Composite Structures
Composite Structures
Composite Structures
for Crashworthiness
Ang Li
Master thesis in Aerospace Engineering
Optimization of
Composite Structures
for Crashworthiness
by
Ang Li
to obtain the degree of Master of Science
at the Delft University of Technology,
to be defended publicly on Tuesday February 26, 2019 at 10:00 AM
iii
Preface
This Master’s thesis project was carried out in the faculty of Aerospace Engineering at TU
Delft, under the supervision of prof. Chiara Bisagni. First of all, I would like to thank prof.
Chiara Bisagni for giving me the opportunity to conduct my master research in the field of
crashworthiness. Furthermore, my colleague Jens Semmelroggen deserves a huge thanks
for his help and support throughout the whole thesis. Lastly, I would also like to thank all the
students and researchers in Chiara Bisagni Group for their encouragement and sustained
support.
Ang Li
Delft, February 2019
v
Contents
1 Introduction 1
1.1 Importance of crashworthiness optimization in the aeronautical industry for composites. . 1
1.2 Optimization strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.1 Single/multi-objective optimization. . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.2 Optimization under uncertainties . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2.3 Topology optimization in crashworthiness . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Objective and methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2 Literature study — crashworthiness of composite structures 7
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2 Experimental investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.2.1 Development of crashworthiness tests on thin walled composite tubes . . . . . . 7
2.2.2 Experimental investigation on key influential parameters for composite tubes . . . 9
2.3 Numerical investigation in LS-DYNA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Modeling approaches in LS-DYNA for composite crashworthiness . . . . . . . . . 12
2.3.2 Material models in LS-DYNA for crashworthiness analysis . . . . . . . . . . . . . 13
2.3.3 Applications of LS-DYNA material models in square tubes . . . . . . . . . . . . . 22
2.3.4 Applications of LS-DYNA material models in other components/structures . . . . . 26
2.4 Surrogate model based optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.1 Surrogate models. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4.2 Heuristic optimization algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3 Material coupon test simulations 35
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.1 Material properties and test standards . . . . . . . . . . . . . . . . . . . . . . . . 35
3.1.2 LS-DYNA settings for modeling the coupons . . . . . . . . . . . . . . . . . . . . . 36
3.2 Coupon-level simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.2.1 Tensile test simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.2 Compression test simulation results . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2.3 Identified material property values . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4 Crashworthiness of composite tube 49
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.2 Experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3 Failure modes of square tubes under crash load . . . . . . . . . . . . . . . . . . . . . . . 50
4.4 Numerical modeling in LS-DYNA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.5 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.5.1 Axial crushing results with MAT-54 . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.5.2 Axial crushing results with MAT-58 . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5.3 Axial crushing results with MAT-262 . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.5.4 Mesh sensitivity analysis in crush simulations with MAT-54 . . . . . . . . . . . . . 63
4.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5 Optimization of composite tube 67
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3 First-stage optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.1 Optimization problem definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.3.2 Mathematical formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
vii
viii Contents
1
2 1. Introduction
Figure 1.1: A selection of structural concepts to enhance the crashworthiness and their position within a CFRP fuselage [7].
of great importance since the relevant data suggested that more stringent crashworthiness
standards could contribute to a significant mitigation in injury severity [6].
Over the past two decades, design optimization has been developed as a powerful tool
to seek a highest possible crashworthiness for various aircrafts. To formulate the optimiza-
tion problem, one has to firstly specify the crashworthiness related metrics. In this study,
crashworthiness is limited to the energy absorbing performance of a composite tube under
impact load. The key parameters of crashworthy composite structures include the energy
absorption (EA), specific energy absorption (SEA), peak crushing force (PCF), mean crushing
force (MCF) and crush efficiency (CE). These key parameters can be expressed as follows,
where 𝐹(𝑠) is the impact force at the crush displacement 𝑠, and 𝑑 is the total crush displace-
ment
𝐸𝐴(𝑑)
𝑆𝐸𝐴(𝑑) = (1.2)
𝑚
where 𝑚 is the mass of the structure the PCF is the maximum crushing load in the load-
displacement curve; the MCF is the mean average crushing force; the crush efficiency can
be interpreted as the ratio of the mean crush forces to the peak crush forces.
𝐸𝐴(𝑑)
𝑀𝐶𝐹(𝑑) = (1.3)
𝑑
𝑃𝐶𝐹
𝐶𝐸(𝑑) = 𝑋100% (1.4)
𝑀𝐶𝐹
They are all very important in the crashworthy design of composite fuselage. For example,
in the design of a fuselage, the SEA combines the effect of energy absorbing capability and
the mass. It determines how advantageous the design is in terms of the crashworthiness
performance and lightweight design. The PCF is also an important parameter to be correlated
to the maximum floor-level accelerations that threatens the occupants’ safety. The MCF is
also a good indicator of the crush response, reflecting the crashworthiness performance.
The CE represents the nature of crush response, and usually it should be constrained within
a prescribed threshold to limit high load levels on the occupants. These crashworthiness
related metrics can be either optimization objectives or constraints, depending on the purpose
of the optimization.
1.2. Optimization strategies 3
𝑚𝑖𝑛 𝑓(𝑥)
{ 𝑠.𝑡. 𝑔(𝑥) ≤ 0 (1.5)
𝑥 ≤𝑥
where 𝑓(𝑥) and 𝑔(𝑥) are the objectives and constraints, respectively; x denotes the design
variables.
The design variables can have multiple choices, usually including the size, shape, layup
and topology of the composite structures. The optimization problem can be performed under
single or multi-objective, with or without uncertainties.
In a optimization problem, deterministic optimization has the tendency to drive the can-
didate solutions toward constraints until they are active. Therefore it leaves no space for
considering uncertainties in a optimization problem. To compensate for this downside, re-
searchers have developed a reliability-based optimization (RBO) approach to use probabilis-
tic constraints instead, representing that probability of design infeasibility is restricted to a
prescribed level [12]. This reliability-based optimization (RBO) has been widely adopted in
crashworthiness problems, and it can be mathematically expressed as:
𝑚𝑖𝑛 𝑓(𝑥)
{ 𝑠.𝑡. 𝑃(𝑔(𝑥) ≤ 0) ≥ 𝑅 (1.6)
𝑥 ≤𝑥
where 𝑓(𝑥) and 𝑔(𝑥) are the objectives and constraints, respectively; x denotes the design
variables; 𝑅 denotes the reliability level; 𝑃 represents the probability function of satisfying
the constraints 𝑔(𝑥) ≤ 0.
Besides the reliability-based optimization (RBO), uncertainties may also appear in the
crashworthiness performance of the structures, which sometimes contributes to significant
fluctuations from the desired performance. These uncertainties could also increase the life-
cycle costs, such as maintenance and repair expenses [13]. To consider this source of un-
certain variability, researchers proposed the concept of robust design optimization (RDO)
to reduce the fluctuations of the structural performance. Recently, the robust design opti-
mization has been adopted in the uncertainty-based multidisciplinary design optimization
for aerospace vehicles [14]. This approach made the structural performance of aerospace ve-
hicles less sensitive to variations of variables beyond the control of design engineers. Math-
ematically, a general robust design optimization (RDO) problem can be expressed as:
heuristic non-gradient method was proposed by reference [18] to vary the density within the
design domain for a prescribed distribution of plastic strains and stresses with a mass con-
straint. Another non-gradient method was proposed to optimize plate or shell structures, by
using thickness as the design variable with one base material [19]. Additionally, researchers
proposed a graph and heuristic based topology optimization method to optimize the profile
cross-sections of crashing structures [20]. Two different loops are characterized in their op-
timization problem. The outer loop is consisted of an expert-knowledge-based optimization
in terms of the topology and shape of the structure, while the inner loop performs the size
and shape optimization. However, this topology optimization method is only limited on the
cross-sections of structures. Furthermore, continuum-based topology optimization for struc-
tural crashworthiness has been addressed by proposing a heuristic (non-gradient) approach
[21]. In their work, all elements of the structure would contribute to the energy absorption
through plastic deformation, and therefore the optimal solution was obtained by determining
a uniform internal energy density in the whole structure.
Topology optimization could be one of the most difficult optimization problem in crash-
worthiness. This could be attributed to considerable complexity of obtaining topological sen-
sitivity or optimality criteria for effectively capturing the dynamic crushing process. Most
recently, a novel approach using evolutionary optimization algorithms and a geometric Level-
Set Method has been proposed in crashworthiness Topology Optimization [22]. Level-set
Methods (LSMs) for structural Topology Optimization define the interfaces between material
phases implicitly by iso-contours of a Level-set Function (LSF) [23]. The Level-set Methods
have been implemented in different Topology optimization problems of crash-loaded struc-
tures, 3D self-supporting structures for additive manufacturing, [24, 25], etc. A energy max-
imization problem for a rectangular beam is considered to evaluate the utilization of standard
Evolution Strategies and the state-of-the-art Covariance Matrix Adaptation Evolution Strat-
egy. The beam is fixed at both ends and under impact loading in the middle by a cylindrical
pole. The results demonstrate that the evolutionary optimization methods are proven to be
efficient in an optimization problem of crash-loaded structures, while explicitly expressing
the objective function.
• Investigate the applicability of single-layer approach in modeling CFRP square tubes for
a balance of accuracy and efficiency.
• Study the applicability of three LS-DYNA material models for crush simulations of CFRP
square tubes, in terms of the correlation to the experiment and minimum calibration
cost.
• Develop different surrogate models to map the input design variables into the output
crashworthiness metrics.
• Search the design space for the optimal set of design variables by using population-
based heuristic optimization algorithms.
Since this study is performed numerically, the material and experimental results involved
will be directly selected from literature. The methodology of this thesis can be divided into
four different steps:
1. The first step is to validate the material models used in LS-DYNA. Corresponding coupon-
level simulations in LS-DYNA will be performed to characterize the material properties
for the carbon/epoxy composite material IM7/8552.
2. The second step is to construct a more complicated square tube model in LS-DYNA
with validated material model in step 1. The numerical model will be established and
6 1. Introduction
calibrated based on the experimental results from literature. Meanwhile, detailed sen-
sitivity analysis of influential parameters in different material models (MAT-54, MAT-58
and MAT-262) will also be performed, to have a more comprehensive understanding of
the complex failure mechanism.
3. The third step is to build different surrogate models to formulate the mapping between
input design variables and output crashworthiness metrics. Two models will be created
for the regression analysis, including Deep Neural Networks (DNN) and Gradient Boost-
ing Regression Trees (GBRT) ensemble. The dataset is extracted from a limited number
of samples from LS-DYNA simulations by using Design of Experiments (DOE). Different
metrics will be evaluated to pursue a model with the best regression performance.
4. The last step is to find the optimal design variables by optimizing the constructed sur-
rogate models in the design space. Gradient-free optimization algorithms will be imple-
mented in this step. The obtained optimal results will be tested in LS-DYNA to check
if the crashworthiness performance is superior to the baseline LS-DYNA model deter-
mined in step 2.
Literature study — crashworthiness of
2
composite structures
2.1. Introduction
The purpose of this chapter is to introduce the literature study of the topic covered in this
study: crashworthiness of composite structures. Three individual sections will be involved,
including experimental investigation, numerical simulations and surrogate-model based op-
timization. The composite structure focused in this study will be component-level square
tubes.
7
8 2. Literature study — crashworthiness of composite structures
and fiber arrangements are well researched. Among all these combinations of fiber and ma-
trix, thin walled CFRP tubes are proven to be efficient, lightweight and controllable in energy
absorption. These papers describe detailed crushing process, efficiency and strain-rate sen-
sitivity of thin walled cylindrical composite tubes, indicating a milestone for crashworthiness
tests. However, the configurations of these tests remain on a limited scope as the influences
of triggering, direction of loading, geometrical parameters, and other factors were not inves-
tigated. Also some of the observations concerning the crushing speed sensitivity could not
be fully understood and interpreted.
Later research on crashworthiness characteristics of composite tubes focused on fiber
types, matrix types, fiber volume, production conditions, test speed, and triggering [1]. The
specific energy absorption and fracture toughness are compared among these combinations
of parameters. This paper mainly reveals the complexity and difficulty of designing such
experiments for optimization, in that considerable number of experiments are required. Fur-
ther research regarding the influence of different laminate stacking sequences, fiber volume,
strain rate and geometric parameters was performed by testing square CFRP tubes under
static and dynamic crush loads [29]. It shows the brittle nature of CFRP tubes under crush
test instead of the buckling failure metallic structures. By comparing the crashworthy pa-
rameters in different failure modes, this paper finds that the stable crushing collapse mode
makes the primary contribution to the energy absorption performance. Besides, the exper-
imental results demonstrate that the peak force has a positive correlation with the number
of fiber layers and the thickness of the tubes. The study provides an extensive insight into
the effect of different influential factors on the crushing response, collapse modes, and the
energy absorbing capability. Limitations of this experimental study are that the influence of
some other geometrical parameters and cross-section shapes was not considered. Further-
more, the parameters investigated in the results do not reach the optimum due to a limited
number of experiments performed.
The laminate properties and the geometry of thin walled composite tubes significantly af-
fect the energy absorbing capability. Depending on different applications, the cross-sections
of these tubes can come in square, circular, conical, and radially corrugated ones. Other ge-
2.2. Experimental investigation 9
Figure 2.2: MD-500 helicopter with energy absorbers attached externally to underside of the A/C [28]
ometrical parameters affecting the crashworthiness performance of the tube include trigger
configuration trigger radius, thickness-to-diameter ratio. To investigate the influence of the
aforementioned parameters on a much deeper level, researchers performed comprehensive
experiments in all aspects on the composite tubes, including laminate properties, strain rate
effect, cross-section, types of triggers, corrugation, and other geometrical parameters.
Figure 2.3: Tearing mode, socking mode, splaying mode and microframentation mode [31].
ously mentioned fiber angles, architectures have effects on the energy absorbing capability
of composite tubes. The resin system was observed as the one of the most influential factors
affecting energy absorption of composite tubes. Significant differences in SEA was found for
carbon or glass NCF (Non-crimp fiber) composite tubes with polyester, vinyl-ester and epoxy
resin systems. Furthermore, this study proved that NCF and biaxially braided reinforce-
ments did not perform well under crushing loads, resulting in a relatively low SEA compared
with metallic specimens.
Synthesizing the advantages of CFRP and GRFP tubes, hybrid configurations showed great
potential for further crashworthiness improvement. The effect of delamination of the hybrid
twill-weave and unidirectional carbon/epoxy composite box structures on the energy ab-
sorption capability was investigated [33]. The hybrid laminate designs with higher fracture
toughness in Mode-I and Mode-II delamination tests, exhibit more advantageous energy ab-
sorption capability. The effect of different layups on the lamina bending crushing mode was
also investigated, as shown in Fig 2.4. The dynamic crush experiments were conducted [34]
on the crashworthy performance of CFRP tubes, GFRP tubes and hybrid tubes. The effect of
ply orientation, UD ply proportion, resin type and thickness on collapse modes and energy
absorption are analyzed. The research concludes that the SEA has an inverse correlation
with the thickness under dynamic load cases, and the hybrid G803/3234 fabric coupled
with G827/3234 axial tapes demonstrate advantageous energy absorbing capability, this is,
SEA=116.9 kJ/kg and SEA=83 kJ/kg under quasi-static and dynamic load case, respec-
tively. Filament wound carbon/aramid epoxy hybrid composite tubes was also tested under
dynamic crushing loads in terms of the crashworthiness [35], due to the potential possibility
of reducing manufacture cost. Different influential parameters including crushing speed,
temperature treatment, raw material, hybrid ratio, fiber orientation and tube thickness have
been investigated. A hybrid carbon/aramid FRP tube after temperature treatment demon-
strates the highest average specific energy absorption with a controllable failure mode of 98
kJ/kg, and 82 kJ/kg for a quasi-static and dynamic impact test, respectively. These papers
demonstrate the advantages of hybrid composite tubes, indicating the possibility of incorpo-
rating hybridization into the composite design for better crashworthy performance. Down-
sides of hybridization are the increased complexity in design and manufacturing, as more
parameters are required to be included which drastically increases the cost of experimental
validation.
Apart from hybridization, laminate properties such as fiber orientations and thickness
have a direct influence on the crashworthy performance. Multiple crushing tests with CFRP
(G827/5224) composite tubes in different fiber orientations and wall thicknesses were con-
ducted under quasi-static and dynamic load cases [36]. The energy absorbing capability of
2.2. Experimental investigation 11
certain structures was investigated and compared, revealing a significant effect of both fiber
orientations and the wall thickness on the crashworthy performance. However, the opti-
mization of these characteristic parameters of crashworthiness is not performed to pursue
the optimum design. The effect of fiber orientation on the energy absorbing capability of
GFRP woven composite tubes under quasi-static axial loads was experimentally investigated
[37]. The paper concludes that 75/-15 layups have advantages in energy absorbing capability
under quasi-static axial load but trigger an unsteady crushing mode under dynamic load-
ing, compared with 0/90 or 45/-45 layups. Therefore, more research regarding the dynamic
crushing cases should be conducted. To this end, the effect of fiber orientations of CFRP
circular tubes on energy absorbing capability was tested under crushing load [38]. Tubes
made of 22 UD pre-impregnated woven fabric layers with epoxy matrix were experimentally
investigated in terms of multiple combinations of different fiber orientations (0∘ , 90∘ , 45∘ ).
Results show that the number of 0∘ layers have a positive correlation with the SEA, while
the 90∘ tubes have the lowest SEA. Completely different collapse modes are also observed for
specimens with different fiber orientations. However there is still much space for laminate
optimization in terms of layup design.
Cross-section
The effect of different cross-sections of the tubes have been investigated for years. A compre-
hensive experimental investigation of these parameters can be found in reference [39], which
experimentally investigates the axial crushing performance of pultruded glass fiber/polyester
and glass fiber/vinylester tubes with circular or square cross-sections. The results show that
tubes with circular cross-sections have a better crashworthiness performance than square
tubes, and yield a 59%, 45% and 28% increase in SEA compared with square tubes under
different impact velocities (9.3, 12.4 and 14 m/s), respectively.
Type of triggers
The triggering concept was originally developed and investigated by researchers to pursue a
controllable collapse mode for maximizing the energy absorbing capability. The controllable
collapse mode means that different triggers can initialize different collapse modes, therefore
the design of a proper trigger will contribute to a favorable failure mode for higher energy
absorption. It should be noted that the triggers have significant influence on the crush
response. The most popular triggering type can be found as a 45∘ chamfering pattern and
a “tulip pattern”, as shown in Fig 2.5. 45 ∘ chamfer around the edge of the tube has been
observed to absorb 7%-9% energy than the tulip pattern trigger.
12 2. Literature study — crashworthiness of composite structures
Figure 2.5: 45∘ chamfering pattern (a) and “tulip pattern” (b) [39]
Self-contained and external triggering are two main characteristics for energy absorbers.
Axial quasi-static tests were performed [40] to evaluate the effect of bevel triggers on the ax-
ial crushing response of CFRP (T700/QY8911) tubes. Three crashworthy parameters: initial
peak specific stress, specific energy absorption and crushing load efficiency are measured for
numerical comparison. The experimental results validate the mechanism of using bevel trig-
gers to control the collapse initialization and propagation, revealing that the energy is mainly
absorbed by delamination as well as bending/fracture of the lamina bundles. The progressive
failure and crashworthiness of G803/5224 composite tubes under dynamic axial impact are
further investigated in reference [41]. This paper sets the trigger mechanism at the top of the
tube and categorizes two phases of crushing failure as initial phase of local failure and stable
phase of progressive failure. Similarly, the effects of [0/90] carbon/epoxy composite tubes
with different trigger types on SEA were investigated and validated by experimental results
[42]. The inward-folding chamfer has a positive effect on peak load reduction compared with
no chamfer or splaying chamfer. In this paper, the optimum radius of the plug-type trigger
mechanism is determined to be 2.7mm, and the optimum thickness-to-diameter (t/D) ratio is
determined to be 0.092 to achieve the maximum SEA. The research indicates the important
role the trigger concept plays in achieving a maximum SEA and more controllable collapse
mode. Further experiments on effects of the triggers are also required to be performed for
optimizing other crashworthy parameters, such as the reduction of initial peak load.
Shell elements are widely implemented in LS-DYNA because they require less compu-
tational efforts and are more suitable for modeling crushing behaviors of composites. The
modeling approaches of composite laminates can be classified into ‘single-layer’ or ‘multiple-
layer’ modeling approaches [44]. The single-layer modeling approach models the laminate
with a single shell element where each ply is represented as an integration point in the
thickness direction. Generally this modeling approach can accurately capture the unstable
collapse modes but lacks accuracy in progressive failure. The multiple-layer modeling ap-
proach models the laminate by multiple shell elements where each ply or a group of plies is
represented as a single shell element. These shell elements are bonded or tied together with
tiebreak contact definition or cohesive elements, providing a better performance in capturing
progressive crushing. However, energy absorbing prediction capability of the multiple-layer
modeling approach still needs to be improved to have a good correlation with the experimen-
tal results. It should be noted that a low-pass filter is usually needed in the post-processing
to eliminate the high-frequency oscillations of crushing response. Furthermore, the impor-
tance of material characterization on modeling the crashworthiness behavior of composite
structures has been addressed in [45].
Historically the modeling of conventional composite tubes in LS-DYNA mostly used mul-
tiple layers of shell elements with a tiebreak contact in the interface due to the possibility of
simulating material damage and remaining in an acceptable computational cost. Each shell
element layer represents one or multiple plies of the laminates. It should be noted that the
triggering mechanism is usually modeled by reducing the thickness or assigning different
material properties in the local triggering region. Different material damage models coupled
with different failure criteria are developed by researchers to accurately capture the crash-
worthy characteristics of conventional composite tubes. In the following sections, detailed
material models in LS-DYNA and their industrial applications will be discussed.
1. The material models formulated in terms of individual ply properties should be consid-
ered since the input mechanical properties can be characterized by standard experi-
ments.
2. The material models only applicable for woven fabrics or thermoplastic matrix composite
should be excluded because thermoset matrix composite has mostly elastic stress-strain
responses.
3. The material models that are not widely used in the crashworthiness analysis are ex-
cluded as well.
Therefore, several material models are selected in this section for a preliminary and general
comparison, including the following material cards:
• MAT-54/55 (*MAT_LAMINATED_COMPOSITE_DAMAGE)
• MAT-58/158 (*MAT_LAMINATED_COMPOSITE_FABRIC)
• MAT-162 (*MAT_COMPOSITE_DMG_MSC)
• MAT-219 (*MAT_CODAM2)
• MAT-261 (*MAT_LAMINATED_FRACTURE_DAIMLER_PINHO)
• MAT-262 (*MAT_LAMINATED_FRACTURE_DAIMLER_CAMANHO)
14 2. Literature study — crashworthiness of composite structures
MAT-54/55
Material models MAT-54 and MAT-55 are enhanced composite damage models, applicable
for arbitrary orthotropic materials and UD layers defined in shell elements. Multiple fail-
ure modes including tensile fiber mode, compressive fiber mode, tensile matrix mode and
compressive matrix mode can be specified, depending on the selection of [Chang-Chang]
or [Tsai-Wu] failure criteria. The former is the default criteria for MAT-54, and the latter
for MAT55. Detailed implementation of these failure criteria can be found in the literature
[46, 47], respectively. According to the literature, the Chang-Chang criteria is given the fol-
lowing form.
𝜎 𝑌 𝜎
𝜎 <0⇒𝑒 =( ) + [( ) + 1]𝜎 𝑌 + ( ) −1 (2.4)
2𝑆 2𝑆 𝑆
where if the damage parameter 𝑒 ≥ 0 ⇒ 𝑓𝑎𝑖𝑙𝑒𝑑, else elastic; 𝐸 and 𝜎 denote the modulus,
respectively; the rest symbols are strength values; 𝑡 and 𝑐 denote tension and compression,
respectively; 𝑎, 𝑏 and 𝑐 denote the direction 1,2 and 3 in the material axis.
In the Tsai-Wu (MAT-55) criteria the tensile and compressive fiber modes are same as in
the Chang-Chang criteria. The failure criterion for the tensile and compressive matrix mode
is given in the following form.
𝜎 𝜎 (𝑌 − 𝑌 )𝜎
𝑒 = +( ) + −1 (2.5)
𝑌 𝑌 𝑆 𝑌 𝑌
where if the damage parameter 𝑒 ≥ 0 ⇒ 𝑓𝑎𝑖𝑙𝑒𝑑, else elastic; 𝐸 and 𝜎 denote the modulus,
respectively; the rest symbols are strength values; 𝑡 and 𝑐 denote tension and compression,
respectively; 𝑎, 𝑏 and 𝑐 denote the direction 1,2 and 3 in the material axis.
The MAT-54/55 assumes a linear elastic stress-strain relation prior to failure, without any
pre- or post-peak softening. Besides the input parameters including mechanical properties,
these aforementioned failure criterion and material behaviors are implemented by introduc-
ing many additional input parameters in the material cards, which usually require further
calibration. A summary of the meaning and the description of these additional parameters
is listed in Table 2.1. In MAT-54, failure can occur in any of four different ways:
2. DFAILT>0, failure occurs if tensile fiber strain is larger than DFAILT or less than DFAILC
4. TFAIL>0, failure occurs if current element time step is smaller than TFAIL
2.3. Numerical investigation in LS-DYNA 15
Parameter Definition
DFAILT Max strain for fiber tension
DFAILC Max strain for fiber compression
DFAILM Max strain for the matrix in tension and compression
DFAILS Max shear strain
TFAIL Element is deleted when the time step is smaller than the given value
EFS Effective failure strain
SOFT Crush front strength reduction parameter in the crashfront elements
SOFT2 Optional transverse softening reduction factor
PFL Percentage of layers which must fail until crashfront is initiated
ALPHA Shear stress non-linear term
BETA Weighting factor for shear term in tensile fiber mode
SLIMT1 Factor to determine the minimum stress limit after stress maximum(fiber ten-
sion)
SLIMC1 Factor to determine the minimum stress limit after stress maximum(fiber com-
pression)
SLIMT2 Factor to determine the minimum stress limit after stress maximum(matrix ten-
sion)
SLIMC2 Factor to determine the minimum stress limit after stress maximum(matrix com-
pression tension)
SLIMS Factor to determine the minimum stress limit after stress maximum(shear)
FBRT Softening factor for fiber tensile strength after matrix failure
YCFAC Softening factor for fiber compressive strength after matrix failure
16 2. Literature study — crashworthiness of composite structures
MAT-58/MAT-158
Material model 58 is a damage mechanics-based model, which can be implemented to model
shell, solid or thick shell elements depending on the type of failure surface. Unlike MAT-54,
MAT-58 integrates pre- and post-failure softening to add non-linear material behaviors in the
model.It is common to utilize this model for UD layers, complete laminates and woven fabrics.
Detailed implementation of MAT-58 can also be found in LS-DYNA user manual [48]. Even
though MAT-58 offers special control of shear behavior of fabrics, but this part will not be
covered since the composite utilized in this study will have UD plies besides the mechanical
properties, the additional parameters are summarized in Table 2.2. Several remarks can be
addressed in MAT-58 as follows:
1. The failure of element layers are controlled by the maximum effective strain where the
layer in the element is completely removed after the ERODS reaches maximum (com-
pression /tension including shear).
2. The SLIMxx parameters denote as the stress limits to limit the stress in the softening
part to a specific value, which captures the elastoplastic material behavior.
3. The crashfront algorithm is started if a value for TSIZE is specified.
4. MAT-158 is a updated version for MAT-58, offers additional input parameters to account
for the strain-rate effects.
Variable Definition
SOFT Crush front strength reduction parameter in the crashfront elements
TSIZE Element is deleted when the time step is smaller than the given value
ERODS Maximum effective strain for element failure. If lower than zero, element fails
when effective strain calculated from the full strain tensor exceeds ERODS
SLIMT1 Factor to determine the minimum stress limit after stress maximum(fiber ten-
sion)
SLIMC1 Factor to determine the minimum stress limit after stress maximum(fiber com-
pression)
SLIMT2 Factor to determine the minimum stress limit after stress maximum(matrix ten-
sion)
SLIMC2 Factor to determine the minimum stress limit after stress maximum(matrix com-
pression tension)
SLIMS Factor to determine the minimum stress limit after stress maximum(shear)
MAT-162
MAT-162 developed by Material Science Corporation and University of Delaware Center for
Composite Materials, is a rate-dependent damage model to capture the progressive failure
for composites made from UD layers and woven fabrics [49], especially widely used in model-
ing GFRP composites. This material model has proven to effectively simulate fiber dominate
failures, matrix damage and considers a stress-based delamination failure criteria, without
the requirement of knowing the locations of delamination in advance. The progressive layers
failure adopts the Hashin failure criterion [50] with a generalization to account for the effect
of highly constrained pressure on composite failure. The MLT damage mechanics approach
in contrast to [51] has been adopted to characterize the softening behavior after damage ini-
tiation. In contrast to MAT-54 and MAT-58, MAT-162 is only applicable for solid elements,
which offers the opportunity to model the delamination behavior under all conditions, in-
cluding opening,closure and sliding of failure surfaces. Since MAT-162 is based on phys-
ical models, a more comprehensive set of input parameters in all directions are required,
besides the in-plane material properties. This is also summarized in Table 2.3, excluding
2.3. Numerical investigation in LS-DYNA 17
the strain-rate related parameters. Detailed formulation of the material model, UD lamina
damage functions and damage progression criteria can be found in the LS-DYNA Keyword
manual[48]. The failure criterion of MAT-162 damage model can be expressed in Figure2.6.
It should be noted that the element erosion in MAT-162 follows three different ways:
1. If fiber tensile failure in a unidirectional layer is predicted in the element and the axial
tensile strain is greater than E_LIMIT.
2. If compressive relative volume (ratio of current volume to initial volume) in a failed
element is smaller than ECRSH.
3. If expansive relative volume in a failed element is greater than EEXPN.
Parameter Description
RO Mass density
EA 𝐸 , Young’s modulus – longitudinal direction
EB 𝐸 , Young’s modulus – transverse direction
EC 𝐸 , Young’s modulus – through thickness direction
PRBA 𝜈 , Poisson’s ratio ba
PRCA 𝜈 , Poisson’s ratio ca
PRCB 𝜈 , Poisson’s ratio cb
GAB 𝐺 , shear modulus ab
GBC 𝐺 , shear modulus bc
GCA 𝐺 , shear modulus ca
SAT Longitudinal tensile strength
SAC Longitudinal compressive strength
SBT Transverse tensile strength
SBC Transverse compressive strength
SCT Through thickness tensile strength
SFC Crush strength
SFS Fiber mode shear strength
SAB Shear strength, ab plane
SBC Shear strength, bc plane
SCA Shear strength, ca plane
SFFC Scale factor for residual compressive strength
AMODEL Material models:EQ. 1: Unidirectional layer modelEQ. 2: Fabric layer model
PHIC Coulomb friction angle for matrix and delamination failure
S_DELM Scale factor for delamination criterion
E_LIMT Element eroding axial strain
ECRSH Limit compressive relative volume for element eroding
EEXPN Limit expansive relative volume for element eroding
OMGMX Limit damage parameter for elastic modulus reduction
MAT-219
Material model 219 is the second generation of the UBC Composite Damage Model(CODAM2)
for brick, shell and thick shell elements developed by the University of British Columbia[53].It
is developed for predicting the non-linear in-plane response of composite laminates using
continuum damage mechanics, addresses the deficiencies in both the numerical and ma-
terial objectivity issues that the original CODAM[54] formulation faced. This model is a
18 2. Literature study — crashworthiness of composite structures
sub-laminate-based continuum damage mechanics model for fiber reinforced composite lam-
inates made up of transversely isotropic layers, using an optional non-local averaging and
element erosion. Furthermore, damage accumulation and reduction of elastic moduli are
tracked separately with functions based on the different failure modes and coordinate direc-
tions. The following assumptions have been made in this model[48]:
1. The material is an orthotropic medium consisting of a number of repeating units through
the thickness of the laminate, called sub-laminates.
2. The nonlinear behavior of the composite sub-laminate is only caused by damage evolu-
tion. Nonlinear elastic or plastic deformations are not considered.
MAT-219 adopts the framework of non-local strain-softening formulations, with the con-
siderations of all failure modes, including intra-laminar (fiber and matrix damage), overall
sub-laminate modes(locking or delamination). These failure modes are integrated into dif-
ferent functions of the non-local equivalent strain parameters. Detailed formulation of the
mapping between the evolution of damage mechanisms and equivalent strain parameters,
as well as the element erosion mechanisms can be found in LS-DYNA user manual[48]. A
summary of the input parameters besides the mechanical properties is listed in Table2.4 for
a better understanding.
MAT-261/262
MAT-261/262 utilizes an orthotropic continuum damage model for shell,thick shell and solid
elements in modeling laminated-reinforced composites [55, 56]. It is based on a physical
model for each failure mode and takes non-linear in-plane shear into considerations. Specif-
ically, this model assumes bi-linear post-peak softening for the longitudinal direction and
linear softening for the transverse direction and shear. The difference between MAT-261 and
MAT-262 is that in MAT-261 different functions are used for each failure mode that are then
combined into a constitutive model through a strength method, while in MAT-262, an energy
method is implemented to generating damage functions in different coordinate directions.
Two different damage evolution laws are implemented in MAT-262: one for the bi-linear rela-
tion in longitudinal direction, the other for the linear relation in the transverse direction and
in-plane shear. The damage evolution law and in-place shear behavior defined in MAT-261
and MAT-262 can be found in the Figs 2.8 and 2.9, respectively.
2.3. Numerical investigation in LS-DYNA 19
Parameters Description
NERODE Failure and erosion flag
NDAM Damge flag
EPS1TF Failure strain in tension along the a-direction
EPS2TF Failure strain in tension along the b-direction
EPS3TF Failure strain in tension along the c-direction
EPS1CF Failure strain in compression along the a-direction
EPS2CF Failure strain in compression along the b-direction
EPS3CF Failure strain in compression along the c-direction
EPS12F Failure shear strain in the ab-plane
EPS23F Failure shear strain in the bc-plane
EPS13F Failure shear strain in the ac-plane
EPSD1T Damage threshold in tension along the a-direction
EPSC1T Critical damage threshold in tension along the a-direction
CDAM1T Critical damage in tension along the a-direction
EPSD2T Damage threshold in tension along the b-direction
EPSC2T Critical damage threshold in tension along the b-direction
CDAM2T Critical damage in tension along the b-direction
EPSD3T Damage threshold in tension along the c-direction
EPSC3T Critical damage threshold in tension along the c-direction
CDAM3T Critical damage in tension along the c-direction
EPSD1C Damage threshold in compression along the a-direction
EPSC1C Critical damage threshold in compression along the a-direction
CDAM1C Critical damage in compression along the a-direction
EPSD2C Damage threshold in compression along the b-direction
EPSC2C Critical damage threshold in compression along the b-direction
CDAM2C Critical damage in compression along the b-direction
EPSD3C Damage threshold in compression along the c-direction
EPSC3C Critical damage threshold in compression along the c-direction
CDAM3C Critical damage in compression along the c-direction
EPSD12 Damage threshold for shear in the ab-plane
EPSC12 Critical damage threshold for shear in the ab-plane
CDAM12 Critical damage for shear in the ab-plane
EPSD23 Damage threshold for shear in the bc-plane
EPSC23 Critical damage threshold for shear in the bc-plane
CDAM23 Critical damage for shear in the bc-plane
EPSD31 Damage threshold for shear in the ac-plane
EPSC31 Critical damage threshold for shear in the ac-plane
CDAM31 Critical damage for shear in the ac-plane
20 2. Literature study — crashworthiness of composite structures
In this study, we mainly focus on MAT-262 that implements the LaRC04 criteria [57] as
damage activation functions in the constitutive material model. The constitutive law of the
material model 𝜖 = 𝐻 ∶ 𝜎 can be in the form of the lamina compliance tensor:
⎡ ( )
− 0 ⎤
𝐻=⎢ − ( )
0 ⎥ (2.6)
⎢ ⎥
⎣ 0 0 ( ) ⎦
It assumes that the elastic domain is enclosed by four sub-surfaces representing differ-
ent failure mechanisms, including longitudinal fiber tension, longitudinal fiber compression
(3D kinking model, transformation to fracture plane), transverse matrix failure:transverse
tension, and transverse matrix failure:transverse compression/shear, as demonstrated in
Fig 2.7. Detailed mathematical formulations of the four failure sub-surfaces are given as
follows:
𝜎 − 𝜈̃ 𝜎 ̃
𝜙 = (2.7)
𝑋
⟨|𝜎̃ | + 𝜈 𝜎̃ ⟩
𝜙 = (2.8)
𝑆
√(1 − 𝑔) ̃ ̃ ̃
+ 𝑔( ) +( ) , 𝜎̃ >0
𝜙 ={ (2.9)
⟨| ̃ | ̃ ⟩
, 𝜎̃ <0
𝜏 𝜏
𝜙 = √( ) + ) (2.10)
x x
where 𝜙 are the activation functions for each orientation 𝑖 = 1, 2 and loading direction𝑠 =
+, −. The failure is assumed to occur when 𝜙 > 1; the effective stress tensor 𝜎̃ = 𝐻 ∶ 𝜖 and
𝐻 is the undamaged laminate compliance matrix; 𝑔 is the fracture toughness ratio defined
as ;𝜏 and 𝜏 are the effective stresses computed as:
The fracture angle 𝛼 is approximately 53∘ in uniaxial compression [55]. The additional pa-
rameters in MAT-262 are listed in Table 2.9.
2.3. Numerical investigation in LS-DYNA 21
Figure 2.7: 4 ply fracture planes considered in the material model 262 [55].
Figure 2.8: Damage evolution law and in-plane shear behavior in MAT-261
Figure 2.9: Damage evolution law and in-plane shear behavior in MAT-262
22 2. Literature study — crashworthiness of composite structures
Variable Description
DAF Flag to control failure of an integration point based on longitudinal (fiber) tensile
failure
DKF Flag to control failure of an integration point based on longitudinal (fiber) com-
pressive failure
DMF Flag to control failure of an integration point based on transverse (matrix) fail-
ure.
EFS Maximum effective strain for element layer failure. A value of unity would equal
100% strain.
SOFT Crush front strength reduction parameter in the crashfront elements.
PFL Percentage of layers which mush fail until crashfront is initiated.
GXCO Fracture toughness for longitudinal (fiber) compressive failure mode to define
bi-linear damage evolution.
GXTO Fracture toughness for longitudinal (fiber) tensile failure mode to define bi-
linear damage evolution.
XCO Longitudinal compressive strength at inflection point (positive value).
XTO Longitudinal tensile strength at inflection point.
Figure 2.10: Typical stress-strain curve comparison among MAT-54,MAT-58 and MAT-262[58]
24 2. Literature study — crashworthiness of composite structures
square tubes with different lengths, wall thicknesses, laminate layups and fiber volume con-
tent was virtually tested to model three collapse mode individually, including progressive
end-crushing with tube wall laminate splaying, local tube buckling and mid-length collapse.
Results indicate that the numerical model can capture the crashworthiness characteristics
on a satisfactory level, compared with the experimental results. The study presents a com-
prehensive study concerning the modeling approaches of different collapse modes and points
out the importance of calibrating the material model parameters in the dynamic load case.
Meanwhile the strain rate effect on the mid-length collapse mode was experimentally investi-
gated, revealing a direct influence on the peak load and SEA. Unfortunately, material model
55 has no means to account for strain rate effects, which requires further improvement.
MAT-58
Researchers also developed another damage model and incorporated it into LS-DYNA for bet-
ter predictions. Reference [61] developed a continuum damage mechanics based model in
LS-DYNA to mimic the axial compression response of plug-imitated braided composite tubes.
The results show good agreement with the experiments but the same model had difficulty in
modeling the crushing response of braided composite tubes without plug initiator. The sim-
ilar findings were observed by reference [62], who implemented MAT58 in LS-DYNA based
upon Matzenmiller’s damage mechanics model and found the SEA value of braided tubes
without plug initiator is 30%-40% lower than he experimental results. Reference [63] later
modified the previously developed model to compensate for the aforementioned limitation
by incorporating a debris wedge and a tiebreak contact interface in the model for capturing
splaying mode and delamination, respectively. The results of the enhanced model presents
a satisfactory level of agreement with the experiments in both the braided tubes with or
without plug initiator. The study offers a great method to understand different response
mechanisms on an in-depth level. Further improvement would be focused on the damage
evaluation in the braided tubes under different crush load conditions. Based upon the pre-
viously developed CDM model with a strain softening approach, reference [64] calibrated
the input parameters of the model by a series of experiments to better capture the energy
absorbing capability of carbon/epoxy [0/±45] braided composite tubes. Experiments for cal-
ibration includes tensile fracture, standard coupon and 4-point bend test. The determined
input material stress-strain relations provides a significant enhancement for reflecting the
constituent damage growth on the LS-DYNA code. This helps the predicted SEA values have
only 4% variance compared to the experimental results, even though the predicted load ratios
(peak load/average load) are not as accurate as the predicted SEA, which can be attributed
to the difficulty in modeling chamfer and debris wedge. The results conclude that the energy
is primarily absorbed by material deterioration, then by friction, and eventually by delami-
nation (only 1-3 percent). A comparison of the experimental and numerical failure mode for
2-ply and 4-ply models without a plug are shown in Fig 2.11, revealing a mirrored failure
morphology and therefore a physical simulation result. The limitations of the model are that
the model has not investigated the effect of very high load rate on the energy absorbing capa-
bility as it makes load rate independency assumptions for simplicity, and the chamfer region
can still be refined for better load ratio predictions.
MAT-54
Besides the material model developed above, material model 54 with Chang-Chang failure
criteria in LS-DYNA has wider applications in predicting crashworthy performance of com-
posite tubes. Pultruded glass fiber tubes overwrapped with ±45∘ braided CFRP or GFRP sub-
ject to quasi-static and dynamic axial crushing load were numerically modeled to investigate
the effect of different parameters on crashworthy performance [65]. As used in the pre-
vious models, Belytschko-Tsay quadrilateral shell elements were still utilized to model the
tube(two layers of 0∘ laminates) and braids(four layers of ±45∘ laminates) where specific con-
tact algorithms were assigned to connect these shell elements to the achieve the designed
architecture. Material model 54 coupled with Chang-chang failure criteria in LS-DYNA was
chosen to account for the material degradation. Numerical results presents a correlation
between the response and influential parameters including the length, braid thickness and
2.3. Numerical investigation in LS-DYNA 25
Figure 2.11: Comparison of experimentally observed and predicted failure morphologies in 2-ply (a and b) and 4-ply (c and d)
tubes crushed without plug initiator [64]
loading conditions. Braiding effect has also been found to have a positive effect on the crash-
worthiness enhancement in the dynamic load cases. However, these numerical results were
not validated by the experiments and thus lacked of robustness.
CFRP composite tubes with bevel triggers were proved to be effectively modeled by LS-
DYNA material model 54 using Chang–Chang failure criteria under axial quasi-static crush-
ing load [40]. The triggering mechanism is modeled by two-layer shell elements with different
geometries, which can effectively initiate a stable progressive crushing response. The results
of the numerical simulation comply with the experimental results in terms of crashworthy
characteristics, as well as collapse modes. Even though the load conditions only involve
quasi-static crushing load, this study is still inspiring in modeling the trigger mechanism by
using different geometries in each layer. The other downside is that the failure mode presents
significant local buckling instead of brittle failure observed in the experimental results.
Reference [33] investigated the effect of Mode-I and Mode-II interlaminar fracture tough-
ness on the energy absorbing capability by using LS-DYNA. Hybrid woven and unidirec-
tional carbon/epoxy composite box was modeled by material model 54 coupled with modi-
fied Chang-Chang failure criteria (including four categorized failure modes) in LS-DYNA. To
capture the delamination behavior, this study constructed two layers of shell elements with
surface-to-surface tiebreak contact for the interface bonding. The results reveal that the
hybrid composite box should have higher fracture toughness in Mode-I and Mode-II delam-
ination tests to pursue a better crashworthy performance. The study presents a thought-
provoking approach to model the delamination failure which is difficult to deal with in con-
ventional shell theory. More research still has to be done on the dynamic load conditions to
verify the observations in multiple load cases.
Similar modeling approach can also be found in reference [44], which developed a FE
model to numerically investigate the energy absorbing ability of graphite/epoxy laminated
circular tubes with chamfer and combined trigger mechanisms. Three blocks of shell element
layers with reduced thickness in the triggering region are used to model the laminated tubes.
Material model 54 with Chang-Chang failure criteria are applied in the model to determine
the failure of each ply associated with the integration point through the thickness direction
in each block. The results show that only the matrix failure strain needs to be calibrated to
obtain a more accurate model. After the adjustment, the model can accurately predict the
26 2. Literature study — crashworthiness of composite structures
load-displacement relations, SEA and deformation of the crushing behavior. However, the
triggering region can still be refined with respect to each layer of the laminates to further
improve the simulation results. Future work might be focused on improving the accuracy
and reducing the computational cost at the same time to optimize the current FE model.
To compare the performances of single-layered modeling and multiple-layered modeling
approach, reference [66] investigated the crashworthiness behavior of the glass-polyester
crash tube under impact loading with these two approaches, respectively. Results indicate
that the single-layered approach shows consistency with the experiments but non-physical
failure modes appear, meanwhile the failure modes predicted by multiple-layered approach
coincide with the experiments but the energy absorption shows 20% variance. This study
also investigated the sensitivities of input parameters (TFAIL, SOFT, etc.) on each modeling
approach. By calibrating the input parameters, the model can be further improved by trial-
and-error calibration based upon the sensitivity investigation. The research reveals the fact
that one should take additional care in modeling composite tubes.
As calibrating input parameters are becoming increasing important for expanding the
potential of material model 54. Reference [38] calibrated the input parameters by flat speci-
men’s tests to model the circular CFRP tubes under axial impact loading. In this model, four
Belytschko-Tsay rectangular shell elements are utilized to represent 22 UD-layers with 6 in
the two outer layers and 5 through-thickness integration points in the two inner layers. Three
tiebreak contacts are created between each shell element to model the delamination failure,
and the trigger is modeled by vertical offsets of outer shell nodes of each layer. Material model
54 coupled with Chang-Chang failure criteria is applied to simulate the material degradation.
This study investigated the effect of individual fiber orientations on the SEA of CFRP crash
tubes, revealing a strong dependency. To pursue a consistent modeling approach, the study
investigated the transferability of the presented approach in changing geometries and ma-
terials. Results indicate the mesh dependency nature in which case the mesh size should
be the same when transferring the calibrated parameters to the model with different geome-
tries. Furthermore, the modeling approach has been proved to be correctly transferred from
CFRP tubes to GFRP tubes with a changed lay-up and a thickness, in that the GFRP tubes
simulation results also demonstrate good agreement with the experimental results after cal-
ibration. The failure mode determined from numerical simulations can be found in Fig 2.12,
showing an existence of axial cracks, transversely cracked fragments and mixed failure be-
havior. This study provides a consistent modeling approach for composite tubes which can
deal with changeable material and geometries. Further research might be related to compu-
tational cost reduction and material model optimization. Also the accuracy of the damage
model can still be improved, since the predicted load-displacement curve sometimes does not
comply with the experiments, especially in the peak loads even though the SEA values are
well captured.
Figure 2.12: Comparison of experimentally observed and predicted failure morphologies [38]
performed in order to reduce the input parameters that are negligible. The SOFT parame-
ter is found to be the most influential parameter in the model which artificially reduces the
strength of the elements at the crush front to retain stability.
Reference [68] validated a previously developed interlaminar continuum damage mechan-
ics based model in Abaqus/Explicit user subroutine VUMAT [69] to capture the crush re-
sponse of composite corrugated semi-circular webs, without requiring any calibration for
input material parameters. Similar CDM model was developed in reference [70], which were
used to determine the material degradation in the corrugated composite beam with a 45-
degree chamfer trigger under quasi-static axial impact. Multiple-layered approach is utilized
in conjunction with surface-based cohesive contact to capture the delamination failure. How-
ever, even though the crush morphology and crashworthy characteristics generally coincide
with the experiments, the study draws the importance of calibrating the deformation gradi-
ent parameter of the separation based criteria by trial-and-error, in order to obtain a better
prediction than the stress based criteria.
DLR composite crush element
Despite that this literature study mainly involves numerical simulation from commercial FE
code LS-DYNA, it has difficulty in finding literature about LS-DYNA models since most of
the open-source literature concerning the numerical simulation of Composite DLR crush el-
ements utilizes PAM-CRASH or ABAQUS/EXPLICIT solver. However, literature about other
FE code also worths mentioning as the modeling approaches share considerable similar-
ities. The carbon fiber/epoxy DLR semi-circular crush elements under axial quasi-static
crushing load were modeled by commercial FE code PAM-CRASH in reference [71], which
provided a significant modeling approach to capture the damage and failure development.
Multiple-layered shell elements were stacked with cohesive interfaces and enhanced with de-
lamination model which utilizes the material tests data from delamination behavior related
experiments conducted in DLR. The results indicate a stable progressive failure where the
simulation approximates the experimental results well in terms of energy absorption and
failure modes. Later, the effect of strain rate on the energy absorbing capabilities was in-
vestigated by reference [72], which modeled the DLR crush elements under axial dynamic
crushing load with different speed. To calibrate the input dynamic material parameters,
28 2. Literature study — crashworthiness of composite structures
coupon-level experiments were also conducted for each load case. Similar stacked shells
modeling methodology is utilized to capture global splaying and local fragmentation crush
mode. Results demonstrate good approximations with experimental crushing response and
physical failure modes. In the latest research, this stacked shell modeling approach has
been comprehensively deployed by the application of ABAQUS/EXPLICIT in reference [73].
It concludes that the calibration of input parameters in three main aspects is the key fac-
tor to improve the modeling techniques, including the interlaminar modeling, intralaminar
modeling and further modeling aspects regarding to mesh-dependency reduction. This pa-
per proposed a mesh modification on the DLR crush elements with off-axis orientated ele-
ments(mesh orientation=15∘ ) preventing elements in a row being damaged and eliminated
at the same time, which significantly improved the crushing response in terms of reducing
oscillations in the load-displacement curve. This implementation can be found in Fig 2.13.
However, the modeling approaches still need further improvement to capture all the failure
details, and be applied on a structural level simulation with acceptable accuracy.
Figure 2.13: Mesh modification on the DLR crush elements with off-axis orientated elements [73]
ment but the overestimates the initial peak acceleration. Meanwhile, excellent agreements
between the predictions and the experiments are observed in terms of average accelerations
and crush displacements.
To compensate for the current limitation of the model, further improvement and develop-
ment of the current model are necessary. As one of the possible enhanced model for impact
simulation, material model 213 developed by reference [76, 77] shows great potential as
it incorporates the rate and temperature dependent plasticity in the model. The material
model 213 is an Orthotropic Elasto-Plastic three-dimensional model with tabulated Input,
not requiring any empirically based input parameters. Unfortunately, the shell elements
formulation for material model 213 is still under developed, thus further investigations and
certifications will be required to implement this model in the composite crashworthiness
simulation.
Large-scale crashworthy composite structures
The large-scale crashworthiness simulation has been developed for long time, especially in
the certification of aircraft fuselage concepts. Originally developed from the crashworthy
fuselage design of the rotorcraft and light fixed-wing aircraft, traditional ”bend frame con-
cept” with multiple energy absorbing components has drawn the attention of researchers for
many years due to its superior crashworthy performance. Early numerical investigation of
such concept can be traced back to a full-scale crash simulation of an E-glass/epoxy sand-
wich composite crashworthy fuselage under 0∘ and 15∘ impact loads, which was performed
by the non-linear FE code MSC DYTRAN [78] The upper section, floor and subfloor of the
fuselage are modeled with a mixture of CQUAD4 shell elements and CHEXA solid elements.
A linear elastic-plastic material model with strain hardening is used to reflect the material
properties from the coupon tests on the FE model. The acceleration responses of the full-
scale fuselage section under 0∘ and 15∘ -roll coincide with the experimental results scaled
from a 1/5-scale model fuselage, respectively. However, limitations of the model exist in
the insufficient material property tests, lacking of modeling facesheet-core disbonding, low
computational efficiency. Following the same full-scale model, researchers investigated the
influence of seats and dummies on the acceleration response of the fuselage section [79], as
well as the seat deformation and occupant motion trends. Similar modeling techniques were
applied to validate the impact experiments of the full-scale fuselage section onto soft soil
and water surface [80]. These studies validated the robustness of utilizing MSC DYTRAN in
simulating crashworthiness tests of full-scale fuselage section, and contributed to the devel-
opment of the LS-DYNA models which provided specific enhancements based upon the MSC
DYTRAN software in terms of crashworthiness.
As the development of LS-DYNA techniques in modeling crashworthiness of large-scale
composite structures, DLR conducted full-scale numerical investigation on the crashworthi-
ness of a rotorcraft barrel section retrofitted with three different composite energy absorbing
subfloor concepts under dynamic crash load [81]. Similar LS-DYNA simulations were per-
formed by NASA Langley to compare the crash responses of barrel sections retrofitted with
”conusoid” and sinusoid sandwich composite [74]. The predicted acceleration responses of
these two energy absorbing concepts demonstrated large increases close to the end of pulses.
The average accelerations of the “conusoid” and sinusoid sandwich energy absorbers were
measured as 15.0 g and 14.2 g respectively, while the crush displacement of ”conusoid”
energy absorber showed approximately 9% increase than that of sinusoid sandwich con-
cept. The non-uniform failure mode of the ”conusiod” subfloor in the results matches the
experimental results. Later, the full-scale numerical simulation of the crash response of
Second Transport Rotorcraft Airframe Crash Testbed (TRACT 2) [74] was conducted, reveal-
ing the comprehensive failure mechanisms and the interactions of different components of
the whole aircraft onto a soil surface. The multi-direction acceleration responses of differ-
ent components were collected to validate the experimental measurements, which show that
the correlation between simulation and experiments has multiple performances ranging from
excellent to poor. Even though the study provides a comprehensive investigation of the full-
scale crash test, the model still lacks of robustness and thus requires further improvement.
Also due to the soil anomalies during the crash test, the energy absorption behaviors of ”conu-
soid” and sinusoid concepts cannot be truly restored. Currently the simulation of large-scale
30 2. Literature study — crashworthiness of composite structures
crash tests, especially the full-scale crash tests has difficulty in capturing the interactions
among different components, as well as dealing with the complicated load cases. Future
work could also be performed in reducing the computational costs as the current large-scale
LS-DYNA models still requires considerable CPU time and infeasible for optimization study
with tremendous design variables.
Figure 2.14: A a flow chart for performing crashworthiness optimization of the tubes based upon surrogate models [94]
the genetic algorithm. The KRG model was found to override the PRS model and RBF
model in the robustness and efficiency with increased space of datasets, while the approx-
imation cost remained on a higher level. Instead of selecting the surrogate models by the
trial-and-error, reference [97] proposed to select the surrogate models under data uncertain-
ties, based upon a Bayesian metric in probability theory together with root mean square error.
This selection method was verified by the crashworthiness optimization of a full-frontal im-
pact event, leading to the observation of objective dependent nature of the surrogate models.
Different surrogate models were found to be suitable for different optimization targets, includ-
ing Chest G, crash distance and toe-board intrusion. Similar conclusion was proposed by
reference [98].Therefore, the synthetization of multiple surrogate models were recommended
by reference [99], since the surrogate model with best accuracy might not contribute to the
most robust results when searching the optimum. This recommendation can be verified
in the application of crashworthiness optimization with MOPSO algorithm of the foam-filled
multi-cell thin-walled structures [100], as well as the multi-objective surrogate-based opti-
mization of the crashworthiness of a hybrid impact absorber [101]. The former performed a
comparison among different surrogate models including PRS, RBF, KRG and SVR in terms of
model accuracy by evaluating the relative error (RE) and the root mean square error (RMSE).
The RBF model was observed to be the best in terms of the minimum peak crushing forces
and maximum SEA, but it also found that the error between the surrogate model and the
FEA results might violate the constraints, which requires further considerations to prevent
such violations. Nevertheless the latter compared the performance of PRS, KRG and MARS
model in towards maximum SEA and minimum load ratio(LR) or mean peak force. In this
case the MARS model was the only solution to successfully capture the complex shapes of
the LR responses, while the SEA responses even can be accurately captured by the sim-
plest quadratic PRS model. The KRG could approximate the noises on a satisfactory level
with a minimum sample size. In summary, it is suggested to simultaneously perform the
crashworthiness optimization by multiple surrogate models, and select the most appropriate
model depending on the required objectives and cases. For complex composite structures,
the large number of design variables and the mixed-discrete nature of the design variables
usually require specific considerations. Furthermore, to take the advantage of all surrogate
32 2. Literature study — crashworthiness of composite structures
models, one can also develop the ensemble of surrogates to extract as much data as possible
without additional FE simulations using the weighted-sum formulation [12]. Since each sur-
rogate model has both advantages and limitations, it can be predicted that in the future the
ensemble of all the existing surrogate models would be of great interest among researchers.
2.5. Summary
The experimental investigation on the crashworthiness of thermoset composite structures
under both quasi-static and dynamic loads has been comprehensively conducted on com-
2.5. Summary 33
ponent level tests on energy absorbers with tubular configurations. The failure mechanism
and energy absorbing characteristics are well studied. Basically different failure modes are
categorized, including global splaying, local fragmentation and delamination failure modes.
Besides, the effects of different influential parameters are investigated at the same time, in-
volving laminate properties, cross-sections, and types of triggers. These experiments provide
significant guidance for the crashworthy design of thermoset composite structures, develop-
ing of test standards and predicting models.
The numerical investigation of the crashworthiness of composite structures in LS-DYNA
has been developed for thermoset composite structures. Generally the types of modeling
approaches can be divided into efficient single-layered models and multiple layered models
with tiebreak contact or cohesive elements. These models can take into account the triggers,
laminate properties, and different failure criterion to approximate the real experiment setups.
Material models 54, 58, and 262 are the most popular material cards in LS-DYNA for cap-
turing the material properties of composite structures. These models usually require input
parameters calibration based on several coupon tests to accurately mimic the crush response.
Unfortunately, limitations still exist for modeling crashworthiness of composite structures,
including the overpredicted crush responses especially the initial peak load, non-physical
failure modes, and lack of transferability among different conditions. The following aspects
can therefore be improved in the future:
1. Develop better material models to achieve a better approximation with lower input pa-
rameter requirements.
2. Conduct more research on improving the crush responses in terms of physical failure
modes, and interpret the mechanism in the aspect of FEM on a comprehensive level.
3. Improve computational efficiency of the FE models.
The crashworthiness optimization of composite structures has becoming increasing im-
portant in this field. Research is mostly focused on the development of surrogate models
and gradient-free optimization algorithms towards small size of dataset. Four different types
of surrogate models have been investigated and compared in the literature, including PRS,
RBF, KRG and ANN models. Observations are found that the best way to use these models
is to simultaneously utilizing all possible surrogate models and selecting the one with best
performance. Due to the existence of both continuous and discrete design variables, it is
recommended to use direct search algorithm, such as genetic algorithm or particle swarm
algorithm. However some limitations exist for the current development of the crashworthi-
ness optimization of composite structures. One is the lack of transferability of surrogate
models in different cases, the other is the lack of the implementation in optimizing laminate
parameters and geometrical parameters simultaneously. Furthermore, the global optimiza-
tion or topology optimization of the fuselage structures still requires tremendous efforts since
most of the current research is concentrated on local optimization of energy absorbers. The
lack of computational efficiency when dealing with large number of variables is also another
concern in this field. Therefore it is recommended to conduct research in the following as-
pects for further study,
1. Investigate the transferability of surrogate models in different cases.
2. Develop related methods and algorithms to improve the performance and reduce the
computational efforts for optimization.
3. Perform local optimization of energy absorbers with both laminate properties and geo-
metrical parameters as design variables.
4. Improvement of current heuristic algorithms.
35
36 3. Material coupon test simulations
Layup and Width Length Thickness BCs and Loads Number Element
test [mm] [mm] [mm] of ele- size [mm]
ments
of the boundary conditions and displacements from nodal outputs. This is to be consistent
with the measuring method of load-displacement curves in the real coupon test experiments.
The corresponding global stress and strain values are then determined by using the load and
displacement values from LS-DYNA as follows.
𝐹 𝑑
𝜎 = ,𝜖 = (3.1)
𝑤ℎ 𝑙
where 𝜎 is the global axial stress of the whole laminate, 𝑤 is the width of the coupon, and
ℎ is the total thickness of the coupon; 𝜖 is the axial mid-plane strain of the coupon and
𝑙 is the undeformed effective length of the coupon. At the meantime, the local stress and
strain values were measured in one element at the center of the coupon, to eliminate the
influence of the boundary conditions. It should be noted that in LS-DYNA, the output stress
and strain values of elements are based upon integration points, which can be interpreted
as the stress and strain values for each ply. To validate whether LS-DYNA modeled the
quasi-isotropic layup composite in a correct way, one needs to transform the element-wise
local stress-strain relationships for each ply to the equivalent stress-strain relationships of
the laminate. This process can be implemented by using the Classic Laminate Theory (CLT)
[109], as shown as follows.
/
𝜎 =∫ 𝜎 𝑑𝑧 (3.2)
/
where the 𝜎 is the axial stress for each ply, the h is the thickness of the laminate.
By plotting the equivalent stress versus axial mid-plane strain relationships, one can com-
pare the peak stress, failure strain, and general trends with the global stress-strain relation-
ships to validate the LS-DYNA model. It should be noted that the slope in the stress-strain
curve should also be satisfied with the laminate stiffness calculated by classic laminate the-
ory without running LS-DYNA, as shown in the following equation [109].
1
𝐸 = (3.3)
ℎ𝑎
where the h is the thickness of the laminate and 𝑎 is the value in the laminate compliance
matrix or the [ABD] inverse matrix, which can be calculated by using the given material
properties for each ply and detailed layups.
38 3. Material coupon test simulations
Furthermore, the modulus and strength values of IM7/8552 determined by LS-DYNA can
also be validated, both globally and locally. The elastic modulus for each test is determined
by calculating the chord modulus in the linear elastic stage of the stress-strain curve. The
strength values are determined by using the maximum stress value up to failure in the lam-
inate. In the following subsections, these curves and the identified material properties will
be demonstrated and discussed.
Figure 3.3: Results for the quasi-isotropic tensile coupon test simulations
3.2. Coupon-level simulation results 43
Figure 3.4: Failure modes of three different material models in quasi-isotropic tensile test simulations
Figure 3.5: Different failure phases of quasi-isotropic tensile test simulations with MAT-54
44 3. Material coupon test simulations
1. Generally the stress-strain relationships in the linear elastic stage and the failure mode
coincide with typical compressive response of CFRP composite [108]. It should be noted
that the oscillations appeared in the post-failure regime for the material model without
damage considerations is caused by buckling instead of any composite failure criteria.
3. MAT-58 exhibits pre- and post-failure softening in the quasi-isotropic test, which can
be attributed to the BGM (Brooming, Gage, Middle) failure mode [108].
4. The maximum tensile strain at failure in Mat-58 is slightly larger than Mat-54 and
MAT-262.
6. Unlike the tensile test simulation, the quasi-isotropic compression test simulation doesn’t
exhibit a sudden drop of stiffness prior to ultimate failure. Instead, the stress goes lin-
early with the strain up to the maximum stress, followed by a sudden drop in stress.
However, the element is not completely failed since the compressive strength in the
post-failure regime is reduced to a constant level instead of zero. This process can be
illustrated by the failure mode demonstration in MAT-54, as shown in Fig 3.9. Similar
trends are found for MAT-58. Unfortunately, the test result using the material model
without damage considerations shows a buckling at the gage region, while the MAT-54
and MAT-58 models with calibrated parameters do not show any buckling. These can
be attributed to the initiation of damage which prohibits the coupon from reach the
critical load for buckling.
Figure 3.8: Results for the quasi-isotropic compressive coupon test simulations
Figure 3.9: Failure modes for MAT54 in quasi-isotropic compressive coupon test simulations
46 3. Material coupon test simulations
Table 3.6: Identified material properties and relative errors for MAT-nodmg
3.3. Summary
IM7/8552 material properties were identified by following the standard test methods in LS-
DYNA. 0,90 and quasi-isotropic laminate specimens were simulated in tension and compres-
sion to validate the material model 54, 58 and 262 in LS-DYNA. The coupon-level simulation
results demonstrated that both MAT-54, MAT-58, and MAT-262 generally showed reasonably
accurate stress-strain relationships in the linear-elastic regime, and failure stress. This con-
tributes to a successful validation on the accuracy of the material models, in terms of basic
material properties. Besides, the differences in post-failure characteristics for each model
have been discussed. The results for the quasi-isotropic compression test will be used in the
later sections for a more complicated model.
3.3. Summary 47
Table 3.7: Identified material properties and relative errors for MAT-54
Table 3.8: Identified material properties and relative errors for MAT-58
Table 3.9: Identified material properties and relative errors for MAT-262
49
50 4. Crashworthiness of composite tube
Figure 4.1: Cross-head displacement-time curve in the experiment 46B and 47B
effects. They were manufactured by the University of Utah and crushed for a distance of 100
mm. This stacking sequence was typically exhibiting a brittle fracture failure mode, which
could be described as intermediate length axial cracks, fronds developing and fractures, as
well as large debris wedges. The failure mechanisms include Mode I and II fracture, and a
combination of fiber and matrix fracture. This type of failure mode in the crushing test usu-
ally can be characterized by the phenomenon that relatively low energy absorbed through
delamination and friction, in contrast with fiber splaying, fragmentation and folding failure
mode.
A 500 kN servo-hydraulic Test Machine for Automotive Crashworthiness (TMAC). [113]
was used in the experiment at Oak Ridge National Laboratory. TMAC is an open loop servo-
hydraulic test machine capable of constant crushing velocities of up to 8 m/s. Tube test
samples were crushed at a constant speed of 5.5 m/s. Samples were not supported or aligned
during loading. Test samples were crushed against 1020 10 ga. cold rolled steel sheet with
a new piece for each test. Detailed cross-head displacement-time curve is shown in Fig4.1.
The curve indicated that the constant speed was only maintained for 15 ms, and nonlinear
behavior appeared in the curve after 15 ms. Therefore, the load-displacement data is only
taken before 15 ms, and the termination time of further LS-DYNA models will also be set as
15 ms. Total energy absorbed by each of the two composite tubes after the crushing initiation
were measured as 3020J (the mean of 46B and 47B, standard deviation -9J).
Figure 4.2: Three categorized stable failure modes. (a) fiber splaying, (b) fragmentation, and (c) brittle fracture. [112]
Figure 4.3: Three categorized unstable failure modes.(a) buckling, (b) interpenetration, and (c) barreling. [112]
52 4. Crashworthiness of composite tube
occur. At this time, the tube will fail into two halves and these halves cleave through
one another.
• Barreling: Delamination can occur unstably, where some layers may bow outward from
the inner and outer tube diameter leaving inner layers of the laminate unsupported to
fail at lower loads.
The stable failure modes in the crushing morphology of composite specimens can be cat-
egorized as fiber splaying, fragmentation and brittle fracture:
• Fiber splaying: The key characteristic of fiber splaying mode is the existence of long
interlaminar, intralaminar, and axial cracks that split the fibers into multiple bundles
or fronds. It is also possible that a debris wedge may form when the pile of crushed
resin and fibers accumulate in the gap between external and internal bending fronds.
• Brittle fracture: Combining the characteristics of fiber splaying and fragmentation mode,
brittle fracture has characteristics common to both. Failure modes for brittle fracture
include all of those associated with both fiber splaying and fragmentation modes of fail-
ure. Fiber and matrix fracture, friction, frond bending, and crack growth all contribute
to the total energy absorbed by this failure mode [112].
2. The loading plate is fixed in all DOFs except for the axial direction
4.4. Numerical modeling in LS-DYNA 53
Figure 4.4: LS-DYNA model of the tube, bevel trigger and loading plate
3. The loading plate is assigned a prescribed motion along axial direction with a constant
velocity of 5.5 m/s.
The contact settings in this model, are represented by LS-DYNA’s contact card: *CON-
TACT_AUTOMATIC_SINGLE_SURFACE_ID, as recommended by the user manual in contact
formulation involving crashworthiness analysis [48]. The Coulomb static friction coefficient
is set as 0.2, and a viscous damping factor (VDC=40) is added to improve the stability of the
crushing.
Material models MAT-54, MAT-58, and MAT-262 are used in this model, the parameters of
MAT-54 and MAT-58 can be found in Tables 3.1, 3.3 and 3.4. Besides the material properties,
other parameters in MAT-262 are summarized in Table 3.5.
Settings of several control cards are also necessary to be addressed in this model.
1. *CONTROL_ACCURACY, INN=4: The invariant node numbering for shell and solid el-
ements are toggled on. This is to minimize the sensitivity of updating of the orientation
of the material coordinate system, as the material coordinate system is automatically
updated following the rotation of element coordinate system[48].
2. *CONTROL_BULK_VISCOSITY, Q1=0, Q2=0, TYPE=-2: The shell bulk viscosity may
aid stability in compressive mode of response.
3. *CONTROL_SHELL, LAMSHT=1: The laminate shell theory corrects for the incorrect
assumption of uniform constant shear strain through the thickness of the shell.
4. *CONTROL_HOURGLASS, IHQ=4, QH=0: Hourglass modes are non-physical, and zero-
energy mode of deformation for non fully-integrated elements, therefore producing no
stress and strain. Type 4 offers stiffness based hourglass deformation modes, aiming
to minimize nonphysical stiffening of the response and at the same time effectively in-
hibiting hourglass modes.
5. The termination time is set as 16 ms, since the constant speed of crushing in the ex-
periment was only maintained 15 ms after initiation.
54 4. Crashworthiness of composite tube
Figure 4.6: The comparison of failure modes in three LS-DYNA material models: (a) MAT-54, (b) MAT-58, and (c) MAT-262.
56 4. Crashworthiness of composite tube
Figure 4.7: Comparison of different material models(MAT54, MAT58 and MAT262) and experimental results
From the curve, we can notice that there exists a small initial peak, which can be corre-
lated to the failure of bevel trigger. After the failure of the trigger, the load increases until
reaching the maximum peak since no failure modes of the tube have yet been triggered. When
the tube elements starts failing, the curve makes a transition to stable crushing regime with
crushing load oscillating at an almost constant level. The ups and downs in the stable crush-
ing regime are caused by fact that the elements of the tube are removed row by row. The
reaction forces exhibit high amplitudes, with some non-physical lower periods where many
elements are detached from the laminate at the same time. The interval between two adja-
cent peaks of the oscillations is exactly equal to the mesh size. Several observations can be
found as follows:
1. The model can accurately capture the initial small peak load that is correlated to the
failure of the trigger in the experiment.
2. The predicted maximum peak load is 22% higher than the experimental value, which is
a common feature for single shell modeling approach [114, 115]. The maximum peak
load has a negligible influence on the final energy absorption ability, thus the energy
absorption ability is barely influenced.
3. The predicted total energy absorption and energy absorption in the stable crushing
regime is 19% and 21% lower than the experimental values. This can be partially at-
tributed to the usage of longitudinal compressive stress limit factor SLIMC1 [58], which
will be studied later. The other contributing factor could be the lack of ability to capture
the delamination failure that might dissipate energy as well.
A sensitivity analysis of the SLIMC1 parameter in MAT-54 is then conducted, and the
load-displacement curve can be found in Fig 4.8, where the SOFT=0.8 is fixed and SLIMC1
is changed from 1.0 to 0.25. The quantitative values are listed in Table 4.2. It can be observed
that the peak forces and the energy absorption are reduced with the reduction of SLIMC1
parameter. This result indicates that post-failure behavior of the composite has significant
effect on the crashworthiness performance. SLIMC1=1 represents for the case where the
compressive stress limit in the post-failure regime is same as compressive strength. This is,
the material can still deform for a large strain value with full compressive strength after failure
occurs. It could be beneficial for the energy absorption since the crashfront elements would
still contribute to the reaction forces of the structure, but will also increase the maximum
peak load. Therefore it is necessary to do the trade-offs among different values. To this end,
the SLIMC1=0.375 demonstrates the best correlation to the experiments, in terms of both
the maximum peak load and energy absorption.
Similarly, another sensitivity analysis of the SOFT parameter in MAT-54 is conducted as
well, resulting in a load-displacement curve as shown in Fig4.9, where the SLIMC1=0.375
is fixed and SOFT is changed from 0.5 to 1.0. The quantitative values are listed in Table
4.3. Obviously, the peak force only has negligible changes with different values of SOFT, but
the average crushing force is increased as the increase of SOFT parameter. The SOFT=0.8
demonstrates the most reasonable agreement with the experimental results. Similarly, after
doing the trade-offs, one can find that SOFT=0.8 will provide the best correlation to the
experiments, in terms of both the maximum peak load and energy absorption.
58 4. Crashworthiness of composite tube
From the curve, similar results are observed for MAT-58, except for the significant over-
predicion of the maximum peak load. The results gives a 19% underprediction in energy
absorption, but 100% overpredcition in maximum peak load. This phenomenon can be in-
terpreted alongside the coupon test results, in that MAT058 predicts significantly higher
load-carrying capacity of the compression-loaded quasi-isotropic laminate then others. This
can be attributed to the inability of MAT058 to account for reduction of XC in the case of
transverse compressive failure [58]. More specifically, it means that MAT-58 is lack of re-
lationship between the matrix failure and longitudinal compressive strength. Therefore it is
not suggested to implement MAT-58 for the crashworthiness analysis of this type of CFRP
composite, due to its lack of accuracy.
2𝐸 𝐺
𝑙 ≤ (4.1)
𝑥
where 𝑀 = 1+, 1−, 2+, 2−, 6; 1,2 and 3 denote longitudinal, transverse, and shear direction; +
and − denote tension and compression. So in this case 𝑙 = 0.68 𝑚𝑚, which is too small
to be practical for industrial simulations. Then the fracture toughness values use in the
simulation will be several times higher than the physically-based value given in the mate-
rial properties. In this case, these parameters in MAT-262 will be considered as additional
parameters and thus requires fine-tuning by trail-and-error.
Additionally, several assumptions have to be made to determine the relations between
some parameters for simplicity. The fracture toughness for longitudinal (fiber) compressive
failure mode and tensile failure mode for bi-linear damage evolution (GXCO and GXTO) are
expressed as functions of strength at the inflection point, characteristic element length (L)
and the fractures strain(𝜖 ):
60 4. Crashworthiness of composite tube
𝐿
𝑋𝐶𝑂𝜖
𝐺𝑋𝐶𝑂 = (4.2)
2
where the value of 𝜖 is assumed to be 0.6, which is larger than the effective failure
strain=0.55; the characteristic element length is approximated by the area associated with
the integration points for each elements, and the angle of the mesh line with crack direction
[116], as follows,
√𝐴
𝑙∗ = (4.3)
𝑐𝑜𝑠(𝛾)
For an unknown crack direction, an empirical expression of the average can be implemented
as [117],
𝜋 𝜋
𝑙 ∗̄ = ∫ 𝑙 ∗ 𝑑𝛾 = 1.12√𝐴 (4.4)
4 0
Besides, the ratios and can be considered as similar expressions in the stress limit
factors of MAT54 and MAT58 (SLIMC1 and SLIMT1). Therefore, the remaining input pa-
rameters to be calibrated include longitudinal tensile and compressive stress limit factors
(SLIMC1, SLIMT1), the fractures angle in pure transverse compression (FIO), and the crash-
front softening factor (SOFT).
Since the material model 262 assumes that the both tensile and compressive damage
variables have contribution to the damage of stiffness for each ply, it is found necessary
to take into the longitudinal tensile stress limit factor SLIMT1. With the optimal set of pa-
rameters given in reference [58]: SLIMC1=0.86, SLIMC1=0.83, SOFT=0.4, and FIO=55, the
force-displacement curve for MAT-262 can be found in Fig 4.10, yielding results as follows:
From the load-displacement curve, we can observe that the predicted energy absorption
is 17% lower than the experimental result. Besides, MAT-262 overpredicted the maximum
peak load but had a generally good correlation to the experimental result in the stable crush-
ing regime. It can be interpreted by the phenomenon that some local instability or buckling
4.5. Numerical results 61
occurs close to the crashfront when the maximum load is reached. To have a better under-
standing of the material model, sensitivity studies of these input parameters are conducted
in this thesis. The sensitivity analysis of SLIMC1, SLIMT1, FIO and SOFT are conducted in
this study, giving the results as shown in Figs 4.11, 4.12, 4.13, and 4.14. The quantitative
results are also summarized in Table 4.4, 4.5, 4.6, and 4.7 .Several observations can be made
as follows
1. The sensitivity of SLIMC1 and SLIMT1 are complicated such that there exists a maxi-
mum at some non-trivial combination of these parameters.
2. The total energy absorption is found to be increased with the fracture angle (FIO).
3. The softening factor (SOFT) can affect the maximum peak load, but have negligible effect
on the energy absorption abilities.
62 4. Crashworthiness of composite tube
FIO 45 50 55 Experiment
Peak force [kN] 138 166 142 71
Average crushing force [kN] 27 27 32 37
Total energy absorption [J] 2107 2173 2520 3020
Energy absorption in the stable crushing 1788 1811 2107 2678
regime [J]
this interpretation seems reasonable, in that, a finer mesh showed a higher sensitivity level
and the failure could not be controlled within the contact region between the tube and the
loading plate. Furthermore, the element length for the finer mesh is shorter, which is easier
to be removed and has smaller critical buckling load. It makes the element erosion progress
faster than the model with a coarser mesh, and has a higher tendency to exhibit numeri-
cal instabilities. Therefore, the fast-propagated failure or the instabilities could cut off the
boundary conditions, or removing a random row of elements at the tube center. However,
the quantitative results prior to the interpenetration failure (crash distance 0 to 30 mm in
Fig 4.16 can still be compared to have an insight of the mesh dependency. These results are
summarized in Table 4.8. It can be observed that the total energy absorption remains at the
same level, while the maximum peak load is much smaller for the finer mesh results. It can be
concluded that finer mesh doesn’t not necessarily contribute to the refinement of the crash-
worthiness performance. Furthermore, it requires high costs of parameters re-calibration
and more computational resources.
Figure 4.15: The comparison of failure modes in MAT-54 with different mesh layouts: (a) 3.5 , (b) 1.75
4.5. Numerical results 65
Figure 4.16: Comparison of the load-displacement curves of two different mesh layouts using the same material data in MAT-54
Table 4.8: Summary of mesh sensitivity results. * The errors of coarser mesh results are calculated based on the finer mesh
results, instead of the experimental results.
4.6. Summary
This section implemented three different material models for crashworthiness, including
MAT-54, MAT-58, and MAT-262. The detailed damage mechanism and their performance
on a square tube under dynamic crushing load have been comprehensively investigated.
The proposed modeling approach for MAT-54 and MAT-58 only contains 2 input parame-
ters (SLIMC1, SOFT) to tune, while 4 input parameters (SLIMC1, SLIMT1, SOFT, FIO) for
MAT-262. This significantly reduces the number of parameters to tune compared with con-
ventional approaches. All three material models require extensive calibration to achieve cor-
relation with experimental data, such that sensitivity studies of these parameters have also
been performed for each model, respectively. The axial crushing results with the calibrated
parameters indicate that MAT-54 demonstrates the best correlation to the experiment with
22% higher in maximum peak load and 18% lower in total energy absorption, while MAT-58
and MAT-262 shows similar results in terms of energy absorption but both overpredict the
maximum peak load. Also MAT-54 provides an intermediate CPU (Intel Xeon CPU E5-2640
v4) time to run the simulations, compared with MAT-58 and MAT-262. Therefore, MAT-54
will be selected as the baseline model for the optimization in the following chapter due to
its robustness and efficiency. However, this single shell layer modeling approach is limited
in capturing the physical failure modes due to the simple representation of the tube. Im-
plementing this method can lead to significant underestimation of energy absorption. Also,
it can be concluded a simple representation of the tube is not sufficient to capture all the
failure modes. This could be the primary limitation of the single-layer modeling approach.
It is suggested by the author that stacked shell modeling approach should be used for more
accurate predictions.
5
Optimization of composite tube
5.1. Introduction
In this chapter, the optimization of composite tubes will be performed by using surrogate
models and the particle swarm optimization algorithm. Detailed formulations of the opti-
mization problem will be introduced first, followed by the optimization procedures, including
sampling using Design of Experiments (DOE), sensitivity studies, building surrogate models
and performing optimization to search the optimal solution.
5.2. Methodology
In the following sections, different types of new surrogate models (deep neutral networks
and gradient boosting regression trees) and optimization algorithms (mixed-discrete particle
swarm optimization) in crashworthiness optimization will be introduced and discussed. The
following methodology will be performed in this study for optimization:
1. Generate sampling points using modified Optimal Latin hypercube design (OLHD).
2. Run the corresponding simulations on the cluster and post-process with SAE 1000HZ
filter to obtain the total energy absorption and maximum peak load for each sample.
3. Check the robustness of the simulation for each sample, and eliminate samples with
irregular results caused by numerical errors or unstable failure modes in LS-DYNA.
4. Construct surrogate model for the mapping between design variables and output using
deep neural networks and gradient boosting regression trees. 80% of the sampling
points will be split into training set, and the rest into test set.
5. Evaluate the performance of the surrogate models with multiple metrics until satisfac-
tory, such as whether the root mean squared error is converged.
6. Search the optimal solutions by mixed-discrete PSO (Particle Swarm Optimization) op-
timizer. This modified version of PSO can be applicable for the both continuous and
discrete design variables, as well as their combinations.
67
68 5. Optimization of composite tube
laminate are selected as the design variables. For simplicity, the thickness and number of
plies are fixed as 2.16 mm and 16, respectively. The LS-DYNA material model in this section
will be the MAT-54 model with calibrated parameters, which has been proved to have the
best correlation to the experimental results. Furthermore, the laminate is assumed to be
symmetric to prevent in-plane and out-of-plane coupling. Therefore, only 8 design variables
need to be considered as input when performing the optimization task. The constraints in-
volve maximum peak crash force and two manufacturing constraints. A summary of the
design variables and constraints is demonstrated as follows,
• 8 discrete design variables: fiber orientation at each ply: [𝜃 /𝜃 /𝜃 /𝜃 /𝜃 /𝜃 /𝜃 /𝜃 ] The
fiber orientation can be selected from the integers in the following set: [0∘ /±15∘ /±30∘ /±
45∘ / ± 60∘ / ± 75∘ /90∘ ].
• 1 optimization objective: Maximizing EA (total Energy Absorption)
• Constraint 1: The maximum peak load during crushing is lower than 120 % of the
prescribed value related to the baseline LS-DYNA model.
• Constraint 2: At least 10% of the fibers in every layup should be lined up with each of
the four principal directions: 0, 45, -45, and 90.
• Constraint 3: The maximum number of unidirectional plies with same orientation next
to each should be no more than 2 to minimize edge splitting.
⎧ 𝑀𝑎𝑥 𝐸𝐴(𝑥)
⎪ 𝑠.𝑡. 𝐺 (𝑥) ≤ 0, 𝑖 = 1, 2, 3
(5.1)
⎨ 𝑥 ∈ ℤ, 𝑗 = 1, 2, 3...8
⎪
⎩ 𝑥 ∈ {0, ±15, ±30, ±45, ±60, ±75, 90}
where 𝐸𝐴(𝑥) denotes the total energy absorption; the 𝐺 (𝑥) are inequality design constraints,
including one constraint regarding the maximum peak crash force 𝑃 (𝑥), and two man-
ufacturing constraints regarding the number of maximum constitutive layers 𝑛 (𝑥), and
the percentage of 4 principle fiber orientations in the laminate, 𝑝𝑐𝑡 (𝑥) as follows,
𝑛 (𝑥) ≤ 2 (5.3)
𝑝𝑐𝑡 (𝑥) ≥ 10%, 𝜃 = 0, 45, −45, 90 (5.4)
It should be noted that 𝑃 = 104 is related to the maximum peak load=86 kN determined
in the axial crushing simulation result with MAT-54, as shown in 4.1. This value is relaxed by
20% such that the maximum peak load value used for checking constraint violations would
be 83.3 × 1.2 = 104 kN.
In summary, a flowchart is made for visualization the optimization process in this thesis,
as shown in Fig 5.1.
between the input and output. Before constructing the surrogate models, the influences of
the design variables on the crashworthiness response of the square tube under axial crash
loading are first studied based on numerical simulations. To this end, FEA are needed to
be performed in LS-DYNA for a comprehensive sensitivity study. Nevertheless, the design
feature space is consisted of 12 discrete integers ranging from -75 to 90. If each simulation
has an average runtime equal to 7 hours, the total CPU time estimated for evaluating all
possible locations in the design feature space would be 7× 12 , which is infeasible to run with
current computational resources. Therefore, the design feature space is too large to sample
over the whole feature space. Alternatively, the use of Design of Experiments (DOE) method
and surrogate models can solve this problem in that only a limited number of samples are
required. The selected sampling points are determined through Design of Experiments (DOE)
method, aimed to explain the effect of design variables on the crashworthiness performance
with acceptable number of simulations. The sampling points in this case are hypothesized
to reflect the variation of the design variables.
Due to the characteristics of design variables, a widely used DOE technique namely Latin
Hypercube sampling (LHS) [119] is selected to identify the locations of the sampling points in
the feature space. LHS is aimed to spread the sample points more evenly across all possible
values. It partitions each input distribution into 𝑁 intervals of equal probability, and selects
one sample from each interval. It also shuffles the sample for each input so that there is
no correlation between the inputs. In this study, a modified version of LHS namely Optimal
Latin Hypercude Design (OLHD) is implemented [120]. It develops an efficient global opti-
mal search algorithm, named as enhanced stochastic evolutionary (ESE) algorithm, found to
be much more efficient in terms of the computation time, the number of exchanges needed
for generating new designs, and the achieved optimality criteria [120]. Consequently, 12𝑋8
widely distributed sampling design points are generated over the design domain. More specif-
ically, these samples are integers with a lower and upper bound as -75 and 90, respectively.
The responses of total energy absorption and maximum peak load are evaluated at those
sampling points such that a total number of 96 LS-DYNA simulations are carried out.
Based on the DoE results, the LS-DYNA models representing the sampling design points
used for crashworthiness analysis with a crushing speed as 𝑉 = 5.5𝑚/𝑠 are created. The
baseline tube model implements MAT-54 material model with calibrated parameters. Only
the composite layup for each sample is varied according to the sampling results. Neverthe-
less, some of the simulation results demonstrate extremely low energy absorption or large
peak load, which can be attributed to unexpected failure modes such as barreling, inter-
penetration and folding. These failure modes are usually caused by the numerical errors
in the model. Therefore it is necessary to eliminate these samples that may contaminate
the dataset. On this purpose, only samples with energy absorption larger than 2000 J and
maximum peak load smaller than 120 kN are retained in the dataset. This leads to a total
number of 39 remaining sampling points with reasonable simulation results.
Figure 5.2: Distribution of the sampling design points between each two design variables and the Kernel density estimation of
each variable, 96 samples
be summarized:
2. Layer 5,8 merely have slight changes. It can be correlated to assumption that fiber
orientation in these layers will always exhibit acceptable crashworthiness performance.
Additionally, to evaluate the linear correlation between the design variables and the ob-
jective, one can use the Pearson correlation coefficient (PCC) matrix as a reasonable metric.
According to the Cauchy–Schwarz inequality it has a value between +1 and −1, where 1
is total positive linear correlation, 0 is no linear correlation, and −1 is total negative linear
correlation. The Pearson correlation coefficient 𝜌 is determined by a given pair of random
variables (𝑋, 𝑌) as follows,
𝑐𝑜𝑣(𝑋, 𝑌)
𝜌 , = (5.5)
𝜎 𝜎
where, 𝑐𝑜𝑣 is the covariance, 𝜎 and 𝜎 are the standard deviations of variable X and Y.
A PCC heatmap is then plotted to display the correlation between the variables, as shown
in Fig 5.4 where the cells are colored depending upon the contained value. It can be ob-
served that the PCCs between each design variable and output parameters are all lower than
0.5, indicating no linear correlation. Furthermore, the PCCs in the last two columns or rows
demonstrate the independent correlation between EA and PCF. These observations can be ex-
plained by the complicated and non-linear correlation between fiber orientations for each ply
and the crashworthiness performance. Therefore building a surrogate model that can reflect
the highly non-linear mapping between design variables and crashworthiness performance
is necessary in this study.
72 5. Optimization of composite tube
Figure 5.3: Distribution of the sampling design points between each two design variables and the Kernel density estimation of
each variable, 39 samples after selection
Figure 5.5: Architecture for Deep Neural Networks (DNN) with hidden units=[16,16]
⎪ 𝑣̂ = ( )
⎪ 𝑤( ) = 𝑤( ) − 𝜂 ̂
⎩ √ ̂
5.4. Surrogate models 75
( ) ( )
where 𝑚 is the biased first moment estimate; 𝑣 is the biased second raw moment
estimate; 𝑚̂ denotes the compute biased-corrected first moment estimate; 𝑣̂ denotes the
compute biased-corrected second raw moment estimate; 𝜖 is a small scalar used to prevent
division by zero, and 𝛽 and 𝛽 are the forgetting factors for the gradients and second moments
of gradients, respectively.
The framework for setting up the DNN is selected as Tensorflow [129], an open source
software library for high performance numerical computation.
Setting up the data and the loss
Before training the neural networks, the input data is normalized into the range [0,1] by
using the min-max scaler as follows . This is to speed up the convergence rate for gradient-
based optimization.
𝑋−𝑋
𝑋 = (5.8)
𝑋 −𝑋
Furthermore the parameters of the neurons (weights and biases) in the network have to be
initialized as well. When working with deep neural networks, initializing the network with the
right weights can be the difference between the network converging in a reasonable amount
of time and the network loss function not going wrong even after hundreds of thousands of
iterations. A properly scaled uniform distribution for initialization called ”Xavier” initializa-
tion [130] is adopted in this study. It initializes the weights in such a way that the variance
of input and output for each layer remains the same. This helps to keep the signal from
exploding to a high value or vanishing to zero.
There are also multiple regularization methods to prevent overfitting in advance, such
as L1 and L2 regularization, Max norm constraints, and dropout. In this study dropout
[131] is utilized in the neural networks due to its effective and simple implementation. While
training, dropout is implemented by only keeping a neuron active with some probability 𝑝 (a
hyperparameter), or setting it to zero otherwise. This process can be illustrated in Fig 5.6.
Dropout can be interpreted as sampling a Neural Network within the full Neural Network,
and only updating the parameters of the sampled network based on the input data. This
prevents units from co-adapting too much, since each batch is trained on different sub-
networks. Theoretically it will greatly improve the resistance of overfitting. It should be
noted that during testing there is no dropout applied, with the interpretation of evaluating
an averaged prediction across the exponentially-sized ensemble of all sub-networks.
The loss function for this study is selected as Mean-squared error (MSE), which is suitable
and most commonly used loss function for regression tasks. MSE is the summation over
squared distances between our target variable and predicted values, as shown below. The
76 5. Optimization of composite tube
reason to use the squared value in the objective is that the gradient becomes much simpler,
without changing the optimal parameters since squaring is a monotonic operation.
∑(𝑦 − 𝑦 )
𝑀𝑆𝐸 = (5.9)
𝑛
where 𝑦 is the label for sample i and 𝑦 is the predicted label for sample i.
Learning and evaluation
In this model, the DNN is trained for a total number of 500 epochs with 100 steps every
epoch. One epoch is when an entire dataset is passed forward and backward through the
neural network only once. A large number of epochs will be advantageous in handling limited
dataset since gradient descent is an iterative process. When training the neural networks,
there are multiple quantities we should monitor. In this study, the Root Mean squared errors
(MSE) are evaluated every 10 epochs to track the training process. A typical training curve
with good learning rate is illustrated in Fig 5.7.
(a) A cartoon depicting the effects of different learning (b) Typical training loss curve:An example of a typical
rates. With low learning rates the improvements will be lin- loss function over time, while training a small network on
ear. With high learning rates they will start to be more ex- CIFAR-10 dataset. This loss function looks reasonable,
ponential. Higher learning rates will decay the loss faster, and also indicates that the batch size might be a too low.
but they get stuck at worse values of loss (green line). This
is because there is too much ”energy” in the optimization
and the parameters are bouncing around chaotically, un-
able to settle in a nice spot in the optimization landscape.
Hyperparameter optimization is another important issue in this study, since training Neu-
ral Networks can involve many hyperparameter settings. Grid search combined with 5-fold
cross-validation [133] is used to determine the best hyperparameters. The strategy for tuning
hyperparameters in the DNN proposed here can follow the order below:
1. Number of neurons in each hidden layer
2. Number of hidden layers
3. Batch size
4. Dropout
Specifically, we first tune the number of neurons in each hidden layer by fixing number of
hidden layers=1, batch size=16, dropout=0. After searching the solution with lowest RMSE
on validation set, the number of neurons in each hidden layer will be fixed. It should be noted
that the training curve also needs to be examined such that the overfitting or underfitting
phenomena can be observed. Then the second hyperparameter will be tuned following the
same process, until all the hyperparameters are fine-tuned. The purpose of this tuning is to
obtain a network with the best performance.
5.4. Surrogate models 77
Results
A detailed setting of the DNN initialization is listed as below,
• Activation functions: Relu
• Loss function: MSE
• Optimizer: Adam optimizer initialized with learning rate=0.001, 𝛽 =0.9, 𝛽 =0.999,𝜖 =
1𝑒 − 08
• Weight initialization: ”Xavier” method
• Hidden units:[16,16] (It means 2 hidden layers and 16 neurons for each hidden layer)
• Batch size = 16
• Dropout = 0.0
Training curve of the loss is found in Fig 5.8, which indicates a typical training curve
with good learning rate. This leads to a convergence with RMSE=429.76 on validation set
and RMSE=327.23 on training set. The performance of the network is obviously not good
enough for the predicting task since the RMSE remains at a high level. Usually this kind
of result demonstrates that the network is not complex enough to reflect the non-linear cor-
relation between inputs and outputs. Therefore we should consider add more complexity
to the network to breakthrough current limit on reducing the loss. But when the network
becomes two complicated, the training curve exhibit irregular response where the training
loss will increase with increased number of steps. This is, during the process of searching
optimal hyperparameters of the networks, the RMSE should be minimized while the training
curve should be reasonable. To have better understanding of some key hyperparameters,
the effect of hidden units, batch size and dropout are investigated in Fig 5.9, Fig 5.10 and Fig
5.11, respectively. Several observations and corresponding interpretations are summarized
as follows:
• Using too few neurons in the hidden layers will result in underfitting. It occurs when
there are too few neurons in the hidden layers to adequately detect the signals in a
complicated data set. Using too many neurons in the hidden layers can result in sev-
eral problems. First, too many neurons in the hidden layers may result in overfitting.
Overfitting occurs when the neural network has so much information processing capac-
ity that the limited amount of information contained in the training set is not enough
to train all the neurons in the hidden layers. A second problem can occur even when
the training data is sufficient. An inordinately large number of neurons in the hidden
layers can increase the time it takes to train the network. The amount of training time
can increase to the point that it is impossible to adequately train the neural network.
Obviously, some compromise must be reached between too many and too few neurons
in the hidden layers.
• No dropout results in overfitting since it produces a fragile model too specialized to the
training data. An appropriate value of dropout can improve the performance of network
on validation set and prevent overfitting. This can be attributed to the fact that dropout
forces a neural network to learn more robust features that are useful in conjunction
with many different random subsets of the other neurons. However, if the dropout rate
is high essentially you are asking the network to suddenly unlearn stuff and relearn it
by using other examples, which makes the performance worse.
• Batch size =1 and batch size= size of the training set are two extremes: Full-batch learn-
ing and online learning. The former computes the gradient based on a single instance
of the dataset, while the latter over computes the gradient the entire dataset, averaging
over potentially a vast amount of information. The online learning will become more
difficult to converge since noisy updates make the training curve oscillating around the
converged value, but it may allow you to bounce out of bad local optima. The full batch
78 5. Optimization of composite tube
learning has a smooth training curve and easy to converge, but it has higher risks to be
trapped in a local minimum. Furthermore it has been observed in practice that, when
using a larger batch, there is a significant degradation in the quality of the model, as
measured by its ability to generalize. Therefore a compromise between them is neces-
sary, which leads to a batch size=16 in the plot.
To this end, the best hyperparameters after tuning is found to be hidden units=[32,32],
dropout=0.2, batch size=16. This yields a train RMSE=200.41, and a test RMSE=123.34.
This model will be used in the further optimization.
Figure 5.9: Training curves with different hidden units, dropout=0, batch size=16
Figure 5.10: Training curves with different dropout, hidden units=[32,32], batch size=16
80 5. Optimization of composite tube
Figure 5.11: Training curves with different batch size, hidden units=[32,32], dropout=0.2
function. Detailed expressions of gradient boosting trees will also be discussed in the follow-
ing subsection.
Regression Tree and Ensemble
In order to make a prediction for a given observation, we typically use the mean of the train-
ing data in the region to which it belongs. More specifically, the decision methods can be
described as predicting a response or class 𝑌 from inputs 𝑋 by growing a binary tree. At
each internal node in the tree, we apply a test to one of the inputs, say 𝑋 . Depending on the
outcome of the test, we go to either the left or the right sub-branch of the tree. Eventually
we come to a leaf node, where we make a prediction. This prediction aggregates or averages
all the training data points which reach that leaf. This is, we can divide the predictor space
(the set of 𝑋) into 𝐾 different non-overlapping regions: 𝑅 , 𝑅 ,...,𝑅 . Theoretically, the regions
could be any shape, but we should choose to divide the predictor space into high-dimensional
boxes for simplicity. Therefore, our goal is to find boxes 𝑅 , 𝑅 ,...,𝑅 that minimize the RSS
given by the following equation:
𝑅𝑆𝑆 = ∑ ∑ (𝑦 − 𝑦̂ ) (5.10)
∈
where 𝑦̂ is the mean response for the training observations with the kth box. Since it is
computationally infeasible to loop over every possible partition of the feature space into k
boxes, a greedy approach called recursive binary splitting is implemented. It begins at the
top of the tree and then successively splits the predictor space, leading to two new branches
further down on the tree. At each step of the tree building process, the best split is greedily
made at that particular split. One alternative method to do this is to use pruning, where we
grow a very large tree 𝑇 and then prune it back to obtain a subtree. The detailed algorithm
for building a regression tree can be found as follows:
1. Use recursive binary splitting to grow a large tree on the training data, stopping only
when each terminal node has fewer than some minimum number of observations.
5.4. Surrogate models 81
2. Apply cost complexity pruning to the large tree in order to obtain a sequence of best
subtrees, as a function of 𝛼.
3. Use K-fold cross-validation [133] to choose 𝛼.
4. Return the subtree from Step 2 that corresponds to the chosen value of 𝛼.
Usually, a single tree is not strong enough to be used in practice. What is generally applied
in the industry is the ensemble model, which sums the prediction of multiple trees together.
The prediction scores of each individual tree are summed up to get the final score. This
ensemble model is designed to make each individual trees complement to each other. The
ensemble model can be expressed mathematically as follows,
𝑦̂ = ∑ 𝑓 (𝑥 ), 𝑓 ∈ ϝ (5.11)
where 𝐾 is the number of trees, 𝑓 is a function in the functional space ϝ, and ϝ is the set of
all possible CARTs.
Different from one sinle CART, ensemble model has a different objective function to be
optimized, shown as follows,
where obj is the objective function in machine learning. The first term is the summation of
training loss over the set of all CARTs; the second term is the regularization term dependent
on the complexity of the ensemble model.
Gradient Boosting
As is always for all supervised learning models, the tree ensemble model can also be trained
by optimizing the objective function. For the objective function at step 𝑡, it can be written as,
( )
𝑜𝑏𝑗 = ∑ 𝑙(𝑦 , 𝑦̂ ) + ∑ Ω(𝑓 ) (5.13)
where 𝑦 denotes the prediction; 𝑙 denotes the loss function, Ω is the regularization term; 𝑓
is a function in the functional space.
From the shape of this objective function, obviously the functions 𝑓 are those we are
trying to learn, since each contains the structure of the tree and the leaf scores. Unlike
the traditional optimization problem where the gradient can be simply used, it is intractable
to learn all the trees at once. Alternatively, an additive training method called boosting is
introduced: iteratively learning weak regressors with respect to a distribution and adding
them to a final strong regressor. This is, fix what we have learned, and add one new tree at
( )
a time. When the prediction value at step 𝑡 is written as 𝑦̂ , we can derive its relation with
prediction value at step 𝑡 − 1 as,
( )
𝑦̂ = 0
( ) ( )
𝑦̂ = 𝑓 (𝑥 ) = 𝑦̂ + 𝑓 (𝑥 )
( ) ( )
𝑦̂ = 𝑓 (𝑥 ) + 𝑓 (𝑥 ) = 𝑦̂ + 𝑓 (𝑥 ) (5.14)
...
( ) ( )
𝑦̂ = ∑ 𝑓 (𝑥 ) = 𝑦̂ + 𝑓 (𝑥 )
( )
𝑜𝑏𝑗( ) = ∑ 𝑙(𝑦 , 𝑦̂ + 𝑓 (𝑥 )) + Ω(𝑓 ) + 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡 (5.15)
82 5. Optimization of composite tube
In the general case, we take the Taylor expansion of the loss function up to the second
order:
( ) 1
𝑜𝑏𝑗( ) = ∑[𝑙(𝑦 , 𝑦̂ ) + 𝑔 𝑓 (𝑥 ) + ℎ 𝑓 (𝑥 )] + Ω(𝑓 ) + 𝑐𝑜𝑛𝑠𝑡𝑎𝑛𝑡 (5.16)
2
After removing all the constants, the specific objective to be optimized at step 𝑡 can be ex-
pressed as
1
𝑜𝑏𝑗( ) = ∑[𝑔 𝑓 (𝑥 ) + ℎ 𝑓 (𝑥 )] + Ω(𝑓 ) (5.18)
2
This is our final optimization objective for the new tree. It should be noted that this function
depends merely on gradient terms 𝑔 and ℎ , and thus this model is called gradient boosted
trees. For the last complexity term in the objective function, we need to first refine the
definition of the tree 𝑓(𝑥) as
where 𝑤 is the vector of scores on leaves; 𝑞 is a function assigning each data point to the
corresponding leaf and 𝑇 is the number of leaves. The complexity is then defined by the
function suggested in reference [136], which works well in practice. This can be expressed
as:
1
Ω(𝑓) = 𝛾𝑇 + 𝜆 ∑ 𝑤 (5.20)
2
where 𝑤 is the vector of scores on leave; 𝑇 is the number of leaves; 𝛾 and 𝜆 are two hyperpa-
rameters representing the strength of 𝑙 and 𝑙 regularization, respectively.
Re-formulating the tree model by using the tree model definition 𝑓 (𝑥) above, the objective
value with the t-th tree can be written as:
𝑜𝑏𝑗( )
=∑ [𝑔 𝑤 ( ) + ℎ𝑤 ( )] + 𝛾𝑇 + 𝜆 ∑ 𝑤
=∑ [(∑ ∈ 𝑔 )𝑤 + (∑ ∈ ℎ + 𝜆)𝑤 ] + 𝛾𝑇 (5.21)
=∑ [𝐺 𝑤 + (𝐻 + 𝜆)𝑤 ] + 𝛾𝑇
where 𝐼 = 𝑖|𝑞(𝑥 ) = 𝑗 is the set of indices of data points assigned to the j-th leaf; 𝐺 = ∑ ∈ 𝑔
and 𝐻 = ∑ ∈ ℎ . Basically, for a given tree structure, we push the statistics 𝑔 and ℎ to
the leaves they belong to, sum the statistics together, and use the formula to calculate the
score. This is also a metric to measure the performance of a tree such, and also do pruning
on the branches. For the real valued data in this thesis, it is then sufficient to calculate the
structure score of all possible split solutions, and we can find the best split efficiently.
Results
This algorithm is implemented on XGBoost [136], an optimized distributed gradient boost-
ing library under the Gradient Boosting framework. The loss function is selected as Mean
squared error (MSE).Two independent regressors are built for predicting the energy absorp-
tion and maximum peak load, respectively. Then the loss of two regressors are averaged to
make a multi-output regressor for predicting two scores. The hyperparameters are tuned
by grid-search based on the mean value of 5-fold cross-validation RMSE. A list of detailed
settings are summarized as below:
• objective=”reg:linear” : linear regression, default
5.4. Surrogate models 83
Figure 5.12: The tree structure for 10th regression tree in the ensemble model that is consisted of 500 trees
Table 5.1: Performance of the gradient boosted tree ensemble model on the training set and test set
Dataset RMSE score [-] Explained variance score [-] R squared score [-]
Train 2 0.99 0.98
Test 34.4 0.81 0.64
𝑋 =𝑋 +𝑉 (5.24)
𝑉 ̃
= 𝛼𝑉 + 𝛽 𝑟 (𝑃 − 𝑋 ) + 𝛽 𝑟 (𝑃 − 𝑋 ) + 𝛾 𝑟 𝑉 (5.25)
where,
5.5. Mixed-discrete Particle Swarm Optimization algorithm 85
• 𝑋 and 𝑋 are the locations of the 𝑖 particle at the 𝑡 and the 𝑡 + 1 iterations,
respectively;
• 𝑉 and 𝑉 are the velocity vectors of the 𝑖 particle at the 𝑡 and the 𝑡 + 1 iterations,
respectively;
• 𝑃 is the best candidate solution found for the entire swarm (current best global solu-
tion);
• 𝛼, 𝛽 and 𝛽 are the prescribed coefficients that determine the inertial, the exploitative,
and the explorative attributes of the particle motion;
The best global and the best local solutions are updated at every iteration using the com-
parison between values of the objective functions, the constraint functions, and the candi-
date solutions. This update process is implemented continuously, irrespective of discrete or
continuous design variables. But the discrete component of the design vector is updated to
nearby feasible discrete locations.
A deterministic approach
A deterministic approximation strategy is used in conventional PSO, which ensures that the
system model can only be evaluated at the allowed discrete domain positions. The positions
of the particle in the discrete domain is assumed to be in a local hypercube as follows,
where, 𝑚 is the number of discrete design variables, 𝑥 is the current position of the candidate
solution in the discrete domain. 𝑥 and 𝑥 denote two consecutive feasible values of the 𝑖
variable (lower and upper boundary respectively).
The values of these two consecutive values can be determined from the discrete vectors
for each discrete design variable. The Nearest Vertex Approach (NVA) is then implemented
to approximate the current position of the candidate solution in the discrete domain to the
corresponding vertex in its local hypercube 𝐻 . This criteria is formulated based upon Eu-
clidean distance, as expressed in the approximated discrete-domain location based on the
NVA: 𝑋̃ = [𝑥 ̃ 𝑥 ̃ ...𝑥 ̃ ] where
𝑥 , |𝑥 − 𝑥 | ≤ |𝑥 − 𝑥 |
𝑥̃ = { , ∀𝑖 = 1, 2, 3, ..., 𝑚 (5.27)
𝑥 , 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
86 5. Optimization of composite tube
A stochastic approach
In this stochastic approach, the discrete variables are updated based on diversity preserva-
tion [139]. The population diversity with respect to the design space is characterized by the
overall spread and the distribution of the particles in the swarm. Among different diversity
metrics, we implemented a kind of estimated metric for the entire population at the begin-
ning of an iteration. It should be noted that this metric uses the enclosing hypercube of the
entire population to improve the efficiency of heuristic optimization algorithms. More specif-
ically, a hypercuboid region close to the best global particle is required to be constructed
in the design space, which is defined by a fraction of the volume of the smallest hypercube
enclosing all the particles. The fraction is prescribed by a factor between 0 and 1. Then the
number of particles within the fractional hypercuboid is estimated and used to adjust the
enclosing-hypercube side length. Lastly, the diversity metrics are updated to capture the
particle distribution. This initial diversity metric (𝐷 ) is a normalized discrete variable ranges
(vector) which span the current population as follows:
, ,
𝑥 −𝑥
𝐷 , = , ∀𝑖 = 1, 2, ..., 𝑚 (5.28)
𝑥 −𝑥
where 𝐷 , is the discrete diversity metric with respect to the 𝑖 discrete variable; 𝑥 ,
and 𝑥 , are the maximum and minimum values of the 𝑖 discrete variable in the population
at 𝑖 ℎ iteration, respectively; 𝑥 and 𝑥 are the upper and lower bounds of the 𝑖 ℎ design
variable, respectively.
To better account for the distribution of solutions, this metric is adjusted as:
𝑁+1
𝐷̅ , = (𝜆 ) ×𝐷 , (5.29)
𝑁 +1
where 𝐷̅ , is the adjusted discrete diversity metric; N is the number of particles in the smallest
hypercube enclosing all the particles; 𝑁 is the estimated number of particles in the fractional
hypercube; n is the number of discrete variables.
After determining the adjusted discrete diversity metric, the diversity coefficient for dis-
crete variables 𝛾 , can be calculated as follows:
−𝐷̅ ,
𝛾 , = 𝛾 𝑒𝑥𝑝( ) (5.30)
2𝜎 ,
1
𝜎 , = , ∀𝑖 = 1, 2, ..., 𝑚 (5.31)
√2𝑙𝑛𝑀
where 𝑀 is the size of the set of feasible values of the 𝑖 discrete variables; 𝛾 is a constant
between 0 and 1 for initialization.
With the diversity coefficient 𝛾 , , the position of 𝑖 ℎ particle can be updated in a stochastic
approach by using a random number 𝑟 between 0 and 1, as follows:
1. If 𝑟 is greater than the diversity coefficient 𝛾 , , update the discrete variables using the
discrete variable using the deterministic approach in equation 5.27.
2. If 𝑟 is smaller than or equal to the diversity coefficient 𝛾 , , 𝑥 is randomly approximated
to either 𝑥 or 𝑥 defined in equation 5.27.
2. Both solutions are infeasible and solution 𝑖 has a smaller net constraint violation than
solution 𝑗 or,
3. Both solutions are feasible; in addition, solution 𝑖 is not worse than solution 𝑗 in any
objective, and solution 𝑖 is better than solution 𝑗 in at least one objective
where 𝑔 ̅ and 𝑔 ̅ denote the normalized values of the 𝑗 inequality constraint and 𝑘 equality
constraint, respectively. 𝜖 denotes the tolerance value to relax each equality constraint.
This approach is favorable in driving solutions towards and into their feasible domains with
increasing iterations.
• 𝛽 =2: Scaling factor to search away from the particle’s best known position
• 𝛽 =2: Scaling factor to search away from the swarm’s best known position
• 𝛾 =0.7: Initial value for the discrete diversity coefficient, a constant between 0 and 1.
• Fractional domain size (𝜆 × 𝑁 )=0.1 × 𝑁: The fractional domain volume is 0.1 of the
smallest hypercube enclosing all the particles.
• minfunc=1e-8: The minimum change of swarm’s best objective value before the search
terminates
Figure 5.13: LS-DYNA simulation results with the optimal layups obtained from XGboost and DNN models, respectively.
the formation of fronds. This failure mechanism demonstrates that the total energy are dis-
sipated by different sources, including fiber and matrix failure, friction and frond bending.
Other types of failure, such as delamination, can not be observed in these simulation results.
The quantitative results of the optimization based on DNN and XGboost models has been
demonstrated in Table 5.2. The 𝐸𝐴_𝑝𝑟𝑒𝑑𝑖𝑐𝑡𝑒𝑑 represents the energy absorption of the op-
timal layup predicted by surrogate models, and 𝐸𝐴 is the energy absorption of the
optimal layup obtained from LS-DYNA simulations. 𝑃 and 𝑃 represent the maximum
crushing force and average crushing force, respectively. Also it should be noted that the
layup for the baseline model is [0 , ±45 , 90 ] . From the results, we can observe that the er-
rors between the predicted energy absorption values and true values determined in LS-DYNA
are relatively small: -1.21% for DNN and 1.9% for XGboost. This improvement has proven
the significance and efficiency of using surrogate models in predicting the crashworthiness
performances with a limited size of dataset. However, the optimum solutions differ in these
two models. The optimal layup predicted by DNN has a 6.06% increase on the total energy
absorption, and the XGboost result shows a 11.52% increase. The average crushing forces
predicted by DNN and XGboost demonstrate a 7.32% and 11.83% increase, respectively. In
contrast, the maximum crushing forces predicted by DNN and XGboost are higher than the
baseline value by 7.97% and 16.05% , respectively. In summary, the total energy absorption
has proven to be improved by using surrogate models and the MDPSO algorithm, while the
maximum crushing force does not violate the constraints prescribed in the previous section.
This indicates a successful optimization.
Table 5.2: Optimization results of DNN and XGboost regression models in terms of energy absorption (EA)
(a) DNN-[0/-45/45/-45/0/45/-75/90]s
(b) XGboost-[0/-45/45/30/0/45/-45/90]s
(c) Baseline-[0/0/45/-45/45/-45/90/90]s
Figure 5.14: The comparison of failure modes among LS-DYNA simulation results with different optimal layups (a) and (b), as
well as baseline (c)
90 5. Optimization of composite tube
Figure 5.15: The comparison between quasi-static and dynamic axial crushing response with MAT-54 and baseline layup
([ / ± / ] )
are demonstrated in Fig 5.15. The quantitative results are also summarized in Table 5.3. It
should be addressed that both of the simulation results are post-processed with a SAE 1000
HZ filter to be in consistence with the previous simulations.
Table 5.3: The comparison of the stiffness calculated from the experiments, quasi-static simulation and dynamic simulation.
We can observe from the Fig 5.15 and Table 5.3 that the stiffness calculated in phase (a) for
the quasi-static and dynamic results are inconsistent. The quasi-static simulation exhibits a
much stiffer behavior than the experiments, while the dynamic simulation exhibits a slightly
softer behavior. For the stiffness calculated in phase (b), quasi-static simulation demon-
strates a stiffer behavior than the experiments, while dynamic simulation demonstrates an
excellent agreement. It should be noted that phase (a) represents the deformation of the
trigger with a reduced thickness, while phase (b) represents the deformation of the square
tube. These results and comparisons can be interpreted in two aspects: (i) the bevel trigger
leads to a lower stiffness in the experiments; (ii) The usage of SOFT parameter in dynamic
simulation can lower the stiffness of the crashfront elements to obtain a more reasonable
result. Therefore, the second-linear elastic phase (b) will be selected to calculate the stiffness
of the composite tube in the further investigation. To prevent running all the quasi-static
simulations to obtain the stiffness values, stiffness values calculated from existing dynamic
simulation are used instead to approximate the real stiffness values of the tube. The stiffness
value in Fig 5.15 is then determined by using the second linear-elastic phase in the dynamic
simulation result, giving a k=64 kN/mm, and this value is used as the lower boundary of the
stiffness constraint 𝑘 = 64. This additional constraint can be incorporated in equation 5.1
and be expressed as:
𝑘(𝑥) >= 𝑘 = 64, 𝑥 ∈ {0, ±15, ±30, ±45, ±60, ±75, 90} (5.33)
To incorporate the stiffness constraint in the optimization, we need to re-build the surro-
92 5. Optimization of composite tube
gate model such that the outputs of the model include the stiffness of the tube. Therefore, the
dimensions of the outputs has increased from 2 to 3, which requires training the surrogate
model again. To this end, the XGboost model is trained for predicting the energy absorption
(objective), maximum peak load (constraint) and stiffness (constraint). After fine-tuning the
hyperparameters in XGboost, it has been found that the same set of hyperparameters as
utilized in section 5.4 gives the lowest evaluation RMSE score (mean value in a 5-fold cross
validation). The obtained test RMSE is equal to 47.29, with an explained variance score =
0.804 and R squared score = 0.59. These metrics indicate a high level of explainability on the
test set. Therefore, this model is then used for optimization with MDPSO. The experiment is
repeated 10 times, giving an optimal layup=[0/ − 45/45/30/0/45/ − 45/90]𝑠 with a predicted
energy absorption=2794 J. After evaluating the optimal layup in the LS-DYNA simulation,
it gives a result with an energy absorption=2740 J, maximum peak load=100 kN, stiffness=
73 kN/mm. These values are identical to the optimization results without the stiffness con-
straint. This can be interpreted by possibility that the lower boundary of the stiffness used in
the optimization is close to the lower boundary of the predicted stiffness from the surrogate
model.
⎧ 𝑀𝑎𝑥 𝐸𝐴(𝑥)
⎪ 𝑠.𝑡. 𝐺 (𝑥) ≤ 0, 𝑖 = 1, 2, 3, 4
(5.34)
⎨ 𝑥 ∈ ℤ, 𝑗 = 1, 2, 3...8
⎪
⎩ 𝑥 ∈ {0, ±7.5, ±15, ±22.5, ±30, ±37.5, ±45, ±52.5, ±60, ±67.5, ±75, ±82.5, 90}
where 𝐸𝐴(𝑥) denotes the total energy absorption; The 𝐺 (𝑥) are inequality design constraints,
including one related to the maximum peak crash force 𝑃 (𝑥), one related to the stiffness of
the tube, and two manufacturing constraints regarding the number of maximum constitutive
layers 𝑛 (𝑥), and the percentage of 4 principle fiber orientations in the laminate, 𝑝𝑐𝑡 (𝑥)
as follows,
𝑃 (𝑥) ≤ 𝑃 = 104 (5.35)
𝑘(𝑥) >= 𝑘 = 64 (5.36)
𝑛 (𝑥) ≤ 2 (5.37)
𝑝𝑐𝑡 (𝑥) ≥ 10%, 𝜃 = 0, 45, −45, 90 (5.38)
Since the discrete design domain is extended to a hypercube with approximately twice
larger number of vertices, MDPSO might have higher chances to produce different optimiza-
tion results for each run, due to the nature of heuristic optimization algorithms. It should be
noted that the 10% rule constraint is easier to be violated in this case, since the design vari-
ables satisfied with the constraint would become more difficult to reach during the updating
process of the discrete variables. It is possible that no feasible design can be found during
the optimization, and MDPSO could handle this scenario by comparing the net constraints
violations for the infeasible designs, which can not be accomplished by conventional PSO.
5.9. Summary 93
The experiment is then repeated 20 times, and the best solution(s) out of 20 will be selected
as our result. In contrast to the first-stage optimization, three optimal layups with same
energy absorption values were determined. These optimal layups are then evaluated in LS-
DYNA to obtain the crashworthiness performances, respectively. The LS-DYNA results are
compared with the first-stage optimization results in terms of the total energy absorption,
maximum peak load, average crushing force and stiffness, as shown in Table 5.4.
Table 5.4: Second-stage optimization results and their comparison with the first-stage optimization results
Comparing with the first-stage optimization results, we can observe that optimal layup
no.2 demonstrates a 0.95% increase in energy absorption, 1.04% increase in average crush-
ing force, while 3.44% decrease in maximum peak load and 2.59% decrease in stiffness. This
indicates that optimal layup no.2 slightly outperforms the layup obtained the first-stage op-
timization, in terms of energy absorption, average crushing force and maximum peak load.
Eventhough the stiffness value for no.2 is slightly lower than that for first-stage optimization,
it is still higher than the prescribed lower boundary of the stiffness. Therefore, optimal layup
no.2 is selected as the final solution of the second-stage optimization.
To have a comprehensive understanding of the obtained second-stage optimal layup, the
load-displacement curves from the LS-DYNA simulations are demonstrated in Fig 5.16. Fig
5.17 presents the tube deformation modes obtained by LS-DYNA for three different layups at
crash distances of 12.25 mm, 34.25 mm and 72.75 mm, respectively. It can be observed that
similar steady crash behaviors were achieved for first-stage and second-stage optimization
results. A brittle fracture characterized by a combination of fiber splaying and fragmentation
is observed for each result. The quantitative comparison is also summarized in Table 5.5.
It can be observed that both of the optimization results have significant improvements in
the energy absorption, average crushing load, and the stiffness, except for the maximum
peak load. Furthermore, the second-stage optimization results outperform the first-stage
optimization results in energy absorption, maximum peak load and average crushing load.
This indicates a good transferability of the optimization framework consisted of XGboost and
MDPSO.
5.9. Summary
Chapter 5 performed a two-stage single-objective optimization of a composite tube, to pursue
a optimal layup of the tube with highest energy absorption. 8 discrete design variables were
implemented to represent the fiber orientation of each ply. 4 constraints were incorporated in
the optimization, including one constraint for the maximum peak load, one for the stiffness
and the rest two for the manufacturing related rules. An optimal latin hypercube design
was used for designing experiments in LS-DYNA, giving a total number of 39 valid sam-
ples for constructing the dataset. Two surrogate models: Deep neural networks (DNN) and
94 5. Optimization of composite tube
Figure 5.16: Load-displacement curves for first-stage optimization result, second-stage optimization result, and their
comparisons with the initial design
Table 5.5: First-stage and second-stage optimization results and their comparisons with the initial design
(c) Baseline-[0/0/45/-45/45/-45/90/90]s
Figure 5.17: The comparison of failure modes among LS-DYNA simulation results with first-stage (a), second-stage optimal
layups (b), as well as baseline (c).
96 5. Optimization of composite tube
Gradient boosting regression trees ensemble in XGboost were built to capture the mapping
between the design variables and output crashworthiness metrics, respectively. Both of the
two models demonstrated good performances on the training and the test set. The surrogate
models were then used for optimizing the layup by using a mix-discrete particle swarm algo-
rithm (MDPSO), respectively. Both results indicated effective solutions in the improvement
of energy absorption, while XGboost outperformed DNN model in the accuracy and the per-
centage of improvement. After evaluating the optimal layups in LS-DYNA, the final solution:
[0/-45/45/30/0/45/-45/90]s with an energy absorption=2740 J (11.52% increase than ini-
tial design) was obtained. The axial crushing simulations of this optimal layup exhibited a
stable crushing response with a brittle fracture failure mode, which can be characterized
by a combination of fiber splaying and fragmentation. Lastly, the transferability of the pro-
posed optimization framework across design domains of different composition and size was
investigated. The candidate values for each design variable in the optimization process was
extended from the set {0, ±15, ..., ±75, 90} to {0, ±7.5, ±15, ..., ±75, ±82.5, 90}. This second-stage
optimization obtained a different optimal layup: [0/-52.5/37.5/37.5/0/45/-45/82.5]s with
an energy absorption=2767 J (12.54% increase than initial design). Therefore, the proposed
optimization framework in this study has been proven to have excellent applicability and
transferability in the crashworhy design of composite structures.
6
Conclusions
In this study, a comprehensive investigation on LS-DYNA modeling and design optimization
of composite square tubes for structural crashworthiness was performed. This thesis imple-
mented an efficient modeling approach in LS-DYNA with single-layer shells and stress limit
factors to describe the element behaviors. Three mainstream LS-DYNA composite material
models-MAT-54, MAT-58 and MAT-262 with calibrated parameters were investigated and
verified by modeling the CFRP square tubes under dynamic crushing load. The three mate-
rial models were validated by coupon-level simulations, followed by sensitivity analysis of the
most influential parameters to understand the complicated failure mechanisms. Simulation
results demonstrate good correlation to the experiments in terms of energy absorption and
maximum peak load, while the computational and calibration related costs remain in a low
level. The following conclusions can be made regarding different material models in modeling
axial crushing of CFRP tubes:
• The proposed modeling approach for MAT-54 and MAT-58 only contains 2 input pa-
rameters (SLIMC1, SOFT) to tune, while 4 input parameters (SLIMC1, SLIMT1, SOFT,
FIO) for MAT-262. This significantly reduces the number of parameters to tune com-
pared with conventional approaches [67] that utilize a complete set of input parameters
(DFAILT, DFAILC, DFAILM, DFAILS, TFAIL, SOFT, ALPHA, BETA, FBRT, and YCFAC).
• All three material models require no calibration to achieve correlation with experimental
data.
• The axial crushing results with the calibrated parameters indicate that MAT-54 mate-
rial model demonstrates the best correlation to the experiment with 22.39% higher in
maximum peak load and 18.07% lower in total energy absorption, while MAT-58 and
MAT-262 shows similar results in terms of energy absorption but both overpredict the
maximum peak load.
• The errors of the predicted maximum peak and total energy absorption still remain at
a high level (around 20%), which is a current limitation of this modeling approach.
For the optimization part, this thesis proposed two surrogate models: Deep Neural Net-
works (DNN) and Gradient Boosting Regression Tree (GBRT) ensemble. These are the state-
of-the-art machine learning models for a regression task. Then a two-stage single-objective
optimization was performed by using Mixed-Discrete Particle Swarm Optimization (MDPSO)
to obtain the optimal set of composite layup for maximum energy absorption. The first-
stage optimization was to investigate the applicability of the optimization framework and the
second-stage optimization was to investigate the transferability of the optimization framework
across design domains of different composition and size. The results of the optimization can
be concluded as follows:
97
98 6. Conclusions
• From the first-stage optimization results, the MSE error between the predicted value and
true value is relatively small: -1.21% for DNN and 1.9% for XGboost. This indicates good
performances for the two surrogate models.
• The optimal layup for DNN is [0/ − 45/45/ − 45/0/45/ − 75/90] and that for XGboost is
[0/ − 45/45/30/0/45/ − 45/90] . The former gives a 6.06% increase on the total energy
absorption, while the XGboost shows a 11.52% increase.
• The second-stage optimization results provided a different optimal layup with more ori-
entations for each ply: [0/ − 52.5/37.5/37.5/0/45/ − 45/82.5] with an energy absorp-
tion=2767 J (12.54% increase compared to initial design).
• The optimization framework proposed in this study have been proved to be both appli-
cable and transferable in the crashworthy design and optimization of composite tubes.
This optimization framework can be implemented in the design and optimization of com-
posite structures towards maximum crashworthiness performance, due to its high efficiency
and acceptable accuracy. In the future, following aspects can be concentrated for improve-
ment:
• Multi-layer modeling approach can be introduced after determining the optimal layup of
the composite by the approach proposed in this study. This can improve the accuracy
of the LS-DYNA model to alleviate the under-estimation of energy absorption due to its
capability of capturing delamination failure.
• More advanced neural networks can be investigated in the future, such as Convolutional
Neural Networks (CNN) and Recurrent Neural Networks (RNN).
• More design variables can be considered in the optimization problem. This will improve
the robustness of the proposed optimization framework.
• Other methods except for k-fold cross validation can be implemented to deal with the
lack of training data in training the regression model, such as artificially increase the
number of the samples via the SMOTE algorithm [140].
• More complex components/structures can be considered to verify the applicability of the
modeling approach, such as corrugated specimens, z-struts or even large-scale subfloor
structures.
• The requirements from the preliminary design phase can be incorporated into the pro-
posed optimization framework.
• More optimization strategies can also be developed, including uncertainties-based op-
timization and topology optimization.
Bibliography
[1] G. C. Jacob, J. F. Fellers, S. Simunovic, and J. M. Starbuck, “Energy absorption in
polymer composites for automotive crashworthiness,” Journal of Composite Materials,
vol. 36, no. 7, pp. 813–850, 2002.
[2] J. Paz, J. Díaz, L. Romera, and M. Costas, “Size and shape optimization of aluminum
tubes with GFRP honeycomb reinforcements for crashworthy aircraft structures,” Com-
posite Structures, vol. 133, pp. 499–507, 2015.
[6] D. D. Boyd, “Occupant injury and fatality in general aviation aircraft for which dy-
namic crash testing is certification-mandated,” Accident Analysis & Prevention, vol. 79,
pp. 182–189, 2015.
[7] H. S. Bautz, B. and T. Bergmann, “Tension crash absorbers for a composite fuselage
application,” SAMPE Seattle 2017, 2017.
[10] C. R. Raquel and P. C. Naval Jr, “An effective use of crowding distance in multiobjective
particle swarm optimization,” in Proceedings of the 7th annual conference on Genetic
and evolutionary computation, pp. 257–264, ACM, 2005.
[11] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective
genetic algorithm: NSGA-II,” IEEE Transactions on Evolutionary Computation, vol. 6,
no. 2, pp. 182–197, 2002.
[12] J. Fang, G. Sun, N. Qiu, N. H. Kim, and Q. Li, “On design optimization for structural
crashworthiness and its state of the art,” Structural and Multidisciplinary Optimization,
vol. 55, no. 3, pp. 1091–1119, 2017.
[13] J. Fang, Y. Gao, G. Sun, C. Xu, and Q. Li, “Multiobjective robust design optimization of
fatigue life for a truck cab,” Reliability Engineering & System Safety, vol. 135, pp. 1–8,
2015.
[14] W. Yao, X. Chen, W. Luo, M. van Tooren, and J. Guo, “Review of uncertainty-based
multidisciplinary design optimization methods for aerospace vehicles,” Progress in
Aerospace Sciences, vol. 47, no. 6, pp. 450–479, 2011.
99
100 Bibliography
[32] K. J. Bottome, The energy absorption of damaged braided and non-crimp fibre composite
material structures. PhD thesis, University of Nottingham, 2006.
[33] H. Ghasemnejad, H. Hadavinia, and A. Aboutorabi, “Effect of delamination failure in
crashworthiness analysis of hybrid composite box structures,” Materials & Design,
vol. 31, no. 3, pp. 1105–1116, 2010.
[34] J. Ma and Y. Yan, “Quasi-static and dynamic experiment investigations on the crash-
worthiness response of composite tubes,” Polymer Composites, vol. 34, no. 7, pp. 1099–
1109, 2013.
[35] J. Xu, Y. Ma, Q. Zhang, T. Sugahara, Y. Yang, and H. Hamada, “Crashworthiness of
carbon fiber hybrid composite tubes molded by filament winding,” Composite Struc-
tures, vol. 139, pp. 130–140, 2016.
[36] Y. Wang, J. Feng, J. Wu, and D. Hu, “Effects of fiber orientation and wall thickness on
energy absorption characteristics of carbon-reinforced composite tubes under different
loading conditions,” Composite Structures, vol. 153, pp. 356–368, 2016.
[37] E. Mahdi, A. Hamouda, , and T. Sebaey, “The effect of fiber orientation on the energy
absorption capability of axially crushed composite tubes,” Materials & Design (1980-
2015), vol. 56, pp. 923–928, 2014.
[38] C. Reuter, K.-H. Sauerland, and T. Tröster, “Experimental and numerical crushing
analysis of circular CFRP tubes under axial impact loading,” Composite Structures,
vol. 174, pp. 33–44, 2017.
[39] S. Palanivelu, W. Van Paepegem, J. Degrieck, J. Van Ackeren, D. Kakogiannis,
D. Van Hemelrijck, J. Wastiels, and J. Vantomme, “Experimental study on the ax-
ial crushing behaviour of pultruded composite tubes,” Polymer Testing, vol. 29, no. 2,
pp. 224–234, 2010.
[40] J. Huang and X. Wang, “Numerical and experimental investigations on the axial crush-
ing response of composite tubes,” Composite Structures, vol. 91, no. 2, pp. 222–228,
2009.
[41] H. Luo, Y. Yan, X. Meng, and C. Jin, “Progressive failure analysis and energy-absorbing
experiment of composite tubes under axial dynamic impact,” Composites Part B: Engi-
neering, vol. 87, pp. 1–11, 2016.
[42] A. Atthapreyangkul and B. G. Prusty, “Experimental and numerical analysis on the
geometrical parameters towards the maximum SEA of CFRP components,” Composite
Structures, vol. 164, pp. 229–236, 2017.
[43] D. Schmueser and L. Wickliffe, “Impact energy absorption of continuous fiber compos-
ite tubes,” Journal of engineering materials and technology, vol. 109, no. 1, pp. 72–77,
1987.
[44] D. Siromani, J. Awerbuch, and T.-M. Tan, “Finite element modeling of the crushing
behavior of thin-walled CFRP tubes under axial compression,” Composites Part B: En-
gineering, vol. 64, pp. 50–58, 2014.
[45] B. G. Falzon and W. Tan, “Virtual testing of composite structures: Progress and chal-
lenges in predicting damage, residual strength and crashworthiness,” in The Structural
Integrity of Carbon Fiber Composites, pp. 699–743, Springer, 2017.
[46] S. W. Tsai and E. M. Wu, “A general theory of strength for anisotropic materials,”
Journal of composite materials, vol. 5, no. 1, pp. 58–80, 1971.
[47] F.-K. Chang and K.-Y. Chang, “A progressive damage model for laminated compos-
ites containing stress concentrations,” Journal of composite materials, vol. 21, no. 9,
pp. 834–855, 1987.
102 Bibliography
[48] J. O. Hallquist et al., “LS-DYNA keyword user’s manual,” Livermore Software Technol-
ogy Corporation, 2016.
[49] B. A. Gama, T. A. Bogetti, and J. W. Gillespie Jr, “Progressive damage modeling of plain-
weave composites using LS-Dyna composite damage model mat162,” in 7th European
LS-DYNA Conference, pp. 14–15, 2009.
[50] Z. Hashin, “Failure criteria for unidirectional fiber composites,” Journal of applied me-
chanics, vol. 47, no. 2, pp. 329–334, 1980.
[51] J. Xiao, B. Gama, and J. Gillespie Jr, “Progressive damage and delamination in plain
weave s-2 glass/sc-15 composites under quasi-static punch-shear loading,” Composite
Structures, vol. 78, no. 2, pp. 182–196, 2007.
[52] B. A. Gama and J. W. Gillespie Jr, “Rate dependent progressive composite damage
modeling using MAT162 in ls-dyna®,” in 13th International LS-DYNA Conference, 2014.
[55] S. Pinho, L. Iannucci, and P. Robinson, “Physically-based failure models and crite-
ria for laminated fibre-reinforced composites with emphasis on fibre kinking: Part I:
Development,” Composites Part A: Applied Science and Manufacturing, vol. 37, no. 1,
pp. 63–73, 2006.
[57] C. G. Davila, P. P. Camanho, and C. A. Rose, “Failure criteria for FRP laminates,”
Journal of Composite materials, vol. 39, no. 4, pp. 323–345, 2005.
[58] A. Cherniaev, C. Butcher, and J. Montesano, “Predicting the axial crush response of
CFRP tubes using three damage-based constitutive models,” Thin-Walled Structures,
vol. 129, pp. 349–364, 2018.
[59] A. Matzenmiller, J. Lubliner, and R. Taylor, “A constitutive model for anisotropic dam-
age in fiber-composites,” Mechanics of Materials, vol. 20, no. 2, pp. 125–152, 1995.
[60] A. Mamalis, D. Manolakos, M. Ioannidis, and D. Papapostolou, “The static and dynamic
axial collapse of CFRP square tubes: finite element modelling,” Composite structures,
vol. 74, no. 2, pp. 213–225, 2006.
[62] X. Xiao, M. E. Botkin, and N. L. Johnson, “Axial crush simulation of braided carbon
tubes using MAT58 in LS-DYNA,” Thin-Walled Structures, vol. 47, no. 6-7, pp. 740–749,
2009.
[63] C. McGregor, R. Vaziri, and X. Xiao, “Finite element modelling of the progressive crush-
ing of braided composite tubes under axial impact,” International Journal of Impact
Engineering, vol. 37, no. 6, pp. 662–672, 2010.
Bibliography 103
[64] C. McGregor, N. Zobeiry, R. Vaziri, A. Poursartip, and X. Xiao, “Calibration and val-
idation of a continuum damage mechanics model in aid of axial crush simulation
of braided composite tubes,” Composites Part A: Applied Science and Manufacturing,
vol. 95, pp. 208–219, 2017.
[65] H. Han, F. Taheri, N. Pegg, and Y. Lu, “A numerical study on the axial crushing re-
sponse of hybrid pultruded and ±45 braided tubes,” Composite Structures, vol. 80,
no. 2, pp. 253–264, 2007.
[66] M. Kiani, H. Shiozaki, and K. Motoyama, “Using experimental data to improve crash
modeling for composite materials,” in Composite Materials and Joining Technologies for
Composites, vol. 7, pp. 215–226, Springer, 2013.
[67] P. Feraboli, B. Wade, F. Deleo, M. Rassaian, M. Higgins, and A. Byar, “Ls-dyna MAT54
modeling of the axial crushing of a composite tape sinusoidal specimen,” Composites
Part A: applied science and manufacturing, vol. 42, no. 11, pp. 1809–1825, 2011.
[69] L. N. Chiu, B. G. Falzon, D. Ruan, S. Xu, R. S. Thomson, B. Chen, and W. Yan, “Crush
responses of composite cylinder under quasi-static and dynamic loading,” Composite
Structures, vol. 131, pp. 90–98, 2015.
[70] Y. Ren, H. Jiang, B. Gao, and J. Xiang, “A progressive intraply material deteriora-
tion and delamination based failure model for the crashworthiness of fabric compos-
ite corrugated beam: Parameter sensitivity analysis,” Composites Part B: Engineering,
vol. 135, pp. 49–71, 2018.
[71] A. F. Johnson and M. David, “Failure mechanisms and energy absorption in compos-
ite elements under axial crush,” in Key Engineering Materials, vol. 488, pp. 638–641,
Trans Tech Publ, 2012.
[72] M. David and A. F. Johnson, “Effect of strain rate on the failure mechanisms and energy
absorption in polymer composite elements under axial loading,” Composite Structures,
vol. 122, pp. 430–439, 2015.
[78] K. E. Jackson and E. Fasanella, “Impact testing and simulation of a crashworthy com-
posite fuselage,” in ANNUAL FORUM PROCEEDINGS-AMERICAN HELICOPTER SOCI-
ETY, vol. 56, pp. 1442–1457, AMERICAN HELICOPTER SOCIETY, INC, 2000.
[79] E. L. Fasanella and K. E. Jackson, “Impact testing and simulation of a crashworthy
composite fuselage section with energy-absorbing seats and dummies,” Journal of the
American Helicopter Society, vol. 49, no. 2, pp. 140–148, 2004.
[80] E. L. Fasanella, K. E. Jackson, C. E. Sparks, and A. K. Sareen, “Water impact test and
simulation of a composite energy absorbing fuselage section,” Journal of the American
Helicopter Society, vol. 50, no. 2, pp. 150–164, 2005.
[81] C. Kindervater, R. Thomson, A. Johnson, M. David, M. Joosten, Z. Mikulik, L. Mulcahy,
S. Veldman, A. Gunnion, A. Jackson, et al., “Validation of crashworthiness simulation
and design methods by testing of a scaled composite helicopter frame section,” in Amer-
ican Helicopter Society 67th Annual Forum, pp. 944–956, American Helicopter Society
International, Inc., 2011.
[82] A. Forrester, A. Sobester, and A. Keane, Engineering design via surrogate modelling: a
practical guide. John Wiley & Sons, 2008.
[83] G. G. Wang and S. Shan, “Review of metamodeling techniques in support of engineering
design optimization,” Journal of Mechanical design, vol. 129, no. 4, pp. 370–380, 2007.
[84] J.-S. Park, “Optimal latin-hypercube designs for computer experiments,” Journal of
Statistical Planning and Inference, vol. 39, no. 1, pp. 95–111, 1994.
[85] M. E. Johnson, L. M. Moore, and D. Ylvisaker, “Minimax and maximin distance de-
signs,” Journal of Statistical Planning and Inference, vol. 26, no. 2, pp. 131–148, 1990.
[86] R. H. Myers, D. C. Montgomery, and C. M. Anderson-Cook, “Response surface method-
ology: Process and product optimization using designed experiments (wiley series in
probability and statistics),” Applied Probability and Statistics, 1995.
[87] S. Hou, Q. Li, S. Long, X. Yang, and W. Li, “Multiobjective optimization of multi-cell
sections for the crashworthiness design,” International Journal of Impact Engineering,
vol. 35, no. 11, pp. 1355–1367, 2008.
[88] C. Qi, S. Yang, and F. Dong, “Crushing analysis and multiobjective crashworthiness
optimization of tapered square tubes under oblique impact loading,” Thin-Walled Struc-
tures, vol. 59, pp. 103–119, 2012.
[89] S. Hesse, D.-J. Lukaszewicz, and F. Duddeck, “A method to reduce design complex-
ity of automotive composite structures with respect to crashworthiness,” Composite
Structures, vol. 129, pp. 236–249, 2015.
[90] J. Paz, J. Díaz, L. Romera, and M. Costas, “Crushing analysis and multi-objective
crashworthiness optimization of GFRP honeycomb-filled energy absorption devices,”
Finite Elements in Analysis and Design, vol. 91, pp. 30–39, 2014.
[91] Z. Liu, J. Lu, and P. Zhu, “Lightweight design of automotive composite bumper system
using modified particle swarm optimizer,” Composite Structures, vol. 140, pp. 630–643,
2016.
[92] J. M. Zurada, Introduction to artificial neural systems, vol. 8. West Publishing Company
St. Paul, 1992.
[93] L. Lanzi, C. Bisagni, and S. Ricci, “Neural network systems to reproduce crash behavior
of structural components,” Computers & structures, vol. 82, no. 1, pp. 93–108, 2004.
[94] E. Acar, M. Guler, B. Gerceker, M. Cerit, and B. Bayram, “Multi-objective crashworthi-
ness optimization of tapered thin-walled tubes with axisymmetric indentations,” Thin-
Walled Structures, vol. 49, no. 1, pp. 94–105, 2011.
Bibliography 105
[97] L. Shi, R. Yang, and P. Zhu, “A method for selecting surrogate models in crashwor-
thiness optimization,” Structural and Multidisciplinary Optimization, vol. 46, no. 2,
pp. 159–170, 2012.
[99] X. Song, G. Sun, G. Li, W. Gao, and Q. Li, “Crashworthiness optimization of foam-
filled tapered thin-walled structure using multiple surrogate models,” Structural and
Multidisciplinary Optimization, vol. 47, no. 2, pp. 221–231, 2013.
[100] H. Yin, G. Wen, Z. Liu, and Q. Qing, “Crashworthiness optimization design for foam-
filled multi-cell thin-walled structures,” Thin-Walled Structures, vol. 75, pp. 8–17, 2014.
[103] P. W. Jansen and R. E. Perez, “Constrained structural design optimization via a parallel
augmented lagrangian particle swarm optimization approach,” Computers & Structures,
vol. 89, no. 13-14, pp. 1352–1366, 2011.
[104] H. Xu, C.-H. Chuang, and R.-J. Yang, “A data mining-based strategy for direct mul-
tidisciplinary optimization,” SAE International Journal of Materials and Manufacturing,
vol. 8, no. 2, pp. 357–363, 2015.
[105] K. Liu, D. Detwiler, and A. Tovar, “Optimal design of nonlinear multimaterial structures
for crashworthiness using cluster analysis,” Journal of Mechanical Design, vol. 139,
no. 10, p. 101401, 2017.
[106] A. Kaddour, M. J. Hinton, P. A. Smith, and S. Li, “Mechanical properties and details of
composite laminates for the test cases used in the third world-wide failure exercise,”
Journal of Composite Materials, vol. 47, no. 20-21, pp. 2427–2442, 2013.
[107] ASTM, “D3039/D3039m-17,Standard test method for tensile properties of polymer ma-
trix composite materials,” ASTM International, 2017.
[110] M. Wisnom, B. Khan, and S. Hallett, “Size effects in unnotched tensile strength of
unidirectional and quasi-isotropic carbon/epoxy composites,” Composite Structures,
vol. 84, no. 1, pp. 21–28, 2008.
106 Bibliography
[111] X. Liu, J. Guo, C. Bai, X. Sun, and R. Mou, “Drop test and crash simulation of a civil
airplane fuselage section,” Chinese Journal of Aeronautics, vol. 28, no. 2, pp. 447–456,
2015.
[113] R. G. Boeman and A. G. Caliskan, “A novel capability for crush testing crash energy
management structures at intermediate rates,” tech. rep., SAE Technical Paper, 2002.
[114] Y. Ren, H. Jiang, W. Ji, H. Zhang, J. Xiang, and F.-G. Yuan, “Improvement of pro-
gressive damage model to predicting crashworthy composite corrugated plate,” Applied
Composite Materials, vol. 25, no. 1, pp. 45–66, 2018.
[115] Y. Ren, H. Zhang, and J. Xiang, “A novel aircraft energy absorption strut system
with corrugated composite plate to improve crashworthiness,” International Journal
of Crashworthiness, vol. 23, no. 1, pp. 1–10, 2018.
[116] Z. P. Bažant and B. H. Oh, “Crack band theory for fracture of concrete,” Matériaux et
Construction, vol. 16, no. 3, pp. 155–177, 1983.
[117] P. Maimí, P. P. Camanho, J. Mayugo, and C. Dávila, “A continuum damage model for
composite laminates: Part II–computational implementation and validation,” Mechan-
ics of Materials, vol. 39, no. 10, pp. 909–919, 2007.
[118] M. Andersson and P. Liedberg, “Crash behavior of composite structures a cae bench-
marking study,” Chalmers University of Technology, Göteborg, Sweden, 2014.
[120] R. Jin, W. Chen, and A. Sudjianto, “An efficient algorithm for constructing optimal de-
sign of computer experiments,” in ASME 2003 International Design Engineering Tech-
nical Conferences and Computers and Information in Engineering Conference, pp. 545–
554, American Society of Mechanical Engineers, 2003.
[122] Z. Jiang and M. Gu, “Optimization of a fender structure for the crashworthiness de-
sign,” Materials & Design, vol. 31, no. 3, pp. 1085–1095, 2010.
[123] Y. Bengio et al., “Learning deep architectures for AI,” Foundations and trends® in Ma-
chine Learning, vol. 2, no. 1, pp. 1–127, 2009.
[127] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep con-
volutional neural networks,” in Advances in Neural Information Processing Systems,
pp. 1097–1105, 2012.
[128] D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” ArXiv Preprint
ArXiv:1412.6980, 2014.
Bibliography 107
[132] A. Karpathy, “Cs231n convolutional neural networks for visual recognition,” Neural
networks, vol. 1, 2016.
[133] S. Arlot, A. Celisse, et al., “A survey of cross-validation procedures for model selection,”
Statistics Surveys, vol. 4, pp. 40–79, 2010.