Computer-Aided Design 92 (2017) 42–57

Contents lists available at ScienceDirect

Computer-Aided Design
journal homepage: www.elsevier.com/locate/cad

A robust and accurate geometric model for automated design of

drawbeads in sheet metal forming✩
Zhen Wang, Qiuchong Zhang, Yuqi Liu, Zhibing Zhang *
State Key Laboratory of Materials Processing and Die & Mould Technology, Huazhong University of Science and Technology, Wuhan, 430074,
People’s Republic of China

article info a b s t r a c t
Article history: A robust and accurate geometric model of real drawbeads that can be used for the automated design
Received 1 January 2017 of drawbeads is presented in the paper. A three-dimensional geometric drawbead is a lofted surface, of
Accepted 23 July 2017 which the section curves are constructed parallel to the stamping direction on the control points. Adaptive
control point interpolation is introduced to simplify the management of the drawbead geometry and avoid
Keywords:
unexpected shapes. Given primitive control points on a drawbead curve, dominant control points are
Lofted surface
adaptively obtained with the shapes of both the drawbead curve and the binder considered. An a priori
Drawbead
Intelligent geometry optimization heuristic parameter adjustment strategy is proposed to correct the parameter errors of section curves,
Automated die design which improves the accuracy and consistency of the drawbead geometry. By incorporating the proposed
Sheet metal forming geometric drawbead with a previously developed intelligent drawbead optimization algorithm, a fully
automated design process for drawbeads is realized that includes geometric modeling, finite element
analysis, intelligent optimization of the drawbead geometry, and die manufacturing. Finally, a fender
example is presented to verify the feasibility and validity of the fully automated drawbead design process.
The simulation results with the optimized geometric drawbeads and equivalent drawbeads are compared
with the experimental results. The proposed geometric drawbead shows remarkable practicability and ac-
curacy in the automated design of drawbeads in sheet metal forming and demonstrates good consistency
with the experimental results while the equivalent drawbead model introduces unneglectable deviations.
© 2017 Elsevier Ltd. All rights reserved.

1. Introduction In recent years, significant developments have been made in

Drawbeads are rib-like projections mounted on binder surfaces
that restrict and control the material flow in the sheet metal
forming process [1]. Drawbeads force the sheet metal to bend
and unbend before entering the die cavity. This action creates a
restraining force on the sheet metal, which causes the material
to enter the die cavity at both a reduced rate and a reduced vol-
ume [2]. The drawbeads are one of the most important parameters
to control the material flow and the part quality in the sheet
forming process. Powerful restraining forces prevent the sheet
from drawing-in and may cause necking, but insufficient forces
may lead to wrinkling [3]. Restraining forces of the drawbeads are
mainly related to the geometry of the drawbeads, so the quality of
sheet metal forming can be improved by changing the shape, size,
and location of the drawbeads [4].
✩ This paper has been recommended for acceptance by Panagiotis Dimitriou
Kaklis.
* Corresponding author.
E-mail addresses: wzyuchen@hust.edu.cn (Z. Wang), zhangqiuchong@sina.com
(Q. Zhang), liuyq@mail.hust.edu.cn (Y. Liu), zhangzb@mail.hust.edu.cn (Z. Zhang).
sheet forming finite element simulation. Finite element analysis
(FEA) is extensively used for drawbead design in sheet metal form-
ing. However, drawbead design is still a highly-iterative process
and requires repeated manual adjustment of the geometric param-
eters. To avoid the disadvantages of the trial and error process,
such as low efficiency and dependence on operator experience,
FEA combined with optimization methods is often used to perform
automated drawbead design. Guo et al. [5] developed a numer-
ical procedure based on the combination of a simplified finite
element method called inverse approach (IA) and a sequential
quadratic programming method to perform shape optimization
of blank contours. Naceur et al. [3] incorporated a mathematical
programming algorithm with the IA to optimize the restraining
forces and then design the drawbead. Liu et al. [6] combined
an improved hybrid optimization algorithm with FEA software
to carry out the optimum design of drawbeads for sheet metal
forming. An analytical equivalent model considering anisotropy
was introduced to determine the actual type and geometric pa-
rameters of drawbeads. Song et al. [7] presented the optimization
design of drawbead restraining forces (DBRF) using the response
surface method in order to reduce springback in the deep stamping

http://dx.doi.org/10.1016/j.cad.2017.07.004
0010-4485/© 2017 Elsevier Ltd. All rights reserved.
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 1. Wrinkles due to lack of blank holder force: (a) the experimental result of
the roof, (b) the simulation of A-region with real drawbeads, (c) the simulation of
A-region with equivalent drawbeads.
process. Liu et al. [8] built an automated optimization integration
system of the process parameters in sheet metal forming, which
gains a seamless interaction between the CAD and CAE software.
Sun et al. [9,10] constructed different surrogate models to improve
formability and shorten the design cycle of drawbeads combined
with FEA, in which the optimal solution of drawbead restraining
forces is found by multiobjective particle swarm optimization and
artificial bee colony algorithm.
However, most of the previous studies are not based on a real
drawbead model but an equivalent drawbead model. These have
the following deficiencies: (1) The equivalent drawbead model
generally only considers the deformation pattern under the con-
dition of plane strain while the sheet metal flows through the
drawbeads. Nevertheless, in the forming process of an automotive
covering panel, the material flow is very complicated, especially
at some corners, where non-plane strain deformations take place.
In addition, the variation of sheet thickness and stress state in
the forming process also leads to low precision of the equivalent
drawbead model [11,12]. (2) The inaccurate calculation of the
uplift force generated by drawbeads on the upper binder further
lowers the precision of simulation analysis. Fig. 1 presents the
experimental forming result and the simulation analyses of a top
roof. As shown in Fig. 1(a), the sheet metal is severely wrinkled
before entering the die groove due to lack of blank holder force
and wrinkles stretch into the side wall. The simulation analysis
with real drawbeads in Fig. 1(b) veritably reflects the defect. How-
ever, the analysis with equivalent drawbeads in Fig. 1(c) cannot
simulate the wrinkle phenomenon because the drawbead counter-
force calculated by the equivalent model is relatively small.
(3) The equivalent drawbead model is also unable to reflect the ac-
tual contact status between the sheet metal and drawbeads. Fig. 2
displays a sheet metal forming simulation with two identical rect-
angular drawbeads. The bending effects of the two drawbeads on
the sheet metal have a great difference with each other. However,
the equivalent drawbead model will provide the same restraining
force on the sheet metal that flows through the drawbeads. (4) The
real drawbead is also a requirement of downstream processes such
as CNC manufacturing, while the equivalent drawbead model is in
opposition to the unified management of the die model.
Compared with the equivalent drawbead model, the sheet
forming simulation with the real drawbead model is closer to
43
Fig. 2. Different contact statuses of two identical real drawbeads.
an actual situation, which will represent the constraint effects of
drawbeads on the sheet metal. Because of the small radius of the
three-dimensional drawbead, which requires a large amount of
computing resources and a long computing time in the sheet metal
forming simulation, the real drawbead model has not been used
extensively in die design process previously. With advances in
computing power, however, engineers are increasingly evaluating
the use of geometric models of real drawbeads.
Although recent commercial CAD systems provide tools for geo-
metric modeling of drawbeads on curved surfaces, the tools are not
sufficiently robust and accurate for practical applications. Siemens
NX uses an offsetting approach to model the geometric shape of
drawbeads. The quality of the surface, however, has a great in-
fluence on the offsetting approach. The offsetting command takes
a long time and may fail if the binder is not of high quality. The
drawbead function that AutoForm provides may cause undercut
shapes in the stamping direction when the normal direction of
the binder at one point is not parallel to the stamping direction.
Chung [2] presents a geometric modeling process of drawbeads
using vertical section sweeping that is fully automated and robust.
The geometric model has a no-undercut shape in the stamping
direction and easily represents the smooth ending at ends of
drawbeads. However, the sampling method of control points is not
covered (to be elaborated on in Section 3.4). In addition, although
the height error has been proven to be negligible, when the height
of the drawbead is lower than a certain mathematical expression
about the width of the drawbead, the section curve intersects with
the binder at the shoulder corner instead of the straight wall. Thus,
geometric parameter errors arise in the height, width, and center
position, which affects the accuracy and consistency of the real
drawbead (to be elaborated on in Section 3.5).
This paper presents an automated design oriented geometric
model of real drawbeads, i.e., user-defined feature (UDF) of real
drawbeads, which is based on the parameter association and UDF
technology provided by most CAD systems. Control points are
prescribed to control the distribution of the restraining force of the
drawbeads. Three-dimensional drawbeads are constructed by loft-
ing the section curves that are generated in terms of the parameters
stored in the control points. To simplify the management of the
drawbead geometry and avoid unexpected shapes, adaptive con-
trol point interpolation is introduced to interpolate new dominant
control points. Moreover, a priori heuristic parameter adjustment
is performed on each section curve to correct the parameter errors
of geometric drawbeads. Thus, the drawbeads are more accurate
and consistent.
Based on the geometric model of real drawbeads and the previ-
ously developed intelligent optimization algorithm of drawbeads,
i.e., iterative learning control (ILC) [13], a fully automated design
process for drawbeads is realized including geometric modeling
(CAD), finite element analysis (CAE), intelligent optimization of the
drawbead geometry, and CNC manufacturing (CAM). The process
improves the automation level and efficiency of the drawbead
design and decreases the dependency on engineering experience.
The remainder of this paper is organized as follows. The fully
automated design process of drawbeads and the UDF of real draw-
beads are introduced in Section 2. The geometric modeling method
of real drawbeads is described in Section 3. The experimental
44 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 3. Automated design process of drawbeads.
results verify the effectiveness of the proposed geometric model
of real drawbeads in Section 4. The conclusions are drawn in the
last section.
2. Framework of automated drawbead optimization
2.1. Process of automated drawbead design
This paper presents a fully automated design process for draw-
bead geometry. Fig. 3 shows the flow chart of the automated
drawbead design. The basic procedure is as follows:
(1) Define the initial parameters of drawbeads consisting of the
locations of control points on the drawbead curve and the
geometric parameters at each control point.
(2) Create a UDF of real drawbeads (to be elaborated on in
Section 2.2).
(3) Generate a finite element model and execute sheet metal
forming simulation with geometric drawbeads.
(4) Judge if the convergence condition is satisfied. If yes, the
optimization is terminated and the drawbeads are exported
to the CAM model for manufacturing. Otherwise, ILC is per-
formed on the CAE result to calculate new parameters for
the drawbeads and then return to Step 2.
In Step 1, the initial parameters of drawbeads are defined by
designer and may influence the number of optimization iterations
in Step 4 but not affect the convergence of the intelligent optimiza-
tion algorithm.
2.2. UDF of real drawbeads
Most commercial CAD systems offer UDF functionality. UDF is
a template of a set of features including geometric, numerical, and
constraint features. A feature is stored in the data structure of the
UDF as a property and the UDF will update as its properties change.
This paper presents the UDF of real drawbeads, of which the data
structure is shown in Fig. 4. The inputs of the UDF are binders,
drawbead curves, control points, and stamping direction. Binders
and drawbead curves are stored in the UDF as geometric features
while the control points and the stamping direction are stored
in the UDF as numerical features. A control point is not a simple
coordinate point in space but a set of the location parameter on
the drawbead curve and the geometric parameters of the section
curve at the control point. The location parameter and geometric
parameters of each control point are stored in the defined data
structure as shown in Fig. 4. To define the dependency relation
clearly, the UDF also uses linked lists to record which binder
each drawbead curve belongs to and which drawbead curve each
control point lies on. The output of the UDF is the upper and lower
binders with real drawbeads and the upper binder is a portion of
the die in Fig. 2.
The UDF of real drawbeads provides a convenient and parame-
terized approach to manage drawbead parameters and geometry.
The UDF of real drawbeads changes the location and geometric
shape of the drawbead by modifying input information, which
leads to a redistribution of the restraining force of the drawbead.
2.3. Seamless integration between CAD and CAE
Seamless integration between CAD and CAE is indispensable in
realizing automated design of drawbeads. The automated design of
drawbeads displayed in Fig. 3 is an iterative process. Massive data
conversion between CAD and CAE is repeated until the CAE result
satisfies the convergence condition. Therefore, the performance of
the automated design of drawbeads is directly influenced by the
accuracy and efficiency of the data conversion process between
CAD and CAE.
Metal sheet forming analysis (MSFA) [14], which is a simula-
tion system for the complete stamping process, is independently
developed on the CATIA platform. Based on UDF technology, MSFA
defines each of the process parameters as a UDF, just like the
original feature provided by CATIA. This seamless integration per-
mits automated updating of the CAE model when the CAD model,
e.g., the drawbead geometry in the paper, is modified. The pro-
posed UDF of real drawbeads is appended onto MSFA. In MSFA,
the UDF of real drawbeads is divided into geometric information
and numerical information. The former is represented by binders
with real drawbeads while the latter is the location and geometric
parameters of each control point. Binders with real drawbeads
are converted into a finite element mesh, which is used for sheet
forming simulation. The location and geometric parameters of the
control points, combined with CAE results, are used to calculate the
location and geometric parameters of new control points by ILC.
3. Geometric modeling of drawbead
3.1. Geometric parameters of drawbead
Fig. 5 shows the geometric parameters of the drawbead. Four
parameters including width, height, shoulder radius, and entrance
radius define the rectangular drawbead while three of these pa-
rameters, all except shoulder radius, define the circular drawbead.
The entrance radius is defined on the female bead and the shoulder
radius is defined on the male bead. The width w1 and height h1 on
the male bead are regarded as the benchmarks and the width w2
and height h2 on the female bead are determined by:
w2 = w1 + 2ts + 2cw , (1)
h2 = h1 + ts + ch , (2)
where ts is the sheet thickness and cw and ch are the width clear-
ance and height clearance between the male bead and female bead,
respectively.
As pointed out by Chung [15], the geometric parameters shown
in Fig. 5 are ambiguous on slant binders. Fig. 6 shows two different
definitions of the geometric parameters. In Fig. 6(a), height is the
length of the drawbead along the stamping direction and width is
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57 45

Fig. 4. Data structure of UDF of real drawbeads.

{ci (u), i = 1, 2 . . . n} (referred to as section curve below) as shown

in Fig. 7 [16,17]. Lofted surfaces have different representations de-
pending on the use of different blending functions. B-splines have
various useful properties for the design of complicated geometries
and have been intensively investigated for the interpolation of
lofted surfaces, so the section curves and lofted surfaces are param-
eterized with the representation of B-spline in this paper [18,19].
Mathematically, the isoparametric section curve ci (u) is expressed
as follows:
m
∑
up−1 ≤ u ≤ um+1 ,
( )
ci (u) = vi,j Mj,p (u) (3)
j=0

where vi,j are control points and Mj,p (u) are the normalized
B-spline
{ basis functions}of order p defined on a common knot
vector u0 , u1 , . . . , um+p . The detailed method for multiple curve
interpolation on a common knot vector can be found in the litera-
ture [20]. The lofted surface is represented by:
⎧ ⎫
n n ⎨ m ⎬
∑ ∑ ∑
s (u, v) = ci (u)Ni,q (v ) = vi,j Mj,p (u) Ni,q (v ), (4)
⎩ ⎭
i=0 i=0 j=0

where Ni,q (v ) are the normalized B-spline basis functions of order

Fig. 5. Geometric parameters of drawbeads: (1) rectangular drawbead, (2) circular
drawbead.
Fig. 6. Two definitions of geometric parameters on slant binders.
the groove length of the drawbead. However, in Fig. 6(b), height is
the perpendicular length of the drawbead relative to the groove
and width is the horizontal length of the drawbead relative to
the stamping direction. The second definition shown in Fig. 6(b)
is adopted in this paper but further studies may be required to
determine which definition is better.
3.2. Lofted surface
Surface lofting, also known as skinning, is a process of
passing a surface through a sequence of cross-sectional curves
q defined on a common knot vector v0 , v1 , . . . , vn+q .
{ }
In general, the planar section curves should be first brought into
3D space by the spine curve g (v ) before constructing the lofted
surface [17]. A section curve pi (u) corresponding to a point g (vi )
on the spine curve is defined on its local section plane as shown in
Fig. 7. The 3D point g (vi ) is the reference point and the origin of the
local coordinate system. A moving coordinate frame is defined at
g (vi ) as shown in Fig. 7. Let H , V , and T denote the x-axis, y-axis,
z-axis of the moving coordinate frame, respectively. T is the unit
tangent vector of g (v ) at v = vi and is obtained by:
ġ (vi )
T = . (5)
|ġ (vi )|
H and V are generally two orthogonal normal vectors of g (v ). The
3D section curve ci (u) is then expressed as a transformation of the
2D section curve pi (u):
H
[ ]
ci (u) = pi (u) V + g (vi ). (6)
T
In addition, the spine curve can be also taken as the guide
line to control the longitudinal shape of the lofted surface. It is
generally desired to interpolate the lofted surface following the
spine curve, and the surface parametrization should correspond to
the distances between adjoining section curves. The obvious choice
for the surface parametrization takes the knot vector {vi } as the
target point parameter values from the spine [16].
Different from the swept surface adopted in the Chung’s article,
the lofted surface is modeled by an interpolating function instead
46 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57

Fig. 8. Orthogonal section plane.

Fig. 7. Lofted surface.

of sweep transformation. The swept surface is not precisely rep-

resentable as a rational B-spline form in many cases while the
loft surface is a natural fit to be defined as a rational B-spline
form. Thus, surface lofting is more compatible with the geometric
modeling system. Meanwhile, the lofting process is flexible and
robust and is able to pass a surface through various complicated
shapes [21]. Surface lofting is employed to generate the drawbead
surface that passes through section curves (to be elaborated on in
Section 3.3) and is guided by the drawbead curve. In the lofted
surface, the transformation in Eq. (6) and the spine curve can
preserve the sections to be normal to the drawbead curve.

3.3. Section curve

Fig. 9. Outline of section curve.

Similar to the approach in Chung [2], this study also uses section
curves to control the drawbead shape. A section curve is located on
a normal plane at one point p on the drawbead curve. As shown in where l is the translation length. As defined in Fig. 6(b), height is
Fig. 8, the normal plane is determined by two orthogonal axes, the the perpendicular length of the drawbead relative to the groove
horizontal axis and vertical axis. The horizontal axis H is defined s1 s2 . Generally, V is not perpendicular to s1 s2 . To guarantee the
by: drawbead height as defined in Fig. 6(b) is h, l is calculated by:
T ×U
H = , (7) l = h/|H · B| , (13)
|T × U |
where U is the unit vector of the stamping direction and T is the where B is the unit vector determined by b1 and b2 and is parallel
unit tangent vector of the drawbead curve at point p. The vertical to the groove s1 s2 .
axis V is defined by: The three segments: b1 s1 , s1 s2 and b2 s2 , determined by the four
points: b1 , b2 , s1 and s2 , draw the outline of the section curve at p.
H ×T In the Chung’s article, the outline in Fig. 9 is defined as the section
V = . (8)
|H × T | curve. The shoulder corner is rounded on two sculptured surfaces
by sweeping. However, the variable fillet, which is unsteady, is
As shown in Fig. 9, the outline of the section curve is bounded required when the shoulder radius of a drawbead is not constant.
by four points: b1 , b2 , s1 and s2 . The base points b1 and b2 are In this paper, the shoulder corner is done on the section curve
determined by two steps. The control point p is first translated with in advance as shown in Fig. 10, which improves the reliability
a length of w/2 along −H and H respectively, and then projected of the drawbead generation because surface lofting is robust and
onto the binder along V , compatible with various section curves [21].
b1 = project (p − w/2H |V ) , (9)
3.4. Adaptive control point interpolation
b2 = project (p + w/2H |V ) , (10)
where w is the drawbead width at p and the projection operator From Eqs. (7) and (8), the section plane, on which the section
(∗|D) projects a point onto the binder along the vector D. curve is located, is determined by the stamping direction U and
The shoulder points s1 and s2 are obtained by translating the the tangent vector T . U is invariant, so the section plane is only in-
base points b1 and b2 along V , fluenced by T . From Eq. (13), the section shape is influenced by the
included angle of the horizontal axis H and the base vector B. The
s1 = b1 + lV , (11)
variation of T and the included angle of H and B probably leads to
s2 = b2 + lV , (12) unexpected shapes and even undercut shapes. As shown in Fig. 11,
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57 47

Fig. 10. Section curve: (a) rectangular section curve, (b) circular section curve.

Fig. 11. Undercut shape due to lack of control points: (a) parameters of drawbead, (b) cross section of binder, (c) section outlines at different points. (For interpretation of
the references to colour in this figure legend, the reader is referred to the web version of this article.)

the binder, denoted by the yellow plane, has an included angle of
45◦ with the stamping direction Z and the drawbead, denoted by
the blue surface, is constructed by lofting the section outlines at CP1
and CP2 guided by the drawbead curve. Fig. 11(c) shows the section
outlines at CP1 , S1 , and CP2 using blue lines. Unfortunately, the
side line of the section outline at S1 is not parallel to the stamping
direction and the drawbead has an undercut shape in the stamping
direction at S1 . In addition, the section outline should be generated
as the shape denoted by the red dashed lines rather than blue lines,
which is in opposition to the consistency of the drawbeads. Thus,
control point interpolation should be conducted to adapt to the
change tendency of the drawbead curve and binder.
This paper develops adaptive control point interpolation (ACPI)
to fix the problem discussed above. First, the drawbead curve is
subdivided into a set of points in terms of strict uniform arc length.
Suppose that pk is an input primitive control point and pki is the ith
discrete point between pk and pk+1 . Then ACPI interpolates new
dominant control points on the drawbead curve by estimating the
direction variation of T and the included angle variation of H and
B. The direction variation of T is measured by the dot product of T
at two control points:
δTki = Tik · Tjk , (14)
where Tik
is the unit tangent vector at pki .
Since B is difficult to be calculated prior to the construction of
the section curves, B is substituted by the tangent vector of the
binder in the section plane M . Then, M is calculated by:
T ×N
M = , (15)
|T × N |
where N is the normal vector of the binder at the control point. The
binder is generally smooth and the interpolation rule is based on
the angle difference between H and B, so the simplification error
can be ignored. The angle variation of H and M is measured by the
48 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 12. Errors of section curve: (a) h ≥ w/2 (1 + sinϕ), (b) h < w/2 (1 + sinϕ).
difference of dot products at two control points:
δik = Hjk · Mjk − Hik · Mik , (16)
where Hik and Mik are the horizontal axis and tangent vector of the
binder, respectively, at pki .
The geometric parameters at new dominant control points are
then obtained by linear interpolation of the two adjacent primitive
control points:
tik − t k (
χik = χ k + χ k+1 − χ k
)
(17)
t k+1 − t k
where χ represents a geometric parameter of the section curve at
i
k
pki , e.g., the height hki . t k , t k+1 and tik are the natural parameters of
pk , pk+1 , and pki on the drawbead curve.
The entire ACPI procedure is as in top of next page.
The magnitude of the discrete step d has a significant influence
on the efficiency of ACPI. Because the binder variation is generally
more stable, the tangent vector variation of the drawbead curve is
considered as the main factor influencing ACPI. As is well-known,
the curvature κ describes the rotation rate of the tangent vector of
a curve:
⏐ ∆θ ⏐ 
⏐ ⏐
κ = lim ⏐ ⏐ = Ṫ (t ) ,

⏐ (18)
∆t →0 ∆t ⏐
where ∆θ is the angle variation of T . Therefore, the following
empirical equation can be adopted to determine d:
εA
d= , (19)
ρ · κmax
where κmax is the maximum curvature of the drawbead curve
and ρ is the margin coefficient and set to a value greater than
1. The curvature of a drawbead curve is generally smaller than
0.1 rad/mm due to the consideration of the uniform deformation
of material. If κmax = 0.1 rad/mm, ρ = 5 and εA = 0.01 rad, then
d = 0.02 mm.
3.5. A priori heuristic parameter adjustment
As mentioned in Section 1, in special circumstances the mod-
eling method in Chung [2] introduces errors into the drawbead
parameters. The causes for the errors are as follows: (1) The height
of drawbead at the point is less than a certain value. (2) The
normal vector of the binder at the point is not perpendicular to the
horizontal axis H . (3) The binder is generally not planar. Without
loss of generality, we presume the binder to be temporarily planar
and the circular drawbead is taken as the example. As shown in
Fig. 12, P is a control point on a drawbead curve and O is the center
of the groove corner. h and w denote the user-defined section
height and width, respectively, of the drawbead at P. ϕ is the
included angle of H and B in Eq. (13). The red dashed line CL2 is
the critical line of height, i.e.,
w
h= (1 + sinϕ) . (20)
2
If the height of the drawbead is greater than this value, the
drawbead section is generated without error as shown in Fig. 12(a).
Otherwise, as shown in Fig. 12(b), the actual width of the drawbead
wa is smaller than w. If the binder is also not horizontal, the central
position of the drawbead Pa deviates from the control point P. In
addition, if the binder is not planar, the height error is included
since the vector I1 I2 determined by the intersection points of the
binder and section curve is not parallel to B. Obviously, the param-
eter errors affect the accuracy and consistency of the drawbead.
Moreover, the drawbead restraining force is remarkably influenced
by the parameter errors, which will further affect the design and
optimization of the drawbead.
Since the binder is usually not planar, exact error calcula-
tion cannot solve the problem completely. In this paper, a priori
heuristic parameter adjustment (PHPA) is presented to correct the
parameter errors with two steps: (1) a priori adjustment (PA),
(2) heuristic parameter adjustment (HPA), and the iterative process
is as follows: (See Fig. 13.)
In Step 1, the a priori input width and control point position
are calculated based on the planar binder assumption. The case
h < w/2 (1 − sinϕ) is taken as the example. The a priori section
width wp is provided by:
(( )
)2
w
wp = +h 2
/ h. (21)
2cosϕ
In addition, the a priori section position Pp is provided by:
( ( (w )) )
p
Pp = project P ± sinϕ −h H |V , (22)
2
where the plus and minus (±) is determined by the included angle
of V and B. The plus sign (+) is set if the included angle is smaller
than π/2, otherwise the minus sign (−) is set. Then, wp and Pp
are set as new width and new central position, respectively, to
construct the section curve.
In Step 2, HPA is applied to repeatedly correct remainder errors.
Suppose that Fig. 14 displays the positional relationship of the
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57 49

section curve and binder after the kth iteration. P is the control aw is equal to the reciprocal of derivative of the above equation
point on the drawbead curve. Psk is the kth input section central in terms of Newton’s method, i.e.,
position. I1 and I2 are two intersection points of the section curve √
and the binder. Pak is the middle point between I1 and I2 i.e., the 1 wsk − h
aw = . (26)
actual central position of the section curve. wsk and wak are the kth 2cosϕ h
′ ′
input section width and actual width, respectively. I1 I2 is parallel The kth central position error is the projection of Pak P on H , i.e.,
to I1 I2 and tangent to the section curve, which is used to measure
the actual section height hka . ekp = Pak P · H . (27)
The kth width error is:
Then the (k + 1)th input section central position is:
ekw = w − wak . (23)
Psk+1 = project Psk + ap · ekp H |V ,
( )
(28)
Then the (k + 1)th input section width is:
where ap is the position adjusting coefficient. Due to the planar
wsk+1 = aw · ekw + wsk , (24) binder assumption, ap is generally set to 1.
The kth height error is:
where aw is the width adjusting coefficient. The case h <
w/2 (1 − sinϕ) is taken as the example again. wp and w are sub- ekh = h − hka . (29)
stituted by wsk and wak in Eq. (21), respectively. The equation can be
transformed to the following form: Then the (k + 1)th input section height is:
√
wak = 2cosϕ wsk h − h2 . (25) hks +1 = ah · ekh + hks , (30)
50 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 13. Iterative process of PHPA.
where ah is the height adjusting coefficient. Since the binder is not
planar and irregular, ah is difficult to be calculated and is gener-
ally set to 1. The height error is relatively smaller and converges
quicker.
This paper adopts the minimax method to construct the evalu-
ation function to solve the multi-objective optimization problem:
minek = min max ⏐ekw ⏐ , ⏐ekp ⏐ , ⏐ekh ⏐ .
( {⏐ ⏐ ⏐ ⏐ ⏐ ⏐})
The terminal iteration condition is:
ek ≤ ε,
where ε is the user-specified maximum error.
Since the width and central position have been preferentially
adjusted in Step 1, HPA is divided into two phases, HPA on height
(HPA_H) and HPA on all parameters (HPA_AP), to coordinate the er-
ror adjustment of three parameters. In the former phase, the height
error is first decreased to a lower level (generally ε ) and the width
and position remain invariant. The latter phase reduces the three
parameter errors below the user-specified maximum error simul-
taneously. Thus, the error adjustment converges more quickly.
3.6. Rounding edges of drawbead
The last step is to round the entrance edges of the drawbeads.
First, the original binder and lofted drawbead are trimmed with
each other and the two trimmed surfaces are assembled together.
Then the entrance edge on the female bead is rounded to the radius
r2 as shown in Fig. 5. Although the entrance corner on the male
bead will not come in contact with the sheet metal, due to the CNC
programming requirement in actual manufacturing, the entrance
edge on the male bead may be rounded to a radius that is smaller
than r2 .
4. Results and discussion
The paper developed a plug-in tool to implement the proposed
geometric model of drawbead on the Dassault CATIA V5 platform
using the CAA V5 C++ API, which provides geometric operations
including projection, lofting, joining, and filleting. The skeleton
curves of the lofted drawbead surface are the section curves illus-
trated in Section 3.3.
Fig. 14. Errors in drawbead section parameters.
Table 1
Number of control points interpolated by ACPI with different angle errors.
εA (rad) 0.01 0.02 0.04 0.06 0.08 0.1
Number of CPs 845 426 212 141 105 84
The developed drawbead generation tool has been tested with
a large number of automotive covering panels. The results show
that the proposed geometric modeling process of real drawbeads is
robust and highly automated. Fig. 15 provides some typical exam-
ples of three-dimensional geometric drawbeads generated by the
tool. They include a fender, a deck lid and a side panel, respectively.
Fig. 15(e) and (f) shows the cross sections of rectangular and circu-
lar drawbeads. The upper section and the lower section of a circular
(31)
drawbead are a circular section and a rectangular section, respec-
tively, while the cross section of a rectangular drawbead consists
of two rectangular sections. The hybrid drawbead (Fig. 15(d)), in
(32) which both the circular section and the rectangular section exist on
the upper part, can improve the flexibility of the drawbead design.
The automated drawbead generation tool, in general, only costs
several minutes to construct the whole geometric scheme of real
drawbeads on a panel, while a design engineer may spend several
hours dealing with repetitive and boring operations to accomplish
the same task.
The part in Fig. 16 is tested by ACPI. The start and end points
(denoted by black points) on the closed drawbead curve (denoted
by white curve) are set as primitive control points. Fig. 16 shows
the distribution of dominant control points (denoted by red points)
interpolated by ACPI with an angle error of 0.1 rad. Table 1 presents
the number of control points interpolated by ACPI with different
angle errors, which are 0.01 rad, 0.02 rad, 0.04 rad, 0.06 rad, 0.08
rad and 0.1 rad, respectively. As seen in Table 1, the number of
control points has a nearly linear relationship with the reciprocal
of the angle error. The ACPI process not only avoids unexpected
shapes but also provides designers with a simple and flexible
method to control drawbead geometry.
PHPA is tested on the cylindrical binder, of which the cross
profile in the section plane has a curvature radius of 500 mm
and an incline angle of π/4 with the stamping direction at the
control point. The drawbead section is circular with a width of
10 mm. The maximum tolerance is set to 0.001 mm. Table 2 shows
the adjustment process with the height of 1 mm. The adjustment
results, for which the height varies from 1 to 6 mm with an
increment of 1 mm, are plotted in Fig. 17. It is observed from
Fig. 17 that PHPA leads to rapid convergence for different heights.
Generally, the smaller h is, the greater the initial errors are, which
results in more iterations. As discussed in Section 3.5, PHPA is
composed of three phases: PA, HPA_H, and HPA_AP. The PA process
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 15. Examples of three-dimensional geometric drawbeads: (a) the overall view of a fender, (b) a part of a side panel, (c) the overall view of a deck lid, (d) hybrid drawbead,
(e) cross section of rectangular drawbead, (f) cross section of circular drawbead.
Fig. 16. Distribution of control points interpolated by ACPI with an angle error of
0.1 rad.
can eliminate most of the width error and the position error in
one adjustment. Then HPA_H is conducted to reduce the height
error to a lower level (the maximum tolerance in this example).
Finally, HPA_AP simultaneously reduces all errors to less than the
maximum tolerance. The parameter adjustment suffers from PA
and HPA_H in Fig. 17(c)–(f) and all three phases in Fig. 17(a) and (b).
51
In the Chung’s article, the parameter errors caused by the
special combination of height and width are neglected. Besides,
this is very common in the design of a drawbead, which can be
observed from the optimized drawbead scheme in the following
experimental result shown in Fig. 20. Taking width as an example,
it can be observed from Fig. 17 that the initial errors are on the
same or similar order of magnitude with the prescribed width and
cannot be ignored. However, PHPA, as presented in this article, can
decrease the errors effectively, as shown in Fig. 17, and improve the
accuracy and consistency of the drawbead geometry. Moreover,
material generally flows along the direction perpendicular to the
drawbead curve, and the section curves of the proposed geometric
drawbeads are generated normal to the drawbead curve (Sec-
tion 3.3). Therefore, the restraining force provided by drawbead
per unit length that is generated by the developed tool agrees with
the nominal restraining force, which lays the foundation for the
future design and optimization of drawbeads.
A fender, which is a typical automobile covering panel, is taken
as an instance to verify the practicability and accuracy of the
proposed geometric model of real drawbeads. The CAD model of
the experimental fender is shown in Fig. 18, in which the final
product is painted green and the drawbead curve is colored ma-
genta. The real drawbeads are constructed by the developed draw-
bead generation tool in the optimization process.
52 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57

Table 2
Iterative process of PHPA with the height of 1 mm.
No. of iterations Width error Position error Height error Maximum error
1 5.839507 2.869339 0.025077 5.839507
2 3.676598 0.448574 0.603959 3.676598
3 0.795793 0.11496 0.154918 0.795793
4 0.197602 0.029483 0.03974 0.197602
5 0.050292 0.007563 0.010194 0.050292
6 0.012876 0.00194 0.002615 0.012876
7 0.003301 0.000498 0.000671 0.003301
8 0.001440 0.006058 0.000596 0.006058
9 −0.000787 0.002155 0.000311 0.002155
10 −0.002184 −0.001829 −0.000032 0.002184
11 −0.002382 −0.004154 −0.000282 0.004154
12 −0.00158 −0.004274 −0.00036 0.004274
13 −0.000333 −0.002655 −0.000276 0.002655
14 0.000755 −0.000353 −0.000102 0.000755

Fig. 17. Effectiveness of PHPA: (a) h = 1 mm, (b) h = 2 mm, (c) h = 3 mm, (d) h = 4 mm, (e) h = 5 mm, (f) h = 6 mm.

The CAE model is generated by MSFA (Fig. 19). The material coefficient K = 574.1 MPa, yield stress σ0 = 221.5 MPa, Lankford
parameters are taken as follows: thickness t0 = 0.7 mm, Young parameter r0 = 1.67, r45 = 1.2, r90 = 1.43, and hardening
modulus E=207 GPa, Poisson’s ratio υ = 0.28, strengthening exponent n = 0.185. Simulation analyses are carried out with a
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57 53

of Hill 1948 is employed to describe the yielding and plastic flow

of the anisotropy metal. The Swift–Krupkowsky hardening law
describes the material flow stress as
)n
σ = K 0.002 + ε p .
(

The stamping process is simulated by the dynamic explicit

Fig. 18. The CAD model of the experimental fender. (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version
of this article.)
friction coefficient of 0.14 and a blank holder force of 140 t. The
Belytschko–Tsay shell element is adopted with five integration
points through the thickness of the material. The yield criterion
method solver of FASTAMP [22]. After each simulation, a new
drawbead scheme is calculated by ILC integrated in FASTAMP. Then
new real drawbeads are constructed by the drawbead generation
tool automatically and the CAE model is updated simultaneously.
The simulation is repeated until the CAE result satisfies the ter-
minal condition. All the initial drawbeads are set as the circular
shape with a width of 10 mm, a height of 1 mm, and an entrance
radius of 2.5 mm, uniformly. After intelligent adjustment of the
geometry of the drawbeads, the optimized drawbead scheme is
obtained as shown in Fig. 20. Figs. 21 and 22 present the forming
limit diagrams (FLD) of the simulations before and after geometry
optimization of the drawbeads, respectively. The initial CAE result
shows that many regions suffer from insufficient stretch due to
lack of restraining force provided by drawbeads, while the CAE
result with optimized drawbeads shows good forming quality in
the product region.

Fig. 19. The CAE model of the experimental fender.

Fig. 20. The optimized drawbead scheme after intelligent adjustment of ILC.
54 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Fig. 21. FLD before the geometry optimization of drawbeads.
Fig. 22. FLD after the geometry optimization of drawbeads.
To verify the results above, an actual experiment is conducted
with the same process parameters with the optimized CAE model.
The drawbeads on the stamping tools are directly manufactured
based on the real drawbeads generated by the proposed tool
without any corrections. In the experiment, no defects, such as
fractures, wrinkles, and insufficient stretches, occur in the product
region. The thinning strain distribution of the actual experiment
(Fig. 23) is measured by ARGUS [23], which is an optical forming
analysis system able to provide full-field information about thick-
ness reduction. The thinning strains on 27 points are sampled on
the actual experimental workpiece and the optimized CAE result
(Fig. 24). The sampled data is further compared in Fig. 25, which
demonstrates that the actual experimental results have good con-
sistency with the optimized simulation analysis. Therefore, the
proposed geometric model of real drawbeads, combined with an
Fig. 23. The thinning strain distribution of workpiece in the actual experiment.
intelligent optimization algorithm (for example ILC) has a high
practicability for the automated design of drawbeads in sheet
metal forming. Moreover, as a benefit of the remarkably accurate
drawbead geometry, the generated real drawbeads can be directly
applied to manufacturing. Thus, a full process for the automated
design of drawbeads is realized.
To verify the difference between the simulations with the real
drawbeads and the equivalent drawbeads discussed in Section 1,
the optimized geometric drawbead scheme in Fig. 20 is converted
into restraining forces with the Stoughton model [24], which is
widely used and detailed in [3,5,6,9]. The simulation with the
equivalent drawbead scheme is performed with other process
parameters unchanged. Unfortunately, the simulation FLD (Fig. 26)
shows that defects occur in the stamping process, such as fractures
(A region) and insufficient stretches (B, C and D regions). Compared
with the actual experiment, the blank boundary in Fig. 26 exhibits
more overall contraction, which indicates the smaller restraining
forces being provided by the equivalent drawbeads. In the A region,
powerful restraining forces prevent the sheet metal from drawing-
in, resulting in fractures. It can be illustrated why the A region
does not fracture in the actual experiment and the simulation with
the optimized real drawbeads by Fig. 27, which shows the sheet
shapes in the simulation with the real drawbeads at the point E
(denoted by the black point in Fig. 26). Due to lack of sufficient
restraining forces provided by the die and binder, the sheet metal
in the drawbead changes in the forming process and does not fit
the rectangular drawbead shape. In other words, the real drawbead
actually provides much smaller restraining forces in the A region
than that provided by the equivalent drawbead.
In practice, the sheet shrinking line generated by the simulation
with the equivalent drawbeads is often used to guide die testing.
The die test is completed when the actual shrinking line is the
same with the simulated one. Fig. 28 shows the simulation result
with its shrinking line almost coinciding with the actual one.
However, the thinning strain distribution of the simulation still
disagrees with the actual experiment. As shown in Fig. 29, the
sampled thinning strains in the product region in this simulation
are generally greater than those in the actual experiment and the
simulation with the real drawbeads. The deviation mainly comes
from the different effects of the real and equivalent drawbeads
on the material flow. The former causes severe thinning while the
latter does not. To verify the suspected cause, the thinning strains
on the cross section 1-1 in Fig. 28 are obtained to compare with
those on the same cross section in the simulation with the real
drawbeads as shown in Fig. 30. The A and D points are located
on the die corner and the B and C points are located on the punch
corner. Compared with the simulation with the equivalent draw-
beads, in the simulation with the real drawbeads, the sheet metal
that flows though the real drawbeads (nearby the points A and D)
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57 55

Fig. 24. The thinning strain distribution of workpiece in the optimized CAE result.

Fig. 26. FLD with the equivalent drawbeads converted by Stoughton model.

Fig. 25. Comparison of thinning strains on 27 sampled points between the actual
experiment and CAE result with optimized geometric drawbeads.

is thinned more drastically and the thinning of the sheet metal that
is further away from the real drawbeads is relatively smaller.
From the two simulations with equivalent drawbeads that are
obtained by different methods, the simulation results show a sig-
nificant difference compared with the actual experiment, which
will lead to an increase in die test time and cost. With the com-
puting capacity significantly improved, the real drawbeads, which
can veritably simulate the deformation of finite elements in draw-
beads, should be employed to guide sheet metal forming design
and testing.
5. Conclusions
A robust and accurate geometric model of real drawbeads is
presented for automated drawbead design in the paper, i.e., UDF
of real drawbeads. Three-dimensional drawbeads are constructed
by lofting section curves, which are generated in terms of the
geometric parameters stored in control points. Control points are
prescribed to control the distribution of the restraining force of
the drawbeads. ACPI is developed to interpolate new dominant
control points to adapt to the change tendency of drawbead curve
and binder and avoid unexpected shapes. The introduction of con-
trol points simplifies the management of the drawbead geometry
and increases the consistency of drawbead generation. Moreover,
PHPA is performed on each section curve to correct parameter
errors of the geometric drawbeads. Thus, error-bounded geometric
Fig. 27. Shapes of sheet metal at the closing point and lower die point.
drawbeads are generated, which can significantly improve the
accuracy and consistency of drawbeads. The geometric model of
real drawbeads is one of the main contributions of this paper.
Another novelty of this paper is the fully automated design
process of drawbeads. Based on the geometric model of real
drawbeads and the previously developed intelligent optimization
algorithm of drawbeads, the fully automated design process of
drawbeads including geometric modeling (CAD), finite element
analysis (CAE), intelligent optimization design, and CNC manufac-
turing (CAM) is realized. The experimental result of the fender
demonstrated good consistency with the CAE result optimized
by the proposed geometric drawbead design process, which ver-
ifies the validity and accuracy of the geometric model of real
drawbeads. In addition, the comparison between the proposed
geometric drawbeads and the equivalent drawbeads obtained by
two different methods is conducted and the result shows that the
proposed geometric drawbeads are much closer to the drawbeads
56 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57

Fig. 28. The thinning strain distribution with the equivalent drawbeads with its shrinking line coinciding with the actual one.

References

[1] Pearce R. Sheet Metal Forming. Bristol: Adam Hilger; 1991.

Fig. 29. Comparison of thinning strains on 27 sampled points between the actual
experiment and CAE results.
Fig. 30. Comparison of thinning strains on cross section 1-1 between CAE results
with equivalent and real drawbeads.
in practical application and should be adopted to guide die design
and testing.
Acknowledgment
This work is financially supported by the National Natural Sci-
ence Foundation of China (grant number 51275184).
[2] Chung YC. Geometric modelling of drawbeads using vertical section sweeping.
Comput-Aided Des 2013;45(12):1556–61.
[3] Naceur H, Guo YQ, Batoz JL, Knopf-Lenoir C. Optimization of drawbead re-
straining forces and drawbead design in sheet metal forming process. Int J
Mech Sci 2001;43(10):2407–34.
[4] Beimesl SM, Mostafapour A. Evaluation of draw bead geometry on deep
drawing of complex rectangular shape part. Adv Mater Res 2012;445:96–101.
[5] Guo YQ, Batoz JL, Naceur H, Bouabdallah S, Mercier F, Barlet O. Recent develop-
ments on the analysis and optimum design of sheet metal forming parts using
a simplified inverse approach. Comput Struct 2000;78:138–48.
[6] Liu G, Lin ZQ, Bao YX. Optimization design of drawbead in drawing tools of
autobody cover panel. J Eng Mater Technol 2002;124(2):278–85.
[7] Song JH, Huh H, Kim SH, Park SH, Springback reduction in stamping of front
side member with a response surface method. In: Proceedings of the 6th
international conference NUMISHEET. 2005. p. 303–8.
[8] Liu Q, Liu WJ, Ruan F, Qiu HY. Parameters’ automated optimization in sheet
metal forming process. J Mater Process Technol 2007;187–188:159–63.
[9] Sun GY, Li GY, Gong ZH, Cui XY, Yang XJ, Li Q. Multiobjective robust opti-
mization method for drawbead design in sheet metal forming. Mater Des
2010;31(4):1917–29.
[10] Sun GY, Li GY, Li Q. Variable fidelity design based surrogate and artificial
bee colony algorithm for sheet metal forming process. Finite Elem Anal Des
2012;59:76–90.
[11] Firat M. An analysis of sheet drawing characteristics with drawbead elements.
Comp Mater Sci 2008;41(3):266–74.
[12] Larsson M. Computational characterization of drawbeads a basic modeling
method for data generation. J Mater Process Technol 2009;209(1):376–86.
[13] Zhang QC, Liu YQ, Zhang ZB. A new method for automatic optimization of
drawbead geometry in the sheet metal forming process based on an iterative
learning control model. Int J Adv Manuf Technol 2016;1–17.
[14] Li G, Liu YQ, Du T, Tong HL. Algorithm research and system development
on geometrical springback compensation system for advanced high-strength
steel parts. Int J Adv Manuf Technol 2014;70(1):413–27.
[15] Chung YC. Development of a geometric modeling tool for rectangular draw-
beads in draw dies. J Mech Sci Technol 2013;27(8):2401–7.
[16] Woodward CD. Skinning techniques for interactive b-spline surface interpola-
tion. Comput-Aided Des 1988;20(10):441–51.
[17] Choi BK. Surface modeling for CAD/CAM. Amsterdam: Elsevier; 1991.
[18] De Boor C. A practical guide to splines. Berlin: Springer; 1978.
[19] Hoschek J, Lasser D. Fundamentals of computer aided geometric design.
London: Peters; 1993.
[20] Park H, Kim K. Smooth surface approximation to serial cross-sections. Comput-
Aided Des 1996;28(12):995–1005.
[21] Piegl L, Tiller W. Surface skinning revisited. Visual Comput 2002;18(4):273–83.
[22] Du T, Liu YQ, Zhang ZB, Li ZG. Fast FE analysis system for sheet metal stamping-
FASTAMP. J Mater Process Technol 2007;187–188:402–6.
[23] Schneiderc M, Friebed H, Galanulisd K, Validation and optimization of numer-
ical simulation by optical measurements of tools and parts. In: Proceedings of
the IDDRG 2008 international conference. 2008. p. 373–84.
[24] Stoughton TB, Model of drawbead forces in sheet metal forming. In: Proceed-
ings of the IDDRG 15th Biennial congress. 1988. p. 205–15.
 Z. Wang et al. / Computer-Aided Design 92 (2017) 42–57
Zhen Wang is a Ph.D. student in State Key Laboratory
of Materials Processing and Die & Mould Technology,
Huazhong University of Science and Technology (HUST).
He received his B.E. from HUST, China, in 2013. His re-
search interests include geometric algorithms and auto-
mated design in CAD/CAM and system modeling and sim-
ulation.
Qiuchong Zhang is a Ph.D. student in State Key Labora-
tory of Materials Processing and Die & Mould Technology,
Huazhong University of Science and Technology (HUST).
He received his B.E. from HUST, China, in 2012. His re-
search interests are intelligent optimization algorithm for
process design and finite element method for sheet metal
forming.
57
Yuqi Liu is a professor in State Key Laboratory of Mate-
rials Processing and Die & Mould Technology, Huazhong
University of Science and Technology (HUST). He received
his B.E., M.S. and Ph.D. from Jilin University of Technology,
Jilin University and Jilin University of Technology, China, in
1990, 1993 and 1995, respectively. His research interests
include embedding intelligence for computer graphics and
CAD and finite element method for sheet metal forming.
Zhibing Zhang is a lecturer in State Key Laboratory of Ma-
terials Processing and Die & Mould Technology, Huazhong
University of Science and Technology (HUST). He received
his B.E., M.E. and Ph.D. from HUST, China, in 2002, 2005
and 2008, respectively. His research interests are com-
puter graphics and finite element method for sheet metal
forming.