Chinese Journal of Aeronautics, (2018), 31(6): 1165–1180

Chinese Society of Aeronautics and Astronautics

& Beihang University
Chinese Journal of Aeronautics
cja@buaa.edu.cn
www.sciencedirect.com

An efficient aerodynamic shape optimization of

blended wing body UAV using multi-fidelity models
Parviz MOHAMMAD ZADEH *, Mohsen SAYADI

Faculty of New Sciences and Technologies, University of Tehran, Tehran 1439957131, Iran

Received 25 July 2017; revised 26 October 2017; accepted 26 December 2017

Available online 12 April 2018

KEYWORDS
Adaptive filter sequential
quadratic programing
(AFSQP);
Adaptive robust meta-model;
Aerodynamic shape opti-
mization;
Blended wing body (BWB);
Move limit strategy;
Unmanned aerial vehicle
(UAV)
Abstract This paper presents a novel optimization technique for an efficient multi-fidelity model
building approach to reduce computational costs for handling aerodynamic shape optimization
based on high-fidelity simulation models. The wing aerodynamic shape optimization problem is
solved by dividing optimization into three steps—modeling 3D (high-fidelity) and 2D (low-
fidelity) models, building global meta-models from prominent instead of all variables, and determin-
ing robust optimizing shape associated with tuning local meta-models. The adaptive robust design
optimization aims to modify the shape optimization process. The sufficient infilling strategy—
known as adaptive uniform infilling strategy—determines search space dimensions based on the last
optimization results or initial point. Following this, 3D model simulations are used to tune local
meta-models. Finally, the global optimization gradient-based method—Adaptive Filter Sequential
Quadratic Programing (AFSQP) is utilized to search the neighborhood for a probable optimum
point. The effectiveness of the proposed method is investigated by applying it, along with conven-
tional optimization approach-based meta-models, to a Blended Wing Body (BWB) Unmanned Aer-
ial Vehicle (UAV). The drag coefficient is defined as the objective function, which is subjected to
minimum lift coefficient bounds and stability constraints. The simulation results indicate improve-
ment in meta-model accuracy and reduction in computational time of the method introduced in this
paper.
Ó 2018 Chinese Society of Aeronautics and Astronautics. Production and hosting by Elsevier Ltd. This is
an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

1. Introduction

* Corresponding author.
In recent years, in order to reduce fuel consumption and
improve the performance, the optimization of UAV shapes
E-mail address: pmohammadzadeh@ut.ac.ir (P. MOHAMMAD
ZADEH).
has been the main focus of the competitive aerospace market.
Peer review under responsibility of Editorial Committee of CJA.
The development of Blended Wing Body (BWB) design is such
an effort. In addition to the elimination of the tail for this par-
ticular kind of UAV and the significant reduction in equivalent
weight, drag force, and radar cross-section, the available space
Production and hosting by Elsevier for installing equipment inside the wing and the effective range
https://doi.org/10.1016/j.cja.2018.04.004
1000-9361 Ó 2018 Chinese Society of Aeronautics and Astronautics. Production and hosting by Elsevier Ltd.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1166
have also been increased. Despite all these mentioned advan-
tages, instability is the negative outcome of eliminating the tail.
Correcting this flaw requires designing a combination of con-
trol surfaces and reflexed wing sections and using sophisticated
computer control systems. Therefore, the aerodynamic shape
design optimization of BWBs, along with the need to meet
the design requirements, has inspired several researchers to
overcome its challenges.
However, BWB and pure flying wing have some differences
in definition; their optimization design approaches and some
principles of aerodynamic design are practical for each other.
The BWB design challenges along with necessity of developing
aircraft efficiency impose more computational effort on the
preliminary design process.
Various objectives and different constraints in the design of
BWB make them a proper candidate for the application of
multi-objective and multi-disciplinary design optimization
techniques.1,2 The characteristics of their shape makes their
geometry parameterization easier. A study on improving the
Concurrent Subspace Optimization (CSSO) structure on the
basis of response surface and Monte Carlo analysis for the
robust, single-objective optimization of the flying wing was
conducted in this field.3 Multidisciplinary Design Optimization
(MDO) architecture of aerodynamic shape optimization was
developed for battery-powered composite BWB with Delta
wing.4 The results of implementing this architecture and con-
ventional optimization process were compared to demonstrate
the presented formulation. Pan et al.5 presented a systematic
technique in aerodynamic and stealthy MDO issue for
double-sweep flying wing. They utilized the hybrid structure
of global optimization and gradient algorithm as an optimiza-
tion strategy in conceptual design. Morris et al.6 devoted atten-
tion to multi-disciplinary multi-level optimization for the
simultaneous optimization of aerodynamic shape and
structure.
The mere design of the aerodynamic shape was the main
objective of optimization in certain studies, while only the air-
foil cross-section was the focus in some others.7,8 The defini-
tion of geometry and surface meshing was investigated by
Truong et al.9 to enhance the quality of the mesh modified dur-
ing optimization. The robust design of airfoil shape optimiza-
tion is investigated to reduce the sensitivity of small random
geometry perturbations and uncertain operational condi-
tions.10 The construction of meta-models based on Kriging
and gradient-enhanced Kriging is based on a relatively small
number of CFD evaluations. Since the optimization of the
fixed geometry aircraft demands satisfying conflicted con-
straints in various flight conditions, aerodynamic shape opti-
mization of morphing wing is the subject of the study by
Hunsaker et al.11 This method increases allowable wingspan
with induced drag reduction for a given structural weight.
In addition to putting forward an optimal Lifting-Fuselage
Configuration (LFC) shape for BWB in the research by Reist
and Zingg,12 the aerodynamic shape was optimized for the best
cruise altitude and reduced fuel consumption. In another
study, the hybrid design of the aerodynamic shape and struc-
ture of the flying wing was optimized by combining the
multi-bump method with automatic optimization and flow
control to increase the lift-to-drag ratio and improve longitu-
dinal static stability.13 The presented approaches differ mainly
in the definition of the geometry of the problem, objective
P. MOHAMMAD ZADEH, M. SAYADI
functions, optimization constraints,14 and finally, the accuracy
of the adopted models.15
In complicated engineering design problems such as BWB,
Surrogate-Assisted Optimization (SAO) methods have been
developed to enhance the accuracy and reliability of optimiza-
tion process.16 In particular, due to the high computational
cost of solving the CFD models, the researchers use the
meta-model in aerodynamic optimizations.14,17,18 However,
constructing accurate meta-model would still be time-
consuming and is often associated with insufficient accuracy
in order to ensure a great degree of change in variables and
the presence of local extrema for the objective and constraint
functions.
The concept of sequential approximation method is intro-
duced to overcome the mentioned limitations of the meta-
models imposed by large-scale and complex design space.19
The simulation outcomes show that building appropriate
low-fidelity model reduces the computation costs and improves
model accuracy. Using the response correction techniques for
aerodynamic shape optimization introduced by Koziela et al.7,
the precision of the alternative models derived from low-
accuracy models is improved. An automated selection of
low-fidelity model for aerodynamic shape optimization is pro-
posed in another study.20 This approach utilizes low- and high-
fidelity model misalignment.
In the variable-fidelity shape optimization, the hierarchical
kriging technique is utilized for modifying low-fidelity kriging
model.21 Since the low-fidelity model is constructed based on a
single design point, some weighted aerodynamic data correct
the meta-model as high-fidelity data. Other studies consider
the effects of boundary layer transition for optimizing the
shape of a lifting body with adjustment of the meta-
models.22 The sample selection method for correcting the
model plays a key role in such architectures.23 Maximizing
the Expected Improvement function (EI),24 Probability of
Improvement function (PI), and the Mean Squared Error
(MSE)25 and minimizing the Lower Confidence Bound
(LCB)26 are some examples of these methods—known as infill-
ing strategies.
K-means algorithm classifies the solutions for selecting
points and modifying the database of the meta-model, whereas
genetic algorithm is tasked with optimization of the aerody-
namic shape.27 The application of the parallel processing capa-
bilities to the optimization of aerodynamic problems is
facilitated by combining these techniques in order to mitigate
the defects of each.23,28
The other method for reclaiming local adaptive meta-model
building is the move-limit strategy. The main merit key of these
approaches is the suppression off design space in the current
optimum point neighborhood and refining of the model in this
space. The vital importance of selecting the move limit strategy
is controlling optimization performance. These strategies—
both fixed29 and adaptive30—differ from one another by differ-
ent bound-adjustment methods.31 Among them, the global
convergence can be achieved by utilizing the trust-region
method.32
On the other hand, the selection of design variables in shape
optimization has an important effect on the appropriate
covering design space and reduction of computational cost.
Poole et al.33 proposed a novel method for the proper orthog-
onal decomposing set of training airfoils, which increase the
An efficient aerodynamic shape optimization
number of the required shape optimization variables based on
an optimal set of airfoil deformation modes.
Although gradient-based algorithms are very popular due
to their higher computational speed in solving aerodynamic
problems, evolutionary algorithms, such as genetic algorithm34
and particle swarm optimization,35 have been employed in a
number of studies. It should be taken into consideration that
the use of these algorithms would be more cost-effective owing
to the great number of evaluations in multi-objective prob-
lems, while they offer a better guarantee of evading the local
minima.36 On the other hand, numerous studies have been
conducted to decrease the computation time of gradient-
based algorithms. The presentation of the adjoining method
was such an effort37 that was associated with the calculation
of the derivative of aerodynamic forces. The Sparse Nonlinear
OPTimizer (SNOPT) algorithm—a combination of this
method with SQP method—has received much attention from
researchers.38,14
To enhance the aerodynamic shape design optimization of
BWB UAV, an adaptive robust design optimization frame-
work is presented in this paper. In the proposed framework,
initializing global meta-models from the 2D low-fidelity model
significantly reduces computational costs, while locally tuning
meta-models with high-fidelity models increases model accu-
racy near the global optimum point. Since the proposed frame-
work employs 3D high-fidelity models to construct local meta-
models, a promising reliability can be accomplished. In addi-
tion, in each optimization iteration, the adaptive trust region
manages space design and the uniformed sampling technique
guarantees the model’s robustness.
2. Methodology
The main aim of this paper is to reduce computational time for
solving aerodynamic shape optimization based on 3D high
fidelity aerodynamic model. Towards this goal, the approach
adopted in this paper involves first utilization of sensitivity
analysis to eliminate the design variables, which have very
low impact on the objective function (i.e., CD) or the con-
straints. Therefore, the design variables, which were eliminated
from the optimization process have very low impact effect on
the objective function (i.e., CD) and constraints and this is
mainly used to reduce the computational time. Second, this
paper presents accurate meta-models using multi-fidelity mod-
elling to reduce the computational time significantly and at the
same time to achieve accurate CD based on high fidelity (3D)
aerodynamic model. The approach is adopted based on two
levels of meta-model building process (i.e. global and local).
One of the key advantages of the multi-fidelity modelling
approach presented in this paper is the use of models of differ-
ent fidelity, i.e. low-fidelity (panel model) and high-fidelity
(3D) aerodynamic models in the optimization process. In the
proposed method, the interaction between 2D and 3D aerody-
namic models is utilized to correct (tune) 2D aerodynamic
model to achieve the same degree of accuracy of the 3D aero-
dynamic model, but at the same time it is computationally very
efficient model to be used in the aerodynamic shape optimiza-
tion process.
Fig. 1 illustrates the framework of the proposed optimiza-
tion technique for the aerodynamic shape optimization of
BWB UAVs. This framework is not limited to aerodynamic
1167
optimization and can be implemented for other types of geo-
metric optimization. According to this, optimization process
is partitioned into three stages. In the first stage, the initial
design geometry is parameterized and high-fidelity and low-
fidelity models are derived. The aim of the second stage is to
construct the global meta-model from the low-fidelity model.
Finally, the local meta-model is tuned in the neighborhood
of the current design point to improve the accuracy in each
optimization iteration.
Regardless of the optimization nature, each optimization
problem begins with the mathematical definition of the prob-
lem’s objectives and constraints. How the problem is defined,
i.e. optimization problem definition, affects the solution pro-
cess and the final optimal solution. Generally, an optimization
problem aims to minimize an objective function fðxÞ and sat-
isfy the constraints of equality (hðxÞ) and inequality (gðxÞ).
According to the problems, the functions may be linear and/
or nonlinear, as defined in the following:
Minimize : fðxÞ
8
< hðxÞ ¼ 0
>
ð1Þ
Subject to : gðxÞ  0
>
:
xlower  x  xupper
where x is the vector of design variables with the length of n,
and xlower and xupper are the upper and lower limits of the vari-
ables, respectively.
The implementation method of the structure introduced for
the problem is elaborated in the subsequent sections.
2.1. Preparing geometric models
The first step in this study is the appropriate definition of 2D
(low-fidelity) and 3D (high-fidelity) geometries. The defined
models should have all the specifications as well as an auto-
matic geometric parameterization. In fact, all the geometric
variables (geometric dimensions) should be defined parametri-
cally, so that the overall geometry changes automatically and
in proportion to other variables, as the value of each variable
changes.
This study considers appropriate control points at the
selected sections. Regardless of the number of changes in the
position of control points, the final shape is a smooth surface
passing through the control points. On the other hand, by
defining appropriate constraints along the third dimension,
taper ratio and the distances of control sections change not
only individually, but also across the whole shape with respect
to the final geometric shape. Therefore, the final shape under-
goes distortions both in the 2D sections and in the 3D space
related to variable change.
However, improving the accuracy of a model involves an
exponential increase in the computational cost. Therefore,
the defined model should be simple and a reflection of the
actual behavior. The main idea of the introduced framework
is the application of a low-fidelity model (2D model), which
leads to a simpler closed form.
The low-fidelity 2D model is extracted by simplifying the
high-fidelity 3D model. In fact, in this model, similar sections
of the 3D model are selected to define shape points, while
design variables are defined corresponding to them. Since the
panel method evaluates each section and the results are
1168
Fig. 1 Efficient aerodynamic shape optimization of BWB UAV using multi-fidelity models.
ultimately collected to obtain the values of objective functions
and constraints, the amount of calculations required is
decreased. By comparing results of 3D and 2D models, the
correction factor (model’s modifier coefficient) is derived. This
modification improves the accuracy.
2.2. Initializing global meta-model
The optimization of aerodynamic shape still takes a long time,
even when low-fidelity models are utilized. Employing the
meta-model is one of the situated strategies to diminish com-
putational time in the optimization process. On the other
hand, in large-scale and complex problems such as aerody-
namic optimization, the number of design variables is another
factor that affects the convergence specification and gradually
increases computational time. Consequently, the non-
prominent variables can be eliminated by means of ANOVA
for simplification of the meta-model. In this technique, it is
assumed that the explicit function f can be decomposed as
expressed in the following form:
X
n n X
X n
^fðxÞ ffi fm þ /i ðxi Þ þ /ij ðxi ; xj Þ þ   
i¼1 i¼1 j¼1
þ /1;2;...;n ðx1 ; x2 ; . . . ; xn Þ
where fm and / are the means of output distribution and low-
dimensional decomposition terms respectively. The contribu-
tion indices pk based on smoothing spline ANOVA are inter-
P. MOHAMMAD ZADEH, M. SAYADI
preted as importance of decomposition terms and are defined
below:
ð/k ; ^fÞ
pk ¼ ð3Þ
k^f2 k
where ð; Þ and ||.|| are the scalar product of vectors and vector
norm respectively. The effective terms have a higher value that
would be kept, while the lower value shows a negligible effect
that should be eliminated.39
The sampling points would be as well-distributed as possi-
ble in design space. Full or fractional factorial design, central
composite design, Taguchi,40 Sobol41 and uniform Latin
Hypercube42 are some of the predominant Design of Experi-
ments (DoE) schemes. The uniform Latin hypercube method
has been suggested for ensuring the evenness of the sampling
points. This DoE approach explores the entire design space
without taking into consideration the dimension of the
problem.
The meta-model features can be significantly affected by the
selected construction method. Polynomial regression,43 mov-
ing least-square method,44 neural networks, fuzzy logic,45 krig-
ing,46 and radial basis function47 are some of these
approaches.48 Kriging method is utilized by several aerospace
researchers to build a high-quality meta-model.49 In this study,
ð2Þ based on the Gaussian process theory, this method is preferred
compared with the other ones. At higher dimensionality like
wing design, this method has promising performance and more
robustness in constructing meta-models and updating them as
well as in lowering the computational time.50
An efficient aerodynamic shape optimization
The key factor in accuracy and convergence rate of opti-
mization methods such as Sequential Quadratic Programming
(SQP) is the selection of initial point. Selection of a feasible
point as the starting point may also prevent confusion of the
optimizing algorithm and reduce the computational costs.
Therefore, DoE sample with the smallest value of the objective
function is selected as the initial point.
2.3. Optimization process
Iterative optimization by updating a meta-model is taken into
consideration to achieve a structure with optimal performance.
In any iteration of optimization process, SQP optimization
algorithm determines the optimal point in proportion to the
problem conditions. On the other hand, meta-model parame-
ters are modified locally in the neighborhood of the optimal
point found in the design space. In other words, the modified
model is optimized in the created trust region of the local meta-
model in the following stage by assuming the current optimiza-
tion point as the initial point. In other words, the optimization
process continues as long as the local meta-model and the opti-
mal value are fixed.
2.3.1. Global optimization algorithm (AFSQP)
Adaptive Filter Sequential Quadratic Programing (AFSQP) is
an algorithm based on SQP with adaptive filter, which is exten-
sively used for the global optimization of single-objective non-
linear constraint problems. The advantage of this method
overshadows the performance of nearly all other nonlinear
programming methods in terms of velocity, accuracy, and per-
centage of feasible solutions.51 Taking the optimization history
into account and inspired by the idea of multi-objective opti-
mization problems and the adaptive adjustment of the number
of inputs to the filter during optimization, AFSQP provides
the only dominant point compared to previous solutions.52
Hence, x*is the optimal point of the problem if it satisfies:
r^fðx Þ þ k rgðx Þ ¼ 0
gðx Þ P 0 ð4Þ
ðk  gðx ÞÞ ¼ 0; k P 0
where k is the optimum value of Lagrange multiplier.
The main steps involved in optimization with AFSQP are
described below:
Step 1. Determine the initial point. Design variable bounds,
the maximum number of iteration, and termination tolerance
on the function value and constraints violation. Define the
scalar-valued ‘‘Lagrange function” in the following form:
Lðx; kÞ ¼ ^fðxÞ þ kgðxÞ ð5Þ
Step 2. Formulate QP problem for the iteration ðkÞ and the
unknown linear search direction d
1 the new design space and new sampling points are distributed
Minimize : ðd; r2X LðxðkÞ ; kðkÞ ÞdÞ þ r^fðxðkÞ Þd
2 ð6Þ
T
Subject to : rgðxðkÞ Þ d þ gðxðkÞ Þ P 0
Step 3. Solve QP problem Eq. (6), d value. The tentative
point is xðkþ1Þ ¼ xðkÞ þ d.
Step 4. If tentative point does not pass the test, appropriate
reduction is applied and a value controlled. The adaptive filter
judges the next iteration safety, anew = raold. The optimal
1169
choice of reduction factor, r2 (0, 1), is discussed by Truco.53
An evolutionary scheme adapts parameter r in a randomly
adaptive perturbation manner. The promising solution is
expressed as
xðkþ1Þ ¼ xðkÞ þ ad a 2 ½0; 1 ð7Þ
Step 5. The constraint violations are reduced with feasibil-
ity restoration. By introducing slack variable, the solution of
iterative trust-region algorithm is transformed.54 This step is
invoked when the filter accepts point or the constraint viola-
tions satisfy threshold criterion.
Step 6. Until any termination condition, such as achieving
termination tolerance on the function value or maximum num-
ber of iteration, is met, the number of iterations is incremented
by +1 and then Step 2 is repeated.
2.3.2. Tuning local meta-model
Improving the accuracy of calculations leads to an improve-
ment in the accuracy of the model. In this way, the meta-
model is updated after each optimization stage. In the pro-
posed method, updating is based on a few solutions of the
high-fidelity model to affect the model accuracy in a better
way. In fact, by limiting the variable bounds and modifying
the database, parameters of the meta-model are adjusted
locally, neighboring the optimal point. Therefore, the modified
e
meta-model of ^fðxÞ is a function of the evaluations of the high-
e
fidelity model of f(x) and the existing meta-model of ^fðxÞ in the
new design.
^fðfðxÞ; e
^f ¼ e
e ^fðxÞÞ ð8Þ
Selection of points to modify the high-fidelity model data-
base was done based on move-limit strategy method, described
in detail below.
2.3.3. Move-limit strategy
Sequential approximation method is one of the most practical
methods used to improve the accuracy of meta-model for
increasing the convergence probability under reduction of
the computational time caused by iterative analysis costs. It
is sometimes very difficult to find a meta-model with sufficient
accuracy throughout the design space.55 The theory of this
method is based on suppressing variable boundaries in the
neighborhood of the current design point. The modified
boundary, called ‘‘trust region”, is determined in each stage
based on solution positions. Therefore, optimization is per-
formed locally and there are limited permissible changes in
design variables. In other words, they are move-limited. Selec-
tion of move-limit strategy, especially when model accuracy is
insufficient, plays a crucial role in controlling convergence
rate.34
In this method, the current optimal point is at the center of
in this space. A new algorithm called ‘‘Adaptive Uniform
Infilling” is presented for the proper distribution of the points
in the new space.
The number of the points selected during optimization is
adaptively determined in relation to the number of optimiza-
tion iterations. Therefore, the number of required sampling
points reduces adaptively by limiting variable boundaries cor-
responding to model accuracy and eliminates unnecessary
1170 P. MOHAMMAD ZADEH, M. SAYADI

computational costs. The number of required sample points is

reduced as the model accuracy improves. The number of the
points is expressed as
" 3 #
ItRs
N¼ 1  10 ð9Þ
max It

where It, maxIt and Rs refer to the number of current iteration,

maximum number of iteration, and meta-model R-square in
optimum point. The operator [.] also denotes ceiling function.
Fig. 2 shows the process of changing the number of sample
points by assuming R-square = 0.85 and MaxIt ¼ 10. It Fig. 3 Adaptive uniform infilling criterion.
should be noted that the meta-model changes into a small
smooth surface at high iterations, while the considerable
reduction of the boundaries in proportion to the introduced
move-limit algorithm and the need for a larger number of sam-
ples to create the new model have been removed. As seen in the
diagram, there is almost no point in modifying the model after
about six iterations.
In this algorithm, the selected points are distributed evenly
on the oval surrounded by the boundaries of the new space,
while the oval is drawn in a way that the optimal point
obtained from the earlier stage is placed at the center. Distribu-
tion angle of the points (hj ) is equal in all ði; i þ 1Þ planes and is
calculated as follows:
2p
Dh ¼
N1 ð10Þ Fig. 4 Move-limit strategy in 2D design space.
hj ¼ ðj  1ÞDh; j ¼ 1; 2; . . . ; N  1
where N is number of the sampling points selected for modify-
ing the new model, which also includes the optimal point of the where xbi is the position of the current best point in
earlier stage. Fig. 3 illustrates span changing and number of i-dimension, and Dui and Dli denote upper and lower bounds
points during optimization in ði; i þ 1Þ plane. of variable i in this optimization iteration respectively.
After determining the points on the oval, the set of feasible Step 2. Modify variable boundaries, Dunewi and Dlnewi .
points within this subspace is also identified and finally all the
points on and inside the oval are used as the set of new sam- Dunewi ¼ xbi þ Dnewi
ð12Þ
pling points for creating a local meta-model. Undoubtedly, Dlnewi ¼ xbi  Dnewi
the probability of the presence of feasible points except the Fig. 4 displays the applied limitation on design boundaries
optimal point reaches almost zero in the design space at higher in 2D space. The figure shows that the variable boundaries
iterations, since the search space continuously becomes smal- during optimization process gradually become smaller around
ler. Algorithm steps are explained as follows: the optimal point.
Step 1. Determine new variables’ amplitudes Dnew . It is for- Step 3. Calculate the number of sampling points (N) in the
mulated for i-variable as new design space Eq. (9) and distribution angles of samples
minðDui  xbi ; xbi  Dli Þ Eq. (10), xsji . The position of j-sample in i-dimension xsji , is
Dnewi ¼ ð11Þ
2 described as follows:
Diþ1 Di
xsji ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi þ xbi j ¼ 1; 2; . . . N  1 ð13Þ
D2iþ1 þ D2i ðtan hj Þ2

where D and xbi are new variable boundaries and position of

the best point respectively. If p2 6 hj 6 3p
2
, the 00  00 is assigned
to Eq. (13); otherwise sign 00 þ 00 has to be used. Furthermore,
the position of this sample in i + 1-dimension is expressed as
xsjiþ1 ¼ xsji tan hj j ¼ 1; 2; . . . N  1 ð14Þ
Step 4. Organize the new set as the design points for high-
fidelity model. This set includes sampling points based on
infilling strategy and all the feasible points in the current opti-
mization iteration. The dispersion of this set across the sub-
region should be sustained in terms of accuracy improvement
guarantee. Evaluation of the optimum point should be
Fig. 2 Behavior of sampling number versus iterations. neglected since it was simulated for model verification.
An efficient aerodynamic shape optimization
3. Benchmark case
Optimization of the aerodynamic shape of a BWB benchmark
case was considered for evaluating the presented method and
comparing the optimization performance. The final objective
of the aerodynamic optimization is to achieve geometry with
the optimal aerodynamic performance. This means a trade-
off between the drag coefficient reduction and lift coefficient
increase. On the other hand, stability constraints should be
established for the optimized design in order to have an
accepted design. In this study, the minimization of drag coef-
ficient (CD ) with no reduction in lift coefficient (CL ) from
designer-admissible level and satisfaction of the static stability
are the major issues for optimization. Static stability is shown
in Cma (three-dimensional pitching moment curve slope) equa-
tion. The optimization formulation of the benchmark case can
be stated as follows:
Minimize : fðxÞ ¼ CD
Subject to : ga ðxÞ ¼ 0:45  CL 6 0 ð15Þ
gs ðxÞ ¼ Cma < 0
The aircraft has a 4380 mm wingspan with an aspect ratio
of 4.45, the middle section of which is fixed for analysis. Fig. 5
shows the overall dimensions of the aircraft.
Tables 1 and 2 summarize the values of parameters and
design variables of the blended wing body UAV with their ini-
tial variable boundaries.
To improve the optimization accuracy, a high-fidelity
model is usually referenced for constructing meta-models.
Therefore, the problem was optimized using the meta-model
based on the 3D (high-fidelity) model to demonstrate the high
capability of the proposed framework in terms of both the
computational cost and the solution accuracy. The results were
then compared with the ones obtained from the introduced
framework. The results prove the points of strength of the pre-
sented framework.
3.1. Conventional meta-model based aerodynamic shape
optimization using high-fidelity aerodynamic model of BWB
UAV
In the first step, a CFD 3D model was created using the vari-
able geometry of a UAV. The high-fidelity model was the same
as 3D UAV model in which several geometric variables are
considered for describing the shape of sections, wingspan,
complexity of airfoil, and sweep angle. In total, 341 geometric
Fig. 5 Initial geometry of blended-wing body UAV.
1171
variables were considered. Eight sections with constant spacing
were considered for controlling the shape and modifying the
model. Fig. 6 shows the names and numbers of the geometric
variables defined in the model.
3.1.1. Mesh generation and CFD solver
Table 3 indicates the boundary layer specifications considered
in the model. The nominal current condition is cruise altitude
of 35000 feet and a cruise Mach number of 0.16. Boundary
layer thickness was applied based on the Reynold’s number
of the flow. The maximum effective length was estimated
and applied based on the safety margin. With respect to the
solver type, height of the first layer is such that y+ falls in
the range of 20–100.
A tetragonal grid was selected in the 3D model (Fig. 7) in
order to achieve a proper accuracy versus reasonability of time
for calculations. There are 3690000 elements on the surface of
the initial model. The major control points are on the points in
the shape of a specified section.
The grid convergence was also studied to ensure the resolu-
tion accuracy of such a mesh. Fig. 8 shows the mesh indepen-
dence in proportion to the numbers and dimensions of
elements for the considered number of meshes (about 3.7 mil-
lion). In addition to the high quality of the aforementioned
grid, changes are applied automatically in the new grid by
changing the geometric points in the optimization process.
Air was considered as the ideal gas and the flow as an invis-
cid fluid for 3D implementation of the model. Table 4 details
other specifications of the used solver.
3.1.2. Optimization process
Due to the uniform distribution and minimum required num-
ber of covering surface, the uniform Latin hypercube method
was chosen to select the sample points in the design space.
Fig. 9 shows dispersion of the design points in a 3D space of
variables. In this figure, a is the rotation angle, Cr0 is the
longitudinal position of the point on the trailing edge (length
of section chord), and b1 is the section span of the first airfoil
section.
Before creating the meta-model, ANOVA detects the effec-
tive variables; 25 of the 330 defined variables in the 3D model
had the maximum impacts on the changes in objective and
constraint functions. Thus, the meta-model was created using
the Kriging method based on dominant variables, see Fig. 10.
The following diagrams and Table 5 show the evaluation of
the created model. The noticeable compatibility between the
results of the meta-model and the high-fidelity 3D model fol-
lows the considerable guarantee of confidence in the optimiza-
tion process.
Selecting the most appropriate point as the initial value of
design, AFSQP algorithm is prepared for optimization. As this
is a gradient-based method, the algorithm attempts to con-
verge to the global optimal point in each optimization step.
Finally, the optimization is terminated after 895 steps. Table 6
shows the summarized optimization solutions that constitute
initial optimal and last solutions.
To ensure the comprehensiveness of the obtained solution,
optimization should begin from another initial point. The opti-
mization process is repeated by selecting the 10 different initial
points. As predicted before, the algorithm was finally con-
verged to the best solution in this case and other initial points
1172 P. MOHAMMAD ZADEH, M. SAYADI

Table1 Design variables.

Design variable Description and unit Initial value Upper bound Lower bound
x1 Sweep angle of leading edge (°) 41.84 46.84 36.84
x2 Minor span (mm) 900 950 850
x3 Semi span (mm) 2190 2290 2090
x4-9 Section twist angles (°) 0 5 5
x10-n Control point coordinate in section (mm) Related to section

Table 2 Design parameters.

Design parameter Description and unit Value
P1 Aspect ratio 4.45
P2 Speed (km/h) 195
P3 Angle of attack (°) 0

Fig. 8 Grid convergence study.

Table 4 CFD solver properties.

Velocity (km/h) 200
Solver and Density based Energy equation K-epsilon of
Fig. 6 Geometric design variables in 3D model. equations viscous model
Time Steady
Turbulence Intensity and viscosity ratio
method

Table 3 Boundary layer specification.

Model type 3D model
First layer height (mm) 0.4
Maximum layer 10
Growth rate 1.2
Boundary layer depth (mm) 10

Fig. 9 DoE points distribution.

were conducted to a worse optimal point. This certifies com-

prehensiveness of the optimization process. Table 7 summa-
Fig. 7 UAV mesh mapping and center plane cells in 3D model. rizes the best results of these optimizations.
An efficient aerodynamic shape optimization 1173

Table 7 Optimal solutions for various initial points.

Number CD CL
1 0.008355 0.425478
2 0.008505 0.437833
3 0.007998 0.487335

Fig. 11 Conventional optimized blended wing body UAV

dimensions.

3.2. Efficient aerodynamic shape optimization using multi-

fidelity models for a BWB UAV

This section focuses on the implementation of the proposed

framework of ‘‘efficient aerodynamic shape optimization” for
the introduced samples. The aerodynamic optimization of
Fig. 10 Meta-model built by Kriging method.
the introduced BWB UAV is implemented step-by-step to
show the impact of the methodology expressed in Section 2.
The 3D model created in the earlier stage is used for deriving
the low-fidelity model, extracting correction factors of the 2D
Table 5 Evaluation of constructed meta-models. model, and updating local meta-models in the optimization
Name MAE MRE R-square process.
C^D 1.58  104 1.90  102 0.904
3.2.1. 2D geometry model
C^L 0.112 0.325 0.898
The 2D model with lower fidelity was created by selecting 15
Note: MAE = Mean Absolute Error; MRE = Mean Relative sections of the UAV 3D model. In fact, airplane sections are
Error.
analyzed in the airfoil 2D model using the panel method. As
the BWB geometry is symmetrical, the sections were evaluated
only in half of the UAV symmetry. Therefore, seven similar
sections were deleted and the number reduced to eight. The
first section was compatible with the UAV’s XOZ plane while
Table 6 Optimal solutions of conventional aerodynamic other sections were parallel to it with an identical distance
shape optimization for BWB UAV. between them. The initial distance of the sections was 300
Iteration number CD CL mm, which varies according to wingspan size during optimiza-
tion. Fig. 12 presents a view of the sections distributed in the
0 0.008354 0.426588
2D model.
652 0.007756 0.464934
895 0.007755 0.467228
The important point in modeling the low-fidelity model
used in this article involves the variables related to wing length
and sweep angle. The earlier works on wing and/or BWB UAV
would overlook these variables to simplify optimization com-
plexities if a low-fidelity model and/or a 2D model are
In addition, the optimal point was evaluated in the high- used.56,57 In these studies, only the 2D sections are analyzed
fidelity 3D model. The evaluation error guarantees the accept- and the result is generated regarding the entire BWB UAV.
ability of the final result. The final shape of the BWB is shown However, the effect of sweep angle on the performance and
in Fig. 11. ultimate stability of a BWB, especially at high speeds, cannot
1174
Fig. 12 Section positioning in 2D model.
be overlooked.58 Therefore, it is expected that low-fidelity
model cannot affect the accuracy of optimization results.
Fig. 13 compares the analytical structure of a 3D model and
a 2D model. The figure shows that the airplane structure is
compared with the main model of the airplane in a 2D model
(gray sections). It is clear that the lift and drag coefficients
extracted from the 2D model differ from those of the main
model and they have lower accuracies. However, the correc-
tion factor is determined as 0.85 for the 2D model after exam-
ining the differences in the analyses of the two models under
similar conditions.
Based on investigation on 2D and 3D models (Figs. 14–17),
there are a number of minor differences between lift and drag
correction factors, which are approximated to 0.85.
The coefficients CL and CD for two-dimensional and three-
dimensional models are shown in Table 8. The coefficients CD
and CL for two-dimensional and three-dimensional models are
shown in Fig. 14. The ratio of the 3D drag coefficient to the 2D
drag coefficient and the ratio of the 3D lift coefficient to the 2D
lift coefficient are shown in Fig. 15.
As shown in Fig. 15(a), drag coefficient changes within the
range of [0.82, 0.865]. Due to the range of 3D/2D drag ratio
values, the correction factor is determined to be 0.85. As
pointed in Fig. 15(b), lift coefficient changes within the range
of [0.84, 0.86]. Due to the range of 3D/2D lift ratio values,
the correction factor is determined to be 0.85.
As explained above, the correction factor is determined as
0.85 for the 2D model after examining the differences in the
analyses of the two models under similar conditions for lift
and drag coefficients.
A total of 20 points are defined as variables in each section,
whose airfoil varies by changing the specifications of the points
Fig. 13 Comparison between 2D and 3D model
implementations.
P. MOHAMMAD ZADEH, M. SAYADI
in X and Z directions (Fig. 16). Changing the angle of attack
for each section also indicates a change in the angle of wing
twist.
Thus, 327 variables are defined in this model, and changes
in them would result in changes in their aerodynamic struc-
tures (Fig. 6). A triangular grid provides further acceptable
results for meshing such a 2D model (Fig. 17).
This model evaluates mesh independence versus the number
and dimensions of the elements. This examination is based on
the diagram in one of the eight sections. The convergence
shown by Fig. 18 indicates the guarantee of accuracy for such
a meshing. It should be noted that BWB’s CD is the result of
drag coefficient in all sections.
Using the boundary layer specifications shown by Table 9,
the model is prepared for the panel method. The constraints
and solvers for analyzing the 2D model are selected completely
similar to the 3D model (Table 4). Therefore, the comparison
of the results obtained from the two optimization frameworks
is valid.
As the 2D model is capable of conducting a high-speed
analysis, it was used for extracting the initial design, determin-
ing dominant variables, and generating the initial global meta-
model.
3.2.2. Constructing global meta-models
To specify the variables’ effectiveness in drag coefficient reduc-
tion and increase the lift coefficient for deriving global meta-
models, CFD solution of the low-fidelity model was collected
for the 200 points obtained from uniform Latin hypercube
strategy.
At this stage, the variables with the maximum effectiveness
on aerodynamic coefficients were specified and other variables
were deleted. ANOVA was performed on about 115 acceptable
points obtained from this solution, showing that 25 variables
in drag coefficient and 28 variables in changes of lift coefficient
had maximum influence among 327 initial variables. There-
fore, the initial variables were replaced by the effective vari-
ables. Comparison between the numbers and types of the
selected variables and the dominant variables in the earlier
optimization process not only indicates an accurate modeling
of the 2D model but also promises an accurate optimization
with shorter computational time.
The meta-models were then created from the feasible points
for determining the values of lift and drag coefficients by uti-
lizing Kriging method. Fig. 19(a) shows the response of the
extracted surface for objective function (CD) versus one of
An efficient aerodynamic shape optimization
Fig. 14 Drag coefficient and lift coefficient comparison for 2D and 3D models.
Fig. 15 Ratio of 3D drag coefficient to 2D drag coefficient and 3D lift coefficient to 2D lift coefficient.
Fig. 16 Shape design variable definition in an arbitrary section.
Fig. 17 Meshing sections in 2D space.
twist angles and chord length of the root section. Fig. 19(b)
shows the constraint (CL) response surface in proportion to
one of the twist angles and the chord length of the middle sec-
tion. As the figures highlight, the appropriate response was
created in such a way that the effect of local extrema could
1175
be overlooked in the final shape and the global extrema would
be the only option for optimization.
In addition, the accuracy and ability of the meta-models are
verified (Table 10). The results indicate that the high efficiency
of the optimization process is guaranteed due to the high accu-
racy of the meta-models. Therefore, the two frameworks are
similar as far as the point selection and model creation method
are concerned.
3.2.3. Implementing efficient aerodynamic shape optimization
At this stage, the meta-models are replaced with the aerody-
namic model in the optimization framework. Fig. 20 illustrates
the differences between initial conceptual design and this
point. The optimum point based on DoE is selected as the ini-
tial design point in the optimization process.
Optimization starts through the initialization of design
variables and determination of their variation boundaries.
The optimization process continues until the value of the opti-
mal point becomes constant in two successive stages of the
optimization. The termination criterion for optimization
framework is defined as a difference smaller than 0.0001
between the optimal point and the initial point and/or an eval-
uation accuracy (Rs ) greater than 0.998. This value is selected
due to the reported accuracy in the studies conducted in this
field. The acceptable value for evaluation error is the minimum
1176 P. MOHAMMAD ZADEH, M. SAYADI

Table 8 Drag and lift coefficients for 2D and 3D models.

Iteration number 2D model 3D model
Drag coefficient Lift coefficient Drag coefficient Lift coefficient
1 0.009864 0.578673 0.007998 0.487330
2 0.010082 0.557035 0.008582 0.475880
3 0.010174 0.529642 0.008613 0.445507
4 0.009925 0.528260 0.008294 0.451906
5 0.009327 0.444217 0.008092 0.374382
6 0.009719 0.458293 0.008301 0.390209
7 0.009447 0.487998 0.008167 0.415850
8 0.009406 0.446446 0.007946 0.377358
9 0.009853 0.512496 0.008426 0.440444
10 0.009845 0.510050 0.008508 0.437833

Fig. 18 Grid convergence study in 2D section.

Table 9 Boundary layer properties in 2D model.

Model type 2D model
Sections Sections Section 8
1 to 3 4 to 7
First layer height (mm) 0.4 0.4 0.1
Maximum layer 10 8 4
Growth rate 1.2 1.2 1.2
Boundary layer depth (mm) 10 6.5 0.5

values of Rs for the objective function (CD) and the constraint

function (CL). For instance, in the first stage, the Rs values for Fig. 19 Initial global meta-model built by Kriging method.
the objective function and constraint function are 0.965 and
0.95 respectively. Finally, the Rs value used in Eq. (9) min
{0.965,0.95} = 0.95. The maximum number of the selected
iterations is seven with respect to the high accuracy of the Table 10 Evaluation of predictive capability of constructed
meta-models (Table 10). In this case, there are six sample meta-models.
points in the first stages of optimization.
The need for updating is minimized due to a specific Name MAE MRE R-square
5 3
method for modeling the low-fidelity model, modifying it for C^D 2.47  10 2.72  10 0.996
a high overlap with the high-fidelity model, and specifically C^L 3.42  103 7.98  103 0.973
selecting the initial point of design. In other words, local
An efficient aerodynamic shape optimization
Fig. 20 Initial design scheme versus initial conceptual design.
modification of the meta-model is merely carried out in three
iterations and the accuracy of the model is reported at an
acceptable level with convergence to the optimal point. This
shows the capability of the optimization algorithm to find
the global optimal point. Otherwise, the meta-model has to
be updated at any optimization iteration due to the major
change in the optimal point neighborhood.
Finally, the cost function is fixed in the minimum value in
the fourth iteration after three steps for updating the meta-
models. Table 11 shows the data related to each stage of the
optimization process. Stage 0 is selected as equal to the starting
point of the design. Table 11 shows the mean variable bound-
aries. With the accuracy improving and variable bounds mov-
ing in the neighborhood of optimal point candidate, the mean
of changes reduces quickly to make the convergence of
response surface to the optimal point possible.
Table 12 shows the selected sample points at the different
stages of optimization. In fact, both the number of sample
points and its corresponding Rs converge to 1. Undoubtedly,
after some optimization iteration, the response surface around
the optimal point becomes limited for finally being mapped to
the optimal point. In this mode, the continuation of the opti-
mization process is meaningless.
Table 11 shows that the optimal point found in the fourth
iteration has a minimum drag coefficient as well as a greater
lift coefficient. Comparison of Figs. 5 and 21 reveals the final
difference between the initial and optimized geometries. In
spite of selecting the initial point near the optimal point
and having an approximately fixed value for wingspan, the
thickness and length of the chord of sections faced some
changes. With respect to the relative changes of the values,
the considerable difference in drag coefficient indicates
the importance of geometric shape in the aerodynamic
Table 11 Optimal solutions of efficient meta-model based
aerodynamic shape optimization for BWB UAV.
Iteration number CD CL D value average (mm)
0 0.008241 0.422954 12
1 0.007835 0.486346 3.57
2 0.007768 0.467067 1.27
3 0.007747 0.466572 0.034
4 0.007748 0.466530 0.018
1177
Table 12 Number of sampling points and Rs in optimization
process.
Iteration number Rs N
1 0.958 6
2 0.984 3
3 0.998 2
Fig. 21 Optimized BWB UAV dimensions.
performance of the UAV. This indicates the importance of
optimization in this phase of aircraft design.
Another aspect of the proposed framework is its accuracy
via the conventional method. Although the 3D model (high-
fidelity) is utilized in conventional strategy, updating low-
fidelity RSM model based on 3D model and modifying vari-
able bounds during optimization are recommended to improve
framework capabilities, which will avoid unfeasible space and
converge to optimal solution in minimum time without affect-
ing accuracy, as shown in Tables 6 and 10.
Fig. 22 shows the process of changing the UAV geometric
model during the optimization process. While airfoil geometric
changes are not noticeable at some points, some effective
changes have been made at some points, especially in the trail-
ing edge. This is true for any stage of optimization. As shown
in Fig. 22, some sections are subject to further changes at any
stage of optimization.
With respect to the uncertain and unpredictable process
concerning the deformation of sections, the changes finally
lead to drag coefficient reduction. The option of a designer
to achieve a proper optimal point in the optimization process
is among the features of this method. In other words, with
respect to the improvement of the accuracy of local models
in each iteration related to the computational costs and time,
the designer could continue or terminate the process. More-
over, the response surface is modified. However, it might not
be the most optimal solution. As compared to the optimal
solution found in the previous method, the result obtained
from the proposed framework has altogether better
performance.
Taking modeling time into account, the computational time
of the proposed optimization method utilizing the high-fidelity
model is approximately 3.5 h. Certainly, this period is much
longer if the main model (high-fidelity model) is used, as seen
in Table 13.
1178
Fig. 22 BWB geometric modification during efficient aerodynamic shape optimization using multi-fidelity models.
Table 13 Comparison of computational time between conventional and proposed aerodynamic optimization processes.
Optimization type 2D model preparation time 2D model evaluation time
Conventional
Proposed 5  60 200  10
Of course, the final period extends to one day if the compu-
tational time for the pre-processing stage (creating the high-
fidelity model) is included. Table 13 shows the computational
time for the proposed framework in the initial preparation and
optimization stage as compared to the conventional optimiza-
tion method. The considerable reduction of the computational
time with no influence on solution accuracy exhibits the supe-
riority of the proposed framework.
4. Conclusions
References
(1) The construction of the initial meta-models is based on
the low-fidelity 2D panel model and the meta-models
in the neighborhood of design point are locally modified
based on the high-fidelity 3D CFD model in case of
insufficient accuracy.
(2) As compared to creating an appropriate model using the
high-fidelity model, the low-fidelity meta-model has
much lower computational costs. Therefore, with
respect to all the variables’ effectiveness in the 2D
model, modeling accuracy is improved considerably sim-
ilar to the 3D model.
(3) The method’s major capability can be considered for
those problems in which it is difficult to create a reliable
meta-model in the whole design space. the points of
strength of the presented framework is highlighted for
those problems in which a designer selects fairly large
variable boundaries for some reasons, such as inappro-
priate conceptual design and a lack of knowledge on
the possible bound of the design optimal point.
(4) The proposed aerodynamic shape optimization frame-
work based on multi-fidelity modelling is compared with
the conventional aerodynamic shape optimization using
high-fidelity (3D CFD) model and evaluated for a BWB
benchmark case. Despite relatively low changes in the
aerodynamic shape of sections and the overall dimen-
sions of the aircraft, drag coefficient is reduced
considerably.
P. MOHAMMAD ZADEH, M. SAYADI
3D model evaluation time Optimization time Total time
200  1260 5 144005
13  1260 10 11670
(5) The comparative analysis between the results of the pro-
posed and conventional approaches shows the absolute
superiority of the proposed framework in terms of over-
all computational time. Meanwhile, the optimal design
point obtained using the proposed approach clearly
shows improvement in the accuracy of the results due
to the compatibility of the local meta-model and cus-
tomization of the appropriate design space.
1. Okonkwo P, Smith H. Review of evolving trends in blended wing
body aircraft design. Prog Aerosp Sci 2016;82:1–23.
2. Wang Z, Huang W, Yan L. Multidisciplinary design optimization
approach and its application to aerospace engineering. Chin Sci
Bull 2014;59(36):5338–53.
3. Liu M, Hu Y. Robust design consideration in the multi-
disciplinary optimization of a flying wing UAV design. AIAA a
erospace sciences meeting. Reston: AIAA; 2014.
4. Boozer CM, Tooren MJL, Elham A. Multidisciplinary aerody-
namic shape optimization of a composite blended wing body
aircraft. AIAA/ASCE/AHS/ASC structures, structural dynamics,
and materials conference. Reston: AIAA; 2017.
5. Pan Y, Huang J, Li F, Yan C. Integrated design optimization of
aerodynamic and stealthy performance for flying wing aircraft.
Proceedings of the international multi conference of engineers and
computer scientists; 2017.
6. Morris A, Arendsen P, LaRocca G, Laban M, Voss R, Hon H.
MOB— A European project on multidisciplinary design opti-
mization. Proceedings of the international congress of the aeronau-
tical sciences; 2004.
7. Koziela S, Tesfahunegnc YA, Leifsson L. Expedited constrained
multi-objective aerodynamic shape optimization by means of
physics-based surrogates. Applied Mathematical Modelling 2016;
40(15-16): 7204–15.
8. Lyu Z, Martins JRRA. Aerodynamic design optimization studies
of a blended-wing-body aircraft. J Aircraft 2014;51(5):1604–17.
9. Truong H, Zingg DW, Haimes R. Surface mesh movement
algorithm for computer-aided-design-based aerodynamic shape
optimization. AIAA J 2014;54(2):542–56.
An efficient aerodynamic shape optimization
10. Maruyama D, Liu D, Görtz S. An efficient aerodynamic shape
optimization framework for robust design of airfoils using
surrogate models. ECCOMAS congress 2016 on computational
methods in applied sciences and engineering. 2016 June 5-10; Crete
Island, Greece: National Technical University of Athens; 2016.
11. Hunsaker DF, Phillips WF, Joo JJ. Aerodynamic shape optimiza-
tion of morphing wings at multiple flight conditions. AIAA
aerospace sciences meeting . Reston: AIAA; 2017.
12. Reist TA, Zingg DW. Aerodynamic design of blended-wing-body
and lifting-fuselage aircraft. Proceedings of the AIAA applied
aerodynamics conference. Reston: AIAA; 2016.
13. Gan W, Zhou Z, Zhang X. Airframe-intake-exhaust integration
design of flying wing using multi-bump strategy. Aerosp Eng
2016;231(13):2396–407.
14. Mader CA, Martins JRRA. Stability-constrained aerodynamic
shape optimization of flying wings. Journal of Aircraft 2013; 50
(5):1431–49.
15. Lyu Z, Kenway GK, Martins J. RANS-based aerodynamic shape
optimization investigations of the common research model wing.
AIAA aerospace sciences meeting. Reston: AIAA; 2014.
16. Li E, Wang H, Ye F. Two-level multi-surrogate assisted opti-
mization method for high dimensional nonlinear problems. Appl
Soft Comput 2016;46:26–36.
17. Paiva RM, Crawford C, Sul A. Robust and reliability-based
design optimization framework for wing design. AIAA J 2014;52
(4):711–24.
18. Rodr´ıguez-Cort´s H, Arias-Montaño A. Robust design optimiza-
tion of a small flying wing planform based on evolutionary
algorithms. The Aeronautical Journal 2012; 116(1176):175–88.
19. Wang GG, Dong Z, Aitchison P. Adaptive response surface
method–A global optimization scheme for approximation-based
design problems. Eng Optim 2001;33(6):707–34.
20. Koziel S, Leifsson LT. Automated selection of low-fidelity models
for rapid aerodynamic shape optimization using physics-based
surrogates. AIAA/ASCE/AHS/ASC structures, structural dynam-
ics, and materials conference. Reston: AIAA; 2017.
21. Huang M, Yang X, Peng X. Efficient variable-fidelity multi-point
aerodynamic shape optimization based on hierarchical kriging.
AIAA aerospace sciences meeting. Reston: AIAA; 2017.
22. Xia C, Tao Y, Jiang T, Chen W. Multi-objective shape optimiza-
tion of a hypersonic lifting body using a correlation-based
transition model. Aerosp Eng 2016;230(12):2220–32.
23. Yao W, Chen X, Huang Y, Michel MV. A surrogate based
optimization method with RBF neural network enhanced by
linear. Optim Methods Softw 2014;29(2):406–29.
24. Jones DR. A taxonomy of global optimization methods based on
response surfaces. J Global Optim 2001;21:345–80.
25. Forrester AI, Keane AJ. Recent advances in surrogate-based
optimization. Prog Aerosp Sci 2009;45(1–3):50–79.
26. Laurenceau J, Meaux M, Montagnac M, Sagaut P. Comparison of
gradient-based and gradient-enhanced response-surface-based
optimizers. AIAA J 2010;45(8):1–24.
27. Toal DJJ, Keane AJ. Efficient multi-point aerodynamic design
optimization via co-kriging. J Aircraft 2011;48(5):1685–95.
28. Liu J, Song WP, Han ZH, Zhang Y. Efficient aerodynamic shape
optimization of transonic wings using a parallel infilling strategy
and surrogate models. Struct Multidiscip Optim 2016;55(3):925–43.
29. Wujek BA, Renaud JE, Batill SM. A concurrent engineering
approach for multidisciplinary design in a distributed computing
environment. Multidiscip Des Optim: Proc Appl Math Ser
1997;80:189–208.
30. Wujek BA, Renaud JE. New adaptive move-limit management
strategy for approximate optimization, Part 1. AIAA J 1998;36
(10):1911–21.
31. Pérez VM. Reduced-oder response surface approximations. In
Homotopy-managed interior-point optimization [dissertation].
Chicago: University of Notre Dame; 2003.
1179
32. Conn AR, Gould NIM, Toint PL. Trust region methods. Philadel-
phia: Society for Industrial and Applied Mathematics; 2000.
33. Poole DJ, Allen CB, Rendall TCS. High-fidelity aerodynamic
shape optimization using efficient orthogonal modal design
variables with a constrained global optimizer. Comput Fluids
2017;143(17):1–15.
34. Ahmed MYM, Qin N. Surrogate-based aerodynamic design
optimization: use of surrogates in aerodynamic design optimiza-
tion. Proceedings of the aerospace sciences & aviation technolog.
2009.
35. Nejat A, Mirzabeygi P, Panahi MS. Airfoil shape optimization
using improved multiobjective territorial particle swarm algorithm
with the objective of improving stall characteristics. Struct
Multidiscip Optim 2014;49(6):953–67.
36. Lyu Z, Xu Z, Martin JRRA. Benchmarking optimization
algorithms for wing aerodynamic design optimization. Proceedings
of the international conference on computational fluid dynamics
(ICCFD8); 2014 July 14–18; Chengdu, Sichuan, China; 2014.
37. Zhang M, Wang C, Rizzi A, Nangia R. Hybrid feedback design
for subsonic and transonic airfoils and wings. AIAA aerospace
sciences meeting. Reston: AIAA; 2014.
38. Kuntawala NB. Aerodynamic shape optimization of a blended-
wing-body aircraft configuration [dissertation]. Toronto: Univer-
sity of Toronto; 2011.
39. Ricco L, Rigoni E, Turco A. Smoothing spline ANOVA for
variable screening. Dolomites Res Notes Approx 2013;6:130–9.
40. Taguchi G, Chowdhury S, Wu Y. Taguchi’s quality engineering
handbook. New York: Wiley; 2004. p. 1736.
41. Hong HS, Hickernell FJ. Algorithm 823: Implementing scrambled
digital sequences. ACM Trans Math Softw 2003;29(2):95–109.
42. Ye KQ, Li W, Sudjianto A. Algorithmic construction of optimal
symmetric latin hypercube designs. Stat Plann Inference
2000;90:145–59.
43. Myers RH, Montgomery A. Response surface methodology.
Process and product optimization using designed experiments. New
York: John Wiley and Sons; 2001.
44. Breitkopf P, Naceur H, Rassineux A, Villon P. Moving least
squares response surface approximation: formulation and metal
forming applications. Comput Struct 2005;83(17–18):1411–28.
45. Ross TJ. Fuzzy logic engineering application. 2nd ed. New
York: John Wiley and Sons; 2004.
46. Couckuyt I, Deschrijver D, Dhaene T. Fast calculation of
multiobjective probability of improvement and expected improve-
ment criteria for pareto optimization. Global Optim 2014;60
(3):575–94.
47. McDonald DB, Grantham WJ, Tabor WL. Global and local
optimization using radial basis function response surface models.
Appl Math Model 2007;31:2095–110.
48. Persson J. Efficient optimization of complex products, a simula-
tion and surrogate model based approach [dissertation]. Linköp-
ing: Linköping University; 2015.
49. Yaoa S, Guob D, Sunb Z, Yang G. A modified multi-objective
sorting particle swarm optimization and its application to the
design of the nose shape of a high-speed train. Eng Appl Comput
Fluid Mech 2015;9(1):513–27.
50. Paiva RM, Carvalho ARD, Crawfordy C, Sulemanz A. A
comparison of surrogate models in the framework of an MDO
tool for wing design. Proceedings of the 50th AIAA/ASME/ASCE/
AHS/ASC structures, structural dynamics, and materials confer-
ence. Reston: AIAA; 2009.
51. Sivasubramani S, Swarup KS. Hybrid SOA-SQP algorithm for
dynamic economic dispatch with valve-point effects. Energy
2010;35:5031–6.
52. Turco A. Adaptive filter SQP-description. Learning and intelligent
optimization. Berlin Heidelberg: Springer; 2010. p. 68–81.
53. Turco A. Adaptive filter SQP - benchmark tests. ESTECO; 2009.
Report No.: Tec. Rep. 2009–003.
1180
54. Macconi M, Morini B, Porcelli M. Trust-region quadratic
methods for nonlinear systems of mixed equalities and inequali-
ties. Elsevier Sci Publ 2009;59(5):859–76.
55. Im J, Park J. Stochastic structural optimization using particle
swarm optimization, surrogate models and Bayesian statistics.
Chin J Aeronaut 2013;26(1):112–21.
56. Rashad R, Zingg DW. Aerodynamic shape optimization for
natural laminar flow using a discrete-adjoint approach. AIAA
Journal 2016; 54(11).
P. MOHAMMAD ZADEH, M. SAYADI
57. Tesfahunegn YA, Koziel S, Gramanzini JR, Hosder S, Han ZH,
Leifsson L. Application of direct and surrogate-based optimiza-
tion to two-dimensional benchmark aerodynamic problems: A
comparative study. AIAA aerospace sciences meeting. Reston:
AIAA; 2015.
58. Mader A, Martins JRRA. Optimal flying wings: a numerical
optimization study. Proceedings of the AIAA/ASME/ASCE/AHS/
ASC structures, structural dynamics and materials conference.
Reston: AIAA; 2012.