Thin-Walled Structures 84 (2014) 302312

Contents lists available at ScienceDirect

Thin-Walled Structures
journal homepage: www.elsevier.com/locate/tws

Optimal shape design of thin-walled tubes under high-velocity axial

impact loads
Niyazi Tanlak, Fazil O. Sonmez n
Department of Mechanical Engineering, Bogazici University, Istanbul, Bebek 34342, Turkey

ar t ic l e i nf o

a b s t r a c t

Article history:
Received 2 October 2013
Received in revised form
7 July 2014
Accepted 7 July 2014
Available online 2 August 2014

In this study, the objective is to maximize the crashworthiness of thin-walled tubes under axial impact
loads by shape optimization. As design variables, parameters dening the cross-sectional prole of the
tube as well as parameters dening the longitudinal prole like the depths and lengths of the
circumferential ribs and the taper angle are used. The methodology is applied to the design optimization
of a crash-box supporting the bumper beam of a vehicle for the loading conditions in standard EuroNCAP
crash tests. The crash event is simulated using explicit nite element method. While the crash-box is
fully modeled, the structural response of the remaining parts during the tests is taken into account by
developing a lumped-parameter model. A hybrid search algorithm combining Genetic and Nelder &
Mead algorithms is developed. The results indicate signicant improvement in the crashworthiness over
the benchmarks designs.
& 2014 Elsevier Ltd. All rights reserved.

Keywords:
Thin-walled tubes
Crash-box
Crashworthiness
Explicit nite element analysis
Parametric system identication
Global optimization

1. Introduction
Thin-walled tubular structures are preferred in applications
requiring high performance under impact loadings. In automobiles, thin-walled tubular parts are used to absorb impact
energy in a potential crash. These parts should be designed to
minimize the damage to the main parts of the vehicle and protect
the occupants from injury by absorbing the collision energy. Their
effectiveness in preventing injury under such impact loads is
called crashworthiness. The performance of these parts can be
signicantly improved by optimizing their shapes.
There are a number of studies in the literature on the optimization of tubes made of metals (usually steel and aluminum) under
axial impact loading. In these studies, tubes are considered as
either empty [114] or lled [12,1523]. Yamazaki and Han [1]
studied square and cylindrical tubes hitting a rigid wall with a
velocity of 10 m/s. They maximized the total energy absorption
while maintaining the mean crushing force at a certain limit by
varying the thickness of the tube and the section radius. Lee et al.
[2] studied tubes with circular cross-section hitting a rigid wall
with a velocity of 10 m/s and additional mass of 500 times the
mass of the tube. Their design parameters were the wall thickness,
radius, and length of the tube. Sheriff et al. [3] used the bottom
diameter, height, and taper angle as design variables to maximize
the total energy absorbed in circular cross-section tubes. Avalle

Corresponding author. Tel.: 90 212 359 7196; fax: 90 212 287 2456.
E-mail address: [email protected] (F.O. Sonmez).

http://dx.doi.org/10.1016/j.tws.2014.07.003
0263-8231/& 2014 Elsevier Ltd. All rights reserved.

and Chiandussi [4] optimized cylindrical tubes with tapered tip for
uniform reaction force distribution. They varied the length of the
tapered tip and the tip diameter. Hou et al. [5,6] optimized square
and hexagonal single-cell and multi-cell tubes using base dimensions and thickness as design variables for minimum peak force
and maximum specic energy absorption, i.e. energy absorption
per unit mass. Acar et al. [8] varied taper angle and number of ribs
on the surface in order to maximize the ratio of the mean crush
force to the peak force and the specic energy absorption. Qi et al.
[9] analyzed single and multi-cell square tubes under oblique
impact. Their objective was to increase the specic energy absorption and minimize the peak crushing force by changing the taper
angle and the wall thickness. Liu [10] optimized the wall thickness
and the side length of a box-shaped column to maximize the
specic energy absorption with a constraint on the peak force. Liu
[11] considered straight and curved octagonal and hexagonal tubes
and selected the side length and the wall thickness as variables.
The objective was to maximize specic energy absorption of the
columns while constraining the peak force. Yang and Qi [12]
studied empty and lled tubes with a square cross-section under
axial or oblique impact. Their objective was to increase the specic
energy absorption and minimize the peak crushing force by
varying the wall thickness, cross-section width, material yield
strength, and ller material density. Zarei and Krger [13] optimized empty cylindrical tubes by taking their length, diameter,
and thickness as design variables for increased total energy and
specic energy absorption. They extended that study to tubes
lled with honeycomb [16] and foam [17] by considering their
densities as variables. Kim and Arora [14] studied representation

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

of tapered tubes with uniform ones with square-sections in the

forcedisplacement domain. With these force representations, a
model with a single degree of freedom that simplied the analysis
of the tube structure was identied. Santosa and Wierzbicki [15]
studied the axial crushing resistance of a square-box column lled
with aluminum honeycomb or foam under quasi-static loading
conditions. Sun et al. [20] optimized crash-boxes with functionally
graded foams for maximum specic energy absorption and minimum peak force. They assumed the foam material as layered and
they varied the density of these layers. Hanssen et al. [18] used
formulas derived based on experimental data relating design
parameters to average force, maximum force, and stroke efciency
in order to obtain optimum designs of columns for minimum
reaction force and maximum energy absorption. Ahmad and
Thambiratnam [19] conducted a parametric study on empty and
foam-lled tubes under axial impact loads using the wall thickness, taper angle, foam density, impact mass, and impact velocity
as variables. Yin et al. [21] studied honeycomb-lled single and
bitubular polygonal tubes. The variables were the wall thickness
and the side length. The objective was to maximize the specic
energy absorption and to minimize the peak force. Bi et al. [22]
studied foam-lled single and multi-cell hexagonal tubes, which
were crushed under a rigid wall moving downward with a velocity
of 2 m/s and penetration depth of 100 mm. The variables were
chosen to be the wall thickness and the side length of the section,
and the foam density. The objective was to increase the specic
energy absorption while keeping the mean crushing force larger
than a certain limit to ensure a certain structural rigidity. Tarlochan et al. [23] conducted a parametric study on foam lled tubes
under axial and oblique impact loading. They compared tubes
having circular, square, hexagonal, octagonal, ellipsoidal crosssections with the same circumference in terms of energy absorption and crush force efciency.
The previous researchers generally developed approximate expressions for the objective functions using response surface methodology
[16,811,13,16,17,2022], Kriging [12], moving least-squares approximation [14], and articial neural network [7]. After obtaining the
surrogate models, they used genetic algorithms [3,16,17], leap-frog [7],
particle swarm optimization [9,20,12,21], non-linear programming
[5,6,8], or multi-rst order method [4] as search algorithms.
The cross-sectional shapes of the tubes considered by the
researchers were circular [14,8,13], square [1,5,7,9,10,12,1620],
hexagonal [6,11,2123], or octagonal [11,21]. Some of the previous
studies focused on straight columns with uniform cross-section
along the length [1,2,57,913,1618,2022], some of them introduced taper angles [3,4,8,19], and some [8,7,16,17] introduced ribs
with predened shapes.
In the present study, a larger number of geometric parameters
are used as optimization variables in comparison to the previous
studies. The parameters dening the shape of the cross-sectional
prole (the coordinates of key points dening the spline curves)
and the longitudinal prole (depths and widths of the circumferential ribs, and the taper angle), and the wall thickness are varied.
To the authors' knowledge, the shapes of the ribs are optimized for
the rst time in this study; in the previous studies, on the other
hand, they were taken as constant. The ribs can be inward or
outward. The taper angle can be positive or negative. Besides, the
loading conditions considered in almost all the previous studies
were either drop tests i.e. an object being dropped on a column, or
a column with a mass at the rear hitting a rigid wall. In this study,
the methodology developed for the optimum shape design of
tubular structures is applied to the crash-boxes (or brackets)
holding the bumper beam of a vehicle. The behavior of the
crash-box is simulated for the loading conditions in a standard
high-speed crash test, European New Car Assessment Program
(EuroNCAP). Because of the difculty in modeling the whole car

303

and the resulting long computational times, a lumped-parameter

car model is developed that accounts for the structural behavior of
the main body of the vehicle as well as the parts in front of the
crash-box. Moreover, in this study, a hybrid of genetic algorithm
(GA) and NelderMead algorithm is developed to nd the globally
optimal design or a near global optimal design.
The goal of global design optimization is to nd the design with
the best possible performance. This requires a denition of the
geometric design that allows signicant changes in shape, i.e. the
solution domain should be large so that it includes the designs
leading to the highest possible levels of performance. This means
the number of geometric parameters and the range of values that
can be assigned to these parameters by the search algorithm
should be large. Global shape optimization of a vehicle for
maximum crashworthiness is infeasible considering the high
number of interacting parts, the high number of parameters used
to dene their geometries, and complex interactions among them
during crash. This is beyond the capabilities of the current state-of
the-art computers and search algorithms. Considering the computational effort to simulate crushing of the whole vehicle, it is not
possible to nd the globally optimum design within such a large
solution domain and with such a large number of design variables
even if a powerful global search algorithm is used. If large changes
are allowed in the values of the optimization variables during
optimization, the accuracy of surrogate models will also be
questionable even for a single part let alone the whole vehicle.
Besides, if the individual parts are separately optimized, loading
conditions on them will be different from that of a drop test. That
means the shape of a part optimized for the loading conditions in a
drop test will not be optimum for the loading conditions in a real
crash test.
The procedure suggested in this study to surmount these
problems is the following: individual parts of the vehicle are
optimized via a reliable global search algorithm by using a high
number of design variables and allowing large changes in the
values of these variables. The remaining parts of the vehicle are
modeled with a system of lumped masses, springs, and dampers
using parametric system identication, therefore computational
times will not be prohibitively long. At the last stage, the whole
vehicle is optimized starting from the optimized shapes of the
individual parts, but this time the ranges of values that can be
assigned to the variables will be small and some of the parameters
may be taken as constant. Then, it becomes feasible to develop a
reliable surrogate model for the vehicle and perform optimization.
In this study, modeling and design optimization of a single part, a
crash-box, is considered.

2. Problem statement
The main objective of this study is to develop a methodology to
obtain the optimum shape design of thin-walled tubes subjected
to high-velocity axial impact loads. The specic structure considered in this study is the bracket that supports the bumper beam of
a car. Two brackets hold the bumper beam at two sides. They are
in turn xed to the main frame of the car. The types of obstacles
that bumper-beam-crash-box system endures during frontal
impact are countless. However, they can be categorized into three
major divisions: full frontal collision, offset frontal collision, and
pole frontal collision. The harshest collision that a bracket endures
is the offset frontal impact, where one of the brackets takes the
impact energy. Accordingly, the crash-box is optimized for offset
collision conditions in accordance with EuroNCAP, IIHS, ANCAP
standard tests, where the vehicle hits a wall with 40% offset and
64 km/h speed (See Fig. 1). At such high speeds, a car incurs
substantial damage, but it is crucial that the occupants do not

304

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

suffer serious injury. Crashworthiness of the vehicle depends on

the transferred impact force and the capacity of the structural
parts to absorb the collision energy. The crash-box is one of the
parts that should be optimally designed in order to achieve the
required crashworthiness of the vehicle.

3. Approach
3.1. The objective function

P P g P mass P en P an

The effectiveness of design optimization depends on the proper

choice of the objective function. Improving its value should
enhance the structural performance of the part in the most
effective way. The following objective function is chosen to be
maximized in the present study:
v
Z t f ZZZ
ZZZ
w1 1
w u1
ij _ ij dV dt  2 u
 2 dV
f obj
t
n 1 V 0
n2 V

 w3

phases of the collision to reduce the jerking effect felt by the

occupants during impact. wi are the weighting constants, their
values are chosen according to the relative importance that
the designer gives to the individual terms, and ni are the normalization constants, which are n1 38:1 J=g, n2 1195, n31
156:7 kN, and n32 991:8 MN=s. These values are obtained from
the simulations conducted for a 70 mm diameter circular tube,
which is taken as benchmark. P is the penalty function, which
includes the following terms:


1
1 F max
P
F max
2n31
2n32 t

Fig. 1. A scheme for EuroNCAP frontal offset crash tests [24].

Pg is the geometric constraint, which is activated when a crosssectional prole generated by the search algorithm does not t the
allowable spacing. If Pg becomes active, a large value is added to
the objective function without calculating the other terms. P mass is
the mass constraint such that

P mass

where V is the volume of the crash-box, tf is the duration for which

the simulation is conducted starting from the initial contact with
the barrier, is the equivalent strain, is its mean value, Fmax
is the maximum force at the rear of the crash-box that occurs in
the rst 0.5 ms of the collision, and t is the time that passes until
the maximum force occurs. The rst term is the specic internal
energy, or the energy absorbed per unit mass of the crash-box
during crash. It is a measure of how effectively it takes the collision
energy. The second term is a measure of uniformity in the
deformation. This term is introduced to avoid shapes that result
in highly localized deformation. The smaller this term is, the more
uniform is the deformation; accordingly, the integrity of the tube
is expected to be better maintained. The third term includes the
magnitude of the peak force during the rst 0.5 ms, Fmax, as well as
the average rate of increase in force, F max =t. During collisions,
generally the force transmitted reaches a peak value at the initial
stages of the impact. It is essential to reduce the magnitude of the
peak force as well as the rate of increase in force during the initial

8
>
<0





m  mben 2
m mben
>
1000

100
:
mben
mben

if m r mben
if m 4 mben

where m is the mass of the crash-box and mben is the mass of the
benchmark structure. The mass constraint is introduced in order to
avoid optimal designs that show increased crashworthiness over
the benchmark design at the expense of increased weight. Pen is
the penalty introduced to eliminate the designs taking a lower
impact energy in comparison to the benchmark case. Its value is
calculated as

P en

8
0
>
>
<
>
>
: 3600

E ben  E acc
E ben

!2
450

E ben  E acc
E ben

if E acc Z E ben
if E acc o E ben

where E acc is the total accumulated energy in the crash-box and

E ben is the benchmark value for energy. Pmass and P en are formulated such that the terms yield small penalty values for small
violations, but the penalty value increases quadratically for large
violations. Pan is the analysis constraint, which is activated when
nite element analysis fails due to sharp corners or some other
reason. By introducing the penalty terms, the constrained optimization problem is transformed into an unconstrained one.
The above metric is formulated to nd the best shape that
maximizes the specic energy absorption, avoids development of
extreme plastic strains leading to rupture in the structure and also
reduces the peak force and its jerk effect while ensuring that this is
achieved by keeping the total energy absorption above certain
limit without increasing the mass.

Fig. 2. Dening the cross-sectional (a) and the longitudinal (b) prole of the tube by spline curves.

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

(4) that

3.2. The optimization variables and denition of the shape

In the present study, the shapes of both the cross-sectional prole
and the longitudinal prole as well as the wall thickness of the crashbox are varied. Making use of the double symmetry of the cross
section with respect to the x- and y-axes (Fig. 2a), only the shape of
one-quarter is dened by a spline curve using three key points. Point
1 is allowed to move only horizontally; point 3 only vertically; on the
other hand, point 2 can move in both directions. The longitudinal
prole is dened by a series of coefcients as well as three length
parameters. The rst coefcient is the taper coefcient, c , which is
used to dene the taper angle, . c is the ratio of the radial
coordinates of the base and end-section key points. Furthermore,
up to three ribs can be created at predened locations along the
crash-box. The proles of the ribs are constructed using cubic spline
curves, which are described by key points dened by depth coefcients, r i , and length parameters, li (See Fig. 2b). ri  c is the ratio of
the radial distance of the key point at the ith rib to that at the base.
With the help of r i  c , the depths of the ribs, hi, are dened. The
taper coefcient, c , can take values larger or smaller than 1.0.
Accordingly, taper angle, , can be positive or negative. Similarly, hi
may take positive or negative values; the ribs may then be inward or
outward. Due to size limitations, constraints are imposed on the
design variables such that the key points may not go outside the
domain shown in Fig. 2a. In addition to the outer boundary of the
design space, the key points dening the base cross-sectional prole
are not allowed to enter the inner circular region with 12 mm radius
in order to avoid numerical difculties. The dimensions of the outer
boundary are chosen based on the dimensions of a crash-box
currently used in a Renault car. If the algorithm assigns a position
for a key point outside this domain, a large penalty value, P g , is
assigned to the objective function.
3.3. Modeling the crash conditions
3.3.1. Modeling of the barrier
EuroNCAP uses a deformable barrier made of aluminum honeycomb in frontal offset-impact tests. The deformable barrier
is designed to simulate an average collision partner during an
accident. Although quite complex structural problems can be
solved by FEM, analysis time should not be long in design optimization studies where thousands of runs may be needed to locate
the globally optimum design or a near globally optimum design.
For this reason, the barrier is modeled as rigid in the simulations.
Because a rigid barrier does not absorb any impact energy, the
effect of collision is severer. Deb et al. [25] used a lower velocity for
the car in their simulations in order to account for the effect of
deformable barrier. However, in this study, a lower mass is used
for the car such that the severity of impact is more or less the
same. The energy conservation equation for collision of a vehicle
with a deformable barrier can be written as
1
mv20 Evint Evke Ebar
2

m0 m 

2
Ebar Evke  E0vke
v20

m0  m 

2
Ebar
v20

According to Deb et al. [25], the ratio of the energy absorbed by

the deformable barrier in a conventional offset test to the vehicle
mass, Ebar =m, does not change much for cars of various masses.
Taking Ebar =m as 82 J/kg as suggested by Deb et al. [25], m0 is found
to be 0.483 m, but to be on the safe side it is taken as 0.5 m. The
mass of the vehicle is 1116 kg (904 kg 2  88 kg 36 kg, which
are the masses of Hybrid-III dummies of 88 kg and a luggage of
36 kg). So the effective mass, m0 , is 558 kg. In the simulations, the
car hits a rigid wall with an initial velocity of 64 km/h (17.8 m/s)
and 40% offset.
3.3.2. Modeling of the main vehicle body
Another concern is the modeling of the car. One option is to
model the whole car as a deformable body in full detail, but this
leads to excessively long computational times. Considering that a
globally optimization procedure may require a high number of
iterations, modeling the whole car is not a feasible option.
In this study, a lumped-parameter model is developed consisting of lumped masses, springs and dampers to account for the
effects of the vehicle components behind the crash-box as well as
the bumper beam in front of the crash-box, which is represented
by a non-linear spring. Fig. 3 shows a depiction of the model used
in this study. Because, at the early stages of the impact, plastic
deformation mainly occurs in the bumper beam and the crashbox, then in the remaining structural parts of the vehicle, the main
body is assumed to be linearly elastic and its mechanical response
can be represented by linear springs during early phases of the
collision during which the simulations are continued.
The model parameters are k1 ; c1 ; k2 ; m1 ; m2 , where k1 and c1
account for the elastic deformation and damping effects behind
the crash-boxes, while the effect of the deforming bumper-beam is
accounted for by a nonlinear spring. The forcedisplacement curve
of this spring is dened by a quadratic polynomial, Fu au2 bu,
up to u un ; after that the relation becomes linear, Fu cu d.
m1 is the mass of the intermediary rigid plate and m2 is the mass
of the point mass such that m1 m2 m0 . Here, the other crashbox is assumed to take no impact energy in offset impact tests.
In order for the lumped-parameter model to represent the
behavior of the car during collision, suitable values for the model
parameters should be chosen. For this purpose, the values of the
model parameters are optimized so that the results of the present
lumped-parameter car model match that of the full nite element

Here, m0 is the equivalent vehicle mass and E0vke is the nal kinetic
energy of the vehicle after impact. It now follows from Eqs. (3) and

Assuming Evke  E0vke , the above relation becomes

where v0 is the initial velocity of the vehicle, m is its original mass,

Evint is the energy absorbed by the vehicle, Evke is the nal kinetic
energy of the vehicle after impact, and Ebar is the energy absorbed
by the deformable barrier.
If the vehicle hits a rigid barrier with the same velocity, v0, but
with a different mass such that the same amount of energy, Evint, is
absorbed by the vehicle, then the equation becomes
1 0 2
m v0 Evint E0vke
2

305

Fig. 3. A depiction of the vehicle model.

306

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

model for 2010 Toyota Yaris (Sedan), which was developed and
validated by The National Crash Analysis Center (NCAC) of the U.S.
[26]. This car model is crashed into a rigid wall with an offset
according to the scenario specied above.
The lumped-parameter model is assembled with the crash-box
as depicted in Fig. 3. Finite element simulations are conducted
according to the collision scenario described above. The resulting
reaction forces on the rigid barrier due to the collision are
calculated as a function of time. The model parameters are
optimized so that the impact forces on the wall obtained using
the full car model and the lumped-parameter car model are as
close to each other as possible. In order to estimate the closeness
of the two outcomes, the following measure is used, which is the
sum of the differences in the impact forces at corresponding time
intervals:
s
 


2
100
t f t 0
tf  t0
i  F~ t 0
i
F t0
7
f obj
100
100
i0
where t0 is the initial time, tf is the nal time, F is the resulting
impact force on the barrier for Toyota Yaris Model, and F~ is the
force for the lumped-parameter model. The values of k1 ; c1 ; m1
together with the parameters dening the nonlinear spring
representing the bumper-beam are optimized to yield the minimum value for fobj. This optimization problem is solved using the
same search algorithm used to optimize the crash-box. After
optimizing the parameters, the forcedisplacement relation is
obtained for the spring representing the bumper-beam.
As depicted in Fig. 4, the lumped-parameter model gives a
response close to that of the full car model. The optimum values of
parameters k1 ; c1 , and m1 are found to be 1690 106 N/m, 145.7
103 N s/m, and 112.79 kg, respectively. The parameters used to
dene the stiffness of the nonlinear spring are found to be
a 2.9 MN/m2, b 143:7 MN=m, c 1.4 MN/m, d 1.6 MN and
the transition displacement is obtained as un 0:32 mm.
3.4. Search algorithm
A search algorithm is utilized to nd the optimum values of the
variables that yield the maximum value for the objective function
expressed in Eq. (1). Considering that typical structural optimization problems contain numerous local optimums, a local search
algorithm may easily get stuck at a worse local optimum rather
than the global optimum. If the problem has a complex solution

domain, multiple restarts may even not work. Heuristic global

search algorithms, on the other hand, may nd the best conguration, but they require a large number of function calls for
convergence. For this reason, they are not feasible for problems
requiring long computational times like crash simulations. In the
present study, in order to search for the globally optimum design
without excessive computational burden, a hybrid algorithm
combining global and local search algorithms is developed. In this
method, the genetic algorithm (GA) is used to nd congurations
potentially close to the global optimum or one of the near global
optimums. Then, these congurations are supplied to a local
optimizer, Sequential Simples (Nelder & Mead) algorithm, which
in turn locates the optimum.
3.5. Optimization procedure
In each iteration, new congurations are generated by the
search algorithm. In order to evaluate the value of the objective
function for these congurations, structural analyses of the corresponding crash events are carried out. For this purpose, a FE model
is developed. The FE model and the optimization algorithm are
integrated using a built-in ABAQUS python script. This code carries
out FE analyses of the congurations generated by the search
algorithm, writes the results on output les, and also evaluates the
results to modify the values of the optimization variables according to the decision criteria of the hybrid search algorithm to obtain
new candidate congurations.
At the start, the optimization code selects random values for the
optimization variables within the feasible domain and creates the
corresponding geometries of the crash-box. In this way, the initial
population of GA is obtained. Using the predened velocities, the
boundary conditions, and the material properties, FE analyses are
conducted for the randomly selected geometries. Based on the FE
analysis results , the values of the objective function are calculated.
The code then compares these values and selects new values for
the optimization variables for the next generation according to the
decision criteria of the GA. This procedure is repeated until
the stopping criterion is satised, which requires no change in the
best value found in three consecutive generations after the rst 19
iterations. The initial population is taken as 300. The population is
dynamic with an elitist approach that means it is initially high, but
towards the end it gets lower; however the ttest member is always
maintained. Crossover and mutation probabilities are also dynamic
like the population. The best points found by GA are used as initial
points by Nelder & Mead algorithm, which then tries to nd the best
local optimum in their neighborhood. Iterations are continued until
the difference between the objective function values of the best and
worst congurations becomes small.

4. Finite element modeling

Explicit FE methods are better in solving structural problems
involving complex contact interactions occurring within a short
duration compared to the implicit ones. Accordingly, commercial
nite element software ABAQUS/Explicit is used in the present
study to simulate the behavior of the crash-box system during
crash tests.
4.1. Model geometry

Fig. 4. Impact forces on the barrier resulting from car crash calculated using the
lumped-parameter and the full car models.

The nite element model includes a deformable model of one

crash-box as shown in Fig. 5. The main vehicle body is represented
by a point mass, a spring, a damper, and an intermediary rigid
block, which also serves uniform transmission of forces between
the crash-box and the rest of the car. The bumper beam is

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

307

Fig. 5. Finite element model.

represented by a nonlinear spring, which is placed between the

rigid plate and the rigid barrier for uniform force transmission.
4.2. Initial and boundary conditions
The initial and boundary conditions dened in the nite
element model should reect the conditions of the crash tests.
Otherwise, the response of the crash-box cannot be correctly
predicted. In the present nite element model, all parts except
the rigid barrier have an initial velocity of 64 km/h (17.8 m/s). The
rigid barrier which the car hits is xed. The rear end of the crashbox and the intermediary rigid plate and block as well as the point
mass are constrained to move only in the direction of the initial
velocity so as to prevent relative movement of the blocks in the
transverse directions considering that the transverse displacements are almost zero during the initial phases of the crash as
shown in Fig. 6. Rotation of the vehicle occurs only at later stages
of the crash. Because the geometry of the crash-box and the
loading are symmetric with respect to the horizontal as well as
vertical mid-planes, only one-quarter of the assembly is analyzed
and the necessary symmetry conditions are applied on the corresponding interfaces.

Fig. 6. Rigid body displacement history of Toyota Yaris model [26] during the initial
stages of offset frontal impact.

4.3. Analysis time

Fig. 7 gives the energy absorbed by a circular crash-box, while
Fig. 8 shows the change in the geometry of the crash-box at 1 ms
intervals. As seen in the gures, the crash-box totally collapses
after 5 ms, after that its stiffness greatly increases like a compression spring closed solid due to an overload, then a much larger
energy is required to induce a small deformation. At this stage, the
remaining portions of the car, which are initially stiffer than the
bracket, start to absorb signicant amount of energy. Accordingly,
continuing the simulations longer than 5 ms is not appropriate,
because the energy absorbing capacity of the bracket is consumed.
The analysis consists of a single explicit dynamic step. In the
simulations, automatic time incrementation is used with elementby-element stable time increment estimates.
4.4. Constitutive model
Yield strength of materials typically increases with an increase
in plastic strain as well as strain rate. During a crash test, the
crash-box severely deforms in a very short time. For this reason, a
realistic simulation of a crash event requires a constitutive model
that accounts for non-linear and strain-rate dependent deformation and also work hardening. In the present FE model, Johnson
Cook constitutive model [27] is used. According to this model, the
equivalent ow stress of the material, , depends on equivalent

Fig. 7. Total energy values accumulated within a typical the crash-box during the
collision.

plastic strain and its rate as

"

_ pl
0 B pl n  1 C ln _
0

!#
8

308

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

Fig. 8. Deformation of the crash-box with a circular base shape.

where pl is the equivalent plastic strain, _ pl is its rate, _ is the

reference strain rate, and 0 is the initial yield stress. The values of
strain hardening coefcient, B, and exponent, n, can be obtained
using quasi-static tension tests at the constant strain rate. The
value of the strain rate parameter, C, is determined through
dynamic tension tests conducted at different strain rates. Note
that temperature effect is excluded in the present model.

4.5. Failure model

In crash tests, severe deformations are induced in the crash-box
that can lead to local failures like cracks. Because failed elements
may not continue to absorb energy, these may affect further deformation behavior and energy absorbing capacity of the crash-box.
For this reason, a cumulative failure model proposed by Johnson
and Cook [28] is adopted in this study.
, is dened,
In the model, a critical equivalent fracture strain, pl
f
which depends on hydrostatic tension, p, and effective stress, .
The failure strain is expressed with four material constants, Di, as
"
!#
_ pl
9
plf D1 D2 eD3  1 D4 ln _
0
A damage parameter is dened as

pl0 pl
plf

10

pl
where pl
is an
o is the initial equivalent plastic strain and
increment of the equivalent plastic strain. is dened at the
integration point of every nite element and is used as a measure
of failure. Failure occurs when exceeds the unity. If an element
fails, very low values are assigned to the element for the mechanical properties.

Fig. 9. Comparison of FEM results with the experimental drop test data [4].

5. Results and discussion

5.1. Comparison of FEM results with a test case
In order to validate the accuracy of the nite element model
used in the present study, the model is adapted to a similar
problem and the results obtained by the FE model are compared
with the experimental results obtained by Avalle and Chiandussi
[4]. The tests were conducted on a drop weight testing machine in
which an object was dropped on a thin-walled tube with tapered
tip at rest. The mass of the hammer was 60 kg and its initial
velocity was 10 m/s.
The test is simulated using the aforementioned approach with
an element size of 2.0 mm. Fig. 9 shows the relation between the
force measured at the supports and the displacement measured.
The numerical results correlate well with the experimental data as
seen in Fig. 9.

4.6. Meshing
5.2. Results of crash-box optimization
Considering that the thickness is small in comparison to the
other dimensions, shell elements are used to model the walls
of the crash-box. The element type used in the model is S4R,
a 4-node quadrilateral shell element with reduced integration and
a large-strain formulation. The elements account for both nite
membrane strains and arbitrary large rotations. These elements
allow transverse shear deformation. Thick shell theory is applied
for thick shell elements while Kirchhoff theory is applied for thin
shell elements. Simpson integration rule is used with ve integration points through the thickness.

4.7. Contact modeling

The contact between the crash-box and the rigid plate and
block is modeled using surface-to-surface contact modeling.
However, in order to account for self-contact of the crash-box,
the general contact algorithm is used in the nite element model.
The value of the friction coefcient is chosen to be 0.15 for all the
contact interactions.

The shape optimization problem is solved using various combinations of weighting factors, wi, in Eq. (1), either only one term
in the objective function is used or more than one term, that
means single and multi-objective optimizations are carried out.
For single objective optimizations, the energy and mass constraints are not used. The optimization process is repeated for a
constant thickness of 2.5 and 2.0 mm as well as variable thickness.
The results are obtained for 6061-T6 aluminum alloy. The material
properties provided by Corbett [29] are used in the simulations.
The optimal shapes obtained by the algorithm and the optimal
values of the optimization variables are given in Tables 1 and 2
respectively. Table 3 provides the results for the circular benchmark design. A comparison of the results for the optimum shapes
as well as the chosen benchmark shapes is given in Table 4 in
terms of the normalized values. Table 5 shows the deformation of
the optimal crash-boxes during the course of collision.
It is noteworthy that the mass constraint (m r 156 g) and the
total energy constraint (E acc Z 6672 J) do not become active in the
multi-objective optimizations except for the design obtained with

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

309

Table 1
Optimum base shapes of the crash-box for AL 6061-T6.

w1;2;3 0:5; 0:1; 0:4 and 2.0 mm thicknesses. In the singleobjective optimizations, where the mass and energy constraints
are not imposed, the masses of the optimal designs are also less
than that of the circular crash-box except for the cases in which
variance in the deformation is minimized with constant thickness.
In contrast, if one tried to increase the static strength of the
column, given the spacing limitations, one would increase the
cross-sectional area, which would in turn decrease the stress.
However, this would also increase its rigidity and thus reduce its
capacity to absorb impact energy. In that case, the rest of the

vehicle, which is represented by the lumped-parameter model in

the present study, would take a signicant portion of the impact
energy. Too exible columns, on the other hand, would quickly
collapse and thus its energy absorbing capacity would be quickly
consumed. One should recognize that simple rules of thumb will
not work in complex problems. The optimization algorithm nds
the optimal shapes that cannot be intuitively known because of
the complexity of the deformation behavior.
The optimal shaped tube that can absorb the largest strain
energy per unit mass is obtained if only the rst term of the

310

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

Table 2
Optimal values of the optimization variables dening the longitudinal prole.
Weights

r1

r2

r3

l1 (mm)

l2 (mm)

l3 (mm)

t (mm)

t 2.5 mm
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

0.8840
0.8931
1.4253
0.8616
0.8608

0.9902
0.8426
0.9691
0.8609
0.8610

0.9333
1.1138
0.8249
0.9859
0.9851

0.9879
0.8723
1.1216
1.0205
1.0203

20.0517
26.9095
15.7988
13.3882
13.3473

22.1671
15.3196
8.8586
17.5132
17.4272

19.7076
24.5560
17.1982
20.8555
20.9374

2.5
2.5
2.5
2.5
2.5

t 2.0 mm
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

0.8294
0.7570
1.0776
1.1313
0.8895

0.9420
0.8798
1.0270
1.0274
0.9881

1.0167
1.1334
1.0406
1.0994
0.9318

1.0844
1.1106
0.8214
0.9873
0.9962

21.0814
10.0435
24.3266
14.6318
19.8434

18.4990
20.3362
9.0648
23.4376
22.5509

22.9199
7.5634
8.7931
28.3280
21.0224

2.0
2.0
2.0
2.0
2.0

Variable thickness
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

0.8800
0.7484
1.2317
0.8056
0.9035

0.9854
0.9715
0.9108
1.0007
0.9946

0.9334
0.9364
0.8244
0.9895
0.9445

0.9892
0.8969
0.8841
1.0589
0.9995

20.0671
17.6808
24.3376
17.0277
19.6417

22.1775
16.0072
12.8680
29.0038
22.5352

19.6712
12.2644
14.1382
27.0100
21.2395

2.3721
1.1077
1.0143
2.7044
2.0195

Bench marks
Circular (D 70)
Rectangular (70  100)

1.0
1.0

1.0
1.0

1.0
1.0

1.0
1.0

2.5
2.5

Table 3
Results for the circular benchmark design.
Weights

Specic energy (J/g)

Variance

Fmax (kN)

F max =t (MN/s)

Mass (g)

Accumulated energy (J)

Circular (D 70)

42.8

1500

117.5

979.1

156

6672

Table 4
Comparison of the optimal and benchmark shapes in terms of normalized values.
Specic
energy

Variance Fmax F max =t Mass Accumulated

energy

t 2.5 mm
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

1.61
0.41
1.02
1.45
1.46

4.79
0.56
1.48
1.85
1.98

0.83
0.95
0.51
0.68
0.68

0.53
0.81
0.12
0.39
0.39

0.72
1.46
0.85
0.82
0.82

1.16
0.59
0.86
1.19
1.19

t 2.0 mm
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

1.63
0.45
0.87
1.11
1.57

3.71
0.46
0.65
1.64
5.94

0.46
0.87
0.31
0.54
0.84

0.29
0.50
0.08
0.32
0.52

0.44
1.23
0.87
0.90
0.72

0.71
0.56
0.76
1.00
1.13

Variable thickness
w1;2;3 1; 0; 0
w1;2;3 0; 1; 0
w1;2;3 0; 0; 1
w1;2;3 0:5; 0:1; 0:4
w1;2;3 0:5; 0:0; 0:5

2.34
0.66
0.72
1.33
1.80

5.85
0.29
0.31
1.91
4.37

0.77
0.17
0.10
0.89
0.64

0.47
0.10
0.02
0.73
0.39

0.67
0.36
0.36
0.85
0.59

1.56
0.24
0.26
1.12
1.06

1.0
0.69

1.0
1.02

1.0 1.0
0.47 0.58

1.0
1.54

1.0
1.06

Weights

Bench marks
Circular (D 70)
Rectangular
(70  100)

objective function is considered i.e. w1 1, w2 0, and w3 0.

Even though, in these runs, the third term, F max , and F max =t, is
not considered, the resulting values are much better than that of
the benchmark case. If only the second term is optimized, that
means if the variance in the deformation is minimized (w1 0,

w2 1 and w3 0), unacceptably low specic energy values are

obtained. If the third term in Eq. (1) is minimized (w1 0, w2 0,
and w3 1), the algorithm typically chooses larger taper angles
and deeper circumferential ribs to minimize the jerk effect, but
specic energy absorption is not improved. Besides, the total
absorbed energies are much lower than that of the benchmark
cases. When the three terms are considered with w1 0:5,
w2 0:1, and w3 0:4, more balanced results are obtained
(Tables 1 and 5). Recognizing that the contribution of low variance
in deformation to crashworthiness is questionable, the crash-box
is optimized considering only the rst and the third terms
(w1 0:5, w2 0 and w3 0:5). For this case, F max and F max =t
values are improved in comparison to the rst case (w1 1, w2 0
and w3 0). The crash-box designs with optimized thickness
outperform the corresponding designs optimized using the same
weights but with a predened thickness.
Optimum design of the crash-box for high-speed collisions
increases the crashworthiness of the vehicle, thus providing
increased protection for passengers; however the crash-box must
also satisfy the requirements for low-velocity impact so that it
sustains limited deformation and prevents damage to the remaining parts of the vehicle. This requirement can be integrated to the
optimization process in two ways. First, it can be used as a
constraint, but this requires additional simulations of the lowvelocity crash event at each iteration and thus signicantly
increases the computational times. Second, recognizing that many
near global optimum designs are generated by the optimization
process, in which the objective function values are very close, the
optimization process is conducted without any regard for lowvelocity impact performance, then the constraint is checked only
for the optimal designs obtained at the end of the optimization
process. Because it is more feasible and cost-effective, the second
approach is adopted in this study. A standard low-velocity crash

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

311

Table 5
Comparison of the deformed shapes.

test, the Research Council for Automobile Repairs (RCAR) oblique

frontal impact test, is used. In the RCAR test, a car with a velocity of
16 km/h hits a rigid wall oriented by 801 with respect to the
direction of movement with 40% offset. The optimal designs
presented in Table 2 and 4 as well as some near optimal designs
are checked for low-velocity performance. Some of them are found
to be better regarding low-velocity oblique impact requirements
in comparison to the benchmark design, while some others are
found to be worse. Table 6 gives a comparison of the performances
of the circular benchmark design and the optimal designs of
variable thickness having weights of w1;2;3 1; 0; 0 and
w1;2;3 0:5; 0:1; 0:4. The optimal design of w1;2;3 1; 0; 0 shows a

worse performance in comparison to the benchmark. On the other

hand, the optimal design of w1;2;3 0:5; 0:1; 0:4 is better than the
circular design in terms of specic energy absorption (SEA), Fmax
and Fmean and similar in performance in terms of accumulated
energy and stroke efciency, which is the displacement of the
crash-box over its initial length. If the comparison is made for
the circular design with the same mass, the performances of the
optimal designs are far better in terms of energy absorption
capability. In conclusion, one may state that optimizing a part for
maximum crashworthiness does not ensure satisfaction of the
low-velocity requirements, but one may nd a design among the
optimal designs that satises these requirements.

312

N. Tanlak, F.O. Sonmez / Thin-Walled Structures 84 (2014) 302312

Table 6
Oblique impact results of selected designs.
Weights

SEA (J/g)

Fmax (kN)

Fmean (kN)

Mass (g)

Stroke efciency

Accumulated energy (J)

Circular (D 70; t 2.5 mm)

Circular (D 70; t 1.683 mm)
Circular (D 70; t 2.15 mm)
w1;2;3 1; 0; 0
w1;2;3 0:5; 0:1; 0:4

39.0
27.2
38.1
37.0
43.7

106.4
59.8
85.7
70.5
99.8

80.4
31.6
65.4
47.3
75.2

156
105
134
105
134

0.76
0.83
0.79
0.92
0.77

6081.3
2854.1
5106.8
3885.3
5859.4

6. Conclusion
In the present study, the shape of a crash-box is optimized
under impact conditions very similar to EuroNCAP tests to maximize its crashworthiness. The parameters dening the shapes of
the cross-sectional and longitudinal proles as well as the thickness are taken as the optimization variables. The crash-box is
modeled as a deformable body in full detail. In order to reduce the
computational time, a lumped-parameter model is developed to
mimic the behavior of the main vehicle body using a parametric
system identication method. The parameters of the lumpedparameter car model are tuned to reect the response of the car
predicted by a full car model.
The resulting optimum shapes depend highly on the formation
of the objective function, the number of variables, and the range of
values that can be assigned to the variables. The optimum shapes
obtained in this study show signicant improvement over the
benchmark designs. By choosing different values for the weighting
factors of the terms in the objective function, different optimal
shapes are obtained. The best results are obtained, if only the rst
and the third terms are considered. When the thickness is allowed
to vary, the crashworthiness of the crash-box is improved.
The designs optimized for maximum crashworthiness at highvelocity impact may not satisfy low-velocity requirements, but it is
possible to nd near global optimums that show satisfactory
performance for low-velocity impacts.
Acknowledgments
Scientic Research Projects of Bogazici University (Code number 5893) is gratefully acknowledged for supporting this research.
References
[1] Yamazaki K, Han J. Maximization of the crushing energy absorption of tubes.
Struct Optim 1998;16:3746.
[2] Lee S-H, Kim H-Y, Oh S-I. Cylindrical tube optimization using response surface
method based on stochastic process. J Mater Process Tech 2002;130131:
490496.
[3] Sheriff NM, Gupta NK, Velmurugan R, Shanmugapriyan N. Optimization of thin
conical frusta for impact energy absorption. Thin-Walled Struct 2008;46:
65366.
[4] Avalle M, Chiandussi G. Optimisation of a vehicle energy absorbing steel
component with experimental validation. Int J Impact Eng 2007;34:84358.
[5] Hou S, Li Q, Long S, Yang X, Li W. Multiobjective optimization of multi-cell
sections for the crashworthiness design. Int J Impact Eng 2008;35:135567.
[6] Hou S, Li Q, Long S, Yang X, Li W. Design optimization of regular hexagonal
thin-walled columns with crashworthiness criteria. Finite Elem Anal Des
2007;43:55565.

[7] Lnn D, Fyllingen O, Nilssona L. An approach to robust optimization of impact

problems using random samples and meta-modelling. Int J Impact Eng
2010;37:72334.
[8] Acar E, Guler MA, Gerceker B, Cerit ME, Bayram B. Multi-objective crashworthiness optimization of tapered thin-walled tubes with axisymmetric
indentations. Thin-Walled Struct 2011;49(1):94105.
[9] Qi C, Yang S, Dong F. Crushing analysis and multiobjective crashworthiness
optimization of tapered square tube under oblique impact loading. ThinWalled Struct 2012;59:10319.
[10] Liu Y. Optimum design of straight thin-walled box section beams for
crashworthiness analysis. Finite Elem Anal Des 2008;44:13947.
[11] Liu Y. Crashworthiness design of multi-corner thin-walled columns. ThinWalled Struct 2008;46:132937.
[12] Yang S, Qi C. Multiobjective optimization for empty and foam-lled square
columns under oblique impact loading. Int J Impact Eng 2013;54:17791.
[13] Zarei HR, Krger M. Multiobjective crashworthiness optimization of circular
aluminum tubes. Thin-Walled Struct 2006;44:3018.
[14] Kim CH, Arora Jasbir S. Development of simplied dynamic models using
optimization: application to crushed tubes. Comput Method Appl M
2003;192:207397.
[15] Santosa S, Wierzbicki T. Crash behavior of box columns lled with aluminum
honeycomb or foam. Comput Struct 1998;68:34367.
[16] Zarei HR, Krger M. Optimum honeycomb lled crash absorber design. Mater
Design 2008;29:193204.
[17] Zarei HR, Krger M. Optimization of the foam-lled aluminum tubes for crush
box application. Thin-Walled Struct 2008;46:21421.
[18] Hanssen AG, Langseth M, Hopperstad OS. Optimum design for energy
absorption of square aluminium columns with aluminium foam ller. Int J
Mech Sci 2001;43:15376.
[19] Ahmad Z, Thambiratnam DP. Dynamic computer simulation and energy
absorption of foam-lled conical tubes under axial impact loading. Comput
Struct 2009;87:18697.
[20] Sun G, Li G, Hou S, Zhou S, Li W, Li Q. Crashworthiness design for functionally
graded foam-lled thin-walled structures. Mater Sci Eng A 2010;527:19119.
[21] Yin H, Wen G, Hou S, Chen K. Crushing analysis and multiobjective crashworthiness optimization of honeycomb-lled single and bitubular polygonal
tubes. Mater Design 2011;32:444960.
[22] Bi J, Fang H, Wang Q, Ren X. Modeling and optimization of foam-lled thinwalled columns for crashworthiness designs. Finite Elem Anal Des 2010;46:
698709.
[23] Tarlochan F, Samer F, Hamouda AMS, Khalid K, Ramesh S. Design of thin wall
structures for energy absorption applications: enhancement of crashworthiness due to axial and oblique impact forces. Thin-Walled Struct 2013;71:717.
[24] http://www.euroncap.com/tests/frontimpact.aspx.
[25] Deb A, Naravane A, Chirwa EC. An offset rigid barrier-based test: equivalence
to the insurance institute for highway safety frontal offset impact safety test.
Int J Crashworthines 2006;11(4):28190.
[26] http://www.ncac.gwu.edu/vml/models.html.
[27] Johnson GR, Cook WH. A constitutive model and data for metals subjected to
large strains, high strain rates, and high temperatures. In: Proceedings of the
7th international symposium ballistics; 1983.
[28] Johnson GR, Cook WH. Fracture characteristics of three metals subjected to
various strains, strain rates, temperatures and pressures. Eng Fract Mech
1985;21(1):3148.
[29] Corbett BM. Hypervelocity impact damage response and characterization of
thin plate targets at elevated temperatures [Ph.D. thesis]. University of
Denver; 2008.