This art icle was downloaded by: [ Gazi Universit y]

On: 07 December 2012, At : 20: 15

Publisher: Taylor & Francis
I nforma Lt d Regist ered in England and Wales Regist ered Number: 1072954 Regist ered office: Mort imer House,
37- 41 Mort imer St reet , London W1T 3JH, UK
Int ernat i onal Journal of Crashwort hi ness
Publ i cat i on det ai l s, i ncl udi ng i nst r uct i ons f or aut hor s and subscr i pt i on i nf or mat i on:
ht t p: / / www. t andf onl i ne. com/ l oi / t cr s20
Crashwort hi ness opt i mi sat i on of vehi cl e st ruct ures
wi t h magnesi um al l oy part s
Andr ew Par r i sh
a
, Masoud Rai s- Rohani
b
& Al i Naj af i
b
a
Depar t ment of Aer ospace Engi neer i ng, Mi ssi ssi ppi St at e Uni ver si t y, MS, 39762, USA
b
Depar t ment of Aer ospace Engi neer i ng and Cent er f or Advanced Vehi cul ar Syst ems,
Mi ssi ssi ppi St at e Uni ver si t y, MS, 39762, USA
Ver si on of r ecor d f i r st publ i shed: 13 Jan 2012.
To ci t e t hi s art i cl e: Andr ew Par r i sh, Masoud Rai s- Rohani & Al i Naj af i ( 2012) : Cr ashwor t hi ness opt i mi sat i on of vehi cl e
st r uct ur es wi t h magnesi um al l oy par t s, Int er nat i onal Jour nal of Cr ashwor t hi ness, 17: 3, 259- 281
To l i nk t o t hi s art i cl e: ht t p: / / dx. doi . or g/ 10. 1080/ 13588265. 2011. 648518
PLEASE SCROLL DOWN FOR ARTI CLE
Full t erms and condit ions of use: ht t p: / / www. t andfonline. com/ page/ t erms- and- condit ions
This art icle may be used for research, t eaching, and privat e st udy purposes. Any subst ant ial or syst emat ic
reproduct ion, redist ribut ion, reselling, loan, sub- licensing, syst emat ic supply, or dist ribut ion in any form t o
anyone is expressly forbidden.
The publisher does not give any warrant y express or implied or make any represent at ion t hat t he cont ent s
will be complet e or accurat e or up t o dat e. The accuracy of any inst ruct ions, formulae, and drug doses should
be independent ly verified wit h primary sources. The publisher shall not be liable for any loss, act ions, claims,
proceedings, demand, or cost s or damages what soever or howsoever caused arising direct ly or indirect ly in
connect ion wit h or arising out of t he use of t his mat erial.
International Journal of Crashworthiness
Vol. 17, No. 3, June 2012, 259281
Crashworthiness optimisation of vehicle structures with magnesium alloy parts
Andrew Parrish
a
, Masoud Rais-Rohani
b
and Ali Naja
b
a
Department of Aerospace Engineering, Mississippi State University, MS 39762, USA;
b
Department of Aerospace Engineering
and Center for Advanced Vehicular Systems, Mississippi State University, MS 39762, USA
(Received 13 August 2011; nal version received 7 December 2011)
This paper explores the effects of replacing the baseline steel with lightweight magnesium alloy parts on crashworthiness
characteristics and optimumdesign of a full-vehicle model. Full frontal, offset frontal and side crash simulations are performed
on a validated 1996 Dodge Neon model using explicit nonlinear transient dynamic nite element analyses in LS-DYNA to
obtain vehicle responses such as crash pulse, intrusion distance, peak acceleration and internal energy. Twenty-two parts of
the vehicle body structure are converted into AZ31 magnesium alloy with adjustable wall thickness while the remaining parts
are kept intact. The magnesium alloy material model follows a piecewise linear plasticity law considering separate tension
and compression properties and maximum plastic strain failure criterion. Six different metamodelling techniques including
optimised ensemble are developed and tuned for predictions of crash-induced responses within the design optimisation
process. The crashworthiness optimisation problem is solved using the sequential quadratic programming method with most
accurate surrogate models of structural responses considering both constrained single- and multi-objective formulations. The
results show that under the combined crash scenarios with the selected material models and design constraints, the vehicle
model with magnesium alloy parts can be optimised to maintain or improve its crashworthiness characteristics with up to
50% weight savings in the redesigned parts.
Keywords: crashworthiness; design optimisation; magnesium alloy; material model; material substitution; nite element
analysis; metamodels
1. Introduction
Stricter regulations on fuel economy and growing concerns
over automobile emissions have led to an increased fo-
cus on vehicle weight reduction through the application of
lightweight materials, especially in auto body structures.
The fuel economy improvements gained through weight re-
duction must also be balanced with the requirements for
crash safety. Therefore, a simple one-to-one substitution of
the original parts in absence of a thorough consideration
of the overall crash characteristics may not result in a safe
design. Figure 1 describes the overall approach considered
in this study, where material substitution and design opti-
misation techniques are combined to reduce vehicle weight
while maintaining the vehicles crash performance or crash-
worthiness.
The U.S. Department of Energy has sponsored many
studies in the area of Automotive Lightweighting Materi-
als [7] under the Vehicle Technologies Program to promote
the application of lightweight materials such as magnesium
alloys and improvements in structural design and manufac-
turing for lightweighting of vehicle structures. The strategic
vision for magnesium [32] advocates opportunities for in-
corporating more magnesium parts into automobile design.
It suggests that magnesium is a good candidate material

Corresponding author. Email: Masoud@ae.msstate.edu

to replace the heavier steel parts due to magnesiums high
strength-to-weight ratio, potential reduction in part count
due to larger castings and the ability to tune magnesium
parts to frequencies related to noise, vibration and harsh-
ness (NVH). That study also indicates the need for more
research to address corrosion and manufacturing issues as
well as brittle fracture at high strain rates before magnesium
is ready for widespread application in auto body structures.
Crashworthiness is the ability of the structure to protect
occupants during crashes. Performing real crash tests to
obtain necessary data for crashworthiness optimisation is
not feasible. Finite element (FE) crash simulations are often
used in place of physical experiments to obtain responses
needed for optimisation or exploratory studies.
Critical responses for crashworthiness optimisation are
often selected to measure or estimate the likelihood of
occupant injury. The U.S. Department of Transportation
places requirements of crash responses felt by test dum-
mies through the Federal Motor Vehicle Safety Standards
(FMVSS) [24]. FMVSS contain specics for multiple crash
scenarios as well as for different classes of vehicles. Crash
scenarios in these requirements include full frontal im-
pact (FFI), offset frontal impact (OFI), side impact (SIDE),
rollover or roof impact and rear impact.
ISSN: 1358-8265 print / ISSN: 1754-2111 online
C
2012 Taylor & Francis
http://dx.doi.org/10.1080/13588265.2011.648518
http://www.tandfonline.com
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

260 A. Parrish et al.
Figure 1. Considerations for a lighter weight design.
Obtaining these values in FE simulations requires an
accurate crash occupant model set correctly in the vehicle
model. Alternative responses must be chosen if an FE
model with an occupant is not available. Internal energy ab-
sorption, intrusion distance and peak acceleration measured
at selected locations are often used as substitutes. These
responses focus on the vehicles crash performance rather
than occupant injury metrics. Fang et al. [8] used internal
energy absorption of selected components at two time
steps along with peak engine top acceleration. Liao et al.
[15] used a combination of occupant- and vehicle-based
responses including an integration of the deceleration curve
and intrusion distance at the toeboard. Intrusion distance of
the front panel for frontal impacts and intrusion distance of
the door for side impacts were considered by Fang et al. [9].
Several studies have considered different approaches
for crashworthiness optimisation. For example, Akkerman
et al. [3] used shape and sizing optimisation to improve
crashworthiness of an automotive instrument panel. Fang
et al. [9] considered sizing optimisation of vehicle struc-
tures whereas Rais-Rohani et al. [27] used shape and sizing
optimisation to improve crashworthiness of a vehicle
by altering the geometry of the side rails. Alternatively,
Mozumder et al. [20] examined the application of cellular
automata for topology-based crashworthiness optimisation.
Anumber of studies have used multi-objective formula-
tion to solve crashworthiness optimisation problems. Fang
et al. [8] improved crash performance while maintaining
weight in a FFI whereas Fang et al. [9] improved crash per-
formance considering multiple impact scenarios. In another
study, Liao et al. [15] solved a two-step optimisation prob-
lem minimising weight and vehicle crash responses before
optimising the occupant restraint systembased on occupant
injury criteria.
Simulation runtime is a major obstacle in optimisation
studies. In many cases, hundreds or thousands of function
evaluations may be required for solving a design optimi-
sation problem depending on the method selected as well
as the numbers of design variables and constraints used.
The computational burden is considerably reduced when
the objective and constraint responses are described by an-
alytical functions but the problem becomes signicantly
more challenging when high-delity FE crash simulations
are required to obtain the response values. We have found
that a single crash simulation using a full-vehicle model
takes anywhere between one and a half and three hours, de-
pending on the crash scenario, using 4 six-core Intel X5660
processors with 48 GB of total RAM. Another important
consideration is the non-smooth (noisy) behaviour of some
of the crash responses that could pose problems in appli-
cation of gradient-based optimisation methods. Hence, ap-
proximate mathematical models or metamodels are needed
to overcome the difculties associated with computational
cost and noisy responses.
Metamodels or surrogate models that approximate func-
tion values are widely used in automotive crashworthiness
studies [9, 8, 15, 30] to provide approximate responses at a
design point. Responses predicted using a metamodel will
have some error associated with them when compared to
simulation or test results at the same design point, but these
errors can be reduced by selecting appropriate values for
the corresponding tuning parameters.
This study explores the application of AZ31 magne-
sium alloy for vehicle body structures. Effects of material
substitution on crash responses at different vehicle sites are
studied considering three major crash scenarios (i.e. FFI,
OFI and SIDE). Tension and compression stressstrain re-
sponses are included to enhance the accuracy of the material
model. A maximum plastic strain criterion is used to iden-
tify failed elements in FE simulations using LS-DYNA.
No previous study has been found comparing steel and
magnesium designs in full-scale automotive crash scenar-
ios. This study provides an example of using equivalent
toughness as a way of determining substituted material
thickness with a magnesium material model based on sepa-
rate stressstrain curves for tension and compression. Strain
rate dependency in crash simulations is also explored. Se-
lected results from designs with the steel and magnesium
structures are compared. The study also introduces optimi-
sation techniques to determine the weight savings of a mag-
nesium design with crashworthiness similar to the heavier
steel design.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 261
Figure 2. Dodge neon FE model (a) exterior view showing elements and (b) with exterior panels removed.
The remaining portion of the paper is organised as fol-
lows: Section 2 introduces the FE model, crash scenarios
and design responses. Section 3 discusses the magnesium
for steel material substitution. Section 4 presents the meta-
model construction. Section 5 shows the optimisation for-
mulations and results, with concluding remarks and future
work appearing in Section 6.
2. Vehicle model and crash simulations
A full-scale 1996 Dodge Neon model, FE model (v07
LS-DYNA) developed at the National Crash Analysis
Center (NCAC) is used in this study. This model, shown
in Figure 2, contains no interior panels or seats and is
publicly available from the NCAC website [23]. Most
structural components are made of steel dened using
Mat 24 (piecewise linear plasticity) in LS-DYNA [17] and
modelled using Belytchko-Tsai reduced integration shell
elements. The vehicle is made up of 270,768 elements in
336 parts and the total vehicle mass is 1333 kg. This model
was validated by NCAC for an FFI scenario [22].
This vehicle model is used for FFI, SIDE and OFI sce-
narios by following the FMVSS for impact velocity and
impact location/angle. The FFI scenario models an impact
into a rigid wall at a speed of 56 km/h. Velocity for vali-
dation and testing for SIDE was 52 km/h to coincide with
test data. The impact occurs at a 27

angle from an impact

vehicle with honeycomb material simulating the front of
another automobile. The OFI simulations were validated at
60 km/h based on available test data and used for this study
at 56 km/h to coincide with FFI simulations. The Neon
model impacts a honeycomb material barrier in front of a
rigid wall at 40% offset. Figure 3 shows the Neon model in
each crash simulation.
Simulation-based acceleration curves at the left rear sill
for OFI and at the middle of the B-pillar for SIDEare shown
in Figure 4 and compared to physical test results [14, 19].
These curves show that the general shapes of the curves are
the same but peak values may differ as a result of ltering
and the method used to capture the data. A Butterworth
lter with a frequency of 60 Hz was used for the simulation-
based results. The location of these observation points was
determined by the position of accelerometers in the actual
testing.
We identied 22 parts as shown in Figure 5 with sig-
nicant contributions to energy absorption and structural
stiffness in our preliminary investigation. These parts ac-
count for approximately 40% of the energy absorption in
all three crash scenarios and have a mass of 105 kg or 8%
of the vehicle mass at 1333 kg.
Responses considered include the intrusion distances
at the toeboard and dashboard for FFI and OFI and at the
Figure 3. Crash scenarios (a) FFI, (b) SIDE, and (c) OFI scenarios.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

262 A. Parrish et al.
Figure 4. Acceleration for model validation (a) OFI a-accel. at left rear sill and (b) SIDE y-accel. at middle b-pillar.
door for SIDE (Int Toe, Int Dash and Int Door), resul-
tant acceleration at a location on the B-pillar in all three
scenarios (Accel) and internal energy absorption of the se-
lected parts in all three scenarios (Int Eng). These responses
were chosen because of their relevance to occupant safety
and the acceleration response location was chosen to be
near the approximate head location of an occupant during
a crash. Mass is determined by dividing the initial mass of
the baseline part by the initial thickness. This coefcient is
multiplied by a new thickness for that part to determine the
parts new mass. Selected response locations can be seen in
Figure 6.
Intrusion distance was calculated by measuring the dis-
tance between 20 nodes at each response location and a ref-
erence node on the opposite side of the car before and after
the crash and then nding the difference. An average over
the 20 nodes was used as an intrusion distance response. A
Butterworth lter at 60 Hz was applied to the acceleration
curve at 20 nodes in each direction, the resultant found, an
average over the 20 nodes taken and the maximum value
found to represent an acceleration response. The internal
energy response is the sum of the internal energy at the end
of the simulation for the selected parts.
Figure 5. Selected vehicle components and associated design
variables.
The simulated response values determined using the
methods described will be referred to as the true or actual
response in the remainder of this paper.
3. Material substitution
The goal of material substitution here is to replace some of
the baseline steel parts of the Neon body with those made
of a lightweight magnesium alloy. Magnesium alloys are
being seriously considered for auto body applications by
the automotive industry as a way of reducing the vehicle
weight and improving the fuel economy. The manufacturing
processes being considered for magnesium parts include
sheet forming, extrusion and die-casting [32]. In this paper,
we are assuming that the parts are made of sheet formed
with AZ31 magnesium alloy.
Experimental studies show that magnesium alloys
demonstrate different behaviours under tension, compres-
sion and shear, and that the failure points for tension and
compression are different due to anisotropic plastic defor-
mation. Magnesium alloys also show considerable sensitiv-
ity to strain rate.
In a vehicle crash, plastic deformation is considered
as the major mechanism for energy absorption in ductile
metallic parts. Studies at the component level show that the
localised regions in the deformed component experience
different load paths [26]. In this study, Material Type 124
(MAT 124) in LS-DYNA[17] is used to model the material
behaviour of magnesiumparts based on the work by Wagner
et al. [34]. This material model is capable of making a
distinction between tension and compression behaviour by
considering the piecewise linear isotropic hardening model.
The strain rate sensitivity is considered through
Cowper-Symonds model that scales the instantaneous yield
stress as:

0
= 1 +
_

C
_
1/p
(1)
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 263
Figure 6. Selected response locations.
where and
0
are the von-Misses stress at equivalent
strain rate and quasi-static yield stress, respectively. The
strain rate sensitivity parameters p and C for tension and
compression are assumed to be identical with C = 24,124
and p = 3.09 for AZ31 [21].
The stressstrain curves for AZ31 under tensile loading
were obtained from experimental data. Since similar data
was unavailable for AZ31 under compressive loading,
the ratio of AM30 magnesium alloy extrusion under
compression and tension along with AZ31 tension data
were used to approximate the stressstrain curve for AZ31
compression as:
AZ31(RD)
comp
=
AZ31 (RD)
ten
AM30 (ED)
ten
AM30(ED)
comp
(2)
where RD and ED represent the rolling and extruded di-
rections, respectively, under tension (ten) and compression
(comp). AZ31 properties in the transverse direction (TD)
for tension and AZ31 (RD) for compression were used
to dene the MAT 124 material card. Figure 7 shows the
stressstrain curves for AM30 and AZ31.
Material failure in the model is captured through fail-
ure strain limit, which is assumed to be identical for both
tension and compression. However, in the material used in
this study, tensile failure strain is higher than compressive
value. In order to model this difference, a softening por-
tion is added to the compression curve. By having three
integration points through the shell thickness, the energy
absorption as a result of bending can be captured well by
using different curves for the region of the shell that is in
tension or compression. In using MAT 124, we are exclud-
ing the presence and effect of anisotropic texture.
Material substitution was performed by changing the
22 parts from steel to magnesium and leaving the other
parts as is in the original model. Whereas seven different
steel materials are used to dene the 22 selected parts in
the original Neon model, the same magnesium material is
used for all of the parts. To establish a baseline model of
Neon with magnesium parts, the thicknesses of selected
parts were adjusted so as to maintain the same total internal
energy absorption as the original steel parts. This was done
by using the relationship between toughness and thickness
given as:
t
Mg
= t
St
T
St
T
Mg
(3)
where t
Mg
and t
St
are the thicknesses of magnesium and
steel parts, respectively, and T
Mg
and T
St
are the toughness
values of steel and magnesium derived from integrating the
Figure 7. Magnesium quasi-static stress-strain curves at 22

C.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

264 A. Parrish et al.
Figure 8. Tensile loading on the single element model.
stressstrain curves for each material from 0 to 0.3 strain.
An average of the thicknesses found using the toughness
under compression and the toughness under tension was
used. Thicknesses of the selected parts are found in
Table 1 for the steel and magnesium designs.
To examine the importance of strain rate dependency
for AZ31 with MAT 124 representation, the deformation
response of a single shell element under tension and com-
pression was examined. To simulate a constant strain rate
loading, a prescribed displacement was applied on one edge
of the element while the other edge was held xed as shown
in Figure 8. The displacement was applied by considering
the actual time step size of the model using the relationship:
x = xl
0
(exp( (t +t ) 1) (4)
where t is some fraction of the solution time step and x
and xl
0
are the displacement and initial element length, re-
spectively. Therefore, the amount of applied displacement is
varied within each time step increment. Figure 9 shows the
stressstrain response of a single element at different strain
rates for tensile and compressive loading conditions. This
material model also allows the denition of plastic strain
to failure, which is important in modelling the response of
parts made of magnesium alloy.
To examine the importance of including strain rate sen-
sitivity parameters for magnesium alloy parts of the Neon
model, the effective value of the strain rate tensor at each
integration point for all of 22 parts were evaluated with the
Table 1. Part thicknesses (in mm) for baseline designs.
Part Design variable St base Mg base
A-pillar x
1
1.611 2.597
Front bump x
2
1.956 5.975
Firewall x
3
0.735 1.072
Front oor panel x
4
0.705 1.136
Rear cabin oor x
5
0.706 1.138
Outer cabin x
6
0.829 1.366
Cabin seat reinf. x
7
0.682 1.099
Cabin mid rail x
8
1.050 1.692
Shotgun x
9
1.524 3.620
Inner side rail x
10
1.895 3.966
Outer side rail x
11
1.522 3.186
Side rail exten. x
12
1.895 3.966
Rear plate x
13
0.710 1.144
Roof x
14
0.702 1.157
Susp. frame x
15
2.606 5.342
results shown in Table 2. For parts that have left and right
components such as the side rail, only that with the largest
measured strain rate is shown in Table 2. The strain rate
changes during the crash simulation; therefore, the maxi-
mum values are checked throughout the crash duration at
0.005 s time intervals for total of 32 points for each element
integration point. It is seen that the maximumvalue of effec-
tive strain rate varies from 12/s to 447/s within the different
parts considered in these simulations. However, the average
of the maximum strain rates in the elements located in the
localised regions are below 10/s. This shows that the strain
rate dependency of the material can be neglected in the car
crash scenarios considered here. Table 2 also lists the parts
that experienced failure based on the failure strain limit of
38% as dened in the material card. It was observed that
the baseline model experiences a severe failure in the outer
cabin components. Additionally, the failure is affected by
the crash scenario. For instance, suspension frame has some
failed elements in OFI whereas it does not fail in the SIDE
or FFI cases.
Table 3 compares the total internal energy absorption
of the selected parts for each crash scenario using 19%
Figure 9. Material behaviour (a) under tensile load and (b) under compressive load for single element model.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 265
Table 2. Effective strain rate and the element failure observed in the base line model with magnesium material.
Maximum effective strain
rate (1/s)
Average of maximum effective
strain rate (1/s) Element failure
Part FFI OFI SIDE FFI OFI SIDE FFI OFI SIDE
A-pillar 73 103 38 3 12 4 No No No
Front bump 142 103 0.5 6 4 N/A No No No
Firewall 269 137 105 9 15 0.2 Yes Yes No
Front oor panel 311 231 158 15 11 2 Yes Yes Yes
Rear cabin oor 53 44 119 2 4 5 No No Yes
Outer cabin 337 223 347 6 15 35 Yes Yes Yes
Cabin seat reinf 12 11 41 2 1 3 No No No
Cabin mid rail 142 86 17 3 1 3 No No No
Shotgun 145 110 46 8 15 10 No No No
Inner side rail 166 216 49 6 10 0.1 Yes Yes No
Outer side rail 132 98 22 5 6 0.2 Yes Yes No
Side rail exten. 113 24 57 3 7 0.001 Yes No No
Rear plate 447 9 118 7 0.5 4 No No No
Roof 39 45 48 0.5 4 2 No No No
Susp. frame 117 39 18 3 0.1 0.5 Yes No No
and 38% plastic strain to failure for magnesium parts and
the baseline steel parts. This table shows that magnesium
with 38% plastic strain to failure has total internal energy
absorption closer to the steel baseline than the 19%.
Simulations of each crash scenario using the mag-
nesium material model dened above were performed.
Figure 10 shows that the acceleration curves at the B-pillar
response location for each scenario generally match in
shape and peak values. Table 4 shows an overview of the
responses for baseline simulations with both materials; see
Figure 5 for a labelled picture of the parts. The magnesium
parts have signicantly less mass than the steel parts
and generally maintain or improve the internal energy
absorption of the replaced parts but the intrusion distances
are signicantly larger than the steel parts. These responses
and designs represent the baseline models for comparison
with the optimum designs found later.
4. Response approximation
Response approximation is necessary in crashworthiness
optimisation as FE crash simulations tend to be very
computationally expensive and many responses have
non-smooth (noisy) behaviour. Surrogate models in the
form of analytical functions can be developed using a
Table 3. Total internal energy of selected parts for steel and
magnesium at 19% and 38% plastic strain to failure.
St Mg 19% Mg 38%
FFI 6.23E + 07 6.03E + 07 6.36E + 07
OFI 3.94E + 07 3.69E + 07 3.94E + 07
SIDE 2.24E + 07 1.94E + 07 2.09E + 07
number of metamodelling techniques. A metamodel is
built using data sets obtained through FE simulations or
physical tests at the training points selected using a design
of experiments (DOE) model. Two DOE types frequently
used for crashworthiness optimisation are Taguchi orthog-
onal arrays [9, 8] and Latin hypercube sampling (LHS)
[15, 25]. In this study we chose LHS because it provides a
more uniform sampling of the design space based on the
specied bounds of the design variables.
The number of training points depends on several
factors. These include the problem type, number of vari-
ables, metamodel technique and the level of effort (cost)
involved in obtaining the desired responses at each training
point. Both Yang et al. [36], using an impact example, and
Fang and Wang [10], using analytic benchmark problems,
Table 4. Comparison of baseline steel and baseline magnesium
designs.
St Mg Mg relative to St
FFI
Int toe (mm) 157 295 87.9%
Int dash (mm) 122 186 52.3%
Accel (gs) 63.5 49.2 22.5%
Int eng (kJ) 62.3 62.3 0.1%
SIDE
Int door (mm) 314 420 33.7%
Accel (gs) 47.9 40.8 14.9%
Int eng (kJ) 22.4 21.4 4.3%
OFI
Int toe (mm) 273 349 27.7%
Int dash (mm) 247 386 56.2%
Accel (gs) 35.0 36.2 3.3%
Int eng (kJ) 39.4 39.2 0.4%
Mass (kg) 105.2 42.7 59.4%
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

266 A. Parrish et al.
Figure 10. Acceleration at b-pillar response location (a) x-dir. FFI, (b), y-dir. SIDE, and (c) x-dir. OFI.
show that metamodels are generally more accurate as the
number of training points increases but this trend is not
true for all functions and metamodel types. In their study,
Wang et al. [35] examined response functions of low to
high nonlinearity consisting of 1101 input variables using
7527 training points, and found that the required number
of response samples for metamodel accuracy depends in
large part on the response considered and not so much on
the number of input variables. For example, they could not
nd a good approximation to a two-dimensional response
even after increasing the number of training points to
1000, whereas in some other examples, they were able
to generate very good surrogates by using three times as
many training points as the input variables.
Different metamodelling techniques have been devel-
oped with different levels of complexity and accuracy.
Turner [31] shows metamodels divided into three main
groups. These groups are: Geometric, containing response
surface models and spline-based models; Stochastic,
containing Kriging models and radial basis functions; and
Heuristic, containing kernel model, frequency domain
methods and neural networks. Jin et al. [13] and Wang
et al. [35] compared several metamodelling techniques
using metrics such as accuracy, robustness, efciency,
transparency, exibility and conceptual simplicity for
benchmark as well as more industry relevant example
problems. Metamodel techniques explored in this study in-
cluded Polynomial Response Surface (PRS), Radial Basis
Function (RBF), Kriging (KR), Support Vector Regression
(SVR) and Gaussian Process (GP). An optimised ensemble
(EN) containing multiple metamodels developed by Acar
and Rais-Rohani [1] was created using the ve models
listed above. This technique was used previously in Acar
and Solanki [2] for a crashworthiness problem and shown
to be more accurate than stand-alone metamodels.
With the thicknesses of selected parts as input variables,
45 training points were found in the range of 50% of the
baseline thicknesses. The baseline design point was added
for a total of 46 design points. Due to similarity of the part
in the right and left sides of the vehicle model, only 15
thickness values are needed for the 22 parts. Three separate
(FFI, OFI, SIDE) LS-DYNA simulations were performed
at each design point to provide the responses necessary for
metamodel construction.
Stand-alone metamodels for each response were
developed and some (i.e. RBF, KR and SVR) were tuned
for maximum accuracy before being used to construct an
optimised ensemble of metamodels. The mathematical
description of each metamodel is provided in Appendix
1 for completeness. The models were tuned by selecting
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 267
Table 5. Metamodel tuning parameters for steel design.
FFI SIDE OFI
Scenario
Response Int toe Int dash Accel Int eng Int door Accel Int eng Int toe Int dash Accel Int eng
RBF
c 1 1 0.2 1 1 0.05 0.05 1 0.95 0.25 0.75
M M TP M M M G IM M TP M
KR
UB, 0.1 0.011 1 1 0.1 0.011 0.1 0.011 0.1 1 0.1
LB, 0.01 0.001 0.001 0.001 0.01 0.001 0.01 0.001 0.001 0.01 0.01
C S Sp E L G L G Sp G S S
RPD 0 1 0 1 0 1 0 0 1 0 0
SVR
c 1 1 100 10 0.01 10 10 10 1 10 1
P 10 2 0.1 2 10 2 5 5 5 0.1 5
Note: M = multiquadric; TP = thin plate; G = Guassian; IM = inverse multiquadric; UB = upper bound; LB = lower bound; C = correlation function;
S = spline; E = exponential; L = linear; Sp = spherical; RPD = regression polynomial degree; P = penalty parameter.
the parameter or combination of parameters that produced
the least error for each response for each material.
Cross-validation generalised mean square error (GMSE)
was used as the error metric. A metamodel is created using
all except one design point and the predicted response
is compared to the actual response at that design point
to measure error. This process is repeated for all design
points and the average is used as the overall error of the
metamodel. GMSE was calculated using:
GMSE
1
N
N

i=1
(f
i


f
i
)
2
(5)
where N is the number of design points, f
i
is the actual
response and

f
i
is the response predicted by the model.
Tables 5 and 6 show the tuned metamodel parameters for
steel and magnesium, respectively, and Table 7 shows the
optimised weight factors for the ensemble.
Normalised GMSE for steel and magnesium can be
found in Tables 8 and 9, respectively. The optimised ensem-
ble metamodel was then used to approximate the responses
at the baseline design point and compared to the simulation
response value at the initial design for both materials. This
served as a baseline error when evaluating the accuracy of
each optimum design.
The maximum per cent error for the steel baseline oc-
curred for the intrusion at the toeboard for OFI with 15%
error. This was followed by 9% error for intrusion at the
dashboard for OFI and 5% error for intrusion at the toe-
board for FFI. The remaining responses for steel had less
than 2% error at the baseline design when predicted with
Table 6. Metamodel tuning parameters for magnesium design.
FFI SIDE OFI
Scenario
Response Int toe Int dash Accel Int eng Int door Accel Int eng Int toe Int dash Accel Int eng
RBF
c 0.05 0.05 0.2 0.05 1 1 1 1 0.15 0.05 1
G G M G M IM M M G M M
KR
UB, 0.011 0.1 0.011 0.011 0.011 0.1 0.011 1 0.1 1 1
LB, 0.01 0.001 0.001 0.001 0.001 0.01 0.001 0.001 0.001 0.01 0.001
C Cu G S E L S Cu Sp S S Cu
RPD 0 0 0 0 1 0 1 1 0 0 1
SVR
c 1 1 10 0.01 10 0.01 10 10 0.01 1 10
P 5 10 0.1 2 2 0.1 0.1 10 10 0.1 2
Note: M = multiquadric; G = Guassian; IM = inverse multiquadric; UB = upper bound; LB = lower bound; C = correlation function; S = spline; E =
exponential; L = linear; Cu = cubic; Sp = spherical; RPD = regression polynomial degree; P = penalty parameter.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

268 A. Parrish et al.
Table 7. Optimised ensemble metamodel weight factors.
FFI SIDE OFI
Int toe Int dash Accel Int eng Int door Accel Int eng Int toe Int dash Accel Int eng
PRS 0 0 0 0 0 0 0 0 0.56 0 0
GP 0.57 0.17 0 0.33 0.48 0 0 0.51 0 0 0
RBF 0.30 0 0 0 0.01 0 0 0 0 0.16 0
KR 0.13 0.59 0.41 0.54 0.11 0.57 0.59 0.14 0.44 0.54 0.58
SVR 0 0.24 0.59 0.13 0.40 0.43 0.41 0.35 0 0.30 0.42
PRS 0 0 0 0 0 0 0 0.01 0 0 0
GP 0 0 0 0 0.34 0 0.14 0 0.07 0.27 0.15
RBF 0.33 0 0 0.09 0.08 0 0 0 0 0 0
KR 0.67 1.00 0.30 0.36 0.55 1.00 0.86 0.57 0.56 0.10 0.31
SVR 0 0 0.70 0.55 0.03 0 0 0.42 0.37 0.63 0.55
Table 8. Normalised GMSE for steel design.
FFI SIDE OFI
Int toe Int dash Accel Int eng Int door Accel Int eng Int toe Int dash Accel Int eng
PRS 1.60 1.41 1.64 1.12 1.56 1.53 1.49 1.31 1.04 2.10 1.79
GP 1.12 2.09 1.57 1.29 1.09 2.26 2.31 1.09 1.49 1.35 1.95
RBF 1.39 1.75 1.32 2.14 1.39 1.60 1.46 1.51 1.26 1.13 1.44
KR 1.13 1.15 1.00 1.12 1.18 1.06 1.12 1.22 1.04 1.05 1.01
SVR 1.34 1.20 1.19 1.13 1.34 1.15 1.19 1.24 1.07 1.31 1.27
ENS 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
the ensemble. The maximum per cent error for the magne-
sium baseline was 4% for the acceleration at mid B-pillar
for FFI. The remaining magnesium responses had less than
2% error.
5. Crashworthiness optimisation problems
5.1. Single-objective
With mass as the objective function, both the steel and
magnesium models are optimised such that the constrained
responses do not exceed the corresponding (material spe-
cic) baseline values. The single-objective (SO) nonlinear
constrained minimisation problem is formulated as:
min F (x)
s.t.
R
j
(x) R
jbase
; j = 1..8
R
j
(x) R
jbase
; j = 9..11
0.5x
base
x 1.5x
base
(6)
where F(x) is the objective function, x is the vector of 15 de-
sign variables (part thicknesses), R
j
(x) is the jth response,
and R
jbase
is the jth response of the baseline model (material
specic) as described by the corresponding optimised en-
semble metamodel. In Equation (6), R
j
, j = 1..8 represents
the intrusion distances at the toeboard and dashboard (Int
Toe, Int Dash) for FFI and OFI and at the door (Int Door)
for SIDE and resultant acceleration (Accel) at a location
on the B-pillar in all three scenarios, whereas R
j
, j = 9..11
represents the internal energy (Int Eng) responses of the
selected parts in all three scenarios.
The optimisation problem in Equation (6) was solved
using the Sequential Quadratic Programming (SQP)
method implemented in the VisualDOC [33] software.
Among the optimisation methods available in VisualDOC,
we chose SQP because of its efciency and accuracy as a
local optimiser. However, to expand the search of feasible
design space, multiple (i.e. eight) initial design points were
Table 9. Normalised GMSE for magnesium design.
FFI SIDE OFI
Int toe Int dash Accel Int eng Int door Accel Int eng Int toe Int dash Accel Int eng
PRS 1.73 1.84 1.50 1.70 1.52 1.07 2.07 1.48 1.91 3.00 1.07
GP 2.71 1.58 1.77 1.84 2.15 1.84 1.38 1.50 2.59 2.44 1.64
RBF 1.30 1.54 1.31 1.59 1.34 1.32 1.65 1.25 1.40 3.33 1.53
KR 1.00 1.01 1.25 1.04 1.00 1.06 1.08 1.13 1.11 1.05 1.06
SVR 1.55 1.61 1.05 1.37 1.26 1.03 1.49 1.16 1.12 2.85 1.07
ENS 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 269
Table 10. Metamodel prediction errors relative to LS-DYNA
results at the SO optimum.
St Mg
FFI
Int toe 4.1% 4.9%
Int dash 13.8% 1.2%
Accel 14.9% 3.0%
Int eng 3.1% 1.6%
SIDE
Int door 4.4% 2.2%
Accel 15.8% 1.9%
Int eng 2.9% 0.1%
OFI
Int toe 5.1% 10.1%
Int dash 24.1% 6.7%
Accel 0.4% 23.3%
Int eng 0.9% 1.8%
used and results are shown only for the best case. We chose a
limited number of initial design points to explore the impact
of non-convex design space on the optimum design as the
search for global optimum would require a wider array of
initial design points. The initial design points were selected
at random from different regions of the design space.
In the steel design, mass of designed parts dropped
from 105.2 kg to 88.0 kg while in the magnesium design,
the same mass dropped from 42.7 kg to 37.2 kg. That
is a 16% and 13% reduction for steel and magnesium,
respectively, relative to the corresponding baseline models.
Metamodel-based responses at the optimum design point
are compared to the LS-DYNA simulation results in
Table 10, where the per cent error is that of the metamodel
predictions relative to the simulation results. For the steel
design, the relative errors are in the range of 0.4% to 24.1%
for an average of 8.1% while those for the magnesium
design vary from 0.1% to 23.3% for an average of 5.2%.
Table 11. Comparison of surrogate-based crash responses at the
SO optimum for St and Mg.
St SO
opt
Relative to
St base
Mg
SO opt
Relative to
Mg base
FFI
Int toe (mm) 163 1.1% 261 11.7%
Int dash (mm) 123 0.1% 165 11.1%
Accel (gs) 59 1.4% 51 0.4%
Int eng (kJ) 64 1.9% 61 1.2%
SIDE
Int door (mm) 319 0.1% 423 0.8%
Accel (gs) 45 7.6% 40 1.0%
Int eng (kJ) 24 7.8% 21 0.3%
OFI
Int toe (mm) 221 5.1% 352 0.1%
Int dash (mm) 205 11.5% 268 28.0%
Accel (gs) 36 1.29% 38 0.7%
Int eng (kJ) 40 0.4% 39 0.0%
Mass (kg) 88 16.4% 37.2 13.0%
It took total of 2483 function calls and nine optimisation
cycles to nd the optimum point for the steel design
whereas for the magnesium design, 2095 function calls
were used for the same number of optimisation cycles.
An optimisation cycle refers to a complete solution of the
quadratic (direction-nding) subproblem and associated
step size calculation in SQP.
Figure 11 shows the normalised design variables at the
SO optimums. All except two parts (outer cabin and outer
side rail) saw their thickness decrease for the steel opti-
mum design while for the magnesium optimum, nine parts
decreased in thickness while the rest became thicker. It is
interesting to note that the variations in part thickness are
quite different in the steel and magnesiumoptimumdesigns.
Surrogate-based response values at the SO optimum points
for both steel and magnesium can be found in Table 11
Figure 11. Normalised design variables for SO optimums (a) steel and (b) magnesium.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

270 A. Parrish et al.
along with a per cent difference relative to their respective
constraint baselines. A per cent decrease for intrusion dis-
tance, acceleration and mass along with a per cent increase
for internal energy is favourable. A constraint violation
tolerance of 1.5% was dened when solving the optimisa-
tion problem. The intrusion distances using the magnesium
parts are signicantly larger than their steel counterparts.
The accelerations are lower or about the same for magne-
sium compared to steel while internal energy absorption is
slightly less. This difference is understandable as the mag-
nesium design was optimised based on its corresponding
baseline model and not that of the steel design.
5.2. Multi-objective
Results in the previous section suggest that the lighter mag-
nesium designs, baseline and SO optimum, maintain or
improve upon the steel baseline responses for acceleration
and internal energy but exceed the steel baseline intrusion
distances. For this reason, another optimisation problem is
formulated and solved to determine the mass of a mag-
nesium design that meets the crashworthiness properties of
the steel baseline. The goal here is to nd the weight savings
offered by the lightweight magnesium to nd a design with
similar crashworthiness characteristics as the steel base-
line design without focusing on alternative combinations
of responses to optimise or constrain.
This optimisation problem uses the baseline steel in-
trusion distances as target values while constraining the
acceleration and internal energy to the steel baseline val-
ues. This multi-objective (MO) constrained optimisation
problem is formulated as:
s.t.
min F (x) = f (R
j
(x) , j = 1..5)
R
j
(x) R
jSt
; j = 6..8
R
j
(x) R
jSt
; j = 9..11
0.5x
Mg
x 1.5x
Mg
(7)
Table 12. Metamodel prediction errors relative to LS-DYNA
results at the MO optimum.
Metamodel LS-DYNA Error
FFI
Int toe (mm) 166 186 10.6%
Int dash (mm) 136 131 4.4%
Accel (gs) 52.0 59.3 12.3%
Int eng (kJ) 65.3 62.0 5.2%
SIDE
Int door (mm) 328 361 9.1%
Accel (gs) 44.2 48.0 8.0%
Int eng (kJ) 22.1 21.9 0.8%
OFI
Int toe (mm) 256 212 20.7%
Int dash (mm) 247 236 4.7%
Accel (gs) 35.0 37.0 5.4%
Int eng (kJ) 39.0 37.9 2.7%
Figure 12. Normalised design variables for MO optimum.
where F(x) is the vector of objective functions representing
the differences between the intrusion distances of the mag-
nesium design and the corresponding values in the baseline
steel model, R
jSt
for j = 6..8 are the baseline steel accel-
erations, R
jSt
for j = 9..11 are the baseline steel internal
energies and x
Mg
are the baseline magnesium part thick-
ness values.
Several approaches may be used to solve a constrained
MOproblemsuch as the one in Equation (7). Due to the con-
ict that often exists among the different objectives, there
are multiple optimal solutions that fall on the Pareto fron-
tier. Depending on the level of preference (weight) given
to an individual objective, a different point on the Pareto
frontier may be selected as the desired Pareto optimum de-
sign. In this problem, the Pareto frontier would be dened
in a ve-dimensional space. Given the nonlinearity of the
constraint functions involved, we chose compromise pro-
gramming formulation to combine the multiple objectives
into a single composite objective function expressed as:
F(x) =

_
5

j=1
W
j
_
f
j
(x) f
T
j
f
W
j
f
T
j
_
2
, j = 1..5 (8)
Figure 13. Intrusion distance at the baselines and MO optimum
(F is FFI, S is SIDE and O is OFI).
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 271
Table 13. Comparison of surrogate-based responses at MO optimum to St constraints.
FFI SIDE OFI
Accel Int eng Accel Int eng Accel Int eng Mass
Mg MO 52 gs 65 kJ 44 gs 22 kJ 35 gs 39 kJ 50.7kg
Relative to St base 12.6% 4.5% 8.6% 0.4% 2.8% 0.1% 51.8%
Table 14. Design variable summary (thicknesses in mm).
Part Part no. Design variable St base Mg base St SO Mg SO Mg MO
A-pillar 310, 311 x
1
1.611 2.597 0.915 2.561 1.984
Front bump 330 x
2
1.956 5.975 1.204 2.987 6.649
Firewall 352 x
3
0.735 1.072 0.545 0.867 1.515
Front oor panel 353 x
4
0.705 1.136 0.612 1.211 1.592
Rear cabin oor 354 x
5
0.706 1.138 0.477 0.569 1.696
Outer cabin 355, 356 x
6
0.829 1.366 0.988 1.482 2.049
Cabin seat reinf 357 x
7
0.682 1.099 0.622 1.649 1.649
Cabin mid rail 358, 359 x
8
1.050 1.692 0.633 1.792 1.636
Shotgun 373, 374 x
9
1.524 3.620 0.762 1.810 1.810
Inner side rail 389, 391 x
10
1.895 3.966 1.887 3.436 4.141
Outer side rail 390, 392 x
11
1.522 3.186 1.834 3.145 2.754
Side rail exten. 398, 399 x
12
1.895 3.966 1.780 4.805 5.950
Rear plate 415 x
13
0.710 1.144 0.417 1.559 1.717
Roof 416 x
14
0.702 1.157 0.351 0.739 0.791
Susp. frame 439 x
15
2.606 5.342 1.303 4.367 4.931
where W
j
is the weight factor for the jth objective, f
j
(x) is
the jth objective, f
T
j
is the target value of the jth objective
taken to be equal to the corresponding response value
in the steel baseline design (R
j
St
) and f
W
j
is the worst
known value of the jth objective taken to be equal to the
corresponding response value in the magnesium baseline
Figure 14. Optimisation response results summary: (a) accelerations and internal energies, and (b) intrusion distances. (F is FFI, O is
OFI, and S is SIDE)
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

272 A. Parrish et al.
Figure 15. Simulation images of the designs: FFI at 150 ms for (a) St Base, (b) Mg Base, (c) St SO, (d) Mg SO, and (e) Mg MO.
design (R
j
Mg
). Given the importance of all intrusion
distances to vehicle safety, we chose equal weight factors
for all the objectives in Equation (8). As in the case of
SO, all responses are represented by the corresponding
optimum ensemble metamodels.
The optimisation problem in Equation (8) is solved us-
ing SQP in VisualDOC software with eight different ini-
tial design points. The crash responses predicted by the
metamodels and the corresponding LS-DYNA values at
the MO optimum design point are shown in Table 12. The
metamodel errors are generally less than 10% with only
two responses above 11%. The MO optimum design was
found in nine optimisation cycles with total of 2909 function
calls.
Figure 12 shows the normalised optimum design vari-
able values with ve reaching the upper bound, one at the
lower bound and the rest scattered between the two bounds.
It appears that the parts dened by design variables x
5
, x
6
,
x
7
, x
12
and x
13
play a more critical role in this MO optimi-
sation problem than the rest. The chart in Figure 13 shows
that the intrusion distance responses at the MO optimum
are considerably less than those of the magnesium baseline
model and most are near to or less than those of the steel
baseline design with IntToe of FFI and IntDoor of SIDE
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 273
Figure 16. Simulation images of the designs: SIDE at 150 ms for (a) St Base, (b) Mg Base, (c) St SO, (d) Mg SO, and (e) Mg MO.
being about 1618% higher. The values shown in Figure 13
are the LS-DYNA simulation results at the MO optimum
rather than the metamodel predictions. This MO optimum
design with magnesiumparts has a mass that is about half of
that of the baseline steel design, 50.7 kg compared to 105.2
kg, and is 8 kg heavier than the baseline magnesiumdesign.
Surrogate-based results of the other responses of inter-
est are found in Table 13 along with a percentage com-
parison to the steel baseline constraint values. A per cent
decrease for intrusion distance, acceleration and mass along
with a per cent increase for internal energy is favourable.
A constraint violation tolerance of 1.5% was dened when
solving the optimisation problem. Table 13 shows that the
MO optimum design produces better results for the accel-
eration constraints and FFI internal energy response with
a minimal reduction in the OFI and SIDE internal energy
responses.
5.3. Optimisation results summary
Results show that the magnesium designs have lower mass
than the steel designs and similar or better crashworthi-
ness when considering acceleration and internal energy as
design constraints. The all steel model was 17.2 kg lighter
when optimised for mass than the steel baseline model. The
optimised model with selected parts replaced with magne-
sium was 54.5 kg lighter than the steel baseline and 37.3 kg
lighter than the optimised steel model with similar crash
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

274 A. Parrish et al.
Figure 17. Simulation images of the designs: OFI at 150 ms for (a) St Base, (b) Mg Base, (c) St SO, (d) Mg SO, and (e) Mg MO.
responses. The LS-DYNA based response values at the
point of optimumfor SOand MOproblems are summarised
in Figure 14.
Design variable values representing wall thickness of
individual parts in mm at the baseline and optimised de-
signs can be found in Table 14. The shotgun wall thickness
was reduced to the lower bound in all three optimisation
problems solved. This suggests that this part is at its en-
ergy absorption limit and altering the thickness within these
bounds does not alter the absorption signicantly. Results
show that magnesium for steel substitution based on only
maintaining energy absorption does not produce a design
that satises the other crashworthiness characteristics. De-
sign optimisation is needed to nd a magnesium design
meeting steel crashworthiness. Figures 15, 16 and 17 show
simulation images of the ve designs for FFI, SIDE and
OFI, respectively. Differences can be observed in the door
for the FFI scenario, the door and roof for SIDE, and the
roof for OFI. The overall crash response of the magnesium
at the point of MO optimum appears to be fairly close to
that of the steel baseline model.
6. Summary and conclusions
This studyusedmetamodellingandoptimisationtechniques
to explore the application of a lightweight magnesium alloy
for a group of energy absorbing parts in a vehicle body
structure.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 275
A full-scale Dodge Neon model developed and vali-
dated for FFI at the NCAC was incorporated into SIDE and
OFI scenarios. Vehicle-based responses from these three
scenarios along with selected part mass were used as con-
straints and objectives during design optimisation with part
thicknesses as the design variables. All crash simulations
were performed using LS-DYNA nite element solver.
AZ31 magnesium alloy was used to replace the base-
line steel material in the FE model. Due to the complex-
ity of the failure characteristics of magnesium, a limit on
maximum plastic strain was used as a failure criterion to
disable failed elements. This material substitution coupled
with design optimisation reduced the mass of the selected
parts by approximately 50%of the baseline steel model and
maintained or improved the crashworthiness for most of the
factors considered.
Adirect substitution of magnesiumfor steel using thick-
nesses dened by maintaining the internal energy absorp-
tion does not give the same crash characteristics or re-
sponses as the steel components. Material substitution can
change the deformation mode and folding mechanism of
energy absorbing parts under crash loading as indicated by
similar internal energies but signicantly larger intrusion
distances for the models using steel and magnesium parts.
Shotgun part thickness was reduced to the lower bound
in all of the optimum designs. This suggests that this part is
at its energy-absorbing limit for these bounds and increas-
ing part thickness does not necessarily increase energy
absorption. Parts such as the roof were chosen as design
parts primarily because of their contributions to vehicle
stiffness. Roof thickness was signicantly reduced in all
of the designs because of its small contribution to energy
absorption. Inclusion of vehicle stiffness responses during
optimisation would likely show less weight reduction in
the optimised designs.
Future work will consider inclusion of vehicle stiffness
(as a measure of overall rigidity as well as NVH charac-
teristic of the vehicle) in the design optimisation process,
including a dummy model to provide dummy-based re-
sponses, adding a microstructure-based material model to
provide a more representative model of magnesium, and
using alternative optimisation strategies.
Acknowledgements
This material is based on the work supported by the U.S. Depart-
ment of Energy under Award Number DE-EE0002323.
Disclaimer
This report was prepared as an account of work sponsored by
an agency of the United States Government. Neither the United
States Government nor any agency thereof, nor any of their em-
ployees, makes any warranty, express or implied, or assumes any
legal liability or responsibility for the accuracy, completeness,
or usefulness of any information, apparatus, product, or process
disclosed, or represents that its use would not infringe privately
owned rights. Reference herein to any specic commercial prod-
uct, process, or service by trade name, trademark, manufacturer,
or otherwise does not necessarily constitute or imply its endorse-
ment, recommendation, or favouring by the United States Gov-
ernment or any agency thereof. The views and opinions of authors
expressed herein do not necessarily state or reect those of the
United States Government or any agency thereof.
References
[1] E. Acar and M. Rais-Rohani, Ensemble of metamodels with
optimised weight factors, Struct. Multidiscip. Optim. 37
(2008), pp. 279294.
[2] E. Acar and K. Solanki, Improving the accuracy of vehicle
crashworthiness response predictions using an ensemble of
metamodels, Int. J. Crashworthiness 14 (2009), pp. 4961.
[3] A. Akkerman, M. Burger, B. Kuhn, H. Rajic, N. Stander,
and R. Thyagarajan, Shape optimisation of instrument panel
components for crashworthiness using distributed comput-
ing, Proceedings of the 6th International LS-DYNA Confer-
ence, Detroit, MI, 2000.
[4] C.M. Bishop, Neural Networks for Pattern Recognition, Ox-
ford University Press, New York, 1995.
[5] V. Cherkassky and Y. Ma, Practical selection of SVM pa-
rameters and noise estimation for SVM regression, Neural
Netw. 17 (2004), pp. 113126.
[6] S. Clarke, J. Griebsch, and T. Simpson, Analysis of sup-
port vector regression for approximation of complex en-
gineering analyses, J. Mech. Des. 127 (2005), pp. 1077
1087.
[7] EERE, Annual progress reports, Vehicle Technologies Pro-
gram, U.S. Department of Energy, Washington, DC, (2011).
Available at http://www1.eere.energy.gov/vehiclesandfuels/
resources/fcvt_reports.html.
[8] H. Fang, M. Rais-Rohani, Z. Liu, and M.F. Horstemeyer,
A comparative study of metamodeling methods for multi-
objective crashworthiness optimisation, Comput. Struct. 83
(2005), pp. 21212136.
[9] H. Fang, K. Solanki, M.F. Horstemeyer, and M. Rais-Rohani,
Multi-impact crashworthiness optimisation with full-scale
nite element simulations, Proceedings of the 6th World
Congress of Computational Mechanics, Beijing, China,
2004.
[10] H. Fang and Q. Wang, On the effectiveness of assessing
model accuracy at design points for radial basis functions,
Commun. Numer. Meth. Eng. 24 (2008), pp. 219235.
[11] S.R. Gunn, Support vector machines for classication and
regression, Technical Report, University of Southampton,
Southampton, 1997.
[12] C.W. Hsu, C.C. Chang, and C.J. Lin, A Practical Guide to
Support Vector Classication, National Taiwan University,
Taipei, 2010.
[13] R. Jin, W. Chen, and T.W. Simpson, Comparative studies of
metamodeling techniques under multiple modeling criteria,
Struct. Multidiscip. Optim. 23 (2001), pp. 113.
[14] Karco Engineering, National Highway Trafc Safety Ad-
ministration Frontal Barrier Forty Percent Offset Impact
Test: 1996 Dodge Neon, U.S. Department of Transportation,
Washington, DC, 1997.
[15] X. Liao, Q. Li, X. Yang, W. Li, and W. Zhang, A two-stage
multi-objective optimisation of vehicle crashworthiness un-
der frontal impact, Int. J. Crashworthiness 13 (2008), pp.
279288.
[16] S.N. Lophaven, H.B. Nielsen, and J. Sndergaard, DACE: A
MATLAB Kriging Toolbox, Informatics and Mathematical
Modeling, Technical University of Denmark, Lyngby, 2002.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

276 A. Parrish et al.
[17] LSTC, LS-DYNA Keyword Users Manual: Volume II, Ma-
terial Models, Version 971, Livermore Software Technol-
ogy Corporation, Livermore, CA, 2007, pp. 110114 and
473475.
[18] Mathworks, Inc., MATLAB: The Language of Technical
Computing, Software Package, Ver. 7.8.0347 (R2009a), The
Mathworks, Inc., Natick, MA, 2009.
[19] MGA, Saftey Compliance Testing for FMVSS 214 Side
Impact Protection Passenger Cars: 1997 Dodge Neon,
prepared by MGA proving Grounds, U.S. Department of
Transportation, Washington, DC, 1997.
[20] C. Mozumder, P. Bandi, N. Patel, and J. Renaud, Thick-
ness based topology optimisation for crashworthiness de-
sign using hybrid cellular automata, Proceedings of the 12th
AIAA/ISSMOMultidisciplinary Analysis and Optimisation
Conference, Victoria, British Columbia, Canada, 2008.
[21] A. Naja and M. Rais-Rohani, Mechanics of axial plastic
collapse in multi-cell, multi-corner crush tubes, Thin-Wall.
Struct. 49 (2011), pp. 112.
[22] NCAC, Finite Element Model of Dodge Neon: Model Year
1996 Version 7. FHWA/NHTSA National Crash Analysis
Center, Ashburn, VA, 2006.
[23] NCAC, Finite Element Model Archive, National Crash Anal-
ysis Center, Ashburn, VA. Available at http://www.ncac.
gwu.edu/vml/models.html.
[24] NHTSA, Federal Motor Vehicle Safety Standards and Regu-
lations, U.S. Department of Transportation, National High-
way Trafc Safety Administration, Washington, DC, 1998.
[25] F. Pan, P. Zhu, and Y. Zhang, Metamodel-based lightweight
design of B-pillar with TWB structure via support vec-
tor regression, Comput. Struct. 88 (2010), pp. 36
44.
[26] N. Peixinho and C. Doellinger, Characterisation of dy-
namic material properties of light alloys for crashwor-
thiness applications, Mater. Res. 13 (2010), pp. 471
474.
[27] M. Rais-Rohani, K. Solanki, E. Acar, and C. Eamon, Shape
and sizing optimisation of automotive structures with deter-
ministic and probabilistic design constraints, Int. J. Vehicle
Des. 54 (2010), pp. 309338.
[28] C. Rasmussen and C. Williams, Gaussian Processes for Ma-
chine Learning, MIT, Cambridge, MA, 2006.
[29] T. Simpson, T. Mauer, J. Korte, and F. Mistree, Kriging
models for global approximations in simulation-based mul-
tidisciplinary design optimisation, AIAA J. 39 (2001), pp.
22312241.
[30] J. Sobieszczanski-Sobieski, S. Kodiyalam, and R.Y. Yang,
Optimisation of car body under constraints of noise, vibra-
tion, and harshness (NVH), and crash, Struct. Multidiscip.
Optim. 22 (2001), pp. 298306.
[31] C. Turner, Design space analysis with hyperdimensional
metamodels, Proceedings of the NSF Engineering Research
and Innovation Conference, Honolulu, HI, 2009.
[32] USAMP, Magnesium Vision 2020: A North American Au-
tomotive Strategic Vision for Magnesium, U.S. Automotive
Materials Partnership, Southeld, MI, 2006.
[33] Vanderplaats R&D, VisualDOC 6.2: Advanced Examples
Manual, Vanderplaats Research and Development, Inc.,
Novi, MI, 2009, p. 93.
[34] D.A. Wagner, S.D. Logan, K. Wang, and T. Skszek, FEA
predictions and test results from magnesium beams in bend-
ing and axial compression, Proceedings of the SAE 2010
World Congress & Exhibition, Detroit, MI, 2010.
[35] L. Wang, D. Beeson, G. Wiggs, and M. Rayasam,
A comparison of meta-modeling methods using prac-
tical industry requirements, Proceedings of the 47th
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dy-
namics, and Materials Conference, Newport, RI, 2006.
[36] R.J. Yang, N. Wang, C.H. Tho, J.P. Bobineau, and B.P.
Wang, Metamodeling development for vehicle frontal im-
pact simulation, J. Mech. Des. 127 (2005), pp. 1014
1020.
[37] L. Zerpa, N.V. Queipo, S. Pintos, and J. Salager, An optimi-
sation methodology of alkaline-surfactant-polymer ooding
processes using eld scale numerical simulation and mul-
tiple surrogates, J. Petrol. Sci. Eng. 47 (2005), pp. 197
208.
Appendix 1
Polynomial response surface (PRS)
PRS is one of the most widely used and simplest metamodelling
techniques, but it may not provide an accurate prediction for cer-
tain responses. The most typical form of PRS is a second-degree
polynomial function of the form:

f (x) = b
o
+
L

i=1
b
i
x
i
+
L

i=1
b
ii
x
2
i
+
L1

i=1
L

j=i+1
b
ij
x
i
x
j
(A.1)
where

f (x) is the metamodel prediction at point x, L is the num-
ber of design variables in the design vector x and b
0
, b
i
, b
ii
,
b
ij
are the unknown coefcients found using the least squares
technique.
As a regression model, PRS may not pass through the train-
ing points. The degree of the polynomial can be changed or
some of the terms appearing in Equation (A.1) can be omitted
depending on the nonlinearity of the response function being
modelled.
Gaussian process (GP)
Descriptions of Gaussian Process (GP) in this discussion fol-
low that in Wang et al. [35] and Acar and Rais-Rohani
[1]. The GP metamodel is a group of output variables f
N
=
{f
n
(x
1
n
, x
2
n
, . . . , x
1
n
)
N
n=1
} with a Gaussian joint probability distri-
bution:
P (f
N
|C
N
, X
N
)
=
1
_
(2)
N
|C
N
|
exp
_

1
2
(f
N
)
T
C
1
N
(f
N
)
_
(A.2)
where X
N
{x
n
}
N
n=1
are N pairs of L-dimensional input variables
x
n
= x
1
n
, x
2
n
, . . . , x
L
n
, C
N
is the covariance matrix with elements of
C
ij
= C(x
i
, x
j
) and is the mean output vector.
Elements of the covariance matrix C
N
are calculated from:
C
ij
=
1
exp

1
2
L

l=1
_
x
(l)
i
x
(l)
j
_
2
r
2
l

+
2
(A.3)
C
ij
=
1
exp

1
2
L

l=1
_
x
(l)
i
x
(l)
j
_
2
r
2
l

+
2
+
ij

3
(A.4)
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 277
where
1
,
2
,
3
and r
l
are referred to as hyperparameters with
r
l
being the length scale.
3
is an independent noise parameter
and
ij
is Kroneckers delta (equal to one when i = j and zero
otherwise). These hyperparameters are selected to maximise log-
arithmic likelihood of the predictions matching the training data.
This is given by:
LL =
1
2
log |C
N
|
1
2
f
T
N
C
1
N
f
N

N
2
log(2) +ln(P())
(A.5)
where P() is the prior distribution of the hyperparameters. This
is usually uniform because no prior knowledge is available and
can be equated to zero for optimisation.
Equations (A.3) and (A.4) dene the interpolation and re-
gression modes of the Gaussian process model, respectively. The
former passes through all training points while the latter provides
a smoother surface to help with noisy data. The prediction surface
with noise ltered out is less complex and might not pass through
all training points but this has better predictions at non-training
points.
The response value at a prediction point x
p
=
(x
1
p
, x
2
p
, . . . , x
L
p
) is estimated as:

f (x
p
) = k
T
C
1
N
f
N
(A.6)
where k = [C(x
1
, x
p
), . . . , C(x
N
, x
p
)]. Standard deviation at the
prediction point is available without requiring additional simula-
tions or tests and can be calculated from:

f
(x
p
) = k
T
C
1
N
k (A.7)
where = C(x
p
, x
p
).
No tuning parameters are explored within GP. The MATLAB
[18] toolbox fromRasmussen and Williams [28] is used to develop
the GP metamodels.
Radial basis function (RBF)
This formulation of RBF requires normalised training and test
points in the range of 0 to 1. This is done by dividing each variable
by the maximumvalue of that variable in the DOE table. The basic
form of RBF is given as:

f (x) =
N

i=1

i
(x x
i
) (A.8)
where

f (x) is the metamodel prediction at point x, N is the number
of training points, x is the input vector of normalised variables, x
i
is vector of normalised design variables at the ith training point
and x x
i
=
_
(x x
i
)
T
(x x
i
) is the Euclidean normor dis-
tance r from point x to the training point x
i
. The
i
parameters are
the unknown interpolation coefcients that must be calculated.
is the radially symmetric basis function that can take on a number
of forms. Equation (A.8) represents a linear combination of a -
nite number of basis functions. Typical radial basis functions are
listed below.
Thin Plate Spline: (r) = r
2
ln (cr)
Gaussian: (r) = exp(cr
2
)
Multiquadric: (r) =

r
2
+c
2
Inverse Multiquadric: (r) = 1/

r
2
+c
2
The value c is a constant that the user determines. Values of
r are between 0 and 1 because the training and test points are
between 0 and 1 resulting in 0 c 1. In general, Multiquadric
with c = 1 gives good results for many function types.
The interpolation coefcients,
i
, can be found by minimising
the residual (the sum of the squares of the deviations) as:
R =
N

j=1
_
f (x
j
)
N

i=1

i
(x
j
x
i
)
_
2
(A.9)
In matrix form, this is expressed as:
[A]{} = {f } (A.10)
where [A] = (x
j
x
i
) with j = 1, N, and i = 1, N. Solving
Equation (A.11) for and inputting into Equation (A.8) generates
predictions. Error analysis typically relies on data at test points
outside of the training set since RBF is an interpolation model that
passes through all the training points.
RBF tuning parameters are c and . All four radial ba-
sis functions discussed here were used with values of c =
{0.05, 0.1, 0.15, . . . , 1} to determine the combination with lowest
cross-validation GMSE.
Kriging (KR)
Descriptions of Kriging followthose in Acar and Rais-Rohani [1],
Lophaven et al. [16] and Simpson et al. [29]. For this discussion,
the design (training) point matrix is s, the corresponding training
point function evaluations are Y, the test point or prediction point
is x and general variables w and v are introduced. This descrip-
tion of Kriging requires the training set to have zero mean and a
covariance of 1. This is done by:
s =
(s
original

s
)

s
(A.11)
Y =
(Y
original

Y
)

Y
(A.12)
where original indicates the un-normalised values,
s
and
Y
are
the means of the training points and their responses, respectively,
and
s
and
Y
are the standard deviations of the training points
and their responses, respectively.
Kriging models assume that the function takes the form of:

f (x) = P(x) +Z(x) (A.13)

where

f (x) is the approximate function, P(x) is a polynomial
function that globally approximates the actual function and Z(x) is
the stochastic component that accounts for deviations or Z(x) =
r(x)
T
.
First, the polynomial portion of the model is explored, i.e.
P(x). This is a linear combination of np polynomial functions,
p(x), and regression parameters

:
P(x) = p(x)
T

(A.14)
The degree of the polynomial is chosen to be 0, 1 or 2 re-
sulting in np polynomial equations, with a 1st degree polynomial
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

278 A. Parrish et al.
generally being used. The 2nd degree (quadratic) formulation is
similar to the 2nd degree PRS model discussed previously. That
is, with L = number of variables using general variable w:
degree 0(constant)
np = 1
p
1
(w) = 1
(A.15)
degree 1(linear)
np = L +1
p
1
(w) = 1, p
2
(w) = w
1
, . . . , p
L+1
(w) = w
L
(A.16)
degree 2(quadratic)
np =
1
2
(L +1)(L +2)
p
2
(w) = w
1
, . . . , p
L+1
(w) = w
L
p
L+2
(w) = w
2
1
, . . . , p
2 L+1
(w) = w
1
w
L
p
2L+2
(w) = w
2
2
, . . . , p
3 L
(w) = w
2
w
L
. . . . . . p
np
(w) = w
2
np
(A.17)
Polynomial function evaluations at each training point must
be performed to t the training data. Dene P(s) or just P as:
P(s) = P = [p(s
1
), . . . , p(s
N
)]
T
(A.18)
where N is the number of training points. This is a vector of ones
if the 0 degree polynomial is used.
The stochastic component Z allows the Kriging model to
interpolate the response value and has a zero mean and covariance
of:
COV[Z(x
i
), Z(x
j
)] =
2
R
_
, x
i
, x
j
_
i, j = 1, ..., N
(A.19)
where
2
is the variance, R(, x
i
, x
j
) is the correlation function
between sample points x
i
and x
j
and represents correlation pa-
rameter that must be calculated. The user denes the correlation
function, R(, x
i
, x
j
), with a Gaussian correlation function gen-
erally chosen. Correlation functions use general variables w and v
and are in the form:
R(, w, v) =
L

k=1
R
k
(
k
, d
k
) (A.20)
where L is the number of design variables and d
k
= w
k
v
k
is
the distance between w
i
and v
j
at the kth component of the points.
Some Kriging correlation functions are listed below:
Gaussian: R
k
(
k
, d
k
) = exp(
k
d
2
k
)
Exponential: R
k
(
k
, d
k
) = exp(
k
|d
k
|)
Exp. General: R
k
(
k
, d
k
) = exp(
k
|d
k
|

L+1
),
0 <
L+1
2
Linear: R
k
(
k
, d
k
) = max{0, 1
L+1
|d
k
|}
Spherical: R
k
(
k
, d
k
) = 1 1.5
2
k
+0.5
3
k
,

k
= min{1,
k
|d
k
|}
Cubic: R
k
(
k
, d
k
) = 1 3
2
k
+2
3
k
,

k
= min{1,
k
|d
k
|}
Spline: R
k
(
k
, d
k
) = (
k
),
k
=
k
|d
k
|
(
k
) =

1 15
2
k
+30
3
k
for 0
k
0.2
1.25(1
k
)
3
for 0.2 <
k
< 1
0 for
k
1

The correlation matrix R is dened as

R(, s) = R = R(, s
i
, s
j
), i, j = 1, ..., N (A.21)
Maximum likelihood estimation is used to estimate the cor-
relation parameter, . is found by solving the following optimi-
sation problem if a Gaussian correlation model is selected:
min () |R|
1/N

2
or
max () =
_
N ln(
2
) +ln |R|
_
/2
s.t. > 0
(A.22)
where |R| is the determinant of R and
2
is the process variance
dened as:

2
= [(Y P

)
T
R
1
(Y P

)]/N (A.23)
The initial value of as well as its upper and lower bounds
inuence the calculation of from Equation (A.22) and the accu-
racy of the resulting Kriging model. usually falls between 0 and
2 with the initial value being taken as the midpoint. Almost any
value of will produce a Kriging model and results will be pre-
dicted, but these predictions are not necessarily accurate. Solving
the optimisation problem in Equation (A.22) will result in a more
accurate Kriging model.
The steps to derive Equations (A.24)(A.27) are not presented
here; see Lophaven et al. [18] for a description of the derivation.
Once regression and correlation functions are chosen, the response
is predicted as:

f (x) = p(x)
T

+r(x)
T
(A.24)
where p(x) is the polynomial term found from Equation (A.15)
to Equation (A.17), r(x) is the correlation vector (see Equation
(A.25)),

comes from a generalised least squares solution and
seen in Equation (A.26) and is computed from the residual and
seen in Equation (A.27).
r(x) = [R(, s
1
, x) . . . , R(, s
m
, x)]
T
(A.25)

= (P
T
R
1
P)
1
P
T
R
1
Y (A.26)
= R
1
(Y P

) (A.27)
Beloware some key points about the discussion of Kriging.
1.

and are functions of training points (not test points). This
means

and are constant for a given training set, so from
Equation (A.24) only the correlation and regression functions
need to be evaluated at the test point(s) to get the Kriging
prediction(s) once

and are found.
2.

and are functions of R(, s) and therefore functions of
correlation parameter theta.
3. The quantity R
1
appears in multiple places. This can be-
come very computationally expensive as the number of training
points (N) increases (R is an N by N matrix), which is common
in a number of metamodelling applications. For this reason, the
inverse should be calculated using linear algebra techniques.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 279
4. The best regression and correlation functions are dependent
on the problem, but a 1st degree polynomial and a Gaussian
correlation function serve as a good starting point.
The MATLAB toolbox developed in Lophaven et al. [16] is
used in this study. The tuning parameters of KR explored are the
polynomial degree, regression function, as well as the upper and
lower bounds on . The upper and lower bound of are the nal
two tuning parameters of KR.
Support vector regression
Descriptions of SVR in this discussion follow that in References
[1, 5, 6, 11, 12]. The literature suggests that the design variables
should be normalised to a range of (1,1) or (0,1). Simply, SVR
constructs a hyperplane that passes near each design point such
that they fall within a specied distance of the hyperplane. In
two dimensions, this hyperplane is simply a line. The hyperplane
is then used to predict other responses. SVR estimates the real
function as:
y = r(x) + (A.28)
where is an independent random noise, x is the multivariable
input, y is the scalar output and r is the mean of the conditional
probability (regression function); see Cherkassky and Ma [5] and
Gunn [11] for more information. SVR technique selects the best
approximate model from a group of selection models that min-
imise the prediction risk. Linear or nonlinear regression can be
performed. When a linear regression is used, the pool of approxi-
mation models is given by:

f (x) = x +b (A.29)
where b is the bias term and x is the dot product of and
x. Minimising empirical risk using the -insensitive loss func-
tion allows regression estimates. It is desirable to have a at
approximation function and this is achieved by minimising ||
2
.
Non-negative slack variables (
i
and

i
) are introduced to account
for training points that fall outside of the -insensitive zone. That
is:
minimize
1
2
||
2
+C
N

i=1
(
i
+

i
)
s.t.

y
i
x
i
b +

i
x
i
+b y
i
+
i

i
,
i
0
(A.30)
where C is a positive constant and is the insensitive zone, both
chosen by the user. C is also referred to as the regression parameter
or penalty parameter. Cherkassky and Ma [5] propose that C be
chosen as:
C = max(|
Y
+3
Y
|, |
Y
3
Y
|) (A.31)
where
Y
and
Y
are the mean and standard deviation of the
training point responses. Hsu et al. [12] suggest a cross-validation
approach to nd C.
The parameter determines the width of the -insensitive
zone and affects the complexity/atness of the model. The values
of should be tuned to the input data, but a reasonable starting
value is found using:
=
c
100
| y
max
y
min
| (A.32)
with c = 1 where | y
max
y
min
| is the range of the responses at
the training points. The value of c can be tuned for the function.
Cherkassky and Ma [5] propose that be chosen as:
= 3
_
ln N
N
(A.33)
where is the standard deviation of the noise associated with
the training point response values and N is the number of training
points. This assumes that the noise is known or can be determined.
Cherkassky and Ma [5] suggest the following to estimate the
unknown variance of noise using a k-nearest neighbour technique:

2
=
N
1/5
k
N
1/5
1

1
N
N

i=1
(y
i
y
i
)
2
(A.34)
where k is in the range [9, 32] and

N
i=1
(y
i
y
i
)
2
is the squared
sum of the residuals.
The optimisation problem in Equation (A.31) written as a
Lagrangian function is
L =
1
2
||
2
+C
N

i=1
_

i
+

i
_

i=1

i
( +
i
y
i
+ x
i
+b)

i=1

i
_
+

i
+y
i
x
i
b
_

i=1
_

i
+

i
_
(A.35)
where
i
and

i
are additional slack variables. From Lagrangian
theory, necessary conditions for to be a solution are listed below:

b
L =
N

i=1
(

i

i
) = 0 (A.36)

L =
N

i=1
_

i

i
_
x
i
= 0 (A.37)

i
L = C
i

i
= 0 (A.38)

i
L = C

i
= 0 (A.39)
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

280 A. Parrish et al.
Substituting Equations (A.36)(A.39) into Equation (A.30)
gives the dual form optimisation problem:
Maximize

1
2
N

i,j=1
_

i
_ _

j
_ _
x
i
x
j
_

i=1
_

i
+

i
_
+
N

i=1
y
i
_

i
_
s.t.

i=1
_

i
_
= 0
_

i
_
[0, C]
(A.40)
Equation (A.37) is rewritten as:
=
N

i=1
_

i

i
_
x
i
(A.41)
The linear regression rst expressed in Equation (A.29) is
written as:
f (x) =
N

i=1
_

i
_
x
i
x
i
+b (A.42)
Clarke et al. [6] (page 1079) summarise the process of trans-
forming the problem into dual form by stating:
Transforming the optimisation problem into dual form
yields two advantages. First, the optimisation problem is
now a quadratic programming problem with linear con-
straints and a positive denite Hessian matrix, ensuring a
unique global optimum. For such problems, highly efcient
and thoroughly tested quadratic solvers exist. Second, as
can be seen in Equation [(A.40)], the input vectors only
appear inside the dot product. The dot product of each pair
of input vectors is a scalar and can be preprocessed and
stored in the quadratic matrix M
ij
= (x
i
x
j
)
ij
. In this way,
the dimensionality of the input space is hidden from the
remaining computations, providing means for addressing
the curse of dimensionality.
A nonlinear regression model can be developed by replacing
the dot product x
i
x
i
with a kernel function, k, rewriting the
optimisation problem in Equation (A.40) as
Maximize

1
2
N

i,j=1
(
i

i
)(
j

j
)k(x
i
, x
j
)

i=1
(
i
+

i
) +
N

i=1
y
i
(
i

i
)
s.t.

i=1
y
i
(
i

i
) = 0
(
i

i
)[0, C]
(A.43)
Replacing the dot product with a kernel function in the ap-
proximation function Equation (A.43) gives the nonlinear SVR
approximation as:
f (x) =
N

i=1
_

i
_
k(x
i
, x
j
) +b (A.44)
Common kernel functions include the following:
Linear: k(x
i
, x
j
) = x
T
i
, x
j
Polynomial: k(x
i
, x
j
) = ( x
i
x
j
+r)
d
> 0
Gaussian: k(x
i
, x
j
) =
exp(
x
i
x
j

2
2p
2
)
Radial Basis Function: k(x
i
, x
j
) =
exp( x
i
x
2
j
) > 0
Sigmoid: k(x
i
, x
j
) = tan h( x
i
x
j
+r)
where , d, p and r are kernel parameters and should be adjusted
by the user for each data set. Listed below are some insights about
setting kernel parameters:
The polynomial degree d is typically chosen to be 2. Gunn
[11] uses r = 1 to avoid problems with the Hessian be-
coming zero.
Hsu et al. [12] use a cross-validation approach to determine
for RBF. This procedure could be applied to for other
kernels as well as other kernel parameters.
Cherkassky and Ma [5] suggest p for the Gaussian kernel
(they call it RBF, the formulation is the same as Gaus-
sian here) as P (0.1 0.5)

range(x) for single variable

problems and P
L
(0.1 0.5) for multivariable problems
where L is the number of variables and all input variables
are normalised to [0,1].
It should be noted that the Gaussian kernel here is some-
times referred to as Radial Basis Function and Gaussian Radial
Basis Function in the literature. Experience shows that SVRmeta-
models are highly sensitive to tuning parameters. The suggestions
expressed in this discussion may not result in an acceptable ac-
curacy level. The reader is encouraged to perform SVR tuning to
ensure an accurate model for the selected response.
The SVRMATLABtoolbox developed by Gunn [11] was used
in this study and the Linear kernel is used. The tuning parameters
explored within SVR are penalty parameter, C, and -insensitive
zone parameter, c.
Optimised ensemble
An ensemble of metamodels combines predictions from several
stand-alone metamodels such as the ones presented above. Ensem-
ble metamodels are more accurate than the individual members
but also more computationally expensive. The general form of an
ensemble is a weighted sum of the predictions of separate meta-
models. In mathematical form, this is expressed as [4, 37]:

f (x) =
M

i=1
w
i
(x)

f
i
(x) (A.45)
where

f (x) is the ensemble prediction, x is the vector of input
variables, M is the number of metamodels used to build the en-
semble, w
i
is the weight factor for the ith metamodel and

f
i
is the
prediction of the ith metamodel. The weight factors must sum to
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2

International Journal of Crashworthiness 281
one:
M

i=1
w
i
(x) = 1 (A.46)
Selection of the weight factors is the most important step
when constructing an accurate ensemble. Acar and Rais-Rohani
[1] developed an ensemble minimising the error by nding the
optimal weight factors. In mathematical form, this is expressed
as:
min = Err{

f (w
i
,

f
i
(x
k
)), f (x
k
), k = 1 N}
s.t.
M

i=1
w
i
= 1 (A.47)
where Err{} is the error metric that nds error of the ensemble
predictions,

f , f (x
k
) is the actual response at the training point x
k
,
and N is the number of training points. In this study the prediction
error was based on the GMSE metric dened as:
GMSE
1
N
N

i=1
(f i

f
i
)
2
(A.48)
where N is the number of design points, is the actual response
and

f
i
is the predicted response from each metamodel.
D
o
w
n
l
o
a
d
e
d

b
y

[
G
a
z
i

U
n
i
v
e
r
s
i
t
y
]

a
t

2
0
:
1
5

0
7

D
e
c
e
m
b
e
r

2
0
1
2