International Journal of Crashworthiness

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/tcrs20

Product performance prediction and optimisation

based on generative design: a case study of the
collision of automotive front bumper beam

Lei Zhang, Ziqi Li & Junkai Huang

To cite this article: Lei Zhang, Ziqi Li & Junkai Huang (23 Sep 2024): Product performance
prediction and optimisation based on generative design: a case study of the collision
of automotive front bumper beam, International Journal of Crashworthiness, DOI:
10.1080/13588265.2024.2406837

To link to this article: https://doi.org/10.1080/13588265.2024.2406837

Published online: 23 Sep 2024.

Submit your article to this journal

Article views: 15

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=tcrs20
INTERNATIONAL JOURNAL OF CRASHWORTHINESS
https://doi.org/10.1080/13588265.2024.2406837

REPORT

Product performance prediction and optimisation based on generative design: a

case study of the collision of automotive front bumper beam
Lei Zhang, Ziqi Li and Junkai Huang
School of Mechanical Engineering, Hefei University of Technology, Hefei, China

ABSTRACT ARTICLE HISTORY

The automotive front bumper beam is of great significance to safety performance. Generally, there are Received 21 August 2023
two ways to obtain its collision performance: experimental testing and computer simulation. However, Accepted 15 September 2024
both of them consume a lot of time and cannot meet the requirements for quickly predicting its per­
KEYWORDS
formance in the early stage of design. This paper proposed a performance prediction model that
Generative design;
establishes a relationship between the design parameters and performance parameters of the product, parameter mapping; neural
which can provide reference during the design process. Firstly, a three-dimensional model of a certain network; multi-objective
automotive front bumper beam is established, and collision simulation is performed through finite optimisation
element analysis to obtain the performance parameters and analyse the deformation. Then, a large
number of design schemes are generated using the method of generative design (GD), and the same
collision simulation is performed for each scheme to obtain its performance parameters. Finally, the
design parameters and performance parameters of the generated design schemes are formed into a
data set to train the neural network, which realises the performance prediction and is applied to prod­
uct performance optimisation. The result indicates that the established mapping model has high
accuracy and can be used for quickly predicting performance when changing design parameters in
the early design stage.

1. Introduction large number of design schemes based on given design rules

and find the optimal solution from them. GD was first pro­
In recent years, with the rapid development of the economy
posed in the 1970s. Nowadays GD has been widely used in
and industrial level, product requirements are continuously
architectural design and visual arts, also the design of indus­
updating. Product design is gradually showing a trend of
trial products. Bentley and Wakefield [2] described a proto­
rapid, personalised, and knowledge-based development.
type design system that uses genetic algorithms to generate
Enterprises need to redesign their products every time they
new conceptual designs and iteratively optimise them, and
face new customer needs, which not only leads to inefficient
design, but also consumes a lot of manpower, material, and then used them for the design of optical prisms. Gunpinar
financial resources. And many historical designs are often et al. [3] generated the required number of design schemes
not used effectively to provide adequate feedback and guid­ based on GD and simulated annealing algorithms in vehicle
ance for new designs. Simulation verification of a new side profile design to provide data for training resistance
design scheme is necessary when determining its perform­ coefficient prediction models. Lohan et al. [4] conducted
ance. However, with the continuous adjustment of the targeted exploration of heat conduction design space in the
design scheme, it needs to be carried out from CAD model­ engineering field based on GD and topology optimisation
ling to CAE analysis. This process needs to be carried out methods combined with genetic algorithms. Bright et al. [5]
repeatedly in the design stage of the product, which means used a GD approach to design the quadcopter frame, result­
that it inevitably consumes a lot of time and reduces the ing in a higher fracture resistance and smaller displacement
efficiency of solving practical problems. Therefore, making compared to the original design. Oh et al. [6] studied the
full use of existing designs and historical cases to study the necessity of adopting deep learning in GD, established a
mapping relationship between product design parameters framework that can generate many design models from lim­
and performance parameters is of great significance for ited design cases, and optimised aesthetic and mechanical
improving design efficiency and shortening the product properties. Jang et al. [7] proposed a GD method based on
development cycle. reinforcement learning and topology optimisation, which
A large number of design cases are the basis for studying seeks the optimal combination of design parameters based
the mapping relationship between parameters. GD can be on a given reference design and can generate a large num­
used as a solution to overcome the limitation of data short­ ber of different designs in a short time. Obviously, GD can
age [1]. As a novel design approach, it could generate a provide a variety of design schemes according to

CONTACT Lei Zhang [email protected]

� 2024 Informa UK Limited, trading as Taylor & Francis Group
2 L. ZHANG ET AL.

requirements, which can effectively expand the design case relationships should finally be used to improve the
data and provide reference for the design. products.
The front bumper beam of a car is a protective device, In this work, the mapping relationship between the
usually hidden behind the car’s front cover. It deforms and design parameters and performance parameters of the car
absorbs energy during collisions to protect other important front bumper beam is established and applied to its per­
components of the car, as well as the driver and passengers. formance prediction and multi-objective optimisation.
Therefore, the main evaluation criteria for the front bumper Firstly, a three-dimensional model of the front bumper
beam are its performance during a collision, including mass, beam of a car is established. According to the relevant
intrusion, peak load, contact time, total energy absorption, standards, the collision process of the bumper beam at a
and maximum stress, and the main optimisation targets are speed of 4 km/h is simulated, and the relevant performance
also these attributes. In relevant research, Yucheng Liu et al. parameters are obtained through finite element analysis.
[8] used ANSYS to study the performance of the bumper Then, a large number of new design schemes are generated
system under low-speed collision conditions, observing the through GD of the bumper beam, and the performance
displacement curve and the impact force curve. The simula­ parameters corresponding to each scheme are obtained by
tion results are compared with the experimental results, analysing these design schemes in the same simulation
which provided a theoretical basis for the numerical analysis environment. In the step of machine learning, the data of
method of bumper collision. Takashi Hasegawa et al. [9] the GD model is used to train a neural network, establishing
researched the collision characteristics of cars and con­ the relationship between the design parameters and per­
ducted offset collision tests on SUVs and small passenger formance parameters of the bumper beam, and realising
cars using the bending beam model to verify the energy performance prediction. Finally, the obtained performance
absorption effect of bumper beams in the collision process. prediction model is applied to product optimisation, using
Duan [10] proposed a new method for structural optimisa­ the non-dominated sorting genetic algorithm-II (NSGA-II)
tion and reliability analysis, which simultaneously considers algorithm to achieve better collision performance of the
longitudinal and angular impacts, improves safety, and bumper beam with a smaller mass.
achieves weight reduction in design. Sun et al. [11] pro­
posed a new optimisation method based on multi-objective
2. Methodology
optimisation for the bumper beam, which takes multiple
objective scenarios into account, allowing for the optimisa­ To establish a mapping relationship between product design
tion of intrusion and energy absorption under different parameters and performance parameters, so as to predict
impact loads. Godara et al. [12] established a collision product performance and adapt to new requirements
model for different cross-sections of the bumper beam and quickly, this paper proposes a mapping method for product
studied the performance of these bumper beams under dif­ design parameters and performance parameters based on
ferent collision conditions. GD and BP neural networks. The method proposed in this
Many researches have been conducted to study the map­ article is shown in Figure 1. The process of the method can
ping relationship between product design parameters and be divided into three steps. First, the bumper beam is mod­
performance, involving different product design processes. elled and its working load and design attributes are defined.
In the optimisation design of compressor blades, Adjei et al. A large number of design schemes are obtained by GD.
[13] established a mapping relationship of blades from per­ Then the performance of these design schemes is obtained
formance space to parameter space based on the parametric by simulation analysis. At this point, the design parameters
free-form deformation method of B-splines. Blades are and performance of the different design schemes have been
mapped from object space to parameter space by transform­ clarified. Finally, with the support of these data, the map­
ation operation. To optimise the aerodynamic shape of tur­ ping model between design parameters and performance is
bines and improve their performance, Li et al. [14] established by neural network. The feasibility of the pro­
developed a parameterised free-form deformation method posed method is verified by taking the front bumper beam
with the function of mapping design parameters. Fan et al. of an automobile as an example.
[15] established mapping models between various engineer­
ing parameters and performance of fan blades, studied the
2.1. Generative design
effects of single parameters and coupling of two parameters,
and improved the efficiency and total pressure ratio of For GD, It takes space, materials, forces, fixed supports, and
ducted fans. other constraints as design constraints, and makes com­
From the above research, it can be found that GD has puters automatically iterate to obtain the optimal solution in
been applied to mechanical products, but most of them are the design space. This design method is often performance-
used for the creation of a single result, without fully utilising oriented, and the designed structure usually presents den­
the large number of design schemes generated by GD. In dritic and complex curved features, with the advantages of
terms of parameter mapping, some scholars have used vari­ good stiffness and even stress distribution. GD modifies an
ous methods to establish mapping relationships between existing solution or creates a new solution based on con­
product design parameters and customer needs, structural straints, such as performance and size of the product to
performance, and even green performance. However, these meet the constraints.
 INTERNATIONAL JOURNAL OF CRASHWORTHINESS 3

Figure 1. Product performance prediction and optimisation method based on GD and BP neural network.

Optimisation problems can generally be described as:

( � �
minf ðxÞ ¼ f1 ðxÞ, f2 ðxÞ, :::, fi ðxÞ
(1)
s:t:gj ðxÞ � 0, j ¼ 1, 2, :::, n

Where f(x) is the optimisation objective and g(x) is the

constraint condition. The selection of variables often
involves considering the parts that significantly impact the
results. Constraints are generally external conditions that
need to be considered during optimisation, such as fixed
support, contact, initial velocity, etc. Some constraints may
conflict with the optimisation objectives of the product,
which requires comprehensive consideration when designing
or optimising.
The numerical simulation problems of GD can generally
be described as follows [16,17]:
Y ¼ HðMjZÞ (2)
Where H is the operator that simulates data Z using Figure 2. Flow chart of GD.
simulation model M. M should be flexible enough to evolve
by changing some of its content. M can be described as:
M ¼ hS, E, fa1:jAj gi (3)
2.2. Mapping relationship establishment
Where S is the functional attribute, fa1:jAj g is the param­
eter set, and E is the relationship between them. The math­ Back Propagation (BP) neural network is one of the most
ematical model of GD can be expressed as: widely used artificial neural networks, which has the ability
to classify arbitrary complex patterns and excellent multi-
PRmax ðM� Þ ¼ max fR ðMjI þ , Tgen � Tg Þ, M ¼ fMp g (4) dimensional function mapping capabilities [19]. Therefore,
M
it was used to establish the mapping relationship between
Where fR is a fitness function, representing the quality of product design parameters and performance parameters.
mathematical generation, PRmax represents the maximum The network proposed in this paper consists of three parts:
value of the fitness function, Iþ Is the actual computing input layer, hidden layer, and output layer. And there are
resource, Tgen is the requirement for model optimisation two processes in network training: forward propagation pro­
with its critical value Tg, and Mp is a possible model struc­ cess and back propagation process.
ture space. (1) Forward propagation process. The main task of the
In essence, the GD method is a dynamic modelling forward process is to calculate the predicted value of the
method, which is designed through a large number of calcu­ network, that is, the output value of the output layer neu­
lations and iterative processes. Therefore, the GD method is rons. For input layers, the output value is equal to the input
suitable for design patterns that support high computational value, so there is no need to calculate. For hidden layers
costs. The main process of GD is shown in Figure 2 [18]. and output layers, the input of each node can be expressed
4 L. ZHANG ET AL.

as a weighted sum of the output values of all nodes in the To maintain diversity among individuals in the popula­
previous layer, as shown below: tion, the concept of crowding degree is introduced into the
ni−1
X NSGA-II algorithm, which is the density of individuals
xij ¼ yði−1Þh � xði−1Þhj þ b (5) around a given individual in the population. For the bound­
h¼1 ary solution, suppose its congestion degree as infinity. For
Where xij represents the input value of the jth node of the intermediate solution, its congestion degree can be cal­
the ith layer, ni-1 represents the total number of nodes culated using the following equation:
of the previous layer, y(i-1)h represents the hth output value Iiþ1, k − Ii−1, k
Idm ¼ (8)
of the i-1 layer, x(i-1)hj represents the weight value when the fkmax − fkmin
output value propagates to the jth node of the ith layer for
Where Idm is the degree of congestion, Iiþ1,k and Ii-1,k are
input, and b is the offset value. The output of a node is the
the values of the i þ 1th and i-1th points on the objective
result of the input being activated by an activation function.
function k, respectively, fkmax and fkmin are the maximum
Common activation functions include the sigmoid activation
and minimum values of the function k. When selecting
function, ReLU activation function, and so on.
individuals, priority should be given to individuals with
(2) Back propagation process. After completing the for­
high crowding levels, which can avoid the situation where
ward propagation process, it is necessary to adjust the
the selected individuals are concentrated in a small portion
weight and bias values of each node based on the output
of the space.
values and expected values, which is the training process of
In optimisation problems, the objective function is often
the network. First, it is necessary to calculate the error of
a specific expression. But in reality, many mapping relation­
each output layer node as ej and calculate the mean square
ships cannot be expressed using equations [21]. Therefore,
error objective function E, as shown below:
in this paper, the neural network is used as the surrogate
ej ¼ dj − yj (6) model to predict the results and use it for optimisation.
1X
E¼ ðej Þ2 (7)
2 3. GD of bumper beam
Where dj is the expected value of the jth node, and yj is
the actual output value. The input value of each node is To establish the link between design parameters and pro­
only related to the output values of all nodes connected to gram performance, a certain amount of schemes are
it in the previous layer, and each weight value is only required as data support. And GD can fulfill this need by
responsible for the data propagation between the corre­ exploring design possibilities to generate a large number of
sponding two nodes. Therefore, when updating the weight potential schemes. This section describes how to obtain a
value, partial derivatives are performed for it: large number of potential schemes through a small number
of initial models with the help of GD to provide a data basis
for the establishment of parameter-performance mapping
2.3. Multi-objective optimisation relationship.
In product optimisation, the improvement of one objective
could weaken the others. Therefore, when a product has 3.1. Bumper beam model
multiple conflicting optimisation objectives, it is necessary
The bumper beam is a device used to reduce the absorption
to comprehensively consider various indicators to obtain an
of collision energy when a vehicle is impacted. It is com­
overall optimal solution [20]. NSGA-II is an improved ver­
posed of a main beam and crash boxes. It can effectively
sion of the non-dominated genetic sorting algorithm(NSGA)
absorb collision energy when a vehicle encounters a low-
and is currently widely used as a multi-objective optimisa­
speed collision, thereby protecting the automobile engine
tion algorithm.
and passengers to a certain extent [22]. To facilitate
research, a simplified vehicle front bumper beam model is
built as shown in Figure 3.
The total length of the bumper beam is 1155 mm [11], and
the top is an arc with a diameter of 6480 mm, with a wall
thickness of 3.8 mm. The cross-section of the beam is rect­
angular, with a width of 102 mm and a height of 45 mm. The
width of the crash box is 72 mm, and the spacing between the
two crash boxes is 785 mm. The bumper beam is made of alu­
minium alloy, and its performance is shown in Table 1. The
Figure 3. Simplified model of automobile front bumper beam. total mass of this simplified model is 4.9437 kg.

Table 1. Mechanical properties of bumper beam materials.

Density (kg�m−3) Young’s modulus (Pa) Poisson’s ratio Bulk modulus (Pa) Shear modulus (Pa) Yield strength (Pa) Tangent modulus (Pa)
2770 7.1 � 1010 0.33 6.96 � 1010 2.67 � 1010 2.8 � 108 5 � 108
 INTERNATIONAL JOURNAL OF CRASHWORTHINESS 5

3.2. The GD setting of the bumper beam value converges, it indicates that the exploration of the
design space is over and the GD result is obtained. In this
In this article, the GD module of Creo 7.0 was used to opti­
paper, the design objective of bumper beam is to minimise
mise the target product. In this article, the GD module of
the strain. With the increase of the iteration, the model will
Creo 7.0 was used to optimise the structure of bumper
be gradually modified until maximum strain value of the
beam. Before starting the GD, constraints, loads, and design
buffer beam finally converges. Figure 5 shows the results for
regions need to be set up to guide the algorithm in adding
different iterations. According to different conditions set,
or removing material to the structure of the bumper beam.
such as volume fraction, quality constraints, material con­
The main design region is the main beam of the front
straints, manufacturing constraints, and so on, different
bumper beam. Therefore, the main beam is set as the design
design results can ultimately be obtained. Figure 6 shows
space, and the crash box and the bracket used for installa­
the GD results under different volume fraction constraints.
tion and fixation remain unchanged. The main performance As can be seen from Figures 5 and 6, the design results
of the bumper beam is its collision performance. For the obtained under different iterations and volume fraction
load setting, since GD module does not support dynamic ratios are different. In addition, changes in the initial model
collision load, the average collision force is used instead. can affect the results of GD. In this paper, the height, width,
This paper focuses on the optimisation design of the and thickness of the cross-section of the beam are selected
bumper beam in the case of low-speed collision, and the as design parameters. By changing the weight and height of
methods and tests involved are referred to E/ECE/324 and the hollow rectangular section, more initial models are built
GB 17354-1998 [23,24]. According to the regulations, the to obtain a large number of generation schemes by GD. The
speed of simulated collision at low speed is 4 km/h [25–27]. range of each parameter value is shown in Table 2.
Combined with the body weight m of about 1500 kg, the A large number of different design schemes are obtained
collision time t is about 0.035 s, according to the acceler­ by GD. After setting up the initial model, the goal of the
ation formula as shown in Equation (9), the average colli­ design is to reduce the weight of the structure, so GD’s
sion force F can be estimated to be about 50 kN. The stress algorithm generates a series of new structures based on the
point of the bumper beam is the centre of the front side of set conditions, such as loads and constraints. Compared to
the beam at the beginning of the collision. Therefore, 50 kN the initial models, the new structures differ mainly in the
force is applied to the centre of the upper surface of the uneven reduction of the material in the inner cavity of the
main beam to simulate the collision force. Fixed constraints beams forming an elliptical cavity.
are applied at the bottom connection of the beam. The con­ When the section width is large or the set volume frac­
straint conditions of the model are shown in Figure 4. tion is low, the central void presents a flat and long shape,
mv where the average wall thickness is small. When the section
F¼ (9)
t width is small or the set volume fraction is large, the central
void presents a narrower shape. Figure 7 shows some of the
generated results of the bumper beam.
3.3. GD results of the bumper beam
By clarifying the design objectives and constraints, GD’s 4. Bumper beam collision simulation
algorithm will be iteratively calculated. When the objective
After acquiring a large number of design schemes, the spe­
cific performance of each scheme needs to be clarified. The

Table 2. The range of parameter values.

Design parameters Initial value Value range
Wall thickness 3.8 0.8–5.8
Section width 102 90–110
Section height 45 45–55
Figure 4. Constraints of the bumper beam optimisation model.

Figure 5. Results for different iterations.

Figure 6. Results under different volume fraction restrictions.

6 L. ZHANG ET AL.

Figure 7. Some of the results of GD for bumper beam.

Figure 9. Cloud chart of bumper beam collision displacement.

a collision on the interior of the beam, the lower the

likelihood of damage to the beam, and the better the
collision safety performance.
Figure 8. Collider model.
To simulate the collision of a car at low speed, according
to the standard requirements, the collider shown in Figure 8
main role of the bumper beam is to protect passengers and is established, with a mass equivalent to the curb weight of
interior components in the event of low-speed collisions. the entire vehicle, which is 1500 kg. The collider is set to a
Therefore, according to the requirements of E/ECE/324 and rigid object during the collision.
GB 17354-1998, the crash performance parameters of the The beam model established in Section 3.1 and collider
front bumper beam for vehicles are set as follows: model are imported into the ANSYS Workbench, and LS-
DYNA module is used for the collision simulation. Set the
average size to 5 mm and the grid type to the quadrilateral,
1. Maximum intrusion amount R. The maximum distance resulting in 11165 nodes and 27400 grid elements. Apply
the collision object enters the beam. The smaller the flat fixed support to the bottom of the bumper beam and
amount of intrusion, the smaller the damage to the apply an initial speed of 4.0 km/h to the collider. The dis­
device behind the beam. The deformation of the beam placement cloud chart of the bumper beam is shown in
should be within a safe distance as much as possible. Figure 9. Figure 10 shows the internal energy curve of the
2. Total energy absorption E. The total energy absorbed beam during the collision.
by the bumper beam in a low speed collision. The As shown in Figure 9, the bumper beam deforms to dif­
beam fully absorbs collision impact energy during the ferent degrees under the impact of the collision object, and
collision process, effectively reducing the collision the maximum deformation is 13.898 mm at the collision
energy transmitted to the rear structure, and protecting contact point in the centre of the beam. As can be seen
the internal people and key components of the vehicle. from Figure 10, the internal energy of the beam continues
3. Maximum stress F. Internal stress of the beam after to increase with the collision process, reaching a maximum
impact. The smaller the stress, the smaller the impact of of 741.48 J. After the separation of the beam and the
 INTERNATIONAL JOURNAL OF CRASHWORTHINESS 7

collision object, due to the elastic recovery of the bumper thickness is taken as the average value. To reduce computa­
beam, its internal energy decreases and subsequently tends tional complexity and obtain the mapping relationship
to stabilise. The maximum internal stress is 282.41 MPa. between parameters faster and more accurately, it is neces­
sary to pre-process the initial dataset.
Normalisation of data is a common method of prepro­
5. Mapping the relationship between design cessing, which involves dimensionless processing of data
parameters and performance parameters and mapping them all to the [0,1] interval. It can solve the
In the previous two sections, the design parameters and the problem of incompatibility between different data while
corresponding performances of the different design schemes accelerating the speed of computation. The commonly used
normalisation methods include the Z-score method, max-
with explicit design parameters are obtained through GD
min normalisation method, crop normalisation, standard
and finite element analysis. With the support of these data,
deviation normalisation, and so on. After eliminating unreli­
this section fits the relationship between design parameters
able data, this paper selects the max-min normalisation
and performance by building a neural network for fast pre­
method. The basic principle of the max-min normalisation
diction of the performance of design solutions.
method can generally be described as the following
equation:
5.1. Data set for mapping x − Min
xn ¼ (10)
In the previous part of the paper, collision analysis was con­ Max − Min
ducted on the automotive front bumper beam to obtain Where x is the initial data, xn is the data after standard­
various performance parameters during the collision. At the isation, Max is the maximum value in the initial dataset,
same time, a GD has been carried out for the bumper Min is the minimum value in the initial dataset. This
beam, and a large number of design schemes were automat­ method obtains the difference between the maximum and
ically generated by GD. The simulation analysis of each minimum values in the initial data, and based on this, estab­
design scheme is carried out under the same conditions, lishes a cardinality that maps the initial data to the interval
and the design parameters (wall thickness, section width, of [0,1].
After data filtering and normalisation, 294 pieces of data
and section height) and corresponding performance parame­
were finally obtained from the initial dataset and used as a
ters (mass, maximum intrusion amount, total energy
training dataset.
absorption, and maximum stress) of each design scheme are
obtained. These data make up the data set, where the wall
5.2. Training process analysis
The BP neural network generalises the trained rules by add­
ing hidden layers to the network model [28]. In this paper,
deep learning framework in Python is adopted to build a
neural network model, with three design parameters of the
front bumper beam as inputs and four performance parame­
ters as outputs. The neural network structure is set as 5
layers with 3-20-50-50-4 nodes per layer. There are 120
nodes in the three hidden layers, as shown in Figure 11.
Data passing through each node will be converted through
the activation function. The commonly used activation func­
Figure 10. Internal energy curve during collision. tions in BP neural networks include the sigmoid function,

Figure 11. Performance prediction network model of bumper crossbeam.

8 L. ZHANG ET AL.

tanh function, ReLU function, and Leaky ReLU function. training process, the Adarad optimisation algorithm is
The ReLU function is selected as the activation function in selected. The learning rate can be adaptively changed accord­
this paper. The ReLU function is a piecewise linear function, ing to the gradient of the parameter, thereby preventing
which can be described by the following equation: shocks when the gradient is large, and accelerating the change
when the gradient is small. The initial learning rate is set as
f ðxÞ ¼ maxð0, xÞ (11) 0.03, the mean square error (MSE) loss is used for prediction,
and 200 training times are performed. Figure 12 shows the
It means changing all negative values to 0 while maintain­
loss variation curve during the model’s training process. It
ing positive values, which allows neurons in the neural net­ can be seen from the figure that the overall training process of
work to have sparse activation, achieving more efficient the model is good, with the loss eventually tending to 0.0021,
gradient descent and lower computing costs. During the and dropping to a lower value after about 75 training times.

5.3. Verification and analysis of prediction results

To verify the accuracy of the trained model, a portion of the
data is required as a test set to test the trained network model,
and this portion of the test set is required not to participate in
the training process of the network. Latin hypercube sampling
is a layered Monte Carlo sampling method that is suitable for
uniform sampling in multi-dimensional spaces, enabling sample
points to uniformly fill the entire sampling space and avoiding
the situation where sample points are concentrated in a certain
area during random sampling. Due to the uniform distribution
of sampling points, Latin hypercube sampling methods are
widely used in various studies. Fifteen sample points were
obtained through Latin hypercube sampling, and these sample
Figure 12. The curve of mean square error loss. points were used as test sets to verify the accuracy of the previ­
ously trained network. The design parameters of the test set
Table 3. Data of test set design parameters. data are input into the trained network model, and the pre­
Number Wall thickness t Section width b Section height h dicted results are compared with the simulated performance
1 0.836735 0.093264 0.74 parameters. The test set data design parameters are shown in
2 0.55102 0.943005 0.99 Table 3. The comparison between the predicted results and the
3 0.673469 0 0.01 simulated performance parameters is shown in Table 4.
4 0.244898 0.80829 0.93
5 0.714286 0.569948 0.82 By analysing the mass, maximum intrusion amount, total
6 0.346939 0.046632 0.39 energy absorption, and maximum stress in the table, the
7 0.653061 0.854922 0.22 variation curves shown in Figure 13 can be obtained.
8 0.77551 0.430052 0.67
9 0.918367 0.455959 0.77 It can be seen from Figure 13, the predicted values of per­
10 1 1 0.54 formance parameters fit well with simulated values. There is a
11 0.510204 0.492228 0.64
12 0.897959 0.284974 0.43
certain prediction error only in total energy absorption, which
13 0.265306 0.367876 0.57 may be due to the low complexity of the analysis model.
14 0.122449 0.336788 0.06 Although there is some absolute error between the prediction
15 0.591837 0.766839 0.15
of total energy absorption and the simulated value, the relative

Table 4. Comparison of prediction results and simulation results in test set.

Mass Maximum intrusion amount Total energy absorption Maximum stress
Number Simulated value Predicted value Simulated value Predicted value Simulated value Predicted value Simulated value Predicted value
1 5.7239 5.7741 11.8240 12.2952 738.8800 737.8658 305.6200 276.5751
2 5.0815 5.0152 12.4380 12.8430 738.8200 738.3816 242.0400 257.2581
3 4.8372 4.8898 14.1600 14.0422 738.5000 736.0464 286.7700 293.5049
4 3.5805 3.6656 14.9750 15.4585 734.3900 737.6479 280.4800 297.9818
5 5.5569 5.5449 12.1040 12.4859 741.7400 740.4837 254.9900 259.2628
6 3.6764 3.7175 15.1320 15.4237 735.1400 736.4966 326.8700 305.8255
7 5.3051 5.3052 13.6070 13.7763 738.8100 738.9364 279.4500 266.6538
8 5.6825 5.7289 12.0310 12.4822 736.9400 739.9930 276.8200 264.8921
9 6.3143 6.2715 11.3300 11.7148 739.6900 739.4120 268.6500 260.8230
10 6.9850 6.8478 11.4800 11.4058 737.0000 735.3019 278.7100 254.4100
11 4.5980 4.6845 13.5150 14.0665 742.2000 741.5054 274.8500 273.6441
12 5.9959 6.0965 11.7850 12.2717 740.5000 738.5655 280.6800 272.3715
13 3.4782 3.4361 15.7840 16.2854 739.0100 738.0162 345.8300 315.7555
14 2.7982 2.6671 18.2990 18.5797 727.4400 729.3384 449.6800 422.5328
15 4.9777 4.9657 13.9200 14.3505 742.0200 740.1901 269.6700 270.5388
 INTERNATIONAL JOURNAL OF CRASHWORTHINESS 9

Figure 13. The curves of parameters of simulation and prediction.

Table 5. Average relative error of each parameter. prediction results during the optimisation process.
Mass Maximum intrusion amount Total energy absorption Maximum stress Therefore, this paper uses the neural network model estab­
1.34% 2.76% 0.20% 4.76% lished in the previous part as a surrogate model for multi-
objective optimisation of the bumper beam.
error is still relatively small. The relative error is calculated
according to the following equation:
8 � � 6.1. Problem description
> �yi − ^y �
> i
< RE ¼
>
yi
The automotive front bumper beam should absorb as much
(12) energy as possible in a collision, so E should be as large as pos­
>
> 1 X n
sible. To comply with the trend of overall lightweight design of
>
: ARE ¼ RE
n i¼1 the automobile, the beam should reduce the mass m as much as
possible. Therefore, energy absorption rate Em is defined as the
Where RE represents the relative error of each sample ratio of total energy absorption E to mass m. In the optimisation
point, and ARE represents the average relative error of each objective, the energy absorption rate should be as large as pos­
performance parameter. According to Equation (12), the aver­ sible. In addition, to protect important components and passen­
age relative error of each parameter is obtained as shown in gers behind the bumper, the intrusion amount R and the
Table 5. From the table, it can be seen that the average relative maximum stress F should be as small as possible.
error of each parameter is small and within the allowable The wall thickness t, cross section width b, and cross-sec­
range. This can prove the feasibility of the prediction method, tion height h of the bumper beam are taken as design varia­
which can be used as a reference in the early design stage and bles, and their range has been defined in Table 2. According
improve the efficiency of product optimisation. to the characteristics of aluminium alloy bumper beams, the
multi-objective optimisation problem of bumper beams
under frontal collision can be expressed as:
6. Optimisation of performance parameters of
bumper beam 8
< 0:8 � t � 5:8
>
Simulation technology has developed relatively mature, but min½Rðt, b, hÞ, − Em ðt, b, hÞ, Fðt, b, hÞ� s:t: 90 � b � 110 (13)
highly nonlinear collision simulation analysis requires high >
:
45 � h � 55
computational costs and is difficult to provide real-time
10 L. ZHANG ET AL.

Figure 14. Distribution of Pareto frontier solutions for 50, 100, and 150 iterations.

Table 6. Comparison of design parameters before and after optimisation. The comparison of performance parameters before and after
Wall thickness Section width Section height optimisation is shown in Table 7.
t (mm) b (mm) h (mm) According to the optimisation results, the mass of the
Before optimisation 3.8 102 45 optimised design scheme is significantly reduced, while the
After optimisation 2.9 99.4 55 total absorption energy is slightly increased, resulting in a
significant increase in specific absorption energy of
Table 7. Comparison of performance parameters before and after 20.3136 J�kg−1, accounting for 13.54% of the simulated
optimisation.
value. The maximum intrusion amount decreased by 1.32%.
Maximum Total Energy
intrusion energy absorption Maximum The maximum stress decreased by 0.61%. Overall, the per­
Mass amount absorption rate stress formance parameters of the optimised bumper beam are
m (kg) R (mm) E (J) Em (J�kg−1) F (MPa) better than those before optimisation, and the effect on
Before 4.9437 13.898 741.48 149.9848 282.41 lightweight optimisation is significant. Therefore, the overall
optimisation
After 4.3627 13.7148 742.9609 170.2984 280.6968 performance is superior to the original design scheme.
optimisation
Variation −11.75% −1.32% 0.20% 13.54% −0.61%
7. Conclusion
6.2. Optimisation results In this paper, a neural network model is built to study the
NSGA-II is currently the most popular multi-objective genetic relationship between design parameters and performance
algorithm. It proposes a fast non dominated sorting algorithm parameters. GD and simulation analysis are used to obtain
with elite strategy, which reduces the computational complex­ the data set required for neural network training. The test
ity on the one hand, and retains the best individuals on the shows that the neural network can predict the performance
other hand, with the advantages of fast running speed and of the design scheme well. The results verify the feasibility
good convergence of the solution set. In this paper, NSGA-II of establishing mapping relationships to predict product
is used for optimisation, with a population size of 50. When performance parameters. Based on the obtained mapping
the number of iterations is 50, 100, and 150, the distribution relationship, a surrogate model is established for the multi-
results of the solution set in the three-dimensional space are objective optimisation of the bumper beam, which reduces
shown in Figure 14. The Pareto frontier solution in the scatter the mass and improves the comprehensive performance of
diagram shows that the Pareto curve becomes smoother and the product while ensuring its collision performance.
smoother as the number of iterations increases, especially There are still some problems and deficiencies in the
when it comes to 150. This proves the optimisation and con­ research that need to be improved. First, the average wall
vergence ability of NSGA-II for multi-objective models [29]. thickness in the performance parameters is the average wall
After 150 iterations, a series of Pareto frontier solutions are thickness of the front bumper beam, which cannot reflect
obtained. The Pareto frontier solution is a group of solutions the real shape of the section well. This will cause a certain
that do not dominate each other. An excellent solution is deviation between the training set data and the actual per­
chosen with the wall thickness t ¼ 2.9 mm, the cross-section formance. The impact of the crash boxes on collision is not
width b ¼ 99.4 mm, and the cross-section height h ¼ 55 mm. considered, but in reality, it is the joint action of the crash
The comparison of design parameters before and after opti­ box and the bumper beam that affects the collision perform­
misation is shown in Table 6. ance. Besides, there are few design parameters involved in
the research, and the anti-collision beam model is simpli­
fied. There may be other parameters that affect bumper
6.3. Data analysis
beam performance but are not taken into account by us. In
The neural network obtained above is used to predict the future work, more complex and more complete models will
optimised performance parameters of the bumper beam. be built in research.
 INTERNATIONAL JOURNAL OF CRASHWORTHINESS 11

Disclosure statement [14] Li L, Jiao J, Sun S, et al. Aerodynamic shape optimization of a

single turbine stage based on parameterized free-form deform­
No potential conflict of interest was reported by the author(s). ation with mapping design parameters. Energy. 2019;169:444–
455. doi: 10.1016/j.energy.2018.12.031.
[15] Fan C, Adjei RA, Wu Y, et al. Parametric study on the aero­
Funding dynamic performance of a ducted-fan rotor using free-form
method. Aerosp Sci Technol. 2020;101:105842. doi: 10.1016/j.
The research contained in this paper is supported by the National ast.2020.105842.
Natural Science Foundation of China (51875156). [16] Kalyuzhnaya AV, Nikitin NO, Hvatov A, et al. Towards genera­
tive design of computationally efficient mathematical models
with evolutionary learning. Entropy. 2020;23(1):28. doi: 10.
References 3390/e23010028.
[17] Wang H, Du W, Zhao Y, et al. Joints for treelike column struc­
0[1] Soyoung Y, Sunghee L, Seongsin K, et al. Integrating deep
tures based on generative design and additive manufacturing. J
learning into CAD/CAE system: generative design and evalu­
Constr Steel Res. 2021;184(6):106794. doi: 10.1016/j.jcsr.2021.
ation of 3D conceptual wheel. Struct Multidiscip Optim. 2021;
106794.
64(4):2725–2747. doi: 10.1007/s00158-021-02953-9.
[18] Jia C, Ming XT. Towards generative systems for supporting
0[2] Bentley PJ, Wakefield JP. Conceptual evolutionary design by a
product design. Int J Des Eng. 2017;7(1):1–16. doi: 10.1504/
genetic algorithm. Eng Des Autom. 1997;3(2):119–131.
IJDE.2017.085639.
0[3] Gunpinar E, Coskun UC, Ozsipahi M, et al. A generative
[19] Zhang Y, Wang H, Wang Y, et al. A novel optimization algo­
design and drag coefficient prediction system for sedan car side
rithm for BP neural network based on RS-MEA. 2nd
silhouettes based on computational fluid dynamics. Comput
International Conference on Image, Vision and Computing;
Aided Des. 2019;111:65–79. doi: 10.1016/j.cad.2019.02.003. 2017 June 2–4; Chengdu, China; 2017. p. 1073–1078. doi: 10.
0[4] Lohan DJ, Dede EM, Allison JT. Topology optimization for 1109/ICIVC.2017.7984719.
heat conduction using generative design algorithm. Struct [20] Liu H, Qi G, Zhang T, et al. Research on multi-disciplinary
Multidisc Optim. 2017;55(3):1063–1077. doi: 10.1007/s00158- process modeling of conceptual design for complex product. J
016-1563-6. Zhejiang Univ. 2009;43(3):517–522. doi: 10.3785/j.issn.1008-
0[5] Bright J, Suryaprakash R, Akash S, et al. Optimization of quad­ 973X.2009.03.023.
copter frame using generative design and comparison with DJI [21] Ya€el B, Mikael B, Michela A, et al. Model-based multi-param­
F450 drone frame. IOP Conf Ser Mater Sci Eng. 2021;1012(1): eter mapping. Med Image Anal. 2021;73:102149. doi: 10.1016/j.
012019. doi: 10.1088/1757-899X/1012/1/012019. media.2021.102149.
0[6] Oh S, Jung Y, Kim S, et al. Deep generative design: integration [22] Du B, Li QC, Zheng CQ, et al. Application of lightweight struc­
of topology optimization and generative models. J Mech Des. ture in automobile bumper beam: a review. Materials. 2023;
2019;141(11):111405. doi: 10.1115/1.4044229. 16(3):967. doi: 10.1016/10.3390/ma16030967.
0[7] Jang S, Yoo S, Kang N. Generative design by reinforcement [23] United Nations. E/ECE/324, Concerning the adoption of uniform
learning: enhancing the diversity of topology optimization conditions of approval and reciprocal recognition of approval for
design. Comput Aided Des. 2022;146:103225. doi: 10.1016/j.cad. motor vehicle equipment and parts. Available from: https://unece.
2022.103225. org/fileadmin/DAM/trans/main/wp29/wp29regs/2009/r042e.pdf
0[8] Liu YMichael, L, D. Experimental analysis and computer simu­ [24] Ministry of Industry and Information Technology, People’s
lation of automotive bumper system under impact condition. Republic of China. GB 17354-1998. Front and rear protective
Int J Comput Methods Eng Sci Mech. 2008;9(1):51–59. doi: 10. devices for passenger cars. Beijing: Standards Press of China;
1080/15502280701759218. 1998.
0[9] Takashi H. Study of vehicle-to-vehicle collision performance [25] Dai C, Yu P, Yin M, et al. Optimization of the crashworthiness
based on balance of front end strength. SAE Trans. 2007;116: design of carbon fiber-reinforced polymer bumper beam. Int J
1171–1179. doi: 10.4271/2007-01-1175. Crashworthiness. 2024;29(1):163–177. doi: 10.1080/13588265.
[10] Duan SY, Han X, Liu GR. Structural optimization and reliability 2023.2230628.
analysis of automotive composite bumper against low-velocity [26] Wang T, Li Y. Design and analysis of automotive carbon
longitudinal and corner pendulum impact. Int J Comput fiber composite bumper beam based on finite element analysis.
Methods. 2018;16(08):1841003. doi: 10.1142/S0219876218410037. Adv Mech Eng. 2015;7(6):168781401558956. doi: 10.1177/
[11] Sun G, Wang X, Fang J, et al. Parallelized optimization design 1687814015589561.
of bumper systems under multiple low-speed impact loads. [27] Davoodi M, Sapuan S, Aidy A, et al. Development process of
Thin Walled Struct. 2021;167:108197. doi: 10.1016/j.tws.2021. new bumper beam for passenger car: a review. Mater Des.
108197. 2012;40:304–313. doi: 10.1016/j.matdes.2012.03.060.
[12] Godara SS, Shiv NN. Analysis of frontal bumper beam of auto­ [28] Herlan S, Muhammad, Z, Sawaluddi. Analysis of accuracy and
mobile vehicle by using carbon fiber composite material. Mater epoch on back-propagation BFGS Quasi-Newton. J Phys Conf
Today Proc. 2020;26:2601–2607. doi: 10.1016/j.matpr.2020.02. Ser. 2017;930(1):012006. doi: 10.1088/1742-6596/930/1/012006.
550. [29] Zhang L, Zhang B, Bao H, et al. Optimization of cutting
[13] Adjei RA, Wang W, Liu Y. Aerodynamic design optimization parameters for minimizing environmental impact: considering
of an axial flow compressor stator using parameterized free- energy efficiency, noise emission and economic dimension. Int
form deformation. J Eng Gas Turb Power. 2019;141(10):101015. J Precis Eng Manuf. 2018;19(4):613–624. doi: 10.1007/s12541-
doi: 10.1115/1.4044692. 018-0074-3.