Professional Documents
Culture Documents
Abstract
This contribution presents some of the capabilities of the Computational
Fluid Dynamics (CFD) code XFlow, which uses a proprietary particle-based
kinetic solver based on the Lattice-Boltzmann Method. Using traditional
CFD software, industrial problems require time consuming meshing process
which often leads to errors or even divergence of the simulation. Due to its
particle-based and fully Lagrangian approach, the complexity of the geometry
surfaces is not a limiting factor in XFlow even in the presence of moving
parts, allowing to solve real industrial problems. The performance of XFlow
will be demonstrated for different industry benchmarks. The first example is
the Ahmed body which is a classical benchmark in the automotive industry.
The second benchmark presented will be the NASA trapezoidal wing. XFlow
results will be described and show good agreement with experimental data.
Keywords: Lattice-Boltzmann, Lagrangian, particle-based, Ahmed body,
NASA trapezoidal wing
1. Introduction
For the past 20 years, the field of Computational Fluid Dynamics (CFD)
has reached a high level of maturity, but it has only been recently that
CFD has been broadly applied to the improvement of several processes at
different stages: research, design, manufacturing, optimization, etc. The
need for robust and reliable analysis tools is therefore growing rapidly, in
proportion to the increasing complexity of simulations. To provide quick,
accurate feedback to realistic engineering problems is consequently essential
for companies to be competitive.
Preprint submitted to ECCOMAS 2012
b
1X
ni
b i=1
(2)
b
1X
ni ei
b i=1
(3)
(4)
where fi is the particle distribution function in the direction i, ei the corresponding discrete velocity and i the collision operator.
The stream-and-collide scheme of the LBM can be interpreted as a discrete approximation of the continuous Boltzmann equation. The streaming
or propagation step models the advection of the particle distribution functions along discrete directions, while most of the physical phenomena are
modeled by the collision operator which also has a strong impact on the
numerical stability of the scheme.
In the most common approach, a single-relaxation time (SRT) based on
the Bhatnagar-Gross-Krook (BGK) approximation is used
BGK
=
i
1 eq
(f fi ),
i
(5)
= wi
ei u u u
1+ 2 +
cs
2c2s
ei ei
c2s
!!
(7)
N
X
fi ekix eliy em
iz
(9)
(10)
xk y l z m =
N
X
i
t = 2f
S
Gd
g
(11)
(12)
(13)
(14)
where f = Cw x is the filter scale, S is the strain rate tensor of the resolved
scales and the constant Cw is typically 0.325.
A generalized law of the wall that takes into account for the effect of
adverse and favorable pressure gradients is used to model the boundary layer
[13]:
U1 + U2
u U1 up U2
U
=
=
+
uc
uc
uc u
uc up
w u
dpw /dx up
+ u
+ up
=
f1 y
f2 y
+
u2 uc
uc
|dpw /dx| uc
uc
uc y
+
y =
uc = u + up
u =
up =
(15)
(16)
(17)
(18)
|w | /
(19)
!
dp 1/3
w
.
dx
(20)
Here, y is the normal distance from the wall, u is the skin friction velocity,
w is the turbulent wall shear stress, dpw /dx is the wall pressure gradient, up
is a characteristic velocity of the adverse wall pressure gradient and U is the
mean velocity at a given distance from the wall. The interpolating functions
f1 and f2 given by Shih et al. [13] are depicted in figure 1.
25
45
f1 (y + u/uc )
40
20
f2 (y + up /uc )
35
30
15
f2
f1
25
10
20
15
10
5
100
101
y + u/uc
102
100
101
y + up /uc
102
Inlet velocity
Density
Dynamic viscosity
Car length
Reynolds number
Slant angles
Turbulence intensity
60 m/s
1 kg/m3
1.46014 105 Pa.s
1044 mm
4.29 106
0 ; 5 ; 10 ; 12.5 ; 15 ; 20 ; 25 ; 30 ; 40 degrees
0.5%
ulation time was two seconds and the time step t = 7.69231 105 s is
automatically estimated by XFlow to ensure the numerical stability.
3.2. Spatial discretization
Since XFlow is a particle based technology it does not require a timeconsuming meshing process. The preprocessor generates the initial octree
lattice structure based on the input geometries and the user-specified resolution for each geometry. The lattice may have several levels of detail which
are hierarchically arranged. Each level solves spatial and temporal scales
two times smaller than the previous level, thus forming the aforementioned
octree structure.
The lattice structure may be modified later by the solver if the computational domain changes (due to the presence of moving parts) or if the
resolution changes dynamically in order to adapt to the flow patterns (adaptive wake refinement). The adaptive wake refinement feature in XFlow is
based on the module of the vorticity field: in the lattice elements where
the vorticity reaches a threshold value the lattice is automatically refined.
Similarly, when the vorticity is lower than another threshold, eight adjacent
lattice elements are merged to form a coarser lattice element. This saves
computational resources and removes the need to refine your solution in advance. Consequently, as in illustrated figure 3, three resolutions are required
by the user: the far field, the wake and the near wall resolutions.
In order to select the best resolution near the walls and within the wake
that allows us to get good results in an acceptable time, a resolution dependency study is conducted before starting the validation of the Ahmed body.
This preliminary study consists in refining the resolutions and seeing how this
affects the accuracy of the results, but also checking if the code is converging
8
Figure 3: Example of lattice structure using the near wall and adaptive wake refinement
Table 2: Near walls and wake resolutions used in the resolution dependency study
Resolution (m)
# of Elements at t = 0.3 s
h
0.04
88,316
h/2
0.02
222,337
h/22
0.01
1,132,292
h/23
0.005
8,316,626
0.65 h
0.60
0.55
Drag Coefficient, Cx
0.50
0.45
0.40
h/2
0.35
0.30
0.25
0.200
h/22
1000
h/23
2000
3000
4000
5000
N (103 nodes)
6000
7000
8000
9000
Figure 4: Drag coefficient against the number of lattice nodes for different resolutions at
= 35
as soon as the wake resolution is the double or quadruple of the near wall
resolution affects the results quite dramatically. Hence, for all our runs, the
spatial discretization chosen for all the different slant angles is done with an
automatic wake refinement with a resolution of 0.08 m for the far field, and
0.01 m around the Ahmed body and within the wake.
3.3. Numerical results
The time required in XFlow to set up the case is about 10 minutes and
mainly consists in geometry importation, the flow and boundary specifications, and the resolution setup. The calculation time is almost the same for
all the slant angles and varies between 6 and 8 hours with the previously
selected resolutions on two Intel Xeon E5620 (2.4GHz).
The first result given by Ahmed is the curve representing the drag coefficient against the slant angle , and gives the drag contributions of every
part of the Ahmed body: the front Ck , the rear vertical surface Cb , the rear
slant surface Cs and the friction drag Cr . The total drag Ahmed found was
Cw and was the sum of the different contributions. Hence, the total drag
obtained from XFlow for the different slant angles is superimposed with the
Cw from Ahmed, as shown in figure 6.
From the figure 6 we observe a good overall drag prediction by the code:
the drag breakdown occurs right after 30 degrees and the minimum drag point
10
0.50
Wake 0.01m
Wake 0.02m
Wake 0.04m
Experimental
0.45
Drag Coefficient, Cx
0.40
0.35
0.30
0.25
0.200.0
0.1
0.2
Time (s)
0.3
0.4
0.5
is the critical angle 12.5 degrees, as measured by Ahmed. The absolute drag
values predicted by XFlow are accurate and the relative error varies from
only 0.4% to 3.2% for most of the angles, except around the drag breakdown
and at 0 degree angle where it reaches a maximum of only 7.1%. These small
discrepancies can be explained, on the one hand, by the complexity around
the flow around 30 degrees of slant angle which is switching from a massive
3D separation in the near-wake region to an almost 2D attached structure at
higher angles [15], and, on the other hand, by stronger gradients produced
by the rear of the car at 0 degree angle.
3.4. Flow field results
The second part of the results analysis is done by analyzing the main
recirculation structures resulting from the flow around the Ahmed body. For
this study, the averaging of the flow fields is required in order to filter the
temporal fluctuations and to identify the main structures of the turbulent
wake. The averaging of the fields started from t = 0.3 s when the flow was
established, as indicated for example by figure 5, to cut off the transient
period.
Ahmed provides pictures of the oil flow on the slanted surface for =
12.5, 25 and 30 degrees. It can be compared with XFlow which features
Line Integral Convolution (LIC) that approximates the surface streamlines
on a body. The figure 7 shows similar structure for the three angles: a quite
11
smooth and attached flow at 12.5 degrees, smooth flow patterns with two
small and symmetric fringes on the sides at 25 degrees, and two large and
symmetric separation bubbles at 30 degrees.
Ahmed also provides different velocity vectors plots in the symmetry
plane of the car, showing the near-wake region. This allows the study of
the separation bubble on the rear slant and within the wake for different
slant angles.
Figure 8 compares the near-wake region for a slant angle of 5 degrees
between the experimental results measured by Ahmed and results obtained
by XFlow at the same scale. This allows us to check the length of the bubble
separation located around the non-dimensional coordinate x/Lref = 0.375,
predicted in an extremely similar way in the two pictures. Two main eddy
structures are detected - highlighted in red boxes on figure 8 - which are
symmetrical from the top and bottom of the separation bubble. The code
tends to locate them slightly further downstream, though with reasonable
overall flow patterns.
The near-wake structure for a slant angle of 25 degrees also show good
similarities. This figure 8 shows an equivalent triangular separation bubble,
ending around the non-dimensional coordinate x/Lref = 0.2 for both cases.
4. NASA trapezoidal wing benchmark
The NASA trapezoidal wing benchmark comes from the 1st AIAA CFD
High Lift Prediction Workshop (HiLiftPW-1), sponsored by the Applied
Aerodynamics Technical Committee, which took place in June 2010 in Chicago,
IL. The challenge was to simulate a half aircraft configuration composed of
a body and a 3-element airfoil with a plane of symmetry as shown in figure
9 for a wide range of angles of attack. The trapezoidal wing is composed
of slat, main element and flap. The latter can be in two different configurations: Configuration 1 at 25 degrees and Configuration 8 at 20 degrees of
angle-of-attack.
The objectives of the benchmark are multiple [16]:
Assess the prediction capability of CFD codes in landing/taking-off
configuration,
Develop practical modeling guidelines for the analysis of high-lift configurations,
12
h
0.04
0.08
201,513
h/2
0.02
0.04
653,211
h/22
h/23
0.01
0.005
0.02
0.01
2,893,687 21,880,186
Table 4: Resolutions used for the 1st High Lift Prediction Workshop
Resolution 1
Resolution 2
Walls (m)
0.005
0.005
Wake (m)
0.01
0.02
Max. # of Particles
Angles
6
25 10
[-4 ; 32 ]
10 106
[34 ; 37 ]
15
[17] C. McGinley, L. Jenkins, R. Watson, A. Bertelrud, 3-d high-lift flowphysics experimenttransition measurements, AIAA Paper 5148 (2005)
2005.
17
18
Figure 7: Averaged Line Integral Convolution (LIC) on the slanted surface from Ahmed
(left) and XFlow (right)
19
20
0.40
XFlow
Experiment
Drag Coefficient, CD
0.38
0.36
h/2
0.34
h/23
0.32
h/22
0.300
15
10
Number of Lattice Nodes (106 )
20
25
Figure 10: Drag coefficient against the number of lattice nodes for different resolutions at
= 13
0.8
3.5
Experimental
Experimental Lower
Experimental Upper
XFlow
3.0
Lift Coefficient, CL
Drag Coefficient, CD
1.0
0.6
0.4
0.2
0.0 10
3.5
10
(deg)
20
30
0.0 10
40
10
10
(deg)
20
30
20
30
40
0.1
2.5
Pitching Moment, Cm
Lift Coefficient, CL
(b)
0
0.0
3.0
2.0
1.5
1.0
0.5
0.00.0
1.5
1.0
0.5
(a)
0
2.5
2.0
(c)
0.2
0.4
0.6
Drag Coefficient, CD
0.8
1.0
0.2
0.3
0.4
0.5
(d)
10
(deg)
40
Figure 11: Drag (a) and lift (b) coefficients against the angle of attack, the polar curve
(c), and the pitching moment coefficient (d)
21
22