Conference PaperPDF Available

An Extensive Validation of an Open Source Based Solution for Automobile External Aerodynamics

Authors:
Abstract
The number of computational uid dynamics (CFD) simulations
performed during the vehicle aerodynamic development process
continues to expand at a rapid rate. One key contributor to this trend
is the number of analytically based designed experiments performed
to support vehicle aerodynamic shape development. A second
contributor is the number of aerodynamic optimization studies
performed for vehicle exterior components such as mirrors,
underbody shields, spoilers, etc. A third contributor is the increasing
number of “what if” exploratory studies performed early in the design
process when the design is relatively uid. Licensing costs for
commercial CFD solutions can become a signicant constraint as the
number of simulations expands. A number of alternative products
(e.g., independently developed, supported and documented forks of
the popular open-source OpenFOAM® toolbox [1]) have become
available in recent years, offering a lower cost alternative to
traditional commercial CFD products. This paper summarizes results
from a broad and deep evaluation of the capability of iconCFD® to
substitute for the more traditional commercial CFD solutions
currently used to support vehicle aerodynamic development early in
the program development cycle. Included in this study were detailed
B-car, sedan, SUV and truck shapes as well as multiple variants of
each shape. The study investigated both static and moving ground
boundary conditions as well as alternative turbulence models. Trends
of the predicted aerodynamic drag coefcients (Cd) are compared
against experimental data. Both transient and steady state simulation
ranking accuracy of total vehicle Cd were found to be equivalent to
that historically observed with more traditional commercial solver
results. A consistent upward bias in absolute Cd values was observed
in the transient results.
Introduction
Aerodynamic loading on vehicles has a signicant inuence on fuel
economy. Wind tunnel testing and computational uid dynamics
simulations are widely used to evaluate the aerodynamic forces
exerted on a vehicle. There are two metrics by which the accuracy of
CFD solvers have traditionally been evaluated. The rst metric is
visualization—the ability to accurately predict and visualize surface
pressures, surface ow lines and ow eld structures for potential
insights into Cd reduction actions prior to test. CFD is now a widely
accepted tool for this type of pretest analysis. The second metric is
the absolute accuracy of the predicted aerodynamic coefcients.
Numerous validation studies have been performed by both academia
and automotive OEMs over the last two decades for the purpose of
determining if the predictive accuracy of CFD could consistently and
reliably replicate experimental results. Accuracy levels have
signicantly improved over the years as a function of increasing
model denition, mesh resolution, etc., however, absolute replication
of experimental force measurements remains somewhat challenging.
It should be noted that repeatability of wind tunnel measurements
between different facilities can also be an issue. In particular, test
section replication, small tire modications, etc. can have a
signicant effect on the measured forces.
Today’s aerodynamicist continues to depend on CFD to visualize
ow eld data for insight into potential Cd reduction actions. As
experimental aerodynamicists acknowledge that results can often be
tunnel dependent and CFD proponents acknowledge that certain
phenomena are extremely challenging to simulate, a third metric,
ranking accuracy, has become increasingly important. Today’s
aerodynamicists depend on CFD to correctly rank aerodynamic
coefcients across both multiple vehicle shapes and conguration
variations within a given shape.
An Extensive Validation of an Open Source Based Solution for
Automobile External Aerodynamics
2017-01-1524
Published 03/28/2017
Robert Lietz, Levon Larson, Peter Bachant, John Goldstein, Rafael Silveira, and Mehrdad
Shademan
Ford Motor Company
Pete Ireland and Kyle Mooney
ICON
CITATION: Lietz, R., Larson, L., Bachant, P., Goldstein, J. et al., "An Extensive Validation of an Open Source Based Solution for
Automobile External Aerodynamics," SAE Technical Paper 2017-01-1524, 2017, doi:10.4271/2017-01-1524.
Copyright © 2017 SAE International
Numerically driven designed experiments and optimization studies
are driving the number of CFD simulations performed to increase at a
rapid rate. As High Performance Computing (HPC) resources
continue to expand, the main constraint on increasing CFD
simulations becomes the cost and availability of commercial CFD
licenses. These cost and licensing constraints are driving an ever-
increasing interest in open source CFD solvers. SAE Technical Paper
2011-01-0163 [2] was a limited investigation into the capability of
open source CFD. That study reported that open source steady state
simulations based on the Reynolds Averaged Navier–Stokes (RANS)
approach did a good job at predicting Cd while transient simulations
(detached eddy simulation; DES) were required for more accurate lift
predictions. This paper expands on that material as it examines the
viability of open source CFD as an alternative to today’s commercial
CFD solvers. The evaluation metrics used in the study were Cd
ranking accuracy and ow visualization. Two questions were
investigated. First, can an open source, nite volume, transient CFD
code provide the same level of ranking accuracy as more traditional
commercial CFD solvers? In other words, can open source transient
analyses correctly rank not only different vehicle shapes but also
variations in the conguration of each shape? Second, is the ranking
accuracy of steady state analyses adequate enough to provide design
direction early in the vehicle development process? The relative
correctness of the ow structures predicted by the steady state
analyses is evaluated by comparison against those from the transient
analyses. The steady state simulations employed the k–ω SST
turbulence model, and the transient simulations employed the
Spalart–Allmaras detached eddy simulation model (SA-DES). Both
turbulence models are commonly recommended for ows with large
rotation, separation and reattachment. A total of 54 vehicle variants
(including B-cars, sedans, SUVs and trucks) were evaluated in this
study. Results from simulations were compared against results from
wind tunnel experiments. Static ground wind tunnel testing was
performed at the Ford Motor Company Dearborn Test Facility.
Moving ground wind tunnel testing was performed at the Windshear
facility in North Carolina. Images of the performance vehicle in the
test section of both facilities are shown in Figure 1.
Figure 1. Wind tunnel test facilities.
The structure of the paper is as follows. The Methods section of this
paper discusses numerical methods, geometry modeling, boundary
condition setup and the governing equations. Ranking accuracy and
ow visualization for the total data set (as well as specic subsets of
the data) are presented in the Results section of this paper.
Observations and potential next steps are discussed in the
Conclusion/Summary section of this paper.
Methods
There have been a growing number of validation studies published on
open source CFD in recent years. Most have been limited in scope to
a small number of either vehicle shapes or variants of a single shape.
This paper reports the results of a study conducted across a wide
range of vehicle shapes, conguration variants and boundary
conditions. All CFD analyses were performed using a single common
modeling and simulation best practice. A high level schematic of the
process is shown in Figure 2.
Figure 2. Schematic of the process involved in each simulation.
Numerical Method
Governing Equations
RANS Method
In the current study, the three-dimensional incompressible Reynolds
Averaged Navier-Stokes (RANS) equations were used for simulating
turbulent ow over different vehicles. The continuity and momentum
equations used are as follows:
(1)
(2)
In these equations ui, p, ρ and μ represent the velocity, pressure,
density and kinematic viscosity, respectively, and the overbar
represents a time-averaged quantity. In RANS in order to calculate
the shear stresses, the Boussinesq hypothesis (which relates the
stresses to the mean velocity gradients) is used, i.e.,
(3)
where μt is the turbulent viscosity, k is the turbulent kinetic energy
and ϵ is the dissipation rate. Among the RANS turbulence models
available, the Spalart-Allmaras, k–ϵ and k–ω use the Boussinesq
hypothesis. These approaches are employed when low computational
cost is required.
For the RANS portion of this study, Menter’s k–ω SST model was
used [3]. The k–ω SST model is designed in a way to activate the
standard k–ω model in the inner region of the boundary layer and the
standard k–ϵ model in the outer region. It also uses a modied
turbulent viscosity equation to take into consideration the transport
effects of the shear stresses. In this model the turbulent viscosity is
calculated from
(4)
where F2 is a blending function, s is the strain-rate magnitude, α1
=0.31 and α* is given by
(5)
where
, Rk = 6.0, , in high-Re ows and
ω is the turbulence specic dissipation rate. These features make the
k–ω SST model more robust and accurate for ows with adverse
pressure gradients and separation and reattachment comparing to the
standard case.
Second-order spatially accurate schemes were used to discretize the
governing RANS equations within the framework of the nite
volume method (FVM). The SIMPLEC algorithm is used for
pressure-velocity coupling, and iconCFD’s iconSimplecFoam version
3.0.17, based on OpenFOAM’s simpleFoam, was used to solve the
governing equations. The RANS simulations were run for about 3000
iterations each on 144 processors on high performance computing
clusters, and the combined meshing and simulation time for each case
was approximately 12 hours. The behavior of the drag force on the
vehicles was monitored, and the solution was assumed to be nalized
and converged when the drag coefcient exhibited either a repeatable
behavior or showed no signicant change with each successive
iteration. Based on this criterion, the simulations were observed to be
converged before reaching the above mentioned iteration number.
DES Method
Three-dimensional incompressible DES simulations were also carried
out to solve the transient ow over different vehicles. DES is a hybrid
method which combines the benet of using RANS-based methods in
the wall proximity region and LES in the rest of the domain. The
popularity of this method is growing rapidly for many industrial
applications due to its comfort for satisfying the mesh requirements
compared to the LES method. The DES approach used in the current
analyses employs a ltered approach where the sub-grid scale model
used is the Spalart–Allmaras turbulence model with a higher
inuence near the wall and a decay far from the wall.
As with the RANS simulations, a second-order nite volume
discretization was used. Second-order temporal discretization was
also employed, and the PISO algorithm was used for pressure-
velocity coupling. A time step of 0.0001 seconds was chosen for the
transient cases, and simulations were conducted on 144 and 240 cores
on the HPC system. Transient simulations were run for 1.5 seconds of
simulated time (which corresponded to 3-5 days of wall clock time),
and statistical averages were computed over the last 1.0 seconds of
simulated time.
Determination of the ranking accuracy of the predicted aerodynamic
drag coefcient was the primary focus of the study. Flow eld
visualization comparisons were conducted in order to conrm the
correctness of the solution and inspect for possible offsetting errors.
Geometry Modeling and Boundary Conditions
An example of one geometry modelled in the current study is
presented in Figure 3. The size of the domain is kept constant in all
cases and only the vehicle is changed. To minimize the inuence of
the inlet, outlet, sides and top boundaries on the pressure and velocity
elds, the computational domain is considered to be 60 m in length,
50 m in width and 30 m in height. Based on this dimensions, the
maximum blockage ratio for all vehicles analyzed is between 1.5% to
2.5%, which is smaller than the 3% maximum suggested by Franke et
al. [4]. Note that a constant numerical setup was applied to all cases.
Figure 3. Computational domain and boundary conditions.
For consistency, a single user independent meshing strategy was
applied to all vehicle models. iconCFD’s hexahedral dominant
unstructured mesh generator iconHexMesh, derived from
OpenFOAM’s snappyHexMesh, was used for creating the hybrid
mesh inside the domain. There are different approaches in CFD for
accounting the wall effect. In the current study, the wall function
approach, which uses the law of the wall, was used for modeling the
ow in the wall region. This approach was used for both k–ω SST
and DES methods. In order to use a wall function, a grid spacing with
30 < y+ < 300 is required. Here, y+ = uτy/ν is a non-dimensional wall
distance, where uτ = (τw/ρ)0.5 is the friction velocity, y is the wall
normal distance, ν is the kinematic viscosity, ρ is the density and τw is
the wall shear stress.
In order to capture the inuence of the high velocity gradients in the
wall regions, a ner mesh is used for regions close to the walls
(ground surface and vehicle parts); while for the rest of the domain, a
coarser mesh is used. Cells adjacent to most parts of the vehicle were
rened 9 times, with cells next to ner-detailed parts such as tires and
grilles being rened 10 times. Near all vehicle parts a boundary layer
mesh with 6 high aspect ratio layers was generated with an expansion
ratio of approximately 1.3, such that the rst cell center was located
in the log law region of the boundary layer, consistent with the high
Reynolds number wall function strategy. Additional renement
regions were added to capture the vehicle wake, an example of which
is shown in Figure 4. On average, the vehicle models were simulated
with about 15 million surface mesh elements, and the entire
computational domain contained approximately 100 million volume
mesh elements.
Figure 4. Constructed mesh around a model.
The boundary conditions considered include a constant velocity of 80
mph at the inlet with a turbulence intensity of 0.5%, which matches
the wind tunnel condition. A symmetry boundary condition is applied
to both the top and side walls. The ground surface is split two
sections, one with slip condition and the other with the no-slip
condition. In order to consider the boundary layer growth in the
upstream of the domain, a distance of 9.6m in the upstream ahead of
the vehicle is set to have no-slip condition. For the rest of the ground
surface, a slip condition is used. A pressure outlet condition that lets
the ow to exit to the atmosphere is applied at the outlet. For static
ground cases, static boundary conditions were used on the oor and
the tires. For other cases, in order to correctly consider the car
movement, moving ground and rotating tire boundary conditions
were used. The numerical wind tunnel was tuned to generate
measured empty tunnel boundary layer growth for static conditions in
order to best replicate wind tunnel measurements.
Table 1. Summary of validation models.
Validation Models
The objective of this study was to determine the capability of an
open-source CFD solver to support aerodynamic development across
all vehicle programs. It was therefore necessary to evaluate the solver
across a wide range of vehicle shapes and variants. Detailed B-car,
sedan, SUV and truck shapes, as well as multiple variants of each
shape, are included in this study. A select set of simplied models
(DrivAer notchback, fastback and wagonback) are also included [5].
A summary of all cases simulated and their setup information is
provided in Table 1.
Model Geometry
Model geometry for each of the base models are presented in Figure
5. The details of each model variant are presented in Figure 5 to
Figure 15.
Figure 5. Validation set of base models.
Figure 6. DrivAer models with configuration variants.
Figure 7. B-car SUV base model and configuration variants.
Figure 8. B-car base model and configuration variants.
Figure 9. Sedan 1 base model and configuration variants.
Figure 10. Sedan 2 base model and configuration variants.
Figure 11. Sedan 3 base model and configuration variants.
Figure 12. Truck 1 base model and configuration variants.
Figure 13. Truck 2 base model and configuration variants.
Figure 14. Performance vehicle base model and configuration variants.
Figure 15. SUV base model and configuration variants.
Metrics
Cd ranking accuracy is the primary metric for comparing simulated
and measured results. It provides a measure of the capability of the
simulation tool to correctly predict the Cd trends across the vehicle
sample set used in this study. Cd development plots and ow
visualization were used to identify any signicant differences
between the steady state and transient simulation results.
Cd Ranking Accuracy
In this paper ranking accuracy is dened as the number of correctly
predicted positional rankings divided by the total number of
positional rankings. A visual example is presented in Figure 16. The
blue bars show ve sample measurements of increasing value. The
red line shows the predicted values for each sample measurement.
Each measurement has a rank order position relative to the other four
samples, leading to a total of 20 positional rankings for this data set.
The rank order of samples 1, 2 and 3 were correctly predicted four
times. The rank order of samples 4 and 5 were correctly predicted
three times. The net result is a total of 18 correctly predicted sample
rank orders and a ranking accuracy of 90%.
Figure 16. Ranking accuracy example.
Two metrics are used in this paper to compare the results of the
steady state and transient simulations.
Cd Development Plots
The rst comparison metric is the Cd development plot, an example
of which is given in Figure 17. This plot is essentially a visual
representation of drag accumulation moving from the front to the rear
of the vehicle. This plot is used to assist in identifying locations on
the vehicle with notable differences between steady state and
transient local drag force predictions.
Figure 17. Cd development plot example.
Flow Visualization
The second comparison metric used in identifying signicant
differences between steady state and transient simulation results is
ow visualization. A combination of velocity distributions and
isosurfaces of wake structure were used in this study.
Results
A total of 54 unique vehicle and vehicle variants were evaluated. Cd
trends versus experimental measurements are plotted in Figure 18. A
positive bias in the absolute Cd values was observed in both the
transient and steady state simulation results. The values in the plot
have been adjusted by subtracting the average bias from the drag
values. The simulation results reveal a high level of correlation to the
experimental trends for both the transient and steady state results.
Figure 18. Normalized Cd trend predictions versus experimental
measurements.
Cd Ranking Accuracy - All Cases
The correlation presented in Figure 18 is further supported by the
ranking accuracy bar graph shown in Figure 19. The steady state and
transient ranking accuracy results from the entire data set were very
good and generally equivalent to similar studies performed on
commercial CFD solver applications.
Figure 19. Steady state and transient Cd ranking accuracy for all data sets.
As Figure 19 shows, the ranking accuracy for the transient results
was 94.3%, while the ranking accuracy for the steady state results
was 91.3%. The ranking accuracies of both methods were
approximately equivalent overall, and were in alignment by
expectations based on previous results from more traditionally
utilized commercial solvers. The broad range of shapes and boundary
conditions within the data set, however, creates the potential for
making lower accuracy results within subsets of the data. To that
extent, the data was divided into different specic groupings as
following and the ranking accuracies recalculated for each case:
1. Static Ground
2. Moving Ground
3. Sedans
4. SUVs
5. Trucks
6. Open AGS
7. Closed AGS
8. Closed Grille
9. B-cars
A total of 39 samples were included for the static ground cases, 15 for
the moving ground cases, 21 for sedans, 5 for SUVs, 12 for trucks, 6
for B-cars, 9 for open AGS, 23 for closed AGS and 13 for closed
grille cases. The following observations were made:
Ranking accuracy for the static ground cases was equal or
slightly higher than the ranking accuracy calculated for the
entire data set. DES results were 1.4% greater. Results from the
k–ω SST model simulations were 1.4% greater.
For the moving ground cases, the ranking accuracy was
approximately equal to the number calculated for the entire data
set. DES results were 0.9% greater. k–ω SST results were 1.8%
less.
For sedan cases the ranking accuracy number was lower than
the one calculated for the entire data set. DES results were 6.7%
lower. k–ω SST results were 12.3% lower.
Ranking accuracy for the SUV subset of cases was lower than
the ranking accuracy calculated for the entire data set. DES
results were 14.3% lower. k–ω SST results were 21.3% lower.
DES ranking accuracy for the truck subset of cases, at 86.4%,
was 7.9% lower than the ranking accuracy of the entire data set.
k–ω SST ranking accuracy, at 77.3%, was 14.0% lower.
For the open AGS cases the ranking accuracy was slightly lower
than the ranking accuracy calculated for the entire data set. DES
results were 2.6% lower. k–ω SST results were 8% lower.
Ranking accuracy for the closed AGS subset of cases was
slightly lower than the ranking accuracy calculated for the entire
data set. DES results were 2.2% lower. k–ω SST results were
3.9% lower.
Ranking accuracy for the closed grille subset of cases was
slightly lower than the ranking accuracy calculated for the entire
data set. DES results were 5.8% lower. k–ω SST results were
6.7% lower.
Ranking accuracy for the B-car subset of cases was higher than
the ranking accuracy calculated for the entire data set. DES
results were 5.7% higher. k–ω SST results were 8.7% higher.
The 100% ranking accuracies are most likely an anomaly
attributable to the small size of the subset and to the types of
variants included in the subset.
Cd Development Plots - Static Ground Cases
Cd development plots for the 12 base models simulated with static
ground conditions (with both the k–ω SST and DES turbulence
models) are presented in Figure 20 to Figure 31. The experimentally
measured total vehicle Cd is included for reference in each plot.
Key observations:
1. Both DES and k–ω SST closed front end simulations generated
similar results (Figure 20 to Figure 22) from the front bumper to
the midpoint of the wheelbase (x = 0).
2. Both DES and k–ω SST open front end simulations (Figure 23
to Figure 31) generated similar results from the front bumper to
the heat exchangers.
3. Both turbulence models capture the same overall Cd
development pattern in the under hood region, although some
differences are seen in absolute Cd values predicted in that
region.
4. Cd development patterns were similar for most vehicles from
the rear of the engine compartment to the B-pillar with the
exception of the performance vehicle (Figure 29). Further
investigation is required to understand the mechanisms behind
these differences.
5. Cd development with the DES model increased at a slightly
greater rate than the k–ω SST model between the B-pillar and
the rear corners of the models. Initial review of the ow elds
indicate that this may be the result of differences in underbody
ow rates.
6. No consistent pattern in the Cd development differences between
DES and k–ω SST simulations was observed in the data at the
base of the models. This result is not totally unanticipated given
the complexity of the ow eld in that region.
7. A positive bias in the DES predicted total vehicle Cd is observed
in the data.
Figure 20. Cd development plot for the fastback variant of the simplified
reference model (detailed floor, static ground condition).
Figure 21. Cd development plot for the notchback variant of the simplified
reference model (detailed floor, static ground condition).
Figure 22. Cd development plot for the wagonback variant of the simplified
reference model (detailed floor, static ground condition).
Figure 23. Cd development plot for Sedan 1 (closed AGS, shields removed,
static ground condition).
Figure 24. Cd development plot for Sedan 2 (open AGS, static ground
condition).
Figure 25. Cd development plot for Sedan 3 (baseline, static ground
condition).
Figure 26. Cd development plot for the performance vehicle (closed AGS,
static ground condition).
Figure 27. Cd development plot for the B-car SUV (baseline, static ground
condition).
Figure 28. Cd development plot for the B-car (baseline, static ground
condition).
Figure 29. Cd development plot for the SUV (open AGS, static ground
condition).
Figure 30. Cd development plot for Truck 1 (baseline, static ground condition).
Figure 31. Cd development plot for Truck 1 (closed AGS, static ground
condition).
Flow Visualization - Static Ground Cases
Isocontours of zero total pressure for the 12 base models simulated
with static ground conditions (with both k–ω SST and DES
turbulence models) are shown in Figure 32 and Figure 33. Centerline
velocities (with the colorbar scale indicating the velocity magnitude
in meters per second) for the 12 base models are displayed in Figure
34 and Figure 35. These gures are useful in highlighting the
differences in the predicted ow structures around vehicles.
Key observations:
1. Based on these evaluations, the RANS-based k–ω SST creates
larger separation vortices (as compared to the transient DES
model) around tires. The k–ω SST model also creates wall-
attached vortices in the wake region.
2. It can be realized from the gures that the RANS-based k– ω
SST model creates higher underbody velocities.
3. For Sedan 1, 2 and 3, higher underbody velocities are observed
(marked with arrows in Figure 34) using the k–ω SST model,
as compared to the DES cases. This might be a possible reason
for the larger separation vortices from the front tires observed in
Figure 32. Low velocity regions can be seen over the backlight
in all three Sedans (k–ω SST cases). This is consistent with the
separation bubbles observed in that region.
4. The velocity contours for the SUV obtained from two models
are similar, except for the mid-underbody (marked with arrows
in Figure 34), where k–ω SST predicts much lower velocity.
5. For Truck 1 and 2, higher underbody velocities can be observed
using k–ω SST model (marked with arrows in Figure 34).
Similar to the phenomenon observed for SUV vehicles, low
velocities can be seen in the mid-underbody region (marked
with arrows in Figure 34) using the k–ω SST model.
Figure 32. Isocontours of zero total pressure generated using the steady, k–ω
SST turbulence model for static ground cases.
Figure 33. Isocontours of zero total pressure generated using the unsteady,
DES turbulence model for static ground cases.
Figure 34. Contours of velocity magnitude generate using the steady, k-ω SST
turbulence model for static ground cases.
Figure 35. Contours of velocity magnitude generated using the unsteady, DES
turbulence model for static ground cases.
Cd Development Plots - Moving Ground Cases
Cd development plots for the 5 base models simulated with moving
ground conditions (with both k–ω SST and DES turbulence models)
are presented in Figure 36 to Figure 40. The experimentally measured
total vehicle Cd is included for reference in each plot.
Key observations:
1. Both DES and k–ω SST open front end simulations generated
similar results from the front bumper to the heat exchangers
(Figure 36 to Figure 40).
2. Both turbulence models capture the same overall Cd
development pattern in the under hood region, although some
differences are seen in absolute Cd values predicted in that
region (Figure 36 to Figure 40).
3. Cd development patterns were similar for most vehicles from the
rear of the engine compartment to the B-pillar. DES and k–ω
SST patterns for the performance vehicle (Figure 38) showed
less variation than that seen under static ground conditions.
4. Cd development with the DES model increased at a slightly
greater rate than the k–ω SST model beginning at the C-pillar.
5. Cd development differences between DES and k–ω SST
simulations observed at the base of the models were much less
than those observed under static ground conditions.
6. A positive bias in the DES predicted total vehicle Cd is observed
in the data similar to that noted in the static ground simulations.
Figure 36. Cd development plots for Sedan 1 (closed AGS, shields removed,
moving ground condition).
Figure 37. Cd development plots for Sedan 2 (open AGS, moving ground
condition).
Figure 38. Cd development plots for performance vehicle (closed AGS,
moving ground condition).
Figure 39. Cd development plots for SUV (open AGS, moving ground
condition).
Figure 40. Cd development plots for Truck 2 (closed AGS, moving ground
condition).
Flow Visualization - Moving Ground Cases
Isocontours of zero total pressure for the ve selected vehicles (Sedan
1 and 2, the performance vehicle, the SUV and Truck 2) simulated
with moving ground conditions (with both k–ω SST and DES
models) are shown in Figure 41 and Figure 43. Figure 42 and Figure
44 show the centerline velocity magnitude contours.
The key observations from these gures were:
1. For all ve vehicles, larger separation vortices from the front
tires can be observed using the k–ω SST model compared to
those from the DES model.
2. For all ve vehicles, higher underbody velocities are observed
(marked with arrows in Figure 42) using the k–ω SST model as
compared to DES.
3. The k–ω SST model predicts lower velocities for regions close
to the mid-underbody for the SUV and Truck 2 (marked with
arrows in Figure 42).
Figure 41. Isocontours of zero total pressure generated using the steady, kω
SST turbulence model for moving ground cases.
Figure 42. Contours of velocity magnitude generated using the steady, k–ω
SST turbulence model for moving ground cases.
Figure 43. Isocontours of zero total pressure generated using the unsteady,
DES turbulence model for moving ground cases.
Figure 44. Contours of velocity magnitude generated using the unsteady, DES
turbulence model for moving ground cases.
Summary and Conclusions
A combination of 54 different vehicle shapes and shape variants were
evaluated against both static and moving ground wind tunnel
experimental measurements. The ability of both transient DES and
steady k–ω SST RANS simulations to correctly rank all 54
congurations was evaluated using the ranking accuracy metric. Both
the DES and SST models achieved ranking accuracies greater than
90% overall, where DES showed an advantage of a few percentage
points.
Transient simulation ranking accuracy of total vehicle Cd was found
to be approximately equivalent to commercial transient lattice
Boltzmann solver results even though a consistent upward bias in
absolute Cd values was observed. This was true across the entire data
set as well as within specic subsets of the data. Steady state ranking
accuracy of total vehicle Cd was also found to be approximately
equivalent to commercial transient lattice Boltzmann solver results.
This was true across the entire data set as well as within specic
subsets of the data.
Cd development plot comparisons conrmed that similar Cd
development patterns were predicted by both steady state and
transient simulations. The most signicant variations were linked to
ow structure differences between steady state and transient
simulations observed in the tire wakes, underbody velocities and
body wakes.
Analysis of specic subsets of the data set highlighted the fact that
the ranking accuracy of transient simulations is slightly better than
steady state across all subsets with the most signicant improvement
seen for trucks and SUVs.
The results of this study indicate that both steady state and transient
simulations are capable of supporting general Cd trend analyses.
Transient analyses are recommended for maximum trend accuracy,
and in particular for truck and SUV models. However, the faster
turnaround time for steady state simulations makes this a very time
efcient option to support proportion studies and theme evaluations
early in the design cycle when absolute accuracy is less critical.
The next steps for this study would include an investigation of the
positive bias in absolute vehicle Cd, expansion of the validation set to
expand the number of samples in each subset to investigate more
subtle vehicle surface changes, and the addition of vehicle lift ranking
accuracy as an evaluation metric.
References
1. OpenFOAM, v2.1.1, 2012. www.openfoam.com.
2. Lietz, R., Hupertz, B., Lewington, N., Silveira, R., ,
“Benchmarking of an Open Source CFD Process for
Aerodynamics Prediction of Multiple Vehicle Types,” SAE
Technical Paper 2011-01-0163.
3. Menter, F.R., 1994. Two-equation eddy-viscosity turbulence
models for engineering applications. AIAA J. 32(8), 1598–1605.
4. Franke, J., Hellsten, A., Schlünzen, H., Carissimo, B., COST
2007. Best practice guideline for the CFD simulation of flows
in the urban environment, Action 732, Quality assurance and
improvement of micro-scale meteorological models.
5. Heft, A., Indinger, T., and Adams, N., "Introduction of a New
Realistic Generic Car Model for Aerodynamic Investigations,"
SAE Technical Paper 2012-01-0168, 2012, doi:10.4271/2012-
01-0168.
Contact Information
Robert Lietz
Vehicle Aerodynamics Technical Specialist and Core Supervisor
Ford Motor Company
rlietz@ford.com
Acknowledgments
The authors acknowledge support for this project from Burkhard
Hupertz and Neil Lewington of Ford Motor Company.
Definitions and Abbreviations
CFD - Computational Fluid Dynamics
HPC - High Performance Computing
DES - Detached-Eddy Simulation
AGS - Active Grille Shutters
RANS - Reynolds Averaged Navier-Stokes
OPENFOAM® - Registered trademark of OpenCFD Limited,
producer and distributor of the OpenFOAM software.
FVM - Finite Volume Method
LES - Large Eddy Simulation
The Engineering Meetings Board has approved this paper for publication. It has successfully completed SAE’s peer review process under the supervision of the session organizer. The process
requires a minimum of three (3) reviews by industry experts.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or
otherwise, without the prior written permission of SAE International.
Positions and opinions advanced in this paper are those of the author(s) and not necessarily those of SAE International. The author is solely responsible for the content of the paper.
ISSN 0148-7191
http://papers.sae.org/2017-01-1524
... As described by Mockett, Knacke and Thiel [4], the effective bandwidth of signal can be calculated by using the Repeat Reference method to calculate s at different times and using Equation 16 to calculate B at each point. The value of B is then chosen by taking the 10/31/2019 minimum of all the calculated values for a signal. ...
... In this section, we will apply the methods discussed above to Cd and Cl signals produced by unsteady CFD simulations of the detailed underbody drivAer fastback and estate back geometries described in [14] and the medium cab, medium bed GTU geometry presented in [15]. Simulations were carried out with IconCFD's unsteady incompressible solver with the simulation set up similar to that described in [16]. Simulations were run for 100 s to create a signal length suitable for use with the repeat reference method, which serves as the benchmark to compare other methods against. ...
... Aerodynamic study of the vehicle platoon through CFD [26] is usually employed to study the drag characteristics of each vehicle in the platoon since experimental testing abounds considerable risk and cost. In particular, the experimental study of drag on platoons using wind tunnel testing is not feasible for platoons greater than 4 cars given the space restriction within a wind tunnel as opposed to greater length platoons for scaled-down vehicle size, let alone the original size of the vehicles. ...
Article
Full-text available
Machine learning is used for extraction of valuable information from data thus helping in exploration of hidden patterns, leading to learning models that can be used for prediction. In the domain of autonomous vehicles machine learning techniques have been applied in several areas, vehicle platooning being one of them. Vehicle platooning is a vital feature of automated highways which provides the key benefits of fuel economy, road safety and environmental protection coupled with safe road transportation. However, high computational cost associated with the numerical simulation of vehicle aerodynamics makes the Computational Fluid Dynamics (CFD) study of vehicle platoon prohibitively expensive and complex. Machine learning, with its high predictive power, has emerged as a promising compliment to CFD studies of external aerodynamics. This paper presents estimation error based performance comparison of five different supervised learning algorithms: Support Vector Regression, Polynomial Regression, Linear Regression and two different models of Neural Networks for prediction of aerodynamic drag coefficient corresponding to each vehicle in a two, three and four vehicle platoon configurations based on the drag coefficients provided by experimental study at different inter-vehicle distances. Predicted drag coefficients are then juxtaposed with CFD data from numerical simulations to evaluate closeness to experimental drag coefficients. Results reveal that polynomial regression model best fits the aerodynamics with 0.0223 estimation error. To the best of our knowledge no machine learning based methods have been applied before for modeling aerodynamic drag on vehicle platoon.
... In this section, we present preliminary CFD simulations of the GTU. Simulations were carried out using iconCFD®'s unsteady incompressible solver with a set up similar to the one described in [43]. Further details are given in Table 7. ...
... In this section, we present preliminary CFD simulations of the GTU. Simulations were carried out using iconCFD ® 's unsteady incompressible solver with a set up similar to the one described in [43]. Further details are given in Table 7. ...
Preprint
Full-text available
Compared to sedans, pickup truck (PU) and sports utility vehicle (SUV) body styles pose additional aerodynamic challenges due to their complex wake structures. The introduction of a realistic generic PU and SUV model as an open access tool is expected to yield benefits to the wider aerodynamics community in characterizing wind tunnel & CFD performance including correlation, development and process control, developing a deeper understanding of the aerodynamic mechanisms unique to PU and SUV body styles, and finally, research and development around a generic realistic PU and SUV model between OEMs, vendors and academia. This paper introduces the Generic Truck Utility (GTU) as a realistic, generic PU truck and interchangeable SUV model. The paper will focus on the design and development of the GTU and will present a summary of preliminary experimental results of the GTU complemented by numerical simulations.
... This includes a detailed look into highly sensitive rear window flow region which will be referred to later in this paper. Lietz et al. [11] have also presented simulation results with qualitative analyses but without reporting any Cd values or detailed quantitative comparisons. Finally John et al. [8] have recently presented comparisons of results from PowerFlow simulations to a new set of experiments from FKFS for the fully detailed DrivAer model with moving ground. ...
Chapter
An unobscured view from the vehicle during rainy weather conditions is essential for occupants’ safety and comfort. With the decrease in the time available for vehicle development and testing, it is becoming even more important to control and predict vehicle water management early in the development cycle to avoid undesired soiling effects. To do so, a transient external aerodynamics airflow coupled with a discrete particle phase is simu-lated, where rain droplets away from the vehicle are modeled as Lagrangian particles and rain droplets that impact the vehicle are represented through a film model. A novel, time-efficient implementation of a windscreen wiper based on [1] within iconCFD [2] is also summarized. A representative example - rain-induced reduced visibility on the side window - is considered to discuss the physics, time scales, and recommended best practices for rain soiling simulations. The simulation results for this example case are compared with experimental data.
Chapter
An unobscured view from the vehicle during rainy weather conditions is essential for occupants’ safety and comfort. With the decrease in the time available for vehicle development and testing, it is becoming even more important to control and predict vehicle water management early in the development cycle to avoid undesired soiling effects. To do so, a transient external aerodynamics transient simulation coupled with a discrete particle phase is explored, where rain droplets away from the vehicle are modeled as Lagrangian particles and rain droplets that impact the vehicle are represented through a film model. A novel, time-efficient implementation of a windscreen wiper based on [1] within iconCFD [2] is also summarized. A representative example - rain-induced reduced visibility on the side window - is considered to discuss the physics, time scales, and recommended best practices for rain soiling simulations. The simulation results for this example case are compared with experimental data.
Article
State of the art aerodynamic research of vehicles often employs strongly simplified car models, such as the Ahmed and the SAE body, to gain general insights. As these models exhibit a high degree of abstraction, the obtained results can only partly be used for the aerodynamic optimization of production vehicles. Aerodynamic research performed on specific vehicles is on the other hand often limited due to their short life span and restricted access. A new realistic generic car model for aerodynamic research-the DrivAer body-is therefore proposed to close this gap. This paper focuses on the development of the model and the first experimental results, namely force and pressure measurements of the different configurations. The experiments were performed in the recently updated Wind Tunnel A of the Institute of Aerodynamics and Fluid Mechanics at the Technische Universität München.