Paper

Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016)

DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491

A Highly Efficient Aeroelastic Optimization Method Based on a

Surrogate Model
Wan Zhiqiang*, Wang Xiaozhe** and Yang Chao***
School of Aeronautic Science and Engineering, Beihang University, Beijing 100191, China

Abstract
This paper presents a highly efficient aeroelastic optimization method based on a surrogate model; the model is verified
by considering the case of a high-aspect-ratio composite wing. Optimization frameworks using the Kriging model and
genetic algorithm (GA), the Kriging model and improved particle swarm optimization (IPSO), and the back propagation
neural network model (BP) and IPSO are presented. The feasibility of the method is verified, as the model can improve the
optimization efficiency while also satisfying the engineering requirements. Moreover, the effects of the number of design
variables and number of constraints on the optimization efficiency and objective function are analysed in detail. The accuracy
of two surrogate models in aeroelastic optimization is also compared. The Kriging model is constructed more conveniently,
and its predictive accuracy of the aeroelastic responses also satisfies the engineering requirements. According to the case of a
high-aspect-ratio composite wing, the GA is better at global optimization.

Key words: Kriging, back propagation neural networks, genetic algorithms, improved particle swarm optimization

1. Introduction tailoring of a forward-swept composite wing and a high-

Requirements regarding weight reduction and the
application of composites are increasing rapidly in modern
aircraft design, resulting in more flexible structures and a
notable aeroelastic effect. Aeroelastic optimization must
be performed to reduce unfavourable impacts and take
advantage of aeroelastic effects, thereby increasing the
overall performance of an aircraft. In the preliminary phase of
aircraft design, aeroelastic optimization can be used as a tool
to reduce weight and increase aeroelastic performance. The
design efficiency is enhanced, which is particularly notable in
composite structures [1].
In the past several years, many aeroelastic optimization
methods have been suggested and applied to practical
aircraft design. Livne et al. suggested the active control of a
composite wing that combined aerodynamic, structure and
control in aeroelastic tailoring [2]. Guo studied the aeroelastic
tailoring design of a composite back-swept wing using a
genetic algorithm (GA) [3]. Wan studied the aeroelastic
aspect-ratio composite wing using a genetic/gradient-based
hybrid algorithm; furthermore, the relationship between the
objective function and design variables was studied [4].
These studies indicate that in aeroelastic optimization,
many constraints, including flutter, divergence, natural
frequency, deformation, aileron efficiency, strength and flight
loads, are considered simultaneously, and the aeroelastic
responses typically have to be computed repeatedly to search
for the optimum objective function. The structural model
becomes more complex with the development of aircraft
design, and the quantities of design variables and constraints
are increasing rapidly. All of these factors will lead to an
increased computational burden and decreased efficiency
[5]. In aeroelastic optimization, excessive computational
consumption affects the optimization efficiency directly.
Therefore, aeroelastic optimization with a proper surrogate
model is crucial to reducing the computational burden. Many
surrogate models have been developed in recent years. The
Kriging model and back propagation neural network model

This is an Open Access article distributed under the terms of the Creative Com- * Professor
mons Attribution Non-Commercial License (http://creativecommons.org/licenses/by- ** Ph. D Candidate
nc/3.0/) which permits unrestricted non-commercial use, distribution, and reproduc- *** Professor, Corresponding author: yangchao@buaa.edu.cn
tion in any medium, provided the original work is properly cited.

Received: April 27, 2016 Revised: November 17, 2016 Accepted: November 21, 2016
Copyright ⓒ The Korean Society for Aeronautical & Space Sciences 491 http://ijass.org pISSN: 2093-274x eISSN: 2093-2480

(491~500)16-067.indd 491 2017-01-02 오전 5:13:25

 2.2 Equation for Flutter Analysis
2.2
2.2Equation
Equationfor
forFlutter
FlutterAnalysis
Analysis
The equation of the p  k flutter analysis method, which is applicable f
2.2 Equation for Flutter Analysis
The equation of the p  k
The equation of the p  k flutter analysis method, which isapplicable
flutter analysis 2.2
method, Equation
which isfor Flutter Analysis
applicable for
forthe theoptimization,
optimization,isis
Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016) stated as [10]:
The equation of the p  k flutter analysis method, which is applicable for the optimiz
stated The equation of the p  k flutter analysis method, which is applicable for the optimiz
statedasas[10]: [10]:
stated as [10]:  V  22 22 V 1 22  p II  
(BP) are widely applied because
VV 2 2
of their high computational stated as [10]:
applicable for   p
theoptimization, M  p B K
isstated  as V [10]:  Q RR
 Q  u  0
11 2 2 RR pp I I
22
VV 
  b b 2  k  

speed and high predictive  accuracy.pp M M ppBBKK VV QQ  QQ uu  0022 (2)
(2)

The Kriging model
bb
and bb
BP model are constructed based
22  kk  V  2
 V  pp 2 M  V pB  K  11 V 22  Q RR  pp Q II  u 
V
00
M  p B  K
bb V is the 22flow speed,   V Q  Q u  (2)
 bb  where  kkb isthe
  length of the reference chord, p is
on the given samples, and then, the values of unknown
where VV isisthe
where theflow speed, bb isisthe
flowspeed, thelength
lengthofofthe thereferencereferencechord, chord, pp isisthe eigenvalue,  isis
theeigenvalue,
individuals can be predicted rapidly based on the models. whereV Vis the
where V the
where is the
is the flow flow
density
flow speed, of air,b b
speed,
speed, bkis is is the
is
the the
the
length
reduced
length
length of the the reference
of frequency,
of the reference
chord,
referenceu chord,
pp is
is the vector the
of eigenvalue
is the the gener
eigenvalue
The Kriging model and BP model are currently k being applied
model are currently inbeing applied thethe density
density
in aircraft optimization of
of air,
air, k isis
design. Shinkyu the
the reduced
reduced frequency,
frequency,
et al. developed athe uu
chord, is is
methodthe
the p vector
vector
is the of of thethe
eigenvalue, generalized
generalized k is structural
structural
the density degree
degreeof air, u is the
model
aircraft areoptimization
currently beingdesign. appliedShinkyu in aircraftetoptimization
al. developed design. Shinkyu density of
et al. developed air, k is
of freedom, aisand
method M is the
the reduced matrix of
frequency, u generalized
is the vectorunsteady aerodynamic
of the generalized force
structura
the density
reduced of air, k u
frequency, the
is reduced
the vector frequency,of theu generalized is the vector of the generalized structura
a method that combinedof of freedom,
freedom,
the and
and
Kriging
that combined the Kriging model and GA in the aerodynamic optimization of a two-dimensional M M modelisisthe
the matrix
andmatrix
GA of
of
in generalized
generalized
the unsteady
unsteady aerodynamic
aerodynamic forces.
forces. The The superscripts
superscripts R R
that combined the Kriging model and GA in the aerodynamic optimization of freedom,
structural
of freedom, ofdegree
andand
and IM
a two-dimensional
Mof isfreedom,
represent the matrix
is the matrix the real of generalized
and andMimaginary
of generalized
is the unsteadyparts,aerodynamic
matrix
unsteady respectively.
of
aerodynamic
forces. The superscr
The commercial
forces. The superscr
aerodynamic optimization of a two-dimensional airfoil
airfoil profile [6]. Mashio et al. studied andand I
multi-objectI represent
represent the
the
optimization real
real and
and
of imaginary
a imaginary
multistage parts,
parts,
profilerespectively.
respectively.
generalized
crevasse The
The commercial
unsteady commercial software
aerodynamicsoftware MSC.
MSC.
forces. Nastran
Nastran
The superscripts
profile
airfoil[6]. Mashio
profile [6]. etMashioal. studied et al. multi-object
studied multi-object optimizationoptimization ofand I represent
a multistage the real
profile
is utilized for andflutter
crevasse
the imaginary analysis. parts, respectively. The commercial software MSC.
and
R and II represent
represent the real
the realand andimaginary imaginary parts,
parts,respectively.
respectively. The commercial software MSC.
of a multistage profile is is crevasse
utilized
utilized forfor theusing
the flutter
flutter the Kriging model
analysis.
analysis.
using the Kriging model [7]. Rai et al. studied the robust optimization design of an aerodynamic The is shape
utilized
commercial for the flutter analysis.
[7].usingRai theet Kriging
al. studiedmodelthe [7]. robust
Rai et al.optimizationstudied the robust design optimization
of design
is utilized the software
of anforaerodynamic flutter analysis. MSC. Nastran is utilized for the
shape
flutter analysis.
using the BP model an[8]. aerodynamic
Hass et al. analysed shape helicopter rotor loads using[8].the BP et model [9].
using the BP model [8]. using
Hass etthe al. BPanalysedmodel helicopter Hass rotoral.loads using the BP model [9].
2.3 Aeroelastic Optimization
analysed helicopter 2.3 rotor
2.3 loads using
Aeroelastic
Aeroelastic the BP model [9].
Optimization
Optimization
In this study, a highly efficient optimization method based on a surrogate model is proposed 2.3Aeroelastic to
Aeroelastic Optimization
InIn thisthis study,
study, a highly
a highly efficient
efficient optimizationmethod
optimization methodbased based on a2.3 surrogate
2.3 Aeroelastic model isOptimization
Aeroelastic proposed
Optimization to
optimization is a typical optimization problem involving the sea
Aeroelastic
Aeroelastic optimization
optimization isisa atypical
typical optimization
optimization problem
problem involving
involving
Aeroelastic optimization thethe search
search
is a typical ofof design
design variables
variablesproblem involving the search of design v
solve the excessiveoncomputational
a surrogate model burden is in proposed
aeroelastictoresponse
solve the excessive computational burden in aeroelastic response Aeroelastic
solve the analysis.
excessive Many optimization Aeroelastic
analysis. Many
in optimization
optimization
the optimization
ndvdv -dimensionalis a typical is a optimization typical
optimization
space to minimizeoptimization
problem involving
the objective the search of design
function whilv
computational burden in aeroelastic response analysis. Many problem involving the search of design variables in the n -
methods are studied, and frameworks using inin the the nndvdv-dimensional
theKriging -dimensional
model and GA, space
spacethetoto minimize
minimize
Kriging model the
the inobjective
and objective
the ndv -dimensional
IPSO function
function while while space satisfying
satisfying
to minimize the nthe
the nconcon objective function while satisfying th
dv
methods are methods
optimization studied, and areframeworks
studied, andusingframeworks
the Kriging using model and GA, inthe theKrigingndv -dimensional
model to and space to minimize the
IPSO objective function while satisfying th
dimensional constraints
space [11]:
minimize the objective function while
and the BP model and theIPSO Kriging are model
presented. and GA,
All
constraints
constraints ofthe
these Kriging
[11]:
[11]: methods model are and
verifiedIPSO by and
considering the
constraints
satisfying case the [11]:n constraints [11]:
and the BP model and IPSO are presented. All of these methods are verified by considering
constraints min F (vthe
[11]: ) case
con
the BP model and IPSO are presented. All of these methods
of a high-aspect-ratio
areofcomposite bywing, and theFthe
F(optimum ofobjective function and efficiency areF (v )
a high-aspect-ratio
verified minmincomposite
considering v(v))case wing, aand the optimum objectivemin
high-aspect-ratio min function F (vs.t ) . efficiency
and c lower
lower
jj  gare jj (v )  c jj
upper
upper
j 1, 2,..., ncon con
lower upper
composite wing, s.t
and lower
s.t. . theccj optimization
lower
optimum ggj (j v(v))objectiveupper
ccj j
methodupper
1,1,on
jfunction
j 2,...,
2,..., s.t
nnconcon s.tmodel .. c lower  g
c jj  g jj (vv)lower ( v )  c upper
 c jj v  vj upper j 
 1, 2,...,
1, 2,..., nconn (3)
(3) (3)
compared. The results indicate that the aeroelastic j based a surrogate lower upper
i 1, 2,..., ndvdv
compared. The results indicate that the aeroelastic optimization method based on a surrogate ii model ii ii con
and efficiency are compared. vvilower The
lower
vresults upperindicate ithat
vi ivvi iupper i the n
1,1,2,...,
2,..., ndvdv
lower
vvilower   vvii 
upper
 vviiupper ii  1,
1, 2,...,
2,..., nndv
has engineering practicability,
aeroelastic and the optimization
optimization method
i
efficiency
based can on beaimproved
surrogate while maintaining i dv
has engineering practicability, and the optimization efficiency can be improvedwhere while F maintaining
(v ) is the objective function (structural mass in this study), gii
where F(v)(v is ))this isthe theobjective function (structural (structural mass in study), g (v ) are the co
mass F
model has engineering practicability,
FF(v(v) ) isis the and the optimization where objective
ggi (i v(v))function mass in this
accuracy. where
where the objective
objective function
function (structural
(structuralwhere mass inin(vthis
F is study),
study),
the objective are
are the
function constraintmass in this study), gii (v ) are the co
constraint
the (structural
accuracy.
efficiency can be improved while maintaining accuracy. this study), g (v) are the constraint functions
functions (such as the deformation of the wing tip, internal stress, aileron
i (such as the
functions (such as wing
the deformation internalof the wing tip, efficiency,
internal stress, aileron efficiency, and
functions
functions (such (such asas the the deformation
deformation ofof the wingdeformation
the wing functions
tip,
tip, internal(such
internal of
stress, the
stress, as aileron
the
aileron tip,efficiency,
deformation
efficiency, ofstress,
the
andandwing aileron
flutter),
flutter), tip, internal stress, aileron efficiency, and
lower
lower upper
upper
andlower flutter), cupper and c jj are
and arethe the lower lower and andupper upperbounds bounds of constraints, anduppervilower
lower
an
a
2. Methodology 2. Methodology lower upper cc lower and
and
c
c
jj
upper are
are
the
the
lower
lower
and
and
upper
upper
bounds
bounds
of
of
constraints,
constraints,
and
and
v
v
lower
and v
i
lower and viupper are th
are th
2. Methodology c lower
cj j and cj j
and c upper
are the lower and upper bounds
are the lower and upper boundsofofconstraints, of j
constraints,
constraints,
j
j
and
and vi i
j and v lower
lower
and v
and vi i are
and
upper
upper
are the
arethe lower
thelower
lower and upper
i
i i

bounds of the and upper

design bounds of the design variables [12], respectively, including th
2.1 Equation for Aeroelastic response Analysis and upper bounds of variables
the design[12], variables respectively,
[12], respectively,including including the skin thickne
2.12.1Equation
Equation for forAeroelastic
Aeroelastic
andand upper response response
bounds Analysis
ofof thetheAnalysis
design
design variables
variables [12], and upper bounds of thethe design skinvariables [12], webrespectively, including the skin thickne
upper bounds therespectively,
[12], respectively,
skin thickness, including
including
web thickness, the skin thickness,
and flangeweb
thickness, section area.
thickness, thickness,
and flange and flange
section area. section area.
The basic equation for the aeroelastic response analysis is [10]: thickness,
The penalty and flange function sectionisarea. a common method for the
The The basic
basic equation
equation for
thickness,thethe
thickness,
for aeroelastic
and andflange
aeroelastic flange response
section
response analysis
area.
sectionanalysis
area. is [10]:
is [10]:
calculation The penalty The
of constraint penalty
function is afitness, function
commonand ismethod
a common
the for function
the method can
calculationfor be
the
of calculation of constraint
constraint fitness, and the ff
 K aa  qQaa  ua  M aa uKa  qqQQax ux u Pa M The u  q
penalty Q u 
function P is a common method
(1) for the
The penalty function is a common method for the calculation of constraint fitness, and the
calculation(1) of constraint fitness, and the function
aaThe a penalty ax xfunction is a common method forexpressed the calculation of constraint fitness, and the function
 as [13]:
aa aa a a
can beasexpressed (1)
as [13]:
can be expressed [13]:
can be expressed as [13]:
awhere
where the subscript where means the structural can bebeexpressed
candisplacementexpressed asas[13]:
vector [13]: analysis freedom, K
set ofdisplacement is the   g j (Kv )  cislower 
the subscript a
the subscript a means
means the thestructural
structural displacement vector set ofaaanalysis freedom, j the 4 4
vector set of analysis freedom, Kaa is the matrix of structural exp  loweraa  g j (v )  clower 4
44   cj  1.0 
j

stiffness, q isisthe
matrix of structural stiffness, thedynamic
dynamic pressure, Qaa Q isaatheis matrix of aerodynamic influence  
matrix of structural stiffness, pressure, q is the dynamic the matrix
pressure, Qofaa is the matrix of aerodynamic  influence
aerodynamic influence coefficients, uaa is the vector  of R j (v ) 1.0 c lower
j  g j (v )  cupper j (4)
coefficients, ua is the vector ofdeformation,
structural deformation, M M is the matrix of structural mass, Q is 
coefficients, ua is the vector
structural aa is of structural deformation, M aa is the matrix of structural
the matrix aa of structural mass, ax
  gmass, j (v )  c j ax
Q is 
upper

Qaa is the matrix of unit aerodynamic loads, ux is the vector exp  upper  cupper
 j  g j (v )
the matrix of unit aerodynamic loads, u is the vector of aerodynamic trim parameters (e.g., angle of   c  1.0 
of theaerodynamic
matrix of unittrim aerodynamic
x parameters loads,(e.g., ux is angle
the vector of attack,
of aerodynamic trim parameters  j
(e.g., angle of
elevator deflection), which is used to define the deflection
attack, elevator deflection), which is used to define the deflection of the aerodynamic control surface
where R j is(vthe ) iscontrol the fitness surfaceof the j th constraint. If g j (v ) satisfies the specific co
where Rj(v) fitness of the j th constraint. If gi(v) satisfies
of attack, elevator deflection),
the aerodynamic controlwhich surface is used andtothe define the deflection
overall rigid of the aerodynamic
the specific constraint, the value of Rj(v) is 1.0, otherwise it
and the overall rigidmotion
motionof ofthethe aircraft,
aircraft, and and PPa isisthe thevector
vectorofofapplied applied loads.
loads. The commercial
and the overall rigid motion aof the aircraft, and Pa is the vector ofbelongs value ofto
applied R
loads. j (vThe
(0.0, )1.0),
iscommercial
1.0, andotherwise
the farther it belongs form to the(0.0,1.0)bounds, , the and the farther form the bound
The commercial software MSC. Nastran is utilized for the
closer to 0.0. The penalty function combines the responses
software MSC. Nastran is utilized
aeroelastic for the aeroelastic
response analysis. response analysis.
software MSC. Nastran is utilized for the aeroelastic response analysis. withto 0.0. the The penalty function
constraint bounds, combines which develops the responses with the constraint bounds, which
an effective
method for the constraint estimation. Fig.1 shows the fitness
2.2 Equation for Flutter Analysis 3 effective method for the constraint estimation. Fig.1 shows the fitness curve of a norm
3 curve of a normal constraint function with the bounds of
The equation of the p-k flutter analysis method, which is function
(-1.0, 1.0). with the bounds of ( 1.0,1.0) .

A variety of algorithms have been studied for aeroelastic optimization; this paper use

DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491 492 IPSO [14, 15].

2.4 Surrogate Model

Surrogate models are constructed by samples (including design variables and respons
(491~500)16-067.indd 492 2017-01-02 오전 5:13:26
 where
value of R j (v ) is 1.0, otherwise it belongs to (0.0,1.0) , and the farther form k (v )
fthe is a polynomial
bounds, the closer expression that is typically divided into zero-order (con
order (linear) and second-order (quadratic polynomial) polynomials, βk is coefficient
order (linear) and second-order (quadratic polynomial) polynomials, βk is coefficient m
to 0.0. The penalty function combines the responses with the constraint order (linear) and develops
second-order (quadratic polynomial) polynomials, βk is coefficient
nr bounds,
is the responsewhich number an (usually including constraints and the objective function).
n is
effective method for the constraint estimation. Fig.1 shows the fitnessr ncurve the response
of a normal numbernumber
constraint (usually including constraints and the objective function). Z
r is the response
nonparametric
(usually including constraints and the objective function).
part, whose covariance matrix can be expressed as:
function with the bounds of ( 1.0,1.0) . nonparametric part, whose covariance matrix can be expressed as:
nonparametric part,
s whose 2 covariance sj matrix can be expressed as:
Wan Zhiqiang A Highly Efficient Aeroelastic
i
), z (v j )]   Method
Cov[ z (v sOptimization G (v si , vBased ) on a Surrogate Model
sj sj
Cov[ zthis
A variety of algorithms have been studied for aeroelastic optimization; si
(v paper ), z (v uses)]s  GA G[4]
2
(vandsi
, v )s
Cov[ z (v si ), z (vs j )]   2G (v si , v j )
where G (v si , v j ) is the correlation equation between two design variable vectors s
IPSO [14, 15]. s
A variety of algorithms have been studied for aeroelastic where G (v si , svi j )s j isis the
where the correlation
correlation equation equationbetween
between design variable vectors s
twotwo
where G (v , v ) is the correlation equation between two design variable vectors i s
optimization; this paper uses GA [4] and IPSO [14, 15]. design variable vectors si and sj in space that determines Gauss correlation equati
space that determines the accuracy of the Kriging model. The
space
the that determines
accuracy the accuracy
of the Kriging model. of the Kriging
Gauss model. The Gauss correlation equation
space
appliedthat duedetermines
its preferable theperformance,
accuracy ofThe
the
and Kriging
correlation
it can bemodel. Theas:
expressed Gauss correlation equati
2.4 Surrogate Model
2.4 Surrogate Model equation is widely applied due its preferable performance,
applied due its preferable performance, and it can be expressed as:
andapplied
it can due expressed
its preferable performance, and it can be expressed as:
Surrogate models are constructed by samples (including design variables andsberesponses) dv as:
 nthat have s 
Surrogate models are constructed by samples (including G(v si , v j )  exp ndv i visi  vi j 
 
s 
 indv1i in 
si vi j sj 
s
design
alreadyvariables and responses)
been computed that have
using the finite already
element methodbeen (FEM), and v si , v j )s j
G(then, exp utilized visi the
G(v si ,theyv ) are
exp 
 i 1 i i vi 
 v  (8)
computed using the finite element method (FEM), and then,  i 1 
aeroelastic analysis instead of the FEM. Because the ofKriging where  is the correlation parameter. In order to improve the accuracy of Kriging m
they are utilized in the aeroelastic analysis instead the model and BPi model are effective
where
where
where θi iis is
thethe correlationparameter.
correlation parameter.InInorder ordertotoimprove
improvethe
the accuracy of Kriging mo
i is thecan correlation parameter. In order to improve the accuracy of Kriging m
FEM. Because methods,
approximate the Kriging
the model
complexand BP model
responses, areaseffective
such non-linear aeroelastic functions, be thereduced,
adjusted to minimized predictive error by an optimization method.
approximate methods, the complex responses, such as accuracy of Kriging model, θi is adjusted to minimized the
adjusted to minimized the predictive error by an optimization method.
and the efficiency
non-linear will functions,
aeroelastic be improved.can be reduced, and the adjusted to
predictive errorminimized the predictivemethod.
by an optimization error by an optimization method.
efficiency will be improved.
2.4.2 BP Model
2.4.2 BP Model
2.4.2 BP Model
2.4.1Kriging
2.4.1 KrigingModel
Model 2.4.2 BP Modelneural
AnAnartificial
artificial neural network
networkinvolves
involvesthethe imitation
imitationof human
of cerebral operation and
The An
human artificial
cerebral neural network
operation involves
and the
consists imitation
of many of human
artificial cerebral operation and c
The Kriging modelhashas
Kriging model preferable
preferable flexibility
flexibility and
and predictive An artificial
performance,
many artificial neural network
thus overcoming
neurons theAninvolves
[17]. artificialthe imitation
neuron of human
is a form cerebral operation
of mathematical and
model that
predictive performance, thus overcoming the limitations of neurons [17]. An artificial neuron is a form of mathematical
many artificial neurons [17]. An artificial neuron is a form of mathematical model that ca
thelimitations
parameterisedof the parameterised model for high-dimensional
model for high-dimensional data. Thedata. The many
Kriging
the
model artificial
model
operation
that neurons
isa human
canofsimulate [17].
a method
the ofAn artificial
neuron, which of
operation neuron
typically
a human is aneuron,
has form of mathematical
multiple model that
inputs. The weighted sum
Kriging model is a method of improved linear regression that the operation
which typically of ahashuman neuron,inputs.
multiple which The
typically has multiple
weighted sum ofinputs. The weighted sum o
improved linear regression that combines the linear regression part andvalues the operation
nonparametric of part
is calculated a human neuron,
(typically
by an whichfunction,
activation typically and
has multiple
then, theinputs.
outputThe weighted
value can besum
o
combines the linear regression part and nonparametric part the input values is calculated by an activation function,
values is calculated by an activation function, and then, the output value can be obta
via the random
(typically via the method)
random[16]:
method) [16]: andvalues
then,isthe
activation calculated
outputshould
function by anbe
value activation
can function,
be obtained.
monotonically Theand then,
activation
increasing, the output
continuous value can The
and derivable. be oB
activation should
function functionbe should be monotonically
monotonically increasing,
increasing, continuous and derivable. The BP
continuous
K (v )  f (v ) + z (v ) (5) activation
be expressed function
as [18]:should be(5)monotonically increasing, continuous and derivable. The B
and derivable. The BP model can be expressed as [18]:
be expressed as [18]:
f(v) is the linear regression part, and it can be expressed be expressed ndv as [18]:
as:
f ( v ) is the linear regression part, and it can be 5expressed as: y k  n(
dv wki vi ) (9)
y   (ni w
dv1
v)
nr
k
Whereyk  yi k(1yykiskiw
Where
Where is
i
the 
kithe ) predictivevalue
vipredictive valueofof thethek thk thresponse,
response,wkiwkiis is
k1 is the predictive value of the k th response, wki is the weight coefficient, a
thethe
weight
weight coefficient, and
coefficient, a
f (v )  f k (v ) β k (6) Where yk is the predictive(6)value of the k th response, wki
i 

the
is theactivation
weight function.
activation function. and φ(v) is the activation6function.
function.
k 1 the coefficient,
the activation
The error function that is generated in 6 the back
where fk(v)
where f k (vis) ais apolynomial
polynomial expression thatis istypically
expression that typicallydivided intoThezero-order
Theerrorerror function (constant),
function that
that is is first-
generated
generated in in
thethe
back back 6
propagation
propagation can bebe
can expressed
expressed as:as:
The
propagation error function
can be that
expressed is generated
as: in the back propagation can be expressed as:
divided into zero-order (constant), first-order (linear) and
ns nsnr nr
second-order
order (linear) (quadratic polynomial)
and second-order (quadraticpolynomials, βk is
polynomial) polynomials, βk 1is11coefficient
ns nr
(d((ksddkssmatrix, and
coefficient matrix, and nk is the response number (usually
  EE E 2n22s nns  1s k1 

1k 1
 y ' ksy)''kss2 ) 22
k  yk) (10)
s s 1  k 1
nr is theconstraints
including response number (usually
and the including
objective constraints
function). Z(v)andis the objective function). Z (v ) is the
the nonparametric part, whose covariance matrix can be whered kd
where
where
where s s
disk is is
s the thethe expectation
expectation
expectation value,
value,
value, and andy 'sky 's'skisisis
and the predictive
the
the values,
predictive
predictive values, andandns nsis isthethe
sample
samp
k is the expectation value, and y k is the predictive values, and ns is the samp
nonparametric
expressed as: part, whose covariance matrix can be expressed as: values, and n s is the sample number.
s s
Therefore,
Therefore,
Therefore,
Therefore, BP
BP BPmodel
BP modeltraining
model
model trainingisisis
training
training isequal
equaltototo
equal
equal tothe theoptimization
the
the optimizationofof
optimization
optimization ofE (E
Ew(()w
w. ))TheTheminimum
.. The minimum
minimum
Cov[ z (v si ), z (v j )]   2G (v si , v j ) (7) (7)
of E(w). The minimum E can be attained by adjusting the
weightw w
Figures attained byby adjusting thethe weight
s
attained
attained
weight w in by adjusting
adjusting
practical the
problems.
weight winin practical
in practical
practical problems.
problems.
problems.
where G (v si , v j ) is the correlation equation between two design variable vectors si and s j in
1.0
space that determines the accuracy of the Kriging model. The Gauss 2.5 Optimization
correlation equation Flowchart
is widely
2.52.5
Optimization
2.5 Optimization
Optimization Flowchart
Flowchart
Flowchart
0.8

applied due its preferable performance, and it can be expressed as: Surrogate models are combined with optimization
Constraint fitness

0.6
Surrogate
Surrogatemodels
Surrogate modelsareare
models arecombined
combinedwith
combined withoptimization
with optimizationalgorithms
optimization algorithmstoto
algorithms toconstruct
constructthethe
construct thehighly
high
hig
algorithms to construct the highly efficient aeroelastic
 ndv
s  aeroelastic
aeroelastic
optimization optimization
method;method;
optimization method; thethe
the flowchart flowchartof of
flowchart the thethe
of method
method
method is is
expressed
expressed in in
Fig. 2. 2.
Fig.
 aeroelastic optimization method;
(8) the flowchart of the method is expressed in Fig. 2.
s
G(v si , v j ) 0.4 exp   i visi  vi j 
 i 1  expressed in Fig. 2.
(1)(1)
Sample
(1) Sample selection and
selection andanalysis
analysis
(1)Sample
Sampleselection and analysis
0.2
selection and analysis
where i is 0.0 the correlation parameter. In order to improve the accuracy of Kriging
Samples are model,byby
selected the is Latin-hypercube
Samples
Samples are selected
Samples are selected by the
are selected i theLatin-hypercube
Latin-hypercubemethod
Latin-hypercube method to to
method
method toensure
ensurethethe
ensure
to thehomogeneous
homogeneousdistr
homogeneous di
di
-10 -5 0 5 10 ensure the homogeneous distribution in variable space. The
adjusted to minimized the predictive
Normal error by an optimization method.variable
constraint function
FEM
variable
variablespace.
space.
space.
is utilized
The
TheFEM
The FEM
FEM is is
to analyse
utilized
is utilized
utilizedto to
analyse
to
the structural
analyse
analysethethe
the
mass,
structural mass,
structural
structural
aeroelastic
mass,
mass, aeroelastic
aeroelastic
aeroelastic responses
responses
responses and
an
an
Fig.
Fig. 1. Fitness of 1. FitnessConstraint
a Normal of a Normal Constraint Function
Function responses
fitness ofof
fitness and
each
each constraint
sample.
sample. fitness of each sample.
fitness of each sample.

2.4.2 BP Model (2)(2)

Surrogate
(2) model
Surrogate
Surrogate construction
model
model construction
construction
493cerebral
The Kriging http://ijass.org
An artificial neural network involves the imitation of human The Krigingmodel
Kriging
operation
The oror
model
and orBPBP
consists
model model
BP modelis is
ofmodel isconstructed
constructedwith
constructed withdesign
with designvariables,
design variables,structural
variables, structuralmass,
structural massa
mass

responses
many artificial neurons [17]. An artificial neuron is a form of mathematical and
responses andconstraint
and
model
responses fitness.
constraint
that fitness.If If
fitness.
can simulate
constraint Ifthethe
themodel
modelaccuracy
model accuracyis is
accuracy isnotnot
notacceptable,
acceptable,then
acceptable, thenthethe
then thesampl
sam
sam

the operation of a human neuron, which typically has multiple inputs.updated

The to to
updated construct
weighted
updated to sum aofnew
construct
construct aathe
new
newmodel.
model.
input
model.
(3)(3)
values is calculated by an activation function, and then, the output Aeroelastic
value
(3) can beoptimization
Aeroelastic
Aeroelastic optimization
obtained. The
optimization
(491~500)16-067.indd 493 2017-01-02 오전 5:13:26
 Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016)

(2) Surrogate model construction
The Kriging model or BP model is constructed with
design variables, structural mass, aeroelastic responses and
constraint fitness. If the model accuracy is not acceptable,
then the samples will be updated to construct a new model.
(3) Aeroelastic optimization
GA: the optimization strategy and parameters are defined
first; the parameters include the population size, coding
mode, convergence criterion, and probabilities of crossover
and mutation. Next, the initial population is generated,
and the objective function and fitness are calculated by the
surrogate model. The next population will be generated from
the current population by individual fitness, fitness scaling,
niche technology, reproduction, crossover, and mutation if
the termination criteria are not met.
IPSO: the optimization strategy and parameters are
defined first; the parameters include the population size,
convergence criterion, acceleration coefficient, learning
factor, and mutation probability. Next, the initial population
is generated, and the objective function and fitness are
calculated by the surrogate model. If the termination criteria
are not met, then the next population will be generated from
the current population by the individual fitness, acceleration
coefficient, and learning factor, and the concentration is
mutated appropriately.
utilized for the construction of upper/lower skins, ribs
and front/rear spar webs. Bar elements are utilized for the
construction of upper/lower stringers, stiffened frames
and front/rear spar flanges. The engine is reduced to bar
elements, and the concentrated mass at a point element
is used to simulate the mass characteristics of engine and
payloads. The subsonic doublet-lattice method is utilized
for aerodynamic computations, and the wing aerodynamic
surface is divided into 8 groups with 266 elements in total.
The structural and aerodynamic models are shown in Fig. 3.
3.2 Optimization Constraints and Strategies
Ten constraints are considered in aeroelastic optimization,
including structural deformation, ply failure of composite
laminate, aileron efficiency and flutter. The specific
constraints are divided into three groups according to the
flight parameters:
(1) for the flight parameters at a 11,200 m height,
0.78 Mach and 2.5 g pull-up maneuver, the maximum
displacement at the wing tip is less than 12% of the half

3. Optimization based on the FEM

3.1 Baseline Model

The structural finite element model considered as

the optimization object is a typical composite wing-box
structure consisting of double spars, multi ribs and stiffened
skins, which is constructed by the commercial software MSC.
Patran. The model consists of 1286 nodes and 3371 elements
(207 point elements, 1768 bar elements and 1396 triangle/
Fig. 3. Structural Model (upper) and Aerodynamic Model (lower)
quadrangle elements). Triangle/quadrangle elements are Fig. 3. Structural Model (upper) and Aerodynamic Model (lower)

Fig. 2. Flowchart of the Aeroelastic Optimization Algorithm based on a Surrogate Model

Fig. 2. Flowchart of the Aeroelastic Optimization Algorithm based on a Surrogate Model

DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491 494

(491~500)16-067.indd 494 2017-01-02 오전 5:13:27

 Wan Zhiqiang A Highly Efficient Aeroelastic Optimization Method Based on a Surrogate Model

span, and the maximum torsion is less than 4.5°, and each
ply of the composite laminates of skins and webs satisfies the
“Tsai-Wu” composite failure theory with the response values
belonging to (-1.0, 1.0);
(2) for the flight parameters at a 11,200 m height, 0.78
Mach, 0° angle of attack and 10° deflection of the aileron, the
aileron efficiency is greater than 60%; and
(3) the flutter speed at sea level is greater than 320 m/s;
The characteristics of the high-aspect-ratio composite
wing and the composite failure theory are considered in
the aeroelastic optimization strategies, which specifically
require the following:
(1) the structural laminate is symmetrical, where the 0°
fibre direction refers to the direction of the rear spar flange.
The ply thickness of 0°, ±45° and 90° fibre directions accounts
for 50%, 40% and 10% of the composite laminate of the skins,
respectively, and the ply thickness of ±45° fibre directions
of front/rear spar webs is equal. The front/rear spar flanges
and stringers are constructed with equivalent isotropic rods
which are converted from the composite laminate.
(2) the upper/lower skins, front/rear spar webs and front/
rear spar flanges are divided into twelve regions, and the
laminate thickness and the sectional area of flanges increase
from the wing tip to the wing kink, whereas it decreases
from the wing kink to the wing root. Therefore, the structural
model can be described by 72 design variables.
(3) the genetic parameters are composed of:
population size 500
crossover probability 0.7
mutation probability 0.4
termination criterion maximum generation (15)
and the parameters of IPSO are composed of:
population size 500
learning factor 1.1
upper/lower limit of acceleration coefficient 0.9/0.3
concentration mutation probability 0.8
termination criterion maximum generation (70)
3.3 Optimization Results
The baseline model is optimized using the GA according
to the optimization strategies, and the individual that has
the best fitness is regarded as the optimum individual. Fig.
4 shows the deformation of the optimized model. Fig. 5 and
Fig. 6 show the g-V curves and F-V curves of the optimized
model, respectively, and Table 1 shows the natural frequency
and mode of the optimized model. The typical bend/twist
coupling flutter is taken into consideration (ignoring Mode
6 of the winglet). Table 2 compares the baseline model and
optimized model.
The results indicate that the weight of the composite wing
can be reduced with the satisfaction of all constraints. A more
appropriate stiffness distribution and reduced structural
mass can be attained through the aeroelastic structural
optimization.
4. Optimization based on Surrogate Models
Surrogate model construction and its accuracy validation
are based on the FEM analysis. The efficient aeroelastic
optimization methods are analysed using the Kriging model

Tables Table 1. Natural Mode Information of the Optimum Model

Table 1. Natural Mode Information of the Optimum Model
Order Mode Description F(Hz)

1 1st vertical bend 2.92

Fig. 4.
Fig. 4. Deformation ofDeformation of the
the Optimized Optimized Model
Model
2 Engine swing 3.25

3 Engine pitching 4.06

4 1st horizontal bend 6.96

5 2nd vertical bend 8.53

6 1st winglet bend 12.83

7 3rd vertical bend 18.25

8 1st twist 21.56

9 4th vertical bend 24.31

10 Aileron twist 28.79

Fig. 5. g-V curves Fig. 5. g-V curves

495 http://ijass.org

(491~500)16-067.indd 495 2017-01-02 오전 5:13:27

 uniform distribution, are utilized for the construction of surrogate models. The structural mass,

composite failure of the skin plies with 90°, 0°, +45° and -45° fibre directions, composite failure of

the web plies with +45° and -45° fibre directions, aileron efficiency and flutter, 11 responses in total,

are computed as the database according to the selected samples.

Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016)

The BP model cannot satisfy all of the samples in the model construction, and the predictive
and GA, the Kriging model and IPSO, and the BP model by these five hundred samples is shown in Fig. 7. The
accuracy validated by these five hundred samples is shown in Fig. 7. The Kriging model can satisfy
and IPSO. The efficiencies of the different methods are also Kriging model can satisfy all of the samples in the model
studied. all of the samples inconstruction;
the model construction; therefore,
therefore, the modelthe model is constructed
is constructed with four with four hundred
hundred samples and verified by the others. The predictive
samples and verified by the others. The predictive accuracy is shown in Fig. 8.
4.1 Model Construction and Accuracy Analysis accuracy is shown in Fig. 8.
The relative error isThe therelative
differenceerror is the the
between difference between of
actual responses theFEMactual
and the predictive
The fitted responses of the surrogate model mainly contain responses of FEM and the predictive responses of Kriging,
an objective function and constraints, and theresponses
appropriateof Kriging, andititisisdefined erel rFEM  rKrig / rFEM . The predictive
as: 
defined as:
and predictiveaccuracy of these
response types are fitted for different constraints. The
accuracy of these two surrogate models is practicable, except
objective function (structural mass) and constraints, such models is practicable, except for some particular individuals, and they are applicable for
two surrogate for some particular individuals, and they are applicable for
as the deformation of the wing tip and aileron efficiency,
aeroelastic optimization. The relative errors of the composite
aeroelastic
can be fitted directly, whereas the other constraints, suchoptimization. The relative errors of the composite failure constraint of the skin with +45°
failure constraint of the skin with +45° fibre direction ply and
as the composite laminate failure, and flutter, are fitted
fibre with ply and the flutter constraint are slightly larger, but they still satisfy the engineering
direction the flutter constraint are slightly larger, but they still satisfy
the constraint fitness. In order to improve the accuracy of
the engineering requirement. The mean predictive errors of
requirement.
surrogate models, only the objective function is fitted directly,The mean predictive errors of the objective function and the constraint fitness values for
the objective function and the constraint fitness values for
and the other responses are fitted with the constraint fitness.
these two surrogate models
these twoare surrogate
shown in Table
models3. are shown in Table 3.
Five hundred samples, whose variables are selected
The mean predictive errors of these two surrogate models
using the Latin-hypercube method to ensure the The mean predictive errors of these two surrogate models are approximately 1%. If the response
uniform
are approximately 1%. If the response satisfies the constraint
distribution, are utilized for the construction of surrogate
satisfies the constraintcompletely,
completely, thenthen
the the
predictive
predictivevalues of constraint
values of constraintfitness
fitness will be one.
models. The structural mass, composite failure of the skin
will be one. Therefore, the errors of some responses (such as
plies with 90°, 0°, +45° and -45° fibre directions, composite
Therefore, the errorsdisplacement
of some responses and (such as angle)
torsion displacement
may beand torsion
zero. The angle)
BP modelmay be zero. The BP
failure of the web plies with +45° and -45° fibre directions,
model has are has higher
higher predictive predictive
accuracy accuracy
compared with thecompared with the
Kriging model. TheKriging
computational burden is
aileron efficiency and flutter, 11 responses in total,
model. The computational burden is analysed, and the
computed as the database according to the selected samples.
analysed, and the comparison is shown in Table 4. 4.
comparison is shown in Table
The BP model cannot satisfy all of the samples in the
model construction, and the predictive accuracyThe BP model is more efficient in model construction and individual prediction. The Kriging model
validated
is a method with multiple inputs and a single output, that is, one Kriging model can only predict one

response value, whereas the BP model is a method with multiple inputs and multiple outputs, and thus,

one model can predict several response values. However, the BP model is a static function with simple

matrix operations, whereas the Kriging model is a dynamic interpolation method that requires the

construction of a new function by calculating the spatial position relationship between the predictive

individual and other samples to obtain the response value. Therefore, the BP model has a lower
11

Fig. 7. Relative Errors of the BP Model

Fig. 7. Relative Errors of the BP Model
Fig. 6. F-V curves Fig. 6. F-V curves

Table 3. Mean Relative Errors of Kriging and BP

Table 2. Comparison between the Baseline Model and Optimum Table 3. Mean Relative Errors of Kriging and BP
Model
Table 2. Comparison between the Baseline Model and Optimum Model Response Kriging BP

Response Constraint Baseline Optimum Weight 0.535% 0.466%

Weight —— M 0.904M Displacement 0.000% 0.000%

Displacement <12% 7.82% 8.55% Torsion angle 0.000% 0.000%

Torsion (degree) <4.5 2.33 2.39 Aileron Efficiency 0.068% 0.034%

Aileron efficiency >60% 64.5% 61.7% Failure Constraints 0.108% 0.097%

Flutter speed (m/s) >320 364 356 Flutter 1.546% 0.964%

DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491 496

(491~500)16-067.indd 496 2017-01-02 오전 5:13:27

 Wan Zhiqiang A Highly Efficient Aeroelastic Optimization Method Based on a Surrogate Model

The BP model is more efficient in model construction
and individual prediction. The Kriging model is a method
with multiple inputs and a single output, that is, one Kriging
model can only predict one response value, whereas the
BP model is a method with multiple inputs and multiple
outputs, and thus, one model can predict several response
values. However, the BP model is a static function with simple
matrix operations, whereas the Kriging model is a dynamic
interpolation method that requires the construction of a
new function by calculating the spatial position relationship
between the predictive individual and other samples to
obtain the response value. Therefore, the BP model has a
lower computational burden.
The Kriging model has an advantage over the BP model
in practical engineering applications. The BP model is
constructed by repeatedly adjusting the network structure
and activation function until the optimum model is attained.
These parameters, which depend largely on experience,
have an important impact on the predictive accuracy; thus,
the BP model construction is highly uncertain. In contrast,
the Kriging model can be utilized without adjustment, and
its accuracy is practicable. Therefore, the Kriging model is
applied widely and is more advantageous than the BP model
in the aeroelastic optimization of the high-aspect-ratio
composite wing.
The optimum results based on the FEM and GA are used
as the reference to verify the optimization of the surrogate
models. The Kriging model is mainly optimized with the GA;
the fitness convergence curves are compared in Fig. 9.
The Kriging method can converge rapidly, whereas the
optimum objective function of the FEM is superior to that of
the Kriging method. The consumption comparison is shown
in Fig. 10.
The FEM has the highest consumption in the initial
generation, and the responses of repetitive individuals can
be adopted directly without analysis; thus, the computational
burden declines in subsequent generations. The Kriging
method has the highest consumption in the initial generation
because it consists of the sample calculation and model
construction. The individual is analysed by the Kriging

4.2 
Aeroelastic Optimization based on the Surro- Fig.
Fig. 9.  9. Convergence
Convergence Curves
Curves of GA:
of GA: FEM FEM
(blue)(blue) and Kriging
and Kriging (red)(red)
gate Model

Table 4. Comparison of Computational Burdens

Table 4. Comparison of Computational Burdens
Type FEM Kriging BP
Response 1050 s 5.23 s 0.013 s
Computation / 500 individuals / 500 individuals / 500 individuals
Model 20 s 18 s
Construction (Existing samples) (Existing samples)

Fig.Fig.
10.10. 
Comparison
Comparisonof the
of Computational Burdens:
the Computational FEM (blue)
Burdens: FEMand Kriging
(blue) and (red)
Kriging (red)

Table 5. Comparison of the Optimum Results

Table 5. Comparison of the Optimum Results
Response Constraint FEM Kriging Validation

Weight ─ M 1.020 M 1.023 M

Displacement <12% 8.55% 8.74% 8.66%

Torsion (degree) <4.5 2.39 2.54 2.52

Aileron efficiency >60% 61.7% 61.6% 61.4%

Fig. 8. Relative Errors of the Kriging Model Flutter speed (m/s) >320 356 358 358
Fig. 8. Relative Errors of the Kriging Model

36
497 http://ijass.org

(491~500)16-067.indd 497 2017-01-02 오전 5:13:28

 Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016)

model in the following generations; thus, the computational based on a surrogate model, and the design efficiency can
burden can be substantially reduced. The model optimized also be increased considerably. Because the surrogate
by the Kriging model and GA is validated by the FEM to models are random due to the randomness of the sample
attain the actual responses. The optimum results of these two selection and model parameters, the responses may be
methods are shown in Table 5. different in each optimization. The method based on the
The responses of these two methods satisfy all of the BP model and IPSO has the best optimum result, and the
constraints, although the objective value of the Kriging global optimization efficiency of the GA is better than that
method results in a slightly heavier wing. The Kriging model of IPSO according to the convergence curve. The results for
and BP model are optimized with IPSO, and the fitness the optimization algorithm efficiency and applicability of
convergence curves are shown in Fig. 11. surrogate models illustrate that the aeroelastic optimization
The Kriging method can converge more rapidly, but the method based on the Kriging model and GA has the highest
optimum objective function of the BP model is better. The efficiency.
weight and computational burden of these methods are
shown in Table 6. 4.3 Analysis of the Optimization Results
The computational burden can be reduced considerably
To study the applicability of the surrogate model in
using the highly efficient aeroelastic optimization method
optimization, the effects of the number of design variables
and constraints on the optimization efficiency and the
objective function of the surrogate models are analysed in
detail, and the optimization method based on GA and FEM
is performed as the reference. Considering the efficiency,
the optimization method based on GA and Kriging model is
used.
The baseline model is optimized with different design
variables. The skin thickness described by 24 variables, the
web and skin thickness describe by 48 variables, and the
entire model described by the same 72 variables mentioned
above are optimized by these two methods with the same
Fig.
Fig. 11. 11. Convergence
Convergence Curves
Curves of IPSO:
of IPSO: Kriging
Kriging (blue)(blue) and(red)
and BP BP (red)
constraints, respectively, and the comparison is shown in
Table 7.
Table 6. Comparison of the Weight and Computational Burden
Table 6. Comparison of the Weight and Computational Burden The computational burden of FEM is not sensitive to the
Method Weight Amount Time increment of variables, while the consumption of the Kriging
method increases obviously because of the expansion of the
500 individuals
GA+FEM M 9.3 h sample space. The optimum objective function of the Kriging
15 generations
model yields a slightly heavier result, and it is not sensitive to
500 individuals
GA+Kriging 1.023 M 1.3 h changes in the number of variables.
15 generations
The two optimizations containing the same 72 design
500 individuals variables are also performed considering different
IPSO+Kriging 0.977 M 1.7 h
70 generations
constraints. Except for the basic case with the same 10
500 individuals constraints mentioned above, the case ignoring 2 constraints
IPSO+BP 0.936 M 1.5 h
70 generations
of the composite failure of the webs (8 in total) and the case

Table 7. Comparison of the Optimum Results with Different Numbers of Design Variables
Table 7. Comparison of the Optimum Results with Different Numbers of Design Variables
Constraint number 12 12 12

Design variable number 24 48 72

Method FEM Kriging FEM Kriging FEM Kriging

Weight 1 1.0291 1 1.031 1 1.0231

Computation time (h) 9.0 0.98 9.0 1.11 9.3 1.3

38

DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491 498

(491~500)16-067.indd 498 2017-01-02 오전 5:13:28

 Wan Zhiqiang A Highly Efficient Aeroelastic Optimization Method Based on a Surrogate Model

Table 8. Comparison of the Optimum Results with Different Numbers of Constraints

Table 8. Comparison of the Optimum Results with Different Numbers of Constraints
Design variable number 72 72 72

Constraint number 8 10 12

Method FEM Kriging FEM Kriging FEM Kriging

Weight 1 1.015 1 1.019 1 1.023

Computation time (h) 8.1 1.24 8.5 1.24 9.3 1.3

supplementing another 2 constraints of the axial stress of the considerably.

flanges (12 in total) are computed, and the comparison is  
shown in Table 8.
The computational burden of the Kriging method is not Acknowledgement
sensitive to the increment of the constraints, whereas the
consumption of the FEM increases notably because of the This work was supported by the National Key Research
increase in the computational sub-cases. The optimum and Development Program (2016YFB0200703) and the
objective function of the Kriging model yields a slightly National Natural Science Foundation of China (grant
heavier result, and the deviation increases with an increasing 11302011, 11672018).
number of constraints.

5. Conclusion
A method based on surrogate models is proposed
for aeroelastic response analysis; the method can solve
the excessively high consumption and low efficiency in
aeroelastic optimization.
1) Optimization frameworks using the Kriging model and
GA, the Kriging model and IPSO, and the BP model and IPSO
are presented. The applicability and feasibility are verified Fiber Composite Wings”, Journal of Aircraft, Vol. 27, No. 12,
using the case of a high-aspect-ratio composite wing.
2) The accuracy of the Kriging model and BP model is
practicable in aeroelastic response prediction. The accuracy
and consumption of the BP model are better than those
of the Kriging model, although the Kriging model is more
widely used because of its convenient model construction.
The GA has a higher efficiency in global optimization than
IPSO; therefore, the method based on the Kriging model and
GA is more efficient and applicable.
3) The effects of the number of design variables and the
number of constraints on the optimum objective function 27 AHS/ASC Structures, Structural Dynamics, and Materials
and computational burden of the Kriging model are studied.
There is a positive correlation between the consumption of
the model construction and the design variables, whereas
the objective function value is not sensitive to increases in
the number of design variables. The consumption of the
model construction is not sensitive to the increment of the
constraints, whereas the objective function value increases
References
[1] Librescu, L. and Khdeir, A., “Aeroelastic Divergence
of Swept-Forward Composite Wings Including Warping
Restraint Effect”, AIAA Journal, Vol. 26, No. 11, 1988, pp.
1373-1377.
DOI: 10.2514/3.10050
[2] Livne, E., Schmit, L. A. and Friedmann, P., “Towards
Integrated Multidisciplinary Synthesis of Actively Controlled
1990, pp. 979-992.
DOI: 10.2514/3.45972
[3] Guo, S., Cheng, W. and Cui, D., “Aeroelastic Tailoring
of Composite Wing Structures by Laminate Layup
Optimization”, AIAA Journal, Vol. 44, No. 12, 2006, pp. 3146-
3150.
DOI: 10.2514/1.20166
[4] Wan, Z., Yang, C. and Zou, C., “Design Studies of
Aeroelastic Tailoring of Forward-Swept Composite Aircraft
Using Hybrid Genetic Algorithm”, 44th AIAA/ASME/ASCE/
Conference, Norfolk, Virginia, 7-10 April 2003, AIAA 2003-
1491.
DOI: 10.2514/6.2003-1491
[5] Isogai, K., “Direct Search Method to Aeroelastic
Tailoring of a Composite Wing Under Multiple Constraints”,
Journal of Aircraft, Vol. 26, No. 12, 1985, pp. 1076-1080.
DOI: 10.2514/3.45883

499 http://ijass.org

(491~500)16-067.indd 499 2017-01-02 오전 5:13:28

 Int’l J. of Aeronautical & Space Sci. 17(4), 491–500 (2016)
[6] Shinkyu, J., Mitsuhiro, M. and Kazuomi, Y., “Efficient
Optimization Design Method Using Kriging Model”, Journal
of Aircraft, Vol. 42, No. 2, 2009, pp. 413-420.
DOI: 10.2514/1.6386
[7] Kanazaki, M., Tanaka, K., Jeong, S. and Kazuomi, Y.,
“Multi-Objective Aerodynamic Optimization of Elements'
Setting for High-Lift Airfoil Using Kriging Model”, 44th AIAA
Aerospace Sciences Meeting and Exhibit, Reno, Nevada, 9-12
January 2006, AIAA 2006-1471.
DOI: 10.2514/6.2006-1471
[8] Rai, M. M. and Madavan, N. K., “Aerodynamic Design
Using Neural Networks”, AIAA Journal, Vol. 38, No. 1, 2000,
pp. 173-182.
DOI: 10.2514/2.938
[9] Haas, D. J., Milano, J. and Flitter, L., “Prediction of
Helicopter Component Loads Using Neural Networks”, AIAA
34th Structures, Structural Dynamics, and Materials Conference,
La Jolla, CA, U.S.A., 19-22 April 1993, AIAA 1993-1301.
DOI: 10.2514/6.1993-1301
[10] Wan, Z., Xiao, Z. and Yang, C., “Robust Design
Optimization of Flexible Backswept Wings with Structural
Uncertainties”, Journal of Aircraft, Vol. 48, No. 5, 2011, pp.
1806-1809.
DOI: 10.2514/1.C000221
[11] Wan, Z., Zhang, B., Du, Z. and Yang, C., “Aeroelastic
Two-Level Optimization for Preliminary Design of Wing
Structures Considering Robust Constraints”, Chinese Journal
of Aeronautics, Vol. 27, No. 2, 2014, pp. 259-265.
DOI: 10.1016/j.cja.2014.02.018
[12] Chen, P. C., Sarhaddi, D., Jha, R., Liu, D. D., Griffin,
DOI: http://dx.doi.org/10.5139/IJASS.2016.17.4.491
(491~500)16-067.indd 500
K. and Yurkovich, R., “Variable Stiffness Spar Approach for
Aircraft Maneuver Enhancement Using Astros”, Journal of
Aircraft, Vol. 37, No. 5, 2000, pp. 865-871.
DOI: 10.2514/2.2682
[13] Zhou, M. and Sun, S. D., Genetic Algorithms: Theory
and Applications, National Defence Industry Press, Beijing,
1999 (in Chinese).
[14] Dhadwal, M. K., Lim, K. B., Jung, S. N. and Kim, T. J.,
“Particle Swarm Assisted Genetic Algorithm for the Optimal
Design of Flexbeam Sections”, International Journal of
Aeronautical and Space Sciences, Vol. 14, No. 4, 2013, pp.
341-349.
DOI: 10.5139/IJASS.2013.14.4.341
[15] Jiang, Y., Hu, T. S., Huang, C. C. and Wu, X. N., “An
Improved Particle Swarm Optimization Algorithm”, Applied
Mathematics and Computation, Vol. 193, 2007, pp. 231-239.
DOI: 10.1016/j.amc.2007.03.047
[16] Lophaven, S. N., Nielsen, H. B. and Søndergaard, J.,
“DACE: A Matlab Kriging Toolbox Version 2.0”, Informatics
Math and Modelling TR 2002-12, Technical Univ. of
Denmark, 2002.
[17] Peng, W., Zhang, J. and You, L., “The Hybrid Uncertain
Neural Network Method for Mechanical Reliability Analysis”,
International Journal of Aeronautical and Space Sciences,
Vol. 16, No. 4, 2015, pp. 510-519.
DOI: 10.5139/IJASS.2015.16.4.510
[18] Pedrycz, W., “Fuzzy Neural Networks with Reference
Neurons as Pattern Classifiers”, IEEE Transactions on Neural
Networks, Vol. 3, No. 5, 1992, pp. 770-775.
DOI: 10.1109/72.159065
500
2017-01-02 오전 5:13:28