Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

SAE TECHNICAL
PAPER SERIES 2002-01-2043

Aerodynamic Design of High Performance Cars:

Discussion and Examples on the Use of
Optimization Procedures
G. Lombardi
Department of Aerospace Engineering, University of Pisa

F. Beux
Scuola Normale Superiore di Pisa

S. Carmassi
Ferrari S.p.A.

Reprinted From: Proceedings of the 2002 SAE International Body Engineering Conference
and Automotive & Transportation Technology Conference on CD-ROM
(IBAT2002CD)

International Body Engineering Conference & Exhibition and

Automotive & Transportation Technology Conference
Paris, France
July 9–11, 2002

400 Commonwealth Drive, Warrendale, PA 15096-0001 U.S.A. Tel: (724) 776-4841 Fax: (724) 776-5760 Web: www.sae.org
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

The appearance of this ISSN code at the bottom of this page indicates SAE’s consent that copies of the
paper may be made for personal or internal use of specific clients. This consent is given on the condition,
however, that the copier pay a per article copy fee through the Copyright Clearance Center, Inc. Operations
Center, 222 Rosewood Drive, Danvers, MA 01923 for copying beyond that permitted by Sections 107 or
108 of the U.S. Copyright Law. This consent does not extend to other kinds of copying such as copying for
general distribution, for advertising or promotional purposes, for creating new collective works, or for
resale.

Quantity reprint rates can be obtained from the Customer Sales and Satisfaction Department.

To request permission to reprint a technical paper or permission to use copyrighted SAE publications in
other works, contact the SAE Publications Group.

All SAE papers, standards, and selected

books are abstracted and indexed in the
Global Mobility Database

No part of this publication may be reproduced in any form, in an electronic retrieval system or otherwise, without the prior written
permission of the publisher.

ISSN 0148-7191
Copyright © 2002 Society of Automotive Engineers, Inc.

Positions and opinions advanced in this paper are those of the author(s) and not necessarily those of SAE. The author is solely
responsible for the content of the paper. A process is available by which discussions will be printed with the paper if it is published in
SAE Transactions. For permission to publish this paper in full or in part, contact the SAE Publications Group.

Persons wishing to submit papers to be considered for presentation or publication through SAE should send the manuscript or a 300
word abstract of a proposed manuscript to: Secretary, Engineering Meetings Board, SAE.

Printed in USA
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

2002-01-2043

Aerodynamic Design of High Performance Cars: Discussion

and Examples on the Use of Optimization Procedures
G. Lombardi
Department of Aerospace Engineering, University of Pisa

F. Beux
Scuola Normale Superiore di Pisa

S. Carmassi
Ferrari S.p.A.

Copyright © 2002 Society of Automotive Engineers, Inc.

ABSTRACT solutions. In this context, the possibility of evaluating

performances of different configurations is of utmost
In the initial phase of the aerodynamic car project, difficulties importance; however, difficulties arise due to the high number
arise from the large number of parameters involved. A of geometrical parameters involved, which are necessary for
systematic aerodynamic analysis taking into account the effects defining each configuration. A systematic analysis taking into
of all these parameters appears to be difficult. The use of a account the effects of all these parameters is very difficult,
direct numerical optimization technique appears very attractive given the complexity related to both aerodynamic load
to solve this complex survey. In this paper a critical analysis evaluation and the assessment of mechanics, stylist,
of the most significant optimization methodologies is commercial and others requirements.
presented. The characteristics and the use of a specific
optimization tool are also described, to highlight the A direct numerical optimization technique may be
capabilities offered by the optimization approach. The use of satisfactorily employed to find one’s way through this
this procedure to increase the car safety, increasing the vertical complex survey. Indeed, with this kind of aid the designer has
down-load without affecting car performance, is detailed. a great flexibility in the choice of the design variables and the
Furthermore, the problems arising from the present problem may be addressed systematically.
implementation and the relevant indications for a more
efficient and effective optimization procedure are discussed. The activity on the study of optimization has significantly
increased over the last years, driven by advances in
1. INTRODUCTION computational methods and improvement in computer
performances, and this aspect has been particularly significant
The traditional design processes cannot longer be met in the aerodynamic design, especially for land and flying
competitive with the increasing performance requirements and vehicles. Several examples of this trend can be found, for
the economical pressure to increase efficiency of ground instance, in Ref. 1. From the analysis of the existing
transportation vehicles. Current practice is to move the design bibliography, two different aspects can be highlighted.
of complex equipments away from a process involving a
sequence of specialist departments and to emphasize its The first aspect is the need to improve the accuracy and the
multidisciplinary nature through the use of integrated product validity range of the results, to obtain a realistic representation
teams. These commercial trends, together with the immense of the aerodynamic flow; this implies using a sophisticated
volume of design, manufacturing and maintenance data flow solver within the optimization procedures.
inherent to complex modern equipments, demand for a heavily
computerized environment. Multidisciplinary Design and The second aspect is the requirement to obtain the results in
Optimizations (MDO) envisions a parametric description short time.
format of input data, which will generate, for a specific set of
values of the parameters, a new vehicle description that in turn These two aspects have the common targets to achieve an
is used to generate input for Computer Aided Engineering, improvement in the configuration performances and to reduce
including Computational Fluid Dynamics (CFD). the “time to market”. Clearly, they act in opposite directions,
because high accuracy flow solutions imply long computing
The aerodynamic design plays a crucial role in the time. As a consequence, an important research activity on the
development phase of new automotive configurations and, due optimization techniques is in progress, especially in the
to its intrinsic complexity, the designer needs as much aids as aeronautical field (see [2] for a review).
possible to strengthen his/her choices and discard unsuitable
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

In recent years, different methodologies have been proposed to constraints, this kind of method gives poor computational
solve this classical accuracy-time dilemma. In general the performances when the functional is expensive to be
results are procedures with high accuracy coupled to long computed. The adjoint approach, based on control theory or
processing time or, conversely, rapid time responses obtained variational analysis, is an interesting alternative procedure for
with simplified aerodynamic solvers. An example of the first the calculation of the functional derivatives, which may use by
type is the FRONTIER procedure [3], partially supported by means of continuous or discretized differentiation, [2]. Indeed,
the EC as a part of the ESPRIT program. An example of the it permits to largely limit the number of cost evaluations, and
second type can be the HIPEROAD procedure [4, 5], partially thus, to greatly reduce the computational cost. However, the
supported by the EC as a part of the ESPRIT program, where hand-code exact derivation is extremely difficult to be
a simple potential flow solver was used. A recent example that performed due to the complexity of both models and
combines sophisticated flow solvers and efficient optimization discretization. The use of automatic differentiation techniques
techniques for aerospace is the Growth AeroShape project [6] is a promising possible approach to overcome these
(currently running). The technical strategy is to merge difficulties.
together computational fluid dynamics (CFD) and numerical
optimization, thereby facilitating a much broader utilization of Optimization problems for industrial applications are also
these simulation technologies in vehicle design. The potential often characterized by the presence of a large number of
impact of this technology extends across many aspects of constraints and several local minima. In this context, an
vehicle engineering. Fluid dynamic analysis, including heat important drawback of gradient-based methods, independently
transfer, is the basis of design not only for the external shape of the particular gradient computation, is their strong
of the vehicle, but also for the prime mover and power train dependence on the initial configuration. Indeed, the procedure
(cylinders, valves, intake and exhaust systems, transmission, generally stalls in the local minimum closest to the starting
and cooling), passenger comfort and climate control (noise point. To overcome this problem, stochastic global search
reduction, heating, ventilation and air conditioning), and techniques, which do not require the use of gradients (see
subsystems (such as windscreen de-icing). Automatic Hajela [7] for a review), have been considered. In particular,
aerodynamic optimization can be used in the context of any of genetic algorithms (GAs), which are based on an imitation of
these design tasks by helping to achieve the best possible the elitist reproduction process of biological evolution [8],
solution in each case, while simultaneously reducing the appear particularly attractive in a complex context because of
duration of the design cycle and time to market. their robustness. Indeed, GAs have the capability to explore
the entire search space, even if it is non-convex or disjointed,
In any case, it should be clear that such a tool could not be and thus, they escape from local minima to search towards a
employed as a black box, since experience, skillfulness and global solution. For reason, actually, GAs has aroused
engineering assessment must ultimately guide the project. For significant interest in engineering applications.
these reasons, in the first part of this paper we describe the
fundamental aspects of the optimization procedures, both from Generally speaking, evolutionary algorithms as GAs require a
mathematical and engineering point of views, then, we present large number of objective function evaluations: consequently,
a first study in which the capabilities of different design the high time-consuming computational effort is an important
optimization strategies are evaluated and, finally, we describe limitation. To overcome this lack of computational efficiency,
a consistent (computer aided design) approach, and will use it several improvements can be envisaged, as for instance, a
with criticism by emphasizing both its potentials ant its coupling with neural networks [9-11] or the use of parallel
limits. computer architectures [12-13].

2. OPTIMUM DESIGN STRATEGIES IN AN Another possibility is to consider a hybrid procedure, which

INDUSTRIAL CONTEXT combines GAs with deterministic local search methods. One
may proceed defining a two-stage strategy which splits the
2.1 A CRITICAL DESCRIPTION OF THE OPTIMUM optimization procedure in two steps: first, the genetic
DESIGN METHODS - For industrial applications, and in algorithms exploring a large search space may localize the
particular for automotive problems, design optimization global optimum region, then a deterministic local approach,
studies are performed in a difficult context due to complex as gradient-based methods, is employed to accurately reach the
and/or three dimensional geometry, multi-objective problems optimum [14-15]. Indeed, the lack of supplementary
(multipoints and/or multidisciplinary optimization) and information about derivatives (which is characteristic of GAs)
complex modeling. Thus, the main limitation is surely the does not permit a fast convergence in the nearness of the
prohibitive computational cost of the optimization procedure. optimum. More complex hybridization have also been
Among the different methodologies that are being pursued for developed, as for instance in [16], where four different
direct optimization, gradient-based methods, in which a optimizers are used.
specified objective is minimized, are often employed. These
procedures require computations of both objective functional In the hybrid approaches, a critical point is to find a well-
and its derivatives with respect to each control variables. suited criterion to switch from an optimizer to another. In a
GA, new generations are constructed by successively applying
The gradient calculations are usually carried out by using three different operators with specific tasks (selection, cross-
black box finite difference methods which are simply to be over and mutation), and another class of hybridization can also
implemented and can be applied in a rather general context; be devised in which the local search method is directly
however, they require for each control variable at least a cost incorporated in the GA as a new, fourth, operator. For
evaluation. Thus, if a large number of control variables is instance, in [17] a particular mutation is defined by using the
used, to take into account geometry and multidisciplinary gradient of the fitness function, whereas in [18] a few conjugate
gradient iterations are carried out.
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

2.2 ANALYSIS OF THE CAPABILITIES OF DIFFERENT applied, i.e. two gradient iterations are performed, with the
APPROACHES: AN EXAMPLE - To better analyze the probability p g=1/ N s on these selected individuals.
capabilities of the different approaches, they are applied, here,
to a specific example, described in details in [19]; in the The TS approach combines a stochastic pre-optimizer with a
present paper the analysis is focused on the optimization gradient-based final-optimizer. As far as the pre-optimizer is
aspects. Although, in the present example, the design concerned, both the classical GA and the previously described
optimization has not been performed for automotive GOGA have been considered and, in the last case, a bi-
applications but for aeronautical ones, the conclusions of this hybridization has been performed.
study, and, in particular, the choice of the most efficient
strategies, can be directly applied to the methodologies used As previously mentioned, the realization of an efficient
in Section 3. Indeed, the two optimum design studies use the coupling is strongly related to the choice of the switch-over
same methodology, i.e. a simplified aerodynamic solver criterion. The first considered criterion is based upon the
(potential flow in the both cases), few control parameters and a improvement of the objective function. More precisely, the
basic approach based on a gradient method (only this approach first stage is stopped when the objective function of the best-fit
is used in the study of Section 3). individual of the current generation cannot decrease more than
a fixed percentage of a specified number, ng, of generations. A
The analysis of optimum design strategies is focused, here, second switching criterion has been defined looking at the
more particularly on the hybrid procedures. More precisely, variance of the population. Indeed, if the variation of the
two classes of hybridization, i.e. a two-stage (TS) procedure design parameters in the population becomes small, it means
and a modified GA including a gradient operator (GOGA), that the individuals tend to gather around a particular
have been considered in our study. The gradient operator configuration, which, hopefully, is close to the global
genetic algorithm is defined in a way close to that proposed in optimum. Accordingly, the genetic algorithm is stopped when
[18]. As for the two-stage procedure, a good compromise must the current variance, σ (k), becomes less than a specified limit,
be found between a sufficiently accurate exploration of the σlim. In the definition of the population variance, only the most
search space and the computational cost of the GA. Thus, significant variables have been taken into account.
particular care has been taken here in the switching criterion
choice. 2.2.2. Analysis of the results - The study has been performed
with six different runs characterized by different starting
2.2.1. Description of the procedure - The optimization populations, in order to make a comparison, between the
procedure involves a deterministic search approach, a genetic different algorithms, less dependent on a particular
algorithm and different kinds of hybridization. The initialization. For each run, N individuals (here, N=20) are
deterministic approach is based on a gradient method, which randomly chosen in the space of feasible solutions and used as
uses conjugate directions when the constraints are satisfied and initial population for all the stochastic approaches. For the
feasible directions otherwise. For the genetic algorithm, a deterministic algorithm, i.e. the gradient-based method, the
classical binary string codification has been adopted in which best-fit individual of the first GA generation has been assumed
each optimization parameter value is converted into a string of as starting point.
15 bits. Accordingly, a chromosome-like string, obtained by
the codification of the design variables, represents a particular Table 1 – Search methods and switching
configuration. The selection has been performed by a biased criteria for the different cases
roulette-wheel, based on chromosome fitness. This first
operator does not create new individuals but just yields a
resulting population with a higher average fitness value. In Ca Short name Description Algorithm
order to increase the diversity among the population, a se
uniform crossover operator is applied with a probability of a GC Conjugate Gradient Deterministic
pc =0.5. Finally, to avoid a premature loss of potential zones b GA Genetic Algorithm Stochastic
of interest in the design search space, a mutation operator is c GOGA Gradient Operator Hybrid
used with a probability of p m = 1/N where N is the population Genetic Algorithm
size. d TS (1) Classical Two-
Stage hybridization Hybrid
In the GOGA algorithm, besides these three classical using strategy 1
processes, a fourth operator is also applied, in which a few e TS-GOGA TS method with
iterations of the local gradient-based search method are (1) GOGA as first step Bi-Hybrid
performed. This additional operator gives the possibility to (strategy 1)
locally improve the fitness of the individuals. The local search f TS-GOGA TS method with
method works in the real space of the design variables and, (2) GOGA as first step Bi-Hybrid
according to [18], an operation of decoding and encoding is (strategy 2)
performed to switch from binary to real and vice versa. The
g TS-GOGA TS method with
use of the gradient-based method should be marginal in the
(3) GOGA as first step Bi-Hybrid
algorithm, which remains strongly stochastic: indeed, as for
(strategy 3)
mutation, this operator is introduced with a low probability of
occurrence. The gradient operator is applied to each generation
as follows. First, the Ns best-fit individuals of the current
population are selected (Ns=0.3 N has been chosen, i.e. the For the two-stage hybridization, three different strategies, i.e.
best 30% of the population). Then the gradient operator is three different switching criteria, have been tested. Two of
them are based upon the improvement of the objective
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

function. More precisely, the switching criterion is activated For the simple GA and the GOGA, the stop criterion was
when the objective function cannot decrease more than a fixed simply chosen by a priori fixing the number of iterations.
percent (in our case 0.14%) over ng generations: strategy 1 Accordingly, the value reported in Fig. 1b, is not the total
corresponds to ng=3 and strategy 2 corresponds to ng =4. The number of evaluations effectively done, but an a posteriori
last case (strategy 3), based on the population variance, is estimate indicating when the objective function has reached a
characterized by an upper limit of the variance value, stable value. As an example, Fig. 2 shows, for the different
σlim =0.03. A number of combinations of the different search approaches, the convergence behavior of two runs.
methods and switching criteria, defined in Table 1, have been
analyzed and results are summarized in Fig. 1. In Fig. 1a the
final best objective function values, obtained with the previous
strategies for each of the six different runs are reported, while
in Fig. 1b the corresponding computational costs are plotted.
In order to be able to compare deterministic and stochastic
methods, the cost is expressed as the total number of objective
function evaluations. Note that, for the gradient-based method,
the cost of the construction of the GA first generation is also
included in the total cost.

a) Example #1

a) Final best objective function values

b) Example #2

Fig. 2 - Examples of convergence history

A few points are worth noting. For the gradient-based

approach, the local minimum is rapidly reached and is often
located on a constraint boundary. The value obtained is
strongly dependent on the initial solution. Indeed, as it is
shown in Fig. 1a, a rather large scattering of the solutions is
obtained. Conversely, the GA method provides almost the
b) Computational costs “same” configuration, noticeably better than the results
obtained by the deterministic local search method, even if
Fig. 1 – Analysis of different approaches characterized by high computational costs.
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

The overall performance of the genetic algorithm is the most significant aspects of the procedure are presented, and
significantly improved by the addition of a gradient operator. the capability of the methodology is shown by means of a
Indeed, GOGA substantially reduces both the objective specific application.
function value and the number of function evaluations. The
obtained solutions are slightly more scattered than the 3.1 DESCRIPTION OF THE METHODOLOGY - In the
corresponding ones obtained by the simple GA, but the analysis through direct numerical optimization, an
scattering is small and the final configurations are always aerodynamic code is coupled with an optimization routine,
better. giving rise to an iterative procedure which is able to
automatically manage the values of the design variables –
The two-stage hybridization, in which strategy 1 is used for typically concerning geometry modifications – by minimizing
the switching criterion, has been tested using as pre-optimizer a given scalar quantity (the objective function). This approach
both the simple GA (case d) and the GOGA (case e). The bi- is extremely flexible, and capable of meeting multidisciplinary
hybridization is a little more expensive than the classical two- requirements.
stage hybridization and gives a larger dispersion in the final
results. However, in case e the final values of the objective
function are globally smaller with, in the best cases, values
close to those obtained using the simple GOGA (case c) and a
large generation number. As an example, Fig. 2a shows that
the convergence behavior of cases e, f and g (bi-hybridization
cases) are close to that of a simple GOGA with a suitable
stop. Indeed, when the switching is performed, the
convergence curve is just arrived on the final plateau, i.e. after
this point the objective function decreases only very slowly.
Thus, at this step, the gradient-based final-optimizer fails to
find a better solution or gives a little improvement. On the
other hand, Fig. 2b shows a particularly favorable situation
(case e) in which the convergence towards a suitable solution
is rapidly reached. However, this is not a general behavior:
other runs (not shown here) may be characterized by an early
stop of convergence for case e. By comparing cases f and g, it
should be noted that case g is slightly more expensive, but
gives a result scattering lower than that obtained in case f.
However, the specified variance limit appears difficult to be Fig. 3 - The optimization loop
settled, because of its sensitiveness to different parameters, as
for instance, the number of variables involved, the population In Fig. 3 a flow chart of the optimization loop is shown; the
size or the type of problem. main components will be shortly described in the following
sections.
The previously described behavior of the different approaches
has been also confirmed by other studies performed to test 3.1.1 CAD input and repair module - The solver acts on a
both the mesh sensitivity and the choice of the type of initial surface mesh derived from a car description provided by the
configuration. engineering department. This CAD description is generally
made of a number of separated patches (as many as a few
In summary, strategy f, i.e. the TS-GOGA approach with hundred), describing different parts of the car. The
strategy 2 as switching criterion, appears to give a good optimization procedure requires a single surface, which must
compromise between costs and effectiveness in reaching the be obtained from the separate patches, correcting geometry
global optimum. Also, it has the great advantage (in imperfections in an automatic way; it is important to note that
comparison to case c) of giving a consistent stop criterion this is one of the most critical aspects in the optimization
which is of great importance for the user. Indeed, due to the process. This task is accomplished by CADRE (CAD input
lack of theoretical results, the stop criterion is often chosen by and REpair) module. This module was developed to produce
simply a priori fixing the number of generations, spending a single surface, described by Bezier points, [5]. This surface
unnecessarily computational time, or on the contrary, causing can be subsequently modified and meshed during the
a too early stop. optimization loop.

3. AN EXAMPLE OF APPLICATION TO 3.1.2 Aerodynamic module - A loop involving several

AUTOMOTIVE PROBLEMS hundreds of optimization steps is possible only if the
evaluation of the pressures over the surface of the car is very
A preliminary application of the optimization approach was fast and accurate. Consequently, a “potential flow” model was
developed. The project, named HIPEROAD (HIgh assumed, which is well suited for evaluating the flow around
PErformance Roadvehicle Optimized Aerodynamic Design), aerodynamic bodies. Naturally, this implies that the present
had the object to make the optimization loop in large part version of the procedure may be applied to study the flow
automatic, in order to reduce the time needed to evaluate and around a car only if a «streamlined» shape characterizes it,
improve the sketch design provided by the stylists. with the boundary layer separation occurring only at its
rearmost end. This may be considered to be the case for the
The complete procedure is described in [4, 5]. In this paper high performance sport cars. However, the effects of the
separated wake on the flow upstream of the separation must be
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

taken into account, by some means, to accurately predict the through a deterministic spline algorithm. However, the set of
pressures on the attached flow region. Bezier points is too large to be a good set of variables to be
changed during optimization. Moreover, changing a single
Previous researches [20] demonstrated that the pressures acting Bezier point induces modifications confined in local portions
on the portions of bluff bodies characterized by attached flow of the geometry, thus easily resulting in parameter evolutions
can be accurately evaluated, by means of a potential flow code, that are not acceptable from the point of view of the constraints
practically up to the separation region. This if the separation and of the general style of the car.
is positioned only at the rear end of the body, is practically
fixed by the geometry, and the wake is modeled as a closed The solution has been the introduction of additional, fewer
continuation of the body and treated as a solid surface with structures, which allow the Bezier points to be controlled,
unknown pressures. The validity of the present approach is producing changes that are more extended over the car
also confirmed by the results shown in [21]. During the geometry. The implemented control structures are points and
project an effort has been made to develop a model for the lines, whose displacements drive those of the Bezier points in
wake, which depends on the characteristics of the car surface a prescribed way.
and which allows the vertical loads up to the end of the car to
be accurately predicted [22]. The hierarchy of the geometry moves from the control
elements, C (points and lines), to the Bezier points and,
The used potential flow code [23, 24] is based on Morino’s finally, to the mesh; while control points can move in three
method [25], which is widely used for the evaluation of the dimensions, control lines can be moved only along two
loads on aerodynamic bodies, and is characterized by dimensions; their role is typically to allow geometry
robustness to variations in surface discretization. As already modifications that are coherent along the whole width or
pointed out, in order to take the effects of the separated wake length of a macro panel element.
into account, a fictitious afterbody is added to the portion of
car characterized by attached flow. The optimizer code acts displacing the control elements.
These displacements change the position of the Bezier points
Wind tunnel tests were carried out to validate the whole with a prescribed influence curve, for instance a gaussian. By
system [4, 5]. A first model was constructed and tested in the tuning the width of the gaussian for the different control points
Ferrari Auto wind tunnel. Force measurements and pressures it is possible to change the locality of the elementary
measurements, with 90 pressure taps, were carried. optimizer steps. More details can be found in [5].

The comparison between the experimental and calculated 3.1.3.2 Cost function - The design variables used in the
pressure coefficients obtained in the longitudinal plane for the optimization process are the Bezier points defining the
upper surface of the car is shown in Fig. 4. As can be seen, the external shape of the car, by means of the hierarchy structure
comparison is very satisfactory. Furthermore, it is seen that, defined in the previous paragraph.
as predicted, the agreement extends up to the very end of the
model, demonstrating a good performance of the wake model. By integrating the pressures acting on the car surface (i.e. on
the car forebody and base), previously computed by the
aerodynamic solver, it is possible to obtain the pressure forces
acting on the car. For this evaluation the lower surface of the
car may or may not be considered, according to the choice of
the user.

The quantities evaluated by the solver, and thus involved in

the optimization process, are Vertical Force, Pitching
Moment, Vertical Force acting on the fore and rear axles,
Pressure Drag force acting on the forebody part of the car and
Base Drag.

The cost function is defined as a linear combination of these

quantities, with the weights given by the user.
Fig. 4 - Comparison between experimental and
calculated pressure coefficients 3.1.3.3 Minimization algorithm - Once the cost function Γ is
defined, by means of a linear combination of the quantities
3.1.3 Optimization module - The optimization module has evaluated by the aerodynamic module, its minimization
the purpose of defining a new car geometry that minimizes a proceeds by a steepest descent gradient algorithm: we note
cost function related with certain aerodynamic characteristics of that each control point counts for 3 degrees of freedom (DOF)
the car, while keeping a set of constraints into account. The and each control line for 2 DOF. Hence we have:
constraints in turn are of two kinds: some are related to
aerodynamic quantities, and some derive from design NDOF = 3 NP + 2 NL
limitations.
The gradient of the cost function along each DOF is
3.1.3.1 Degrees of freedom for geometry update - The mesh of approximated by a finite difference. The direction d=(-∇Γ), is
the car surface, namely a set of points and a connectivity chosen for an optimization step; therefore, the new value of the
matrix, is completely determined by a set of Bezier points control point vector is
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

Ci+1 = Ci + µ d obtained together with a slight reduction in drag force, was

considered to be more than satisfactory.
where the parameter µ controls the size of the step.
In conclusion, the validation tests showed that the code is
As described in the previous paragraphs, the steepest descent capable of satisfactorily carrying out its required task, i.e.
algorithm, adopted for the minimization of the cost function, producing in a very small time a modified configuration with
is not in general the most effective. However, it presents the small geometrical modifications but technically significant
advantage that at each step the evolution of the car shape can improvements in the aerodynamic loads.
easily be followed and understood in terms of the geometry of
the cost function surfaces, thus allowing the designer to gain 3.3 EXAMPLES OF APPLICATION - The described
insight in the aerodynamics, which is one of the objectives of procedure of optimization was used in the development of the
this project, and, therefore, it was assumed as the most Ferrari 360 Modena. The basic geometry analyzed is shown
appropriate in an initial stage of the development of an in Fig. 5.
optimization procedure.

3.1.3.4 Constraints - To carry out the optimization procedure

following chosen requirements, several constraints may be
assigned to the different quantities that are evaluated.
Moreover, the modifications to the external shape of the car are
not completely free, but they are clearly limited by esthetic
and functional requirements (volume specification for the
engine, passengers, wheels, etc.). These constraints are
considered in the optimization procedure by defining certain
specific zones of the car and imposing, with respect to the
initial car geometry, a maximum displacement and/or a
maximum mean displacements, for each defined zone. The
choice of the constraints to be activated is left to the user.

The geometrical constraints are in turn defined as a

modification of the cost function Γ. A repelling “potential” is Fig. 5 - The basic geometry for the analysis of
introduced, which results in a very small modification of Γ the 360 Modena
when the mesh is inside the boundaries, while close to the
boundary and outside a high gradient drives the evolution As a first example, a run carried out with the vertical load as
back within the limits. More details can be found in [5]. cost function is considered. The car could be modified on the
upper surface, with a centered longitudinal control line and
Aerodynamic constraints can be activated for each of the one control point, placed on the car side, at a position
aerodynamic quantities evaluated by the solver. These corresponding to the rear axle: 5 degrees of freedom were then
quantities are clearly available only at the end of a certain step used. A maximum number of 100 iterations were considered,
and, therefore, these constraints are not included in the cost and, in a first run, no constraints were taken into account.
function, and are checked on the output of each optimization
step, by means of a recovery procedure described in [5]. The results show a significant increase in the download: at a
speed of 280 Km/h, after 100 iterations there is an increase in
3.2 EXPERIMENTAL VERIFICATION - The HIPEROAD the download of about 1200 N.
code was then used to produce, starting from the previously
described geometry, a new optimized shape having as cost Unfortunately, this solution can not be considered a suitable
function the reduction of the vertical load. However, the one: indeed, the modification to the geometry, which is not
allowed geometry modifications were strictly constrained constrained, is found to be not acceptable. As could be
(maximum allowed displacement of 3 cm at full scale), in foreseen, it is evident that, to have a suitable solution, the
order to obtain a new shape maintaining all the main stylistic constraints (in particular the geometric one) must be activated.
features of the original one, and a constrain on the drag (that
could not increase) was imposed. A calculation on the same case, with a geometric constraint
defined by a maximum displacement of the car surface of 5 cm,
It is important to note that the entire optimization procedure was then carried out. The final shape (after 100 iterations)
was carried out within one day. The obtained shape produced showed a very small modification in the car shape, and a final
a variation of the vertical force coefficient due to the pressures shape that is acceptable from the esthetic point of view.
acting on the upper part of the car of _CZ=-0.042. A wind
tunnel model with this geometry was then constructed and The results in terms of download, for both the described
tested with the same methodology already used for the analyses, are shown in Fig. 6. It is possible to note that up to
original model. The experimental results gave _CZ =0.035, iteration 40 the behavior is not affected by the constraints.
with a reduction slightly smaller than predicted, and After this iteration the geometric constraint is violated, and
CD=0.181, smaller than 0.185 given by the original shape. the increases in download became less important, being
Considering the very small geometrical modifications with smaller the modifications allowed to the geometry. In any case
respect to the first model, the reduction in vertical upload, we can observe an increase in the download, after 100
iterations, of about 1000 N at 280 Km/h.
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

Fig. 6 – Object function for basic and Fig. 7 - Pressure drag (Forebody drag + Base
geometrically constrained evaluations drag) on the car

Clearly, it is necessary to verify also the other aerodynamic

characteristics, for instance the drag or the axles balance. As an
example, the pressure drag is reported in Fig. 7, together with
the object function for a better comprehension. As can be
observed, this quantity decreases up to iteration 70, and,
successively, increases.

It is interesting to note that a single operator carried out the

described analyses in two days. At this point different
strategies can be used, depending on the specific problem.

It is possible, for instance, to make a further optimization run

with the same or a different cost function, taking both the
vertical force and the pressure drag into account. Another
possibility is to retain the shape obtained after 100 iterations,
characterized by a high download and a not particularly low
drag, or an intermediate shape. For instance, the one
corresponding to iteration 70, characterized by an increase in
the download with respect to the initial shape, and the
minimum value of the pressure drag.

A further step could be to impose a defined load distribution

between the wheels axles, that is a constraint of aerodynamic
type. Clearly, this parameter changes in the optimization
process. A calculation of the same case, with the same Fig. 8 – Object function for basic and
constraint on the displacement of the car surface, and an aerodynamically constrained evaluations
additional constraint on the load distribution (variation in the
fore/total load less than 1%) has been carried out. The final It should be noted that the results of the presented procedure
shape (after 100 iterations) shows again a very small cannot be assumed as definitive, but must be critically
modification in the car shape, but the improvement in the reviewed, since a significant number of different configurations
object function, as can be seen from Fig. 8, appears must be considered and several points need to be more
significantly less important. This is important information to accurately analyzed. The time necessary to make a procedure
the designer, i.e. that the significant increases in the of this type: a complete optimization was carried out in about
performances are related to modification in the longitudinal one week, giving an important support to the analysis of the
load distribution. aerodynamic characteristics of the car.

Finally, we can stress that, from a practical point of view,

although many parameters must be constrained, it is
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

interesting to obtain quick estimates of the order of magnitude REFERENCES

of the improvements, which would be related to a relaxation of
these constraints. In this case, the designer has a powerful tool 1. RTO Meeting Proceedings 35, “Aerodynamic Design and
to estimate the real advantages given by the relaxation of a Optimization of Flight Vehicles in a Concurrent Multi-
certain parameter, and this aspect can eventually be considered Disclipinary Environment”, June 2000.
in the design. 2. Newman III J.C., Taylor III A.C., Barnwell R.W.,
Newman P.A., Hou G.J.-W., "Overview of Sensitivity
4 CONCLUSIONS
Analysis and Shape Optimization for Complex
Aerodynamic Configurations", Journal of Aircraft, Vol.
A study of the capabilities of different optimum design
strategies has been presented. The hybrid procedures, which 36, No. 1, Jan. 1999, pp. 87-96.
combine stochastic global search and deterministic local 3. Fenwick S. V., Harris J.P., “The Application of Pareto
search techniques, are more particularly considered. In an Frontier methods in the Multidisciplinary Wing Design
industrial context, the computational cost of a design of a Generic Modern Military Delta Aircraft”, RTO
optimization procedure appears as a critical point and, thus, a Meeting Proceedings 35, “Aerodynamic Design and
suitable strategy must give a good compromise between costs Optimization of Flight Vehicles in a Concurrent Multi-
and effectiveness in reaching the global optimum. As Disclipinary Environment”, June 2000.
expected, an important improvement on the final optimized 4. Lombardi G., Manacorda G., Paap H.G., Vicerè A.,
solution is obtained using Genetic Algorithms instead of “Optimized Aerodynamic Design for High Performance
classical gradient-based methods. Moreover, the hybridization Cars”, AIAA Paper 98-4789, 7th
process applied with an adequate switching criterion also
AIAA/USAF/NASA/ISSMO Symposium on
permits to considerably reduce the computation costs. The
efficient strategies and consistent stop criterions presented in Multidisciplinary Analysis and Optimization, Sept.
this study can be directly incorporated in car optimization 1998, St. Louis.
procedure like the aerodynamic design optimization described 5. Beccaria M., Buresti G., Ciampa A., Lombardi G.,
in the last part of the paper. Gentzsch W., Paap H.G.,Vicere' A., “High Performance
Road-vehicle Optimised Aerodynamic Design:
By using the developed procedure it is evident that the Application of Parallel Computing to Car Design”,
reduction of industrial costs is significant: in principle it is no Journal of Future Generation Computer Systems, 1999.
longer needed to build many different models, to be subject to 6. Bischof C.H., Carle A., Corliss G.F., Griewank A.,
wind tunnel measurements: it is sufficient to test the final Hovland P., "ADIFOR: Generating derivative code from
optimized ones. FORTRAN programs", Scientific Programming, Vol. 1,
1992, pp. 11-29.
The described optimization procedure satisfies the main 7. Hajela P., “Nongradient methods in multidisciplinary
requirements given by the end user. In particular, it is capable
design optimisation - status and potential”, Journal of
of yielding shapes having more favorable aerodynamic
characteristics with very small geometry modifications, thus Aaircraft, Vol. 36, No. 1, Jan. 1999, pp. 255-265.
keeping the style substantially unchanged. Obviously, by this 8. Holland J.H., “Adaptation in natural and artificial
means it is not expected that very high decreases of the systems”, Univ. Of Michigan Press, Ann Arbor, MI,
vertical load may be obtained. The objective is to find rapidly 1975.
the best configuration within a certain small range of shapes 9. Giotis A.P., Giannakoglou K.C., Périaux J., “A reduced-
that are considered to be acceptable from the style point of cost multi-objective optimization method based on the
view, without resorting to time-consuming and expensive Pareto front technique”, Neural Networks and PVM.
wind tunnel tests. In other words, while it is relatively easy, European Congress on Computational Methods in
for an experienced operator, to devise significant changes to a Applied Sciences and Engineering, ECCOMAS 2000,
shape producing substantial increases in the vertical load, but Barcelona, September 2000.
also considerably altering the style of the car, it is almost
impossible to predict the effects of very small modifications of 10. Karakasis M.K., Giotis A.P., Giannakoglou K.C.,
the geometry. In any case, the cost and the time for a wind “Efficient genetic optimization using inexact information
tunnel campaign for this purpose would be prohibitive. and sensitivity analysis. Application in shape
optimization problems”, European Congress on
The use of more complex mathematical procedures, as Computational Methods in Applied Sciences and
described in paper, could give significant improvements in the Engineering, ECCOMAS 2001, Swansea, September
optimization procedure. 2001.
11. Tse D.C.M., Chan L.Y.Y., “Application of micro
Finally, It should be emphasized that a numerical procedure of genetic algorithms and neural networks for airfoil design
this kind can never be regarded as a completely automated optimization”, Symposium on Aerodynamic Design and
tool; user’s experience and external control on the process are Optimisation of Flight Vehicles in a Concurrent Multi-
always needed to obtain the best results. Disciplinary Environment, Ottawa, October 1999. RTO-
MP-35, 2000; pp. 23-1/23-11.
 Downloaded from SAE International by Univ of Nottingham - Kings Meadow Campus, Wednesday, September 19, 2018

12. Marco N., Lanteri S., Desidéri J.A., Périaux J., “A 19. Lombardi G., Mengali G., Beux F., “An hybrid genetic
parallel genetic algorithm for multi-objective optimization based optimisation procedure for aircraft conceptual
in computational fluid dynamics”, Evolutionary analysis”, submitted to Aircraft Design.
Algorithms in Engineering and Computer Science, 20. Tabacchi L., “Valutazione delle Pressioni su Corpi di
Miettinen et al Eds, John Wiley and Sons, 1999, pp. interesse automobilistico: Applicazione di un Codice di
445-456. Flusso Potenziale e Sperimentazione in Galleria”, Thesis
13. Lee J., Hajela P., “Parallel genetic algorithm in Aeronautical Engineering, Università di Pisa, June
implementation in multidisciplinary rotor blade design”, 1993.
Journal of Aircraft, Vol. 33, No. 5, Sept. 1996; pp. 21. Gruen N., “Simulating External Aerodynamics with
962-969. CARFLOW”, SAE Paper 960679, 1996.
14. Syrjakow M., Szczerbicka H., “Efficient parameter 22. Buresti G., Lombardo D., “Prediction of Pressure Loads
optimization based on combination of direct global and on Streamlined Bluff Bodies through Potential Flow
local search methods Evolutionary Algorithms”, the IMA Solvers”, Proc. XIV National Congress of the Italian
volumes in Mathematics and its Applications, Davis et al Association for Aeronautics and Astronautics, Napoli,
Eds, Springer 1999; 111, pp. 227-249. 1997, Vol. I, pp. 63-72.
15. Poloni C., Giurgevich A., Onesti L., Pediroda V., 23. Polito L., Lombardi G., "Calculation of Steady and
“Hybridization of a multi-objective genetic algorithm, a Unsteady Aerodynamic Loads for Wing-Body
neural network and a classical optimizer for a complex Configurations at Subcritical Speeds", AIDAA Conference
design in fluid dynamics”, Comput. Methods Appl. Proceedings, Napoli, 1983, Vol. 1, pp. 209-222.
Mech. Eng., 2000; 186, pp. 403-420. 24. Baston A., Lucchesini M., Manfriani L., Polito L.,
16. Dulikravich G.S., Martin T.J, Dennis B.H., Foster N.F., Lombardi, G., "Evaluation of Pressure Distributions on
“Multidisciplinary hybrid constrained GA optimization an Aircraft by Two Different Panel Methods and
Evolutionary Algorithms in Engineering and Computer Comparison with Experimental Measurements", 15th
Science”, Miettinen et al Eds, John Wiley and Sons, ICAS Congress, London, Sept. 1986.
1999, pp. 233-259. 25. Morino L., Chen L.T., Suciu E.O, “Steady and
17. Burczynski T., Beluch W., Dlugosz A., Nowakowski oscillatory subsonic aerodynamics around complex
M., Orantek P., “Coupling of the boundary element configurations”, AJAA Journal, Vol. 13, No. 3, March
method and evolutionary algorithms in optimization and 1975, pp. 368-374.
identification problems. European Congress on
Computational Methods in Applied Sciences and
Engineering, ECCOMAS 2000, Barcelona, September
2000.
18. Vicini A., Quagliarella D., “Airfoil and wing design
through hybrid optimization strategies”, AIAA Journal,
Vol. 37, No. 5, May 1999; pp. 634-641.