Structural and Multidisciplinary Optimization (2024) 67:62

https://doi.org/10.1007/s00158-024-03780-4

RESEARCH PAPER

A crashworthiness optimization method of subway underframe

structures based on the differential evolution of the weighted graph
representation
Shuguang Yao1,2,3 · Yili Zhou1,2,3 · Jie Xing1,2,3 · Ping Xu1,2,3 · Qi Huang1,2,3 · Fan Zou1,2,3

Received: 8 June 2023 / Revised: 16 December 2023 / Accepted: 26 February 2024

© The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature 2024

Abstract
The crashworthiness optimization of the subway end structure is essentially a topology–shape–size collaborative optimiza-
tion problem of a composite structure under collision conditions. To this end, a rapid multivariate co-optimization method
based on weighted graph model and differential evolution algorithm is proposed. The underframe configuration is mapped
as a weighted graph and then represented by mathematical matrices. By parameter inversion, a basic parametric simplified
finite element model is established and validated by the detailed finite element model and the impact test. Three different
configuration optimizations are accomplished and compared to prove the efficiency of the proposed method. The research
results of this paper provide a reliable method and a reference for the further development of the crashworthiness optimiza-
tion for complex structures.

Keywords Subway underframe structure · Crashworthiness · Collaborative optimization · Weighted graph · Differential
evolution algorithm

1 Introduction Taghipoor et al. 2020), and optimization of energy-absorb-

ing elements(Xu et al. 2016a, b, 2017; Xing et al. 2020).
The energy absorbed by the underframe structure at the end However, the ideal crashworthy underframe structure not
of the subway vehicle is the largest during the collision acci- only needs to fully absorb the impact energy, but also be
dent, which indicates that the crashworthiness design of the able to transfer the impact forces to the vehicle behind in an
underframe structure plays an important role in the safety orderly and controlled manner, so as to protect the driver and
protection of the subway trains (Tyrell 2002). Relevant stud- passenger area from serious damage (Xie and Zhou 2014;
ies on the crashworthiness of structures have focused on the Xie et al. 2016; Yang et al. 2018). Therefore, it is of great
material property (Albooyeh et al. 2022), parameters analy- significance to carry out research on force transmission path
sis (Gao et al. 2020; Eyvazian et al. 2022; Taghipoor et al. planning and comprehensive configuration optimization to
2022, Taghipoor and Eyvazian 2022; Taghipoor and Sefidi improve the crashworthiness of the underframe structure (Lu
2022), innovative design (Taghipoor and Nouri 2018a, b; et al. 2017).
The configuration optimization of structures generally
includes three levels: size (Passos and Luersen 2018; Wang
Responsible Editor: Axel Schumacher
and Sun 2018; Liu et al. 2018), shape (Herskovits et al.
* Jie Xing 2000; Forsberg and Nilsson 2005; Gustafsson and Ström-
[email protected] berg 2008), and topology (Bakhtiarinejad et al. 2017; Wei-
1 der and Schumacher 2018; Bahramian and Khalkhali 2020).
Key Laboratory of Traffic Safety on Track (Central South
University), Ministry of Education, Changsha 410075, China Size optimization is the simplest because structural perfor-
2 mance and size variables generally exhibit linear or simple
Joint International Research Laboratory of Key Technology
for Rail Traffic Safety, Changsha 410075, China nonlinear relationships. In contrast, shape optimization is
3 more complex because there is usually a complex nonlin-
National & Local Joint Engineering Research Center
of Safety Technology for Rail Vehicle, Changsha 410075, ear relationship between structural performance and shape
China variables. Structural topology expresses the distribution

Vol.:(0123456789)
 62 Page 2 of 19 S. Yao et al.

of materials and is difficult to introduce as an independent the performance optimization under three conditions is com-
design variable in the optimization process. The configura- pleted using the genetic algorithm (Yoshimura et al. 2019).
tion optimization of the underframe structure is essentially Ortmann et al. proposed a method to optimize the configura-
a complex problem considering three levels and is therefore tion of a structure under crushing condition. The structure
more challenging. is fully parameterized by using the concept of mathematical
For such complex co-optimization problems, the current graph. Simple structural examples are displayed to verify the
solutions such as hybrid cellular automata (HCA) method efficiency of the method (Ortmann and Schumacher 2013).
(Chia et al. 2006; Duddeck et al. 2016; Afrousheh et al. In our previous work, the shape -size co-optimization
2019), equivalent static load (ESL) method (Lee and Cho of a collision frame structure was carried out based on the
2018; Bai et al. 2019; Karev et al. 2019), and surrogate weighted graph representations (Yao et al. 2022). The results
model (SM) method (Chen et al. 2018; Shu et al. 2020) are show that the introduction of the weighted graph model can
based on the idea of hierarchy, which decouples and simpli- effectively solve the collaborative problem in structural
fies the optimization problem. However, these methods can- optimization. In this paper, a further development of the
not realize the collaborative optimization of different levels weighted graph representation method is proposed and
of problems, which lead to a long search cycle and make it applied to a subway underframe structure configuration map-
difficult to obtain the ideal optimization solution at one time. ping and connectivity detection, and the crashworthiness
As an important sub-discipline of applied mathematics, optimization is carried out by using equivalent graph model
the graph theory can also be applied to structural optimiza- in parallel with a novel DE algorithm.
tion design. By using graph to characterize the size, shape The rest of this paper begins with the detailed description
and topology of the structure, the mechanical model of a of the proposed methodology in Sect. 2. Subsequently, the
specific structure is converted into a graph model, thus the basic parametric finite element model is verified in Sect. 3.
synergistic organization of different types of optimization In Sect. 4, three configuration optimizations of the under-
variables is realized. Some exploratory studies have been frame are carried out to prove the efficiency of the method
carried out by scholars. proposed in this paper. Finally, conclusions are drawn in
In 2004, Kawamoto et al. first propose the concept of Sect. 5.
graph to describe the structure (Kawamoto et al. 2004).
Based on the graph theory, the nodes and beam element in
the structure are described by the vertices and edges in the
graph, respectively. Thus, the equivalent mapping relation- 2 Methodology
ship between the structure and the graph is established. A
graph is a mathematical model that can be expressed and 2.1 Structure description and graph‑based
operated by a sequence of numbers or vectors. Therefore, representation
with the help of the graph, structural parameters such as size,
shape, and topology can be converted into specific numerical 2.1.1 Geometry of the underframe structure
vectors, which can be easily applied to various optimization
algorithms. The partition design scheme of a subway underframe struc-
Giger et al. equated the truss structure to a mathemati- ture is displayed in Fig. 1. Each part of the underframe is
cal graph and applied it directly to the genetic operator in described as follows:
the genetic algorithm for optimization. The efficiency of the
concept is demonstrated by three structural optimization (1) Primary honeycomb: The honeycomb is located at
examples (Giger and Ermanni 2006). Sauter et al. performed the forefront of the underframe. Its main function is
graph modeling of beam structures with variable cross sec- to absorb a small amount of energy using limited dis-
tions and optimized the structural configurations using placement under the low-speed crushing condition.
genetic algorithms (Sauter et al. 2008). Giger et al. solve (2) Secondary crushing tubes: After the dissipation of the
the global parameter optimization problem of a laminated primary honeycomb, the secondary crushing tubes act
structure using a graph model. An equivalent graph model sequentially to absorb the remaining kinetic energy
is established by performing structural feature extraction, under the high-speed crushing condition.
and the optimization is accomplished using the evolutionary (3) Guidance: Auxiliary guide structures are designed on
algorithm (Giger et al. 2008). both sides of the underframe to constraint the lateral
Yoshimura et al. carried out a structural optimization of a displacement.
miniature infusion structure to obtain better flow field con- (4) Bearing beams: The bearing beams have a large stiff-
trol performance. The layout parameters of the upper struc- ness and their function is to distribute the force reason-
ture are used as design variables through a graph model, and ably to the rear-end during the crushing process.
A crashworthiness optimization method of subway underframe structures based on the differential… Page 3 of 19 62

Fig. 1  Partition design scheme

of the underframe

Obviously, the topological configuration and energy vertex. The matrix D=(dij) with size n×n is called a weighted
absorption characteristics of secondary crushing tubes and adjacency matrix, where:
bearing beams have an important influence on the transmis- {
sion of crushing loads. Therefore, the above two parts of kij vi is adjacent to vj
dij = (1)
the underframe are selected as the optimization area in the 0 vi is not adjacent to vj
paper.
To construct a weighted graph model, the necessary
2.1.2 Weighed graph representation features are extracted from the underframe structure,
firstly. As can be seen in Fig. 2, considering the symmetry
A graph with a number on its edge is called a weighted of the underframe, the left side of( the structure
) is repre-
graph. The weight can be understood as a mathematical sented as a weighted graph G0 = V0 , E0 . The vertex set
abstraction that represents a physical attribute belong- V0 = {v1 , v2 , … , v15 } includes {
12 active vertices
} and 3 fixed
ing to an edge in the weighted graph. Therefore, any par- vertices. The edge set E0 = e1 , e2 , … , e17 includes 15
ticular structure can be described using a weighted
{ graph} active edges and 2 fixed edges. Furthermore, the correspond-
G = (V(G), E(G), C(G), D(G)). The{V(G) = v1 , v}2 , … , vn ing coordinate matrix can be determined by the position of
is called a vertex set. The E(G) = e1 , e2 , … , em is called each vertex, as displayed in Eq. 2. It is worth mentioning that
an edge set. The matrix C=(cij) with size n×3 is called a since the underframe can be simplified as a planar structure,
coordinate matrix. The three elements in the ith column of the coordinate matrix C0 only contains x and y coordinate
the matrix correspond to the x, y, and z coordinates of the ith information.

Fig. 2  Geometry description and graph representation of the underframe

 62 Page 4 of 19 S. Yao et al.

[ ]
0 510 510 510 510 510 510 716 716 716 716 716 906 906 906
C0 = (2)
485 1160 610 485 418 272 0 1160 610 418 272 0 1160 610 418

adjacency matrices D01, D02, and D03. Therefore, the struc-

The weight of each edge can be determined by the ture optimization can be seen as the graph or matrices
mechanical properties of the corresponding structure ele- optimization.
ments. For the secondary crushing tubes, the mean crush- In this paper, a novel DE algorithm with a dual self-adap-
ing force can be treated as the weight, directly. As for the tive mutation operator, which is named DSADE, is applied for
bearing beams, the axial compression and lateral bending graph optimization (Xing et al. 2021).
coexist during the crushing process. Thus, the section width
and thickness are selected as the weights, respectively. The 2.2.1 DSADE algorithm
axial force and bending moment can be analytical calculated
by the section dimensions. DE algorithm is a simple and efficient population-based
Through the above analysis, as the weight information of method invented by Storn and Price (1996) and has been
the edge contains three features, the complete representation explored by many scholars to derive several variants (Opara
of the underframe structure is G0 = (V0, E0, C0, D01, D02, and Arabas 2019). In DSADE, a mutation strategy “DE/tar-
D03). Considering the sparsity of the weighted adjacency get-to-best” in a dual self-adaptive manner is implemented. In
matrix, the corresponding ternary array representation is addition, an elitist selection technique is applied to the selec-
displayed in Table 1. tion phase to increase convergence speed. This technique has
According to the information indicated in C0, D01, D02, been successfully applied to numerous studies (Zhang and
and D03, the dimension, shape, and topology characteris- Sanderson 2009; Islam et al. 2012; Yu et al. 2015; Cui et al.
tics of the underframe structure are well organized by the 2016). A brief introduction to the graph optimization process
weighted graph model. of DSADE can be described as following 5 steps: initialization,
mutation, crossover, selection, and convergence judgment.
2.2 Graph optimization Generally, the core parameters including number of popula-
tion (NP), scaling factor (F), and crossover rate (CR) need to
Based on the weighted graph model, the underframe struc- be defined, firstly.
ture is represented by the coordinate matrix C0 and weighted
Step 1: Initialization. The initial population is randomly
Table 1  The ternary array representation of the weighted adjacency generated based on the value limits of each matrix,
matrix which can be expressed as:
Row No. Column No. Weighted adjacency matrix ( )
Xi,j (0) = XLj + randj (0, 1) XU (3)
L
j − Xj i = 1, … , NP, j = 1, … , Dim
D01 D02 D03
where X is the population with dimension of Dim. The
1 4 950 – –
subscripts i and j represent the ith individual and jth matrix
2 3 – 80 8
variables in the current population. The superscripts U and
2 8 – 160 20
L represent the upper and lower bounds for the matrices,
3 4 – 80 8
respectively. rand(0,1) represents a random matrix arbitrar-
3 9 – 80 6
ily set within the range of 0 to 1.
4 5 – 80 8
5 6 – 80 8
Step 2: Mutation. This operation aims to obtain mutated
5 10 – 80 6
individuals by performing interpolation operations
6 7 – 80 8
6 11 – 40 4
on different individuals in the current population,
8 9 – 80 8
which can be expressed as Eq. (4).
8 13 – 160 20
( ) ( )
V i (g + 1) = Xr1 (g) + Fc Xbest (g) − Xr1 (g) + Fg Xr2 (g) − Xr3 (g)
9 10 – 80 8 (4)
9 14 – 80 6
where Fc and Fg are control parameters that adaptively
10 11 – 80 8
change with the iteration process. The detailed definitions
10 15 – 80 6
are displayed in Eqs. (5) and (6).
11 12 – 80 8
A crashworthiness optimization method of subway underframe structures based on the differential… Page 5 of 19 62

−1
Fc = 1 − e |fbest −fmean | (5)

[ ( )]/
𝜋 g − 0.5gm
Fg = 1 − sin ⋅ 4 + 0.3 (6)
2 0.5gm

Compared to the regular mutation operator, this

strategy can better balance the search ability and
convergence speed of the algorithm.

Step 3: Crossover. This operation aims to obtain candi-

date individuals by cross information between the
current individuals and the mutated individuals,
which can be shown in Eq. (7).
{
V i,j (g + 1) if randj < CR
Ui,j (g+1) = (7)
Xi,j (g) otherwise

where Ui,j(g+1) represents the jth matrix contained in the

ith candidate individual. randj is a random matrix between
Fig. 3  The sorting method of the selection operator
0 and 1, and CR is a pre-defined real number.

Step 4: Selection. By sorting the candidate population and be easily achieved by sorting the fitness function values of
the current population, the individual with a better all individuals.
objective function value will enter the next popu- Step 5: Convergence judgment. After selection, if the new
lation as the updated individual. This elitist tech- population has already met the convergence criterion, the
nique can further improve the convergence effi- algorithm ends and the optimal individual is the final opti-
ciency (Padhye et al. 2013). It is worth mentioning mization result. Otherwise, the algorithm repeats steps 2 to
that for optimization problems with constraints, 4 until the convergence condition is met or the maximum
the selection needs to comprehensively consider iterations is reached.
the objective function value and constraint value
of the individuals. In this paper, a comprehensive
sorting method proposed by Deb is adopted, which 2.2.2 Configuration check
can be described in Eq. (8) and Fig. 3 (Deb 2000).
In the iterations of the DSADE algorithm, finite element
⎧ f (X) if wk (X) ≥ 0 ∀k = 1, 2, … , m
⎪ calculations are performed for each individual, i.e., the
F(X) = ⎨ �K
(8) structure corresponding to the weighted graph matrices, so
⎪ fmax + ⟨wk (X)⟩ otherwise
⎩ k=1
as to extract the mechanical response as the objective and
constraint values. Since the individuals are generated based
where F(X) is the fitness function value, f(X) is the objective on a stochastic evolutionary strategy, some of the randomly
function value, wk(X) is the constraint violation, m repre- generated structural configurations are not directly usable
sents the number of constraints, fmax is the objective func- for crushing analysis.
tion value of the worst feasible solution in the population, Two typical useless configurations are indicated in Fig. 4.
operator ⟨ ⟩ denotes the absolute value of the operand, if the In Fig. 4a, the beam unit {v6, v11} is not arranged on the
operand is negative and returns a value zero, otherwise. force transmission path, thus the existence of this beam unit
As displayed in Fig. 3, the population is divided into fea- is meaningless. In Fig. 4b, the constrained v15 becomes an
sible and nonfeasible domains based on the worst feasible isolated vertex, which causes the structure to lose a force
solution. In the feasible domain, each individual is sorted transmission channel. Therefore, it is necessary to exclude
according to its objective function value, while in the non- unreasonable configurations before performing a crushing
feasible domain, each individual is sorted according to its analysis of the individual. This checking process is carried
constraint violation value. Therefore, the elite selection can out based on the concept of connected components in the
weighted graph theory.
 62 Page 6 of 19 S. Yao et al.

Fig. 4  Two typical useless underframe configurations

{ }
For a division V1 , V2 , … , Vm of a weighted graph the data of the weighted adjacency matrices D01
G = (V(G), E(G)), vertices x and y are defined as connected and D02, the starting node and the ending node of
only when they belong to the same subset Vi. If Vi consists the beam elements are determined according to the
of only one isolated node, then subgraph G[Vi] is called row and column where the nonzero elements are
the zero connected component; otherwise, it is called the located. For the configuration shown in Fig. 2, 15
nonzero connected component. nodes and 17 1D elements are created.
Based on the above definition, it can be seen that in
Fig. 4a, 2 nonzero connected components are included, and
in Fig. 4b, the constrained node v15 constitutes a zero con- Step 2: Properties. For the secondary crushing tubes, the
nected component. In this regard, the criterion for configura- nonlinear spring element is used to simulate its
tion check can be obtained as follows: axial collapse process, thus the section proper-
ties do not need to be defined specifically. As for
(1) For each individual, its corresponding weighted graph the bearing beams, the section properties can be
contains at most 1 nonzero connected component; defined using the keyword “*SECTION_BEAM”
(2) Load and constraint vertices should all belong to this on the software. The section type “Tubular Box”
nonzero connected component. is chosen to describe the rectangular section of the
bearing beams, the corresponding section proper-
In this paper, Depth First Search Method (DFSM) is ties can be assigned according to the matrices D02
applied to split the connected components of the graph indi- and D03.
viduals and filter the useless configurations. The pseudocode
of the DFSM is indicated in Table 2. Step 3: Materials. In LS-DYNA, Belytschko element and
According to the above method, for the individual that special material type "MAT_FORCE_LIMITED"
meets the check criterion, further crushing model and cal- are provided to simulate the elastic–plastic prop-
culation are performed to obtain its crushing response; oth- erties of beam structures. The main basic param-
erwise, the responses are directly set to “∞.” eters (density, elastic module, Poisson's ratio) can
be obtained by calculating the cross-sectional
parameters of the beam element and input into the
2.2.3 Parametric simplified finite element model (SFEM) model. The bending moment-turning angle curves
and collapse forc–-displacement curves obtained
After completing the configuration check of the individual, from well-established theoretical methods (Wier-
a parametric SFEM can be built in LS-DYNA based on the zbicki 1983; Abramowicz and Jones 1984) are
weighted graph matrices, which mainly includes the follow- added as inputs to the material model to define
ing steps: the bending and collapsing characteristics of the
beam elements.
Step 1: Nodes and elements. Create nodes based on the
data in coordinate matrix C0, with each column
representing the coordinates of a node. Based on
A crashworthiness optimization method of subway underframe structures based on the differential… Page 7 of 19 62

Table 2  The pseudocode of the DFSM

20 km/h velocity is built to simulate the impact

source.

Finally, the SFEM of the underframe structure is shown

in Fig. 5.

2.2.4 Optimization framework

According to the above analysis, the algorithm framework

Fig. 5  The SFEM of the underframe structure
of the graph optimization is shown in Fig. 6. It can be seen
that the optimization process consists of the optimization
Step 4: Boundary Conditions. After completing the mod- module and the analysis module. The optimization module
eling of the optimization area, it is necessary to constitutes the core process of the algorithm, and the analy-
import the external model of the underframe pre- sis module provides the objective function and constraint
pared in advance and establish the connections. values of the individuals for the optimization module.
In addition, a rigid wall with 37.56 t mass and
 62 Page 8 of 19 S. Yao et al.

Fig. 6  The framework of the

underframe graph optimization

Table 3  Material foundation parameters involved in the FEM

Parameters ASTM AISI 301 AISI 301 AISI 301
A588-A LN-1/16 H LN-1/8 H LN-1/4 H

Density (kg/m3) 7700 8000 8000 8000

Young’s modulus 206 186 186 186
(GPa)
Yield stress (MPa) 345 345 415 515
Ultimate stress (MPa) 485 690 760 825
Elongation (%) 21 40 35 25
Tangent modulus 672 867 992 1254
(MPa)
Poisson's ratio (–) 0.30 0.30 0.30 0.30

Fig. 7  Detailed finite element model of the underframe structure in

impact test condition
Considering the whole structure of the test sled is rela-
tively large and the main plastic deformation occurs in the
3 Model validation front-end area, different mesh sizes are used for different
parts of the entire model after the sensitive analysis. Finally,
3.1 Detailed finite element model (DFEM) the mesh size of the underframe structure and the test sled
are chosen to be 5 mm and 20 mm, respectively.
In order to verify the accuracy of the parametric model, a In order to simulate the elastic–plastic behavior of
DFEM of the underframe structure is established. As dis- structures during impact condition, the material type
played in Fig. 7, the model consists of a track, a test sled, an “MAT_PLASTIC_KINEMATIC” is selected for the test
underframe structure, and a rigid wall. The total mass of the sled and the underframe structure. This model simu-
test sled and underframe structure is 37.56 t, and they have lates the elastic–plastic behavior of materials by defining
the same initial velocity of 5.59 m/s. stress–strain curves. The input curve data can be obtained
by performing quasi-static test on material samples.
A crashworthiness optimization method of subway underframe structures based on the differential… Page 9 of 19 62

Table 3 shows the material foundation parameters involved stable. However, local plastic buckling occurs in the rear-end
in the finite element model. of the underframe. According to the sequence comparison,
On the other hand, the “AUTOMATIC_SINGLE_SUR- the simulation results of both the SFEM and the DFEM are
FACE” contact algorithm is selected to simulate the self- very close to the experimental phenomenon.
contact of the entire model. The static and dynamic fric- The statistical comparison of the displacement of vari-
tion coefficients are defined as 0.3 and 0.2, respectively. ous parts of the underframe is shown in Table 4. The results
The “CONTACT_TIED_SURFACE_TO_SURFACE” indicate that the crushing displacement calculated by the
contact algorithm is chosen to simulate the fixed connec- simplified model is highly consistent with the experimental
tion between the underframe structure and the test sled. results. The relative error of the total displacement between
the simplified model and the experiment is 0.42%.
3.2 Impact test The comparison of the crushing force of the underframe
is displayed in Fig. 10. It can be clearly seen that the crush-
The impact test is conducted on the vehicle crash test bench ing force histories obtained by the two finite element models
at the Central South University. The test set-up is shown in are excellent agree with the experimental result. The entire
Fig. 8. crushing process is clearly divided into three stages: (i)
During the test, synchronous trigger signal is used the first stage is when the honeycomb collapses, which is
between various data acquisition systems to provide the relatively stable, with an average crushing force of about
same time axis for all data. Therefore, each type of data has 1200 kN; (ii) at the beginning of the second stage, due to
the same time axis to facilitate analysis and calculation of the inability of the honeycomb to further compress, a large
key characteristics such as mean crushing force (MCF) and peak force occurred, resulting in plastic buckling of the rear
energy absorption (EA). beams; and (iii) in the third stage, the energy uniformly dis-
sipated by crushing tube, with an average force of about
3.3 Comparison results 2000 kN.
The energy absorption results of the underframe are com-
Firstly, the deformation mode of the underframe structure pared in Fig. 11. During the crushing experiment, the energy
during the collision are compared in Fig. 9. The results indi- absorbed by the underframe is 577.48 kJ. The calculation
cate that the collapse process of the underframe is relatively

Fig. 8  Impact test set-up

(a) Schematic diagram of test scenario

(b) Load cells (c) High-speed camera (d) Speed sensor

 62 Page 10 of 19 S. Yao et al.

Fig. 9  Comparison of the defor-

mation mode of the underframe
structure

(a) t = 0ms (b) t = 20ms

(c) t = 120ms (d) t = 137ms

Table 4  Comparison results of the crushing displacement

Displacement (mm) Experiment DFEM SFEM

Honeycomb 86.50 87.25 89.09

Crushing tube 216.00 214.92 212.76
Bearing beam 27.00 26.88 26.26
Total 329.50 329.05 328.11

Fig. 11  Comparison of the energy absorption of the underframe

force, and energy absorption calculated by SFEM are excel-

lent agree with the corresponding experimental results.
On the other hand, it is difficult to directly measure the
force distribution of the three main force transmission paths
during the experiment. However, it can be extracted from the
calibrated SFEM, which is indicated in Fig. 12. As can be
Fig. 10  Comparison of the crushing force of the underframe seen from the figure, the crushing force is mainly transmit-
ted to the rear-end through vertices v15 and v14. The average
forces on the three main transmission paths are 415.86 kN,
result of the SFEM is 576.02 kJ, with an error of only 0.25% 356.04 kN, and 74.56 kN, respectively. It is precisely
compared to the experimental data. because the crushing force is mainly concentrated at vertices
Based on the above analysis, it can be fully demonstrated v15 and v14 that local buckling and unexpected deformation
that the SFEM established based on the weighted graph occurs at the rear-end of the underframe.
model is accurate and reliable. The displacement, crushing
A crashworthiness optimization method of subway underframe structures based on the differential… Page 11 of 19 62

Therefore, the Sef can also be defined as the propor-

tional relationship between Sc and Ss.
(3) Load transfer distribution. As indicated in Fig. 12,
the concentrated crushing force can lead to the local
buckling and unexpected deformation. Therefore, a uni-
formly distributed impact force should be obtained in
the optimization. The uniformity of force distribution
(UFD) is defined to provide a quantitative evaluation
of the load distribution:
{ }
max R13 , R14 , R15
0.9 ≤ UFD = { } ≤ 1.1 (11)
min R13 , R14 , R15

where R13, R14, and R15 represent the crushing force

through the vertices v13, v14, and v15, respectively.
Fig. 12  The force distribution of the rear-end of the underframe (4) Limited displacement. In an ideal crushing process, the
impact energy should be fully absorbed by the energy-
absorbing elements, and plastic deformation should not
4 Optimization occur in the bearing beam area. Thus, the maximum
longitudinal displacement 𝛿max of the bearing beams is
4.1 Problem definition limited to less than 1 mm.

According to the previous section, the optimization in this On the other hand, the geometry constraints should
paper focuses on the energy absorption characteristics of also be considered in the optimization: the crushing tubes
the secondary tubes and the geometry parameters of the should be arranged in the longitudinal direction to avoid
rear bearing beams. Therefore, the optimization objective instability caused by eccentric loading; the section width
is defined to minimize the total mass of the bearing beams. and thickness of the bearing beams are constrained in the
The design performance constraints can be summarized range of [40 mm, 80 mm] and [4 mm, 8 mm], respectively.
as follows: Based on the above analysis, the mathematical model of
the optimization problem can be expressed as:
(1) Energy dissipation. During the collision condition, min Mass = f (D31 , D32 , D33 , C3 )
the underframe structure must first ensure that it can
fully absorb the impact kinetic energy. Therefore, the ⎧ 0.95 ≤ E ≤ 1
⎪ f
energy absorption efficiency Eef is the first important ⎪ 0.50 ≤ Sef ≤ 1
constraint, which can be defined as: ⎪
⎪ 0.9 ≤ UFD ≤ 1.1
Et ⎪
0.95 ≤ Eef = ≤1 (9) ⎪ 𝛿max ≤ 1 (12)
Ed s.t.⎨ i,j
⎪ 400 ≤ D31 ≤ 1200
where the Et is the absorbed energy, and the Ed is the ⎪ 40 ≤ Di,j ≤ 80
⎪ 32
total kinetic energy. In this optimization problem, the ⎪ −1 ≤ Di,j ≤ 8
Eef should not be less than 0.95. ⎪ 33
(2) Stroke efficiency. This parameter reflects the utilization ⎪ CL ≤ C ≤ CU
⎩ 3 3 3
ratio of the energy-absorbing tubes to design space,
which can be described as: where the f (D31, D32, D33, C3) is the function to calculate
the structure mass from the weighted adjacency matrix and
Sc 1.43Sc
0.5 ≤ Sef = = ≤1 (10) coordinate matrix; CL3 and CU
3
represent the lower and upper
Sd Ss limits of the coordinate matrix.
where Sc and Sd are the crushing stroke and the design
stroke of the tubes, respectively. Generally, the design
stroke is 70% of the total length S s of the tubes.
 62 Page 12 of 19 S. Yao et al.

Fig. 13  Initial layout configurations with different number of crushing tubes

Table 5  The optimization results obtained from different DE variants

Applied algorithm Optimized mass Average
iteration
Best Worst Mean

DE-Rand/1 101.79 109.27 105.88 77

DE-Best/1 119.61 155.72 136.85 28
DSADE 101.79 108.52 103.64 50

4.2 Initial layout configurations

As mentioned in the previous section, the arrangement and

layout of energy-absorbing components not only deter-
mine the energy absorption characteristics of the under-
frame structure, but also have an important influence on
the crushing load transmission. Therefore, as displayed
Fig. 13, the optimization is carried out based on three ini- Fig. 14  The iterations of the calculated individuals
tial configurations (CT-2, CT-3, and CT-4) with differ-
ent crushing tube layouts. In order to achieve fully global
search, sufficient beam elements have been established DE variants with different mutation operator (Rand/1 and
between adjacent vertices in the bearing beams area. Best/1) are compared and displayed in Table 5. Each algo-
rithm was run independently 5 times. The results indicate
4.3 Results and discussion that the optimized mass values obtained from DSADE agree
well with those given by “Rand/1” DE but the iterations
4.3.1 CT‑2 configuration required are significantly reduced. On the contrary, the
“Best/1” DE has a fast convergence speed but with a local
Firstly, in order to prove the effectiveness of the DSADE result. the “Rand/1” DE process has a strong exploration
algorithm, the optimization results of the DSADE and two ability, and thus, the optimized values of 10 runs are quite
A crashworthiness optimization method of subway underframe structures based on the differential… Page 13 of 19 62

of two standard DE processes. Thus, the DSADE will be

used in all subsequent optimizations.
Subsequently, the details of the DSADE algorithm solu-
tion process are analyzed and presented in Figs. 15, 16, 17.
Due to the integration of configuration check in the algo-
rithm process, the underframe structures with unreasonable
configurations are filtered directly during the iterations. The
variation of the calculated individuals during the best opti-
mization is displayed in Fig. 14. In the first 28 iterations, the
average proportion of calculated individuals is only 12.08%,
indicating that the algorithm mainly performs large-scale
searches across the entire domain; thus, a large number of
unreasonable configurations are filtered. As the local opti-
mization begins, the proportion of calculated individuals
gradually increases. During the entire iteration process, the
total calculated individuals are 4689, accounting for 29.85%
Fig. 15  The iterations of the optimization objective
of all populations. It can be concluded that the checking
technique proposed in this paper can effectively screen out
stable with a limit deviation. Therefore, it can be seen that unreasonable configurations, thereby significantly improving
the DSADE successfully combines the respective advantages the optimization efficiency.

Fig. 16  The iterations of the constraints

 62 Page 14 of 19 S. Yao et al.

Fig. 17  The iterations of topology configuration with two tubes

Table 6  The beam section dimensions of the optimized CT-2 con-

figuration

Code Width (mm) Thickness (mm)

1 40.00 4.00
2 40.00 5.54
3 40.00 4.72
4 40.00 4.00
5 40.00 4.00
6 40.00 4.00
7 40.00 4.00
8 40.00 4.00
9 40.00 6.02
Fig. 18  The critical geometry parameters of the optimized CT-2 con-
figuration

The variation of the objective mass is shown in Fig. 15. It

can be seen that the mass fluctuated and gradually decreased
after 20 iterations. Finally, the optimized solution of
101.79 kg is obtained in the 51st iteration.
The iterations of the constraints are shown in Fig. 16.
The results show that the energy absorption efficiency Ef
and the stroke efficiency Sef always meet their defined limits
during the optimization. The variation of the force distribu-
tion UFD is most significant, gradually decreasing from the
initial value of 2.47 to 1.08. The maximum displacement
𝛿max is 1.45 mm at the beginning, which gradually decreases
and drops below 1 mm after 14th iteration.
The iterations of topology configuration with two tubes
are displayed in Fig. 17. The crushing tubes moves away
from the symmetry line, and the topology configuration of Fig. 19  The rear-end force distribution of the optimized CT-2 con-
the bearing beams is gradually simplified. Finally, the direct figuration
A crashworthiness optimization method of subway underframe structures based on the differential… Page 15 of 19 62

Fig. 20  The iterations of topology configuration with three tubes

Table 7  The beam section dimensions of the optimized CT-3 con-

figuration
Code Width (mm) Thickness (mm)

1 40.00 5.32
2 40.00 4.63
3 57.04 4.20
4 53.08 4.36
5 49.12 5.98
6 40.00 4.42
7 40.00 4.42
8 40.00 4.42
Fig. 21  The critical geometry parameters of the optimized CT-3 con- 9 40.00 6.00
figuration 10 40.00 4.03

load transmission paths from tubes to rear-end vertices are

obtained. The critical geometry parameters and beam sec-
tion dimensions of the optimized CT-2 configuration are
displayed in Fig. 18 and Table 6, respectively.
The optimized force distribution is shown in Fig. 19.
The mean crushing force through v9–v11 are 286.45 kN,
280.93 kN and 308.29 kN, respectively. The results indicate
that the optimized configuration can fully utilize the limited
beams to transmit crushing forces, thereby achieve the light-
weight design goal.

4.3.2 CT‑3 configuration

The topology iterations of configuration with three tubes

are displayed in Fig. 20. It can be seen that the two sym-
metrically arranged tubes move away from the center Fig. 22  The rear-end force distribution of the optimized CT-3 con-
line, and their design stroke also decreases accordingly. figuration
 62 Page 16 of 19 S. Yao et al.

Fig. 23  The iterations of topology configuration with four tubes

Table 8  The beam section dimensions of the optimized CT-4 con-

figuration

Code Width (mm) Thickness (mm)

1 40.00 4.00
2 40.00 4.00
3 40.00 4.00
4 40.00 4.00
5 40.00 4.00
6 40.00 4.00
7 40.00 4.00
8 40.00 4.00
Fig. 24  The critical geometry parameters of the optimized CT-4 con- 9 40.00 4.00
figuration
10 40.00 4.00
11 40.00 4.00
12 40.00 4.00
Compared with the optimized configuration with two
tubes, due to the set-up of an additional crushing tube at
the center line, corresponding force transmission paths
from center tube to rear beams are created. The geometry
details are given in Fig. 21 and Table 7, respectively. tubes are more uniform and the topology configuration of
The force distribution of optimized configuration with the rear beams is greatly simplified. Similarly, the detailed
three tubes is indicated in Fig. 22. The crushing force structural parameters are provided in Fig. 24 and Table 8,
is transmitted along the four main transmission paths respectively.
to the three constraint vertices. The distributed forces As can be seen in Fig. 25, the optimized force through
through v9–v11 are 268.81 kN, 251.75 kN, and 267.10 kN, v 11–v 13 are 278.75 kN, 260.06 kN, and 267.29 kN,
respectively. respectively.

4.3.3 CT‑4 configuration 4.3.4 Comparisons

Compared to the above two solutions, the optimized con- To further illustrate the influence of initial configuration
figuration with four tubes is the most complex, as shown on optimization results, the optimized configurations are
in Fig. 23. After 30 iterations, the intervals of the crushing compared in this section. Firstly, the force histories of
A crashworthiness optimization method of subway underframe structures based on the differential… Page 17 of 19 62

Fig. 25  The rear-end force distribution of the optimized CT-4 con-

figuration
Fig. 27  The comparison of performance indicators between different
optimization results

advantages in different parameters. The overall comparison

shows that the optimized mass of CT-2 is the smallest, and
all performance indicators are relatively balanced, which
can be considered as the most ideal result among the three
configurations.

5 Conclusions

In this paper, a crashworthiness optimization method of the

subway underframe structures based on weighted graph rep-
Fig. 26  The comparison of force history between different optimiza-
resentation in parallel with DSADE algorithm is provided.
tion results Based on the validated SFEM, three configuration optimiza-
tions are carried out and compared. The following conclu-
sions are drawn:
different optimization results are compared in Fig. 26. It
can be seen that the crushing force of the three optimized (1) The subway underframe structure is successfully rep-
solutions are clearly divided into two stages, correspond- resented by a weighted graph model. The size, shape,
ing to the energy dissipation of the honeycomb and the and topology information of the structure are organized
crushing tubes. This phenomenon demonstrates that the and characterized synergistically by the mathematical
optimization improves the topology configuration of the matrices.
rear beams, resulting in no further plastic deformation (2) To provide the crushing responses of the underframe
occurred. Therefore, the crushing tubes are collapsed in structures, the SFEM based on graph matrices is estab-
an orderly manner with a stable force history. lished and validated. The crushing displacement, force,
The performance indicators are compared in Fig. 27. As and absorbed energy calculated by SFEM are excellent
can be seen in the figure, the parameters of the configura- agree with the DFEM and corresponding experimen-
tion CT-2 and CT-4 are relatively similar. The 𝛿max and the tal data. The results fully demonstrate that the SFEM
UFD of the solution CT-3 are both optimal, while the Sef established based on the weighted graph model is accu-
is significantly lower than the other two configurations. rate and reliable.
Based on the above analysis, it can be concluded that (3) To illustrate the effectiveness of the proposed method in
the optimization results of the three initial configurations application of the crashworthiness optimization prob-
can meet all performance constraints and have their own lem, three optimizations based on different configura-
 62 Page 18 of 19 S. Yao et al.

tions are carried out. It is found that the present method Chen M, Xiao X, Tong J et al (2018) Optimization of loading path in
can effectively deal with the optimization problem for hydroforming of parallel double branched tube through response
surface methodology. Adv Eng Softw 115:429–438. https://​doi.​
underframe structures with complex crashworthiness org/​10.​1016/j.​adven​gsoft.​2017.​11.​003
constraints. Furthermore, by doing the configura- Chia CM, Rongong JA, Worden K (2006) Structural optimisation using
tion check, the optimization efficiency significantly a hybrid cellular automata (HCA) algorithm. Appl Mech Mater
improved. In addition, according to the comparisons, 56:93–100
Cui L, Li G, Lin Q et al (2016) Adaptive differential evolution algo-
the optimized configuration CT-2 can be considered as rithm with novel mutation strategies in multiple sub-populations.
the most ideal result among the three configurations. Comput Oper Res 67:155–173. https://d​ oi.o​ rg/1​ 0.1​ 016/j.c​ or.2​ 015.​
09.​006
Deb K (2000) An efficient constraint handling method for genetic algo-
Acknowledgements The research presented in this paper was con- rithms. Comput Methods Appl Mech Eng 186:311–338. https://​
ducted with the support of the National Key Research and Development doi.​org/​10.​1016/​S0045-​7825(99)​00389-8
Program of China (Grant No. 2021YFB3703801), the Hunan Provincial Duddeck F, Hunkeler S, Lozano P et al (2016) Topology optimiza-
Natural Science Foundation of China (Grant No. 2021JJ30853), and tion for crashworthiness of thin-walled structures under axial
the Leading Talents of Science and Technology of Hunan Province impact using hybrid cellular automata. Struct Multidisc Optim
(Grant No. 2019RS3018). 54:415–428. https://​doi.​org/​10.​1007/​s00158-​016-​1445-y
Eyvazian A, Taghipoor H, Tran T (2022) Analytical and experi-
Funding Funding was provided by the National Key Research and mental investigations on axial crushing of aluminum tube with
Development Program of China (Grant No. 2021YFB3703801), Natu- vertically corrugated. Int J Crashworthiness 27(4):1032–1045.
ral Science Foundation of Hunan Province (Grant No. 2021JJ30853), https://​doi.​org/​10.​1080/​13588​265.​2021.​18929​54
and the Leading Talents of Science and Technology of Hunan Province Forsberg J, Nilsson L (2005) On polynomial response surfaces and
(Grant No. 2019RS3018). Kriging for use in structural optimization of crashworthiness.
Struct Multidisc Optim 29:232–243. https://​doi.​org/​10.​1007/​
Declarations s00158-​004-​0487-8
Gao GJ, Zhuo TY, Guan WY (2020) Recent research development
Conflict of interest The authors declare that they have no conflict of of energy-absorption structure and application for railway vehi-
interest. cles. J Cent South Univ 27:1012–1038. https://​doi.​org/​10.​1007/​
s11771-​020-​4349-3
Replication of results The results reported in this research were per- Giger M, Ermanni P (2006) Evolutionary truss topology optimization
formed in MATLAB (main procedure) and Tcl/Tk (pre/post-procedure using a graph-based parameterization concept. Struct Multidisc
of the crushing analysis). The authors will help interested researchers Optim 32:313–326. https://​doi.​org/​10.​1007/​s00158-​006-​0028-8
reproduce the results given in the article. Interested readers can contact Giger M, Keller D, Ermanni P et al (2008) A graph-based parameter-
the corresponding author for basic codes of this research with reason- ization concept for global laminate optimization ETH Library
able requests. A graph-based parameterization concept for global laminate
optimization. Struct Multidisc Optim. https://​doi.​org/​10.​3929/​
ethz-b-​00001​5175
Gustafsson E, Strömberg N (2008) Shape optimization of cast-
ings by using successive response surface methodology.
References Struct Multidisc Optim 35:11–28. https://​d oi.​o rg/​1 0.​1 007/​
s00158-​007-​0114-6
Abramowicz W, Jones N (1984) Dynamic axial crushing of square Herskovits J, Dias G, Santos G, Mota Soares C (2000) Shape structural
tubes. Int J Impact Eng 2:179–208 optimization with an interior point nonlinear programming algo-
Afrousheh M, Marzbanrad J, Göhlich D (2019) Topology optimization rithm. Struct Multidisc Optim 20:107–115
of energy absorbers under crashworthiness using modified hybrid Islam SM, Das S, Ghosh S et al (2012) An adaptive differential evolu-
cellular automata (MHCA) algorithm. Struct Multidisc Optim tion algorithm with novel mutation and crossover strategies for
60:1021–1034. https://​doi.​org/​10.​1007/​s00158-​019-​02254-2 global numerical optimization. IEEE Trans Syst Man Cybern B
Albooyeh A, Soleymani P, Taghipoor H (2022) Evaluation of the 42:482–500. https://​doi.​org/​10.​1109/​TSMCB.​2011.​21679​66
mechanical properties of hydroxyapatite-silica aerogel/epoxy Karev A, Harzheim L, Immel R, Erzgräber M (2019) Free sizing opti-
nanocomposites: optimizing by response surface approach. J mization of a front hood using the ESL method: overcoming chal-
Mech Behav Biomed Mater 136:105513. https://d​ oi.o​ rg/1​ 0.1​ 016/j.​ lenges and traps. Struct Multidisc Optim 60:1687–1707. https://​
jmbbm.​2022.​105513 doi.​org/​10.​1007/​s00158-​019-​02285-9
Bahramian N, Khalkhali A (2020) Crashworthiness topology optimiza- Kawamoto A, Bendsøe MP, Sigmund O (2004) Planar articu-
tion of thin-walled square tubes, using modified bidirectional evo- lated mechanism design by graph theoretical enumeration.
lutionary structural optimization approach. Thin-Walled Struct. Struct Multidisc Optim 27:295–299. https://​doi.​org/​10.​1007/​
https://​doi.​org/​10.​1016/j.​tws.​2019.​106524 s00158-​004-​0409-9
Bai YC, Zhou HS, Lei F, Lei HS (2019) An improved numerically- Lee J, Cho M (2018) Efficient design optimization strategy for struc-
stable equivalent static loads (ESLs) algorithm based on energy- tural dynamic systems using a reduced basis method combined
scaling ratio for stiffness topology optimization under crash loads. with an equivalent static load. Struct Multidisc Optim 58:1489–
Struct Multidisc Optim 59:117–130. https://​doi.​org/​10.​1007/​ 1504. https://​doi.​org/​10.​1007/​s00158-​018-​1976-5
s00158-​018-​2054-8 Liu Y, Liu Z, Qin H et al (2018) An efficient structural optimization
Bakhtiarinejad M, Lee S, Joo J (2017) Component allocation and sup- approach for the modular automotive body conceptual design.
porting frame topology optimization using global search algo- Struct Multidisc Optim 58:1275–1289. https://​doi.​org/​10.​1007/​
rithm and morphing mesh. Struct Multidisc Optim 55:297–315. s00158-​018-​1949-8
https://​doi.​org/​10.​1007/​s00158-​016-​1468-4
A crashworthiness optimization method of subway underframe structures based on the differential… Page 19 of 19 62

Lu Z, Li B, Yang C et al (2017) Numerical and experimental study on Schumacher A, Vietor T, Fiebig S, Bletzinger K-U, Maute K
the design strategy of a new collapse zone structure for railway (eds) Advances in structural and multidisciplinary optimization.
vehicles. Int J Crashworthiness 22:488–502. https://​doi.​org/​10.​ Springer, Cham, pp 1601–1614
1080/​13588​265.​2017.​12810​80 Wierzbicki T, Abramowicz W (1983) On the crushing mechanics of
Opara KR, Arabas J (2019) Differential evolution: a survey of theoreti- thin-walled structures. J Appl Mech 50:727–734
cal analyses. Swarm Evol Comput 44:546–558. https://d​ oi.o​ rg/1​ 0.​ Xie S, Zhou H (2014) Impact characteristics of a composite energy
1016/j.​swevo.​2018.​06.​010 absorbing bearing structure for railway vehicles. Composites B
Ortmann C, Schumacher A (2013) Graph and heuristic based topology 67:455–463. https://​doi.​org/​10.​1016/j.​compo​sitesb.​2014.​08.​019
optimization of crash loaded structures. Struct Multidisc Optim Xie S, Liang X, Zhou H, Li J (2016) Crashworthiness optimisation
47:839–854. https://​doi.​org/​10.​1007/​s00158-​012-​0872-7 of the front-end structure of the lead car of a high-speed train.
Padhye N, Bhardawaj P, Deb K (2013) Improving differential evolution Struct Multidisc Optim 53:339–347. https://​doi.​org/​10.​1007/​
through a unified approach. J Glob Optim 55:771–799. https://d​ oi.​ s00158-​015-​1332-y
org/​10.​1007/​s10898-​012-​9897-0 Xing J, Xu P, Yao S et al (2020) Crashworthiness optimisation of a
Passos AG, Luersen MA (2018) Multiobjective optimization of lami- step-like bi-tubular energy absorber for subway vehicles. Int J
nated composite parts with curvilinear fibers using Kriging-based Crashworthiness 25:252–262. https://​doi.​org/​10.​1080/​13588​265.​
approaches. Struct Multidisc Optim 57:1115–1127. https://​doi.​ 2019.​15775​22
org/​10.​1007/​s00158-​017-​1800-7 Xing J, Xu P, Yao S et al (2021) A novel weighted graph representa-
Sauter M, Kress G, Giger M, Ermanni P (2008) Complex-shaped beam tion-based method for structural topology optimization. Adv Eng
element and graph-based optimization of compliant mechanisms. Softw. https://​doi.​org/​10.​1016/j.​adven​gsoft.​2021.​102977
Struct Multidisc Optim 36:429–442. https://​doi.​org/​10.​1007/​ Xu P, Yang C, Peng Y et al (2016) Crash performance and multi-
s00158-​007-​0182-7 objective optimization of a gradual energy-absorbing structure
Shu W, Qiu B, Yang J et al (2020) Loading path optimization of shaft for subway vehicles. Int J Mech Sci 107:1–12. https://​doi.​org/​10.​
clinching forming assembly using finite element simulation and 1016/j.​ijmec​sci.​2016.​01.​001
response surface methodology. Proc Inst Mech Eng C 234:734– Xu P, Yang C, Peng Y et al (2016) Cut-out grooves optimization to
745. https://​doi.​org/​10.​1177/​09544​06219​887770 improve crashworthiness of a gradual energy-absorbing structure
Storn R (1996) On the usage of differential evolution for function opti- for subway vehicles. Mater Des 103:132–143. https://​doi.​org/​10.​
mization. In: Proceedings of North American fuzzy information 1016/j.​matdes.​2016.​04.​059
processing. IEEE, pp 519–523 Xu P, Xing J, Yao S et al (2017) Energy distribution analysis and multi-
Taghipoor H, Eyvazian A (2022) Quasi-static axial crush response objective optimization of a gradual energy-absorbing structure for
and energy absorption of composite wrapped metallic thin-walled subway vehicles. Thin-Walled Struct 115:255–263. https://d​ oi.o​ rg/​
tube. J Braz Soc Mech Sci Eng 44:158. https://​doi.​org/1​ 0.​1007/​ 10.​1016/j.​tws.​2017.​02.​033
s40430-​022-​03449-3 Yang C, Xu P, Yao S et al (2018) Optimization of honeycomb strength
Taghipoor H, Nouri MD (2018) Experimental and numerical investiga- assignment for a composite energy-absorbing structure. Thin-
tion of lattice core sandwich beams under low-velocity bending Walled Struct 127:741–755. https://​doi.​org/​10.​1016/j.​tws.​2018.​
impact. J Sandwich Struct Mater 21:6. https://​doi.​org/​10.​1177/​ 03.​014
10996​36218​761315 Yao S, Zhang P, Xing J et al (2022) Application of a weighted graph
Taghipoor H, Nouri MD (2018) Axial crushing and transverse bending representation method to the crashworthiness optimization of sub-
responses of sandwich structures with lattice core. J Sandwich way collision frame structures. Structures 45:1095–1109. https://​
Struct Mater 22:3. https://​doi.​org/​10.​1177/​10996​36218​761321 doi.​org/​10.​1016/j.​istruc.​2022.​09.​065
Taghipoor H, Sefidi M (2022) Energy absorption of foam-filled corru- Yoshimura M, Shimoyama K, Misaka T, Obayashi S (2019) Optimiza-
gated core sandwich panels under quasi-static loading. Inst Mech tion of passive grooved micromixers based on genetic algorithm
Eng 237:1. https://​doi.​org/​10.​1177/​14644​20722​11104​83 and graph theory. Microfluid Nanofluidics 23:30. https://​doi.​org/​
Taghipoor H, Eyvazian A, Musharavati F et al (2020) Experimental 10.​1007/​s10404-​019-​2201-6
investigation of the three-point bending properties of sandwich Yu X, Cai M, Cao J (2015) A novel mutation differential evolution for
beams with polyurethane foam-filled lattice cores. Structures global optimization. J Intell Fuzzy Syst 28:1047–1060. https://d​ oi.​
28:424–432. https://​doi.​org/​10.​1016/j.​istruc.​2020.​08.​082 org/​10.​3233/​IFS-​141388
Taghipoor H, Ghiaskar A, Shavalipour A (2022) Crashworthiness per- Zhang J, Sanderson AC (2009) JADE: Adaptive differential evolution
formance of thin-walled, square tubes with circular hole discon- with optional external archive. IEEE Trans Evol Comput 13:945–
tinuities under high-speed impact loading. Int J Crashworthiness 958. https://​doi.​org/​10.​1109/​TEVC.​2009.​20146​13
27(6):1622–1634. https://​doi.​org/​10.​1080/​13588​265.​2021.​19811​
25 Publisher's Note Springer Nature remains neutral with regard to
Tyrell DC (2002) US rail equipment crashworthiness standards. Proc jurisdictional claims in published maps and institutional affiliations.
Inst Mech Eng F 216:123–130. https://​doi.​org/​10.​1243/​09544​
09026​00823​62 Springer Nature or its licensor (e.g. a society or other partner) holds
Wang J, Sun Z (2018) The stepwise accuracy-improvement strategy exclusive rights to this article under a publishing agreement with the
based on the Kriging model for structural reliability analysis. author(s) or other rightsholder(s); author self-archiving of the accepted
Struct Multidisc Optim 58:595–612. https://​doi.​org/​10.​1007/​ manuscript version of this article is solely governed by the terms of
s00158-​018-​1911-9 such publishing agreement and applicable law.
Weider K, Schumacher A (2018) A topology optimization scheme
for crash loaded structures using topological derivatives. In: