Proceedings of the 11th Annual Conference of the CFD Society of Canada

Numerical modeling of airflow over the Ahmed body

Yunlong Liu, Alfred Moser

Air and Climate Group, Swiss Federal Institute of Technology, ETH-Zentrum WET A1, CH-8092 Zurich,
Switzerland, Tel: +41 1 6326915, Fax: +41 1 6321023

Email: [email protected], URL: http://www.airflow.ethz.ch/

ABSTRACT performance of the SST model and some other

turbulence models.
Airflow over the Ahmed body is investigated by
means of transient RANS turbulence models. The
simulations have been performed using two different
differencing schemes. The performances of several
RANS turbulence models have been compared. It has
been found that Durbin’s k-ε-v2 model is more
accurate than the other turbulence models for the
wall-bounded cases with separation and
reattachment. A wall function for k-ε-v2 model has
been introduced to avoid the divergence when very Figure 1: Schematic of the Ahmed body model[1]
fine mesh is employed for complex geometries.
Numerical results agree well with the reported
experiments.

1. INTRODUCTION
In order to investigate the behavior of newly
developed turbulence models for complex geometry
cases, a simplified car model, known as the Ahmed
body, has been tested by Ahmed, et al[1] in the early
1980s. The Ahmed body is made up of a round front
part, a moveable slant plane placed in the rear of the
body to study the separation phenomena at different
angles, and a rectangular box, which connects the
front part and the rear slant plane, as shown in Figure
1. All dimensions listed in figure 1 are in mm.
Several researchers have worked on the experiments
and numerical modeling of the flow over the Ahmed Figure 2:Characteristic drag coefficients for the
body. Ahmed studied the wake structure and drag of Ahmed body for various rear slant angles ϕ
the Ahmed body[2-3], Lienhart and his colleagues[4] measure by Ahmed[1]
conducted the experiments for two rear slant angles
(25°, 35°) at LSTM. The velocities and turbulence
kinetic energies have been measured by LDA at As the wake flow behind the Ahmed body is the
several key locations. This paper will take the LSTM main contributor to the drag force, accurate
test results as the validation data. Craft[5] compared prediction of the separation process and the wake
the performance of linear and non-linear k-ε model flow are the key to the successful modeling of this
with two different wall functions. Basara[6] case. To simulate the wake flow accurately,
conducted the numerical modeling of this case by resolving the near wall region using accurate
means of large eddy simulation (LES), Menter[7] turbulence model is highly desirable. This paper
compared the will study the effectiveness of three different
turbulence models, including the k-ε-v2 model[8-9],

507
508
 Proceedings of the 11th Annual Conference of the CFD Society of Canada

the k-ε model and the full stress model, for the been reported[8-9, 11-13] that this model is a
modeling of the flow over the Ahmed body, and significant improvement over the two-equation
shows the behavior of different turbulence models, model for several test cases, such as channel flow,
as well as the effect of the grid layout and backward facing step, etc. Governing equations
differencing schemes on the numerical results. read:
Dk ∂  ν t ∂k 
= Pk − ε + ( ν + )  (4)
Dt ∂x j  σ k ∂x j 
2. TURBULENCE MODELS
Airflow over the Ahmed body is governed by the 
Dε C ε1 Pk − C ε 2 ε ∂ ν t ∂ε 
Navier-Stokes equations. As turbulent flow is made = + ( ν + )  (5)
up of a spectrum of vortex scales, the turbulence Dt T ∂x j  σ ε ∂x j 
energy is distributed through the whole spectrum
based on the wavelength. Ideally, resolving all the
scales can offer the best insight into the Dv 2 ε ∂ ν ∂v 2
= kf 22 − v 2 + [(ν + t ) ] (6)
understanding of the turbulent flow, which can be Dt k ∂x j σ k ∂x j
accomplished by direct numerical simulation (DNS).
However, it is not practical to resolve all the scales
∂  ∂f 22 
 − f 22 = (1 − C1 ) [ 2 / 3 − v / k ] − C 2 Pk (7)
2
for engineering problems such as the flow over the 
Ahmed body. While TRANS, a transient Reynolds ∂x j  ∂x  T k
 j
averaged Navier-Stokes approach(RANS), offers a
very promising approach because the large scales can f 22 is a quotient of the pressure strain Φ 22 by the
be resolved while the small scales, which carry less
turbulence energy compared to the large scales, are turbulent kinetic energy k
modeled by RANS sub-scale models. The averaged
Navier-Stokes equations take the following form: ν t = Cµ v 2T
∂ρU i (1)
=0 Cµ = 0.19 , C ε1 = 1.44 , C ε 2 = 1.9 ,
∂x i
DU i 1 ∂P ∂  ∂U i ∂U j  σ k = 1.0 , σ ε = 1.3
= gi − + ν ( + ) − uiu j  (2)
Dt ρ ∂x i ∂x j  ∂x j ∂x i 
According to the Boussinesq assumption, the ∂U j ∂U i ∂U j
Pk = ν t ( + )
isotropic eddy viscosity/diffusivity formulation for ∂x i ∂x j ∂x i
Reynolds stress reads:

2 ∂U ∂U j The time scale T and length scale L can be obtained

uiu j = kδ ij − ν t ( i + ) (3) from the following:
3 ∂x j ∂x i
In the standard k-ε turbulence model, the Boussinesq k ν
assumption is applied together with wall functions. It T = max{ ,6( )1 / 2 }
is widely used and the convergence is stable.
ε ε
However, the above-mentioned assumption is not k3/ 2 ν3
always true because of the an-isotropic nature of the L = 0.3 max{ , C η ( )1 / 4 } ,
ε ε
flow in specific cases, such as the cases that involve
the flows in the near wall region. 20ν 2 v 2
f 22 w = − ( ) p , where index w and p
According to Launder[10], the normal stress v ,
2
εw y4
perpendicular to the wall, plays the most important denote the values at the wall and that in the first
role to the eddy viscosity. Motivated by this idea, cell above the wall, respectively.
Durbin[8-9] devised a three-equation model, known as
the k-ε-v2 model, or v2f model. The idea is to resolve To study the performance of the k-ε-v2 turbulence
2
the normal stress v , along with solving the model for application in complex geometries,
several different turbulence models, including the
modified k and ε equations. The near wall region is
resolved exactly and the wall-reflection is considered k-ε model with wall function, the k-ε-v2 model and
by means of elliptic relaxation in the model. It has full Reynolds stress model with wall function have
been investigated in this paper. In this study, a wall

508
509
 Proceedings of the 11th Annual Conference of the CFD Society of Canada

function for f 22 is introduced in the near wall region computation of several blocks. 16 processors are
employed for parallel computation. To limit the
for k-ε-v2 model.
total number of element cells, a non-uniform
Based on the numerical results for the channel flow
structured grid is constructed, with the near wall
DNS data of Kim[14] et al (1987), in the near the wall
region using smaller grid size to control the first y+
region, we can get the non-dimensional value of f22
at around 20 to 50. Total element number is
as a function of y+, so the f22 value for the first near
460,000.
wall cell can be obtained. The following formula is
It has been found that the grid quality is crucial for
used to calculate the f22 value for the first near wall
convergence, and smaller time step is required in
element:
the initial stage to ensure the computation not to
f 22 ( y + ) = C µ k (0.65 /( y + + 13.0) − 4.44 /( y + + 13.0)1.58 ) / ν
0.5
diverge. The differencing scheme can affect the
+
( y ≥20) accuracy of the solution, in the initial stage, the
0 .5 upwind differencing scheme is employed in the
f 22 ( y + ) = C µ k[0.00199 + 2.511 × 10 −4 ( y + − 20)] / ν initial computation stage because it is robust, but it
( y+ <20) is not as accurate as second order central difference
scheme. At time=0.6s it is switched to the second
It should be pointed out that the difference from order central difference scheme to get a good final
Durbin’s original model is that a value is given on solution.
the first point near the wall for f22 equation, along
with the wall functions for the other equations, just
like the case for the k-ε model. The purpose of this
boundary condition for f22 equation is to improve the
robustness of the original model.

3. PHYSICAL CASE AND NUMERICAL

METHOD
The size of the computation domain is 7.044m long,
2.0m wide and 1.05m high, with the down stream Figure 3: Grid distribution near the Ahmed body
portion extended five meters behind the rear of the
Ahmed body. The Ahmed body is 1.044m long,
0.288m high and 0.389m wide, with a rear slant 4. RESULTS AND DISCUSSION
angle of 35°, as shown in figure 1. The projected The results shown below are the averaged value
area A x of the Ahmed body in the mainstream over several cycles of the transient flow field,
direction is 0.112m2, which corresponds to a unless it is specially stated.
blockage ratio of 5.33%. The bottom surface of the At upstream air velocity of 40m/s, an unsteady
Ahmed body is located at 0.05m above the ground. wake in the downstream is generated downstream.
The incoming flow, located at one meter upstream of The unsteady wake comprises two vortices behind
the front surface, is at 40 m/s with 0.2% free stream the rear with the larger one in the higher part, and
turbulence level, the corresponding Mach number is the smaller one in the lower part, as shown in the
about 0.115, which gives a vehicle height based streamline in figure 4. The background color in
Reynolds number of 7.68×105. Airflow is assumed to figure 4 shows the contour of the turbulent kinetic
be incompressible, heat transfer is not considered in energy k. It is found that the peak value of k is
this study. Outflow is assumed fully developed and located in the center of the small vortex
the zero-gradient velocity boundary condition is downstream of the body, as observed in the
imposed. The ground and the body surfaces are experiment[4].
treated as no-slip smooth walls. All the other Figure 5 shows the mean velocity and flow field in
boundaries are symmetry, because the size of the the symmetry plane of the Ahmed body. Vortical
wind tunnel is much larger than the computation structures extend more than 0.5m beyond the end
domain. Figure 3 shows part of the grid distribution the body rear. The reverse flow climbs up to the
near the Ahmed body symmetrical plane. Because of rear slant, as observed in the experiment[4].
the complexity of this Ahmed body model, the
computation domain has been split into 46 blocks, so
that each processor is responsible for the

509
510
 Proceedings of the 11th Annual Conference of the CFD Society of Canada

Figures 7a-d show the downstream development of

the counter-rotating vortex system and contours of
the predicted turbulent kinetic energy k at four
different sections: 80mm, 200mm, 500mm and
1500mm downstream of the body, respectively. It
confirms that two counter-rotating trailing vortices
are generated downstream of the Ahmed body, and
the vortices effect, though very small, still remains
at more than 1.5m away from the rear, and the
Figure 4: unsteady wake behind the body predicted velocity deficit still visible at more than 4m behind
using the k-ε-v2 model the Ahmed body.
Figures 8 shows the drag coefficients predicted by
the k-ε-v2 model. The time averaged drag force FD
is found by integration of surface pressure and the
shear stress, the drag coefficient is defined as
following:
FD
CD = 2
0.5ρu 0 A x
Where ρ is the air density, u 0 is the upstream bulk
velocity, A x is the projected area of the Ahmed
body in x direction.
Figure 5: Flow field in the symmetry section colored It should be pointed out that before time=0.6s, the
by the stream-wise u-velocity predicted by the k-ε-v2 upwind differencing scheme is used. At time=0.6s,
model the differencing scheme is switched to the second
order central differencing scheme. Its averaged
value of the drag force coefficient is 0.262, which
shows that the numerical result agrees quite well
with the experiment data of Ahmed[1], which is
about 0.26, if the second order central differencing
scheme is employed. This verified that, not only the
turbulence model and the near wall grid resolution
are affecting the accuracy of the predicted drag
coefficient, the proper selection of the differencing
scheme also plays an important role to make CFD
more accurate.
Table 1 gives a list of the measurement data and the
predicted drag coefficient and its components,
Figure 6: Comparison of velocity profile of where Ck, CB, CS, CD represents the drag coefficient
numerical results and experiment at the nose, back, the rear slop and the total.

Figure 6 gives a mean velocity profile comparison of Pref is the reference pressure which is set to 0 Pa,
the numerical results with the experiments for the when comparing the drag coefficients, the force
separation zone. Geometrical parameters are components are sensitive to its location. In this
normalized by the height of the Ahmed body, L paper the reference pressure is located at the outlet.
(0.288m). Compared with the standard k-ε model Table 1: Validation of drag force and force
and full Reynolds stress model, the k-ε-v2 model gets components
better results for velocities above the rear slant and
behind the Ahmed body, because the velocities Ck CB CS CD Error%
predicted by the k-ε-v2 model fit well with the Ahmed[1] 0.020 0.095 0.090 0.260 -
LSTM[4] - 0.129 0.121 - -
experimental data[4]. The numerical results of the full k-ε+WF 0.026 0.105 0.111 0.242 -6.8
stress model and the k-ε model predicted a wake k-ε-v2 0.020 0.124 0.120 0.264 +1.5
being recovered too soon at the downstream, and SSG 0.010 0.102 0.098 0.210 -19.2
predicted velocities have larger discrepancies when RSM 0.013 0.093 0.178 0.282 +8.5
compared to the experiment data. SST 0.026 0.107 0.108 0.241 -7.3

510
511
 Proceedings of the 11th Annual Conference of the CFD Society of Canada

Based on the comparison of drag coefficients with

LSTM experiment data, it can be concluded that
Durbin’s k-ε-v2 models gives the best result,
followed by k-ε, SST and RSM model.

7c:500mm

7a: 80mm

7d: 1500mm
Figure 7a-d: Velocities and contours of the
predicted turbulent kinetic energy k at 80mm,
200mm, 500mm and 1500mm behind the body
predicted by the k-ε-v2 model

7b: 200mm

Figure 8: Drag force coefficient predicted by

the k-ε-v2 model

511
512
 Proceedings of the 11th Annual Conference of the CFD Society of Canada

5. CONCLUSIONS [9] Durbin P.A.: Separated Flow Computations

with the k-ε-v2 Model, AIAA Journal, V33, N4,
The flow field and drag force of the flow over the PP659-664, 1995
Ahmed body can be simulated by computational [10] Launder, B. E., Low Reynolds Number
approach. Compared with the standard k-ε model Turbulence Near Walls, UMIST Mechanical
and the full stress model, the k-ε-v2 model performs Engineering Dept, Rept.TFD/86/4, University of
better, and the second-order central differencing Manchester, England, UK, 1986.
scheme is more accurate than upwind scheme. [11] M. Behnia, S. Parneix and P. A. Durbin,
Prediction of heat transfer in an axisymmetric
ACKNOWLEDGEMENTS turbulent jet impinging on a flat plate. International
journal of heat and mass transfer, V41, N12,
This study is supported by MOVA, a project PP1845-1855, 1998
undertaken by thermo-fluids section directed by
professor K. Hanjalic. Part of this work is finished at [12] Georgi Kalitzin, Towards a robust and
TU Delft under the supervision of professor K. efficient v2f implementation with application to
Hanjalic. Academic discussion with W. Khier, O. transonic bump flow, CTR Annual Research Briefs,
Ouhlous and M. Hadziabdic at TU Delft, The PP291-299, 2000
Netherlands, is gratefully acknowledged. Special [13] Lien F S, Durbin, P.A., Non-linear k-ε-v2
thanks are extended to the Cray T3E computer center modeling with application to high lift. Proc. of the
at Delft University of Technology, The Netherlands. summer program, CTR, NASA Ames/Stanford
Univ. 5-22, 1996
[14] Robert D. Moser, John Kim and Nagi N.
Mansour, DNS of turbulent channel flow up to
REFERENCES Re τ =590, Physics of Fluids, V111, N4, PP943-
945, 1999
[1] Ahmed, S.R., Ramm G.: Some Salient Features
of the Time-Averaged Ground Vehicle Wake. SAE
Technical Paper 840300, 1984.
[2] Ahmed, S.R.,Wake Structure of typical
automobile shapes. Transactions of the ASME,
Journal of Fluids Engineering V103, P162-169, 1981
[3] Ahmed, S.R., Influence of base slant on the wake
structure and drag of road vehicles. Transactions of
the ASME, Journal of Fluids Engineering V105,
P429-434, 1983
[4] Lienhart H., Stoots C.,Becker S., Flow and
turbulence structures on the wake of a similified car
model(Ahmed model), DGLR Fach. Symp. der AG
ATAB, Stuttgart University, 2000.
[5] T. J. Craft, S. E. Gant, H. Iacovides, B.E.
Launder, Computational study of flow around the
Ahmed car body, 9th ERCOFTAC workshop on
refined turbulence modeling, Darmstadt University
of Technology, Germany, 2001
[6] B. Basara, Numerical simulation of turbulent
wakes around a vehicle, FEDSM 99-7324, 1999
[7] L. Durand, M.Kuntz, F. Menter, Validation of
CFX-5 for the Ahmed Car Body, 10th joint
ERCOFTAC (SIG-15) -IAHR-QNET/CFD
Workshop on Refined Turbulence Modelling,
October 10-11, 2002
[8] Durbin P.A.: Near Wall turbulence closure
modeling without “damping functions”. Theoretical
and Computational Fluid Dynamics 3, P1-13, 1991

512
513