Struct Multidisc Optim

DOI 10.1007/s00158-016-1592-1

INDUSTRIAL APPLICATION

Multi-objective and multi-case reliability-based design

optimization for tailor rolled blank (TRB) structures
Guangyong Sun 1,2 & Huile Zhang 1 & Jianguang Fang 2 & Guangyao Li 1 & Qing Li 2

Received: 14 April 2016 / Revised: 10 September 2016 / Accepted: 14 September 2016

# Springer-Verlag Berlin Heidelberg 2016

Abstract Light weight and crashworthiness signify two main Keywords Reliability based design optimization .
challenges facing in vehicle industry, which often conflict with Crashworthiness . Tailor rolledblank . Multiple loading cases .
each other. In order to achieve light weight while improving Multi-objective optimization
crashworthiness, tailor rolled blank (TRB) has become one of
the most potential lightweight technologies. To maximize the
characteristics of TRB structures, structural optimization has 1 Introduction
been adopted extensively. Conventional optimization studies
have mainly focused on a single loading case (SLC). In prac- The research attention in automotive industry has been paid to
tice, however, engineering structures are often subjected to vehicle lightweight and crashworthiness due to ever-growing
multiple loading cases (MLC), implying that the optimal design requirements in environmental concerns, government legisla-
under a certain condition may no longer be an optimum under tions and consumer demanding (Sun et al. 2010).
other loading cases. Furthermore, traditional deterministic op- Unfortunately, these two performances always conflict with
timization could become less meaningful or even unacceptable each other. To maintain crashworthiness performance during
when uncertainties of design variables and noises of system vehicular lightweighting, high strength steel (HSS) and ultra-
parameters are present. To address these issues, a multi- high strength steel (UHSS) have been widely adopted to re-
objective and multi-case reliability-based design optimization place conventional steel. In this regard, the Ultralight Steel
(MOMCRBDO) was developed in this study to optimize the Auto Body (ULSAB) project achieved a body-in-white
TRB hat-shaped structure. The radial basis function (RBF) (BIW) weight reduction of 68 kg (from 271 kg to 203 kg)
metamodel was adopted to approximate the responses of ob- by using HSS (Kim et al. 2010). Zhang et al. (2006) developed
jectives and constraints, the non-dominated sorting genetic al- a rule to carry out lightweight design of automobile parts by
gorithm II (NSGA-II), coupled with Monte Carlo Simulation replacing mild steel with HSS, which was validated by numer-
(MCS), was employed to seek optimal reliability solutions. The ical simulations. Jiang et al. (2012) designed the door beam
optimal results show that the proposed method is not only with UHSS to realize full marks of crash tests, with a stiffness
capable of improving the reliability of Pareto solutions, but also increase of 2.5 times, strength increase of 3.8 times, and
enhancing the robustness under MLC. weight reduction of 9.32%.
Although the aforementioned methods of material substi-
tution can reduce vehicle weight effectively, the high prices of
these materials hinder their large-scale application in a highly
* Guangyao Li competitive market. Moreover, conventional uniform thick-
[email protected] ness structures may not exert their maximum capacities of
crashworthiness and light weight. Therefore, tailor welded
1
State Key Laboratory of Advanced Design and Manufacture for blank (TWB) process has been developed as an advanced
Vehicle Body, Hunan University, Changsha 410082, China manufacturing technology. TWB structures are manufactured
2
School of Aerospace, Mechanical and Mechatronic Engineering, The by welding metal sheets with different materials and/or thick-
University of Sydney, Sydney, NSW 2006, Australia ness prior to the forming process (Merklein et al. 2014). Thus,
 Sun et al

not only does it save materials appreciably, but also provides a many nondeterministic optimization algorithms have been pro-
more flexible combination of different sheets. In this regard, posed to take into account the effects of various uncertainties
frontal side rails, the inner door panels and B-pillars are some (Cheng et al. 2006; Youn and Choi 2004) and effectively re-
typical examples (Li et al. 2015; Pan et al. 2010; Shi et al. solved design problems in real-life (Du and Chen 2004; Fang
2007; Xu et al. 2013). However, since the material properties et al. 2015b; Gu et al. 2013b; Li et al. 2011; Zhu et al. 2011). In
in the thermal influence zone can be rather different from base this regard, Du and Chen (2004) proposed the sequential optimi-
materials, potentially causing stress concentration and leading zation and reliability assessment (SORA) method and used in
to high risk of fatigue failure. The transverse movement of reliability-based design for vehicle crashworthiness of side im-
seam during deep drawing is disadvantageous and the seam pact. Chen et al. (2013) developed an optimal shifting vector
could abrade the tools significantly (Hirt et al. 2005). approach to enhancing the efficiency of reliability-based design
Recently, a new rolling technology, namely tailor rolled optimization (RBDO) for the design of honeycomb cellular
blank (TRB), is developed to overcome these defects of structures. Fang et al. (2013) presented a reliability-based multi-
TWB. Compared with TWB, TRBs allow a continuous tran- objective design optimization for the design of a vehicle door. To
sition between the thick and thin zones leading to better form- the authors’ best knowledge, however, there have been limited
ability and a higher surface quality. Furthermore, the produc- studies on the multi-objective reliability-based design optimiza-
tion cost of TRBs is independent of the number of thickness tion (MORBDO) for multiple loading case problems.
transitions (Merklein et al. 2014). To understand the crushing The rest of the paper is organized as follows: Section 2 pro-
behaviors of TRB, TRB thin-walled structures under axial and poses the mathematical modeling of multi-objective and multi-
lateral impacts have been studied recently (Sun et al. 2015b; case reliability-based design optimization (MOMCRBDO) pro-
Sun et al. 2014). For example, Sun et al. (2014) investigated cedure and relevant algorithms, including Monte Carlo simula-
the crashworthiness of a square TRB tube under an axial im- tion (MCS), radial basis function (RBF) metamodeling and the
pact. Chuang et al. (2008) adopted the TRB manufacturing non-dominated sorting genetic algorithm II (NSGA-II) optimi-
technology for designing a vehicle underbody. Duan et al. zation. In Section 3, the finite element modeling of TRB hat-
(2016) redesigned a front longitudinal beam using the TRB shaped (TRBHS) structures are developed and then the quasi-
technology, and claimed that the weight of FLB was reduced static axial crushing and drop-hammer impact tests are per-
by 15.21%, whereas the crashworthiness was improved com- formed to validate the accuracy of the FE models. Section 4
pared with the baseline design. depicts the optimization process of MOMCRBDO for
While there have been some studies on the design of TRB TRBHS, followed by the results and discussions. Finally,
structures, most of them were conducted under a specific load- Section 5 draws some conclusions.
ing case. As a matter of fact, practical engineering structures are
often subjected to multiple load cases (MLC), implying that an
optimal design under a certain loading condition may not meet
the performance requirements under other loading conditions 2 Methods for design analysis and optimization
(Fang et al. 2015a). The literature indicated that although struc-
tural optimization allows enhancing crashworthiness of vehicle 2.1 Multi-objective and multi-case reliability-based design
in a specific loading case, the effectiveness of optimization optimization
could be challenged if different crushing velocities and direc-
tions are considered (Duddeck and Wehrle 2015; Zhang et al. Generally speaking, engineering structures likely have to be
2014). From a design perspective, an optimum is expected to operated in MLC, and thus a design is expected to be an
accommodate not only a single loading case (SLC), but also optimum under MLC. A multi-objective deterministic design
MLC. In this regard, Zhang et al. (2014) proposed a dual optimization (MODDO) under MLC can be formulated as
weight factor method to optimize the hollow and conical tubes follows (Qiu et al. 2015):
under MLC. Qiu et al. (2015) investigated the multi-cell hex-
agonal tube under MLC by using complex proportional assess- 8
> X
q
>
> λk f i;k ðxÞ i ¼ 1; 2; …; m
ment and multi-objective optimization approaches. >
> min
>
>
Furthermore, these abovemntioned MLC designs were large- >
< s:t:
k¼1
g j; k ðxÞ ≤ 0 j ¼ 1; 2; …; n
ly restricted to deterministic optimization, in which all design ð1Þ
>
> X q
parameters involved are certain. However, engineering problems >
> λk ¼ 1 λk ≥0
>
>
>
>
inevitably involve uncertainties in loads, geometry, material : k¼1
xL ≤ x ≤ xU
properties and operational conditions, etc., in which a determin-
istic optimization could lead to unreliable or unstable designs ð1Þ
thereby increasing risk of design failure (Choi et al. 2006; Fang where x denotes the vector of design variables, fi,k(x) is the ith
et al. 2014; Yang and Gu 2004; Fang et al. 2016). Therefore, objective under the kth load case; gj,k(x) is the jth inequality
Multi-objective and multi-case reliability-based design

constraint function; m and n are the numbers of objective where ^p f is the estimated failure probability, and N is the
functions and inequality constraints, respectively; λk is the number of sample points, xi for i = 1, 2, …, N are the sample
weighting factor to reflect the relative importance and/or oc- points generated according to f(x).
currence probability of the kth loading case, q represents the The accuracy of MCS estimation can be quantified with the
number of load cases considered; xL and xU are the lower and standard error, defined as:
upper bounds, respectively. Eq. (1) is a typical deterministic
optimization problem.
σ f ðxÞ
To obtain the multi-objective reliable optimization under error ¼ pffiffiffiffi ð5Þ
MLC, MOMCRBDO is explored in this study. Different from N
the deterministic optimization, reliability-based design seeks an The error is therefore unrelated to the problem dimension
optimum subject to certain probabilistic constraints, mathemat- (i.e., the number of design variables), which is very appealing
ically expressed as: for large-scale problems. And the error is proportional to
pffiffiffiffi
8 1= N , implying that the improvement of accuracy by one
> X
q
>
> λk f i ;k ðxÞ i ¼ 1; 2; …; m order of magnitude will require 100 times more samples.
>
> min
>
> Such computational cost can be prohibitive in application
>
> hk¼1 i
< for complex and highly nonlinear problems.
s:t: p g j; k ðxÞ ≤0 ≥R j;k j ¼ 1; 2; …; n
ð2Þ On the other hand, the minimum sampling size required for
>
> X
q
>
> the failure probability level p[g(x) ≤ 0] as suggested by (Tu
>
> λk ¼ 1 λk ≥0
>
> et al. 1999) is:
>
: k¼1
x ≤ x ≤xU
L
10
L¼ ð6Þ
p½ g ð x Þ ≤ 0 
where the design feasibility is formulated as the probability
which indicates that for a 10% estimated probability
(p[·]) of constraint satisfaction (i.e., gj,k(x) ≤ 0) bigger than or
of failure; about 100 function evaluations (e.g., FEA
equal to a desired probability Rj,k.
runs) are required with some confidence on the first
digit of failure prediction. To verify an event having a
2.2 Monte Carlo simulation (MCS) method
1% failure probability; about 1000 structural analyses
are required, which would be usually considered too
The MOMCRBDO problems in Eq. (2) always involve a pro-
expensive for some engineering applications. To im-
cedure to evaluate the failure probability or reliability of con-
prove computational efficiency, MCS combined with
straints. Many reliability analysis methods have been devel-
metamodels, has been employed to quantify the failure
oped, including the approximation methods (e.g., the first order
probability, which makes a great number of model eval-
and second order reliability analysis methods), direct integra-
uations feasible (Abdessalem and El-Hami 2015; Fang
tion, and sampling methods (e.g., MCS). Among which, MCS
et al. 2013; Jansson et al. 2008).
is the most commonly used approach for its accuracy and ap-
plicability (Gu and Yang 2005).
2.3 Metamodeling
Mathematically, the failure probability pf can be formulated
by the multivariate integration, given as (Melchers 1987):
Crashworthiness optimization requires a considerable
Z Z þ∞
number of nonlinear finite element (FE) runs which
pf ¼ f ðxÞdx ¼ I ½g ðxÞ ≤ 0f ðxÞdx ¼ E fI ½g ðxÞ ≤ 0g ð3Þ
gðxÞ ≤ 0 −∞ typically leads to high computational cost, especially
for a MOMCRBDO problem in Eq. (2). As an alterna-
where x is the vector of random parameters, g(x) is the tive, the metamodeling techniques have been widely
performance function defined such that failure occurs when used, which can largely reduce the number of FE runs.
g(x) < 0, f(x) is the joint probability density function, I[·] is In this regard, the radial basis function (RBF) has ex-
the indicator function for event g(x) < 0, having the value 1 hibited a fairly good accuracy for highly nonlinear prob-
if event g(x) < 0 and value 0 otherwise. E{·} is the expec- lems (Fang et al. 2005; Sun et al. 2011; Xu et al. 2013)
tation of the indicator function. and was utilized to construct the metamodel for re-
According to Eq. (3), pf can be evaluated using MCS as sponses in this study.
follows: To construct a metamodel accurately, design of exper-
iment (DOE) needs to be first employed to sample the
1 XN design space. Among many available DOE approaches,
^
pf ¼ I ðx Þ
i¼1 ½gðxi Þ ≤ 0 i
ð4Þ
N the optimal Latin hypercube sampling (OLHS) (Park
 Sun et al

1994; Sun et al. 2012) has proven effective and was im- where K is the number of newly generated validation points, yi
plemented here to generate sample points. Then, the re- is the true values, ŷ i is the corresponding approximate
sponses of these sample points need to be evaluated by metamodel value, and y is the mean of yi. In general, the larger
FEA for constructing RBF models. the R2 values, the more accurate the metamodel. The smaller
A typical RBF model can be formulated as (Hardy 1971): the RAAE and RMAE, the better the metamodel.

X
N
X
M
2.4 Non-dominated sorting genetic algorithm II
f ð xÞ ¼ wi ϕðkx‐xi kÞ þ c j p j ð xÞ ð7Þ
i¼1 j¼1
(NSGA-II)

In this study, the NSGA-II (Deb et al. 2002) was

where N is the number of sampling points, ∥x ‐ xi∥ is the employed to obtain the non-dominated optimal
Euclidean norm of design variable vectors x and the ith solutions for the multi-objective optimization problem.
sampling point xi, ϕ is a basis function, wi is the un- The main features of NSGA lie in that it ranks
known weighting factor positioned at the ith sampling solutions with non-dominated sorting and assigns them
point; M is the number of polynomial terms, pj(x) are in fitness based on their ranks. While the crossover and
the polynomial terms and cj is the corresponding coef- mutation operators remain similar to a simple GA, the
ficient for p j (x), usually M < N. Obviously, an RBF selection operator distinguishes itself. As an improve-
model is actually a linear combination of N radial basis ment of NSGA, NSGA-II is characterized by a fast
functions and M polynomial terms with the weighted non-dominated sorting procedure that is an elitist strat-
coefficients. egy, a parameter-less diversity-preservation mechanism
Of those feasible basis functions, the multi-quadric and a simple yet efficient constraint-handling method.
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi Many crashworthiness design problems (Gu et al.
formulation (specifically ϕðrÞ ¼ r2 þ c2, where c is
the free shape parameter) was chosen for its prediction 2013a; Gu et al. 2013b; Liao et al. 2008; Xu et al.
accuracy as well as commonly linear and possibly ex- 2013) have been successfully solved using NSGA-II.
ponential rate of convergence with increasing sampling The details of NSGA-II can be found in Deb et al.
points (Acar et al. 2011). More discussion about RBF in (2002).
crashworthiness design can be found in literature (Fang During the optimization progress, accurate RBF
et al. 2005). models were first constructed with the given weighting
It should be noted that RBF passes through all sam- factors for different loading case. Then, the NSGA-II
ple points, meaning that the fitness accuracy of an ap- algorithm was employed to perform the deterministic
proximate function from the existing sampling points and non-deterministic optimization with MCS. For clar-
cannot be checked directly. For this reason, a series of ification, the proposed MOMCRBDO procedure for the
additional validation points were generated to verify the crashworthiness optimization of TRB structures is fur-
accuracy of the constructed metamodels. Three different ther depicted in the flowchart (Fig. 1).
fitting indicators, namely R-square (R2), relative average
absolute error (RAAE) and relative maximum absolute
error (RMAE), are employed here (Jin et al. 2001) and 3 Numerical modeling and experimental validation
given as
3.1 Description of geometrical features
XK  2
i¼1
y i −^y i
R2 ¼ 1− X K  2 ð8Þ The front longitudinal beam is the most significant
y −y energy-absorbing component for frontal impacts, and
i¼1 i
its collapse modes and energy absorbing capability can
X K  

greatly influence the full vehicle crashworthiness and
i¼1
 y i −^
y i  safety of passengers. To simplify the complexity of real
RAAE ¼ X K   ð9Þ front longitudinal beam, a representative TRB hat-
 
i¼1
 y i −y i  shaped (TRBHS) structure was extracted for design
analysis and optimization, as shown in Fig. 2.
n   o
    The TRBHS specimen consists of a hat-shaped sheet
max y1 −^y1 ; …; yK −^yK  and a bottom sheet which are joined by spot-welding
RMAE ¼ X K  . ð10Þ
 along the center line of the flange with a spot diameter

i¼1 i
y −^
y i K
of 5mm at an interval of 30 mm, as shown in Fig. 3a.
Multi-objective and multi-case reliability-based design

Fig. 1 The flowchart of multi-objective and multi-case reliability-based design optimization

These two sheets are divided into three different thick- quasi-static loading case and dynamic loading case
ness zones, namely thin zone, thick zone and thickness were established. The material of hat-shaped specimen
transition zone (TTZ), respectively. The total length considered herein is high-strength steel HSLA 340,
and the flange width of TRBHS are L = 400 mm and whose density, Poisson’s ratio and Young’s modulus
30 mm, respectively. The corner radii near the weld are 7.8 × 10 3 kg/m 3 , 0.3 and 210 GPa, respectively.
flange and at the top edges are R = 4 mm. The other To consider the non-uniform material properties of
detailed dimensions of the TRBHS are illustrated in TRB due to the variable thickness, Duan et al. (2016)
Table 1. The transition slope of the TTZ are set as have developed the effective stress vs. effective strain
1:100 to meet the economic requirement (Hirt et al. relation of TRBs made of HSLA 340 by the Lagrange
2005). Herein, the length of the TTZ, l2 in Fig. 3b, polynomial interpolation, as illustrated in Fig. 4. Based
is 40 mm. on the interpolation surface in Fig. 4b, the material
properties can be obtained for any thickness.

3.2 Numerical modeling 3.2.2 Finite element modeling

3.2.1 Material constitutive model for TRB The explicit nonlinear FE code LS-DYNA was
employed to simulate the crashworthiness of the
To systematically understand the crashworthiness char- TRBHS. Figure 5 shows the FE model of the TRBHS
acteristics of TRBHS, the finite element models under subjected to axial impact. The 4-node quadrilateral
Belytschko-Lin-Tsay shell elements with reduced inte-
gration were used to model the TRBHS walls. To model
the thickness variation more realistically, different thick-
nesses were assigned to four nodes of the shell elements
using the keyword *ELEMNET_SHELL_THICKNESS
in LS-DYNA (Hallquist 2007), as shown in Fig. 6.
Five integration points were employed across the thick-
ness to capture the local element bending, and stiffness-
type hourglass control was utilized to eliminate spurious
zero energy models. The constitutive behavior of
HSLA340 was modeled via the piecewise linear plastic-
ity material model, MAT 24, in LS-DYNA (Hallquist
2007; Kopp et al. 2005). To accurately consider the
different material properties in the different thicknesses,
Fig. 2 The typical hat-shaped structure in vehicle the elements with nearly the same thickness were
 Sun et al

Fig. 3 Description of a
manufacturing process, b the
dimensions of the TRBHS

(a)

(b)

defined as one component and each component was rigid platen impacted onto the specimen at an initial velocity
assigned its own mechanical properties that were inter- of 8 m/s with an additional mass of 706 kg. The “automatic
polated from the effective stress vs. effective strain re- single surface” algorithm was defined to simulate the self-
lationship (i.e. Fig. 4) as illustrated in Fig. 6. To deter- contact of the specimen in buckling and “automatic node to
mine the proper mesh size, several simulations were surface” algorithm was defined between the specimen and the
carried out and an element size of 3 × 3 mm was found two rigid walls. The static and dynamic coefficients of fric-
sufficient for the numerical simulations. The two parts tions were set as 0.35 and 0.25, respectively.
of the TRBHS were welded together along the center Since HSLA340 is a strain-rate sensitive material, the
line of the flange by employing a constrained spot-weld strain-rate effect should be taken into account. In this
option. study, the effect was accounted for using the Cowper-
The FE model of quasi-static axial crushing tests is shown Symonds model given as (Hallquist 2007):
in Fig. 5a, where the two platens were both modeled to be
rigid and the bottom one was fixed. A constant velocity V = 5 2 0 p 11=p 3
    εeff
mm/min was applied to the top platen to gradually crush the p
6 7
σy εpeff ; εeff ¼ σsy εpeff 41 þ @ A 5 ð11Þ
specimen. The contact between the specimen and two rigid C
walls was modeled using “automatic node to surface” algo-
rithm. The “automatic single surface” algorithm was  
employed to the specimen itself to avoid interpenetration dur- where σy εpeff ; εpeff is the dynamic yield stress, σsy(εpeff) is the
ing axial collapse. static stress and εpeff is the strain rate. C and p denote two
Figure 5b shows the FE modeling of dynamic impact tests. strain-rate parameters.
The bottom of the specimen was fixed to a rigid wall and a

Table 1 The dimensions of the TRBHS 3.3 Experimental validation

Parameters t1 t2 l1 w h
To validate the TRBHS modeling accuracy under different
Dimension (mm) 1.2 1.6 180 105 85 loading cases, the quasi-static crushing test and the drop-
hammer impact test were carried out in this study.
Multi-objective and multi-case reliability-based design

were respectively placed at the upper and lower end

of the specimens without other constraints. The thin
zone of the TRBHS specimens was placed at the inci-
dent (top) end for all the tests to generate progressive
folding collapses. A constant crushing speed of 5 mm/
min was set. The total compression displacement was
set as a constant of 270 mm, which represents 67.5%
of the specimen length.
The drop-hammer impact tests were performed in a
dynamic test rig, as illustrated in Fig. 8. To prevent the
toppling of the specimens, a steel plate was welded to
the bottom end of the specimens. The other steel plate
was placed at the top end of the specimens to ensure a
central loading. The dynamic procedure was conducted
at the initial impact velocity 8 m/s with an additional
mass of 706 kg.

3.3.2 Validation of the FE model

To systematically study the crashworthiness of structures,

crashworthiness indicators, i.e., peak crash fore (Fmax) and
energy absorption (EA), were usually used, as shown in
Fig. 5c.
The EA of a structure subjected to the axial compression
loading can be calculated as:
Z dmax
Fig. 4 Effective stress–strain relationship a Different thicknesses, b
Interpolation surface (Duan et al. 2016) EA ¼ F ðxÞdx ð12Þ
0

where dmax is the maximum crushing distance, F(x) de-

3.3.1 Experimental details notes crashing force. The peak value of F(x) during the
whole crashing progress is denoted as Fmax.
The quasi-static crushing tests were performed in a stan- The experimental and corresponding simulation re-
dard universal testing machine MTS647, as shown in sults are given in Figs. 9 and 10, respectively. It is clear
Fig. 7. To ensure a central crushing, two steel plates that the force versus displacement and energy versus

Fig. 5 Schematic of a quasi-static axial crushing tests b dynamic impact tests c indicators of axial crushing
 Sun et al

Fig. 6 Schematic of FE modeling of TRB with a variable thickness

displacement curves of the FE simulation agree well parameter was changed each time to quantify the parametric
with the experimental results. Moreover, the deformation influence by comparing with Case 0. All the cases in this
modes of the simulation match very well with those of study were summarized in Table 2.
the experiments. The satisfactory correlations suggested Figures 11 and 12 plot the histograms of crashworthiness
that the FE models is able to well predict the crash indicators (i.e., Fmax and EA), respectively. From these two
process of the TRBHS specimen and can be employed figures, it is easily found that the structural parameters signif-
for the further design analysis and optimization. icantly affect the TRBHS crashworthiness. Specifically, Fmax
is strongly influenced by the thickness of thin zone (t1) while
3.3.3 Numerical results and discussion EA is affected by all the structural parameters considered.
Note that greater values of the thicknesses of the thin and thick
To better understand the crashworthiness of TRBHS under zones (t1, t2) can improve EA, but a larger value of the thick-
axial crushing loading, a range of TRBHSs with different ness of the thin zone (t1) influences Fmax negatively. Besides,
dimensions were investigated using the FE simulation. As EA decreased with the increase in the length of thin zone (l1)
shown in Table 2, Case 0 is the experimentally-validated and the effect of sectional dimensions on EA is far greater than
model. Other cases were modeled to investigate the effects that on Fmax. More importantly, the change trends of design
of the thicknesses of the thin and thick zones (t1, t2), the variables to the crashworthiness indicators significantly differ
position of TTZ (l1) and sectional dimensions (w, h) of under the quasi-static and dynamic crashes, further reinforcing
TRBHS on crashworthiness. For Cases 1 to 4, only one the strategy of design optimization under MLC.

Fig. 7 Quasi-static axial crushing

tests setup
Multi-objective and multi-case reliability-based design

Fig. 8 Drop-hammer impact test

setup

4 Multi-objective and multi-case reliability-based absorption and minimize the mass, while limiting the
design optimization peak force during collapse (White and Jones 1999).
Thus, the EA and mass M of TRBHS are chosen as
4.1 Definition of optimization problem objective functions, whilst the peak force Fmax should
be constrained within a certain level. To account for the
For addressing crashworthiness and lightweight criteria, the effects of different loading cases, the deterministic
design optimization presented here aims to maximize energy multi-objective optimization problem can be defined

Fig. 9 Comparison between the quasi-static FE simulation and corresponding experimental test: a impact force versus displacement curves, b energy
versus displacement curves, c deformation modes under quasi-static axial crushing
 Sun et al

Fig. 10 Comparison between the dynamic FE simulation and corresponding experimental test: a impact force versus displacement curves, b energy
versus displacement curves, c deformation models under drop-hammer impact tests.

mathematically as To take into account the effects of uncertainties of design

variables and obtain the reliable optimal design under MLC, the
8 ( )
>
> Xq
X
q MOMCRBDO can be formulated as
>
> min − λk EAk ðt 1 ; t 2 ; l 1 ; w; hÞ; λk M k ðt1 ; t 2 ; l1 ; w; hÞ
>
> 8 ( )
>
> k¼1 k¼1
X
q
X
q
< >
>
s:t: F max; k ðt 1 ; t 2 ; l 1 ; w; hÞ ≤ F 0max;k >
> min − λk EAk ðt1 ; t 2 ; l1 ; w; hÞ; λk M k ðt 1 ; t 2 ; l 1 ; w; hÞ
>
>
>
> X q >
> h k¼1 ik¼1
>
> >
<
>
> λk ¼ 1 λk ≥ 0 p F max; k ðt1 ; t 2 ; l1 ; w; hÞ ≤ F 0max;k ≥ Rk
>
>
s:t:
: k¼1
>
1 ≤t 1 ≤ t2 ≤2 50 ≤l 1 ≤ 250 80 ≤ w≤ 2 80 60 ≤ h ≤ 100 >
> X
q
>
>
>
> λk ¼ 1 λk ≥ 0
ð13Þ >
>
: k¼1
1 ≤t 1 ≤ t2 ≤2 50 ≤l 1 ≤ 250 80 ≤w ≤2 80 60 ≤ h ≤ 100
ð14Þ
where F0max,k is the
limit to the peak force; the thicknesses
of thin and thick zones (t1, t2), the location of TTZ (l1) where the design feasibility is formulated as the probability
and sectional dimensions (w, h) are chosen as the design (p[·]) of constraint satisfaction Fmax,k(t1, t2, l1, w, h) ≤ F0max,k
variables, seen in Fig. 3b. bigger than or equal to a desired probability Rk, which was set
When the number of loading case, q in Eq. (13), is set as different levels of 90%, 95%, 99% in this study.
as 1, the optimization problem would downgrade to a To solve the problems defined in Eq. (13) and Eq. (14), the
SLC problem, which represents either the quasi-static multi-objective optimization procedure was employed based on
loading case or the dynamic loading case herein. F0max,k the RBF metamodels, NSGA-II algorithm with MCS. The de-
for the quasi-static and the dynamic loading cases were tailed parameters of NSGA-II employed herein were summa-
set as 90 kN and 320 kN herein, respectively. When load rized in Table 3. It is assumed that all the design variables are
case number q is set as 2, λ1 and λ2 represent the weight normally disturbed with a coefficient of variation of 0.05. For
factors for the quasi-static loading case and the dynamic the probabilistic constraints defined in Eq. (14), MCS were
loading case, respectively. performed to estimate the reliability by employing metamodels
Multi-objective and multi-case reliability-based design

Table 2 Parametrical analysis of the TRBHSs 5. Although the free shape parameter c has certain effect on the
Case t1 t2 l1 w h RBF performance, it is hard to say which one is the most suit-
able for all the cases. In general, regardless of c value, these
Case 1 1.2 1.6 180 105 65 (↓) metamodels are considered sufficiently accurate to be
1.2 1.6 180 85 (↓) 85 employed in the crashworthiness design. Furthermore, some
1.2 1.6 160 (↓) 105 85 researchers (Gu et al. 2013b; Xu 2015; Xu et al. 2013) found
1.2 1.4 (↓) 180 105 85 that c = 0.5 is suitable for most crashworthiness problems, thus
1.0 (↓) 1.6 180 105 85 c = 0.5 is adopted herein.
Case 2 1.2 1.6 180 105 75 (↓)
1.2 1.6 180 95 (↓) 85
4.2 Result of MORBDO under different SLC
1.2 1.6 170 (↓) 105 85
1.2 1.5 (↓) 180 105 85
Figure 13 presents the Pareto fronts for the deterministic and
1.1 (↓) 1.6 180 105 85
reliable designs with different reliability levels under the
Case 0 1.2 1.6 180 105 85
quasi-static loading case and dynamic loading case, respec-
Case 3 1.3 (↑) 1.6 180 105 85 tively. It can be clearly observed that these two objective EA
1.2 1.7 (↑) 180 105 85 and M conflict with each other for all the cases, indicating that
1.2 1.6 190 (↑) 105 85 any increase in EA always leads to an undesirable increase in
1.2 1.6 180 115 (↑) 85 M, and vice versa. Although the shapes of the Pareto fronts
1.2 1.6 180 105 95 (↑) under different reliability levels look similar, their ranges
Case 4 1.4 (↑) 1.6 180 105 85 change fairly evidently. That is to say, the uncertainties of
1.2 1.8 (↑) 180 105 85 design variables have an important influence on the crashwor-
1.2 1.6 200 (↑) 105 85 thiness of the TRBHS. More specifically, the Pareto front of
1.2 1.6 180 125 (↑) 85 the reliable design was further away from the deterministic
1.2 1.6 180 105 105 (↑) design when the reliability level increases.
Furthermore, it can be noted that the Pareto fronts can be
divided into an insensitive region in the upper-left and a sen-
instead of FEA runs, where the OLHS was employed to gener- sitive region in the lower-right. The insensitive region pos-
ate 100 sample points and then their responses were extracted sesses a smaller mass, which would possibly result in a lower
from FEA. Based on the simulation results, the RBFs of EA, M peak force and push the design far away from the constraint
and Fmax under different loading cases can be easily construct- boundary. Thus, this region has a higher reliability, where the
ed. To analyze the influence of the free shape parameter c on the Pareto fronts of deterministic design and reliable design al-
accuracy of RBFs, c = 0, 0.25, 0.5, 0.75 and 1 were adopted to most coincide. On the contrary, the sensitive region had a
construct the metamodels for quasi-static and dynamic loading lower reliability and the Pareto front of reliable design was
cases, respectively. 18 new validation points were generated by pushed away from the deterministic counterpart.
OLHS and used to evaluate the accuracy of the RBF models. Although the Pareto fronts provide the designer with
The results of R2, RAAE and RMAE were listed in Tables 4 and a great number of solutions over the Pareto space,

Fig. 11 Variation of a Fmax and b EA for different design parameters under quasi-static crushing
 Sun et al

Fig. 12 Variation of a Fmax and b EA for different structural parameters under dynamic crushing

Table 3 Details of the NSGA-II parameters used in this study Tables 7 and 8. The relative error of the deterministic
optimal design was summarized in Table 8. It can be
NSGA-II parameter name Value seen easily that the maximum error was less than 6.6%,
which indicates that the metamodel-based optimization
Population size 100
was reasonably accurate. From Tables 6 and 7, it can
Generation 100
be seen that both the deterministic design and reliable
Probability of crossover 0.9
design can improve the crashworthiness of the TRBHS.
Distribution index for crossover 10
However, the reliability of the deterministic optimum
Distribution index for mutation 20
was very low. The crashworthiness of TRBHS gradually
decreased when the reliability requirement increases
decision must be made for the most satisfactory solution from 90% to 99%. Therefore, a compromise should be
from the Pareto-sets. The minimum distance selection made between the desired reliability level and the ob-
method (TMDSM) (Sun et al. 2010) was employed jective performances in practical applications.
herein to obtain the most satisfactory solution (namely
a Knee point) for different loading cases. A brief over- 4.3 Result of MOMCRBDO
view of the TMDSM was presented in the Appendix.
The reliable optima for R = 90%, 95%, 99% along with Because the impacting velocity was really unpredictable
the deterministic optimum and baseline design were in car accidents, Therefore, it is essential and practical
summarized in Tables 6 and 7. And MCS with 100, to consider the uncertainty of the collision velocity in
500, 800 and 1000 samples were conducted, respective- designing the energy-absorbing structures. To take into
ly, at each optimum to obtain the corresponding reliabil- account the effects of multiple load cases, the weighting
ities. It turned out that 1000 MCSs was adequate herein. factor approach, as formulated in Eqs. (13) and (14),
The reliabilities of the constraints were also presented in was adopted. To investigate the effect of weighting

Table 4 Accuracy evaluation for different free shape parameter c under quasi-static loading case

c EA M Fmax

R2 RAAE RMAE R2 RAAE RMAE R2 RAAE RMAE

0 0.9832 0.1361 2.081 0.9996 0.0207 1.775 0.9858 0.1393 1.789

0.25 0.9854 0.1326 1.889 0.9997 0.0193 1.987 0.9887 0.1305 1.632
0.5 0.9873 0.1273 1.698 0.9997 0.0173 1.144 0.9877 0.1355 1.375
0.75 0.9886 0.1222 1.677 0.9993 0.0287 1.905 0.9906 0.1350 1.311
1 0.9865 0.1175 1.706 0.9998 0.0134 1.742 0.9855 0.1360 1.457
Multi-objective and multi-case reliability-based design

Table 5 Accuracy evaluation for different free shape parameter c under dynamic loading case

c EA M Fmax

R2 RAAE RMAE R2 RAAE RMAE R2 RAAE RMAE

0 0.9646 0.2037 2.1421 0.9974 0.3590 2.230 0.8938 0.3690 2.012

0.25 0.9644 0.1999 2.2365 0.9633 0.2989 2.583 0.9072 0.3427 2.063
0.5 0.9648 0.1972 2.256 0.9946 0.2281 1.755 0.9263 0.3219 2.117
0.75 0.9647 0.1962 2.257 0.9487 0.1731 2.851 0.9227 0.3057 2.168
1 0.9675 0.1899 2.206 0.9697 0.1317 2.884 0.9276 0.2917 2.218

Fig. 13 Pareto fronts for a quasi-static loading case; and b dynamic loading case

factors on the optimum, three design cases were consid- factors for different loading cases had considerable ef-
ered here to emphasize static and dynamic crash differ- fect on the optimal solution. Taking Case 1 (Fig. 14a)
ently, as follows: as an example, the Pareto front located fairly closely to
the optimal solution generated from the dynamic loading
8 case, while the Pareto fronts generated by the other
< Case 1 : λ1 ¼ 0:25; λ2 ¼ 0:75
cases were far away. This was because a heavier
Case 2 : λ1 ¼ 0:5; λ2 ¼ 0:5 ð15Þ
: weighting factor was placed to the dynamic loading
Case 3 : λ1 ¼ 0:75; λ2 ¼ 0:25
case (i.e., λ2 = 0.75). Similarly, Case 3 assigned a heavi-
To compare the deterministic design under SLC and er weight to the quasi-static loading case, thus its Pareto
MLC, the corresponding Pareto fronts were plotted in front was close to the quasi-static loading solution, as
Fig. 14a together. Obviously, the selection of weight shown in Fig. 14a. The trend of Case 2 was not obvious

Table 6 MCSs at deterministic and reliable optimal design under Table 7 MCSs at deterministic and reliable optimal design under
quasi-static loading case dynamic loading case

Response Baseline Deterministic 90% 95% 99% Response Baseline Deterministic 90% 95% 99%
Reliable Reliable Reliable Reliable Reliable Reliable

EA 8.8120 10.199 9.5471 9.5594 9.0928 EA 15.933 19.918 18.851 18.667 18.380
M 2.2480 1.8610 1.7840 1.8084 1.7473 M 2.2480 1.5590 1.5173 1.5390 1.5944
Fmax 87.143 89.523 79.179 77.416 73.000 Fmax 336.17 319.46 300.73 296.05 287.55
Feasibility / 48.42% 90.87% 94.71% 99.14% Feasibility / 52.47% 90.24% 95.00% 99.15%
 Sun et al

Table 8 Error analysis of the deterministic optimal design optimal solution under quasi-static loading case may
Description EA M Fmax not be an optimum in dynamic loading case, and vice
versa. The specific solution to the quasi-static loading
Quasi-static loading case Optimal 10.199 1.8610 89.523 case was able to simultaneously reduce the M and
FE model 9.6800 1.8560 94.653 enhance the EA compared to the dynamic loading
error % 5.08% 0.27% −5.73% case. That is to say, the optimal designs for SLC only
Dynamic loading case Optimal 19.918 1.5590 319.46 favored a specific loading case. When the designs
FE model 18.519 1.6410 327.18 were subjected to other loading case, the crashworthi-
Error (%) −5.06% −6.60% −3.93% ness would deteriorate to a certain extent. It is inter-
esting to note that the Pareto fronts for MLC were all
located not far away from the corresponding Pareto
because it emphasized all the loading cases equally. fronts for SLC in the design space as shown in
The difference of these Pareto fronts indicates that Fig. 15. That is to say, although the optimal designs
the selection of weighting factors for different loading under MLC are not able to obtain the optimal solu-
cases was critical in the MOMCRBDO solution. tions for a specific single case, they can obtain a
Therefore, weighting factors for different loading case range of compromise solutions for the MLC.
should be determined based on the occurrence frequen- To obtain the reliable optimal design for MLC, the
cy or statistical data in real life. MOMCRBDO were performed at different reliability
Fig. 15 compared the Pareto solutions of MLC for levels (i.e., R = 90%, 95%, 99%) and the optimal re-
different weighting factors under quasi-static and dy- sults were plotted in Fig. 14c-d. It can be easily seen
namic loading. From which, it is easily found that the that the range of Pareto front with MOMCRBDO was

Fig. 14 Pareto fronts for a deterministic design b reliable design of Case 1 c reliable design of Case 2 and d reliable design of Case 3
Multi-objective and multi-case reliability-based design

Fig. 15 Pareto Solutions of MLC for different weighting factors: a quasi-static loading case b dynamic loading case

Table 9 MCSs at deterministic and reliable optimal design (Case 1)

Response Baseline Deterministic 90% Reliable 95% Reliable 99% Reliable

EA 14.152 15.584 14.711 14.499 14.220

M 2.2480 1.6381 1.5769 1.5556 1.4917
Fmax,1/Fmax,2 87.248/336.17 89.975/279.83 81.214/272.33 79.016/270.74 74.951/271.27
Feasibility / 50.43%/94% 90.4%/99% 95.34%/100% 99.04%/99.98%

significantly less than the MODDO for MLC, and the means that although the objective performances were
Pareto fronts of the reliable design moved gradually sacrificed, the reliability of design increased significant-
toward right as the reliability requirement increases. ly through MOMCRBDO.
In other words, the EA needs to be decreased to ac-
commodate the uncertainty of design variables in
MOMCRBDO. This phenomenon is consistent with 5 Conclusions
the results of MORBDO for SLC.
To obtain the most satisfactory solution from the Vehicular crashworthiness design without considering
Pareto-fronts, the TMDSM was employed and the opti- the random nature of design variables and unpredictabil-
mum results were listed in Tables 9, 10 and 11, respec- ity of collision would reduce the reliability of structures.
tively. The MCS with 1000 samples were also per- Based on the reliability-based multi-objective design
formed at each optimum to obtain the corresponding strategy, the tailor rolling blank hat-shaped (TRBHS)
reliability. Compared with the baseline design, it can structure was optimized in this study. To systematically
be seen that the MOMCDDO solution can improve the investigate the crashworthiness of the TRBHS structure,
crashworthiness of the TRBHS, but is worse than the the FE modeling was developed and the corresponding
deterministic optimal solutions. In addition, the reliabil- FE analysis was conducted by employing the explicit
ity of the reliable optimal solution was higher. This nonlinear FE code LS-DYNA. Then the quasi-static

Table 10 MCSs at deterministic and reliable optimal design (Case 2)

Response Baseline Deterministic 90% Reliable 95% Reliable 99% Reliable

EA 12.373 13.601 13.669 13.920 13.328

M 2.2480 1.6881 1.7690 1.8093 1.7368
Fmax,1/Fmax,2 87.248/336.17 89.999/276.85 79.397/280.92 76.781/271.26 71.945/270.33
Feasibility / 50%/96% 90.24%/94.34% 95.16%/99.88% 99.06%/99.93%
 Sun et al

Table 11 MCSs at deterministic and reliable optimal design (Case 3)

Response Baseline Deterministic 90% Reliable 95% Reliable 99% Reliable

EA 10.592 12.263 11.625 11.506 11.145

M 2.2480 1.8527 1.7821 1.7771 1.7343
Fmax,1/Fmax,2 87.248/336.17 89.766/300.29 79.524/287.68 76.928/275.72 72.119/273.69
Feasibility / 48.62%/87.37% 89.78%/98.07% 95.10%/99.83% 99.31%/99.88%

axial crushing and drop-hammer impact tests were per- Sun is a recipient of Australian Research Council (ARC) Discovery
Early Career Researcher Award (DECRA) in the University of Sydney.
formed to validate the FE models. The uncertainties in
key geometric dimensions and weighting factors for dif-
ferent loading cases were considered, and NSGA-II and
MCS were integrated to seek the Pareto fronts for the
multi-objective and multi-case reliability-based design
Appendix: brief of the minimum distance selection
optimization (MOMCRBDO) problems. To provide the
method (TMDSM)
decision-maker with insightful information, the MODDO
In this study, TMDSM is employed to obtain a most satisfac-
and MORBDO were conducted to maximize the energy
tory solution from the Pareto-sets. Brief description of the
absorption and minimize the mass simultaneously under
TMDSM is provided below.
single and multiple loading cases respectively. From the
The TMDSM can be mathematically formulated as (Sun
Pareto fronts, it can be seen that the proposed method is
et al. 2010):
not only capable of improving the reliability of Pareto
solutions, but also generates more competent solution to vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
u n
the crashworthiness for MLC. Finally, the selected opti- u
N X
mums from these Pareto fronts demonstrate that the ob- min D ¼ t ð f cτ −minð f τ ðxÞÞÞN ðA1Þ
jective performances could be sacrificed somewhat to τ¼1

satisfy the reliability constraints compared with

MODDO, indicating that compromise should be made where fcτ is the τ th objective value in the c th Pareto set and n
between the desired reliability level and objective per- is the number of the objective functions, N = 2, 4, 6,…, D is
formances in crashworthiness design. the least distance from the “Knee point” to a “utopia point (see
Nevertheless, there exist some limitations in this work. Fig. 16), which is composed of the optimum value of each
Firstly, this study considered only the uncertainties in geomet- objective and normally not attainable in real life problems
rical dimensions as well as unpredictability of collision veloc- with conflicting objectives.
ity. Despite their critical importance to optimization of TRB
structure, other uncertainties, such as collision direction, can
affect the reliability of the optimal solutions to some extent,
which should be taken into account. Secondly, the weighting
factor technique was adopted to combine different loading
cases. It turned out that the weighting factors have significant
influence on the Pareto’s location, range and shape. For this
reason, the weighting factors for different load cases should
follow the statistical data and/or occurrence frequency under
multiple impact velocity in real life. Thirdly, the
MOMCRBDO presented in this paper used the RBF
metamodels to approximate the responses. However, for
multi-variable problems, constructing a fine RBF metamodel
requires substantial number of FE analyses, whose computa-
tional cost may become prohibitive in more complex
problems.

Acknowledgments This work is supported by National Natural Fig. 16 The knee point on the Pareto front having the least distance from
Science Foundation of China (51575172, 61232014). Dr Guangyong the utopia point
Multi-objective and multi-case reliability-based design

References Jansson T, Nilsson L, Moshfegh R (2008) Reliability analysis of a

sheet metal forming process using Monte Carlo analysis and
metamodels. J Mater Process Technol 202:255–268
Abdessalem AB, El-Hami A (2015) A probabilistic approach for Jiang C, Shan Z, Zhuang B, Zhang M, Xu Y (2012) Hot
optimising hydroformed structures using local surrogate stamping die design for vehicle door beams using ultra-high
models to control failures. Int J Mech Sci 96:143–162 strength steel International. Int J Precis Eng Manuf 13:1101–
Acar E, Guler M, Gerceker B, Cerit M, Bayram B (2011) Multi- 1106
objective crashworthiness optimization of tapered thin-walled
Jin R, Chen W, Simpson TW (2001) Comparative studies of
tubes with axisymmetric indentations. Thin-Walled Struct 49:
metamodelling techniques under multiple modelling criteria.
94–105
Struct Multidiscip Optim 23:1–13
Chen Z, Qiu H, Gao L, Li P (2013) An optimal shifting vector approach for
Kim HJ, McMillan C, Keoleian GA, Skerlos SJ (2010)
efficient probabilistic design. Struct Multidiscip Optim 47:905–920
Greenhouse Gas emissions payback for lightweighted vehi-
Cheng G, Xu L, Jiang L (2006) A sequential approximate pro-
cles using aluminum and high‐strength steel. J Ind Ecol 14:
gramming strategy for reliability-based structural optimiza-
929–946
tion. Comput Struct 84:1353–1367
Kopp R, Wiedner C, Meyer A (2005) Flexibly rolled sheet metal
Choi SK, Grandhi R, Canfield RA (2006) Reliability-based struc-
and its use in sheet metal forming. Adv Mater Res 6–8:81–
tural design. Springer Science & Business Media
92
Chuang C, Yang R, Li G, Mallela K, Pothuraju P (2008)
Multidisciplinary design optimization on vehicle tailor rolled Li F, Luo Z, Sun G (2011) Reliability-based multiobjective design optimi-
blank design. Struct Multidiscip Optim 35:551–560 zation under interval uncertainty. Comput Model Eng Sci 74:39–64
Deb K, Pratap A, Agarwal S, Meyarivan T (2002) A fast and Li G, Xu F, Huang X, Sun G (2015) Topology optimization of an auto-
elitist multiobjective genetic algorithm: NSGA-II. IEEE motive tailor-welded blank door. J Mech Des 137:055001–055001
Trans on Evolutionary Computation 6:182–197 Liao X, Li Q, Yang X, Zhang W, Li W (2008) Multiobjective
Du X, Chen W (2004) Sequential optimization and reliability optimization for crash safety design of vehicles using step-
assessment method for efficient probabilistic design. J Mech wise regression model. Struct Multidiscip Optim 35:561–569
Des 126:225–233 Melchers RE (1987) Structural reliability. Horwood
Duan L, Sun G, Cui J, Chen T, Cheng A, Li G (2016) Crashworthiness Merklein M, Johannes M, Lechner M, Kuppert A (2014) A re-
design of vehicle structure with tailor rolled blank. Struct view on tailored blanks-production, applications and evalua-
Multidiscip Optim 53:321–338 tion. J Mater Process Technol 214:151–164
Duddeck F, Wehrle E (2015) Recent advances on surrogate modeling for Pan F, Zhu P, Zhang Y (2010) Metamodel-based lightweight design of B-
robustness assessment of structures with respect to crashworthiness pillar with TWB structure via support vector regression. Comput
requirements. 10th European LS-DYNA Conference 2015, Struct 88:36–44
Wurzburg, Germany Park JS (1994) Optimal Latin-hypercube designs for computer experi-
Fang H, Rais-Rohani M, Liu Z, Horstemeyer MF (2005) A com- ments. J Stat Plan Inference 39:95–111
parative study of metamodeling methods for multiobjective Qiu N, Gao Y, Fang J, Feng Z, Sun G, Li Q (2015)
crashworthiness optimization. Comput Struct 83:2121–2136 Crashworthiness analysis and design of multi-cell hexagonal
Fang J, Gao Y, Sun G, Li Q (2013) Multiobjective reliability- columns under multiple loading cases. Finite Elem Anal Des
based optimization for design of a vehicle door. Finite 104:89–101
Elem Anal Des 67:13–21. Shi Y, Zhu P, Shen L, Lin Z (2007) Lightweight design of automotive
Fang J, Gao Y, Sun G, Zhang Y, Li Q (2014) Crashworthiness design of front side rails with TWB concept. Thin-Walled Struct 45:8–14
foam-filled bitubal structures with uncertainty. Int J Non Linear Sun G, Li G, Zhou S, Li H, Hou S, Li Q (2010) Crashworthiness design of
Mech 67:120–132 vehicle by using multiobjective robust optimization. Struct
Fang J, Gao Y, Sun G, Qiu N, Li Q (2015a) On design of multi-cell tubes Multidiscip Optim 44:99–110
under axial and oblique impact loads. Thin-Walled Struct 95:115–126
Sun G, Li G, Gong Z, He G, Li Q (2011) Radial basis functional model for
Fang J, Gao Y, Sun G, Xu C, Li Q (2015b) Multiobjective robust
multi-objective sheet metal forming optimization. Eng Optim 43:
design optimization of fatigue life for a truck cab Reliability.
1351–1366
Reliab Eng Syst Saf 135:1–8
Sun G, Li G, Li Q (2012) Variable fidelity design based surrogate
Fang J, Sun G, Qiu N, Kim NH, Li Q (2016) On design optimization for
and artificial bee colony algorithm for sheet metal forming
structural crashworthiness and its state of the art. Struct and
process finite. Finite Elem Anal Des 59:76–90
Multidiscip Optim. doi:10.1007/s00158-016-1579-y
Gu L, Yang R (2005) On reliability-based optimisation methods Sun G, Xu F, Li G, Li Q (2014) Crashing analysis and
for automotive structures International. Int J Mater Prod multiobjective optimization for thin-walled structures with
Technol 25:3–26 functionally graded thickness. Int J Impact Eng 64:62–74
Gu X, Sun G, Li G, Huang X, Li Y, Li Q (2013a) Multiobjective optimi- Sun G, Fang J, Tian X, Li G, Li Q (2015a) Discrete robust
zation design for vehicle occupant restraint system under frontal optimization algorithm based on Taguchi method for structur-
impact. Struct Multidiscip Optim 47:465–477 al crashworthiness design. Expert Syst Appl 42:4482–4492
Gu X, Sun G, Li G, Mao L, Li Q (2013b) A comparative study Tu J, Choi KK, Park YH (1999) A new study on reliability-based design
on multiobjective reliable and robust optimization for crash- optimization. J Mech Des 121:557–564
worthiness design of vehicle structure. Struct Multidiscip White M, Jones N (1999) Experimental quasi-static axial crushing
Optim 48:669–684 of top-hat and double-hat thin-walled sections. Int J Mech
Hallquist JO (2007) LS-DYNA keyword user’s manual. Livermore Sci 41:179–208
Software Technology Corporation 971 Xu F (2015) Enhancing material efficiency of energy absorbers through
Hardy RL (1971) Multiquadric equations of topography and other graded thickness structures. Thin-Walled Struct 97:250–265
irregular surfaces. J Geophys Res-Atmos 76:1905–1915 Xu F, Sun G, Li G, Li Q (2013) Crashworthiness design of multi-
Hirt G, Abratis C, Ames J, Meyer A (2005) Manufacturing of sheet metal component tailor-welded blank (TWB) structures. Struct
parts from tailor rolled blanks. J Technol Plasticity 30:1–12 Multidiscip Optim 48:653–667
 Sun et al

Yang R, Gu L (2004) Experience with approximate reliability- Zhang Y, Sun G, Xu X, Li G, Li Q (2014) Multiobjective crashworthiness
based optimization methods. Struct Multidiscip Optim 26: optimization of hollow and conical tubes for multiple load cases.
152–159 Thin-Walled Struct 82:331–342
Youn BD, Choi KK (2004) A new response surface methodology for Zhu P, Zhang Y, Chen G (2011) Metamodeling development for
reliability-based design optimization. Comput Struct 82:241–256 reliability-based design optimization of automotive body structure.
Zhang Y, Lai X, Zhu P, Wang W (2006) Lightweight design of automo- Comput Ind 62:729–741
bile component using high strength steel based on dent resistance.
Mater Des 27:64–68