Latin American Applied Research 39:295-306(2009)

NUMERICAL SIMULATION OF THE FLOW AROUND THE AHMED

VEHICLE MODEL

G. FRANCK†‡, N. NIGRO‡, M. STORTI‡ and J. D’ELÍA‡

† Universidad Tecnológica Nacional (UTN), Facultad Regional Santa Fe (FRSF)
Lavaise 610, 3000-Santa Fe, Argentina, [email protected], http://www.utn.edu.ar
‡ Centro Internacional de Métodos Computacionales en Ingeniería (CIMEC)
Instituto de Desarrollo Tecnológico para la Industria Química (INTEC)
Universidad Nacional del Litoral -CONICET, Güemes 3450, 3000-Santa Fe, ARGENTINA
ph.: (54342)-4511 594/5, fx: (54342)-4511 169
(nnigro, mstorti, jdelia)@intec.unl.edu.ar http://www.cimec.org.ar

Abstract— The unsteady flow around the Ahmed three dimensional regions of separated flows where the
vehicle model is numerically solved for a Reynolds largest one is the wake. In a time-averaged sense, the
number of 4.25 million based on the model length. A separated flow regions exhibit complicated kinematic
viscous and incompressible fluid flow of Newtonian macro structures and those present in the wake deter-
type governed by the Navier-Stokes equations is as- mine mostly the body drag. Nowadays, numerical simu-
sumed. A Large Eddy Simulation (LES) technique is lations, wind tunnel and road tests are jointly used in the
applied together with the Smagorinsky model as automotive industry for the aerodynamic study from
Subgrid Scale Modeling (SGM) and a slightly modi- several perspectives.
fied van Driest near-wall damping. A monolithic The Ahmed vehicle model is a very simplified bluff-
computational code based on the finite element me- body which is frequently employed as a benchmark in
thod is used, with linear basis functions for both vehicle aerodynamics. It has been used in several ex-
pressure and velocity fields, stabilized by means of periments (Ahmed et al., 1984; Sims-Williams, 1998;
the Streamline Upwind Petrov-Galerkin (SUPG) Bayraktar et al., 2001; Spohn and Gilliéron, 2002;
scheme combined with the Pressure Stabilizing Pet- Lienhart and Becker, 2003) and numerical studies (Han,
rov-Galerkin (PSPG) one. Parallel computing on a 1989; Basara et al., 2001; Basara, 1999; Basara and
Beowulf cluster with a domain decomposition tech- Alajbegovic, 1998; Gilliéron and Chometon, 1999; Ka-
nique for solving the algebraic system is used. The padia et al., 2003). A slightly modified version was also
flow analysis is focused on the near-wake region, studied by Duell and George (1999); Barlow et al.
where the coherent macro structures are estimated (1999); Krajnović and Davidson (2003).
through the second invariant of the velocity gradient The shape of this body is free from all accessories
(or Q-criterion) applied on the time-average flow. It and wheels but it still retains the primary behavior of the
is verified that the topological features of the time- vehicle aerodynamics, as seen in Fig. 1. Special atten-
average flow are independent of the averaging time tion is focused on the time-averaged flow in the near-
T and grid-size wake and the dependence of drag on the slant angle α at
Keywords— Ahmed vehicle, bluff aerodynamics, the top rear end. The rear end is a simplification of a so-
incompressible viscous fluid, large eddy simulation called fastback one such as on a Volkswagen Golf I.
(LES), time-average flow, finite element method, This model was tested in open jet wind tunnels and no
fluid mechanics. velocity distributions were reported in the inlet and out-
let boundaries (Ahmed et al., 1984). Previous numerical
I. INTRODUCTION simulations had assumed the incoming flow to be lami-
A. The Ahmed vehicle model nar and steady (Han, 1989; Krajnović and Davidson,
Ground vehicles can be classified as bluff-bodies that 2004; Basara and Alajbegovic, 1998). Wind tunnel ex-
move close to the road surface and are fully submerged periments on this model show two critical slant angles:
in the fluid. In general, for the usual velocities of com- αm≈12.5º and αM≈30.0º, called first and second critical
mercial passenger cars, buses and trucks, compressible angles, respectively, where the main topological struc-
effects can be neglected and an incompressible viscous ture of the time-averaged flow in the near-wake changes
fluid model can be assumed. As the Reynolds numbers significantly as summarized in Table 1 (Ahmed et al.,
based on the body length are usually too high, the flow 1984; Sims-Williams, 1998; Bayraktar et al., 2001;
regimes are fully turbulent. In addition, a key feature of Spohn and Gilliéron, 2002; Lienhart and Becker, 2003).
the flow field around a ground vehicle is the presence of Among the experimental tests, Janssen and Hucho
several separated flow regions, while the net aerody- (1975) showed the dependence of the flow on the slant
namic force is the result of complicated interactions angle at the rear end for an industrial vehicle model,
among them. Even simple basic vehicle configurations while Morel (1978) performed tests on the Morel body
with smooth surfaces, free from appendages and wheels, which is previous to the Ahmed one. Both models rep-
generate a variety of quasi two dimensional and fully
295
Latin American Applied Research 39:295-306(2009)

Figure 1: Ahmed model, dimensions in mm.

Table 1: Time-averaged flow in the near-wake of the Ahmed predicted for these cases leading to a poor estimation of
vehicle as a function of the slant angle α, where αm≈12.5º andthe mean aerodynamic forces. Physical turbulence con-
αM ≈ 30.0º. sists of an intricate interplay between random and co-
slant angle α time-averaged flow in the near-wake herent structures, whose intermittency (an increasing
α < αm almost 2D attached sparseness of the small-scale eddies) is a typical prop-
αm < α < αM massively separated 3D erty. Moreover, it was shown that solutions of RANS
α > αM almost 2D attached equations agree with the time-averaged solutions of the
unsteady N-S equations only if the Reynolds stress ten-
resent two benchmarks widely used in the automotive sor used in the RANS equations is well approximated
industry (Basara and Alajbegovic, 1998). (McDonough, 1995). However, this requirement is not
On one hand, among numerical simulations, a LES generally reached at present, which leads to large differ-
around a bus-like vehicle is developed by Krajnović ences between the numerical results obtained with both
(2002), Krajnović and Davidson (2003), while Gilliéron strategies. Krajnović and Davidson (2004) have noted
and Spohn (2002) showed a Detached Eddy Simulations that previous RANS numerical simulations of the flow
(DES) over another simplified vehicle model. Other around the Ahmed model performed either relatively
sources of related work can be found in Hucho (1998) well or poorly according to the slant angle α, which in-
and ERCOFTAC (2007). fluences the main topology of the time-averaged flow in
Regarding numerical benchmarks with the present the near wake, as it was summarized in Table 1. They
model, several computations were performed using the argued that possible reasons of the failures could be due
Reynolds Averaged Navier-Stokes (RANS) equations to (i) higher level of flow unsteadiness; (ii) the fact that
(Han, 1989; Basara et al., 2001; Basara, 1999; Basara most RANS turbulent models miss the flow separation
and Alajbegovic, 1998). On the other hand, LES were at the rear part for such slant angles; and (iii) the fact
employed by Gilliéron and Chometon (1999) for a slant that optimized RANS turbulent models for predicting
angle of 35◦ with an unstructured mesh, while Kapadia the separation onset fail in the massively separated flow
et al. (2003) and Krajnović and Davidson (2004) have regions. For instance, even though the k - ω simulation
considered a slant angle of 25◦. As a more recent review of Han (1989) used good quality grids, an accurate nu-
of the subject, see Krajnović and Davidson (2005a,b). merical scheme and proved grid convergence, a drag
B. The RANS approach coefficient 30 % higher than the experimental one was
The RANS equations determine mean flow quantities predicted (Wilcox, 1998). From the numerical results,
but they require turbulence models to close the system, Krajnović and Davidson (2004) conclude that the wide
i.e. an equal number of equations and unknowns. There range of the turbulent scales around the rear end which,
are many closure models proposed (Wilcox, 1998) but together with very unsteady reattachments in this re-
unfortunately it is very difficult to find one that can ac- gion, are possible reasons for failures.
curately represent the Reynolds stresses in the detached C. The LES approach
flow regions of bluff bodies, where the flow regime is In LES only the larger unsteady turbulent motions are
strongly dependent on the geometry, specially at the computed while the effect of the small (unresolved)
rear part, and they often lead to complex flow patterns. scales on the large (resolved) ones is modeled. The lar-
A deficiency of the most used closure models for RANS ger scales contain most of the kinetic energy of the flow
schemes is their inherent inability to deal with mas- and they are affected by the geometry and forcing.
sively separated flows containing many coherent struc- Thus, they are in general highly inhomogeneous and
tures. Thus, the mean pressure could not be accurately irregular, making their flow description most relevant
296
 G. FRANCK, N. NIGRO, M. STORTI, J. D’ELÍA

from an engineering perspective. The smallest scales of account for anisotropy.

fluid motion are considered “universal”, in the sense From a practical point of view, most of the LES
that the dynamics of the small scales is independent schemes use structured grids but in these the unknown
from the flow geometry and forcing, and solely con- numbers easily become too large in wall-bounded flows
trolled by energy transfers due to scale interactions (Ca- (Krajnović and Davidson, 2004). As the mesh points
lo, 2005). Most of the computational effort in Direct should be concentrated where they are needed, e.g. on
Numerical Simulation (DNS) is spent on solving the boundary layers and separated flow regions, some strat-
smallest dissipative scale, while LES models focus on egies are proposed. On structured meshes, for instance,
their global effect. The computational efficiency of LES this was achieved combining structured multi-block
versus DNS arises as trade-off with accuracy since such using O and C grids (Krajnović and Davidson, 2004),
a modeling reduces the accuracy of the method. In con- but other options can be proposed with the same aim. A
trast to RANS, it can be shown that LES procedures recent work is the Krajnović and Davidson (2004) one,
generally converge to DNS (McDonough, 1995) and where a LES-Smagorinsky model for the SGS stresses
thus their computed solutions can be expected to con- was used with a three dimensional Finite Volume Me-
verge to Navier-Stokes solutions, as space and time step thod (FVM) on a fairly structured mesh and a slant an-
sizes are refined. gle of 25º.
On the other hand, many theoretical problems in deal- The main objectives of the present work are: (i) to
ing with the boundary conditions for LES can be identi- perform large eddy simulations with unstructured mesh-
fied (Sagaut, 2001). From a mathematical point of view, es and finite elements of the unsteady flow around the
two problems dealing with boundary conditions for LES same model but for a slant angle of 12.5◦ (a rear end
are: (i) incidental interaction between spatial filter and variant) using mixed meshes, i.e. unstructured meshes
exact boundary conditions; and (ii) the specification of plus several layers of prismatic elements close to the
boundary conditions leading to a well-posed problem. A boundary layers; (ii) an assessment of the LES turbu-
common approach consists in considering that the filter lence model combined with a SUPG-PSPG stabilization
width decreases nearby the boundaries so that the inter- method to predict the main flow dynamic structures that
action terms cancels out. Then, the basic filtered equa- are involved in vehicle aerodynamics without a deep
tions are left unchanged and the remaining problem is to analysis of sensitivity of the results with the turbulence;
define classical boundary conditions for the filtered (iii) a visualization of the time-averaged flow in the
field. Another option is the use of “embedded” bound- near-wake and determination of coherent macro struc-
ary conditions which consists of filtering through the tures (i.e. high space localized and time-persisting vor-
boundary leading to a layer along the boundary whose tex flows) by means of the second invariant of the gra-
width is order of the filter cutoff length scale, but some dient velocity or Q-criterion (Dubief and Delcayre,
additional source terms must be computed or modeled. 2000); and (iv) a topological validation of the number
From a physical point of view, the boundary conditions and type of the singular points (where velocity vanishes)
represent the whole flow domain beyond the computa- on the symmetrical vertical section.
tional one and they must be applied to all scales (to all II. GEOMETRICAL PARAMETERS
space-time modes). The Ahmed model has a length L = 1044 mm, which is
There are some problems with the boundary condi- approximately a quarter of a typical passenger car
tions at solid walls and inflow sections which are more length. The height H and the width B are defined ac-
closely related to LES while the corresponding ones at cording to the ratio (L:B:H) = (3.36:1.37:1). It has three
the outflow section are not specific to this method (Sa- main geometrical sectors: the front one, with boundaries
gaut, 2001). In the case of inflow conditions, several rounded by elliptical arcs to induce an attached flow, a
strategies are proposed (Sagaut, 2001), for instance: (i) middle sector which is a box shaped sharp body with a
“stochastic reconstructions” from statistical one-point rectangular cross section and, finally, a rear end sector.
descriptions; (ii) “deterministic computations”, as the A set of interchangeable rear ends with sharp edges
“precursor simulation”, where a first computation is were tested, where the slant length is kept fixed to ls =
performed for the attached boundary layer flow without 222 mm. The 12.5º slant angle was analyzed here.
the body and the data from some extraction plane are Figure 2 shows the dimensions of the computational
used as inlet boundary condition; and (iii) “semi- flow domain relative to L. A Cartesian coordinate sys-
deterministic reconstructions”: they propose an inter- tem O(x,y,z) is employed whose origin is placed on the
mediate approach between the previous ones to recover
ground at the front sector. The body is suspended to zbase
two-point correlations of the inflow with no preliminary
=0.048L from the wind tunnel ground. The parallelepi-
computations. Also, mixed strategies have been de-
pedic domain has 10L× 2L× 1.5L in the streamwise y,
signed as the hybrid RANS-LES approaches (e.g. “zon-
al decompositions”, “nonlinear disturbance” equations spanwise x and stream-normal (vertical) z Cartesian
and “universal modeling” (Sagaut, 2001)), which are directions respectively. The inlet flow section is placed
employed to decrease the rather large computational 2.45 L upstream of the model front while the outlet flow
cost of LES due to (i) the goal of directly capturing all section is placed 6.6 L downstream from the model rear
the scales motions of turbulence production and (ii) an end.
observed inability of most subgrid models to correctly
297
Latin American Applied Research 39:295-306(2009)

Figure 2: Dimensions of the computational flow domain as function of the body length L.
III. FINITE ELEMENT DISCRETIZATION
A. Mesh generation
The meshing process is performed with the Meshsuite
mesh generator (Calvo, 1997). It involves a basic tetra-
hedral generation and the addition of layers of wedge
elements for a better resolution close to the body sur-
face. The meshing is done following four stages:
1. Lines meshing: a subdivision of the lines defining
the boundaries (wire model) using CAD resources;
2. Surfaces meshing: they are generated from the
geometrical definition of the surface boundaries and
a refinement function implicitly defined by the spac-
ing of the one-dimensional mesh and, if necessary,
Figure 3: Unstructured finite element mesh for the computa-
some interior surface points. The bluff body and its tional flow domain around the Ahmed model.
wake are surrounded by an auxiliary surface which
is also meshed; Table 2: Length scales of Taylor ηT = 265.68× 10-5L and Kol-
3. Volume meshing: the auxiliary surface defines a mogorov ηK =1.27×10-5L, relative to the averaged mesh size
hm. The normal element spacing of the first wedge-layer on the
sub-domain which is used for better user-control on
body surface are hm ≈ {63.50, 31.75}×10-5L
the volumetric meshing, where refinement in zones
mesh # 1 mesh # 2
close to the body surface and wake is chosen.
4. Wedge elements addition: wedge elements (pris- ηT/hm 4.18 8.37
matic elements of triangular basis) are added to the ηK/hm 0.02 0.04
tetrahedral mesh in order to obtain better simulations
of boundary layers without excessive refinements in rahedral mesh with 955 k-elements and 198 k-nodes.
the remaining flow domain. Since the final mesh has The mesh generator eliminates major degeneracies
both wedge and tetrahedral elements, special consid- with exception of slivers and cups and then, as a mesh
eration is needed for parallel implementation. quality assessment, is sufficient to measure how near is
A general view of one of the meshes for the whole a node from the opposite face. This is computed through
computational flow domain is given in Fig. 3. the quality factor γ e = hmin
e e
/ d min , for e =1, 2, ..., E,
e e
B. Some parameters of the combined wedge- where hmin and d min are the minimum height and edge
tetrahedral meshes length, respectively, on each element. In the present
For a better comparison, it is assumed that each wedge case, the quality factors, without smoothing, are
is nearly equivalent to 2.5 tetrahedrons. Two combined
wedge-tetrahedral meshes are employed and they have
γ min =0.02 and γ max =0.95 which are taken as accept-
the following features: able.
a) mesh # 1: it has 151 k-wedges close to the body sur- C. Turbulent scales and mesh spacing
face, 450 k-tetrahedrons, and 170 k-nodes (83 k- The Kolmogorov length scale ηK is the length scale for
nodes on the wedge part plus 87 k-nodes on the tet- the smallest turbulent motions, while the Taylor one ηT
rahedral part), which is equivalent to a pure tetrahe- gives more emphasis to intermediate turbulent motions
dral mesh with 828 k-elements and 170 k-nodes; whose length scales are near some integral length scale
b) mesh # 2: it has 202 k-wedges close to the body sur- of the mean flow. Empirical correlations for both are
face, 450 k-tetrahedrons, and 198 k-nodes (111 k- ~ ~
given by ηK = A−1/ 4 Re−L~3/ 4 L and ηT =151/ 2 A−1/ 2 Re−L~1/ 2 L ,
nodes on the wedge part plus 87 k-nodes on the tet-
rahedral part), which is equivalent to a pure tet- respectively, e.g. see Howard and Pourquie (2002),

298
 G. FRANCK, N. NIGRO, M. STORTI, J. D’ELÍA

~ + νt, which is the sum of the molecular part calculated

where A is an O(1) constant, L is some integral scale
~ as ν = µ/ρ, plus a “turbulent” one νt. The last one is es-
and ReL~ = U∞ L / v is the corresponding Reynolds num-
timated in the PETSc-FEM (Sonzogni et al., 2002) code
ber, where ν is the kinematic viscosity of the fluid and by means of the Smagorinsky sub-grid model coupled
U∞ the non-perturbed speed. In the present case, the with a modified van Driest near-wall damping factor
~ ~
following values are adopted: A =0.5, L ≡ L = 1.044 m fv , and given by
(i.e., the body length), while the remaining values are vt = CS2he2 fv ε (u) : ε (u);
defined in Sec. D. Then, ReL~ = 4.25×106 ,
fv =1− exp(− y+ / A+ ); (1)
ηT ≈ 2.66 ×10−3 L and ηK ≈ 1.27 ×10−5 L . On the other ~
hand, the normal element spacing of the first wedge- fv = fv H(d − y + );
layer (on the body surface) are in which CS is the Smagorinsky constant (CS ≈ 0.10 for
hm ≈ {6.350,3.175}×10−4 L . Therefore, close to the body flows in ducts) and ε (u) : ε (u) is the trace of the strain
surface, the mesh spacing normal to the body surface rate ε (u) , while H (d − y + ) is the Heavside function and
are smaller than the Taylor length scale but bigger than d is a certain threshold distance from the body surface.
the Kolmogorov one (see Table 2). It can be remarked A constant value A+=25 is adopted, while y+= y/yw is the
that the turbulent eddies down to the Taylor length scale non-dimensional distance from the nearest wall ex-
size should be well modeled. pressed in wall units yw= ν/uτ, where uτ=(τw/ρ)1/2 is the
D. Fluid flow parameters and boundary conditions local friction speed and τw is the local wall shear stress.
A Newtonian viscous fluid model is adopted, with non- The van Driest near-wall damping factor fν reduces the
perturbed speed U∞ = 60m/s and kinematic viscosity “turbulent” kinematic viscosity close to the solid walls,
but it introduces a non-local effect in the sense that the
ν=14.75×10-6 m2/s (air at 15 C). The Reynolds number
“turbulent” kinematic viscosity νt inside an element also
(based on the model length L) is Re = 4.25×106. At the
depends on the state of the fluid at the closest wall. It is
inflow section a uniform axial velocity profile U∞ is
a near-wall modification of the Prandtl’s mixing length
imposed, while a non-slip boundary condition at the turbulence model and was found to be useful in attached
ground is prescribed which, in turn, is moving with the flows. It is written in terms of y+ so that in regions with
same stationary velocity U∞ . These boundary condi- significant wall friction the factor is relevant only in a
tions emulate road tests or wind tunnel tests with a thin layer near the body skin. On the other hand, in re-
boundary layer control and they have been employed in gions where the local friction speed uτ takes very low or
many related studies (Han, 1989; Krajnović and David- null values (usually in the separation and reattachment
son, 2004; Basara and Alajbegovic, 1998) (wind tunnels regions) the influence of the damping factor fν is signifi-
have several mechanisms to ensure that the boundary cant at large distance from the body. In order to prevent
layers are well attached to the tunnel walls, e.g. by suc- this drawback, it is further restricted to act within the
tion techniques). Under these conditions the equilibrium ~
threshold distance d, giving the modified one fv .
flow is uniform flow, and no boundary layers are devel-
oped. Thus, the vehicle model is assumed at rest and the B. Parallel computing
boundary conditions are: (i) uniform flow at inlet U∞ ; The numerical simulations were performed using a do-
(ii) no-slip at the ground which in turn is moving with main decomposition technique (Paz and Storti, 2005) in
the PETSc-FEM (2008) code, which is a parallel mul-
velocity u= U∞ ; (iii) null pressure at outlet; (iv) no-slip
tiphysics finite element library based on the Message
at the body surface (u=o) and (v) slip at the lateral and Passing Interface MPI (2007) and PETSc-2.3.3 (2007),
upper sides. and used in several contexts, for instance, for modeling
IV. STABILIZED FINITE ELEMENTS free surface flows (Battaglia et al., 2006; D’Elía et al.,
The finite element method is used to solve the momen- 2000), added mass computations (Storti and D’Elía,
tum and continuity equations for the velocity and pres- 2004) or inertial waves in closed flow domains (D’Elía
sure at each node and at each time step. For simplicity, et al., 2006). The present problem was solved using the
equal order spatial discretization for pressure and veloc- Beowulf Geronimo cluster (2005), with 11 P4 nodes
ity is highly attractive, e.g. linear tetrahedral or pris- and 2 GBytes of RAM memory each.
matic elements, but it is well known that this kind of V. EXPERIMENTAL NEAR-WAKE FLOW
basis functions needs to be stabilized to achieve stable Even though the wake flow of any bluff body is basi-
solutions. This is performed with a SUPGPSPG scheme cally unsteady, the time-averaged flow exhibits some
whose implementation details are given in Garibaldi et coherent macro structures (i.e. high space localized and
al. (2008). time-persisting vortex flows) that appear to govern the
A. Large Eddy Simulation drag generated at the rear end by massive flow separa-
In LES techniques, the momentum balance equations tions. The experimental information obtained collecting
are solved with an “effective” kinematic viscosity νe = ν several wind tunnel tests on this model (Ahmed et al.,
1984; Sims-Williams, 1998; Bayraktar et al., 2001;
299
Latin American Applied Research 39:295-306(2009)

Spohn and Gilliéron, 2002; Lienhart and Becker, 2003)

show two critical slant angles: αm≈12.5º and αM≈30.0º,
where the main topological structure of the time-
averaged flow is summarized in Table 1. These experi-
mental tests show the following overall behavior of the
time-averaged flow as a function of the slant angle:
1. When the slant angle is lower than the first critical
one (α<αm), see Fig. 4, I and II, the time-averaged
flow in the near-wake is almost two-dimensional and
it remains attached on the top and slanted surfaces. It
only separates past the vertical rear surface. In the
near-wake there are two re-circulatory time-
averaged sub-flows A and B located one over the
other (see sketch in Fig. 5). The shear layer coming Figure 4: Sketch of the time-averaged flow in the near-wake
off the slant side edge rolls up into a longitudinal as a function of the slant angle α, being I y II for α<αm, III for
vortex. At the top and bottom edges of the vertical αm <α<αM, and IV for α>αM.
surface, the shear layers roll up into the two re-
circulatory time-averaged sub-flows A and B. Oil-
flow pictures on the vertical surface suggest that
these re-circulatory time-averaged sub-flows are not
bounded but they extend downstream. Therefore,
they can be thought as generated through two horse-
shoe vortices placed one over the other, inside the
separation bubble D. The bound legs of these vor-
tices are almost parallel to the vertical surface while
the trailing legs of the upper vortex A align them-
selves in the direction of the onset flow and merge
with the vortex C coming off the slant side edges.
The streamlines of the toroidal vortices on the verti-
cal surface produce the singular point N (see sketch
in Fig. 5) and when the slant angle increases, the
singular point N moves downward;
2. When the slant angle is nearly equal to the first criti- Figure 5: Sketch of vortices and bubble of the time-averaged
cal one (α = αm), another bubble appears on the up- flow in the near-wake of the Ahmed model.
per slanted surface, while the streamlines over the
slanted surface are more aligned with the slant side relatively high concentration of vorticity intensity ω, so
edges enforcing the vortex C sketched in Fig. 5; they typically show a spiraled motion and they persist
3. When the slant angle is between both critical ones for a time scale size τc longer than the typical eddy time
(αm <α<αM ), see Fig. 4 III, the time-averaged flow turnover ω-1. Previous studies of coherent structures by
in the near-wake quickly shows a strong three di- means of the pressure isosurfaces suggest that a pressure
mensional behavior, where the streamlines inside the criteria can be better than a vorticity one (Comte et al.,
first bubble are fed by a centered rotation motion 1998). Nevertheless, the threshold value for the pressure
and they result in a helical motion, while the bubble isosurfaces shows a strong dependence on the mean
boundary increases when the slant angle increases, pressure around the coherent structures and it can fail in
and the side vortex C is stronger and the flow de- a flow subregion with a high vortex concentration. The
taches faster, while a separation point is formed be- so-called Q-criterion (Hunt et al., 1988) is another re-
tween the flows close to the slanted and vertical sur- source for its study. It is based on both pressure and
faces; vorticity intensities and it can be defined as Q=(ΩijΩij -
4. When the slant angle is greater than the second criti- SijSij)/2, with Ωij=(ui,j- ui,j)/2 and Sij =(ui,j+ui,j)/2, are the
cal one (α> αM), see Fig. 4 IV, there is a flow sepa- symmetrical and skew-symmetrical components of the
ration on the slanted surface and the singular point N velocity gradient ∇u respectively. It can be shown that
moves upwards, while the flow is again almost two- the scalar Q is the second invariant of the velocity gra-
dimensional. dient ∇u and that it measures the ratio between rotation
and deformation rates. Thus, isosurfaces with Q>0 show
VI. THE Q-CRITERION regions where rotation rates are bigger than the strain
Coherent structures in a turbulent flow are sub-flows ones, so they can develop coherent structures. Since
with some “cohesive” character, where the turbulence is ω2=2Ω2, another equivalent definition is Q=(ω2-
“self-reorganized” in such a way that the resulting mean 2SijSij)/4, showing that the coherent structures have a
motion is rather more regular than chaotic. They have a relatively high concentration of vorticity intensity ω.

300
 G. FRANCK, N. NIGRO, M. STORTI, J. D’ELÍA

VII. NUMERICAL RESULTS (Krajnović and Davidson, 2004). The symmetry of the
A. Instantaneous flow in the near-wake time-averaged flow around the vertical symmetrical
A view of the instantaneous flow on the front, top, one plane was used as another indicator of the convergence
lateral side, and partially over the slant surface is shown in time of the average flow. The visualizations shown in
in the upper plot of Fig. 6, where the flow structure is Figs. 7-11 are all computed through a time-averaged
highlighted through the isosurfaces of the Q-criterion procedure of the large eddy simulations of the unsteady
using a threshold of 40 Hz2, while at the bottom an ex- flow, where the averaged-time must be big enough to
perimental observation in a tunnel is included (Beau- prevent spurious effects. In this case, the averaged-time
doin et al., 2003). The flow pattern is very sensitive to is chosen as T=6TL where TL=L/U∞ is the elapsed time
the curvature radii and to the Reynolds number. There required by a fluid particle to travel one model length.
are detachments on the curved surfaces at the front part The visualizations on the body surface are performed by
as well as on the intersection between front and top means of trace lines (or friction ones) of the time-
middle surfaces. The transversal Kelvin-Helmholtz vor- averaged flow. The detachments and re-attachments of
tices are convected downstream from the front side and the time-averaged flow on the body surface are ex-
converted into the hairpin vortices λ, see Fig. 6 (top). tracted using the Q-criterion. From these numerical re-
The flow separates at the two tilted edges, between the sults the following features can be observed:
slanted surface and the lateral ones producing two large
counter-rotating cone-like trailing vortices. The same
figure shows another flow separation on the sharp edge
between the body roof and the slant surface, where vor-
tex parallel to the separation line is formed. As it may
be observed there is good agreement between numerical
and experimental results.
B. Time-averaged flow in the near-wake
The time used for the averaging of the instantaneous
flow must be long enough to produce a steady flow.
This situation is reached after several hundred of time
steps because of stability and accuracy requirements

Figure 7: Stream-ribbons of the toroidal horseshoes vortex

generated in the separation bubble of the time-averaged flow.

Figure 8: Trace lines of the time-averaged flow showing the

singular point N.

Figure 6: Transversal Kelvin-Helmholtz vortices are down-

stream convected from the front side and converted into hair-
pin λ vortices: present computation (top) and experimental of Figure 9: A bubble on the slanted surface of the time-averaged
Beaudoin et al. (bottom). flow and some singular points.
301
Latin American Applied Research 39:295-306(2009)

1. In the near-wake there are two large re-circulatory

sub-flows A and B situated one over the other which
agree with the experimental ones (Ahmed et al.,
1984; Gilliéron and Chometon, 1999), see stream-
ribbons in Fig. 7. The vortex cores are obtained
through a phase plane technique (Kenwright, 1998);
2. The streamlines of the toroidal vortices over the ver-
tical surface produce the singular point N shown in
Fig. 8;
3. There is a bubble on the slanted surface shown in
Fig. 9, where S10,S11 are half-saddle points while N6 Figure 10: Trace lines of the time-averaged flow on the body
surface. The flow run left to right.
is a nucleus point;
4. Flow detachments occur on the sharp edges of the
model, see Fig. 10. There are three flow detachments
on the frontal part: one on the upper side (T) and two
more on both lateral sides (L). On the same figure,
on both lateral sides, there is a separation zone Z on
the top side which encloses the triple border point
formed by the front, upper and lateral sides, with a
small flow channel due to a high adverse pressure
gradient. Also, there is a separation zone L on the
middle lateral side;
5. Fig. 11 shows trace lines (or friction ones) of the
time-averaged flow showing (i) on the lateral surface
(top-left): a center C, a saddle Ss, an unstable focus
D and two separation lines; (ii) on the slant surface
(top-right): a saddle, a stable focus and a bubble; and Figure 11: Trace lines of the time-averaged flow showing
(iii) on the top surface (bottom): a saddle, a stable singular points on the lateral surface (top-left) and on the slant
focus, an unstable focus and two separation lines. surface (top-right).
C. Topological validation at the symmetrical flow Table 3: A taxonomy for singular points P on a cut plane in a
section three dimensional flow (Chong et al., 1990).
A topological validation of the time-averaged flow vi- point type z=eig(∇u)|P={z1,z2}
sualizations is made on the symmetrical section (x=0) in zi= ai=+i bi, for i=1,2
order to assess if all the main flow structures in the repulsion a 1, a 2 > 0
symmetrical section have been identified. attraction a 1, a 2 < 0
center a1 = a2 > 0
As it is known (Chong et al., 1990), a singular point
saddle S a1·a2 < 0
P in a cut plane ξ, η in a three dimensional flow is a half-saddle when S is at a boundary
point where velocity vanishes. For its classification, focus b1, b2 ≠ 0 (rotation)
Chong et al. (1990) scheme is adopted here, which is node b1, b2 = 0 (no-rotation)
based on the eigenvalues z1 = a1 +ib1 and z2 = a2 +ib2 of
the Jacobian matrix H = ∂(u,v)/∂(ξ,η) evaluated at the
singular point P, where (u,v) are the flow velocities on
the cut plane. The resulting taxonomy is summarized in
Table 3. It can be noted that in a focus point there is a
flow rotation around it as opposed to a node case. Fig-
ure 12 shows streamlines as well as singular points of
the time-averaged flow that were obtained in the sym-
metrical section (x=0), where S0 to S13 are half saddle
points, N1 to N7 are focus points (in the vortex nuclei)
and D is another saddle point. Figures 13 and 14 show

Figure 12: Streamline sketch in a time-averaged sense ob-

tained from the present numerical simulations on the (vertical)
symmetry plane x = 0 showing singular points.

302
 G. FRANCK, N. NIGRO, M. STORTI, J. D’ELÍA

Figure 13: Trace lines of the time-averaged flow on the front

part at the upper side showing some singular points.

Figure 16: Tiny vortices of the time-averaged flow on the

vertical base surface of the rear end at the upper side B,N3.

Figure 14: Trace lines of the time-averaged flow on the front

part at the bottom side showing some singular points.

the streamlines projected on the ZY -plane at x=0. There

are half-saddle points S1,S2, a focus one N1 on the top
side and the corresponding S3, S4 and N4 on the bottom
side. Figure 15 shows the two toroidal vortices F1 (up
per) and F2 (bottom), whose nucleus are N4 and N5, re-
spectively. Figures 16 and 17 show four tiny and small
vortices near the vertical surface at the rear end. Two
vortices are on the upper region and the remaining two
ones on the inferior one. There are vortices whose nu-
Figure 17: Tiny vortices of the time-averaged flow on the
cleus B, N2 and B, N3 rotate in a clockwise and counter vertical base surface of the rear end at the bottom side B,N2.
clockwise fashion respectively, and they also have half-
 1   1 ;
saddle points. Near the F2 vortex a stagnation point P is T = ∑+ ∑  − ∑+ ∑  (2)
identified. A bubble formed behind the upper edge of  N 2 N*   S 2 S* 
the slanted surface is shown in Fig. 9. On the other
hand, on the symmetrical vertical section, the topologi-
while ∑ N
is the sum of nodes and focus points, ∑ N*

cal relation T=1-n must be verified (Hunt et al., 1988),

where
is the sum of the half-node points, ∑ S
is the sum of
saddle points, ∑ S*
is the sum of half saddle points, and
n is the connectivity of the two-dimensional flow sec-
tion. For instance, the connectivity of any two-
dimensional section in a simply connected three dimen-
sional flow region is n=1 (e.g. a flow without a body)
while for the flow around a single body, as in the pre-
sent case, is n=2. From the present visualizations, the
following values are obtained: ∑
= 7,
N
=0,∑ N*
=1 ∑ S

and ∑S*=14. Therefore T≡1, i.e. the topological restric-

tion is satisfied. It can be concluded that all the main
flow structures in the symmetrical section have been
Figure 15: Streamlines of the time-averaged flow projected identified. These topological features of the time-
onto the vertical symmetry plane x = 0 showing some singular average flow are found independently of the averaging
points. time T and grid-size of both meshes.

303
Latin American Applied Research 39:295-306(2009)

Ahmed vehicle model were performed for a slant angle

of 12.5º. Two meshes were employed, showing mesh
convergence through the drag coefficient (see Fig. 19).
The symmetry of the time-averaged flow was used as
another indicator of the solution convergence. Flow
visualizations of the time-averaged flow in the near-
wake qualitatively agree with the experimental observa-
tions. Most of the features of the flow around this bluff
model in ground proximity were predicted, such as the
formation of trailing vortices, massive separation and
re-circulatory flows. The turbulent eddies up to the Tay-
lor length scale were well resolved. The time-averaged
flow was dominated by large coherent macro structures
Figure 18: Pressure coefficient Cp as a function of the y-
formed by massive flow separation. The mean drag co-
streamwise coordinate at the top and bottom body surface on
efficient CD agrees quite well the experimental one, with
the symmetry plane for a slant angle α=12.5º.
a percentual relative error of +6 % for the coarser mesh
and -1 % for the finer one. It is found that the topologi-
cal features of the time-averaged flow are independent
of the averaging time T and grid-size. It is concluded
that according to the current computational resources
LES may be used as a feasible turbulence model for real
vehicles aerodynamics. Future modeling efforts would
be focused on the representations of the structure of
turbulence and some average profiles, as velocities and
turbulent stresses, which could be useful for RANS
studies.
ACKNOWLEDGMENTS
This work was performed with the Free Software Foun-
dation GNU-Project resources as Linux OS and Octave,
Figure 19: Pressure coefficient CD: time-evolutions, mean as well as other Open Source resources as PETSc,
values and experimental measured (Ahmet et al., 1984), for a
MPICH and OpenDX, and supported through Consejo
slant angle α=12.5º.
Nacional de Investigaciones Científicas y Técnicas
D. Drag and pressure coefficients (CONICET, Argentina, grants PIP-02552/2000, PIP
The drag coefficient is computed as CD = Dp/γ, with 5271/05) Universidad Nacional del Litoral (UNL, Ar-
γ = ρU∞2 Aproy / 2 , where Dp is the drag force, Aproy is the gentina, grant CAI+D 2005-10-64) and Agencia Na-
cional de Promoción Científica y Tecnológica (ANP-
frontal surface area, ρ is the fluid density and U∞ is the CyT, Argentina, grants PICT 12-14573/2003, PME
non-perturbed speed. The pressure coefficient is defined
( )
209/2003, PICT 1506/2006). N. Calvo has participated
as Cp = 2( p − p∞ ) / ρU∞2 , where p∞ is the undisturbed in fruitful discussions on mesh quality and mesh genera-
pressure at the inlet boundary. The pressure coefficient tion.
for a slant angle α = 12.5º is plotted in Fig. 18 as a func- REFERENCES
tion of the streamwise coordinate at the top and bottom Ahmed, S., G. Ramm and G. Faltin, “Some salient fea-
body surface on the symmetry plane. In Fig. 19 the un- tures of the time-averaged ground vehicle wake,”
steady nature of the drag coefficient CDnum(t) is shown SAE Report, 840300, 1–31 (1984).
for the coarser mesh (#1) and the finer one (#2). In eachBarlow, J., R. Guterres, R. Ranzenbach and J. Williams,
case, oscillations in its values are observed during star- “Wake structures of rectangular bodies with radi-
tup but, after some time-steps, these oscillations become used edges near a plane surface,” SAE Congress,
small and CDnum(t) approaches a constant value. The 1999-01-0648 (1999).
Basara, B., “Numerical simulation of turbulent wakes
mean drag coefficient measured in the wind tunnel tests
around a vehicle,” ASME Fluid Engineering Divi-
(Ahmed et al., 1984) for a slant angle of 12.5º was
sion Summer Meeting, FEDSM99-7324, San Fran-
CDexp = 0.230 while the numerical are CD=0.246 for the cisco, USA (1999).
coarser mesh, and CD=0.228 for the finer one, with a Basara, B. and A. Alajbegovic, “Steady state calcula-
percentage relative error of εr%≈ +6% and εr%≈-1%, tions of turbulent flow around Morel body,” 7th Int.
respectively. Symp. of Flow Modelling and Turbulence Measure-
ments, Taiwan, 1–8 (1998).
VIII. CONCLUSIONS
Basara, B., V. Przulj and P. Tibaut, “On the calculation
Large Eddy Simulations (LES) with unstructured mesh-
of external aerodynamics industrial benchmarks,”
es and finite elements of the unsteady flow around the
SAE Conf., 01–0701 (2001).
304
 G. FRANCK, N. NIGRO, M. STORTI, J. D’ELÍA

Battaglia, L., J. D’Elía, M. Storti and N.M. Nigro, Howard, R. and M. Pourquie, “Large eddy simulation of
“Numerical simulation of transient free surface an Ahmed reference model,” Journal of Turbulence,
flows,” ASME-Journal of Applied Mechanics, 73, 3, 12 (2002).
1017– 1025 (2006). Hucho, W., Aerodynamics of road vehicles. Society of
Bayraktar, D., D. Landman and O. Baysal, “Experimen- Automative Engineers (1998).
tal and computational investigation of Ahmed body Hunt, J., A. Wray and P. Moin, “Eddies, stream, and
for ground vehicle aerodynamics,” SAE Report, 01- convergence zones in turbulent flows,” Report
2742 (2001). CTRS88, Center For Turbulent Research (1988).
Beaudoin, J., K. Gosse, O. Cadot, P. Paranthoën, B. Janssen, L. and W. Hucho, “Aerodynamische Formop-
Hamelin, M. Tissier, J. Aider, D. Allano, I. Muta- tienierung der Typen VW -Golf und VW –
bazi, M. Gonzales and J. Wesfreid, “Using cavita- Scirocco,” Volkswagen Golf I ATZ, 77, 1–5 (1975).
tion technique to characterize the longitudinal vor- Kapadia, S., S. Roy and K. Wurtzler, “Detached eddy
tices of a simplified road vehicle,” Technical report, simulation over a reference Ahmed model car,” 41st
PSA Peugeot Citroën Directions de la Researche, Aerospace Sciences Meeting and Exhibit, AIAA-
France (2003). 2003-0857, 1–10 (2003).
Beowulf Geronimo cluster, http://www.cimec.org.ar/ Kenwright, D., “Automatic detection of open and closed
geronimo (2005). separation and attachment lines,” Proceedings of
Calo, V., Residual-based Multiscale Turbulence Model- IEEE Visualization (1998).
ing: finite volume simulations of bypass transition. Krajnović, S., Large-Eddy Simulation for computing the
Ph.D. thesis, Stanford University (2005). flow around vehicles. Ph.D. thesis, Chalmers Uni-
Calvo N.A., Meshsuite: 3d mesh generation, versity of Technology, Sweden (2002).
http://wwww.cimec. org.ar (1997). Krajnović, S. and L. Davidson, “Numerical study of the
Chong, M., A. Perry and B. Cantwell, “A general classi- flow around the bus-shaped body,” ASME-J. of Flu-
fication of three-dimensional flow field,” Phys. Flu- ids Engineering, 125, 500–509 (2003).
ids A, 2, 765–777 (1990). Krajnović, S. and L. Davidson, “Large-eddy simulation
Comte, P., J. Silvestrini and P. Bégou, “Streamwise of the flow around simplified car model,” SAE
vortices in large eddy simulation of mixing layers,” World Congress, Detroit, USA, 2004–01–0227
Eur. J. Mech, 17, 615–637 (1998). (2004).
D’Elía, J., M. Storti and S. Idelsohn, “A panel-Fourier Krajnović, S. and L. Davidson, “Flow around a simpli-
method for free surface methods,” ASME-Journal of fied car, part 1: Large eddy simulation,” Journal of
Fluids Engineering, 122, 309–317 (2000). Fluids Engineering, 127, 907–918 (2005a).
D’Elía, J., N. Nigro and M. Storti, “Numerical simula- Krajnović, S. and L. Davidson, “Flow around a simpli-
tions of axisymmetric inertial waves in a rotating fied car, part 2: Understanding the flow,” Journal of
sphere by finite elements,” Int. Journal of Computa- Fluids Engineering, 127, 919–928 (2005b).
tional Fluid Dynamics, 20, 673–685 (2006). Lienhart, H. and S. Becker, “Flow and turbulence struc-
Dubief, Y. and F. Delcayre, “On coherent-vortex identi- tures in the wake of a simplified car model (Ahmed
fication in turbulence,” J. Turbulence, 1, 1–22 model),” SAE Report 2003-01(0656) (2003).
(2000). McDonough, J., “On intrinsic errors in turbulence mod-
Duell, E. and A. George, “Experimental study of a els based on Reynolds-Averaged Navier-Stokes equ-
ground vehicle body unsteady near wake,” SAE Re- ations,” Int. J. Fluid Mech. Res., 22, 27–55 (1995).
port 1999-01-0812, 197–208 (1999). Message Passing Interface MPI. http://www.mpiforum.
ERCOFTAC, European Research Community on Flow, org/docs/docs.html (2007).
Turbulence and Combustion. http://www.ercoftac. Morel, T., “Aerodynamics drag of bluff body shapes.
org, (2007). characteristics of hatch-back cars,” SAE Report,
Garibaldi, J., M. Storti, L. Battaglia and J. D’Elía, 780267 (1978).
“Numerical simulations of the flow around a spin- Paz, R. and M. Storti, “An interface strip precondition
ning projectil in subsonic regime,” Latin American for domain decomposition methods. application to
Applied Research, 38, 241-248 (2008). hidrology,” Int. J. for Num. Meth. in Engng., 62,
Gilliéron, P. and F. Chometon, “Modeling of stationary 1873–1894 (2005).
three-dimensional separated air flows around an PETSc-2.3.3. http://www.mcs.anl.gov/petsc (2007).
Ahmed reference model,” ESAIM, 7, 173–182 PETSc-FEM. A general purpose, parallel, multi-physics
(1999). FEM program, GNU General Public License,
Gilliéron, P. and A. Spohn, “Flow separations generated http://www.cimec.ceride.gov.ar/petscfem (2008).
by a simplified geometry of an automotive vehicle,” Sagaut, P., Large eddy simulation for incompressible
Technical report, Technocentre Renault - CNRS, flows, an introduction. Springer, Berlin (2001).
Lab. d’ Etudes A´erodynamiques de Poitiers (2002). Sims-Williams, D., “Experimental investigation into
Han, T., “Computational analysis of three-dimensional unsteadinless and instability in passanger car aero-
turbulent flow around a bluff body in ground prox- dynamics,” SAE Report, 980391 (1998).
imity,” AIIA Journal, 27, 1213-1219 (1989). Sonzogni, V., A. Yommi, N. Nigro and M. Storti, “A
parallel finite element program on a Beowulf Clus-
305
Latin American Applied Research 39:295-306(2009)

ter,” Advances in Engineering Software, 33, 427– ing hemisphere at very-low and very-high frequen-
443 (2002). cies,” ASME Journal of Fluids Engineering, 126,
Spohn, A. and P. Gilliéron, “Flow separations generated 1048–1053 (2004).
by a simplified geometry of an automotive vehicle,” Wilcox, D., Turbulence Modeling for CFD. DCW, 2nd
IUTAM Symposium: unsteady separated flows. Tou- edition (1998).
louse, France (2002).
Storti, M.A. and J. D’Elía, “Added mass of an oscillat-

Received: February 3, 2008

Accepted: August 15, 2008
Recommended by Subject Editor: Eduardo Dvorkin

306