence models available are employed to
predict the flow behavior in a planer asymmetric diffuser.
Most of these models are derived from standard k model
and vary in complexity and robustness from two equation
turbulence models to more elaborated turbulence model. Five
of the used turbulence models are based on the Boussinesq
assumption. In which the Reynolds stress tensor is computed
from the effective viscosity formulation, which is a direct
extension of the laminar deformation law. It is given by:
2
2
3
ij ij t ij
k S t o v = (4)
Where,
i j
k u u ' ' = is the turbulent kinetic energy,
ij
is the
Kronecker delta and
t
v denotes turbulent kinematic viscosity.
In order to obtain the turbulent viscosity, other transport
equations are needed. Theses equations differ from model to
another.
1. The Standard k model (SKE)
The k model is well described in the literature and has
been widely used. This model was derived by assuming that
the flow is fully turbulent and the effects of molecular
viscosity are negligible [6]. For locations near walls, the
World Academy of Science, Engineering and Technology 53 2009
770


standard k- model, therefore, demands an additional model,
which comprises the effects of molecular viscosity. In this
situation, wall functions based on semi-empirical formulas
and functions are employed.
2. Low Reynolds Number k Model (LRNKE)
The Low-Reynolds-number k model of Launder and
Sharma [7] is similar to the standard k model except that is
uses damping function instead of the wall-function and
contains extra source terms in its turbulent kinetic energy and
dissipation rate equations. Patel et al. [13] reviewed several
low-Re k- model and they found that the model of Launder
and Sharma performs better than the others. Furthermore, the
model of Launder and Sharma uses the turbulent Reynolds
number in the damping function instead of the dimensionless
wall distance, Y
+
. This makes the model suitable for
simulating flow when separation is expected. While, the
models which use Y
+
are not the right choice for separated
flow problems (since Y
+
vanishes at separation and
reattachment locations driving f

to zero, creating laminar

spots in the middle of a turbulent flow). The turbulent kinetic
energy, k, equation for the standard and the low-Re k- models
reads
( )
eff
j
j j k j
k
u k G D
x x x

c
o

c c c
= +
`
c c c

)
(5)
Where, the turbulent production rate is
j
i i
eff
j i j
u
u u
G
x x x

| | c
c c
= + |
|
c c c
\ .
(6)
The dissipation rate, , equation for the standard and the
low-Re k- models reads

1 1 2 2
( )
( )
eff
j
j j j
u
x x x
C f G C f E
k
c
c c
c
c
o
c
c

c c c
=
`
c c c

)
+ +
(7)
The models constants C
1
and C
2
, the damping functions f
1
,
f
2
and f

, and the extra source terms D and E for the low-Re k

model can be found in [6, 7].
3. The Standard k- model (SKW)
The standard k- model is one of the most common
turbulence models. It includes two extra transport equations to
represent the turbulent properties of the flow. The first
transported variable is turbulent kinetic energy, k, similar to
the turbulent kinetic energy equation of the standard k-
model. The second is the specific dissipation, , which can
also be thought of as the ratio of to k [9]. The model
incorporates modifications for low-Re effects, compressibility
and shear flow spreading. Detailed derivations for the closure
equations are provided by Wilcox [9].
4. The Shear Stress Transport k- (SST)
The SST k- model was developed by Menter [10], which
combined the robustness of k- turbulence model near walls
with capabilities of the k- model away from the walls. The
definition of the turbulent viscosity is modified to account for
the transport of turbulent shear stress. The model equations
are provided in [10].
5. The Reynolds Stress Model (RSM)
The Reynolds stress model (RSM) is a higher level,
elaborate turbulence model. It is usually called a Second
Order Closure. This modeling approach originates from [12].
In RSM, the eddy viscosity approach has been discarded and
the Reynolds stresses are directly computed. The exact
Reynolds stress transport equation accounts for the directional
effects of the Reynolds stress fields. Detailed derivations for
the closure equations are provided in [14].
6. v
2
-f Turbulence Model (V2F)
This model is a simplification of the elliptic relaxation
Reynolds stress model developed by Durbin [11], which
requires the solution of three transport and one elliptic
(relaxation) equations. The system of Reynolds stress
equations is replaced by a transport equation for a velocity
scalar (
2
v ' ) and an elliptic equation for (f). The model was
reformulated to avoid the numerical oscillations of wall
boundary for f, as given in [15]. The equations for turbulent
kinetic energy and the dissipation rate are the same as those of
the standard k- model, while the equations for
2
v'
and f can
be written as given in [15] as follows
( )
2 2
2
6
j t
j j j
v v
u kf v
x k x x
c
v v
(
' ' c c c
' = + + (
c c c
(

(8)
2 2
2
1 1 2 2
1 2
( 6) ( 1)
3
j
f v G
L f C C C
T k k x
(
' c
= (
c
(

(9)
max , 6
k
T
v
c c
(
=
(

(10)
1/ 4
3/ 2 3
max ,
L
k
L C C
q
v
c c
(
| |
( =
|
( \ .

(11)
The model constants are taken as given in [15]
C. Wall Functions
In the region near the wall, the gradient of quantities is
considerably high and requires fine grids close to the wall to
capture the changes of quantities. For complex flows where
separation flow and reattachment occur, the conventional
logarithmic wall-function proposed by Launder and Spalding
[6] becomes less reliable. The non-equilibrium wall-function
proposed by Kim and Choudhury is proven to give better
predictions due to the fact that it accounts for the effects of
pressure gradient and departure from equilibrium [17]. The
standard k- model and the RSM model employ the non-
equilibrium wall-function is applied to the wall-adjacent cells,
while the low-Re k- model uses damping functions instead of
the wall-function. The v
2
-f model treats the near-wall
turbulence without the use of exponential damping or wall
functions. For the standard and SST k- models, if the
World Academy of Science, Engineering and Technology 53 2009
771


transitional flows option is enabled in the viscous model
panel, low-Reynolds-number variants will be used, and, in
that case the near-wall grids have to be very fine to obtain
better predictions for the near wall modeling. If transitional
flows option is not active as in the present study, the near wall
grids follow a rule of the wall function [14].
The use of a wall function in a computational flow solver
allows fewer points to be placed near the walls where as
points are typically placed to Y
+
= 1 for a wall integrated grid.
In the present study three wall functions with initial grid point
spacing ranges from Y
+
= 1 to 30 are used near the wall in a
wall function grid. These values of Y
+
were chosen to asses
FLUENTs wall function capabilities across several positions
in the boundary layer logarithmic region. The effect of the
initial grid point spacing is investigated in this paper.
D. Diffuser Geometry and Computational Grids
The test case analyzed in this study is a two-dimensional
turbulent flow in an asymmetric planar diffuser. Due to the
adverse pressure gradient the flow is separated and a large
recirculation bubble is generated. This problem has been
selected because a very reliable experimental database is
available [17 and 18]. Moreover, a detailed Large Eddy
Simulation study is also available for comparison [19].
The diffuser geometry is presented in Fig. 1 and the
computational domain is shown in Fig.2. The tested diffuser
can be divided into three sections: an inflow channel, the
asymmetric diffuser, and an outflow channel. The upstream
channel was made sufficiently long to obtain fully developed
turbulent channel flow at the inlet of the diffuser section. The
Reynolds number based on the bulk velocity and the upstream
channel height, H, is 18000 matches the experimental
configuration of Obi et al. [17] and Buice & Eaton [18].
In order to assess the grid sensitivity of the present results,
Simulations were performed on three different meshes. The
mesh is stretched in the streamwise and wall-normal directions
and designed such that the streamwise spacing gradually
decreases towards the diffuser. The refined grid was obtained
by approximately doubling the number of points in y-
direction. A detailed view of the coarse and fine grids at the
diffuser inlet section is shown in Fig. 3.
E. Boundary Conditions
There are three faces bounding the calculation domain
namely: the inlet boundary, the wall boundary and the outlet
boundary. No-slip boundary conditions are applied along the
solid walls and wall functions were used as described earlier.
At the outlet, 60H downstream the diffuser exit, the boundary
was adjusted as a pressure outlet boundary condition. At the
inlet, 74H upstream the diffuser entrance, flat velocity and
turbulent quantities profiles are specified.
F. Solution Strategy and Convergence
A second-order upwind discretization scheme was used for
the momentum equation while a first-order upwind
discretization was used for turbulent quantities. These
schemes ensured, in general, satisfactory accuracy, stability
and convergence. SIMPLE algorithm described by Patankar
[20] was used for pressure-velocity coupling. The discretized
equations are solved implicitly in sequence, starting with the
pressure equation followed by the momentum equations, by
the pressure correction equation, and finally by the equations
for the scalars (turbulence variables). Within this loop, the
linearized equations for each variable are integrated using a
linear system solver. FLUENT code allows implementing
customized models through user defined functions, UDF,
which is treated for v
2
-f turbulence model. Four user defined
scalars, UDS, are used for this purpose. The wall-boundary
condition of the dissipation rate equation depends on the
turbulent kinetic energy, k, near the wall; therefore, it is
necessary to initialize the solution before hooking the
boundary condition.
The convergence criterion consisted of monitoring skin
friction values and variation of velocity profiles with iteration,
reduction of several orders of magnitude in the residual errors.


Fig. 1 Schematic of Buice-Eaton diffuser, Ref. [18]
III. RESULTS AND DISCUSSION
Steady flow in a planer asymmetric diffuser shown in Fig. 1
is investigated. Separated flow simulations in the tested
diffuser were carried out using several turbulence models and
three different meshes. The measured skin friction coefficient,
C
f
, in the fully developed entrance region is 0.0061. This
value of C
f
is used to calculate the initial wall spacing, y
p
, as
follows:
v
t
u y
Y
p
=
+

Where,
2
1
/ ;
2
w w f b
u C U
t
t t = =
The variation of the dimensionless wall distance, Y
+
, for
the lower wall adjacent cell is presented in Fig. 4. It can be
seen from this figure that Y
+
drops in the tail duct to
approximately one-third its value in the upstream entrance
region. In addition,
World Academy of Science, Engineering and Technology 53 2009
772


-60 -40 -20 0 20 40 60 80
0
2
4
y/H
x/H


Fig. 2 Computational domain


a- fine grid (Y
+
1.0) 42179

b- coarse grid (Y
+
15) 42141

c- coarse grid (Y
+
30) 42141
Fig. 3 Computational grid at diffuser inlet section


it was difficult to obtain a converged solution using LRNKE
with the coarse girds. Therefore, the results presented here
study for LRNKE was obtained with fine grid only
(i.e., Y
+
= 1).
Figs. 5-6 show Comparisons between present predictions of
pressure recovery coefficient using all the tested turbulence
models and the experimental data of [17 and 18], and LES,
results of [19]. It can be seen from the figures that most of the
pressure increase occurs within the first third of the diffuser
with the steepest rise at about x/H = 1.5. The figures also
indicate that, in the case of fine grid, the pressure recovery
coefficient obtained with V2F and SSTKW models agree
reasonably well with experimental data of [17] and LES
results of [19], while, LES predictions are higher than
experimental data of [18] by about 5%. Also the SKE and
LRNKE models predictions are closed to each other and are
higher than the experimental data of [17 and 18] and LES
results by about 20% in the separation region. The RSM
predictions are close to experimental data of [18] rather than
that of [17] and LES results, while the SKW model slightly
over-predicts the pressure recovery coefficient when the fine
grid is used. In the case of coarse grid all models predictions
expect the SKE model are close to each other. The later still
over-predict the pressure recovery coefficient, as shown in
Fig. 6.
For incompressible, inviscid flow the total pressure
recovery coefficient, C
po
= C
p
+ (U/U
b
)
2
, is constant along a
stream tube. A similar relation holds for viscous flow,
declaring that the total pressure decreases in the flow direction
due to frictional losses [19]. Fig. 7 shows comparisons
between predicted total pressure recovery coefficient and
normalized maximum velocity, U
max
/U
b
, and published
experimental data and LES results of [19]. It can be seen from
the figure that the SKE and LRNKE models under-predicts
the maximum velocity in the rear part of the diffuser and in
the downstream duct, while the RSM over-predicts the
maximum velocity up to x/H = 13 and slightly under-predicts
it after x/H = 15. The higher pressure coefficient and the lower
maximum velocity obtained by the SKE and LRNKE models
results in close agreement in the total pressure recovery
coefficient, as shown in Fig. 7. The figure indicates also that
the diffuser causes a decrease in the total pressure by about
35% due to viscous losses.
In order to provide a direct comparison with experimental
data, results are presented by showing axial velocity, turbulent
kinetic energy and Reynolds stresses profiles at several axial
stations. Axial and vertical positions are non-dimensionalized
World Academy of Science, Engineering and Technology 53 2009
773


-20 0 20 40 60
-0.3
0.0
0.3
0.6
0.9
C
p
x/H
V2F
SKE
SKW
RSM
LRNKE
SST
, Experimental
LES, Ref. [ ]
-20 0 20 40 60
0.0
0.5
1.0
1.5
2.0
Y
+
x/H
V2F
SKE
SKW
RSM
LRNKE
SST

a- fine grid (Y
+
1.0) 42179.

-20 0 20 40 60
0
5
10
15
20
Y
+
x/H

b- coarse grid (Y
+
15) 42141

-20 0 20 40 60
0
10
20
30
40
Y
+
x/H

c- coarse grid (Y
+
30) 42141

Fig. 4 Variations of Y
+
values for the tested turbulence models
by the upstream channel height, H, while the velocity profiles
are normalized by the bulk velocity, U
b
. The turbulent kinetic
energy and Reynolds stresses are normalized by
2
b
U . Fig. 8
shows a comparison between present predicted and measured
velocity profiles at various axial locations along the diffuser
for several turbulence models. The figure indicates that the
SST, SKW and V2F models predict the axial velocity profiles
very well, where. The RSM failed to predict the variation of
axial velocity near the lower wall (i.e., in the separation zone).
Similar observation was reported by Iaccarino [21]. This may
be attributed to the use of the law of the wall for calculating
the turbulence. On the other hand, the predicted axial velocity
profiles using the LRNKE and the SKE models are in poor
agreement compared with excremental data of [19]. This may
be due to the inaccuracy of the isotropic turbulence models in
predicting anisotropic turbulent flows, while the SST, SKW
and V2F models contains some near-wall turbulence
anisotropy.


a- upper wall
-20 0 20 40 60
-0.3
0.0
0.3
0.6
0.9
C
p
x/H

b- lower wall

Fig. 5 Comparison between present predictions, LES results of [19]
and experimental results of [17, 18] in terms of pressure recovery, Cp
for Y
+
=1.0 (closed symbols Obi et al. data [17], open symbols
Buice-Eaton data [18])
[19]
World Academy of Science, Engineering and Technology 53 2009
774



a- upper wall
-20 0 20 40 60
-0.3
0.0
0.3
0.6
0.9
C
p
x/H

b- lower wall

Fig. 6 Comparison between present predictions, LES results of [19]
and experimental results of [17, 18] in terms of pressure recovery, Cp
for Y
+
=15 (closed symbols Obi et al. data [17], open symbols Buice-
Eaton data [18])

The turbulent kinetic energy profiles are presented in Fig. 9.
It can be seen from this figure that, at x/H = 5.2 all models
except the V2F model over-predict the turbulent kinetic
energy. Downstream at x/H = 11.2, 15.2, the SKW and V2F
predictions are closed to each other and in a good agreement
with the experimental data, while the SST and RSM over-
predict the turbulent kinetic energy. Further downstream
x/H=15.2, only the V2F model gives the best agreement
between predicted and measured turbulent kinetic energy. It
can be seen also from this figure that the LRNKE and SKE
models completely fail to capture the asymmetric development
of turbulent kinetic energy and underestimate its magnitude in
the diffuser. The development of Reynolds stresses profiles is
presented in Fig.10. Since the Reynolds stresses are available
only for RSM in FLUENT, therefore, a user defined
subroutine is written to calculate and store the Reynolds
stresses for other models in three user defined memories,
UDM. The figure shows that all models under-predict u u ' '
and over-predict v v ' ' . Despite the SKW and V2F models
predict the turbulent kinetic energy better than the RSM, the
latter predicts the Reynolds stresses better than other models.
This can be attributed to the solution of separate transport
equation for each component of the Reynolds stresses in the
RSM model while they are calculated from the Boussinesq
assumption in the other models.


Fig. 7 Comparison of present predictions of pressure recovery C
p
and
U
max
/U
b
with experimental and LES results.
(Caption as in Fig. 6)

The skin friction coefficients on both the upper and lower
wall for different turbulence models used in the case of fine
and coarse grids are presented in Figs. 11-12. Firstly, from the
figure it is seen that there is a notable difference between the
prediction of friction coefficient using different turbulence
models on both the upper and lower wall of the diffuser. The
SKE and RSM models fail to predict the boundary layer
separation and consequently the friction coefficient. This may
be due to the use of the low of the wall in both simulations.
However, the use of damping functions instead of the law of
the wall in LRNKE model does not introduce any significant
improvement. The V2F predicts a bubble in very close
agreement with the experimental data of [18] and LES results
of [19] at Y
+
= 1. The SKW and SST models predictions are in
acceptable agreement. The figure also indicates that increasing
the dimensionless wall distance, Y
+
to 15 enhances the
predictions of the SKE and RSM models. However, further
increase in Y
+
does not improve the predictions. The best
prediction is obtained at Y
+
= 1 for the V2F model and at
Y
+
= 15 for the SKW and SST models. At Y
+
= 15, the V2F
under-predicts the skin friction coefficient in the upstream
duct and along the upper wall of the diffuser. Fig. 13 shows a
comparison between the present predicted skin friction
coefficient along the upper and lower walls of the diffuser and
that obtained by WIND code, Ref. [22]. The comparison
shows a good agreement when SST model is used while the
predictions from both codes are different when the SKE is
used. The predictions obtained by the WIND code show a
separation bubble at the upper of the diffuser wall which it
was not observed experimentally. The very good agreement
obtained when the SST model is used suggests that the
-20 0 20 40 60
-0.3
0.0
0.3
0.6
0.9
C
p
x/H
V2F
SKE
SKW
RSM
SST
, Experimental
LES, Ref. [ ]
[19]
2
max
|
|
.
|

\
|
b
U
U
0 5 10 15 20 25 30
0.0
0.3
0.6
0.9
1.2
1.5
x/H
2
max
|
|
.
|

\
|
+
b
p
U
U
C
World Academy of Science, Engineering and Technology 53 2009
775


difference is not related to the numerical techniques used to
discretize the equations but to the implementation of the SKE
model. The evaluation of the wall normal distance, which is
required in the law of the wall, may introduce such
discrepancy.
The most common definition of the separation point is the
location where the wall shear stress is zero. Table I shows a
comparison between the experimental data of [18] and the
present calculated separation point as well as the reattachment
point using different turbulence models at different values of
Y
+
and different grids based on C
f
profiles. It can be seen
from this figure that V2F model shows a good agreement
compared with the other models. The SKW and SST models
predicted the separation point earlier than that predicted by the
V2F model and than the measured one. As Y
+
increases the
RSM predicts a small separation region while the SKE and
LRNKE models do not predict any separation. There are two
other definitions found in the literature for the separation
bubble. The first one is; the separation bubble is the mean
recirculating region within the dividing streamline (also called
separation streamline) reaching between the stagnation points
on the wall at the separation and reattachment points. The
second definition is; the separation bubble is the region with
mean backflow (i.e. region below the curve of zero mean
velocity) [23]. Fig. 14 presents comparisons between the
predicted position where the streamwise velocity crosses zero
and the experimental data of [17 and 18] and LES results of
[19], while the predicted dividing streamline is compared with
LES results of [19] in Fig. 15. It can be seen from these
figures that the separation bubble obtained from the
experimental data is slightly larger than that obtained from the
present predictions LES results. The present predictions using
SKW and V2F models are in a close agreement with LES
results when the finer grid is used, while a small discrepancy
is introduced as Y
+
increase. The SST model gives the larger
separation bubble compared with other models and LES
results, while the RSM develops inaccurate results.



-5 0 5 10 15 20 25 30 35
0
2
4
y/H
x/H+10u/U
b
Experimental
V2F
SST
SKW
RSM
SKE
LRNKE

Fig. 8 Development of axial velocity profile through the diffuser for the tested turbulence models compared with and experimental results of
Buice-Eaton data [18]


0 5 10 15 20 25 30
0
2
4
y/H
x/H+500 k/U
c
Experimental
V2F
SST
SKW
RSM
SKE
LRNKE
2
Fig. 9 Development of turbulent kinetic energy profiles through the diffuser for the tested turbulence models compared and experimental
results of Obi et al. data [17]

World Academy of Science, Engineering and Technology 53 2009
776


-5 0 5 10 15 20 25 30 35
0
2
4
y/H
x/H+300 u`u`/U
b
Experimental
V2F
SST
SKW
RSM
SKE
LRNKE
2 -

a- u u ' '
-5 0 5 10 15 20 25 30 35
0
2
4
y/H
x/H+300v`v`/U
b
Experimental
V2F
SST
SKW
RSM
SKE
LRNKE
2 -

b- v v ' '

-5 0 5 10 15 20 25 30 35
0
2
4
y/H
x/H+300u`v`/U
b
Experimental
V2F
SST
SKW
RSM
SKE
LRNKE
2 -

c- u v ' '

Fig. 10 Development of Reynolds stresses profile through the diffuser for the tested turbulence models compared with experimental of Buice-
Eaton data [18]

The computational effort and cost in terms of CPU time and
number of iterations is shown in Table II. Generally, four
factors influence the computing time namely, grid resolution,
discretization scheme, degree of nonlinearity of the model,
and number of PDEs the model contains. When fixing the first
two factors, the difference in computing time is mainly
attributed to the turbulence model itself. If the SKE model is
taken as the baseline, then using the LRNKE and SKW
models requires slightly more computation time and number
of iteration due to the extra terms and functions in the
governing equations. Since the functions associated with the
SST model are extra than that with SKW, it requires about 26
% grater time and about 13 % greater number of iteration than
the SKE. Unlike the two-equation models, the V2F and RSM
models require the largest time and number of iterations due
to the extra transport equations (the number of differential
equations to be solved is the same for two-dimensional
problems). Despite of the comparable CPU time per iteration,
World Academy of Science, Engineering and Technology 53 2009
777


the RSM requires about 77% grater time than that of V2F due
to the strong coupling between equations and the high degree
of non linearity when the RSM is used.

-20 0 20 40 60 80
0.000
0.003
0.006
0.009
C
f
x/H
Experimental
SKE
RSM
SKW
V2F
SST
LES, Ref [ ]

a - upper wall

b - lower wall

Fig. 11 Comparison between present predictions, LES results [19]
and experimental results of [17, 18] in terms of skin friction
coefficient Cp for Y
+
= 1.0 (closed symbols Obi et al. data [17], open
symbols Buice-Eaton data [18])
IV. CONCLUSION
The turbulent flow through a planer asymmetric diffuser
was investigated numerically using the commercial CFD code
FLUENT 6.3.26. The performance of six different turbulence
models is compared with published experimental and LES
results. The standard k-, low-Re k-, standard k-, SST k-
and RSM models are available as standard features in the
code, while the v2-f model was implemented through the User
Defined Functions in the code. The simulations was carried
out on three grids having different spacing for the near wall
points and different resolutions. The comparisons showed that
V2F turbulence model indicates the best agreement with
experimental data followed by the SKW and SST turbulence
models. The SKE and LRNKE turbulence model give very
poor results. Also, the RSM model gives unexpected poor
results compared with those obtained by V2F, SKW and SST
model. In addition, the computational time and number of
iterations required by each model are compared. The
comparison showed that the RSM requires the greatest
number of iterations and hence the largest computational time.
-20 0 20 40 60 80
0.000
0.003
0.006
0.009
C
f
x/H
Experimental
SKE
RSM
SKW
V2F
SST
LES, Ref [19]

a - upper wall

b - lower wall

Fig. 12 Comparison between present predictions, LES results [19]
and experimental results of [17, 18] in terms of skin friction
coefficient Cp for Y
+
= 15 (closed symbols Obi et al. data [17], open
symbols Buice-Eaton data [18])
REFERENCES
[1] Wikipedia, "Computational fluid dynamics [online]", Available at
http://en.wikipedia.org/wiki /Computational_ fluid_dynamics, [accessed
Dec. 20, 2008].
[2] Bradshaw P., "Understanding and predictions of turbulent flow 1996",
Int. J. Heat and Fluid Flow, Vol. 18, 1997, pp. 45-54.
[3] Anderson D. A. and Tannehill, J. C. and Pletcher, R. H., "Computational
Fluid Mechanics and Heat Transfer", Heimsphere Publishing
Corporation, 1984.
[4] http://www.cfd-online.com/Wiki/ Turbulence _ modeling [accessed Dec.
20, 2008].
[5] Versteeg H. K. and Malalasekera W., "An introduction to computational
fluid dynamics, the finite volume method", Longman group Ltd, 1998.

0 20 40
0.0
-20 0 20 40 60 80
0.000
0.003
0.006
0.009
C
f
x/H
0 20 40
0.0
-20 0 20 40 60 80
0.000
0.003
0.006
0.009
0.012
C
f
x/H
World Academy of Science, Engineering and Technology 53 2009
778


TABLE I
COMPARISONS BETWEEN PRESENT PREDICTED SEPARATION AND REATTACHMENT POINTS USING DIFFERENT MODELS AND EXPERIMENTAL DATA
Grid Exp. SKW SST V2F SKE LRNKE RSM
Separation (x/H) 7.4 2.91 2.05 6.34 N/A N/A N/A Y
+
1.0
(42179) Reattachment (x/H) 29.2 28.95 30.15 29.25 N/A N/A N/A
Separation (x/H) 7.4 3.51 3.38 5.32 N/A 17.04 Y
+
15.0
(42141) Reattachment (x/H) 29.2 30.05 29.71 30.42 N/A 20.13
Separation (x/H) 7.4 3.62 2.94 5.44 N/A 18.97 Y
+
30.0
(42141) Reattachment (x/H) 29.2 28.92 29.53 30.72 N/A 20.25



-20 0 20 40 60 80
-0.002
0.000
0.002
0.004
0.006
0.008
C
f
x/H
Experimental
Lower wall Upper wall
FLUENT
WIND

a - SST Model
-20 0 20 40 60 80
-0.002
0.000
0.002
0.004
0.006
0.008
0.010
0.012
0.014
C
f
x/H
Experimental
Lower wall Upper wall
FLUENT
WIND

b - SKE model
Fig. 13 Comparison between present FLUENT predictions and
predictions by WIND code, Ref. [22], using SST and SKE
models and Y
+
1.0

[6] Launder B. E. and Spalding D. P., "The numerical computation of
turbulent flows", Computer Methods in Applied Mechanics and
Engineering, Vol. 3, 1974, pp. 269-289.
[7] Launder B. E. and Sharma B. I., "Application of the energy-dissipation
model of turbulence to the calculation of flow near a spinning disc",
Letters in Heat and Mass Transfer, Vol. 1, 1974, pp. 131-138.
[8] Yakhot, V. and Smith, L. M., "The Renormalization Group, the Epsilon-
Expansion and Derivation of Turbulence Models", J. Scientific
Computing, Vol. 7, 1992, pp. 35-51.
[9] Wilcox D. C., "Turbulence Modeling for CFD", DCW Industries, La
Canada, California, 1998.
[10] Menter F. R., "Two-Equation Eddy-Viscosity Turbulence Models for
Engineering Applications", AIAA Journal, Vol. 32, 1994, pp. 1598-
1605.
[11] Durbin P. A., "Near-wall turbulence closure modeling without damping
functions", Theoretical and Computational. Fluid Dynamics, Vol. 3,
1991, pp. 1-13.
[12] Launder B. E. and Spalding D. B., "Mathematical Models of
Turbulence", Lectures Notes, Imperial College of Science and
Technology, London, England, 1972.
[13] Patel V. C., Rodi W. and Scheuerer G., "Turbulence models for Near-
wall and low Reynolds number flows: a review", AIAA J., Vol. 23,
1985, pp. 1308-1318.
[14] Fluent, "Users Guide Fluent 6.3.26", Fluent Incorporated, Lebanon,
NH, 2006.
[15] Lien F-S and Kalitzin G, "Computations of transonic flow with the v2-f
turbulence model", Int. J. Heat Fluid Flow, Vol. 22, 2001, pp. 5361.
[16] Kim J-Y, Ghajar A. J., Clementang and Foutch G. L., "Comparison of
near-wall treatment methods for high Reynolds number backward-facing
step flow", Int. J. Computational fluid dynamics, Vol. 19, 2005, pp. 493-
500.
[17] Obi, S., Aoki, K., and Masuda, S., Experimental and Computational
Study of Turbulent Separating Flow in an Asymmetric Plane Diffuser,
Ninth Symposium on Turbulent Shear Flows, Kyoto, Japan, August 16
19, pp. 3051 to 3054, 1993.
[18] Buice, C.U. & Eaton, J.K., Experimental Investigation of Flow
Through an Asymmetric Plane Diffuser, Journal of Fluids Engineering,
vol. 122, pp. 433435, 2000.
[19] Kaltenback H. J., Fatica M., Mittal R., Lund T. S., and Moin P., "Study
of the Flow in a Planar Asymmetric Diffuser Using Large Eddy
Simulations", J. Fluid Mech., Vol. 390, 1999, pp. 151185.
[20] Patankar S. V., "Numerical heat transfer and fluid flow", McGraw-Hill,
New York, USA, 1983.
[21] Iaccarino G., "Predictions of a turbulent separated flow using
commercial CFD codes", Trans. ASME, J. Fluids Engineering, Vol. 123,
2001, pp. 819-828.
[22] DalBello T., "Computational study of separating flow in a planar
subsonic diffuser", NASA TM 2005-213894, 2005.
[23] Trnblom O., "Experimental and computational studies of turbulent
separating internal flows", Doctoral thesis, KTH Mechanics, Stockholm,
Sweden, 2006.

0 10 20 30
0
2
4
SST
SKW
V2F
RSM
LES, Ref. [19]
, Experimental
y/H
x/H

a - Y
+
=1.0
World Academy of Science, Engineering and Technology 53 2009
779


0 10 20 30
0
2
4
y/H
x/H

b- Y
+
=30.0
Fig. 14 Position where streamwise velocity crosses zero compared
with LES results [19] and experimental results
of [17, 18] (closed symbols Obi et al. data [17],
open symbols Buice-Eaton data [18])

0 10 20 30
0
2
4
y/H
x/H

a- Y
+
=1.0
0 10 20 30
0
2
4
RSM
SST
SKW
V2F
y/H
x/H
LES, Ref. [19]


b- Y
+
=30.0

Fig. 15 Dividing (separating) streamline compared with LES
results [19].


TABLE II
NUMBER OF ITERATIONS AND CPU TIME NORMALIZED BY THOSE OF STANDARD K- MODEL



SKE LRNKE SKW SST V2F RSM
No. of iterations 1 1.03 1.07 1.26 1.2 2.06
CPU time 1 1.09 1.13 1.42 1.75 3.11
CPU time/iteration 1 1.06 1.06 1.13 1.46 1.51
World Academy of Science, Engineering and Technology 53 2009
780