Advances in Engineering Software 76 (2014) 56–68

Contents lists available at ScienceDirect

Advances in Engineering Software

journal homepage: www.elsevier.com/locate/advengsoft

A CAD/CAE integrated framework for structural design optimization

using sequential approximation optimization
Donghui Wang ⇑, Fan Hu, Zhenyu Ma, Zeping Wu, Weihua Zhang
College of Aerospace Science and Engineering, National University of Defense Technology, Changsha, Hunan 410073, PR China

a r t i c l e i n f o a b s t r a c t

Article history: This paper presents an open and integrated framework that performs the structural design optimization
Received 8 October 2013 by associating the improved sequential approximation optimization (SAO) algorithm with the CAD/CAE
Received in revised form 28 April 2014 integration technique. In the improved SAO algorithm, a new estimate of the width of Gaussian kernel
Accepted 28 May 2014
functions is proposed to enhance the surrogate models for SAO. Based on the improved surrogate models,
Available online 1 July 2014
an adaptive sampling strategy is developed to balance the exploration/exploitation in the sampling pro-
cess, which better balances between the competence to locate the global optimum and the computation
Keywords:
efficiency in the optimization process. Fewer function evaluations are required to seek the optimum,
Sequential approximation optimization
CAD/CAE integration
which is of great significance for computation-intensive structural optimization problems. Moreover,
Integrated framework based on scripting program languages and Application Programming Interfaces (APIs), integration
Surrogate model between commercial CAD and CAE software packages is implemented to expand the applications of
Adaptive sampling strategy the SAO algorithm in mechanical practices. Two benchmark tests from simple to complex, from low-
Structural design dimension to moderate-dimension were performed to validate the efficacy of the proposed framework.
Results show that the proposed approach facilitates the structural optimization process and reduces
the computing cost immensely compared to other approaches.
Ó 2014 Elsevier Ltd. All rights reserved.

1. Introduction In the CAD/CAE integrated framework for structural design,

optimization techniques are important in that they affect the accu-
Structural design optimization has been an important and racy, reliability, usability, and computing cost [6]. Over the past
challenging topic in the field of engineering design for obtaining decade a number of optimization algorithms from gradient-based
more efficient and lighter structures [1]. The aim of the design algorithms to evolutionary algorithms (EAs) as well as approxi-
optimization is to determine the optimal shape of a structure to mate-based optimization algorithms [7] have been widely used
maximize or minimize a given criterion, such as maximizing the in structural optimization problems. Every foregoing optimization
stiffness or minimizing the weight subjected to the constraint con- approach has its own advantages and disadvantages.
ditions of the stress or displacement. Modern commercial CAD Examples of gradient-based optimization applied to structural
(Computer Aided Design) and CAE (computer Aided Engineering) design problems can be found in the literatures [8,9]. Gradient-
tools have become a quintessential part of structural design pro- based optimization algorithms need the gradient information
cesses, which always require a loop of model-evaluate-remodel. which is usually difficult to compute [10] or even unavailable.
However, many of them are usually stand-alone systems that are Moreover, they are prone to being trapped in local optima. A num-
not designed for collaborative use [2,3]. Mostly, these tools are ber of EAs such as genetic algorithm [11], simulated annealing [12],
confined to a single simulation because a designer usually has to particle swarm optimization algorithm [13], immune algorithm
work with two or more independent software packages for [14], and artificial bee colony algorithm [15] have been widely
modeling and analysis, and the gap between the CAD and CAE used to solve structural design optimization problems. However,
domains is hard to get through. Hence an efficient CAD/CAE these methods require a huge number of function evaluations
integrating approach is requisite to facilitate the structural design and the computational cost is extremely high [16,17], which makes
optimization [4,5]. it difficult to apply these optimization techniques to the structural
design directly. In recent years, approximation techniques such as
response surface method (RSM) [18,19], neural network (NN)
⇑ Corresponding author. Tel.: +8613808464745. [20,21], polynomial regression models [22], Kriging methods [23]
E-mail address: lightblue117@gmail.com (D. Wang). and radial basis function (RBF) [24] method are explored and

http://dx.doi.org/10.1016/j.advengsoft.2014.05.007
0965-9978/Ó 2014 Elsevier Ltd. All rights reserved.
 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 57

exploited intensively. These techniques in association with the EAs from the surrogate. Results from the surrogate must be evaluated
are able to improve the computational efficiency [25] in structural with the true function evaluations. Additional calls to the true
optimization problems, while the disadvantage is the error function are not only used to validate the surrogate, but also to
between the meta-model and the true model, which will reduce enhance its accuracy. Thus, a sequential approximation optimiza-
the reliability of approximation-based optimization methods. tion (SAO) approach is developed. A general procedure of the
Lower computational cost, generality, robustness, and accuracy SAO algorithm is presented in Fig. 1. Applying a series of new infill
are all required for structural optimization processes [26]. Thus, a points, based on some infill criteria which is also known as sam-
sequential approximate optimization (SAO) approach [27,28] has pling strategy, we sample the objective function using a constantly
gained popularity recently. Firstly, a small-size design of experi- changing surrogate model [29]. In SAO, the approximation tech-
ment (DOE) is employed to construct the surrogate model through nique and the sampling strategy are the key elements for a suc-
various approximate techniques, then the global optimum of the cessful optimization. In this paper, an improved RBF network and
response surface can be found by global optimization techniques an adaptive sampling strategy are employed to implement the
such as EAs. If the termination criterion specified by the designer SAO algorithm. SAO-based structural optimization can be mathe-
is satisfied, the SAO algorithm will be terminated. Otherwise, a matically expressed as
few new sampling points will be added to improve the accuracy
for n ¼ 1; 2; . . .
of the response surface. Through the iterative process, a highly
accurate global optimum can be found with a small number of find x
function evaluations in comparison with pure EAs. The high min f ðnÞ ðxÞ
computational efficiency makes the SAO algorithm a promising s:t:
ðnÞ
g i ðxÞ 6 0 i ¼ 1; 2; . . . ; l ð2Þ
optimization technique for the computation-intensive structural ðnÞ
design problems. hj ðxÞ ¼ 0 j ¼ 1; 2; . . . ; k
However, the study of the SAO algorithm is still relatively SðnÞ ðxÞ 6 0
young. A reliable and practical SAO algorithm is still lacking in
xL 6 x 6 xU
the engineering design area. In SAO, the approximation technique
and the sampling strategy are the two key elements for a success- the m-dimensional design variable X is scaled into an
ful optimization. In Section 2 the general formulation of the SAO m-dimensional unit hypercube x by
approach is first presented, then a new method to determine the
width of basis function for the RBF network and an adaptive sam- X i  X Li
xi ¼ i ¼ 1; 2; . . . ; m ð3Þ
pling strategy are proposed to enhance the SAO algorithm. Further- X Ui  X Li
more, in order to exploit the SAO algorithm thoroughly in
mechanical engineering, a framework that permits the integration where X Ui and X Li denote the upper and lower bounds of the ith
between commercial CAD and CAE tools in conjunction with the design variable, respectively.
SAO algorithm is developed. This approach widens the applications
of the SAO method in structural design optimization problems. 2.2. Surrogate models for SAO
Moreover, this approach also facilitates the complex structural
optimization process and makes it more convenient for engineer- 2.2.1. The fundamental question of surrogate models
ing designers. Section 3 focuses on the general structural design Radial basis function (RBF) network [30], support vector regres-
optimization framework based on the SAO approach in association sion (SVR) [31], and Kriging function [32], etc. are widely employed
with CAD/CAE integration. Section 4 presents different case stud- surrogate models in various areas. The equivalence between SVM
ies, which are used to demonstrate the efficacy of the integrated and ordinary Kriging has been reported under the assumption that
framework. Finally, the paper closes with some concluding
remarks.

2. Sequential approximation optimization technique

2.1. General framework of the SAO approach

A general structural optimization problem with constraints is

considered as follows:

find X
min f ðXÞ
s:t: g i ðXÞ 6 0 i ¼ 1; 2; . . . ; l ð1Þ
hj ðXÞ ¼ 0 j ¼ 1; 2; . . . ; k
XL 6 X 6 XU
Here XL and XU is the upper and lower bound of design vari-
ables, respectively. For structural design problems, the objective
function and constraints are usually implicit functions of design
variables, usually obtained by finite element analysis (FEA).
In the classical approximation-based optimization process sum-
marized in [28], accuracy of the surrogate model could possibly be
deteriorated by a particularly poor or unlucky initial sample, lead-
ing to a very deceptive optimum. Indeed, at any stage we take our
optimum design to be the best result of the true function, not that Fig. 1. General framework of the sequential approximation optimization.
58 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

the covariance function is used as the kernel function [33]. The Hence, a more effective estimate of the width of the kernel func-
equivalence between SVM and RBF network can be established tions is required.
based on the suggestions of Poggio and Girosi [34]. Thus, the equiv-
alence between Kriging and RBF network is naturally inferred. 2.2.2. Estimate of the width in Gaussian kernel
Consider a function f observed without error, according to the Based on the local density of sampling points, a new determina-
sampling plan X = {x1, x2, . . . , xn}T, yielding the responses tion method to estimate the width of the kernel function is
Y = {y1, y2, . . . , yn}T. The basic forms of RBF, Kriging and SVR predic- proposed to enhance the surrogate models for the SAO algorithm.
tions take the same form of the sum of kernel functions /i, with The density of sampling points is defined by the following
weightings xi, added to a base term b, which is outlined as follows equation:
[29]:
X
N

kxxj k

X
n qðxÞ ¼ e r2 ð7Þ
^f ðxÞ ¼ xi /ðkx  xi kÞ þ b ¼ U  x þ b ð4Þ j¼0
i¼1
where r is given by [36].
where x is the vector of design variables, / is a basis function and xi
d
is the coefficient for the ith basis function, U = [/1, /2, . . . , /n] r ¼ pffiffiffiffimax
ffip
m
ffiffiffi ð8Þ
(/j = /kx  xjk), x = [x1, x2, . . . , xn]T. m n
The following Gaussian kernel is commonly employed as the dmax represents the maximum distance between sampling points.
basis function (basis function in RBF, covariance function in Kriging The illustrative example of the density function is shown in Fig. 3.
and kernel function in SVR, respectively) by the three predictions vi is the influence volume of the ith sampling point. Based on
above: the meaning of width of the kernel function ci, vi can be given by:
!
ðx  xj ÞT ðx  xj Þ V i ¼ cm
i ð9Þ
/j ðkx  xj kÞ ¼ exp  ð5Þ
c2j Summation of all sampling points’ influence volume should
In Eq. (5), cj is the width of the jth kernel function /j. The phys- cover the whole m-dimensional unit hypercube, which leads to:
ical meaning of c can be defined as the influence extent of the basis X
N

function /. As shown in Fig. 2, a higher c means that the corre- Vi ¼ Vt; Vt P 1 ð10Þ
sponding basis function’s influence may extend further. Because i¼1

every basis function corresponds to a sampling point, c can also where vt denotes the coupling effect of influence between sampling
be defined as the influence extent of the sampling point. It is points. The determination of vt is the crux of an approximation with
believed that determination of the width of the basis function high precision, which will be elaborated in the next section.
plays an important role in a good approximation [35]. It is assumed that the influence volume of the sampling point
In order to determine the width in a simple manner, Nakayama located in the sparse area should be sufficiently great while in
proposed the following estimate [36]: the intensive area relatively small, which is given by the following
dmax equation:
c1 ¼ c2 ¼    ¼ cn ¼ pffiffiffiffiffi p ffiffiffi ð6Þ
mmn V i 1=qðxi Þ qðxj Þ
¼ ¼ ð11Þ
V j 1=qðxj Þ qðxi Þ
where dmax represents the maximum distance between sampling
points. Eq. (6) implies that all basis functions are assigned the same When vt is determined, the estimate of cj (i = 1, 2, . . . , n) is
width. The above method suffer from two main disadvantages: (i) it implemented by solving the simultaneous Eqs. (7)–(11).
cannot exploit the information of the sampling points thoroughly
and usually requires redundant points to approximate the true 2.2.3. Determination of Vt
model accurately and (ii) this estimate performs not so well with In the estimate of the width of Gaussian kernel above, Vt is the
respect to the non-uniform sampling points. The above disadvan- key parameter. A lower Vt predicates that the influence of sampling
tages are fatal for the surrogate models in the SAO algorithm. points is not exploited thoroughly, while a higher Vt may aggran-
dize the influence coupling between sampling points and lead to
1
c=0.05
c=0.1
c=0.2
0.8 c=0.5
c=1

0.6
(r)

0.4

0.2

0
0 0.2 0.4 0.6 0.8 1
r

Fig. 2. Influence extent of / with respect to c. Fig. 3. Illustrative example of density function in one dimension.
 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 59

a severe deviation. In this section, a method based on cross valida- method, the indexes of MR2 do not exceed 0.80, which indicates
tion is proposed to determine the value of Vt. that the proposed method has an excellent performance and it is
The procedure of cross validation is depicted as follows: (1) The applicable to various surrogate models which satisfy the form of
surrogate model sampling data are split randomly into q roughly Eq. (4) and use Gaussian kernel as basis functions. The fluctuations
equal subsets. (2) Each of these subsets is removed in turn from of R2 for RBF processes in both test cases are demonstrated in Fig. 6.
the complete sampling data and the surrogate model is fitted Even in times that Nakayama’s method deviates severely, the pro-
to the remaining data. (3) The removed subset is predicted using posed method still presents a good fit, which indicates that it has
the model which has been fitted to the remaining data. an excellent and stable performance.
The whole procedure repeats until all subsets have been What distinguish the proposed estimate method from other
removed, when n predictions ðy ^1 ; y
^2 ; . . . ; y
^n Þ of the n observed data methods is that the proposed one makes a difference among influ-
points (y1, y2, . . . , yn) will have been calculated. The cross-validation ences of the sampling points. Thus the width ci of every kernel
error is then calculated as function /i for a fit process is variable. The proposed method solves
the variable ci through a key parameter Vt, which is handled with
1X n
the cross validation method.
eCV ¼ ^ i Þ2
ðy  y ð12Þ
n i¼1 i
2.3. Sampling strategy for SAO
With q = n, an almost unbiased error estimate can be obtained, but
the variance can be very high. Thus somewhat larger subsets with
2.3.1. The adaptive sampling strategy
q = 5 or 10 is suggested by Hastie et al. [37].
Since the advent of the SAO algorithm, a number of sampling
It is obvious that a perfect Vt should be able to eliminate the
strategies have been studied and applied [37–39]. These strategies
cross-validation error eCV. In the mathematical realization, deter-
can be divided into three categories: the exploitation technique,
mination of the value of Vt is achieved by solving the following
the exploration technique and the balanced exploitation/explora-
optimization problem:
tion technique. The exploitation-based sampling strategy is an
min : eCV ðV t Þ V t P 1 ð13Þ attractive method for SAO. However, it is clear that this kind of
sampling strategy is prone to being trapped in the local optimum
[29]. Hence, an exploration sampling technique is proposed to
2.2.4. Evaluation of the proposed estimate method build an accurate approximation of the entire design space to help
The proposed method and the foregoing method by Nakayama the designer locate the global optimum [29]. However, explora-
are applied to the three surrogate models (RBF, Kriging and SVR), tion-based sampling requires a great deal of evaluations in the
respectively, which are used to fit the test cases shown in Fig. 4. non-promising areas for global optimum which reduces the con-
The approximate accuracy of the surrogate models is assessed by vergence efficiency. Exploiting the surrogate after the design space
the index of R-squared (R2), which is depicted as follows: has been explored sufficiently may lead to the global optimum
Pn lying undiscovered, while over exploration is a waste of resources.
^ i Þ2
ðyi  y Therefore Jones [38] proposed an error-based balanced explora-
R2 ¼ 1  Pi¼1
n
ð14Þ
2
i¼1 ðyi  yÞ tion/exploitation sampling strategy to find the correct balance
between them. However, Jones’ method requires the error
The index of R2 is between 0 and 1, where R2 = 1 means no error information on sampling points and thus is inapplicable to the
between observed and approximated values. To evaluate the RBF network and other error-free techniques, etc. In this section,
efficacy of the proposed estimate method thoroughly, a special val- we propose an adaptive sampling strategy to be acclimatized to
idation procedure is used. In the procedure, 15 points for test case 1 various approximation techniques.
and 50 points for test case 2 are sampled randomly each time, and In the proposed sampling strategy, the simplest way of
then the RBF, Kriging and SVR processes with both estimate meth- exploitation sampling is to find the optimum of the surrogate
ods of the width of kernel are performed, respectively. 1000 points model s(x) and the pure exploration is to maximize the minimum
are selected to calculate R2 between the approximation curve and Euler distance d(x) between sampling points, which is given by
the true curve. The whole process repeats for 50 times and the rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi!
mean R2 (MR2) is calculated.  T  
ðnÞ ðnÞ
The fit accuracy comparison for both test cases by the index of dðxÞ ¼ min x  xi x  xi ði ¼ 1; 2; . . . ; NðnÞ Þ ð15Þ
MR2 is shown in Fig. 5. With the proposed method, the indexes of
MR2 for the three surrogate models including RBF, Kriging and SVR where N(n) is the number of sampling points before the nth
are all above 0.96 for both test cases while with Nakayama’s sequential sampling. The adaptive sampling method of solving the

Fig. 4. Test cases.

60 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

Fig. 5. Comparison of fit accuracy by the index of MR2.

Fig. 6. The fluctuations of R2 for RBF processes.

qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
multi-objective optimization problem (16) is therefore proposed to
d ¼ min ðxi  xj ÞT ðxi  xj Þ ði; j ¼ 1; 2; . . . ; M; i–jÞ ð18Þ
balance the exploitation and exploration.
where M is the number of sampling points before the nth sequential
max : sðnÞ ðxÞ; dðxÞ xmin 6 x 6 xmax
ðnÞ
sampling. It is concluded from Eq. (18) that when sampling points
s:t: g i ðxÞ 6 0 i ¼ 1; 2; . . . ; l ð16Þ are sparse, the value of d ought to be great. The knowledge about
ðnÞ
hj ðxÞ ¼0 j ¼ 1; 2; . . . ; k the response space is relatively inadequate. Hence, the sampling
strategy will emphasize particularly on exploration. As more points
where l is the number of inequality constraints and k is the number are sampled, the value of d will inevitably decrease. Meanwhile, the
of equality constraints of the original optimization problems, s(n)(x) knowledge of the global response space is accumulating and the
is the meta-model constructed before the nth sequential sampling. promising areas for global optimum will finally be located. There-
Then the optimal solution of (17) together with the real response fore, d will take a relatively small value to bring exploitation over
evaluated by the original model will be regarded as a new sampling exploration. New sampling points are added around the promising
point to update the surrogate model. areas in order to approximate them precisely for global optimum
To simplify the solving process and obtain a preferable solution, while the details of non-promising areas will be ignored. In this
the objective d(x) is turned into a constraint and the multi- way, the number of true function calls will be remarkably reduced,
objective problem is thus converted to a single-objective optimiza- which is of great significance for the computation-intensive struc-
tion problem as follows: tural design optimization tasks.

max : sðnÞ ðxÞ xmin 6 x 6 xmax 2.3.2. Evaluation of the proposed sampling strategy
ðnÞ
s:t: g i ðxÞ 6 0 i ¼ 1; 2; . . . ; l The validity of the adaptive sampling strategy is examined
ðnÞ
ð17Þ through the following problem.
hj ðxÞ ¼ 0 j ¼ 1; 2; . . . ; k
   
dðxÞ P d f ðxÞ ¼ 4  2:1x21 þ x41 =3 x21 þ x1 x2 þ 4x22  4 x22 þ x1 þ x2

In this paper, the determination of d is given by: ! min 2 6 xi 6 2 ð19Þ

 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 61

Fig. 8. The distribution of sampling points and the final surrogate model.
Fig. 7. The true objective function.

Fig. 7 illustrates the contour line of the true objective function,

Table 1
and the global minimum xG is f(0.02986, 0.7613) = 1.7395. In the Part of the sequential sampling points.
implementation of the proposed sampling strategy, the method
of Optimal Latin Hypercube Design is employed to select the initial Sampling x1 x2 f(x1, x2)
sequence
five sampling points which are marked with circles in Fig. 6 and a
genetic algorithm (GA) [11] is used to solve the optimization prob- Initial sampling points 1 0.80 1.20 2.962
2 1.60 0.60 4.308
lem(17). Besides, the surrogate model is constructed based on the 3 1.60 0.60 0.09181
improved RBF network in Section 2.2. As shown in Fig. 8, 24 4 0.00 0.00 0.000
sampling points are generated by the adaptive sampling method Sampling points in 1 0.637 1.50 11.48
and distinguished in triangle. Sequence of the sampling points is optimization 3 2.000 0.594 3.036
shown in Table 1. It is clear that the sampling points are exploring iterations 5 1.06 0.0933 1.216
the response space in the beginning, when the promising area is 7 2.00 1.50 11.48
9 0.133 0.801 1.678
located, exploiting strategy will take over control and new sam-
11 0.801 0.565 0.005769
pling points will gather around the promising area. The contour 13 1.32 0.702 0.1759
line of the final surrogate model is shown in Fig. 8 and it is clear 15 0.0794 0.414 0.9093
that the promising area around xG is captured precisely while other 17 0.151 0.662 1.505
areas are relatively coarse. 19 0.0892 0.824 1.648
21 0.00358 0.716 1.717
22 0.0289 0.757 1.739
23 0.0198 0.801 1.723
3. Structural optimization framework based on SAO and CAD/ 24 0.0318 0.761 1.740
CAE integration

3.1. CAD/CAE integration for structural design optimization (High Level Topology) based on the mixed shape representation
to simultaneously support the B-Rep and polyhedral model, which
SAO is proved to be a powerful optimization technique in creates a robust link between CAD and CAE models, and finally a
solving structural optimization problems [40]. In mechanical new software environment for CAD/CAE integration based on the
practices, the structural analysis is usually based on the FEA in HLT is presented. Hughes et al. [44] proposed the Isogeometric
the high fidelity level (HFL), which is performed by professional Analysis using complex NURBS (Non-Uniform Rational Basis
CAE systems associated with specific CAD systems product design. Spline) geometry in the FEA applications directly, which allows
However, CAD and CAE systems usually use different data formats models to be designed, calculated and adjusted in one step, using
to represent the design geometry [41,42]. Though the design and a common data set. By using splines in FEA instead of polynomial
analysis models in essence describe the same object, the descrip- approximations, this method is the first computational approach
tions are so different that preparing a qualified CAE model from offering the possibility of integrating FEA in conventional NURBS-
the CAD geometry is usually a laborious task. Therefore, the based CAD design tools. Albers et al. [45] proposed a strategy for
integration between CAD and CAE is requisite to enhance the appli- the development of the engine crankshaft with the integration
cation of the proposed SAO to engineering practices. for CAD, CAE and the genetic algorithm. Matin et al. [46] presents
Several in-house integrated frameworks aimed to close the gap a knowledge-based, parametric, modular and feature-based inte-
between CAD and CAE and reduce the product development cycle grated CAD/CAE system for the mold design. Xu and Chen [47]
have been established. Park and Dang [6] presented a framework developed a completely automated product design system with
that performs the integration between the commercial CAD/CAE the CAD/CAE integration and multi-objective optimization. Chang
software, in which meta-modeling techniques including response and Joo [48] presents an integrated system that supports optimiza-
surface methodology and radial basis function were applied to tion of general mechanical systems, including ground vehicles,
the structural optimization. Hamri et al. [43] introduced the HLT which supports engineers to effectively search, in a CAD solid
62 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

model, for a mechanism design with optimal kinematic and on these APIs and PDLs, the collaboration between CAD and CAE
dynamic performance. systems is greatly facilitated.
The most common and effective communication between CAD
and CAE systems relies on external data exchange. A typical CAD/ 3.2. Structural optimization framework based on SAO associated with
CAE integration process is that a CAD software package provides CAD/CAE integration
an exported geometry model in the standard formats such as IGES,
STL or STEP, which can be parsed and divided into grids by the CAE As shown in Fig. 9, the structural design optimization
analysis packages. But before the geometry can be fed to the CAE framework is developed based on the SAO algorithm associating
analysis, the original model has to be modified and enriched with with the integration between CAD and CAE. The framework con-
some additional information to make it suitable to analysis. Most sists of two main parts: the proposed SAO algorithm and the
commercial CAD systems have provided special mechanisms that CAD/CAE integration mechanism. The collaboration between these
enable the users to implement repeated modeling procedures two parts contributes to an effective structural design optimization
automatically. These automatic mechanisms can be executed framework.
based on the Application Programming Interfaces (APIs), such as
the Pro/Toolkit of Pro/Engineer, the OpenAPI of Unigraphics and 3.2.1. Implementation of the SAO algorithm
the CAA of CATIA. In addition, most of the high-end CAE packages As shown in Fig. 9, the implementation of the proposed SAO is
are equipped with a parametric design language (PDL), e.g. the roughly divided into four stages including Initial stage, Approxima-
Adaptive Programming Design Language (APDL) of Ansys, the tion stage, Termination criteria and Sampling stage, which are
Patran Command Language (PCL) of Nastran et al. These PDLs are elaborated as follows:
usually structured scripting languages by which the designers
can import the model, assign materials, loads, and boundary condi- Initial stage: When the optimization problem is determined, the
tions. Moreover, the required analysis results can be obtained from m-dimensional design variable is first scaled into an m-dimen-
the output database automatically by programming as well. Based sional unit hypercube, then Optimal Latin Hypercube Design

Fig. 9. Structural optimization framework based on SAO associated with CAD/CAE integration.
 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 63

(OLHD) method is used to sample in the unit hypercube. The model automatically based on the setting parameters (Material,
number n of sampling points is generally estimated based on the loads, B.C. and Mesh properties), this setting parameters are set
following rule: up by designers manually beforehand and keep invariable during
optimization process. After the FEA model is generated, the CAE
5m  10m m 6 10
n¼ ð20Þ solver is invoked and the predefined structural performance
100 m > 10 parameters are abstracted from the analysis results of the CAE sol-
At last of this stage the objective function and constraints of the ver. As an external controller, the integration junction-box controls
sampling points are evaluated by invoking the integration the execution sequence of CAD and CAE activities and smooths the
junction-box to run the CAD/CAE integration based FEA analysis. information transfer between them.
The responses and the corresponding sampling points constitute On the other side, the integration junction-box keeps the com-
the initial sample set. plicated integration process between CAD and CAE inside a ‘‘black
box’’, which hides the implementation details from the user, and
Approximation stage: Based on the sample set, surrogate models of presents only the interfaces (input/output) to it, as shown in
the objective function and constraints are constructed. Because of Fig. 10. The integration junction-box provides a straightforward
the high fit efficiency, the RBF network is utilized to construct mechanism to link various optimization techniques with HFL
the surrogate models. The proposed estimate of the width of basis structural analysis. When the proposed SAO is applied to the struc-
function based on the local density of sampling points has good tural design problems, it only needs to designate design variables
performance for both uniform and non-uniform sampling points. from the input interfaces and objectives as well as constraints from
At last, GA algorithm [11] is employed to obtain the optimum of the output interfaces. The originally laborious task to integrate
the surrogate-based optimization problem. The constraints are CAD and CAE software systems is now deposited to the junction-
handled with the penalty method. box, which handles it in an automatic fashion. Therefore, the asso-
ciation between the proposed SAO with HFL structural analysis is
Termination criteria: Termination of the proposed SAO is deter- greatly facilitated by the integration junction-box.
mined by the following criteria:

(1) If the relative distance between the optimal solutions of two 3.3. Realization of the proposed framework
successive iterations is less than 1‰, then evaluate criterion
(2). Otherwise, SAO will proceed to the Sampling stage. In order to realize the proposed framework, the chosen CAD and
(2) If the relative error between the two successive iterations’ CAE software packages should support automation, which means
optimal objective functions with the constraints imposed that all the tasks can be programmed in APIs and performed in
on using a penalty method is less than 0.0001, then evaluate batch mode [6]. Most modern high-end CAD and CAE software
criterion (3). Otherwise, SAO will proceed to the Sampling packages satisfy the functional requirements, and the engineering
stage. designers can implement the proposed integrated framework
(3) If the relative error between the objective functions of the depending on their available tools and individual preference.
surrogate model and the true model is less than 0.0001, In this study, we choose the combination of Pro/ENGINEER and
the convergence is reached and the proposed SAO algorithm ANSYS to establish the proposed integration framework. The API
will be terminated. Otherwise, SAO will proceed to the library of Pro/Toolkit provided by Pro/ENGINEER helps users to
Sampling stage. compile a dynamic link library (dll) file, which extracts a couple
of main geometry parameters to drive the modeling process auto-
matically. The CAE analysis tasks can be recorded and replayed by
Sampling stage: In this stage, the minimum distance d between APDL, a macro programming language provided by ANSYS. The
points in the sample set is first calculated. Based on the surrogate neutral data format IGES is selected to transfer the geometric
model constructed in Approximation stage, the adaptive sampling information from CAD to CAE. Besides, the SAO algorithm and
is performed by solving the optimization problem (17) using the the integration junction-box are developed using Java to make an
GA algorithm. Finally the objective function and constraints at open and automatic integration system for solving the structural
the optimal solution are evaluated by invoking the integration optimization problems.
junction-box to run the CAD/CAE integration based FEA analysis.
The optimal solution together with its true response of the objec-
tive function and constraints is added to the sample set to update
4. Benchmark tests
the surrogate model in the next iteration.
In this section, two case studies from simple to complex, from
3.2.2. CAD/CAE integration mechanism low dimension to moderate-dimension are taken from the
To facilitate the integration between CAD and CAE software engineering practices to investigate the efficacy of the proposed
packages and application of SAO, an integration junction-box is framework. The effectiveness and robustness of the proposed
developed, as shown in Fig. 9. SAO algorithm is validated in comparison with other optimization
On the one side, the proposed junction-box provides an integra- techniques in this section as well.
tion mechanism to bridge the gap between CAD and CAE domains.
The integration junction-box invokes the CAD package to generate
the model according to the geometry parameters and export the
model to the designated format suitable to CAE analysis in an auto-
matic fashion based on the software APIs. The junction-box saves
the exported geometry data file to the specified working directory,
and then executes the CAE system to play the predefined journal
file complied by its embedded PDL, which imports the exported
geometry data file from the working directory, generates the FEA Fig. 10. A component perspective of the Integration junction-box.
64 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

4.1. Test case 1

The design of an automobile component [1] taken from the

automotive industry is selected as the first test case. The proposed
structural design optimization framework is used to perform the
optimization problem. The CAD model generated in the CAD soft-
ware Pro/Engineer is present in Fig. 11. As shown in Fig. 12, four
geometry parameters including d1, d2, d3 and d4 are selected as
design variables to drive the modeling process. The feasible range
of the design variables is present in Table 2. The FEA analysis in
ANSYS is recorded by an APDL file, in which all analysis setting
parameters are embedded. Fig. 13 depicts the corresponding CAE
model from ANSYS. All degree of freedom on upper and lower con-
nection regions are restricted. The component is subjected to the
force F = 5.4 kN. The material’s Young’s modulus and Poisson’s
ratio are specified as 72 GPa and 0.3. The objective function of
the optimization problem is minimization of the volume while
the maximum stress is constrained to be less than 280 MPa. Fig. 12. Description of the selected design variables.
The dll file for CAD modeling and APDL file for CAE analysis are
deposited to the integration junction-box to construct the connec-
tion between them. To testify the efficacy of the presented Table 2
approach, the proposed SAO algorithm as well as other optimiza- Design space.
tion techniques including GA and PSO is connected to the integra- Design variable (mm) Lower bound Upper bound
tion junction-box in an easy way to carry out the test case,
D1 4.0 30.0
respectively. The number of initial sampling points for the SAO is D2 10.0 370.0
set as 40. In addition, the number of individuals of the population D3 10.0 180.0
for GA is 40 and the swarm for PSO is 20. Ten independent runs are D4 1.0 15.0
performed for the three methods to eliminate statistical influence.
The best and worst optimization results are elaborated in Table 3.
With respect to the best results, the optimal volumes of
320.99 cm3 and 326.71 cm3 are found after 55 iterations (2200
FEA evaluations) and 90 iterations (1800 FEA evaluations), respec-
tively. Comparatively, the proposed approach requires only 157
FEA evaluations to find the optimal result of 312.89 cm3. Figs. 14
and 15 records the iteration histories of the convergence criterions
and the objective of volume in the optimization process of the SAO
algorithm, respectively. It is clear from Fig. 14 that the termination
criteria are satisfied after 114 iterations and Fig. 15 shows that the
minimal volume of 312.89 cm3 is found. The total FEA evaluations
during the optimization consist of three parts: 40 sampling points,
114 optimization iterations and 3 evaluations of the termination
criterion 3. The stress distribution of the optimal structure is
depicted in Fig. 16.
It is clear from this test case that the proposed SAO algorithm is
provided with the global optimizing ability as GA and PSO, but the

Fig. 13. The CAE model.

number of computation-intensive FEA evaluation calls is reduced

from the order of 103 by GA or PSO to 102 by SAO, thus the struc-
tural design efficiency is well improved. Moreover, the dedicated
CAD/CAE integration junction-box facilitates the cooperation
between optimizers with the HFL structural analysis and extends
the application of SAO to engineering practices.

4.2. Test case 2

This section presents the design of a moderate-dimensional

case study of a bracket structure based on the proposed structural
design framework. The bracket serves as the connection structure
in the satellites as shown in Fig. 17. As depicted in Fig. 18, eleven
parameters that significantly affect the performance of the bracket
are selected as design variables. Table 4 presents the feasible range
Fig. 11. The CAD model. of the design variables. The CAE model of the bracket is shown in
 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 65

Table 3
Comparison of the design optimization results.

GA (The best) GA (the worst) PSO (the best) PSO (the worst) SAO (the best) SAO (the worst)
D1 24.35 22.99 16.14 16.28 17.79 23.25
D2 359.45 357.62 350.16 350.56 359.89 359.25
D3 160.51 156.48 157.95 156.88 161.21 158.25
D4 5.12 4.68 4.65 4.92 2.34 4.87
Volume (cm3) 320.99 327.89 326.71 329.71 312.89 318.56
Stress (MPa) 271 267 265 263 275 273
FEA evaluations 2200 2800 1800 2600 157 184

Fig. 16. Stress distribution of the optimal structure.

Fig. 14. The convergence history.

Fig. 17. Details of the CAD model.

Fig. 15. The iteration history of the objective function. settled as 100. In this test case, the proposed approach is also
benchmarked against other optimization techniques as shown in
Table 5. Pure GA with a population of 50 individuals and PSO with
Fig. 19. Young’s modulus and Poisson’s ratio are settled as 200 GPa a swarm of 40 individuals are also used to solve the optimization
and 0.3. The bracket is subjected to a stretching force and a bend- problem. Due to the computation costs, the optimization processes
ing moment induced by the forces P1 = 4.7 kN and P2 = 4.2 kN by GA and PSO are terminated before convergence. As shown in
loaded at the center of such screws as L1, L2, L3 and L4. In addition, Table 5, the optimal volume of 31493.33 cm3 is found after 120
the bracket is fixed at the screws including R1, R2, R3 and R4. The iterations (6000 FEA evaluations) by GA and 31510.26 cm3 after
optimization objective is to minimize the volume while the maxi- 140 iterations (5600 FEA evaluations) by PSO. Ten independent
mum stress is constrained to be less than 200 MPa and the total runs are performed by the proposed SAO algorithm and the best
displacement less than 2 cm. and worst results obtained are listed in Table 5. Figs. 20 and 21
The proposed structural design framework is used to solve this record the iteration histories of the design optimization process
optimization problem and the number of initial sampling points is for the best result, which indicates that the proposed SAO requires
66 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

Fig. 18. Description of the selected design variables.

Table 4
Design space.

Design variable (cm) Lower bound Upper bound Table 5

D1 1.0 3.5 Comparison of the design optimization results.
D2 12.0 40.0
Optimization Volume Displacement Stress FEA
D3 10.0 25.0
method (cm3) (cm) (MPa) evaluations
D4 2.0 6.0
D5 12.0 40.0 GA 31493.33 1.94 198.3 6000
D6 10.0 20.0 PSO 31510.26 1.96 195.4 5600
D7 1.5 4.0 SAO (the best) 31081.48 1.95 199.6 381
D8 5.0 13.0 SAO (the worst) 31381.48 1.93 197.3 412
D9 25.0 35.0
D10 35.0 55.0
D11 8.0 50.0

Fig. 19. Details of the CAE model. Fig. 20. The convergence history.
 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68 67

CAD/CAE integration. The approximation technique and the

sampling strategy are the two key factors for a successful SAO algo-
rithm. In this paper, a new method based on the local density of
sampling points and the technique of cross validation is proposed
to determine the width of Gaussian kernel functions, which
enhances the performance of the RBF network. Moreover, an
adaptive sampling strategy is developed to make a balance
between global exploration and local exploitation. The improved
SAO algorithm offers such benefits as easy connections to simula-
tion programs, ability to render a global optimum of the design
space and high computational efficiency, which makes it an ideal
optimization technique for structural design problems.
In order to widen the applications of the SAO algorithm in the
mechanical practices, an open framework is developed to associate
SAO with CAD/CAE integration, which allows the incorporation
between popular commercial CAD and CAE packages in a collabo-
rative environment. The CAD/CAE integration helps to exploit each
of them thoroughly and expand the SAO algorithm to various
complex structural design optimization problems. The feasibility,
convenience and efficacy of the proposed structural design
Fig. 21. The iteration history of the objective function.
optimization framework have been investigated through two case
studies. Furthermore, the effectiveness and computational
efficiency of the improved SAO algorithm are benchmarked against
other optimization approaches as well. The number of the
time-consuming FEA evaluations to find the optimum is reduced
by 1–2 orders of magnitude by the enhanced SAO algorithm. In a
word, the proposed integrated framework proves to be an
adequate strategy for handling structural design optimization
problems effectively and efficiently.

Acknowledgements

The authors acknowledge Prof. Derek Ingham in College of

Environmental and Materials Engineering of Leeds Uni. for his
advices and comments throughout this work.

References

[1] Yildiz AR. Comparison of evolutionary-based optimization algorithms for

structural design optimization. Eng Appl Artif Intell 2013;26:327–33.
[2] Tsuen B, Lin C, Kuo C. Application of an integrated CAD/CAE/CAM system for
stamping dies for automobiles. Int J Adv Manuf Tech 2008;35:1000–13.
[3] Park HS, Lee GB, Kim H. Collaborative engineering system supporting product
Fig. 22. Deformation and stress distribution of the optimized structure. development process. In: ASME design engineering technical conferences,
Chicago, USA, vol. 1, Part B; 2003. p. 873–80.
[4] De Martino T, Falcidieno B, Hassinger S. Design and engineering process
integration through a multiple view intermediate modeler in a distributed
only 278 iterations to converge to the optimal result of objected-oriented system environment. Comput Aided Des
31081.48 cm3. The number of FEA evaluations is 381 in sum (100 1998;30(6):437–52.
sampling points, 278 optimization iterations and 3 evaluations of [5] Gordon S. An analyst’s view: STEP-enabled CAD/CAE integration. In:
Presentation materials of NASA’s STEP for aerospace workshop. Pasadena
the termination criterion 3). The deformation and stress distribu-
(CA): Jet Propulsion Laboratory; January 16–19, 2001.
tion of the optimized bracket is depicted in Fig. 22. [6] Park HS, Dang XP. Structural optimization based on CAD/CAE integration and
As for this test case, the FEA analysis is a bit time-consuming meta-modeling techniques. Comput Aided Des 2010;42:889–902.
and more design variables are selected, which makes the optimiza- [7] Bonte MHA, Boogaard AH, Huétink J. An optimization strategy for industrial
metal forming processes. Struct Multidiscip Optimiz 2008;35:571–86.
tion techniques such as GA and PSO hard to operate. Due to the [8] Edke MS, Chang KH. Shape optimization of heavy load carrying components for
advantages of the proposed SAO algorithm, it preserves the ability structural performance and manufacturing cost. Struct Multidiscip Optimiz
to find the global optimum at quite low costs, indicating the pro- 2006;31:344–54.
[9] Dai L, Guan ZQ, Chen BS, Zhang HW. An open platform of shape design
posed SAO an effective optimization method for structural design optimization for shell structure. Struct Multidiscip Optimiz 2008;35:609–22.
optimization in mechanism practices. [10] Hare W, Nutini J, Tesfamariam S. A survey of non-gradient optimization
methods in structural engineering. Adv Eng Softw 2013;59:19–28.
[11] Goldberg D. Genetic algorithms in search, optimization, and machine
learning. New York (NY): Addison-Wesley; 1989.
5. Conclusion
[12] Kirkpatrick S, Gelatt C, Vecchi M. Optimization by simulated annealing.
Science 1983;220(4598):671–80.
Modern engineering design processes often rely on complex [13] Kennedy J, Eberhart R. Particle swarm optimization. In: IEEE international
computer-aided simulations to evaluate candidate designs. This conference on neural networks, vol. IV. Piscataway (NJ); 1995. p. 1942–8.
[14] Yildiz AR. An effective hybrid immune-hill climbing optimization approach for
research presents a structural design optimization framework solving design and manufacturing optimization problems in industry. J Mater
based on the improved SAO algorithm in conjunction with the Process Technol 2009;209:2773–80.
68 D. Wang et al. / Advances in Engineering Software 76 (2014) 56–68

[15] Karaboga D, Basturk B. A powerful and efficient algorithm for numerical [31] Vapnik V. The nature of statistical learning theory. New York: Springer; 1995.
function optimization: artificial bee colony (ABC) algorithm. J Glob Optimiz [32] Alexandrov N, Dennis JE, Lewis RM, Torczon V. A trust region framework for
2003;39:459–71. managing the use of approximation models in optimization. Struct Optimiz
[16] Yildiz AR. A novel hybrid immune algorithm for global optimization in design 1998;15:16–23.
and manufacturing. Robot Comput Integr Manuf 2009;25:261–70. [33] An W, Sun Y. An equivalence between SILF-SVR and ordinary Kriging. Neural
[17] Corriveau G, Guilbault R, Tahan A. Genetic algorithms and finite element Process Lett 2006;23:133–41.
coupling for mechanical optimization. Adv Eng Softw 2009;41(1):422–6. [34] Poggio T, Girosi F. Networks for approximation and learning. Proc IEEE
[18] Barthelemy JFM, Haftka RT. Approximation concepts for optimum structural 1990;78(9):1481–97.
design – a review. Struct Optimiz 1993;5:129–44. [35] Kitayama S, Yamazaki K. Simple estimate of the width in Gaussian kernel with
[19] Haftka RT, Scott EP. Optimization and experiments – a survey. Theor Appl adaptive scaling technique. Appl Soft Comput 2011;11:4726–37.
Mech 1996:303–21. [36] Nakayama H, Arakawa M, Sasaki R. Simulation-based optimization using
[20] Nikolaidis E, Long L, Ling Q. Neural networks and response surface computational intelligence. Optimiz Eng 2002;3:201–14.
polynomials for design of vehicle joints. In: Proceedings of the 7th AIAA/ [37] Hastie T, Tibshirani R, Friedman J. The elements of statistical learning. New
USAF/NASA/ISSMO symposium on multidisciplinary analysis & optimization. York: Springer; 2001.
American Institute of Aeronautics and Astronautics, AIAA 1998–1777; 1998. [38] Jones DR. A taxonomy of global optimization methods based on response
[21] Berke L, Hajela P. Application of artificial neural nets in structural mechanics. surfaces. J Glob Optimiz 2001;21:345–83.
Struct Optimiz 1992;4:90–8. [39] Xiong Y, Chen W, Apley D, Ding X. A non-stationary covariance-based Kriging
[22] Giunta A, Watson LT, Koehler J. A comparison of approximation modeling method for meta-modeling in engineering design. Int J Numer Methods Eng
techniques: polynomial versus interpolating models. In: Proceedings of the 2007;71:733–56.
7th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary analysis & [40] Wang DH, Wu ZP, Fei Y, Zhang WH. Structural design employing a sequential
optimization. American Institute of Aeronautics and Astronautics, AIAA approximation approach. Comput Struct 2014;134:75–87.
1998–4758; September 2–4, 1998. [41] Deng YM, Lam IC, Tor SB, Britton GA. A CAD-CAE integrated injection molding
[23] Simpson TW, Mauery TM, Korte JJ, Mistree F. Comparison of response surface design system. Eng Comput 2002;18:80–92.
and Kriging models for multidisciplinary design optimization. In: Proceedings [42] Lee SH. A CAD-CAE integration approach using feature-based multi-resolution
of the 7th AIAA/USAF/NASA/ISSMO symposium on multidisciplinary analysis & and multi-abstraction modeling techniques. Comput Aided Des
optimization. American Institute of Aeronautics and Astronautics, AIAA 1998– 2005;37:941–55.
4755; September 2–4, 1998. [43] Hamri O, Léon JC, Giannini F, Falcidieno B. Software environment for CAD/CAE
[24] Broomhead DS, Loewe D. Multivariate functional interpolation and adaptive integration. Adv Eng Softw 2010;41:1211–22.
networks. Adv Complex Syst 1988;2:321–55. [44] Hughes TJR, Cottrell JA, Bazilevs Y. Isogeometric analysis: CAD, finite elements,
[25] Wang GG, Shan S. Review of meta-modeling techniques in support of NURBS, exact geometry and mesh refinement. Comput Methods Appl Mech
engineering design optimization. J Mech Des 2007;129(2):370–80. Eng 2005;194:4135–95.
[26] Sakata S, Ashida F, Zako M. Structural optimization using Kriging [45] Albers A, Leon RN, Aguayo H, Maier T. Development of an engine crankshaft in
approximation. Comput Methods Appl Mech Eng 2003;192:923–39. a framework of computer-aided innovation. Comput Ind 2009;60(8):604–12.
[27] Huang D, Allen TT, Notz WI, Zeng N. Global optimization of stochastic black- [46] Matin I, Hadzistevic M, Hodolic J, Vukelic D, Lukic D. A CAD/CAE-integrated
box systems via sequential Kriging meta-models. J Glob Optimiz injection mold design system for plastic products. Int J Adv Manuf Tech
2006;34:441–66. 2012;42:1140–52.
[28] Kitayama S, Arakawa M, Yamazaki K. Sequential approximate optimization [47] Xu B, Chen N. An integrated method of CAD, CAE and multi-objective
using radial basis function network for engineering optimization. Optimiz Eng optimization. In: Proceedings of the 2009 IEEE 10th international conference
2011;12(4):535–57. in computer-aided industrial design & conceptual design (CAID & CD 2009);
[29] Forrester AIJ, Keane AJ. Recent advances in surrogate-based optimization. Prog 2009. p. 1010–14.
Aerosp Sci 2009;45:50–79. [48] Chang KH, Joo SH. Design parameterization and tool integration for CAD-based
[30] Hardy RL. Multi-quadratic equations of topography and other irregular mechanism optimization. Adv Eng Softw 2006;37:779–96.
surfaces. J Geophys 1971;76:1905–15.