See discussions, stats, and author profiles for this publication at: [Link]

net/publication/266237648

Optimization in the Automotive Industry

Article · January 2003

DOI: 10.1007/978-1-4613-0233-9_3

CITATIONS READS

4 2,204

1 author:

Donald R. Jones

24 PUBLICATIONS 13,621 CITATIONS

SEE PROFILE

All content following this page was uploaded by Donald R. Jones on 17 October 2020.

The user has requested enhancement of the downloaded file.

Chapter 3

OPTIMIZATION IN THE
AUTOMOTIVE INDUSTRY

Donald R. Jones
General Motors Corporation
don .jones@[Link]

Abstract Today, most successful applications of optimization in the automotive

industry are at the detailed-design level. Examples include structural
optimization of metal gauges and beam cross sections, topology opti-
mization of cast parts, and engine calibration. Despite these successes,
it is fair to say that optimization is used less often, and has had a
smaller impact, than most researchers might have expected. The reason
is the presence of numerous barriers, such as poor connections between
computer aided design (CAD) and computer aided engineering (CAE)
software, insufficient predictive accuracy of computer simulations, dif-
ficulty in automating and iterating computer analyses, long simulation
times, and poor availability of data. As a result, until recently, the
focus has just been on getting to the point where we can do optimiza-
tion. Fortunately, many of these barriers are being overcome, and the
next decade is likely to witness a fast growth in the use of optimization
within the industry. Accordingly, the focus will shift to more interesting
challenges, such as robust design, multi-disciplinary optimization, and
the roll-down of vehicle-level requirements to subsystem specifications.

Keywords: Optimization, automotive, industry, CAE, engineering

Introduction
Designing an automobile is a complicated affair, involving manufac-
turing, distribution, and many different engineering disciplines. Designs
must meet myriad performance criteria-styling, cost, safety, aerody-
namics, durability-just to name a few. With such complexity, one
would think that there would be ample opportunity for optimization.
39
P.M. Pardalos and V. Korotkikh (eds.), Optimization and Industry: New Frontiers, 39-58.
© 2003 Kluwer Academic Publishers.
40 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

And one would be right. But this same complexity has also slowed the
penetration of optimization methods. It has taken time for computer
simulations to become sufficiently accurate that it seems worthwhile to
use them for design optimization. It has taken time for computer hard-
ware and software to get to the point where optimization can be applied
on a routine basis. It has taken time for data acquisition and manage-
ment systems to get to the point where the data required for optimization
is readily available. The good news is that many of these barriers are
being removed, and the future for optimization is bright.
In this paper, I will focus on applications of optimization within engi-
neering. I will review some success stories, identify the barriers that have
slowed the penetration of optimization, discuss new developments that
are removing these barriers, and highlight some of the key challenges
and opportunities going forward.

1. Well-Established Applications
Optimization of the sizes of structural parts has become fairly routine.
The starting point for these analyses is a structural model consisting of
beams and panels that can be analyzed with commercial finite-element
solvers, such as NASTRAN or OptiStruct. An example is shown in
Figure 3.1. This software allows users to select the cross-section for the
beam elements from a library and further define the beam by specifying
its height, width, and metal gauge. Based on this information, structural
properties of the beam elements (stiffness, etc.) can be computed and,
using these properties, performance measures for the full model can be
derived (e.g., modal frequencies). A typical optimization problem might
be to choose the value for the metal gauges to minimize weight subject
to a constraint on the modal frequencies.
In these problems, the loads are assumed to be small enough that
the material only temporarily deflects from its initial shape (elastic de-
formation), returning to its original shape when the loads are removed.
Mathematically, the relationship between load and deflection of an ele-
ment is linear, and this assumption so simplifies the analysis that one
can compute analytic derivatives for most performance measures with re-
spect to simulation inputs. As a result, even though the corresponding
optimization problems are large, they are easily handled by standard
algorithms for large-scale, gradient-based local optimization. All the
major software vendors provide built-in facilities to perform such "size"
optimization.
After size optimization, perhaps the next most common use of opti-
mization in engineering is the optimization of the shape of a component.
Optimization in the Automotive Industry 41

Beams

Beams & Panels

Beam
cross-
Q [)
sections d i)
selected
from library ~:::j IJ
-y

Figure 3.1. Size optimization in linear structural analysis.

42 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

Control arm optimized with OptiStruct

This part subjected to

parameterized shape variation.

Figure 3.2. In Altair Engineering's Optistruct code, the shape of a component such
as the above control arm can be parameterized in terms of perturbation vectors, and
then adjusted to solve some optimization problem.

In one approach to shape optimization, variations in the shape of a com-

ponent, such as the control arm shown in Figure 3.2, are parameterized
via perturbation vectors. Various optimization problems can then be for-
mulated, such as changing the shape to minimize mass subject to limits
on stress. In this case, gradient-based optimization is employed. This
approach is embodied, for example, in Altair Engineering's OptiStruct
code.
A second, quite innovative, approach to shape optimization avoids the
need to parameterize shape perturbations, but instead allows the shape
to "evolve" in a way that mimics biological growth (Harzheim et al.,
1997). Bones and trees both grow most where they are stressed, and
grow least, or even atrophy, where stress is low. As a result, they nat-
urally evolve to a fully and equally stressed structural design. This is
mimicked in the computer by starting from an initial shape and com-
puting the stress of the surface elements. Where stress is above a pre-
determined reference level, material is added (the boundary expands in
the normal direction). Where stress is below the reference level, mate-
rial is removed (the boundary contracts in the normal direction). The
process is iterated until the stress is equalized at the reference level.
Figure 3.3 shows an example of how this process works for the case of a
steering knuckle. The key advantage of the "biological growth" approach
is the elimination of the need to choose a (somewhat arbitrary) way to
parameterize the shape perturbations. The disadvantage, however, is
that the only possible objective function is that of a uniform, equally
stressed design, and one can not add constraints (e.g., one can not add
a constraint having to do with modal frequencies).
Optimization in the Automotive Industry 43

Initial shape
of steering
knuckle

After optimization with

biological growth rule,
stress reduced by 22%.

Figure 3.3. In shape optimization based on the "biological growth rule," the bound-
ary shape is systematically perturbed to yield a fully and equally stressed design.

Perhaps the most exciting application of optimization to engineering

is so-called "topology optimization." Here the interest is in finding the
optimal three-dimensional form or topology of a part. For example, con-
sider the engine bracket shown in Figure 3.4. We start with a "design
space" of material that could be potentially used. This would be com-
parable to starting with a large block of marble for making a sculpture.
Next, material is removed using an iterative process in a way that leads
to a high performing design. The resulting topological form is usually
somewhat rough and not easily manufactured, and so some interpreta-
tion of the form is made to yield a final design that can manufactured.
As with shape optimization, there are at least two approaches for
removing the material in topology optimization. Figure 3.4 uses the
"soft kill" option of General Motors' Opel division (Harzheim et al.,
1997). This method is similar to the biological growth rule discussed
earlier; in particular, the Young's modulus of each element is varied
in an iterative fashion to yield a fully and equally stressed design. In
the end, elements with very small Young's modulus are interpreted as
being absent (they are not really deleted, hence the name "soft kill").
In so-called homogenization methods, several parameters of each three-
dimensional element are varied to optimize an objective function subject
to constraints. At the end, certain elements are interpreted as absent
based on the values of the optimized parameters.
The intuitive graphical user interfaces of many of these programs
makes using optimization easy. In fact, most of the initial successes of
optimization have happened when optimization capabilities were added
to analysis codes that engineers were already using to evaluate the per-
formance of designs. Another good example is the GT-Power software
44 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

Design space

After topology
optimization
& smoothing

Final design

Figure 3.4. In topology optimization one searches for the best three-dimensional
form a part, here an engine bracket.
Optimization in the Automotive Industry 45

produced by Gamma Technologies. This software allows users to build

flow and combustion models with a point-and-click graphical user in-
terface that draws on a library of components. An optimizer is built
into the software, and allows users to solve problems such as "adjust
cam timing, cam duration, or intake manifold length to maximize en-
gine torque at given speeds" or "adjust spark timing to maximize fuel
economy, subject to constraints on torque and emissions."
While adding optimization algorithms to engineering analysis codes
on a one-at-a-time basis is effective, it is a slow process due to the large
number of analysis codes used in engineering. This is, nevertheless, how
optimization got started within engineering. In addition to being slow,
this one-at-a-time approach has limits: what does one do if one wants to
perform multi-disciplinary optimization in which the goal is to maximize
energy absorption in a crash (nonlinear structural analysis) subject to
constraints on modal frequencies (linear structural analysis)? In this
case, the objective function would be computed by an entirely different
analysis code (e.g., LS-Dyna) than the constraint (e.g., NASTRAN). As
a result, even if both of these codes had built-in optimizers, we would
still have no way to solve the multidisciplinary problem. Progress in this
area has come only recently, and will be discussed later.

2. Barriers to the Use of Optimization

Despite the above successes, the penetration of optimization in engi-
neering has been slow due to several barriers: insufficient accuracy of
math models, too little time for optimization, difficulty in iterating, and
difficulty in setting objectives.
The culture of the automotive industry is one based heavily on the use
of physical models and testing. For example, it was only in the last few
years that designers began to do initial design work with CAD applica-
tions as opposed to using hand-built clay models. To overcome this bias
towards physical testing, math models have had to show considerable
accuracy in matching test results. It may be argued (correctly) that
a math model does not have to be highly accurate to provide correct
directional guidance, but due to the automotive culture, a good deal
of accuracy is nevertheless required to gain acceptance of any recom-
mendations based on computer simulations. Fortunately, this barrier is
being slowly but surely overcome, helped greatly by the improvements
in computer software and in processing speeds.
Another, somewhat related problem is that there is often very little
time to do optimization studies. Designers have been used to design-
ing for toolmakers, not computer analysts. They treat hardware build
46 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

events as serious deadlines, and they are used to providing information

(usually on paper drawings) at the highly detailed level required to build
parts. But it has taken time for designers to begin to take seriously the
need to give engineers math data to facilitate engineering analyses. As
a result, data is often late and in the wrong form. For example, for an
aerodynamic analysis, only the outer surface of the vehicle is required,
but designers may deliver a fully-detailed model. The engineer will then
have to spend considerable time stripping away the unneeded detail be-
fore he can proceed. Similarly, data may "look" OK on the computer
screen but not be topologically correct. For example, line segments that
should meet end-to-end to describe the perimeter of a part might, on a
zoomed-in view, not quite connect to each other, or perhaps may cross
each other. These problems require time-consuming manual clean up.
Once we have clean geometry, one has to translate this geometrical
data into a "CAE model" that can be used by one of the different analysis
programs. Often, this involves extracting midplane surfaces from solid
parts and then meshing them. While modern software programs have
the facility to do these things, they do not always work flawlessly if the
initial geometry is messy. A good deal of time must therefore be spent
manually correcting poor auto-meshes. Time-consuming manual steps
have also been required to add information on welds that connect parts.
Once a good CAE model is obtained, the next obstacle has been
facilitating the iterations required by an optimization procedure. For
example, changing a design parameter at the geometry level might re-
quire re-tracing some of the previously mentioned manual steps to get
the altered CAE model. Another difficulty is that computer analyses
may fail for numerical reasons, presenting problems for optimization al-
gorithms that expect to get objective and constraint values at every trial
point they suggest. Analysis jobs often produce huge output files; hence,
unless some care is taken, a few iterations of an optimization algorithm
might fill up a user's hard drive and cause a system crash. Even if the
analysis code runs flawlessly, the tedious tasks of setting up input files,
launching the jobs, monitoring their progress, and extracting the results
from output files can take substantial time. Finally, some analysis codes
take a very long time to run, greatly limiting the number "function
evaluations" that can be obtained.
Even if one manages to solve all of the above problems, a few more
obstacles remain. For example, what are the relevant load cases? What
should be our target performance? At the full-vehicle level, we may
know the load cases impinging (say) on the tires as a vehicle drives over
a simulated test track. Likewise, we may know how we would like the
entire vehicle to perform. But what does this mean for the loads on the
Optimization in the Automotive Industry 47
engine mounts or some other part? Without a full-vehicle model, it may
be hard to come up with load cases to use for such isolated components.
Similarly, it may be hard to "roll-down" or translate vehicle-level targets
to targets for subsystems. Even if data on load cases is available, it may
not be stored in a central, easily assessable location, forcing engineers to
use valuable time chasing down the required data using email and phone
calls.
The message of this section is not that optimization is impractical in
the auto industry, but rather that there have been a slew of mundane
issues that the industry has had to work through to get to the point
where optimization is possible. As discussed in the next section, the
good news is that much progress has been made.

3. New Developments
Most engineering analysis codes can be run in batch mode, taking
data from a text-based input file and writing results to an output file.
Early attempts to do automation with such codes involved writing in-
house "wrapper codes" or "scripts" that would create input files with
new parameter values, launch the analysis code, parse the output files
to obtain objective and constraint values, and then use an optimization
algorithm to decide the next point to sample. These codes were success-
ful, but were usually hardwired to solve very specific problems and, like
all in-house software, were hard to maintain.
Recently, several software vendors have produced generic versions of
such wrapper codes. For example, consider the OPTIMUS software
from LMS Technologies. To use OPTIMUS, one first outlines the flow
of data through the analysis codes. This is accomplished in a point-
and-click fashion using the intuitive graphical user interface shown in
Figure 3.5. The particular flow diagram in Figure 3.5 can be interpreted
as follows. First, parameters t1 and t2 are inserted into the input file
[Link] is then run through the NASTRAN analysis code. This
results in two output files, shown by the icons for oilpan_struct.J06 and
oilpan_struct. pch. From the former file, the performance measure Mass
is extracted. The later file then becomes an input into a second analysis
code, SYSNOISE. The file oilpan_acous.cmd is also an input to SYS-
NOISE. The SYSNOISE code produces the output file results. out, from
which the performance measure Integrated_power is directly extracted,
and four other values are extracted and combined to form the perfor-
mance measure called Output-power. This entire flow is created via the
use of drag-and-drop icons.
48 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

oi/PBn_ struct.f06

12

Integrated"'power

Figure 3.5. Describing the process flow from inputs to outputs using LMS OPTI-
MUS.
Optimization in the Automotive Industry 49

S El_ot. anti El ...ent Propedies foe ':.'Jioo ; psooll.l ::'1

PSREtt 1 1 ~tl(8;(f8.6))' 1 1 '1
$ [Link] and El~nt Propeetios for reqion : pshell.r1bl '
PSIlELt. 2 1 U2(O; (fO.6))\ 1 1 ~
$ Elaaante and EIWlltlut proportios for r0910n ; psh<>ll. rib2
P8HELL 5 1 0.0012 l ' 1

Figure 3.6. Parsing of input and output files using LMS OPTIMUS.

By double-clicking on these icons, further details can be provided. For

example, if one double-clicks on the NASTRAN icon, a window pops up
that allows the user to specify the UNIX commands necessary to launch
the NASTRAN code. Similarly, if one double-clicks on the icon for the
input file props. bdj, a window appears that allows one to specify how
the parameters tl and t2 should be inserted into this file. This is done
by bringing up an input file from a previous run and specifying, using
point-and-click methods, which numbers in this file should be replaced
by the values of t1 and t2 selected by the optimization algorithm. In
the example shown in the upper part of Figure 3.6, the numbers in the
file are replaced by format templates such as "%tl[8;(f8.6)]%" which tell
OPTIMUS to put the value of t1 in this location in a field of width 8
with FORTRAN format f8.6.
Similarly, by double-clicking on the icon for the output file oilpan_
struct.j06 one gets a window, shown in the lower part of Figure 3.6, in
which one can tell OPTIMUS how to extract the value of the perfor-
mance measure Mass. To do this, one brings up an output file from a
previous run and gives instructions for how to find the value for Mass.
In this case, the parsing rule is to search for "MASS AXIS SYSTEM",
then go down one line, and take the second (space separated) value on
the line to be the value of Mass.
Other convenient interfaces are provided for setting up a constrained
optimization problem involving the design parameters and performance
50 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

measures. Bounds on the variables can be specified, as well as the par-

ticular optimization algorithm that will be employed. Once everything
is set up, the optimization job is launched and OPTIMUS takes care of
all the details: formatting input files, launching jobs, monitoring their
progress, parsing output files, and iterating until convergence. Other
programs that work similarly are iSIGHT from Engineous Software and
Study Wizard from Altair Engineering.
The flexibility of these commercial wrapper codes makes it possible
to do optimization with analysis codes from different disciplines, such as
crash and structures, so that competing performance measures can be
explicitly traded off against each other. Moreover, these codes can work
with virtually any analysis code that has text-based input and output
codes. These are distinct advantages over the earlier successes that in-
volved building proprietary optimization algorithms into specific codes
on a one-at-a-time basis. These specific optimizers are still extremely
convenient for solving single-discipline problems, but the generic wrap-
per codes have taken optimization to a much higher level.
Another development that holds promise for optimization is the in-
creasing use of "parametric geometry" in CAD systems and the improved
connections between CAD systems and CAE analysis codes. In paramet-
ric CAD, features such as the wheel base are controlled by parameters.
Associations in the model specify how other parameters must adjust if
the wheel base is increased (e.g. the front overhang may increase or the
interior cabin may grow). Ideally, optimization would be done by chang-
ing such high-level parameters, re-meshing to create a new CAE model,
performing an analysis, and iterating with the aid of one of the wrapper
codes. While this ideal has not been achieved everywhere, improvements
in the ability to share parametric data between CAD and CAE codes
has made this possible in certain cases, such as in computational fluid
dynamics (Qian and Yuan, 2001).
For the cases where parametric CAD is not available, or not easily
linked to CAE codes, morphing software is allowing engineers to study
parameterized changes to vehicle geometry. Morphing codes operate
on the CAE mesh. The user specifies a few control points that define
target surfaces or otherwise control how the morphing is done. The
mesh is then distorted to create a new shape. A good example is shown
in Figure 3.7. The points controlling the morphing can then be made
design parameters for an optimization study.
When we discussed the barriers to optimization, we mentioned that
the long time to build the underlying CAE models was itself an obstacle,
because it left little time for optimization. One of the time consuming
steps has been modeling the spot welds between metal parts. Until
Optimization in the Automotive Industry 51

New styling profile

Original
Mesh

,
Precise Morphing to ...;ei.,Si.'I~~
the new profile

Figure 3.7. Example of a morphed mesh. Source: Detroit Engineered Products, Inc.
52 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

recently, this required that the finite element meshes for the two parts
be aligned, so that weld connections could be made node to node. Any
changes to one part that would require remeshing would then require
remeshing of the other part and then recreating the node-to-node welds.
Recent developments, however, have allowed welds to be modeled and
located independent of the meshes of the parts being connected. This
reduces the time to create the math model by up to 40% (MCS Software,
2001).
Another development that can speed up model building is the intro-
duction of "process automation" software. Much of model building is
repetitive. In many cases, the work for this year's model is essentially
the same as last year's model, except that some of the data has changed.
The idea of process automation is to provide an intelligent tool that will
guide an engineer through this long process, prompting only for data
that has changed, and automating the remaining steps.
One such tool is the EASi-PROCESS product from EASi Engineering.
This software provides a library of "process blocks" that automate many
of the routine steps in building CAE models for use with specific anal-
ysis codes. Using a special development environment (EASi-PROCESS
builder) one assembles these components in a drag-and-drop fashion to
create the sequence of steps that need to be done. The blocks will prompt
the user for any information that is needed, such as the location where
a load is to be applied, but then automate the actual creation of input
decks.
Once the process is built, it is executed using the EASi-PROCESS
Executor. This program guides the user through the planned steps,
displaying results, prompting for input, and so on, as the user works
through the process. Such process automation software can not only au-
tomate the process of building the model, but also the process of setting
up analysis runs or even optimization runs. EASi Engineering claims
that a major automotive manufacturer has achieved a productivity gain
of 3 to 1 using this software.
So far we have talked about new developments that have removed bar-
riers. We will now turn our attention to some of the new developments
that present exciting applications for optimization.
The advent of "vehicle configurators" is a development that uses op-
timization to generate a set of high-level vehicle characteristics that will
achieve desired performance targets. Today, most vehicle configurators
are proprietary. They are based on coarse models that relate vehicle
characteristics (wheel base, powertrain, etc.) to performance measures
(safety, fuel economy, etc.). By applying simple optimization methods
to these coarse models, one can determine what (if any) combination of
Optimization in the Automotive Industry 53

vehicle features will allow us to some specified performance targets. If

this can not be done, the performance targets can be adjusted until a
set of targets is obtained that is both desirable and feasible. This re-
duces the chance that, sometime late in the vehicle program, detailed
engineering analyses will show that the original performance targets can
not be met.
Vehicle configurators can also help carmakers decide which of their ex-
isting vehicle architectures is best for the vehicle. Automakers try to use
only a few architectures, or standard assembly processes, because this al-
lows vehicles to share production facilitates and parts, thereby reducing
cost. The downside, however, is that the choice of any particular stan-
dard architecture will impose constraints on the vehicle. Configurators
can clarify how the choice of architecture (and associated constraints)
impacts the ability to meet performance targets.
Configurators only provide high-level vehicle specifications. Deter-
mining what these specifications imply for the specifications of subsys-
tems is often referred to as the "roll-down of requirements." It is not
easy, and is usually done based on past experience. When it is done
poorly, costly design iterations are required to correct it. As we dis-
cuss next, optimization has the potential to play an important role in
facilitating the roll-down of requirements.
In order to use optimization to help with the roll-down of require-
ments, one must start with a coarse, full-vehicle model. This coarse
model can be build up from a variety of component models. For ex-
ample, some features of the vehicle might be modeled using reduced
versions of subsystem finite-element models. Other features might be
represented using test-based frequency response functions. Still other
features may be represented as carry-over components with some ad-
ditional structure. There is no single recipe here; the goal is simply
to get to a coarse full-vehicle model using component models that can
be constructed quickly and that are not too computationally intensive.
Figure 3.8 gives a picture of such a full-vehicle model.
Once a full-vehicle model is developed, optimization can be used to
vary the characteristics of the components in order to achieve perfor-
mance targets at the total-vehicle level. The wrapper codes mentioned
earlier can be used to perform this optimization. Assuming the opti-
mization is successful, the results will tell us how the subsystems must
perform in order to achieve the desired vehicle-level targets. In short,
the optimized characteristics of the subsystems become subsystem per-
formance targets for detailed engineering design and analysis of the sub-
systems themselves.
54 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

Figure 3.8. An example of a full-vehicle model for use in rolling-down vehicle-level

requirements to subsystem specifications.

Another challenge for optimization is to address the need to make

a design robust to variation. Variation can come from many sources,
including materials, assembly, and usage in the field by customers. The
straightforward way to assess the impact of variation is a Monte Carlo
analysis in which design parameters are varied according to the expected
distributions. From these simulations one can estimate the probability
that a design will meet its target specifications. In a subsequent step,
once can then try to alter the nominal design, or change tolerances, so
that the performance will be acceptable even in the presence of varition.
The difficulty with the Monte Carlo approach is that it extremely
time-consuming, especially if is to be performed multiple times as part
of an optimization run. This problem is especially severe in crash anal-
ysis, where a single run may take from a few hours to several days of
computation time. To address this problem, many people are running
their analysis codes over a carefully selected set of points (experimental
design), fitting response-surface approximations to the resulting input
and outputs, and then using these fast surface approximations for opti-
mization and robust design
For a phenomenon as nonlinear as crash, simple linear or quadratic
polynomials are unlikely to yield approximations that are sufficiently
accurate to be useful. But good results have been obtained using more
sophisticated approximation techniques. Engineers at Ford have been
pursuing the use of response surfaces in crash for several years and have
published papers comparing different approximation methods. In an
early paper, they recommended sampling via a uniform Latin hyper-
cube design and fitting surfaces using a combination of multiadaptive
regression splines (MARS) and kriging (Li et at., 1999). In a later paper
(Yang et at., 2000a), they compared approximations based on MARS,
stepwise polynomial regression, artificial neural networks, and moving
Optimization in the Automotive Industry 55

least squares, and concluded that stepwise regression and moving least
squares perform best. In their latest publication (Yang et al., 2000b),
they seem to have decided in favor of stepwise regression. At General
Motors, good results have been reported using a combination of data
reduction techniques and polynomial surfaces (Harzheim et al., 1997).
In all these cases, Monte Carlo analysis is applied to the surface to
access the robustness of the design. For example, in a crash problem,
variables such as the angle of impact, driver location, and air bag inflator
output might be varied according to their estimated distributions. The
result of this Monte Carlo analysis gives us a distribution for the re-
sponse. The optimization formulation can then acknowledge then need
for robustness by (say) making the objective to minimize (say) the 95%
quantile of an injury index as opposed to minimizing its nominal value.
The response surface approach to engineering optimization has many
benefits besides facilitating robust design. For example, it can handle
analysis codes whose outputs have numerical noise that makes finite-
difference derivatives unreliable. It also makes it easy to speed up the
computation, because all the runs in the experimental design can be done
in parallel. Runs that fail due to numerical reasons do not undermine
the procedure, but only rob it of one data point for fitting the response
surface. Surfaces can also be used to quickly compute trade-off curves
between competing objectives, something which can be quite helpful to
decision makers. Finally, the surfaces fit for a component system can be
used as part of a larger vehicle-level model to facilitate multi-disciplinary
optimization or roll-down of vehicle requirements.
There is, of course, a danger in using response surface approximations
to guide an optimization process. Due to the inaccuracy of a surface, it is
entirely possible that when the computed optimum based on the surfaces
is actually run through the analysis code-the so-called "confirmation
run" -it will not confirm due to inaccuracies in the response surface.
The current method for handling this problem is to simply add this new
point to the data upon which the surfaces are based, re-fit the surfaces
(which will now will be better due to the additional point), and re-
solve the optimization problem on the surface. Iterating in this way
is often sufficient to find a local optimum, but the procedure will not
always work, and more sophisticated approaches have been suggested
(see Jones, 2001).

4. Opportunities for the Field of Optimization

What can researchers in optimization do that would be helpful to the
automotive industry? As far as general algorithm development is con-
56 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

cerned, the industry is mostly in need of algorithms that can handle

black-box, non-smooth, non-linearly constrained problems. It is impor-
tant that the method not assume that the problem functions can be
evaluated at every suggested point, that is, the method must be able
to recover from failed or "bombed" runs. To promote parallel process-
ing, it is also desirable for an algorithm to suggest several search points
per iteration. A feature that is often requested is to not only find the
global optimum, but also find any other local optimum that are close in
function value but quite different in the parameter space. While this is
relatively easy to do in closed form global optimization (via branch-and-
bound methods), the challenge for the field is to determine how to do
this in the case of black-box optimization.
The new wrapper codes present an opportunity for algorithm develop-
ers to quickly introduce new methods to the industry. In particular, the
LMS OPTIMUS wrapper code has a very simple procedure for integrat-
ing user-defined optimization strategies. User optimization algorithms
can also be integrated into Engineous's iSIGHT, though the process is
a bit more cumbersome. I encourage algorithm developers to contact
managers of these codes and volunteer their algorithms (for OPTIMUS,
contact Nick Tzannetakis, [Link]@[Link]; for iSIGHT, contact
Michael Sheh, sheh@[Link]).
The response-surface approach to optimization is, in my view, ex-
tremely promising. It could certainly profit from more work on exper-
imental design methods, approximation methods, and associated opti-
mization methods. Key areas for additional work are in assessing the
accuracy of a surface and in determining how to improve it if the initial
approximation suggests an optimum point that is not confirmed by a
follow up run on the analysis code.
The growing use of vehicle configurators also presents opportunities
for optimization work. Unlike detailed engineering models, the coarse
models used in configurators are often quite simple and in closed form.
Hence, it might be possible to develop rigorous methods that could,
for example, enumerate all vehicle configurations that would meet the
performance targets.
Using optimization for roll-down of requirements is another area that
could profit from additional work. When we discussed this previously,
we described how optimization of a full-vehicle model could lead to sub-
system specifications. But we can hardly expect engineers to design sub-
systems that perform exactly to these specifications. Instead, it would
be desirable to modify the procedure so that the end result is a specifica-
tion range for each subsystem, with the property that if every subsystem
performs within its range, then the vehicle-level targets will be met.
REFERENCES 57

In summary, until very recently, the use of optimization within the

auto industry has been slowed by many practical and mundane barri-
ers. Fortunately, many of these have been eliminated, and the focus is
shifting to more interesting challenges: robust design, multi-disciplinary
optimization, vehicle configuration, and roll-down of requirements. We
are likely to see a fast-growing use of optimization within the industry
over the next decade.

References
Harzheim, Lothar, and Graf, Gerhard. "Shape and Topology Optimiza-
tion in the Automotive Industry," in Proceedings of the Third Interna-
tional Conference on High Performance Computing in the Automotive
Industry (M. Sheh, editor), Cray Research, Inc., 1997, pp. 167-182.
Jones, Donald R. "A Taxonomy of Global Optimization Methods Based
on Response Surfaces." Journal of Global Optimization, 21, 2001,
pp. 345-383.
Li, D. J., Liu, X., Soderborg, N., Sudjianto, N., Vora, N., and Wang, S.
C. "An Approach for Computer Experiment Based Robust Design."
In Hoang Pham and Ming-Wei Lu, editors, Proceedings: Fifth ISSAT
International Conference on Reliability and Quality in Design, 1999,
pages 98-107.
MSC Software. Article on "MSC Nastran.2001" in evolution newsletter,
volume 1, number 1, August 2001, page 1.
Qian, Cheng and Yuan, Charles. " An Integrated Process of CFD Analy-
sis and Design Optimization with Underhood Thermal Application."
SAE paper 2001-01-0637
"Time Saving Robust Design Considering Vehicle Crash Peformance."
Paper presented at the conference Optimization in Industry III, Barga,
Italy June 17-22, 2001. To appear in 1. Parmee, editor, Proceedings of
the International Conference on Optimization in Industry III, Springer,
Berlin, 2001.
Valenti, Michael. "Detroit, We Are Here! Automakers Are Using French-
Born Manufacturing Software to Improve the Machining and As-
sembly of Their Vehicles." Mechanical Engineering Magazine, March
2001.
"Production Line Optimization." Vehicle News Magazine, No. 250, Au-
gust 31, 2001.
Yang, R. J., Gu, L., Liaw, L. Gearhart, C., Tho, C.H., Liu, X. "Approxi-
mations for Safety Optimization of Large Systems." Proceedings of the
ASME 2000 Design Engineering Technical Conferences and Comput-
 58 OPTIMIZATION AND INDUSTRY: NEW FRONTIERS

ers and Information in Engineering Conference, Baltimore, Maryland,

September 10-13, 2000.
Yang, R. J., Akkerman, A., Anderson, D., Faruque, M. and Gu, Lei. "Ro-
bustness Optimization for Vehicular Crash Simulations." Computing
in Science and Engineering, November/December 2000, pages 8-13.

View publication stats