Journal of Applied Fluid Mechanics, Vol. 10, No. 4, pp. 1029-1043, 2017.

Available online at www.jafmonline.net, ISSN 1735-3572, EISSN 1735-3645.

DOI: 10.18869/acadpub.jafm.73.241.26311

Computation of Developing Turbulent Flow through a

Straight Asymmetric Diffuser with Moderate Adverse
Pressure Gradient
S. Salehi1†, M. Raisee1 and M. J. Cervantes2,3
1 Hydraulic Machinery Research Institute, School of Mechanical Engineering, College of Engineering,
University of Tehran, Tehran, 11155/4563, Iran
2 Division of Fluid and Experimental Mechanics, Luleå University of Technology, Luleå, 97187, Sweden
3 Water Power Laboratory, Norwegian University of Science and Technology, Trondheim, 7491, Norway

†Corresponding Author Email: s.salehi@ut.ac.ir

(Received March 15, 2016; accepted March 6, 2017)

ABSTRACT

In this paper, numerical investigation of three-dimensional, developing turbulent flow, subjected to a

moderate adverse pressure gradient, has been investigated using various turbulence models, namely: the low-
Re k   , the SST k   , the v 2  f and a variant of Reynolds stress model. The results are compared with
the detailed velocity and pressure measurements. Since the inlet condition is uncertain, a study was first
performed to investigate the sensitivity of the results to the inlet boundary condition. The results showed the
importance of including the contraction effects. It is seen that the developing flow inside the straight duct, is
highly sensitive to the inlet boundary condition. The comparisons indicate that all turbulence models are able
to predict a correct trend for the centerline velocity and pressure recovery inside the straight duct and diffuser
but the low-Re k   and RSM turbulence models yield more realistic results. The SST k   model largely
overpredicts the centerline velocity and boundary layer thickness in the straight duct. The comparisons of the
numerical results also revealed that the RSM model, due to its anisotropic formulation, is able to reproduce
the secondary flows. As expected, the RSM model demonstrates the best performance in prediction of the
flow field and pressure recovery in the asymmetric diffuser.

Keywords: Moderate adverse pressure gradient; Asymmetric diffuser; Turbulent developing flow;
Computational fluid dynamics; RANS models.

1. INTRODUCTION separation could also have destructive effects on

performance of hydraulic machinery equipment
Developing turbulent flow under adverse pressure such as draft tube where the pressure recovery is
gradient (APG) occurs in various fluid flow related restricted by attachment of flow to the walls.
engineering applications. Turbulent flow around Turbulent flow under APG has been the subject of a
aircrafts, automobiles and inside the draft tube of large number of experimental and numerical
hydraulic turbines, are a few examples of such investigations. Among all the previous studies, the
applications. Hence, accurate prediction of experimental investigations conducted by Clauser
turbulent flow under adverse pressure gradient can (1954) and Bradshaw (1967) can be considered as
contribute in the optimization of fluid machineries. the first studies on this topic. Both studies measured
Indeed, the presence of APG can greatly affect the characteristics of equilibrium turbulent boundary
performance and efficiency of hydraulic machines. layers where the pressure gradient parameter is kept
If a turbulent boundary layer flow encounters a constant. Nagano et al. (1993) experimentally
large APG, flow becomes unstable and if the APG studied the behavior of turbulent boundary layer
is sufficiently large, flow separates from the wall under moderate and strong adverse pressure
and forms a recirculating region. Such recirculation gradients. They showed that with increase in
region often has negative consequences such as: adverse pressure gradients, the near wall velocity
increase in the pressure drag, decrease in lift and profile shifts below the standard log-law, indicating
increase in head losses. More specifically, flow a reduction in the viscous sub-layer thickness.
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

Furthermore, the presence of an APG could computations of Obi diffuser flow. They used three
significantly affect the distributions of turbulence different mesh sizes and different modeling
quantities. Rai (1986) conducted experiments in a approaches to examine the influence of the mesh
wind tunnel for three arbitrary adverse pressure resolution and four subgrid models, namely: no
gradient flows. The experimental observations model (implicit LES), the standard Smagorinsky
clearly indicate that wall-wakes in adverse pressure model, the dynamic Smagorinsky model and the
gradient can be adequately described by the two- dynamic localization model. The mesh refinement
layer model proposed. study demonstrated improvement in the predictions.
Among the subgrid strategies examined, the
Direct numerical simulation (DNS) has been dynamic Smagorinsky model performed the best.
employed by Lee and Sung (2008) to investigate the Iaccarino (2001) used three commercial CFD codes,
effects of an adverse pressure gradient on a namely: CFX, Fluent, and Star-CD, and two
turbulent boundary layer. Their numerical results
showed that the mean flow quantities are greatly turbulence models (the low-Re k   and
the v  f ) for the numerical simulation of Obi
2
affected by an APG, and the coherent structures in
the outer layer of the APG flows were more diffuser. The numerical results showed that
activated than those in the zero pressure gradient the v2  f model produces more accurate results
flows. This was attributed to increased turbulence
intensities, shear stresses and pressure fluctuations than the low-Re k   model when comparisons
in the APG systems. More recently, Inouea et al. were made with the experimental data and LES
(2013) reported large eddy simulation (LES) of a predictions. The k   calculations do not show any
turbulent boundary layer at high Reynolds number recirculation region, while the v2  f model
subject to an adverse pressure gradient. The
reproduces the length of the separation bubble
stretched-vortex model (Chung and Pullin, 2009)
within 6 percent of the measured value. More
was used for the subgrid-scale modeling. The
recently, El-Behery and Hamed (2011) employed
results showed self-similarity in the velocity
the commercial code Fluent to examine the
statistics over a wide range of Reynolds numbers. It
capabilities of several turbulence models in
was concluded that the boundary layers under
prediction of turbulent flow in the planar
adverse pressure gradient are far from an
asymmetric diffuser of Obi diffuser. The
equilibrium state.
comparisons showed that the results of the
Among numerous applications of adverse pressure v2  f turbulence model agree best with the
gradient flows, diffusers are one of the most experimental data while the RSM model was shown
commonly used flow devices in the industry, to give unexpected poor results.
especially in hydraulic machines. Therefore,
investigation of turbulent flow through diffusers has Cherry et al. (2008 and 2009) performed
been an important research topic for fluid experiments to measure the mean velocity field in
mechanics researchers. In the following, some of two separate 3D asymmetric diffusers using the
the investigations related to turbulent flow through magnetic resonance velocimetry method. The
diffusers will be reviewed. measured flow fields in both diffusers showed 3D
boundary layer separation but the structure of the
Obi et al. (1993) performed laser Doppler separation bubble exhibited a high degree of
anemometry (LDA) measurements in an sensitivity to the diffuser geometrical dimensions.
asymmetric diffuser with an expansion ratio of 4.7, In their experiment, a fully-developed flow
a single deflection of 10o and a turbulent fully condition was established at the diffuser entry
developed inlet condition. Buice and Eaton (1996) which is suitable for the turbulence modeling
argued that the experimental data from Obi et al. validation proposes. Similar to Obi diffuser, the
(1993) had several deficiencies when comparing Cherry diffuser has been used as a benchmark for
with their numerical results. The most noticeable the investigation of performance of turbulence
problem was that the experimental data of Obi et al. models and numerical approaches in recent years. A
(1993) did not appear to satisfy the mass direct numerical simulation of turbulent flow in the
conservation. Hence, they performed new detailed Cherry diffuser (Cherry et al., 2008) was conducted
measurements in the reproduction of the Obi by Ohlsson et al. (2010) with a massively parallel
experiments using hot-wire and pulsed-wire high-order spectral element method and their results
measurements. The experimental works by Obi et were in good agreement with the experimental data.
al. (1993) and Buice and Eaton (1996) have Jakirlić et al. (2010) applied LES and a zonal
received much attention because of the fully- hybrid LES/RANS scheme to predict the turbulent
developed inlet condition, presence of flow flow through the 3D diffuser investigated by Cherry
separation and flow development downstream of et al. (2008). Both modeling strategies gave
the reattachment point. For example, Kaltenback et acceptable results for the time-averaged quantities.
al. (1999) conducted a numerical investigation on Jeyapaul (2011) also performed numerical
Obi diffuser using LES. They showed that a simulations on Cherry diffuser using RANS
detailed representation of the inflow velocity field models. The results showed that the linear eddy-
is critical for accurate prediction of the flow inside viscosity models fail to predict separation on the
the diffuser. They also found that the sub-grid scale correct wall of the 3D diffuser. However, the
model plays a major role for prediction of both explicit algebraic RSM (EARSM) is able to predict
mean momentum and turbulent kinetic energy. separation accurately. The EARSM predicted
Schlüter et al. (2005) also conducted LES

1030
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

quantitatively the mean flow field, however the 2.102 m straight rectangular duct with a rectangular
Reynolds stresses were incorrect and the wall cross section of 0.1×0.15 m2 is used upstream the
pressure was under predicted. diffuser. The flow is still developing at the end of
the straight duct as found in hydropower systems.
Cervantes and Engström (2008) experimentally The hydraulic diameter of the straight duct
investigated the turbulent flow inside an
asymmetric diffuser with a moderate adverse calculated by 4 Aduct / p is Dh  0.12 m. Hence, the
pressure gradient using LDA technique. The duct length is 17.5 Dh . The Reynolds number based
diffuser has a diverging upper wall, designed to on the bulk velocity and duct hydraulic diameter is
yield an approximately constant adverse pressure
2  104 (flow rate of 2.47 103 m3/s). A contraction
gradient, which opens up from an angle of 2.5 at
(9:1) precedes the duct. In addition, the flow is
the beginning to 7.5 at the end. The diffuser is a tripped at the exit of the contraction by 2 mm
generic model of the rectangular diffuser found at plates, extruding 9.2 mm from the walls,
the end of most hydropower turbines of the Francis corresponding to about 18% of the duct height and
and Kaplan types with an elbow draft tube. The 12% of the duct width. The tripping was necessary
elbow draft tube is found immediately after the to achieve repeatable conditions in the test section,
runner and, as shown in Fig. 1, composed of a as also found by Durst et al. (1998). Following the
conical diffuser, an elbow and a straight rectangular duct is an asymmetric diffuser with a diverging
diffuser. Although the Reynolds number was small upper wall designed to yield an approximately
compared to full-scale turbines, the flow is still constant adverse pressure gradient. The main part of
fully turbulent and the large viscous length-scale this experimental setup, from the duct inlet to the
allows detailed measurements in the boundary layer diffuser outlet, is numerically modeled here. Fig

up to y  1 . They carried out measurements in 3(a) shows a 3D view of the geometry. In addition,
steady regime and three different pulsating regimes Fig. 3(b) represents a 2D schematic of the diffuser
based on the dimensionless frequency with the variation of the diffuser angle. The diffuser
opens up to a cross section of 0.15 × 0.15 m2 at the
(    / u )
 2
namely: quasi-steady
outlet (x= 2.772 m). The diffuser angle increases
(    0.005 ), relaxation ( 0.005     0.02 ) and from 2.5 at the beginning to 7.5 at x= 2.772 m.
quasi-laminar ( 0.02     0.04 ). This test case The flow field inside the diffuser has been
provides accurate experimental data for validating investigated at three stations, namely: x=2.082 m,
CFD calculations with focus on adverse pressure 2.357 m and 2.632 m. As shown in Fig 3. (b), the
gradient effects and non-trivial boundary first station is 20 mm before the start of the diffuser
conditions: both important for hydropower and the second and third stations are at about 40%
simulations. and 80% of the diffuser length downstream,
respectively.

Fig. 1. Schematic of a draft tube. Fig. 2. Experimental setup used by Cervantes

and Engström (2008) (dimensions in mm).

In the present study, numerical investigations are

performed on the diffuser studied experimentally by
Cervantes and Engström in statistically stationary
condition. The main objective of the present study
is to examine the effectiveness of various
turbulence models in prediction of developing
turbulent flow subjected to the adverse pressure
gradient. To the best of our knowledge authors, the (a)
present paper is the first attempt to investigate the
3D asymmetric diffusers with the developing inflow
condition using different turbulence models.

2. PROBLEM DESCRIPTION

The geometry investigated in this paper has been

studied experimentally by Cervantes and Engström (b)
(2008). A schematic of the experimental setup used Fig. 3. (a) 3D l view of the domain. (b) 2D
for their measurements is shown in the Fig. 2. A Schematic of examined diffuser.

1031
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

3. MATHEMATICAL FORMULATION        

(U j )    t    c 1 Pk
x j x j    x j  
3.1 Mean Flow Equations (7)
 2
For an incompressible turbulent flow, the time c 2 f 2 E
k
averaged equations of continuity and momentum
are written as: The empirical constants of the model are presented
in Table 1.
U j
0 (1)
x j The homogeneous dissipation rate can be related to
the true dissipation rate through:
 (U iU j ) 1 P   U i 
    ui u j  (2)  k
 x j      2 ( ) (8)
x j  xi x j   x j

where  and  are fluid density and kinematic The damping functions f  and f 2 are given by:
viscosity and uiu j represents the Reynolds stress
f   exp 3.4 / (1  0.02Rt )2 
tensor.
(9)
3.2 Turbulence Modeling Equations f 2  1  0.3exp( Rt 2 )
The turbulence models employed for computations
where Rt  k /  is the local turbulent Reynolds
2
are the low-Re k   model (proposed by Launder
and Sharma, 1974), the SST k   model (Menter, number.
1994), the v2  f model (Durbin, 1995) and a The term E is first introduced by Jones and Launder
variant of Reynolds stress model (Gibson and (1972) and is expressed as:
Launder, 1978). 2
  2Ui 
3.3.1 Low-Re k   Model E  2 t   (10)
 x j xk 
 
In this turbulence model, the unknown Reynolds
stresses are obtained from the Boussinesq Table 1 Empirical constant of linear k   model
hypothesis:
c c 1 c 2 k 
Ui U j 2
uiu j   t (  )  kij (3) 0.09 1.44 1.92 1.0 1.3
x j x j 3

where the turbulent viscosity is calculated via: 3.3.2. SST k   Model

2
k The main idea of the SST k   model is to
 t  c f  (4)
 combine the robustness of the k   turbulence
To obtain  t , transport equations for the turbulence model near walls with the capabilities of the
k   model away from the walls. The model uses
kinetic energy, k and its dissipation rate are solved.
the Boussinesq hypothesis (equation (3)) to obtain
The transport equation for the turbulent kinetic
the Reynolds stresses. The turbulent viscosity is
energy is written as:
calculated via:
     k  a1k
(U j k )    t    Pk   t  (11)
x j x j   k  x j  max(a1 , SF2 )
2 (5)
 k 
2   where k and ω are obtained by solving the
 x j 
  following transport equations:

where Pk , the generation rate of turbulent kinetic    k  

(U j k )     k t    Pk   k
*
(12)
energy, is obtained from: x j x j  x j 
Ui
Pk  uiu j (6)     
x j (U j )      t     S  
2 2

x j x j  x j 
(13)
The homogeneous dissipation rate of turbulent 1 k 
kinetic energy,  , is obtained by solving the 2(1  F1 )  2
 x j x j
following equation:
The blending function F1 is defined by:

1032
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

 4
 Helmholtz type. The Boussinesq approximation is
   k 500  4   2 k    still used for the evaluation of the Reynolds
F1  tanh min  max  * , 2  , 
2   (14)

    y y   CDk y    stresses. The eddy viscosity  t is given by:
 

 1 k    t  c v2T (17)
where CDk  max  2   2 ,1010  and
  x j x j 
  where the turbulent time-scale T and the turbulent
y is the distance to the nearest wall. length-scale L are obtained from the following
expressions:
The blending function F1 is zero away from
k  
surfaces ( k   model) and switches to one inside T  max  ,6  (18)
  
the boundary layer ( k   model). The S term in 
the turbulent viscosity expression is the invariant
 k 3/ 2  3/ 4 
strain rate and the second blending function F2 is L  cL max  , c 1/ 4 
defined by:     (19)

  2 k 500 
2
 The v 2 transport equation is expressed as:
 
F2  tanh max  * , 2   (15)
  y y 
        v2 

(U j v 2 )    t    kf  v 2  (20)
x j x j   v2  x j  k
The production term has a limiter to prevent the   
build-up of unrealistic high turbulence energy in the
stagnation regions: An elliptic equation is solved for the relaxation
function f:

Pk  min Pk ,10 * k  (16)
2 f 1  v2 2  P
L2  f   c1  1     c2 k (21)
All constants are computed by a blend from the x j
2
T  k 3 
 
corresponding constants of the k   and the k  
model via   1F1  2 1  F1  .The constants of the The coefficients of v2  f turbulence model are
SST model are given in Table 2. given in Table 3.

Table 2 Empirical constant of k   model Table 3 Coefficients of v2  f model

 k1 0.85
0.19
c
k2 1.0
 1.0
v2
 1 0.5
 1.3
2 0.856 4
c 1 1.3  0.25 / 1   cL d / 2 L  
2

 * 0.09  

1 3/40 c 2 1.9

2 0.0828 c1 1.4

1 5/9 c2 0.3

2 0.44 cL 0.3

c 70
3.3.3 v  f Model
2

3.3.4 RSM Model

The v2  f model (developed by Durbin, 1995) is
similar to the standard k   model. In addition, it The Reynolds Stress Model (RSM) is the most
incorporates some near-wall turbulence anisotropy detailed and elaborate RANS turbulence model. In
as well as non-local pressure-strain effects. Instead contrast to isotropic Boussinesq hypothesis based
models, which use algebraic expressions for the
of turbulent kinetic energy, k , the v2  f model evaluation of Reynolds stresses, the RSM models
uses a velocity scale v 2 (velocity fluctuation solve a partial differential transport equation for
normal to the streamlines) for the evaluation of the each Reynolds stress component.
eddy viscosity. The anisotropic wall effects are The exact form of the Reynolds stress transport
modeled through the elliptic relaxation function, f , equations, shown below, can be derived via a
by solving a separate elliptic equation of the mathematical manipulation on momentum equation:

1033
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

were discretized using the central differencing

 (U k ui u j ) 1   u j p  ui p 
 (u k u i u j )
   scheme. The pressure field is linked to the velocity
xk xk   xi xj  field through the SIMPLE pressure correction
   
 
convection
turbulent  diffusion algorithm. The convergence criteria were set to 10-5
for all equations.
U j U i p  u j u i 
 ui u k  u juk    
x k x k   xi x j  4.2 Boundary Conditions
  
production pressure  strain Accurate representation of inlet boundary condition
 ui u j2
u u j is critical for numerical results inside a diffuser
  2 i (Kaltenback et al., 1999). Since the contraction
 xk xk x k x k
    geometry is not fully known and boundary
viscous  diffusion viscous  dissipation
conditions are not available from experimental data,
(22) simulations are performed from section x=100 mm
Clearly the diffusion, the pressure strain and the which is placed right after the tripping part. At this
viscous dissipation terms are not explicitly defined section, the velocity field for the streamwise and
and need to be modeled. The turbulent diffusion normal components is available from experimental
term is modeled using (Gibson and Launder, 1978): data. Fig. 4 shows six different vertical and
spanwise locations where the velocity components
  uiuj  were measured using the LDA technique.

DT ,ij   t  (23)
xk   k xk 
 
Also the viscous dissipation term can be simply
modeled through  ij  (2 / 3) ij .

Here the variant of RSM proposed by Gibson and

Launder (1978) is employed where the pressure
strain correlation (or redistribution term) is
expressed as:
ij  ij ,1  ij ,2  ij , w (24)

where
 2 
ij ,1  C1 uiuj   ij k  (25) Fig. 4. The location of available experimental
k  3  data at section x=100 mm.
 2 
ij ,2  C2  Pij  Cij    ij  P  C   (26) Table 4 Coefficients of RSM model
 3 
C 0.09
 3 3  C k 3/2
ij ,w  C1  uk um nk nmij  uiuk n j nk  uj uk ni nk  l
k 2 2  d C1 
1  2.58 AA20.25 1  exp    0.0067 Ret  

2

 
 3 3 Ck 3/2
C2  km,2 nk nmij  ik ,2 n j nk   jk ,2 ni nk  l
 2 2  d C2 0.75 A
(27)
C1 2 / 3C1  1.67
The Pij and Cij terms are production and convection
C2 max  2 / 3 C2  1 / 6  / C2 ,0 
terms respectively (defined in equation (22)); also
P  1 / 2  Pkk and C  1 / 2  Ckk . The model A 1  9 / 8  A2  A3  
coefficients are given in Table 4. Ret and aij are
A2 aik aki
turbulent Reynolds number and Reynolds-stress
anisotropy tensor, respectively. A3 aik akj a ji

4. NUMERICAL ASPECTS Cl C3/4 / 

 0.4187
4.1 Solver
All equations are discretized using finite-volume
methodology on a collocated grid system. The open A code was developed to evaluate the inlet
source C++ CFD code OpenFOAM (2011) was boundary condition using a 3D surface fit on the
used to perform the computations. The second order experimental data with a polynomial interpolation
upwind differencing scheme was employed for scheme. Then, a moving average filter is applied on
discretization of the convective terms in all the interpolated data. The streamwise velocity is
transport equations. Gradient and Laplacian terms scaled by a factor to match the experimental flow

1034
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

rate after integration. Finally, the velocity profiles (I) and turbulent length-scale ( l ). Turbulent
are calculated on the cell centers at the inlet of the intensity is the ratio of velocity fluctuations to the
computational grid and are imposed as inlet mean velocity (U0). Therefore, assuming an
boundary condition. The 3D surfaces obtained for isotropic field, the turbulent kinetic energy is
both the streamwise and normal velocity obtained by:
components after smoothing are presented in Fig. 5. 3
The spanwise component is obtained using the k  ( IU 0 )2 (28)
2
continuity equation assuming a negligible
streamwise variation of u-velocity ( u / x  0 ),
with respect to v / y and w / z because of the
duct length. The experimental data for the
streamwise velocity are not entirely symmetric.
This may be induced by uncertainty in the
contraction geometry and inflow condition.

(a)

(a)

(b)

(b)
Fig. 5. 3D surfaces fitted to the experimental
data. (a) streamwise and (b) normal velocity
components.

Computations of the flow inside the described

geometry are performed with uniform and non-
uniform velocity boundary conditions to study their
effects on the results. Figs. 6(a)-(c) illustrate the
development of the mean streamwise velocity and
fluctuating part of the streamwise and normal
velocities along the centerline of the duct and
diffuser for both inlet boundary conditions. It is
observed that the numerical results of Case 1 (non- (c)
uniform inlet velocity) are significantly closer to the
experimental data than those obtained in Case 2 Fig. 6. Comparison between results of Case 1
(uniform). Hence, based on this comparison the first (non-uniform inlet B.C.) and 2 (uniform inlet
approach is used for the predictions presented in the B.C.). (a) Mean streamwise velocity and
following sections. fluctuating part of (b) streamwise and
(c) normal velocities.
An isotropic turbulence field is assumed at the inlet
of all cases due to the lack of experimental
turbulent data at the inlet. The turbulent quantities The homogeneous dissipation rate and specific
at the inlet are calculated using turbulent intensity dissipation rate are also obtained from:

1035
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

(a) (b)

(c) (d)
Fig.7. Results of the grid study; development of (a) mean streamwise velocity and (b) its fluctuations
along the duct and diffuser centerline. Streamwise (c) velocity and (d) fluctuations profiles at (c) section
x=2082 mm.

c3/ 4 k 3/ 2 k 1/ 2 homogeneous dissipation rate and all components

  ,  (29) of Reynolds stress tensor are also set to zero. The
l c1/ 4l pressure on wall is also obtained using the
assumption of zero-gradient along normal to the
where l (turbulent length-scale) is set to 0.1Dh . wall.
Hence, by choosing a turbulent intensity at the inlet,
all other quantities can be calculated. Since there is 4.3 Grid Study
no available experimental data on velocity To study the sensitivity of the computational
fluctuations and turbulent intensity at the duct inlet, results to the grid resolution, a grid study was
a study was performed on the effect of inlet performed using the RSM turbulence model. For
turbulent intensity on the results of the mean flow this purpose, four different grids (from coarse grid
and turbulent stresses in the straight duct and of 470,000 cells to fine grid of 3,700,000 cells)
diffuser. The computations were performed using have been generated. All of these grids are non-
the RSM model and different turbulence intensities uniform and fully structured. As it was mentioned
at the inlet and intensity of 3% was found to give earlier all of turbulence models used in this
accurate results. The pressure value at the inlet is investigation are “low-Reynolds number” and
computed using the zero-gradient condition normal these grids are fine enough to resolve the near wall
to boundary. At the exit boundary, zero gradient viscous sub-layer region. In addition, the grid
boundary conditions along normal to the outlet points have been refined along the stream-wise
were imposed for all variables (flow and direction around the diffuser because of high
turbulence) except pressure, which is considered as stream-wise gradients in these regions. Table 5
a constant value. presents the specifications of the five
Since the models used in the present work are computational grids generated for the grid study;
“low-Reynolds number” turbulence models, the nx, ny and nz represent the number of nodes along
near-wall region is resolved up to y+1. Hence, no stream-wise (x), normal (y) and cross-stream (z)
special treatment (such as wall functions) is directions, respectively.
necessary for the wall boundary conditions of the The results obtained from the grid study for four
mean flow and turbulence transport equations, i.e., different grids are presented in Fig. (7). The
the velocity components, turbulent kinetic energy, development of mean streamwise velocity and its
homogeneous dissipation rate, specific dissipation corresponding fluctuation along the centerline of
rate. In addition, for the RSM model, the wall the duct and diffuser are illustrated in Fig. 7(a)
boundary condition of velocity components, and 7(b), respectively. Figs. 7(c) and 7(d) show

1036
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

the variation of streamwise velocity and its flow accelerates immediately and the shear layers
fluctuation along a vertical line placed on the induced by the tripping produce fast growing
symmetry plane at section x=2082 mm. It can be boundary layers.
noted from these four figures that although grid
refinement slightly affects the velocity field inside Figure 9 shows the mean streamwise velocity
the diffuser, it has more noticeable effects on the (normalized with U0  Q / Aduct ) along the centerline
mean and fluctuating velocity in the developing in the straight duct and the diffuser. In both the
region (straight duct). The results obtained on numerical and the experimental results, the
Grid 3 with 1.4106 nodes are almost grid streamwise velocity starts with an overshoot
independent and a finer mesh would not give very induced by the tripping and the sudden decrement
different results. Thus, the subsequent of flow area. After the trip, velocity decreases, but
computational results are obtained using Grid 3. the rate of velocity decrease reduces along the
Figs. 8(a) and 8(b) show the chosen grid in centerline due to the boundary layer growth. From
symmetry plane and inside the diffuser. x=1500 mm to the start of the diffuser, the velocity
gradient becomes smaller, but flow does not show
clear signs of reaching the fully-developed
Table 5 Specifications of grids produced for the
condition. As the flow enters the diffuser, the
grid study
velocity drops rapidly due to the cross-sectional
Grid No. nx ny nz Distance from area increase. It is seen that all turbulence models
wall have successfully predicted the general variation of
Grid 1 156 45 68 y   O(1) the streamwise velocity along the centerline. There
are significant differences between the results of
Grid 2 213 52 78 y   O(1) these models. The SST k   model largely
overpredicts the centerline streamwise velocity due
Grid 3 265 60 90 y   O(1) to an overestimation of the boundary layer
thickness. It is observed that the results of low-Re
Grid 4 390 80 120 y   O(1)
k   and RSM turbulence models are closest to the
experimental data both inside the straight duct and
the diffuser.

(a)

(b)
Fig. 8. Main grid used for the computations. Fig. 9. Development of the streamwise velocity
Cells distribution on (a) Symmetry plane and (b) along centerline.
diffuser.

5. RESULTS AND DISCUSSION

In this section, the numerical results obtained using

four turbulence models are presented and compared
with the experimental data of Cervantes and
Engström (2008).
The variation of the velocity along the centerline of
the duct flows is generally sensitive to the
development of the boundary layers in the duct.
Non-tripped boundary layers will grow along the
duct length, thereby forcing the bulk flow to
accelerate along the centerline. However, in the Fig. 10. Pressure coefficient distribution along
present geometry, the flow is tripped right before the centerline of the upper wall.
the straight duct. With the current tripping, the bulk

1037
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

k  SST k  

v2  f RSM
Fig. 11. Vectors of secondary flow (right) and Contours of streamwise velocity (left) in section
x=2082 m.

The distribution of pressure coefficient, defined by: vectors in all numerical results seem to have
upward directions, which is obviously due to the
p  patm geometry (the lower wall of diffuser is fixed, while
Cp  (30)
1 the upper wall is diverging). In addition, the RSM
U 02
2 model has predicted two rotating vortices
(generated by normal Reynolds stress anisotropy).
along the centreline of the upper wall is illustrated in To understand the structure of these secondary
Fig. 10. The pressure rises to a peak value and then flows it is best to first look at such turbulence
decreases due to the viscous-losses. As the flow driven secondary flows generated in the upstream
enters the diffuser, the pressure starts to increase straight duct. As shown in Fig. 12, with the RSM
again. It is observed that the numerical and computations, two counter-rotating vortices are
experimental results of pressure coefficient observed in each corner at the section x=1500 mm
distributions show the same trend. As mentioned that force the flow to move from the center of the
earlier, the diffuser was designed to yield an duct to the corner and then back to the center. By
approximately constant adverse pressure gradient, moving in the streamwise direction, it is seen that
which is confirmed by the linear variation of the the upper vortices become weaker in section
pressure in the numerical and experimental results. x=2632 mm due to the geometry (diverging upper
In addition, as expected the inviscid theory (which wall). The lower vortex is also merged to the main
the diffuser is designed based on) over-estimates the upward secondary flow forcing the fluid core
pressure coefficient along the diffuser upper wall. displacement to the upward direction in a vortical
Figure 11 shows the predicted secondary flows and pattern. The streamwise velocity contours in Fig. 13
the streamwise velocity contours using different indicate flatter velocity profiles for k   and
turbulence models at section x=2082 mm, (i.e. 20 v2  f models.
mm before the start of diffuser). As expected only
the RSM model is able to predict fairly strong
secondary flow patterns. Using alternative vorticity
form of the Reynolds-averaged Navier-Stokes
equations, one can show that the secondary flows
are generated by the axial mean vorticity, which
becomes non-zero in a straight duct or pipe only if
there are differences between cross-stream normal
Reynolds stress ( v2  w2 ) or Coriolis forces arising
from a spanwise rotation (Speziale et al. 1992).
Therefore, the linear eddy viscosity turbulence
models, based on the Boussinesq hypothesis, are
not able to predict secondary flows in a straight Fig. 12. Vectors of secondary flow (right) and
non-circular duct. Consequently, the results Contours of streamwise velocity (left) in section
presented in Fig. 11 are reasonable. Secondary flow x=1500 m predicted by RSM model.

1038
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

k  SST k  

v2  f RSM
Fig. 13. Vectors of secondary flow (right pictures) and Contours of streamwise velocity (left pictures) in
section x=2632 m.

The predicted velocity and the turbulence field 

holds in the viscous sublayer ( y  5 ) regardless of
inside the diffuser using different turbulence models
the adverse pressure gradient. However, with the
are further presented and compared to the
flow development, the wake region grows both in
experimental data in the following. These quantities
strength and size. Hence, the outer part of the log
are reported at three different sections of the
region becomes smaller. The increase in the wake
diffuser symmetry plane. The sections are placed at
component is due to the significant decrease in the
2082 (20 mm before the diffuser), 2357 (255 mm
skin friction along the diffuser. This feature of
into the diffuser) and 2632 mm (530 mm into the
adverse pressure gradient on the fluid velocity
diffuser), respectively.
development has been also reported in the literature
Numerical and experimental profiles of the (see for example Nagano et al., 1993).
streamwise component of the velocity are compared
The predicted normal velocity profiles inside the
in Fig. 14. It is observed that the SST k   model
diffuser are compared with the measured data in
largely over-predicts the velocity field inside the Fig. 16. The normal velocity profiles inside the
diffuser, which is consistent with the centreline diffuser are affected by both the diffuser geometry
velocity predictions shown in Fig. 9. The other and the secondary flows. Since the first section
turbulence models demonstrate better streamwise (x=2082 mm) is just before the diffuser, the normal
velocity predictions inside the diffuser. As velocity distribution is more affected by the
mentioned earlier, the flow entering the diffuser is secondary flow motion. However, in the
not fully-developed and thus the significant downstream stations the geometry mainly
differences seen in velocity predictions depends on influences the normal velocity profile. As expected,
the capability of the turbulence models in the normal velocities in the first section are lower
reproducing the developing boundary layer in the than those in the downstream stations. As shown
straight duct. Fig. 14 reveals that among the earlier the eddy viscosity turbulence models fail to
turbulence models, as expected, the RSM predict correct turbulence driven secondary flows,
turbulence model returns the best predictions for the and as a result they are not able to predict accurate
streamwise velocity. In addition, a comparison of normal velocity profiles in the first section. As
velocity profiles at the three sections illustrates the expected the RSM model produces a more accurate
rapid boundary layer growth along the diffuser cross velocity profile in the first station due its
initiated by the trip (i.e. the mean streamwise ability in prediction of turbulence driven secondary
velocity profile becomes less flat). The streamwise motion as already shown in Fig. 11. In the two
velocity predictions of the RSM model are further downstream sections, the contribution of the
compared in log-scale in Fig. 15 in order to geometry on the normal velocity distribution
investigate the effect of the adverse pressure becomes dominant. Hence, the numerical results of
gradient on the boundary layer development. It is different turbulence models are more similar and
 
observed that the linear law of the wall, U  y , are closer to the experimental data. Nevertheless,

1039
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

the results of RSM model are still in superior

agreement with the measured data.

x=2082 mm

x=2082 mm

x=2357 mm

x=2357 mm

x=2632 mm
Fig. 16. Normal velocity along vertical direction
x=2632 mm inside the diffuser.
Fig. 14. Streamwise velocity along a vertical
direction inside the diffuser.
Figure 17 shows comparisons for the streamwise
normal Reynolds stress ( u 2 ) in the diffuser. The
comparisons show that the eddy-viscosity based
turbulence models, as expected, significantly under-
estimate the turbulent normal levels. On the other
hand, the RSM model is able to produce accurate
predictions for this quantity. The comparison of
results in the three sections reveals that the
streamwise normal Reynolds stress develops an
outer plateau in the presence of an adverse pressure
gradient. The intensity of this plateau increases in
the streamwise direction. This property of a
turbulent boundary layer flow subjected to an
adverse pressure gradient is more obvious in a
Fig. 15. “Law of the wall” plots of RSM model. semi-log plot as shown in Fig. 18.

1040
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

x=2082 mm
x=2082 mm

x=2357 mm
x=2357 mm

x=2632 mm

Fig. 17. Streamwise component of normal x=2632 mm

Reynolds stress along vertical direction inside Fig. 19. Cross-stream component of normal
the diffuser. Reynolds stress inside the diffuser.

Comparisons v2 predictions using various

turbulence models and experimental data are
presented in Fig. 19. As can be observed the all
eddy-viscosity based turbulence models over-
predict v2 levels. As expected, the RSM model
returned more reliable v2 . Numerical and
experimental results of shear Reynolds stresses are
illustrated in Fig. 20. It is seen that the RSM model
is able to produce acceptable results for the
turbulent shear Reynolds stress distributions in all
three locations. It is noted that the k   and
v 2  f models are also yield acceptable results for
Fig. 18. Streamwise component of the normal
Reynolds stress for the RSM model. this quantity.

1041
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

measurements were employed for the numerical

predictions. Four RANS turbulence models are
employed for the numerical predictions. The results
of centerline velocity in the straight duct
demonstrated the best predictions can be obtained
using the low-Re k   and RSM turbulence
models. The SST k   largely overpredicts the
centerline velocity and boundary layer thickness.
On the other hand, the RSM model produces most
precise results for the pressure recovery. It was
observed that the normal velocity distribution inside
the diffuser is affected by geometry and secondary
flows and the RSM model yields the best
predictions for this quantity due to its anisotropy
nature. It was further shown that streamwise normal
x=2082 mm Reynolds stress is significantly underestimated by
the isotropic based turbulence models. These
turbulence models overpredict the cross-stream
normal Reynolds stress. In contrast, the RSM model
produces more precise results for the turbulent
Reynolds stresses.

ACKNOWLEDGEMENTS

The authors gratefully acknowledge the financial

support of Hydraulic Machinery Research Institute
of University of Tehran and Swedish Hydropower
Centre (www.svc.nu).

REFERENCES
x=2357 mm
Bradshaw, P. (1967). The turbulence structure of
equilibrium boundary layers. J. Fluid Mech.
29, 625-645.
Buice, C. U. and J. K. Eaton (1996). Experimental
investigation of flow through an asymmetric
plane diffuser. CTR Annual research briefs-
1996, 243-248.
Cervantes, M. J. and T. F. Engström (2008).
Pulsating turbulent flow in a straight
asymmetric diffuser. J. Hydraulic Res., 49 (1),
112-128.
Cherry, E. M., C. J. Elkins and J. K. Eaton (2008).
Geometric sensitivity of three-dimensional
separated flows. Int. J. Heat Fluid Flow 29,
803–811.
x=2632 mm
Cherry, E. M., C. J.Elkins and J. K. Eaton (2009).
Fig. 20. Shear Reynolds stress along vertical
Pressure measurements in a three-dimensional
direction inside the diffuser.
separated diffuser. Int. J. Heat Fluid Flow 30,
1-2.
6. CONCLUSION Chung, D. and D. Pullin (2009). Large-eddy
simulation and wall-modeling of turbulent
The 3D developing turbulent flow through a channel flow. Journal of Fluid Mechanics 631,
rectangular asymmetric diffuser with moderate 281–309.
adverse pressure gradient was numerically
investigated and the numerical results of the Clauser, F. H. (1954). Turbulent boundary layers in
predictions are compared with the LDA data. Such adverse pressure gradients. J. Aero. Sci. 21(2),
calculations are important in hydraulic machinery 91–108.
flows, due to importance of adverse pressure Durbin, P. A. (1995). Separated flow computations
gradient in these flows. It was observed that the
developing flow inside the diffuser is significantly with the k    v2 model. AIAA Journal 33
sensitive to the inlet boundary condition. Therefore, (4), 659-664.
the inlet conditions extracted from experimental Durst, F., M. Fischer, J. Jovanović and H. Kikura

1042
S. Salehi et al. / JAFM, Vol. 10, No. 4, pp. 1029-1043, 2017.

(1998). Methods to Set Up and Investigate Application of the Energy Dissipation Model
Low Reynolds Number, Fully Developed of Turbulence to the Calculation of Flow near
Turbulent Plane Channel Flows. J. Fluids Eng. a Spinning Disc, Letter in Heat Mass Transfer
120, 496–503. 1, 131-138.
El-Behery, S. M. and M. H. Hamed (2011). A Lee, J. H. and H. J. Sung (2008). Effects of an
comparative study of turbulence models adverse pressure gradient on a turbulent
performance for separating flow in a planar boundary layer. Int. J. Heat and Fluid Flow
asymmetric diffuser. Computers and Fluids 44, 29, 568–578.
248–257.
Menter, F. R. (1994). Two-Equation Eddy-
Gibson, M. M. and B. E. Launder (1978). Ground Viscosity Turbulence Models for Engineering
Effects on Pressure Fluctuations in the Applications. AIAA Journal. 32(8), 1598–
Atmospheric Boundary Layer. J. Fluid Mech. 1605.
86. 491–511.
Nagano, Y., M. Tagawa and T. Tsuji (1993).
Iaccarino, G. (2001). Prediction of a turbulent Effects of Adverse Pressure Gradients on
separated flow using commercial CFD codes, Mean Flows and Turbulence Statistics in a
J. Fluids Eng. 123(4), 819-828. Boundary Layer. Turbulent Shear Flows 8, 7-
21.
Inouea, M., D. I. Pullina, Z. Harunb and I. Marusicb
(2013). LES of the adverse-pressure gradient Obi, S., K. Aoki and S. Masuda (1993).
turbulent boundary layer. Int. J. Heat and Experimental and computational study of
Fluid Flow 44, 293–300. turbulent separating flow in an asymmetric
plane diffuser. Ninth symposium on turbulent
Jakirlić, S., G. Kadavelil, M. Kornhaas, M. Schäfer, shear flows, Kyoto, Japan, 305-1–4.
D. C. Sternel and C. Tropea (2010). Numerical
and physical aspects in LES and hybrid Ohlsson, J., P. Schlatter, P. F. Fischer and D. S.
LES/RANS of turbulent flow separation in a 3- Henningson (2010). Direct numerical
D diffuser. Int. J. Heat and Fluid Flow 31, simulation of separated flow in a three-
820–832. dimensional diffuser. J Fluid Mech. 650, 307-
318.
Jeyapaul, E. (2011). Turbulent flow separation in
three-dimensional asymmetric diffusers, PhD Rai, S. P. (1986). Wall-wakes in moderate adverse
Thesis. Iowa State University, United States. pressure gradients. J. Hydraulic Res., 24(5),
377-390.
Jones, W. P. and B. E. Launder (1972). The
Prediction of Laminarization with a Two Schlüter, J. U., X. Wu and H. Pitsch (2005). Large-
Equation Model of Turbulence, Int. J. Heat Eddy Simulations of a Separated Plane
and Mass Transfer 15, 301-314. Diffuser. 43rd AIAA Aerospace Sciences
Meeting and Exhibit, AIAA 2005-0672.
Kaltenback, H. J., M. Fatica, R. Mittal, T. S. Lund
and P. Moin (1999). Study of the flow in a Speziale, C. G., R. M. C. So and B. A. Younis
planar asymmetric diffuser using large eddy (1992). On the prediction of turbulent
simulations. J Fluid Mech. 390, 151–185. secondary flows, NASA Contractor Report
189722 OpenFOAM 2.0.0 User Guide.
Launder, B. E. and B. I. Sharma (1974).

1043