Materials and Design 31 (2010) 1917–1929

Contents lists available at ScienceDirect

Materials and Design

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

Multiobjective robust optimization method for drawbead design

in sheet metal forming
Guangyong Sun a,b, Guangyao Li a, Zhihui Gong a, Xiangyang Cui a, Xujing Yang a, Qing Li b,*
a
State Key Laboratory of Advanced Design and Manufacturing for Vehicle Body, Hunan University, Changsha 410082, China
b
School of Aerospace, Mechanical and Mechatronic Engineering, The University of Sydney, Sydney NSW 2006, Australia

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

Article history: It is recognized that fracture and wrinkling in sheet metal forming can be eliminated via an appropriate
Received 3 February 2009 drawbead design. Although deterministic multiobjective optimization algorithms and finite element
Accepted 25 October 2009 analysis (FEA) have been applied in this respect to improve formability and shorten design cycle, the
Available online 28 October 2009
design could become less meaningful or even unacceptable when considering practical variation in
design variables and noises of system parameters. To tackle this problem, we present a multiobjective
Keywords: robust optimization methodology to address the effects of parametric uncertainties on drawbead design,
Multiobjective robust optimization
where the six sigma principle is adopted to measure the variations, a dual response surface method is
Sheet metal forming
Particle swarm optimization
used to construct surrogate model and a multiobjective particle swarm optimization is developed to gen-
Multicriteria erate robust Pareto solutions. In this paper, the procedure of drawbead design is divided into two stages:
Dual response surface method firstly, equivalent drawbead restraining forces (DBRF) are obtained by developing a multiobjective robust
Drawbead design particle swarm optimization, and secondly the DBRF model is integrated into a single-objective particle
swarm optimization (PSO) to optimize geometric parameters of drawbead. The optimal design showed a
good agreement with the physical drawbead geometry and remarkably improve the formability and
robust. Thus, the presented method provides an effective solution to geometric design of drawbead for
improving product quality.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction
With increasing requirements in functions and aesthetics,
highly sophisticated designs of sheet metal workpieces have made
its manufacturability more and more demanding. Design of sheet
metal forming process has become one of the major concerns in
manufacturing engineering, which determines quality and cost of
product. Conventional process design is largely based upon experi-
ence available by incorporating with a trial and error procedure. As
a result, the development of a new die often requires numerous
prototype tests, leading to a long design cycle and significant cost.
Development of advanced computational technology, represented
by finite element analysis (FEA), has been changing such philoso-
phy, which enables us to precisely predict a forming process and
detect such defects as wrinkling and fracture in a design stage,
thereby reducing design and prototyping costs to a considerable
extent. In this respect, Makinouchi [1] adopted finite element
method (FEM) to predict the defects of fracture, wrinkling and
springback of the sheet successfully. Panthi et al. [2] utilized FEM
to analyze springback in sheet metal bending process. Dong and
* Corresponding author. Tel./fax: +61 2 9351 8607.
E-mail address: Q.Li@usyd.edu.au (Q. Li).
Lin [3] explored the formability of a Santana 200 exterior panel
and achieved well-correlated results with the experiment. Chen
et al. [4] investigated the effects of blank holder gap and shell ele-
ment type on the formability of a washing-trough. It was found
that the extent of wrinkling in the flange region of the blank be-
came severer with the increase in the blank holder and thus an
optimal blank holder gap was recommended for the given process
parameters. Nevertheless, these abovementioned practical indus-
trial applications appeared to usually employ FEA in an iterative
fashion, where the improvement of forming process still somewhat
relied on designer’s experience. To attain a satisfactory result,
many FEA runs may be needed to manually alter the design model
parameters and then re-evaluate the results, whereas this by no
means guarantees a global optimum. In this sense, the capacity
of computer aided engineering (CAE) may have not be fully taken
up yet. Therefore, some research attempts have been made on
how to transform FEA from a passive verification tool to a more ac-
tive design tool in the sheet metal forming process recently.
Computational optimization signifies a more effective tool by
seeking for an optimal design systematically, which helps engi-
neers to attain the best possible formability of sheet metal produc-
tion. For example, Ohata et al. [5] incorporated the sweeping
simplex method with FEA to optimize the punch travel and

0261-3069/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.matdes.2009.10.050
1918 G. Sun et al. / Materials and Design 31 (2010) 1917–1929
forming stages for obtaining a uniform thickness distribution. Nac-
eur et al. [6] employed a mathematical programming algorithm in
the inverse procedure to optimize the restraining forces and then
to design the drawbead. Azaouzi et al. [7] developed an automatic
design procedure within commercial FEA program and adopted a
Heuristic Optimization Algorithm (HOA) for the blank shape design
of high precision metallic parts in a special stamping process. Guo
et al. [8] combined an inverse procedure with a sequential qua-
dratic programming (SQP) technique to optimize the blank shape.
Although these algorithms that directly incorporate with sheet me-
tal forming FEA are capable of optimizing the forming processes,
the implicit relationship and the great complexity of sensitivity
analysis in a context of material and geometric nonlinearities as
well as frictional contact dynamics could largely compromises
the feasibility and precision in practical applications of prevalent
mathematical programming techniques [9].
As an effective alternative, such surrogate or metamodel tech-
niques as response surface method (RSM) have been exhaustively
adopted in sheet metal forming optimization. In this regard, Chen-
gzhi et al. [10] presented a new method integrating the finite ele-
ment method with adaptive response surface method (ARSM) to
determine an optimum blank holder force during box deep draw-
ing process. Naceur et al. [11] developed a new RSM involving
Moving Least Squares regression models and pattern search opti-
mization for the rapid design of aluminum sheet metal forming
parameters. Kok and Stander [12] adopted successive response sur-
face method (SRSM) to optimize the preforming die shape in terms
of the blank weight, thereby minimizing the difference in work-
piece thickness. Kayabasi and Ekici [13] integrated FEA, RSM and
genetic algorithm together to find the most appropriate values of
forming process parameters. Huang et al. [14] used RSM to opti-
mize the intermediate tool surfaces in the multi-step sheet metal
stamping process to obtain improved quality of a final product.
Ohata et al. [15] employed RSM to optimize the annealing temper-
ature and time for thickness uniformity of stamping part. Hu et al.
[16] adopted an adaptive RSM to design blank shape and blank
hold forces for thickness uniformity, in which a significant
improvement was made. Breitkopf et al. [17] developed a new
RSM involving Moving Least Squares regression models and pat-
tern search algorithms for achieving uniform thickness, where a
one step solver is used in optimization procedure and the incre-
mental solver for final verification. Jansson et al. [18] employed
RSM to optimize draw bead restraining force, whose combination
with a space mapping technique largely enhanced computing effi-
ciency. Later, they further improved this by using an iterative RSM
[19]. Tang et al. [20] also addressed the design of restraining force
for minimizing the thickness difference subject to the constraints
of failure criteria, where the one step method was adopted in con-
structing response functions. Jakumeit et al. [21] utilized Kriging
model to optimize blank holder force, whose objective was to form
fracture and wrinkling free workpiece with an acceptable thinning
and springback.
These above mentioned sheet metal forming studies are re-
stricted on deterministic optimization, where it is assumed that
all the design variables and parameters involved are certain [22].
It is noted that the optimal solutions are often pushed to limits
of design constraints, leaving no room for tolerances/uncertainties
in modeling, simulation and/or manufacturing capabilities avail-
able. Practically, all real-life sheet metal forming are indeed nonde-
terministic, which involve some degree of uncertainties in
lubricative situation, material properties, geometries, manufactur-
ing precision and actual usage, etc. Consequently, nondeterministic
optimization problems solved by deterministic optimization algo-
rithms may result in unreliable designs. In order to take into ac-
count various uncertainties in sheet metal forming process, Li
et al. [23–25] presented a CAE-based six sigma robust optimization
procedure, which integrated probability optimization, six sigma
criterion and robust design concept. Demir et al. [26] proposed
an effective design strategy, which integrates FEA, approximate
model, numerical optimization algorithm and probabilistic design
method (Monte Carlo simulation) into an automated design tool
to design sheet metal die for reducing stress and increase fatigue
life of sheet metal die. However, these algorithms involve only sin-
gle-objective function. Sheet metal forming is typically character-
ized by a number of quality and/or performance indices, for
example, cracking, wrinkling and springback [27], some of which
could conflict with each other. Despite its significant practical va-
lue, the development of multiobjective robust optimization meth-
ods for sheet metal forming process has been under-studied.
In sheet metal forming process, the initial blank (thickness, con-
tour and surface), process parameters (boundary conditions, hold-
ing forces, lubrication conditions, drawbead types and positions,
etc.) and the material properties (yield stress, hardening, anisot-
ropy, etc.) can also affect the forming quality of workpiece [19].
In these parameters, the drawbead is one of the most important
parameters to control the material flow and final quality. Very hea-
vy restraining forces can prevent the sheet from drawing-in but
may cause fracture, whereas insufficient forces may lead to wrin-
kling (Fig. 1). Therefore, drawbead should be optimized in die de-
sign. It is noted that there has been some published work
available about the optimization design of drawbead. For example,
Wang et al. [28] developed response surface methodology based on
a so-called intelligent sampling to optimize drawbead. Naceur
et al. [29] integrated an inverse approach, BFGS algorithm and ana-
lytical sensitivity analysis into optimization of restraining forces.
However, these studies are mainly optimizing drawbead restrain-
ing forces with single-objective and without considering fluctua-
tions of variables and design parameters. In practice, a series of
modifications on the drawbead positions and dimensions are nec-
essary in order to adjust these restraining forces. Furthermore,
sheet metal forming process is essentially a multiobjective optimi-
zation problem, in which the requirements of avoiding fracture and
wrinkling are often contradictory from each other. Moreover, vari-
ations of design variables and system parameters affect the form-
ability. Under the circumstance, the effectiveness of a single and/
or deterministic design optimization may be problematical.
This paper aims to address two abovementioned major issues of
(1) multicriteria and (2) nondeterministic design for sheeting me-
tal forming process. An effective multiobjective robust optimiza-
tion method will be developed and applied to a real-life
drawbead design based on dual response surface models. To deal
with this problem, the design procedure is divided into two stages
herein: firstly, the optimal equivalent drawbead restraining forces
are obtained through a multiobjective robust optimization; and
secondly, the equivalent restraining force model is integrated into
a particle swarm optimization to optimize the geometric parame-
ters of drawbead.
2. Methods and materials
To develop multiobjective robust optimization for sheet metal
forming, we firstly establish the objective functions with respect
to the drawbead forces by using a dual response surface approach,
one for mean and another for standard deviation in each objective.
Following this, a multiobjective particle swarm optimization pro-
cedure is applied to optimize the drawbead forces. Then, a sin-
gle-objective particle swarm optimization is carried out to
inversely determine the geometric parameters of drawbead for
generating the desirable optimal drawbead forces. This section will
introduce these relevant methods.
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929 1919

Fig. 1. Fracture and wrinkling defects in the forming process.

(
2.1. Dual response surface methodology (DRSM) El ðbÞ ¼ eTl el ¼ ðyl  ul bÞT ðyl  ul bÞ
ð2Þ
Er ðcÞ ¼ eTr er ¼ ðyr  Wr cÞT ðyr  Wr cÞ
Although the development of advanced numerical techniques
and powerful computers have been allowed us to simulate sophis- where b = [b1, b2, . . . , bV]T and c = [c1, c2, . . . , cW]T denote the
ticated sheet metal forming, incorporation of optimization algo- unknown coefficients, ul and Wr are the matrices consisting of ba-
rithms still present several challenges at present. Firstly, each sis functions evaluated at these selected design points, respectively,
forming simulation must involve highly nonlinear FEA and often as:
require considerable computational cost. Optimization character-
izes an iterative process. If a design needs hundreds of FEA to con- 2 3 2 3
/1 ðx1 Þ    /V ðx1 Þ w1 ðx1 Þ    wW ðx1 Þ
verge, its practical value will be largely compromised. For this 6. 7 6. 7
6
ul ¼ 4 .. .. .. 7; 6
Wr ¼ 4 .. .. .. 7
reason, the sheet metal forming simulations should be performed . . 5 . . 5
as few as possible. Secondly, due to nonlinear characteristics of /1 ðxm Þ    /V ðxm Þ w1 ðxm Þ    wW ðxm Þ
sheet metal forming simulation and numerical inaccuracies, the
ð3Þ
behaviors of objective and constraint functions may not be smooth
mathematically. As a result, sensitivity based optimization meth- The use of the least square technique can determine b and c as
ods often face significant difficulty to efficiently generate a correct follows [31]:
optimum. Thirdly, the deformation behaviors involved can be very
complicated, where there are some difficulties in calculating ele- (
b ¼ ðuTl ul Þ1 uTl yl
ment and contact forces, sometimes making sheet metal forming ð4Þ
simulation unstable. As such, collapse of single simulation may c ¼ ðWTr Wr Þ1 WTr yr
lead to the whole optimization procedure be terminated. For these
The choice of sample points can follow the Optimal Latin Hyper-
reasons, the metamodel technique appears effective and will be
cube sampling (OLHS) approach [32]. The errors in the dual RS
adopted in this study.
model can be caused by the selection of basis functions, selection
A traditional response surface (RS) method usually deals with
of sampling points and the least squares [33]. Several typical mea-
the mean value of response Y without considering any variance.
sures, such as relative error (RE) and statistical analysis methods
However, this may be inadequate and sometimes the optimization
can be as applied to evaluate the accuracy of the surrogate models
could become even pointless [30]. The DRSM, on the other hand,
[31,33].
presents two models: one for the mean (l) and another for the
The DRSM simultaneously considers the estimated mean and
standard deviation (r), as:
the standard deviation. The specific steps of generating the dual re-
8 sponse surfaces can be summarized as follows:
>
> P
V
~ l þ eu ¼
< yl ¼ y
> bv /v ðxÞ þ el
v ¼1 (1) Set up the mathematical model of optimization problem
ð1Þ
>
> P
W
with objective functions, constraints and definition of the
> ~ r þ er ¼
: yr ¼ y cw ww ðxÞ þ er
w¼1 design space.
(2) Determine the design variables (controllable factors) and
where yl and yr denote the mean and standard deviation of the true noise factors (uncontrollable factors).
responses from FEA, y ~l and y ~r are the dual RS models for the (3) Construct cross product arrays (see Table 1) to sample the
mean and standard deviation, el and er are the modeling errors of design space and capture variability, where the noise factors
mean and standard deviation, respectively. x = [x1, x2, . . . , xn]T de- sampled with design of experiment are arranged in an outer
notes the vector of design variables, bv (v = 1, . . . , V) the vth un- array, and the control factors sampled with design of exper-
known coefficient corresponding to the vth basis function uv(x), iment are arranged in an inner array.
cw (w = 1, . . . , W) the wth unknown coefficient corresponding to (4) Run numerical experiments using FEA and extract response
the wth basis function ww(x). In theory, basis function can be in values from simulation results.
any form, but in most cases, the polynomial functions have proven (5) Calculate the mean and standard variation of each response
effective for both mean and standardh deviation responses.
iT h and select basis function to establish dual response surface
ð1Þ ð2Þ ðmÞ ð1Þ
If FEA results yl ¼ yl ; yl ; . . . ; yl and yr ¼ yr ; model.
ð2Þ ðmÞ T
yr ; . . . ; yr  are obtained at m carefully selected sample design (6) Evaluate the errors of the dual response surface
points (m P V and m P W), the total squared errors between the model.
analyses and dual RS models can be computed as: (7) Repeat above steps until the errors become acceptable.
1920 G. Sun et al. / Materials and Design 31 (2010) 1917–1929

Table 1 where Yl1(x), Yl2(x), . . . , Ylk(x) and Yr1(x), Yr2(x), . . . , Yrk(x) are the
Cross product arrays design. mean values and standard deviations of these k objectives, respec-
Outer array tively. glj(x) and grj(x) denote the mean value and standard deviation
z1 –1 –1 –1 –1 of the jth constraint respectively, vectors xl and xr are the mean and
z2 –1 –1 1 1 standard deviation of x respectively, xL and xU are the lower and upper
x1 x2 x3 z3 –1 1 –1 1
bounds of x respectively, and g = 6 denotes the six sigma design.
–1 –1 –1 y11 y12 y13 y14 Although the robust optimization targets on controlling stan-
Inner array

–1 0 0 y21 y22 y23 y24

dard deviation, the main objective functions in Eq. (5) often involve
–1 1 1
0 –1 0 different types of mathematical components, where some require
0 0 1 minimization, some maximization, and others seek for specific
0 1 –1 objectives. For this reason, three different robust objective func-
1 –1 1 y71 y74 tions are formulated as follows:

Case 1: Minimization: aiming to minimize the response mean

by expressing the objective function as:

2.2. Multiobjective robust optimization fI ðY lI ðxÞ; Y rI ðxÞÞ ¼ kY 2lI ðxÞ þ ð1  kÞY 2rI ðxÞ ð6Þ

Unlike deterministic optimization that does not consider uncer-

tainties of design variables, the robust optimization develops a where k denotes the weight to emphasize the mean or the stan-
solution that is insensitive to variations of the nominal design dard deviation.
and is feasible in an uncertainty range around the nominal design.
As shown in Fig. 2, the x-axis represents the uncertainty parame- Case 2: Maximization: aiming to maximize the response mean
ters, including design variables and noise factors, and the vertical by expressing the objective function as:
axis represents the value of objective function f(x) to be mini-
mized. Of these three optimal solutions 1, 2, and 3 pointed, Solu- fI ðY lI ðxÞ; Y rI ðxÞÞ ¼ kY 2lI ðxÞ þ ð1  kÞY 2rI ðxÞ ð7Þ
tion 3 is considered robust as a variation of ±Dx in design
variables does not alter the objective function too much and main-
tains the solution within the design space when the design vari- Case 3: Target: on some specific response goals by expressing
ables are perturbed. Although Solution 2 is also within the design the objective function as:
space when the design variables vary in ±Dx, the perturbation
causes a larger change in objective function. Solution 1 is highly  2
fI ðY lI ðxÞ; Y rI ðxÞÞ ¼ k Y lI ðxÞ  Y lI þ ð1  kÞY 2rI ðxÞ ð8Þ
sensitive to the parameter perturbation and often cannot be rec-
ommended in practice, even though it has a better mean value 
where Y lI denotes the target for the Ith objective mean.
than Solutions 1 and 2. It is noted that there are more robust solu-
tions than solution 3 on the right of solution 3, but the objective
2.3. Particle swarm optimization algorithm
performance must be compromised. Therefore, the robustness
and objective performance should be taken into account simulta-
Inspired by the emergent motion of a flock of birds and fish
neously in the mathematical formulation.
searching for food, the particle swarm optimization (PSO) method
To measure performance variability, the six sigma (i.e. ±6r, the
was introduced by Kennedy and Eberhart [35]. Due to its simplicity
term ‘‘sigma” r denotes the standard deviation) criterion is consid-
in implementation and high computational efficiency, it has been
ered adequate, which represents the level of defects per million
used in a range of applications, e.g. neutral network, pattern recog-
parts is 0.002 for not taking into account the shift and 3.4 for taking
nition, fuzzy system control and other fields [36–41]. The PSO
into account the shift [34]. Both are fairly acceptable in engineering
method has been shown in certain instances to outperform other
practice and thus adopted as a robust design criterion in this paper.
stochastic optimization method like genetic algorithm (GA) [42].
Mathematically, a multiobjective robust optimization problem
For this reason, it was chosen to design geometric parameters of
can be formulated as:
drawbead of automobile inner panel in this paper.
8
< min ff1 ðY l1 ðxÞ; Y r1 ðxÞÞ; f2 ðY l2 ðxÞ; Y r2 ðxÞÞ;    ; fk ðY lk ðxÞ; Y rk ðxÞÞg
> In PSO, each particle represents a candidate solution associated
s:t: g lj ðxÞ þ gg rj ðxÞ 6 0 with two vectors: position (Xi) and velocity (Vi). In a b-dimensional
>
: search space, the position of particle i at iteration t can be repre-
xL þ gxr 6 xl 6 xU  gxr
sented as Xti ¼ ðxti1 ; xti2 ;    ; xtib Þ, the velocity of particle i at iteration
ð5Þ t can be described as Vti ¼ ðv ti1 ; v ti2 ;    ; v tib Þ. Let Pti ¼ ðpti1 ; pti2 ;    ; ptib Þ
denote personal best (pbest), which is the best solution that parti-
cle i has obtained until iteration t; and Ptg ¼ ðptg1 ; ptg2 ;    ; ptgb Þ repre-
± Δx ±Δ
± Δx constraint sent the global best (gbest), which is the best solution obtained
Function f(x)

from Pti in the population at iteration t. To search for an optimal

±
±ΔΔx
solution, each particle updates its velocity and position according
to the following equation:
Function minimum (
Vtþ1 ¼ wVti þ c1 v1 ðPti  Xti Þ þ c2 v2 ðPtg  Xti Þ
Δ f1 Reliability solution
i
ð9Þ
Δ 3 Robust solution
Xtþ1
i ¼ Xti þ Vtþ1
i
Δ f3
1 2 where c1 and c2 are the acceleration factors, v1 and v2 are the uni-
form random numbers in the range (0, 1), w is the inertia weight
x
that controls the influence of previous velocity in the new velocity.
Fig. 2. Schematic of robust optimization design. The flowchart of PSO algorithm is shown in Fig. 3.
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929 1921

Specify the parameters for PSO

Randomly initialize population

positions and velocities
Produce next swarm of particles

For every particle

Evaluate the fitness value Next particle

If particle fitness >particle pbest fitness,

update pbest Update particle position

If particle fitness >particle gbest fitness,

Update particle velocity
update gbest

No
Termination?

Yes

Output the optimized results

Fig. 3. The flowchart of PSO algorithm.

In this paper, fracture and wrinkling defects to be addressed

would conflict with each other in sheet metal forming process. In
other words, to some stage, any further improvement in fracture
may worsen wrinkling. A multiobjective optimization appears
essential for dealing with such a problem. As an extension to
PSO, the multiobjective particle swarm optimization incorporating
the mechanism of crowding distance computation (MOPSO-CD) Fig. 4. The flowchart of MOPSO-CD algorithm.
has drawn some attention recently as it exhibits a relatively fast
convergence and well-distributed Pareto front compared with
sented by Stoughton [47] proved fairly satisfactory. This model is
other multiobjective optimization algorithms, such as NSGA, PEAS
based upon the principle of virtual work, where the work required
[43,44]. Therefore, MOPSO-CD is used to obtain the robust draw-
to pull (or to restrain) a sheet through a drawbead is equal to the
bead restraint forces in the paper. A flowchart showing the proce-
work required to overcome the sheet/tool friction as well as bend
dure of MOPSO-CD is provided in Fig. 4. The details of MOPSO-CD
and unbend the sheet. The virtual plastic work due to the bend-
can be consulted in Ref. [43].
ing-unbending gives the restraining force Fi (per unit of the draw-
bead length) at each point i (i = 1, 2, 3, 4, 5, 6) with change in radius
3. Results and discussion (Fig. 5) as:
Z Z !
þti =2 eþi
3.1. Drawbead restraining force Fi ¼ rde dz ð10Þ
t i =2 ei
At present, there are some difficulties in simulating the effects
of blank on flowing by true drawbead. Firstly, drawbead size is where ti is the thickness of the blank, e and r are the strain and
þ
very small and shape can be very complex compared with other stress along the pulling direction, e
i and ei denote the strains be-

parts. To precisely calculate the contact force between blank and fore and after the ith radius change (Fig. 5).
drawbead, the drawbead must be discretized in very fine elements. Therefore, the total drawbead restraining force including fric-
This may not only result in low computational efficiency, but also tion can be obtained as follows:
cause some numerical problems. Secondly, modifications on the
F ¼ ½ðF 1 eju þ jHe þ F 2 þ F 3 Þe2ju þ jHe þ F 4 þ F 5 eju þ F 6 ð11Þ
drawbead need to redesign addendum of die and reconstruct finite
element model, which can be time consuming and may be incon- where j is the friction coefficient, u is the angle of contact, He is the
venient to incorporate with optimization algorithm. Thirdly, if geo- elastic part of the holding force required to close the drawbead. He
metric parameters of drawbead are directly taken as design can be obtained by the classical beam theory:
variables, the dimensions of optimization problem could become
too high. To address these issues, an equivalent drawbead restrain- 16Ebdt 3
He ¼ ð12Þ
ing force approach is adopted to replace modeling the true draw- ðr1 þ 2r 2 þ r 3 þ g 1 þ g 2 Þ3
bead details. Researches have been conducted on this subject to
find empirical or analytical formulae for modeling drawbead where d = min [h, (r1 + 2r2 + r3 + g1 + g2)(r2ry/Et)], h is the height of
restraining forces [45,46]. Among these, the drawbead model pre- the drawbead, ry is the yield point, g is the gap between drawbead
1922 G. Sun et al. / Materials and Design 31 (2010) 1917–1929

ments and their restraining forces are denoted as Db1, Db2, Db3,
respectively. The location of each segment is expressed in Fig. 8
with curved lines on the die. It is assumed that the restraining
s forces are evenly allocated in each segment of the drawbead line.
g b Incremental FEA is used to verify the formability of the die pre-
designed. The material properties of blank material are obtained
rb from a number of tensile tests. The initial parameters include
t0 = 1.2 mm, Db1 = 100 N/mm, Db2 = 160 N/mm, Db3 = 180 N/mm,
rg ϕ Young’s modulus E = 207GPa, Poisson’s ratio m = 0.28, friction coef-
h
rg ficient j = 0.15, Lankford coefficient b = 2.12, holder force
H = 100kN, initial yield stress r0 = 133.30 MPa, strength coefficient
K = 542.40 MPa, strain hardening exponent n = 0.19, the Swift-
Krupkowsky hardening law describes the material flow stress, as

Fig. 5. Drawbead geometry.

r ¼ Kð0:0042 þ ep Þn ð14Þ
Fig. 9 depicts the convergence test of the internal energy versus
punch distance in different element sizes, which shows that there
and the groove, E is the Young’s modulus and r1, r2, r3 are the radii in
is no difference between element sizes 1 mm  1 mm and
the zones 1–2, 3–4 and 5–6, as in Fig. 5, respectively.
3 mm  3 mm. Therefore, element size 3 mm  3 mm is adopted
in the finite element model.
3.2. Identification of geometrical parameters of drawbead
The initial deep drawing analysis indicates that the elements in
zone A tend to fracture while elements in zone B tend to wrinkle
The optimization results of the first stage are the drawbead
(as shown in Fig. 10). Therefore, drawbead restraining forces need
restraining forces. However, design engineers are concerned with
to be designed to decrease or eliminate these potential defects.
the shape of the drawbead (as in Fig. 5). Although the formulae
Since traditional optimization does not consider the randomness
by Stoughton [47] can present the exclusive restraining forces from
of parameters, hence the results often either violate the constraints
the known drawbead geometry and sheet property, a restraining
or are very sensitive to parametric variations. In order to overcome
force may correspond to different drawbeads. Therefore, the iden-
this drawback, the multiobjective robust optimization method is
tification of drawbead shape based upon the given restraining
adopted to design drawbead restraining forces herein. The flow-
force presents the second optimization problem in this study.
The shape of drawbead is described geometrically by the radius
of bead rb, the radius of shoulder rg, the depth h and the bead gap g
(Fig. 5). Our aim is to find the appropriate geometry of the draw-
bead, whose restraining force is equal to the optimal force gener-
ated from the previous multiobjective robust optimization
procedure (i.e. Fdesign = Fopt). The sheet will undergo thinning when
it flows through the drawbead. To limit thinning, the optimization
problem is formulated as follows:
8
> P
6
2
>
< min J ¼ ðti ðxð2Þ Þ  t0 Þ
>
i¼1
ð13Þ
>
> s:t: F design ¼ F opt Fig. 6. The CAD model of the automobile inner panel.
>
:
xLð2Þ 6 xð2Þ 6 xUð2Þ

where ti is the thickness of the sheet inside the drawbead at posi-

tion i (Fig. 5) and t0 is the initial thickness of the sheet. x(2) is the
vector of design variables in Stage two, which represent the geo-
metrical parameters of drawbead (Fig. 5), xLð2Þ and xUð2Þ denote the
lower and upper bounds of design variable x(2), respectively. In this
paper, we adopt the single-objective PSO to solve Eq. (13).

3.3. Description of demonstrative example

To demonstrate the present optimization algorithm for a practi- Fig. 7. The addendum surface of the die.
cal problem with industry interest, the design of automobile inner
panel forming is taken into account in this paper, as shown in
Fig. 6, which is of certain geometric complexity and whose form-
ability is critical yet problematical. Although the parameters of
the addendum surface have a great influence on the material flow,
it is very difficult and time consuming to adjust addendum surface
to improve sheet metal formability. Fortunately, the introduction
of drawbead can facilitate the control of the rate of metal flow in
the forming process. According to engineering experience, the
addendum surface of the die is predesigned, as shown in Fig. 7.
In fact, the drawbead restraining force can be different in different
areas. For this reason, the drawbead line is divided into three seg- Fig. 8. The layout of the drawbead.
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929 1923

3.4. Definition of objective functions

For automotive panel forming process, any rupture and/or wrin-

kling defects should be avoided. To evaluate the forming quality in
the sheet metal forming process, it is critical to reflect such
requirements using the mathematical function that can be deter-
mined from the simulation results. In this paper, the forming limit
diagram (FLD) is used for its proven ability to predict the rupture
and wrinkling [29] .
The forming limit curve (FLC) plotted in the principal strain (e1,
e2) plane was determined by either experiment or numerical stud-
ies for each material. In order to evaluate the crack, firstly, the FLC
Fig. 9. Internal energy versus punch distance. is expressed by an explicit function u0(e2) in terms of polynomial
approximation defined by the following forms [10]:
chart shown in Fig. 11 describes how to optimize the geometrical (
parameters of drawbead by minimizing the fracture and wrinkling.
a0 þ a1 e2 þ    þ an1 en21 e2 6 0
u0 ðe2 Þ ¼ ð15Þ
b0 þ b1 e2 þ    þ bm1 em
2
1
e2 > 0

Fig. 10. Before optimization with initial parameters.

Fig. 11. Flowchart of robust optimal geometrical parameters of drawbead.

1924 G. Sun et al. / Materials and Design 31 (2010) 1917–1929

where n1 and m1 are the polynomial orders, e2 is the minor principal
strain. Based upon this function, a so-called ‘‘secure FLC” function
u(e2) is then defined as:
uðe2 Þ ¼ u0 ðe2 Þ  s
where s represents a ‘‘safety” margin from FLC, which can be a con-
stant chosen by the user.
However, FLC is difficult to measure the wrinkling [48,49]. In
this paper, the wrinkling limit curve (WLC), in which the material
ð16Þ
is in a pure shear strain state, is defined by the function w(e2) as
follows [50]:

wðe2 Þ ¼ e2 e2 6 0 ð17Þ

Similarly, a so-called ‘‘secure WLC” function g(e2) is also defined
as:

gðe2 Þ ¼  tanð450 þ hÞe2 ð18Þ

where h is a user chosen ‘‘safety” marginal angle in WLC. Thus the

area between u(e2) and g(e2) indicates a safety zone (Fig. 12). When
the major principal strain e1 locates in the safety zone, the rupture
and wrinkling values go to zero. When e1 is beyond the safety zone,
there is a risk of rupture and/or wrinkling. In general, the further the
distance from the safety zone, the higher the risk of wrinkling and/
or rupture. In order to quantify different extent of crack and wrin-
kling in terms of the distance to the safety zone in the forming limit
diagram, an exponential weight criterion is presented. Therefore,
Fig. 12. Definition of the objective function. the rupture and wrinkling objective functions can be formulated,
respectively, as:
8
N 
P   
Table 2 >
>
Noise factors and their level. < Robj ¼
> ee1  uðee2 Þ exp ee1  uðee2 Þ ee1 P uðee2 Þ
e¼1
ð19Þ
Level A B C D E F >
> N 
P   
>
: W obj ¼ gðee2 Þ  ee1 exp gðee2 Þ  ee1 ee1 6 gðe22 Þ
E (GPa) m r0 (MPa) K (MPa) b j
e¼1
1 206.00 0.26 123.30 532.40 2.02 0.13
2 207.00 0.28 133.30 542.40 2.12 0.15 where N is the total number of elements. It is noted again that (e1,
3 208.00 0.30 143.30 553.40 2.22 0.17 e2) is the functions of design variable vector x(1) = (Db1, Db2, Db3)T
and can be determined via FEA.

Table 3
Experimental design using L18 orthogonal array and experimental results.

Run Noise factors Objective functions

E m r0 K b j Robj Wobj
1 1 1 1 1 1 1 6.08 7.03
2 1 2 2 2 2 2 6.38 6.44
3 1 3 3 3 3 3 7.53 5.96
4 2 1 1 2 2 3 7.27 6.13
5 2 2 2 3 3 1 6.00 7.17
6 2 3 3 1 1 2 7.33 6.04
7 3 1 2 1 3 2 6.05 6.72
8 3 2 3 2 1 3 9.27 5.83
9 3 3 1 3 2 1 6.02 6.75
10 1 1 3 3 2 2 6.25 6.61
11 1 2 1 1 3 3 6.86 6.15
12 1 3 2 2 1 1 6.19 6.61
13 2 1 2 3 1 3 8.39 6.04
14 2 2 3 1 2 1 6.00 6.98
15 2 3 1 2 3 2 6.20 6.45
16 3 1 3 2 3 1 6.00 7.28
17 3 2 1 3 1 2 7.03 6.25
18 3 3 2 1 2 3 9.09 5.77
Robj K1 39.29 40.04 39.46 41.40 44.29 36.29
K2 41.19 41.54 42.10 41.31 41.01 39.24
K3 43.46 42.36 42.38 41.22 38.64 48.41
k1 6.55 6.67 6.58 6.90 7.38 6.05
k2 6.87 6.92 7.02 6.89 6.84 6.54
k3 7.24 7.06 7.06 6.87 6.44 8.07
R 0.69 0.39 0.48 0.03 0.94 2.02
Wobj K1 38.80 39.81 38.76 38.69 37.80 41.82
K2 38.81 38.82 38.75 38.74 38.68 38.51
K3 38.60 37.58 38.70 38.78 39.73 35.88
k1 6.47 6.64 6.46 6.45 6.30 6.97
k2 6.47 6.47 6.46 6.46 6.45 6.42
k3 6.43 6.26 6.45 6.46 6.62 5.98
R 0.04 0.38 0.01 0.01 0.32 0.99
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929 1925

3.5. Screening noise factors
In the sheet metal forming process, the dimension of addendum
surface and drawbead geometry can be defined in accordance with
the design requirements. Manufacturing errors can be reduced by
using fine machining/grinding processes and the holder forces
can be adjusted easily by a hydraulic system. So these factors are
taken as control factors. However, material properties, including
Young’s modulus E, Poisson’s ratio m, Lankford coefficient b, initial
yield stress r0 and strength coefficient K, can be affected by rolling
process. Friction coefficient j is closely related to lubricate condi-
tion, surface roughness of blank and die and velocity of punch
movement. Actually, these factors are uncontrollable and may be
taken as noise factors.
Orthogonal arrays (OA) [51] are widely used in industrial exper-
iments to screen important factors from a large number of poten-
tial factors. In the present study, L18 was applied to screen the
significant noise factors according to the preliminary experiments.
Tables 2 and 3 present the factors and their levels, and the exper-
imental orthogonal design, respectively, where the effects of these
factors on Robj and Wobj were calculated as in Table 3 and Fig. 13. As
a result, the order of effects of factors on Robj was:
j > b > E > r0 > m > K; and the order of effects of factors on Wobj
was: j > m > b > E > r0 > K.
On the basis of the abovementioned analysis, it is easy to ob-
serve that the variations of noise factors j, m and b are relatively
more sensitive to objective functions Robj and Wobj. If the effects
Robj Wobj
of these noise factors are ignored, the accuracy of results would
be more significantly compromised.
3.6. Multiobjective robust optimization model
The ranges of design variables x(1) = (Db1, Db2, Db3)T in Stage one
are [80, 160], [80, 160] and [120, 200], respectively. The ranges of
noise factors, j, m and b are [0.13, 0.17], [0.26, 0.30] and
[2.02, 2.22], respectively. The noise factors sampled with OLHS
are arranged in an outer array with the sample points of 4, and
the control factors sampled with OLHS are arranged in an inner ar-
ray with the sample points of 15. Experiments in the inner array
are repeated at four points corresponding to the outer array to sim-
ulate the variability due to the uncertainties of these three noise
factors, totaling number of simulations are 60, as in Table 4.
Based upon the results in Table 4, the dual response surface
models of fracture Robj and wrinkling Wobj are formulated as in
Appendix A. The error measures of X2, X2adj and max(RE) are used
to evaluate the accuracy of these dual RS models. The max(RE) is
obtained by randomly generating five other sampling points in
the design space and then the maximum REs are evaluated. The er-
ror results are listed in Table 5, which showed the adequate accu-
racies of the dual response models that allow us to carry on the
design optimization.
After introducing the six sigma scheme, the multiobjective ro-
bust optimization is thus specificially formulated as:
8
>
> min : ðR; WÞ
>
>
>
< R ¼ kR2l þ ð1  kÞR2r
ð20Þ
>
>
> W ¼ kW 2l þ ð1  kÞW 2r
>
>
: L U
s:t: Dbi þ 6Dbri 6 Dbli 6 Dbi  6Dbri ;

Table 5
Error analysis of dual response surface models.

Rl (%) Rr (%) Wl (%) Wr (%)

X2 99.80 99.90 99.43 99.23
X2adj 98.66 97.55 98.42 94.60
max(RE) 1.07 5.43 1.32 3.21
Fig. 13. Plots of noise factor effects on Robj and Wobj.

Table 4
The arrangement of cross product arrays and FEM simulation results.

Design variables Noise factors Mean and standard deviation

j 0.13 0.14 0.16 0.17
m 0.27 0.30 0.26 0.29
b 2.15 2.09 2.02 2.22
No. 1 2 3 4
No. Db1 Db2 Db3 Robj Wobj Robj Wobj Robj Wobj Robj Wobj Rl Rr Wl Wr
1 80.00 142.86 125.71 6.36 8.79 6.33 8.63 7.29 7.92 7.46 7.20 6.86 0.60 8.14 0.73
2 85.71 108.57 177.14 6.00 8.72 6.00 8.75 6.00 7.89 6.15 7.13 6.04 0.08 8.12 0.77
3 91.43 85.71 148.57 6.00 9.35 6.00 9.24 6.00 8.16 6.04 7.60 6.01 0.02 8.59 0.85
4 97.14 160.00 182.86 7.27 7.19 7.57 7.17 8.43 6.48 8.45 5.66 7.93 0.60 6.63 0.72
5 102.86 114.29 131.43 6.00 8.07 6.00 8.13 6.00 7.11 6.26 6.39 6.07 0.13 7.43 0.83
6 108.57 131.43 194.29 6.00 6.91 6.00 7.09 6.28 6.21 6.75 5.63 6.26 0.35 6.46 0.67
7 114.29 102.86 188.57 6.08 7.43 6.00 7.49 6.08 6.87 6.33 6.25 6.12 0.14 7.01 0.58
8 120.00 148.57 137.14 6.66 6.49 6.41 6.41 7.34 5.83 7.93 5.20 7.09 0.69 5.98 0.60
9 125.71 80.00 165.71 6.13 8.22 6.02 8.21 6.30 7.37 6.90 6.70 6.34 0.39 7.63 0.73
10 131.43 125.71 154.29 6.33 6.49 6.03 6.50 6.39 5.84 7.26 5.30 6.50 0.53 6.03 0.58
11 137.14 97.14 120.00 6.37 7.54 6.21 7.48 6.76 6.75 7.47 6.07 6.70 0.56 6.96 0.69
12 142.86 154.29 160.00 7.20 5.58 6.88 5.59 8.48 5.09 9.22 4.45 7.95 1.10 5.18 0.54
13 148.57 120.00 200.00 7.61 6.15 6.72 6.19 7.27 5.56 8.13 6.19 7.43 0.59 6.02 0.31
14 154.29 137.14 142.86 7.35 5.65 6.90 5.72 8.02 5.10 8.87 4.48 7.79 0.86 5.24 0.58
15 160.00 91.43 171.43 8.46 6.81 8.79 6.77 9.14 6.12 9.89 5.55 9.07 0.61 6.31 0.60
1926 G. Sun et al. / Materials and Design 31 (2010) 1917–1929

3.7. Optimization results of Stage one

We first used MOPSO-CD to solve the deterministic multiobjec-

tive optimization problem without considering the perturbations
of design variables and noise factors with inertial weight w = 0.4,
acceleration constants c1 = c2 = 0.5, population size a = 100, and
external archive A = 50. The optimal Pareto fronts with the
30, 50, 80, 100 and 200 generations are respectively plotted in
Fig. 14, which indicates that the 100 generations converged fairly
stably and is considered adequate.
In order to take into account the perturbations of design vari-
ables, it is assumed that the design variables are normally distrib-
uted, whose standard deviations are [0.5, 0.5, 0.5]T in the
multiobjective robust optimization process. Fig. 15 presents the
Fig. 16. Pareto optimal front of standard deviation for the multiobjective robust
optimal Pareto fronts for the mean values of the deterministic mul- optimizations with k ¼ 0:01 and k ¼ 0:1.
tiobjective optimization and robust multiobjective optimizations
with k = 0.01 and k = 0.1, respectively. It is interesting to note that
the consideration of randomness of the parameters leads to sacri- Although the Pareto-set can provide designer with a large num-
fice of Pareto optimum, i.e. the robust solution is farther to the ori- ber of design solutions for their decision-make in the beginning of
gin in the objective space than the deterministic counterpart. design stage, decision must be made for the most satisfactory solu-
Fig. 16 depicts the optimal Pareto fronts of standard deviation in tion (termed as ‘‘knee point”) from Pareto-set finally. Traditionally,
the multiobjective robust optimizations with k = 0.01 and k = 0.1. the most satisfactory solution is often decided by weight method
The ranges and shape of the Pareto optimal front change with var- which aggregates many objectives into a single cost function in
iation of k as compared to the deterministic case. The smaller the k terms of weighted average to enable comparison of importance.
(i.e. the less emphasis on the mean objective), the higher the However, it can be difficult to assign proper weight factor to each
robustness of the objective functions (see Fig. 16, where the stan- objective even for an experienced engineer. In this paper, we pres-
dard deviation is lower). However, the performances of objective ent the minimum distance selection method (TMDSM) to define
functions drop (see Fig. 15, where the Pareto front moves further). the most satisfactory solution from the Pareto-set. Mathematically,
Hence, compromise must be made between robustness and nomi- the method is given as,
nal performance in practice.
( )1=d
XU
 d
min D ¼ fcs  minðfs ðxð1Þ ÞÞ ð21Þ
s¼1

where U is the number of the objective components, fcs is the sth

objective value in the cth Pareto solution, d = 2, 4, 6, . . . , D is the dis-
tance from knee point to an ‘‘utopia point” which is given by the
optimal values of each individual objective (refer to Fig. 17). In gen-
eral, such a ‘‘utopia point” is not attainable in practice with pres-
ence of conflicting objectives.
The deterministic knee point and robust knee point are ob-
tained by using TMDSM from the Pareto-sets, respectively. The re-
sults are summarized in Table 6, where a comparison is made with
the baseline model. It can be seen that both the deterministic mul-
tiobjective optimization and robust multiobjective optimization
can improve the performance of sheet metal forming. Comparing
Fig. 14. Pareto optimal front of MOPSO-CD with different generations.
xDð1Þ with xRð1Þ , the mean objective performance of xDð1Þ is better than

Fig. 15. Pareto optimal front of mean for deterministic multiobjective optimization, Fig. 17. The knee point on the Pareto front having the shortest distance from utopia
multiobjective robust optimization with k ¼ 0:01 and k ¼ 0:1. point.
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929 1927

Table 6
Comparison deterministic multiobjective optimization, robust multiobjective optimization and baseline design.

Rl Rr Wl Wr
T 7.80 0.69 6.57 0.79
Baseline design x0ð1Þ ¼ ½100; 160; 180
T 6.27 0.50 6.07 0.66
Deterministic multiobjective optimization xD
ð1Þ ¼ ½128:95; 121:55; 160
T 6.32 0.40 6.31 0.61
Robust multiobjective optimization xRð1Þ ¼ ½124:25; 122:76; 155:92

Table 7
Ranges of geometrical parameters of drawbead.

rg (mm) rb (mm) h (mm)

Upper bound 3 3 6
Lower bound 8 8 12

Table 8
Optimal parameters.

Drawbead rg (mm) rb (mm) h (mm)

1 4.9 7.0 5.1
2 4.9 6.9 5.0
3 4.3 4.9 4.9

Fig. 19. Simulation result from Autoform.

DBRF model of cylindrical drawbead has the advantage of high pre-

Fig. 18. The CAD model of addendum surface of die with true drawbeads.
that of xRð1Þ , but the robustness of xDð1Þ is worse than that of xRð1Þ ,
which appear logical.
3.8. Optimization results of Stage two
In the second stage, we determine the geometrical parameters of
the drawbeads by using the drawbead design procedure discussed
in Section 3.2. It is noted that cylindrical drawbead can provide ade-
quate restraining force that is optimized above. Meanwhile, the
cision and small number of geometrical parameters. Therefore, we
adopt cylindrical drawbead herein. The design variables of the
cylindrical drawbead are x(2) = (rb, rg, h)T, rb is the radius of the bead,
rg is the radius of the shoulder and h is the depth of drawbead (see
Fig. 3). The bead gap g = 1.1t0.
The ranges of geometrical parameters of drawbead are listed in
Table 7. The optimization is performed using the PSO algorithm.
The swarm size chosen is 40, and the PSO is run 100 generations.
The optimal geometrical parameters of drawbead are listed in Ta-
ble 8. Then we redesign addendum surface of die according to the
optimal geometrical parameters of drawbead, shown as in Fig. 18.
Autoform is used to check the formability of new design addendum
surface with true drawbead, as shown in Fig. 19. It is clear to ob-
serve that the formability quality is satisfactory from the view of
simulation analysis.
Figs. 20 and 21 are the photos of drawing mould and actually-
formed part made from the optimal drawbead geometrical param-
eters. The quality of automobile inner panel with the optimal geo-
metrical parameters of drawbead is rather satisfactory, so the
presented method proved effective.

Fig. 20. Set of drawing mould.

1928 G. Sun et al. / Materials and Design 31 (2010) 1917–1929

2
Rl ¼ 17:7225  0:6576758Db1 þ 0:0028709Db1 þ 0:4084321Db2
2 2
 0:0030654Db2 þ 0:0923776Db3 þ 6:17  104 Db3
2
 1:80  105 Db1 Db2 þ 5:53  105 Db1 Db3
2
 0:0048132Db2 Db3 þ 2:67  105 Db2 Db3
2 2
 2:35  108 Db2 Db3 þ 0:0041691Db1 Db2 ðA1Þ

2 3
Rr ¼ 14:1799  0:3245196Db1 þ 0:0027185Db1  5:94  106 Db1
2
þ 0:0356962Db2  1:83  104 Db2  0:0368768Db3
4 6 2
 9:25  10 Db1 Db2 þ 6:08  10 Db1 Db2
3 3
Fig. 21. Actually-formed workpiece.  9:12  109 Db1 Db2  3:56  109 Db1 Db2
2
þ 5:31  104 Db1 Db3  2:21  106 Db1 Db3
5
4. Conclusion þ 4:29  10 Db2 Db3 ðA2Þ

Fracture and wrinkling are predominant defects in sheet metal W l ¼ 26:4409  0:1062943Db1 þ 2:43  104 Db1
2

forming process. The existence of these defects may damage sur- 2

face quality, reduce dimensional precision, cause local crack of
component and lead directly to waster. In order to improve prod-
uct quality and reduce cost, various optimization techniques have
been successfully applied to sheet metal forming process. How-
ever, traditional optimization has focused on deterministic and/
or single robust objective. Real-life engineering problems are typi-
cally nondeterministic and come with multiple objectives. To ad-
dress such a multiobjective nondeterministic problem, this paper
presented a multiobjective robust optimization method, where
the six sigma principle is adopted to measure the variations, the
dual response surface method is used to construct the surrogate
models, and the multiobjective particle swarm optimization is
developed to generate the robust Pareto solution.
In this paper, the rupture and wrinkling criteria based on expo-
nential weight are adopted in the drawbead design for the sheet
metal forming, and a multiobjective robust optimization method-
ology is presented. The design procedure is divided into two
stages: firstly, equivalent drawbead restraining forces (DBRF) are
obtained using the developed multiobjective robust optimization
method; and then the DBRF model is integrated into the particle
swarm optimization (PSO) to optimize geometric parameters of
drawbead. To validate the design, a try-out verification of the opti-
mized die was conducted and the quality of the produced work-
piece was found rather satisfactory without any rupture and
wrinkling observed. It is thus justified that the presented method
does provide an effective solution to geometric design of drawbead
for improving product quality.
Acknowledgments
The support from National 973 Project of China (2010CB328005),
The Outstanding Youth Foundation of NSFC(50625519), The Key
Project of NSFC(60635020) and The Program for Changjiang Scholar
and Innovative Research Team in Chinese Universities are acknowl-
edged. The first author is also grateful the supports from China Schol-
arship Council and School of Aerospace, Mechanical and
Mechatronic Engineering, The University of Sydney.
Appendix A
The response surface model of rapture and wrinkling objectives
are obtained from the least square procedure as follows:
 0:0724295Db2 þ 2:02  104 Db2  0:0703924Db3
2
þ 1:59  104 Db3  2:45  105 Db1 Db2
þ 1:17  10 Db1 Db3 þ 1:56  105 Db2 Db3
4
ðA3Þ
2
W r ¼ 24:4467 þ 0:2039124Db1  0:0018535Db1
3 2
þ 5:15  106 Db1  0:1851205Db2 þ 0:0014489Db2
3 2
 4:06  106 Db2  0:4526954Db3 þ 0:0028229Db3
3
 5:97  106 Db3 þ 4:42  105 Db1 Db2
þ 2:74  105 Db1 Db3 þ 7:30  105 Db2 Db3 ðA4Þ
References
[1] Makinouchi A. Sheet metal forming simulation in industry. J Mater Process
Technol 1996;60(1–4):19–26.
[2] Panthi SK, Ramakrishnan N, Pathak KK, Chouhan JS. An analysis of springback
in sheet metal bending using finite element method (FEM). J Mater Process
Technol 2007;186(1–3):120–4.
[3] Dong HZ, Lin ZQ. Investigation of sheet metal forming by numerical simulation
and experiment. J Mater Process Technol 2000;103(3):404–10.
[4] Chen L, Yang JC, Zhang LW, Yuan SY. Finite element simulation and model
optimization of blankholder gap and shell element type in the stamping of a
washing-trough. J Mater Process Technol 2007;182(1–3):637–43.
[5] Ohata T, Nakamura Y, Katayama T, Nakamachi E, Nakano K. Development of
optimum process design system by numerical simulation. J Mater Process
Technol 1996;60(1–4):543–8.
[6] Naceur H, Guo YQ, Batoz JL, Knopf-Lenoir C. Optimization of drawbead
restraining forces and drawbead design in sheet metal forming process. Int. J.
Mech. Sci. 2001;43(10):2407–34.
[7] Azaouzi M, Naceur H, Delameziere A, Batoz JL, Belouettar S. An Heuristic
optimization algorithm for the blank shape design of high precision metallic
parts obtained by a particular stamping process. Finite Elem Anal Des
2008;44(14):842–50.
[8] Guo YQ, Batoz JL, Naceur H, Bouabdallah S, Mercier F, Barlet O. Recent
developments on the analysis and optimum design of sheet metal forming
parts using a simplified inverse approach. Comput Struct 2000;78:138–48.
[9] Liao XT, Li Q, Yang XJ, Zhang WG, Li W. Multiobjective optimization for crash
safety design of vehicles using stepwise regression model. Struct
Multidisciplinary Optim 2008;35(6):561–9.
[10] Chengzhi S, Guanlong C, Zhongqin L. Determining the optimum variable blank-
holder forces using adaptive response surface methodology (ARSM). Int J Adv
Manuf Technol 2005;26(1–2):23–9.
[11] Naceur H, Ben-Elechi S, Batoz JL, Knopf-Lenoir C. Response surface
methodology for the rapid design of aluminum sheet metal forming
parameters. Mater Des 2008;29:781–90.
[12] Kok S, Stander N. Optimization of a sheet metal forming process using
successive multipoint approximations. Struct Optim 1999;18(4):277–95.
[13] Kayabasi O, Ekici B. Automated design methodology for automobile side panel
die using an effective optimization approach. Mater Des
2007;28(10):2665–72.
[14] Huang Y, Lo ZY, Du R. Minimization of the thickness variation in multi-step
sheet metal stamping. J Mater Process Technol 2006;177(1–3):84–6.
[15] Ohata T, Nakamura Y, Katayama T, Nakamachi E. Development of optimum
process design system for sheet fabrication using response surface method. J
Mater Process Technol 2003;143(Sp. Iss. SI):667–72.
 G. Sun et al. / Materials and Design 31 (2010) 1917–1929
[16] Hu W, Yao LG, Hua ZZ. Optimization of sheet metal forming processes by
adaptive response surface based on intelligent sampling method. J Mater
Process Technol 2008;197(1–3):77–88.
[17] Breitkopf P, Naceur H, Rassineux A, Villon P. Moving least squares response
surface approximation: Formulation and metal forming applications. Comput
Struct 2005;83(17–18):1411–28.
[18] Jansson T, Andersson A, Nilsson L. Optimization of draw-in for an automotive
sheet metal part – an evaluation using surrogate models and response
surfaces. J Mater Process Technol 2005;159(3):426–34.
[19] Jansson T, Nilsson L. Optimizing sheet metal forming processes – using a
design hierarchy and response surface methodology. J Mater Process Technol
2006;178(1–3):218–33.
[20] Tang BT, Sun JX, Zhao Z, Chen J, Ruan XY. Optimization of drawbead design in
sheet forming using one step finite element method coupled with response
surface methodology. Int J Adv Manuf Technol 2006;31(3–4):225–34.
[21] Jakumeit J, Herdy M, Nitsche M. Parameter optimization of the sheet metal
forming process using an iterative parallel Kriging algorithm. Struct
Multidisciplinary Optim 2005;29(6):498–507.
[22] Mourelatos ZP, Liang JH. A methodology for trading-off performance and
robustness under uncertainty. J Mech Des 2006;128(4):856–63.
[23] Li YQ, Cui ZS, Ruan XY, Zhang DJ. Application of six sigma robust optimization
in sheet metal forming. In: Proceedings of the 6th international conference and
workshop on numerical simulation of 3D sheet metal forming processes,
Detroit (MI, USA); 2005.
[24] Li YQ, Cui ZS, Ruan XY, Zhang DJ. CAE-based six sigma robust optimization for
deep-drawing process of sheet metal. Int J Adv Manuf Technol 2006;30(7–
8):631–7.
[25] Li YQ, Cui ZS, Zhang DJ, Ruan XY, Chen J. Six sigma optimization in sheet metal
forming based on dual response surface model Chinese. J Mech Eng (English
Edition) 2006;19(2):251–5.
[26] Demir I, Kayabasi O, Ekici B. Probabilistic design of sheet-metal die by finite
element method. Mater Des 2008;29(3):721–7.
[27] Zhang DJ, Cui ZS, Ruan XY, Li YQ. Sheet springback prediction based on non-
linear combined hardening rule and Barlat89’s yielding function. Comput
Mater Sci 2006;38(2):256–62.
[28] Wang H, Li EY, Li GY. Optimization of drawbead design in sheet metal forming
based on intelligent sampling by using response surface methodology. J Mater
Process Technol 2008;206(1–3):45–55.
[29] Naceur H, Delameziere A, Batoz JL, Guo YQ, Knopf-Lenoir C. Some
improvements on the optimum process design in deep drawing using the
inverse approach. J Mater Process Technol 2004;146(2):250–62.
[30] Yeniay O, Unal R, Lepsch RA. Using dual response surfaces to reduce variability
in launch vehicle design: a case study. Reliab Eng Syst Safety
2006;91(4):407–12.
[31] Hou SJ, Li Q, Long SY, Yang XJ, Li W. Design optimization of regular hexagonal
thin-walled columns with crashworthiness criteria. Finite Elem Anal Des
2007;43(6–7):555–65.
[32] Yang RJ, Wang N, Tho CH, Bobineau JP. Metamodeling development for vehicle
frontal impact simulation. J Mech Des 2005;127(5):1014–20.
1929
[33] Hou SJ, Li Q, Long SY, Yanga XJ, Li W. Multiobjective optimization of multi-cell
sections for the crashworthiness design. Int J Impact Eng
2008;35(11):1355–67.
[34] Koch PN, Yang RJ, Gu L. Design for six sigma through robust optimization.
Struct Multidisciplinary Optim 2004;26(3–4):235–48.
[35] Eberhart R, Kennedy J. A new optimizer using particle swarm theory. In:
Proceedings of the sixth international symposium on micro machine and
human science, Nagoya (Japan); 1995, p. 39–43.
[36] Fourie PC, Groenwold AA. The particle swarm optimization algorithm in size
and shape optimization. Struct Multidisciplinary Optim 2002;23(4):259–67.
[37] Gaing ZL. A particle swarm optimization approach for optimum design of PID
controller in AVR system. IEEE Trans Energy Convers 2004;19(2):384–91.
[38] Parsopoulos KE, Vrahatis MN. On the computation of all global minimizers
through particle swarm optimization. IEEE Trans Evol Comput
2004;8(3):211–24.
[39] Robinson J, Rahmat-Samii Y. Particle swarm optimization in electromagnetics.
IEEE Trans Antennas Propag 2004;52(2):397–407.
[40] Salman A, Ahmad I, Al-Madani S. Particle swarm optimization for task
assignment problem. Microprocess Microsyst 2002;26(8):363–71.
[41] Yoshida H, Kawata K, Fukuyama Y, Takayama S, Nakanishi Y. A particle swarm
optimization for reactive power and voltage control considering voltage
security assessment. IEEE Trans Power Syst 2000;15(4):1232–9.
[42] Lee KC, Jhang JY. Application of particle swarm algorithm to the optimization
of unequally spaced antenna arrays. J Electromagnet Waves Appl
2006;20(14):2001–12.
[43] Raquel C, Naval P. An effective use of crowding distance in multiobjective
particle swarm optimization. In: Proceedings of the 2005 conference on
genetic and evolutionary computation, Washington (DC, USA); 2005.
[44] Liu DS, Tan KC, Goh CK, Ho WK. A multiobjective memetic algorithm based on
particle swarm optimization. IEEE Trans Syst Man Cybernet Part B – Cybernet
2007;37(1):42–50.
[45] Wang NM. A mathematical model of drawbead forces in sheet metal forming. J
Appl Metal Work 1982;2(3):193–9.
[46] Levy BS. Development of a predictive model for draw bead restraining force
utilizing work of Nine and Wang. J Appl Metal Work 1983;3(1):38–44.
[47] Stoughton TB. Model of drawbead forces in sheet metal forming. In:
Proceedings of the 15th biennial IDDRG congress, Dearborn (USA); 1998.
[48] Obermeyer EJ, Majlessi SA. A review of recent advances is the application of
blank-holder force towards improving the forming limits of sheet metal parts.
J Mater Process Technol 1998;75(1–3):222–34.
[49] Szacinski AM, Thomson PF. Investigation of the existence of a wrinkling-limit
curve in plastically-deforming metal sheet. J Mater Process Technol
1991;25(2):125–37.
[50] Sun CZ, Chen GL, Lin ZQ, Zhao YX. Novel algorithm for determining optimal
blankholder forces in deep drawing of aluminum alloy sheet. Trans Nonferrous
Metals Soc China 2004;14(4):675–80.
[51] Mohan LV, Shunmugam MS. An orthogonal array based optimization
algorithm for computer-aided measurement of worm surface. Int J Adv
Manuf Technol 2006;30(5–6):434–43.