Design Optimization of a Two-Dimensional Subsonic

Engine Air Intake

N. Hoyle,

N. W. Bressloff,

and A. J. Keane

University of Southampton, Southampton, SO17 1BJ England, United Kingdom

DOI: 10.2514/1.16123
A novel geometry modeling technique is dened for the optimization of pressure recovery through a two-
dimensional subsonic diffuser based on that of a Formula One race car airbox. The airbox design procedure involves
considering the expansion of the owentering the airbox coupled with a bend through 90 deg. Both of these features
are discussed separately in terms of parameterization approaches before the most suitable techniques are united in
the nal optimization study producing an airbox harboring strong local geometric control and with a set of design
variables compact enough to retain optimizer efciency. A global Krig response surface model is employed to
support our optimization studies, comprising design of experiments and updates based upon the expected
improvement to the objective function followed by a local exploration in a reduced area of the design space. We nd
that we can efciently converge to an optimumairbox design. The geometry modeling technique discussed allows for
potentially radical designs with high pressure recoveries.
Nomenclature
A = two-dimensional diffuser width, m
a = HicksHenne bump amplitude, m
C
p
= static pressure coefcient
E = two-dimensional diffuser total expansion ratio, A
e
=A
in
E(x) = local expansion ratio, A(x)=A
in
f = objective function
k = number of dimensions
N = diffuser axial length, m
n = number of observed responses
p = mass-averaged static pressure, Pa
R = correlation between two sample points
r = vector of correlations
T = dimensionless width factor of HicksHenne bump
U = mass-averaged inlet velocity, m=s
x
+
= untried point
x
p
= normalized distance of HicksHenne bump peak
along duct centerline from duct entry, x
p
[0; 1|
= wall-contouring parameter

ij
= Kronecker delta where
ij
=1 if i =j and
ij
=0 if
i j
, p, = Kriging hyperparameters
= mean
^ = maximum likelihood estimator of the mean
= uid density, kg=m
3

2
= variance
^
2
= maximum likelihood estimator of the variance
Subscripts
in = diffuser inlet location
e = diffuser exit location measured at the engine lter
location
I. Introduction
O
VER the past century, the design of engine air intakes has
played a signicant role in both the aerospace and automotive
industries as engines have developed in sophistication and
performance.
Aircraft intake design requires a steady, high-quality owinto the
engine to maintain reliable engine performance through conven-
tional ight to high maneuverability situations.
Automotive intake design seeks to maximize static pressure acting
on the intake stroke of the engine cylinders. High static pressure over
the cylinders increases the cylinder charge density and hence engine
power.
Aerodynamic design for this purpose slowly began to appear on
race cars during the rst few decades of the 20th century. It was
known by this time that diffusers could convert kinetic energy at the
diffuser entry into static pressure at the exit but with a lowefciency.
Early work started on the improvement of general diffuser efciency
in 1938 [1]. In the cases studied here, we focus on the engine air
intakes of race cars, in particular those used in Formula One (F1),
namely airboxes. The design of the airboxs geometry, including its
bend through 90 deg and the position of the air lter element, all have
an impact on cylinder-to-cylinder air distribution and thus engine
performance. Hence, aerodynamicists have studied air intakes on F1
race cars since the 1950s. At that time, engines were positioned in
front of the driver with a small air vent in the engine cover bodywork
over the cylinders. Ten years later, rear-engined cars were introduced
in which the engines were left exposed with no covering bodywork.
By 1972, the teams had designed large scooplike airboxes sitting
above the drivers head. However, safety was increasingly becoming
an issue from whence roll bar structures were introduced. Two large
scoops either side of the roll bar then became the norm, reducing in
size through the early eighties until, in 1989, airboxes appeared akin
to those seen today with a single modest-sized entry above the
drivers head.
These particular air intakes pose an interesting design challenge.
Because of the roll bar specications and the engine layout
conguration [2], they need to expand the ow over a short distance
while turning the owthrough 90 deg. Presently, F1 teams place a 3-
liter V10 engine behind the driver, as seen in Fig. 1 The position of
the airbox thus takes advantage of the ramming effects of the
oncoming air at high speeds. The exit of the diffuser is located over a
trumpet tray at the bottom of which sit an offset array of 10 engine
inlet trumpets.
The race car industry relies heavily on the use of computer aided
design (CAD). Hence, it is clearly important that our studies use
commercially available toolkits within the restraints of an
Received 16 February 2005; revision received 5 June 2006; accepted for
publication 16 June 2006. Copyright 2006 by the authors. Published by the
American Institute of Aeronautics and Astronautics, Inc., with permission.
Copies of this paper may be made for personal or internal use, on condition
that the copier pay the $10.00 per-copy fee to the Copyright Clearance Center,
Inc., 222 Rosewood Drive, Danvers, MA 01923; include the code $10.00 in
correspondence with the CCC.

Ph.D. student, Computational Engineering and Design Research Group.

Senior Research Fellow, Computational Engineering and Design

Research Group.

Professor of Computational Engineering, Chair of Computational

Engineering and Design Research Group.
AIAA JOURNAL
Vol. 44, No. 11, November 2006
2672
appropriate amount of compute power, similar to that existing within
the race teams of today, to ensure that a realistic solution may be
offered. The aimof this paper is to produce a comprehensive study of
the design process for a two-dimensional model of the airbox, with
special focus on geometric parameterization. Much research has
been performed on effectively parameterizing aerofoils by choosing
one, or a number of design variables fromfamiliar parameters such as
camber, planform, twist, shear, and so on. However, at present there
exists no such metric from which we can parameterize an internal
owduct such as an airbox. In this paper, we test a number of simple
techniques within an optimization procedure to determine the most
effective method for this particular industrial case. Here,
optimization is based upon tting a response surface model (RSM)
[3,4] to objective function data generated from CFDsimulations at a
small number of initial points. A design of experiments (DoE)
approach [5] forms the basis upon which these initial points are
chosen. The response surfaces created are used to predict objective
function values throughout the design space and, consequently, to
estimate the position of the optimum. Update points are then chosen
from promising regions of the design space, based on the
improvement expected to our objective function, so as to increase the
accuracy of the response surface t and, thus, improve the estimated
optimum location. For high dimensional problems, it becomes
increasingly difcult to perform large numbers of updates due to the
computational expense. However, after a reasonable number of
update points have been calculated to ensure the current best point is
near a promising optimumvalue, we can perform a local exploration
[6] to speed up the convergence and improve the optimum location
further to obtain a higher optimum value.
We rst discuss geometry parameterization in Sec. II, the design
objective in Sec. III and the detail of our CFD model, followed by a
further discussion of response surface methodology and local
exploration in Sec. IV. In Sec. V, three different geometry
parameterizations have been tested for a two-dimensional straight
diffuser within an optimization process to determine the most
effective method. We investigate the design of two-dimensional
elbows of constant width turning through 90 deg in Sec. VI. Again, a
variety of simple parameterization methods have been tested for use
in optimization. The expansion and bend are then fused together in a
novel parameterization method which is employed within the RSM
optimization procedure in Sec. VII. This approach has strong local
geometric control along the airbox walls allowing the upper and
lower wall to be uncoupled. The design variable ranges are chosen
such that the geometry also allows for designs with small bumps
occurring on the walls. The use of bumps on diffuser walls has been
investigated by Zhang et al. [7] and have been seen to be benecial in
reducing pressure distortion over the exit of a curved subsonic S-duct
diffuser while maintaining a high pressure recovery.
Following the RSMupdate process in Sec. VII, a local exploration
further enhances the accuracy of the model allowing us to quickly
converge to an optimum design. Our approach is successful in
producing an efcient optimization strategy with radical results.
II. Geometry Parameterization
Geometry parameterization plays a signicant role when used in
conjunction with an optimizer. By using a set of variables to dene
the diffuser in such a way that radical designs are capable of being
produced through strong local control on the diffuser walls, the
optimizer can discover unconventional designs with the potential to
produce superior results. It is clear that using a technique with strong
local control is important and our aim is to maximize this control
using as small a number of design variables as possible. This will in
turn allowus to performan efcient optimization with the knowledge
that our geometry is not overly restricted and thus permitting less
intuitive designs.
The literature on geometry parameterization techniques is
substantial. Samareh [8] surveyed the available techniques and
assessed each on their suitability in dealing with complex models. As
we are dealing with curve contours to dene the overall geometry of
our air intake, we must ensure that these curves are smooth. B-spline
and polynomial spline approaches to curve contouring have the
advantages of providing a compact set of design variables and are
naturally smooth. For simple geometries, Bezier curves are equally
as effective with again smooth and accurate properties represented
concisely through a small number of design variables. Braibant and
Fleury [9] demonstrated that Bezier curves are well suited to
geometric parameterization when used in optimization studies and
Farin [10] describes some of the more useful properties of this
particular technique. Hicks and Henne developed a global shape
function to efciently modify aerofoil sections [11]. Each curve can
be smoothly represented through just three design variables: the
amplitude of the bump on the curve, the location of the bump on the
curve, and the height of the bump. We have applied this technique
very effectively to a simple case discussed in Sec. V.
To obtain an understanding of how wall geometry affects the
efciency of expansion and of how the centerline bend affects the
efcient turning of the ow, we have initially considered the two
functions of the airbox separately: that of expanding the owthrough
a straight diffuser and that of turning the owthrough 90 deg with no
expansion. From the conclusions of these studies we choose the best
method for parameterizing our airbox, and this is discussed in
Sec. VII.
III. Design Objective
Our design problem is the maximization of pressure recovery
based on the internal ow through the diffuser. We assume steady,
two-dimensional, incompressible ow and so the dimensionless
static pressure coefcient, referred to as our pressure recovery value,
may be dened by
C
p
=
p
e
p
in
q
in
(1)
where
q
in
=
1
2
U
2
(2)
denotes the dynamic pressure.
IV. CFD Analysis and Optimization Strategy
Our geometry is constructed using the CADengine CATIAV5
TM
(Dassault Systmes) and imported into a meshing tool. Both mesh
and ow simulations are executed using the commercial CFD
package FLUENT [12]. To ensure that solutions yield sufcient
accuracy within FLUENT, a mesh dependency study was performed
on a straight-walled diffuser before the optimization studies. With a
xed inlet mass ow rate of 10:8 kg=s (Re =6 10
6
) and using a
paved quad/tri structure for the mesh, various mesh sizes were tested
and solved with the same CFD model each time. Starting from a cell
count of approximately 1500 cells, the mesh was progressively
rened testing approximately 3, 6, 12, 24, 39, 75, and 115 thousand
cell meshes. The static pressure recovery value was calculated for
each of the different mesh sizes. The percentage difference in C
p
value between the meshes comprising approximately 1500 cells and
the nest mesh was 437%, whereas the percentage difference in C
p
ENGINE
air filter
airbox
roll bar
airbox inlet
(airbox exit)
Fig. 1 Airbox positioning within the F1 car.
HOYLE, BRESSLOFF, AND KEANE 2673
value between the meshes comprising approximately 39,000 cells
and the nest mesh was 1.7%. Hence a mesh with approximately
39,000 cells was chosen for our optimization studies. It is difcult to
control the exact cell count when using an automated process to mesh
each design discovered by the optimizer such as the one described
here. However, the meshes used in this initial dependency study were
created with a size function capability within the meshing tool which
allows the same growth rate fromthe boundary layer to be developed
regardless of the shape of the diffuser. Using this functionality, ner
meshes are produced by decreasing the distance between the mesh
nodes on the diffuser walls while maintaining a consistent boundary
layer depth across meshes. For cell count consistency within the
automation process, the xed boundary layer depth, xed
distance between wall nodes at the value required to produce an
approximately 39,000 cell mesh and a xed cell growth rate mean
that the meshes produced during the optimization process will only
vary a small amount in cell count given varying wall shapes.
Our CFD analysis for all the studies carried out in this paper
involves solving the two-dimensional steady-state Reynolds-
averaged NavierStokes equations based on the assumption of
two-dimensional steady incompressible ow. The k turbulence
model [13] is usedfor our straight diffuser study and our nal study in
Sec. VII which is in keeping with the references studied, and the
SpalartAllmaras turbulence model [14] is applied in the constant
width turning elbow. The change to using a SpalartAllmaras model
is due to the fact that this model is more economic than the standard
k model and more accurate for wall-bounded ows and ows with
mild separation and recirculation [15]. The boundary conditions for
studies conducted in Secs. V and VI comprise a xed mass ow rate
of 10:8 kg=s at the inlet and a pressure outow positioned at the exit
of the duct. The exit of the duct is situated downstream of where the
engine lter would sit, at the end of the diffuser expansion for the
straight diffuser study and at the end of the bend in the elbow study.
The end of this constant width extension is classied as the exit. This
is to ensure that any separation arising within the diffuser does not
pass through the outow boundary. Mass-averaged static pressure
values are taken at the inlet and at the position of the lter for the
straight-walled diffuser and at the end of the bend for the elbow.
For the study carried out in Sec. VII, the large expansion required
over such a short distance coupled with the 90 deg bend means that a
pressure outow boundary condition at the lter would be
insufcient due to the expected unstabilized owat this point caused
by separation. We could extend the diffuser downstream of the lter
creating a long constant width outow duct as carried out in Secs. V
and VI, but this is unrealistic in terms of the nature of the airbox set up
within an F1 race car. To ensure we obtain accurate converged
solutions, we have extended the airbox model to include the engine
lter, represented in FLUENT by a one-dimensional porous jump,
and the trumpet tray which is an area of xed width situated between
the lter and the engine trumpets. The area represented by the engine
trumpets is classied as a velocity inlet boundary condition. A
breathing engine sucks the air out of the airbox through the trumpets
and so the velocity inlet condition at the trumpets has a negative
velocity value. We assume that the four-stroke engine is at wide open
throttle running at 18,000 rpm, i.e., we have 9000 intake strokes per
minute or 150 intake strokes of the pistons per second. The 3-liter
engine requires 0:003 m
3
of air and hence the engine requires a
volume ow rate
_
V =0:45 m
3
=s. This ow rate is also equal to the
product of the total cylinder area and the velocity required by the
engine. Fromthis the engine velocity can be calculated given the total
cylinder area. For the two-dimensional model to be representative of
the real 3-D case, the total trumpet length is calculated by matching
the 3-Dtotal cylinder area tolter area ratio. However, within the 3-D
airbox there is a second expansion ratio of total cylinder area to total
inlet area to consider. For this 2-D case, we cannot consider both.
Thus, the length of the diffuser inlet and length of the lter are xed
so that the complete shape represents the center plane of a 3-Dairbox
and the geometry parameterization techniques of dening the walls
can be potentially carried over into an airbox design strategy in three
dimensions. Then, because our 2-D diffuser is effectively doing half
the work needed to expand the ow through a 3-D airbox for the
required breathing engine velocity, we need to increase the velocity
of the owbeing sucked out of the 2-Dtrumpets to represent sensible
inlet speeds akin to those seen in the 3-D case. The velocity through
the engine trumpets is such that it is required to have Re =2 10
6
. A
pressure inlet is imposed at the diffuser inlet. The porous jump values
used are given by the thickness and permeability of the engine lters
typically used by F1 teams and have the following values:
face permeability =1:3 10
8
m
2
, filter thickness =15 mm and
pressure jump coefficient =1000.
The optimization strategy used in this work is illustrated in Fig. 2.
The purpose of this methodology is to conduct a series of CFD
evaluations of carefully chosen designs in the search for the optimum
design.
A parametric geometry is rst specied. Then, a set of design
points are created using a DoE [5] approach which supplies the
training data for the construction of a response surface model [4]
produced in these studies by OPTIONS [16]. This is a form of curve
t through or near the data. Because we do not have any previous
knowledge of what this surface looks like, it is considered benecial
to use a technique which provides good coverage of the design space.
In our case, we use the LP method [17,18]. These points are
evaluated using the CFDanalysis as described and can be performed
within a parallel computing architecture to reduce the computational
time required.
A reasonable size DoE should be implemented to build a
sufciently accurate model to initiate the update process. Typically,
Build an RSM
Evaluate RSM
search methods using RSM
Population and gradient based
Optimum Design
CFD
CFD
CFD
CFD
CFD
CFD
Best Design
Parallel computing
structure
process
Update
Local Exploration
Construct a set of input geometries using a DoE approach
Design
adequate?
CFD on resulting geometry
Specify a global parametric geometry
RSM modification
via hyperparameter tuning
Fig. 2 Typical optimization strategy.
2674 HOYLE, BRESSLOFF, AND KEANE
10 times the number of design variables is sufcient. For further
discussions on this topic refer to Sbester et al. [19].
A. The Update Process
Having sampled the problem space, a Kriging [20] model is used
to construct the RSM [4], see fourth box in ow chart illustrated in
Fig. 2. The original purpose of Kriging was to use prior knowledge
about the spatial distribution of a mineral which encapsulates the
behavior of mineral distribution within the given sample space [21].
Given these sample values and distributions, Kriging can
approximately predict the level of mineral concentration at
unsampled points. We apply this technique here with the mineral
concentration replaced by our objective function, C
p
.
Kriging is a technique which provides a statistical interpretation so
that, in addition to the predictor, a measure of the possible errors in
the model is ascertained, which in turn may be used to position any
further design points more prudently. In all cases studied in
subsequent sections, the response surface represents an approx-
imation of the expected improvement in the objective function value
that can hypothetically be attained over the design space.
We must be aware that Kriging is not suited to all practical
applications. Its efciency is dependent largely upon the number of
design variables dening the problem and also the data set size. We
have already mentioned in Sec. II that strong local control of the
diffuser walls is important. In many cases, local control can be
strengthened by adding more points through which splines pass, for
example, or control points dening Bezier curves or number of
HicksHenne functions. This would be deemed useful if we were to
be studying an already optimized design and further ne-tuning in
specic areas were required. This is not the case for this airbox study
and so by potentially enhancing the geometry in terms of its local
control strength, the number of design variables, i.e., problem
dimension size, increases. Therefore, we face the decision as to
whether we can trade off the geometric complexity of the model with
the computational time required to obtain an adequate solution.
Typically, Kriging is computationally practicable up to approx-
imately 20 design variables [22]. In this paper, we feel it necessary to
explore the simplest techniques with small numbers of design
variables so that we can compile a parameterization technique for the
nal study which falls within this 20 design variable limit.
Mathematically, the update process is performed as follows. From
the DoE we have a vector of n initial sample points. Each of the n
sample points has a response y determined by the objective function
y =f(x), where in this paper the function f determines the pressure
recovery value through the airbox as dened in Sec. III, and x is a
vector (x
1
; . . . ; x
k
) where k denotes the dimension size of the
problem. Intuitively, assuming continuity of f, the difference
between the responses y(x
i
) and y(x
j
) will be small if the distance
between x
i
and x
j
is small. This has a statistical interpretation being
that y(x
i
) and y(x
j
) are highly correlated if [x
i
x
j
[ is small. We
can express this correlation as
R(x
i
; x
j
) =exp
_

k
s=1
10

s
[x
is
x
js
[
p
s
_
10

ij
(3)
satisfying R =1 10

if x
i
=x
j
and where
s
, p
s
, and are
unknown coefcients and are dened as the Kriging hyper-
parameters which provide a good statistical interpretation on the
quality of the surface being built and, once tuned, they can be used to
rank the design variables in accordance to their relative dominance
[22,23].
Now suppose that our set of responses y =(y
1
; y
2
; . . . ; y
n
)
T
follows a normal distribution with mean and variance
2
. To
estimate the values of ,
2
,
s
, p
s
, and , we look to choose values
for these parameters which will maximize the likelihood of our
responses.
The likelihood can be dened as
1
(
2
)
n=2
(2)
n=2
det R
1=2
exp
_

(y 1)
T
R
1
(y 1)
2
2
_
(4)
In practice, however, it is more convenient to choose the
parameters ^ and ^
2
to maximize the log-likelihood function

n
2
log(
2
)
1
2
log([R[)
(y 1)
T
R
1
(y 1)
2
2
constant terms (5)
as performed in the Evaluate RSM box in Fig. 2.
The maximum likelihood estimators ^ and ^
2
of this are then
dened by
^ =
1
T
R
1
y
1
T
R
1
1
(6)
^
2
=
(y 1 ^ )
T
R
1
(y 1 ^ )
n
(7)
By substituting the values for ^ and ^
2
into Eq. (5) we obtain the
concentrated log-likelihood function

n
2
log( ^
2
)
1
2
log([R[) constant terms (8)
We can maximize this concentrated log-likelihood function to nd
estimators and
^
,
^

s
, and ^ p
s
, s =1; . . . ; k. This is known as tuning
the hyperparameters, see Fig. 2, and by doing this we can increase the
predictive accuracy of the RSM. Furthermore, we can readily assess
the relative dominance of the design variables using the
hyperparameters, log
s
>0, 0 <p 2. For example, a large
log
s
value indicates a function whereby the objective value can
change signicantly over a small distance. p determines the
smoothness of the function, the smoother the function the closer to
2p gets. For an interpolating model =o, or in practice 6 or
less to prevent ill-conditioning of the R matrix. Regression is
allowed to occur for values of greater than 6.
In Secs. V, VI, and VII we tune the hyperparameters for every 10
update points due to the computational expense incurred by tuning,
under the assumption that the character of the surface will remain
largely unchanged at intermediate updates.
When exploiting the RSM, we choose a search method to nd the
best candidate point on the model that maximizes the expected
improvement we could achieve over the predicted f values. This
occurs in the Population and gradient based search methods using
RSM box in Fig. 2.
To arrive at a prediction for the objective function at some untried
candidate point x
+
, an objective function value is estimated and
augmented to our initial n-dimensional data set. An intuitive
predictor of y
+
=y(x
+
) would be the value which maximizes the
augmented log-likelihood function
_
1
2
^

2
(1 r
T
R
1
r)
_
(y
+
^ )
2

_
r
T
R
1
(y 1 ^ )
^
2
(1 r
T
R
1
r)
_
(y
+
^ )
terms independent of y
+
(9)
for which the maximum of this is our Kriging predictor:
^ y(x
+
) = ^ r
T
R
1
(y 1 ^ ) (10)
where r is the vector of correlations r =[R(x
+
; x
1
); . . . ; R(x
+
; x
n
)|
T
.
Expected improvement can be dened as the improvement we can
expect toachieve when the model is sampled at an untried point. Let a
random variable Y ~N[ ^ y(x); s
2
| where ^ y is our Kriging predictor
dened in Eq. (10) and s
2
is our mean square error. For a
maximization problem, let f
max
be the current best objective function
value, then we will achieve an improvement if I =Y(x) f
max
>0.
HOYLE, BRESSLOFF, AND KEANE 2675
The expectation of I can be dened as
E(I) =
_
I=o
I=0
I
_
1

2
_
s(x)
exp
_

[I f
max
^ y(x)|
2
2s(x)
2
__
dI (11)
which can be integrated to give
E(I) =s[u cdf (u) pdf (u)| (12)
where cdf (u) is the normal cumulative distribution function and
pdf (u) the normal probability density function and where
u =( ^ y f
max
)=s.
This is the value which is searched and maximized using a
population based method, the Dynamic Hill Climber [24], to nd the
next best update point.
This iterative updating process continues until we have an
adequate design, i.e., we have obtained a sufciently high pressure
recovery value or we run out of time. Generally, the minimum
number of updates needed to yield a convergent solution has been
found to be roughly 23 times the size of the initial DoE, provided the
DoE contains sufcient data for the initial model to be representative
[19].
B. Local Exploration
An efcient RSM approach will yield a good global
approximation of our objective function. For high dimensionality
problems, a large initial DoE is performed but as the number of
update points increases, it becomes impracticable to build and tune
the RSM using Kriging. Once we have reached the limit of sensible
RSM construction times we can move to performing local
exploration instead [6]. Provided a reasonable number of update
points have been calculated we can ensure that the current best point
is near a promising optimum value. We can then perform a local
exploration at this point to speed up the convergence upon an
optimum value and improve the optimum location further.
Sections Vand VI do not contain studies with high dimensionality
problems and so their optimization strategies conclude at the Best
Design box illustrated in Fig. 2. However, the high dimensional
problem studied in Sec. VII means that the limit of practical RSM
building times is reached and hence the continuation of the strategy
into performing a simple local exploration is warranted.
For the combined study, the search of the design space up until this
point has been performed within a k-dimensional hypercube dened
by the upper and lower bounds of the k design variables. The design
space in which the local exploration is performed in this study is xed
in a reduced area of the design space. A small percentage of the
design space for each variable is taken around the current best point.
We choose this exploration region to be 20% of each of the design
variable ranges with the center of this smaller hypercube at the
current best point. Within this exploration region we construct a
regional set of geometries via a further 50 point LP DoE and
evaluate these points using the CFD code to obtain their objective
function values. No RSMis built using these values as the purpose of
this region is to nd potentially superior designs through the dense
sampling of points within a small area. Efcient local convergence
can be assured through this approach.
V. Two-Dimensional Straight Diffuser
Early experimental work classied the major ow regimes within
straight diffusers [25,26]. Relationships were deduced between these
ow regimes and the diffuser characteristics [27] while,
concurrently, a simple geometry parameterization gave room for
more efcient designs [28].
Once sufcient computer power became available in the early
1970s, computational design through shape optimization and
prediction of uid ow behavior became feasible. One of the rst
computational solves of an optimal condition described by
mathematical theory was performed by Glowinski and Pironneau
[29]. Increased levels of compute power combined with the
enhanced efciency of optimization techniques has enabled us to
search larger design spaces and many more designs. Research has
been carried out on the optimal shape design of two-dimensional
diffusers in turbulent ow using alternative methods to provide an
accurate prediction of the owseparation [30,31]. These cases began
from a widely studied optimal diffuser design.
As we start with no previous optimized design, we rst center our
attention on that of a straight diffuser so as to understand the effect of
wall geometry on the diffuser. In this section, we consider three
different parameterization techniques for diffusers with straight
centerlines. Our two-dimensional diffusers have a total expansion
ratio E =4:5 and aspect ratio N=A
in
=1:6. The ow regimes
classied by Fox and Kline [26] established a correlation between
diffuser performance and ow separation occurrence for varying
expansion ratio and aspect ratio values and given Reynolds number,
Re =1:6 10
5
. Ideally, to obtain the best possible diffuser
performance, the diffuser would be designed so that no areas of
separation occurred. This is because separation regions obstruct the
effective ow passage and therefore the diffuser would experience
increased losses from the excessive nonuniformity of the exit ow.
However, if we were to impose a straight wall for our diffuser, the
expansion and aspect ratios are such that, referring to Foxand Klines
ow regimes, we would expect to see an inevitable onset of
separation. Hence, the purpose of this parameterization exercise
within the design optimization strategy is to allow contouring of the
wall so that any separation could be reduced or even eliminated given
a Reynolds number of 6 10
6
.
Wall contouring was rst tested for two-dimensional straight
diffusers using experimental means by Carlson et al. [28]. It was
found that for E [1:5; 4:5| and N=A
in
[3; 18|, bell-shaped
optimum wall geometries returned the highest C
p
value within their
optimization study. Madsen et al. [32] studied this problem using
CFD and modern optimizer codes, maximizing the static pressure
rise, adopting a B-spline parameterization with ve master points
along the wall. They compared computed results through the use of
CFDwith those obtained experimentally by Reneau [27] imposing a
recreation of the same inlet conditions for a straight-walled diffuser.
It was shown that for the straight-walled benchmark diffuser, the
CFD code consistently overestimated the pressure recovery values
found via experiment, explained by the use of a stationary CFD
model where owseparation occurs. However, it was conrmed that
with the diffuser geometries falling within the bounds E [1:5; 3|
with a constant aspect ratio N=A
in
=3, Madsen et al.s parameter-
ization technique coupled with their optimization strategy produced
bell-shaped diffusers which were found to return the highest static
pressure recovery (C
p
values), the same conclusion made by Carlson
et al. [28].
With the CFD analyses discussed here we feel it important to
validate our CFD model setup with the results obtained
computationally by Madsen et al. [32] for the straight-walled
diffuser. Therefore, a geometry was created, matching an expansion
ratio described in the reference to validate our CFD model setup. It
was found that our pressure recovery values converged to within 2%
of those in the reference.
A. Geometry Parameterization
1. Parameterization One: Polynomial Splines
First we used a model comprising four separate piecewise
polynomial splines, referred to from here on as simply splines, that
pass through ve points along the wall. The advantage of using
piecewise splines instead of a continuous cubic spline is that a
continuous spline is governed by both C
1
and C
2
continuity at each
point. This observation leads us to conclude that this provides less
local control of the wall shape than we would ideally like. By
demanding the continuity of the second order derivative at each point
we are ensuring that at these points, the rate of change of gradient
must remain the same. This could lead to the spline overshooting
the points through which the continuous cubic spline passes and
causing a rippling effect: a situation that we would wish to avoid.
We dene the duct to be symmetrical about its centerline with a
cross-sectional area given by
2676 HOYLE, BRESSLOFF, AND KEANE
A(x) =A
in
(A
e
A
in
)(x) (13)
where 0 (x) 1.
The parameter is varied along the duct by treating it as a function
of the x-coordinate of the centerline at that particular cross section.
To avoid the problem of rippling through these points by using
cubics, we have chosen the rst and last splines to be cubic and the
middle two to be quadratic curves. This approach means that we are
not restricting our problem by having too many constraints over the
geometry to maintain continuity at the joints between the curves.
Ghate et al. [33] followed this approach to parameterize a duct using
three piecewise cubics.
In our case a ve-dimensional problem can then be set up as in
Fig. 3.
The piecewise variation of is then prescribed as
(x) =
_

_
c
11
c
12
x c
13
x
2
c
14
x
3
0 x x
1
;
c
21
c
22
x c
23
x
2
x
1
x x
2
;
c
31
c
32
x c
33
x
2
x
2
x x
3
;
c
41
c
42
x c
43
x
2
c
44
x
3
x
3
x x
4
(14)
Using the boundary conditions, setting the derivative at the end
points to be zero, and the continuity conditions on we can express
the coefcients c
ij
entirely in terms of x
1
; . . . ; x
5
and
1
; . . . ;
3
, with
the area of each cross section being calculated using Eq. (13).
2. Parameterizations Two and Three: HicksHenne Functions
A suitable curve can also be modeled using the bump functions
introduced by Hicks and Henne [11]. Our second parameterization
involves a wall dened via a single HicksHenne function. These are
global shape functions which allowfor good local control and can be
written in a general form [34]:
f =a[sin(x
(log 2)= log x
p
)|
T
(15)
These functions always guarantee smooth curves and also have the
appropriate end constraints. Here we superpose the function on a
simple straight-sided duct, normalized such that the wall distance
x [0; 1|. Using one of these functions to describe the wall geometry,
we have three design variables: a, x
p
[0; 1|, and T.
Finally, our third parameterization involved two HicksHenne
bump functions summed together to give a total of six design
variables. This provides us with further geometries where greater
local control of the curve is possible.
B. Results
For our problem, convergent-divergent diffusers began to
materialize with competitive C
p
values for all of our tested
techniques. We found that by converging the owslightly at the inlet,
the acceleration of the ow from the increase in local Reynolds
number increases the turbulence of this ow and hence increases the
turbulence of the boundary layer. This increase in turbulence is
sufcient to prevent the onset of separation. The contours of velocity
magnitude in a convergent-divergent diffuser, here found as the
optimum via the single HicksHenne approach, and in a straight-
walled diffuser for comparison, are illustrated at the top and bottom
of Fig. 4, respectively. This gure illustrates the whole
computational domain of the symmetry half for each. The diffuser
section has been magnied to showthe difference in geometry. In the
straight-walled gure at the top, we can see that above the main core
of ow, there is an area of recirculation. Here, the ow separated
close to the diffuser entry. Because there is a region of high velocity
ow and a region of very low velocity ow within the recirculation
area at the lter point, the mass-averaged static pressure recovery
returns a low overall value of C
p
=0:21. In contrast, the ow in the
converging-diverging wall diffuser pictured at the bottom of Fig. 4
expands with no separation and has a more even pressure uniformity
across the lter. Because of this, the ow is generally at a relatively
low velocity through the lter, even with the slight acceleration near
the inlet caused by the converging wall, and hence returns a higher
mass-averaged static pressure recovery value of C
p
=0:36. The
optimum diffuser for both the double HicksHenne function and the
spline technique also yielded a convergent-divergent diffuser almost
exactly matching that of the one shown at the bottom of Fig. 4.
Essentially we have shown that, given an appropriate optimization
strategy in terms of an accurately represented RSM, the varying
parameterization techniques tested are capable of producing results
with the same geometrical features. We can conclude that given a
straight diffusion of ow, we can choose the single HicksHenne
approach which uses the fewest design variables. However, we feel
that for a turning expanding diffuser as studied in Sec. VII, a
polynomial spline parameterization would be the appropriate option
to pursue for two reasons: rst, a single HicksHenne function
applied upon a straight line connecting the entry to the lter of the
airbox would only allowfor one bump to contour the shape, whereas
the spline technique will be capable of producing a convergence near
the entry and a bulge near the exit, for example, giving us a stronger
local control of the wall shape; second, constructing spline dened
walls is an inbuilt function within the CAD engine used and hence
easier to implement from an industrial designers point of view.
X
= 1
= 0

x
X
entry
diffuser
exit
diffuser
centerline
X
Fig. 3 Spline parameterization of the symmetry half of the straight
diffuser.
Fig. 4 Filled contours of velocity magnitude in a straight wall and
optimum convergent-divergent diffuser.
HOYLE, BRESSLOFF, AND KEANE 2677
With our inlet speed and expansion conditions, the point of
separation occurring within the diffusers studied here are different to
those studied by Carlson et al. [28] and Madsen et al. [32]. This
explains why none of our three parameterization techniques yielded a
bell-shaped optimal design. The physical reasoning behind the bell-
shaped designs is that provided the boundary layer is sufciently
turbulent at the inlet, this will ensure that the boundary layer will
remain attached during the prime area enlargement. However, our
studies show that our ow has too high a velocity, given the
expansion required, for the boundary layer to remain attached.
Hence, bell-shapes produced poor designs as our ow could not
reattach within the specied axial length, where the reattachment
occurred between 0.8 and 1.2 mdownstreamof the lter, resulting in
a lowC
p
value. Because of the xed engine position and strict overall
car dimensions required by the race regulatory bodies, we cannot
allow for a longer diffuser to follow the ow streamlines in order for
the ow to reattach.
Instead, during the optimization process used here an interesting
design began to form. Prior studies such as the ve-dimensional case
using B-splines tested by Madsen et al. [35] constrained the
geometries such that a positive wall slope was seen along the wall.
However, considering our different required speed and expansion,
we did not stipulate this condition. Although we specied tangency
conditions at both the entry and exit points of the diffuser, for the
HicksHenne dened parameterizations we did not limit the variable
bounds such that a negative bump amplitude would be rejected for
points placed near the entry point. Although the spline technique
limited our wall height variable to the interval [0, 1], see Fig. 3, the
value of and the position of the rst control point coupled with an
appropriate -value for the second control point could lead to a wall
shape with a negative wall gradient seen near the inlet.
VI. Two-Dimensional Elbow
Following the study of the preceding section we next consider two
different parameterization methods for a duct of constant width
turning through 90 deg. The purpose of this study, as in the preceding
section, is to determine whether different parameterization schemes
offering different numbers of design variables can affect the
optimum design output from the same optimization process.
A. Geometry Parameterization
Over the distance within which an F1 diffuser turns, an
optimization cycle implemented on a nondiffusing elbow over the
same distance would return similar C
p
values no matter how the
geometric shape varied. This is because the owwould not separate.
By shortening the distance over which the elbow bends through
90 deg we can ensure an amount of separation in the poorer designs
forcing the optimizer to work harder to nd an optimum.
1. Parameterization One: Bezier Curves
Our rst curved duct employed a Bezier curve [36] with six overall
control points dening the centerline, of which two are considered as
variables, see Fig. 5. Two control points are xed on the inlet and
lter position planes and two are placed in a xed position to impose
the tangency condition at the entry and exit to the bend. Two parallel
walls were constructed equidistant fromthe centerline on each side to
construct the constant width elbow. Horizontal and vertical tangency
conditions were implemented at the entry and exit, respectively, by
xing control points along the required tangency. These conditions
prevent designs harboring a sharp point at the diffuser exit onto
which the trumpet tray would join.
2. Parameterization Two: Polynomial Splines
The alternative parameterization was set using a cubic spline
passing through three points positioned along the centerline. Again,
two parallel equidistant walls were placed either side of the
centerline. Although we increase our design variable count by two,
an advantage of this method over the Bezier curve approach is that
continuity conditions are implemented at the entry and exit to the
bend without the need of extra xed control points.
B. Results
In this study, all of the designs tested within the optimization
process returned a negative pressure recovery. This is because there
is no expansion and hence no pressure being recovered due to the
reduction of the ows kinetic energy. So, the static pressure over the
lter or exit of the bend is less than the static pressure at which the
ow enters the bend. However, both parameterization methods
returned similar optimum pressure recoveries of C
p
=0:0413 for
the Bezier curve method and C
p
=0:0415 for the spline method,
with similar geometries. These geometries turn the ow through
90 deg and at a gradual and even rate throughout the bend. This
ensures that there is no boundary layer separation.
Although both designs produced similar optimum C
p
values, in
looking forward towards the fusing of the expansion and the bend,
the piecewise spline approach appears to be the most efcient when
dening the centerline bend. Bezier curves have an inherent problem
in that we cannot set a tangency at an end point of the curve without
requiring extra xed control points next to the entry and exit points in
the direction of the required tangency through which the Bezier curve
must pass. The degree of tangency is then based upon the distance of
these extra control points away from the points positioned on the
entry and lter planes, which could become extra variables if wanted.
This is easily dealt with when using polynomial splines. Given this
extra complication within the geometry, we opt to pursue our
remaining study with the piecewise cubic spline dened centerline.
VII. Two-Dimensional Diffuser with Bend
A. Combined Geometry Parameterization
The geometric fusion of the expansion and bend parameterizations
from the preceding two sections is next used to create a two-
dimensional airbox model. It is important to note that to maximize
the amount of local control of the overall diffuser shape given to the
optimizer, so as to allowthe production of potentially radical results,
the upper and lower walls need to be completely uncoupled. Thus,
the positioning of the control points through which the upper wall
spline passes must be independent of the control points dening the
lower wall spline. This is ensured by using a model where there are
no links between the centerline bend and spline control point
positioning variables. This then prevents the problem of wrinkles or
loops forming on the lower wall spline leading to geometrically
infeasible designs.
The resulting parameterization involves the use of piecewise cubic
splines for all three sections. As noted in Sec. V, the most efcient
simple expansion materialized as a convergent-divergent diffuser
using a polynomial spline parameterization. Hence, here the upper
and lower wall variables were left free to produce convergent-
divergent diffusers should good C
p
values be returned for these
exit
entry
X
X
X
X
Bezier curve defined centerline
parallel walls
variable
control points
X control points for Bezier curve
X
X
Fig. 5 Geometry parameterization shown for a Bezier curve dened
centerline.
2678 HOYLE, BRESSLOFF, AND KEANE
design points. FromSec. VI it was shown that efcient turning of the
owoccurs through a regular and gradual bend and hence the ranges
allowed for the three centerline control points are chosen with this in
mind. The following parameterization technique for our airbox
model features uncoupled walls together with the capability of
designs allowing converging walls or walls featuring bumps for
example. This technique, with a variable count within the 20 design
variable limit, has been devised to give us the potential to produce
possibly radical results with good pressure recovery values.
This technique is illustrated in Fig. 6. This requires a total of 16
design variables: six for the centerline control points CP1(x; y),
CP2(x; y), CP3(x; y); ve for the upper wall r
u1
, r
u2
, r
u3
(being the
normal distances from the centerline) positioned by two variable
ratios, D
1
and D
3
, along the centerline with the middle ratio, D
2
,
xed at 0.5; and ve for the lower wall r
l1
, r
l2
, r
l3
(being the distances
along the lines intersecting the centerline) dened by two variables

1
and
3
with
2
xed at 45 deg.
B. Results
Figure 7 shows the development of the optimization process using
this parameterization, showing C
p
values for design points 1 to 200
representing the initial 200 DoE points, design points 201 through
300 representing the subsequent update points and the nal 301 to
350 design points from the 50-point local exploration in 20% of the
design space. The bold line indicates the current best optimum as
each update point was added. We have adhered to the convention of
performing roughly 10 times as many experiments as design
variables to cover the design space sufciently within the initial DoE.
Despite mentioning earlier that the minimum number of updates is
generally at least twice the number of points used in the initial DoE,
we have found that for a 16 design variable problem, large compute
times are needed to construct a response surface with upwards of 200
points. For this reason, we have limited our updates tojust 100 points.
A wide range of interestingly shaped diffusers were produced
within the rst 300 design points from convergent-divergent wall
shapes to diffusers with one bulge feature on the lower wall and
various bulges on the upper wall. Many designs, such as the
convergent-divergent ones, contained no areas of separation.
However, due to the increase in kinetic energy of the owthrough the
slight convergence the pressure recoveries were lower than the best
value seen, the best ranging between C
p
=0:7 and C
p
=0:8.
Designs that are similar to those seen within the actual race cars were
also found within our design space, with straight upper walls to t
within the present design of roll bar structure, i.e., that had no radical
geometric features such as bulges seen on either wall. One such
airbox can be seen in Fig. 8. Again, we can see that separation has
been completely eliminated. However, with a pressure recovery of
C
p
=0:7805 it was signicantly lower than the best design found.
The best design had a pressure recovery of C
p
=0:9316 and is
shown in Fig. 9. Here, completely different geometric features have
emerged. Two small bulges have been formed along the lower wall
inducing separation within these bulges, the bubbles of separation
being completely contained within the bulges. The upper wall also
contains a bulge within which forms another bubble of separation. It
is important to note here that these bulges formed are not due to the
case of rippling as described in Sec. V.A. Here, all three control
points along the lower spline sit at the inexion points of the curve
and so the spline is not overshooting the point which is what leads
to the rippling situation.
After 300 design points were tested, our process continued with a
50-point local exploration around the current best point found with
C
p
=0:9316, reducing the range of each variable to 20% of the
domain centered around this point. From Fig. 7 these are shown as
l3
(0,0)
r
r
r
r
r
l2
x
x
x
x
l1
+
+
+
r

2
D
D
D
1
CP2(x,y)
CP3(x,y)
CP1(x,y)
Upper x ratio distance along centerline

3
2
1
x
X
X
O
u1
u2
u3
x
X
O
O
Fig. 6 Geometry parameterization of airbox model.
0 50 100 150 200 250 300 350
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Number of design points
C
p
Fig. 7 Optimization history.
Fig. 8 Filled contours of velocity in a design found which contains no
ow separation with C
p
0:7805.
Fig. 9 Filled contours of velocity in the best design found having
completed 100 update points with C
p
0:9316.
HOYLE, BRESSLOFF, AND KEANE 2679
design points 301350. It is clear that by focusing on a small region
within such a large design space and by performing a dense search in
this area we can converge our model more quickly to a better design
yielding a higher C
p
value. Our optimum design was found with
C
p
=0:9658 and is illustrated in Fig. 10. As we can see here, the
upper bulge has been reduced slightly, eliminating separation on this
wall. The lower bulges, however, remain present and capture small
bubbles of separation within them. The upper and lower bounds of
the whole design space, together with the design variable values for
the best design after 300 design points and after the local exploration
are shown in Table 1.
An interesting observation is that after the completion of the
update points and analysis of the best design attained at that point,
one could intuitively think that a better design could be achieved by
eliminating separation, creating a design with walls following the
ow streamlines of the airbox illustrated in Fig. 9. However, the
subsequent local exploration has produced a design without
separation different to that of following the streamlines of the airbox
shown in Fig. 9. This airbox has a pressure recovery value C
p
=
0:9452 and is shown in Fig. 11. If we followed the streamlines of the
airbox in Fig. 9, the upper wall would have been straight. What we
have seen is that by a small manipulation of the design variables, we
can eliminate separationwithout following the streamlines of the best
design after the update process to produce a design with a higher
pressure recovery value. Another interesting design found within the
local exploration returning a high pressure recovery value of C
p
=
0:9555 is shown in Fig. 12. This airbox does feature a straight upper
wall as well as two small bumps on the lower wall. This is of
particular interest to the F1 aerodynamicist as this design would not
require a redesign of the current roll bar structure in which the airbox
sits.
We can see that for all the airboxes illustrated in Figs. 1012, the
elimination of separation on the upper wall is desired to return a high
pressure recovery. It is also clear that the lower wall bulges are useful
features in delivering an efcient expansion of the ow without
experiencing losses from extensive separation. These bulges are
allowed to exist due to the parameterization technique proposed, the
key attribute of which is the geometric independence of the upper
wall from the lower wall. Bulges containing small regions of
separation have previously been seen to be benecial in terms of
reducing excessive nonuniformity of the ow at the diffuser exit
when applied to a curved subsonic S-duct diffuser [7]. Increased
losses in the airbox would occur from separation when the ow
experiences excessive nonuniformity of the ow through the lter.
This is not the case here as the small separation bubbles are
completely contained within the lower wall bumps. In this study, we
see a benet of the two bulges containing small separation bubbles
returning a higher pressure recovery over that of a design which
eliminates separation entirely. It is clear that the balance between the
size of bump on the upper wall and size of bumps on the lower wall,
which together control the expansion of the ow as it turns, is very
ne. This balance has been explored computationally through the
capability of the parameterization technique proposed as it has strong
local control. However, we would accept that there is a need for
further research into such features including experimental testing due
Fig. 10 Filled contours of velocity in the optimum design with
C
p
0:9658.
Table 1 Designparameters andtheir corresponding bounds withthe variable values for the best designfound
after 100 update points and again after a further 50 point local exploration
Variable Lower bound Upper bound Best value after update Best value after local exploration
CP1x 0.1 0.175 0.1408 0.1399
CP1y 0:015 0 0:0122 0:0120
CP2x 0.25 0.325 0.2501 0.2481
CP2y 0:08 0 0:0675 0:0685
CP3x 0.375 0.55 0.4181 0.4290
CP3y 0:202 0:09 0:2018 0:1920
Upper x ratio 1 0.05 0.45 0.1853 0.1703
Upper x ratio 2 0.5 0.5 0.5 0.5
Upper x ratio 3 0.55 0.9 0.6162 0.5943
r
u1
0.055 0.08 0.0651 0.0673
r
u2
0.07 0.14 0.1359 0.1333
r
u3
0.145 0.28 0.2128 0.2111

1
5 40 17.9152 18.3529

2
45 45 45 45

3
50 85 84.8928 83.5803
r
l1
0:08 0:055 0:0734 0:0712
r
l2
0:14 0:07 0:1295 0:1234
r
l3
0:2 0:15 0:1506 0:1462
Pressure recovery 0.9316 0.9658
Fig. 11 Filled contours of velocity in a design which contains no
separation with C
p
0:9452.
2680 HOYLE, BRESSLOFF, AND KEANE
to the uncertainty accompanying accurate prediction of separation
and reattachment using RANS solvers.
VIII. Concluding Remarks
We have approached the design optimization of a two-
dimensional F1 airbox by breaking down the expanding turning duct
into its separate features. Using the knowledge gained through
studying varying geometry parameterization techniques to design a
straight diffuser and a constant width turning elbow, we have fused
the best ideas fromeach to forma exible parameterization technique
that models a two-dimensional turning diffuser. The resulting
parameterization approach for the 2-D airbox has uncoupled upper
and lower walls, offering strong local geometric control. This in turn
increases the design variable count, but, by applying a sophisticated
response surface based optimization strategy using Kriging, we can
efciently optimize such a high dimensional problem. Although the
number of update points that can be calculated is related to the
compute power available, once the computational limit has been
reached for building the response surface models we can conclude
the update process and performa local exploration around the current
best point. This has allowed us to converge the model quickly to nd
small changes to the design which will yield higher pressure recovery
values.
Such a two-stage search process coupled with a parametric
geometry offering strong local control has allowed us to efciently
obtain superior designs that are nonintuitive to a designer.
References
[1] Patterson, G. N., Modern Diffuser Design, Aircraft Engineering,
Vol. 10, Sept. 1938, pp. 267273.
[2] 2005 Formula One Technical Regulations, Federation Internationale de
LAutomobile, 2005.
[3] Myers, R. H., and Montgomery, D. C., Response Surface Methodology:
Process and Produce Optimization Using Design of Experiments,
Wiley, New York, 1995.
[4] Jones, D. R., Schonlau, M., and Welch, W. J., Efcient Global
Optimization of Expensive Black-Box Functions, Journal of Global
Optimization, Vol. 13, No. 4, 1998, pp. 455492.
[5] Mead, R., The Design of Experiments, Cambridge Univ. Press,
Cambridge, England, U.K., 1988.
[6] Alexandrov, N., Dennis, J. E., Lewis, R. M., and Torczon, V., A Trust
Region Framework for Managing the Use of Approximation Models in
Optimization, Structural Optimization, Vol. 15, No. 1, 1998, pp. 16
23.
[7] Zhang, W.-L., Knight, D. D., and Smith, D., Automated Design of a
Three-Dimensional Subsonic Diffuser, Journal of Propulsion and
Power, Vol. 16, No. 6, 2000, pp. 11321140.
[8] Samareh, J. A., A Survey of Shape Parameterization Techniques,
NASA CP-1999-209136, June 1999, pp. 333343.
[9] Braibant, V., and Fleury, C., Shape Optimal Design Using B-Splines,
Computer Methods in Applied Mechanics and Engineering, Vol. 44,
Aug. 1984, pp. 247267.
[10] Farin, G., Curves and Surfaces for Computer Aided Geometric Design,
Academic Press, New York, 1990.
[11] Hicks, R. M., and Henne, P. A., Wing Design by Numerical
Optimization, Journal of Aircraft, Vol. 15, No. 7, July 1978, pp. 407
412.
[12] CFD Flow Modeling Software, Fluent, Shefeld, U.K., 2005, www.
uent.com.
[13] Launder, B. E., and Spalding, D. B., The Numerical Computation of
Turbulent Flows, Computer Methods in Applied Mechanics and
Engineering, Vol. 3, No. 2, 1974, pp. 269289.
[14] Spalart, P. R., and Allmaras, S. R., AOne-Equation Turbulence Model
for Aerodynamic Flows, AIAA Paper 92-0439, Jan. 1992.
[15] Fluent User Guide, Fluent, Shefeld, U.K., 2003.
[16] Keane, A. J., The Options Design Exploration System Reference
Manual and User GuideVersion B3.1, Southampton, England, U.K.,
http://www.soton.ac.uk/~ajk/options.ps, 2002.
[17] Sobol, I. M., On the Systematic Search in a Hypercube, SIAMJournal
on Numerical Analysis, Vol. 16, No. 5, 1979, pp. 790793.
[18] Statnikov, R. B., and Matusov, J. B., Multicriteria Optimization and
Engineering, Chapman and Hall, New York, 1995.
[19] Sbester, A., Leary, S. J., and Keane, A. J., On the Design of
Optimization Strategies Based on Global Approximation Models,
Journal of Global Optimization, Vol. 33, No. 1, 2005, pp. 3159.
[20] Sacks, J., Welch, W. J., Mitchell, T. J., and Wynn, H. P., Design and
Analysis of Computer Experiments, Statistical Science, Vol. 4, No. 4,
1989, pp. 409435.
[21] Cressie, N., The Origins of Kriging, Mathematical Geology, Vol. 22,
No. 3, 1990, pp. 239252.
[22] Keane, A. J., Wing Optimization Using Design of Experiment,
Response Surface and Data Fusion Methods, Journal of Aircraft,
Vol. 40, No. 4, 2003, pp. 741750.
[23] Jones, D. R., ATaxonomy of Global Optimization Based on Response
Surfaces, Journal of Global Optimization, Vol. 21, No. 4, 2001,
pp. 345383.
[24] Yuret, D., and de la Maza, M., Dynamic Hill Climbing: Overcoming
the Limitations of Optimization Techniques, Proceedings of the 2nd
Turkish Symposium on AI and ANN, 1993, pp. 254260, http://home.
ku.edu.tr/dyuret/pub/tainn93.html.
[25] Kline, S. J., Abbott, D. E., and Fox, R. W., Optimum Design of
Straight-Walled Diffusers, Journal of Basic Engineering, Vol. 81,
No. 3, 1959, pp. 321331.
[26] Fox, R. W., and Kline, S. J., Flow Regimes in Curved Subsonic
Diffusers, Journal of Basic Engineering, Vol. 84, Sept. 1962, pp. 303
316.
[27] Reneau, L. R., Performance and Design of Straight, Two-Dimensional
Diffusers, Journal of Basic Engineering, March 1967, pp. 141150.
[28] Carlson, J. J., Johnston, J. P., and Sagi, C. J., Effects of Wall Shape on
Flow Regimes and Performance in Straight Two-Dimensional
Diffusers, Journal of Basic Engineering, March 1967, pp. 151160.
[29] Glowinski, R., and Pironneau, O., On the Numerical Computation of
the Minimum-Drag Prole in Laminar Flow, Journal of Fluid
Mechanics, Vol. 72, No. 2, 1975, pp. 385389.
[30] Zhang, J., Chu, C. K., and Modi, V., Design of Plane Diffusers in
Turbulent Flow, Inverse Problems in Engineering, Vol. 2,
March 1995, pp. 85102.
[31] Lim, S., and Choi, H., Optimal Shape Design of a Two-Dimensional
Asymmetric Diffuser in Turbulent Flow, AIAA Journal, Vol. 42,
No. 6, 2004, pp. 11541169.
[32] Madsen, J. I., Olhoff, N., and Condra, T. J., Optimization of Straight,
Two-Dimensional Diffusers by Wall Contouring and Guide Vane
Insertion, Proceedings of 3rd World Congress on Structural and
Multidisciplinary Optimization, edited by C. Bloebaum, Buffalo, NY,
1999, pp. 629631.
[33] Ghate, D., Isaacs, A., Sudhakar, K., Mujumdar, P. M., and Marathe, A.
G., 3d-Duct Design Using Variable Fidelity Method, AIAA
Paper 2004-4427, Sept. 2004.
[34] Sbester, A., and Keane, A. J., Empirical Comparison of Gradient-
Based Methods on an Engine-Inlet Shape Optimization Problem,
AIAA Paper 2002-5507, 2002.
[35] Madsen, J. I., Shyy, W., and Haftka, R. T., Response Surface
Techniques for Diffuser Shape Optimization, AIAA Journal, Vol. 39,
No. 9, 2000, pp. 15121518.
[36] Bezier, P. E., Emploi des Machines a Commande Numerique, Masson
et Cie., Paris, 1970.
A. Messac
Associate Editor
Fig. 12 Filled contours of velocity in a design with a straight upper wall
and C
p
0:9555.
HOYLE, BRESSLOFF, AND KEANE 2681