International Journal of Solids and Structures 300 (2024) 112893

Contents lists available at ScienceDirect

International Journal of Solids and Structures

journal homepage: www.elsevier.com/locate/ijsolstr

Machine learning predictions on the compressive stress–strain response of

lattice-based metamaterials
Lijun Xiao *, Gaoquan Shi , Weidong Song *
State Key Laboratory of Explosion Science and Technology, Beijing Institute of Technology, Beijing 100081, China

A R T I C L E I N F O A B S T R A C T

Keywords: Predicting the stress–strain curve of lattice-based metamaterials is crucial for their design and application.
Machine learning However, the complex nonlinear relationship between the mesoscopic structure of lattice materials and their
Additive manufacturing macroscopic mechanical behavior makes prediction challenging. In this study, beam element models of over
Lattice-based metamaterial
20,000 lattice structures were established using Python scripts, and calculations were performed by ABAQUS to
Sequential features in mechanical response
Spatial features
obtain training and testing datasets. The spatial features of each lattice-based metamaterial were then encoded
into a graph, a data structure recognizable by machine learning algorithm. Utilizing machine learning methods, a
Structure to Sequence Neural Network was constructed and trained, achieving rapid prediction of the
compressive stress–strain curves for lattice-based metamaterials. Afterwards, several lattice structures were
randomly selected and 3D printed. The accuracy of the simulation results as well as machine learning predictions
was validated through quasi-static compression experiments. It is revealed that the proposed Neural Network
model outperforms the traditional Artificial Neural Networks as the errors are reduced while the Coefficient of
Determination is higher. The results demonstrate the accurate fitting between the complex spatial features of the
lattice-based metamaterials and their stress–strain curves, which provides a potential methodology for inverse
optimization of the lattice-based metamaterials in the future.

1. Introduction research by Berger et al. (Berger et al., 2017) and Tancogne-Dejean et al.
(Tancogne-Dejean et al., 2018), the elastically-isotropic response of
Lattice-based metamaterials are a type of porous mechanical meta­ plate-lattices can reach the theoretical limits provided by Hashin and
materials arranged periodically by unit cells (Xiao et al., 2024; Zok et al., Shtrikman (Hashin and Shtrikman, 1963). At the same relative density,
2016). Most lattice structures possess excellent mechanical properties the stiffness of isotropic shell-lattices typically falls within the range
such as ultra-lightweight (Chougrani et al., 2017), high specific stiffness defined by optimal truss- and plate-lattices (Bonatti and Mohr, 2019).
and strength (Yin et al., 2021; Portela et al., 2018), exceptional dura­ Although the mechanical properties of truss lattice structures are not
bility (Ling et al., 2019), excellent damage tolerance (Li et al., 2022; optimal, they are still widely studied by researchers for the following
Wang et al., 2024), and outstanding energy absorption rate (Yin et al., reasons. From the manufacturing perspective, truss lattice structures
2023), which have been widely used in industries such as aerospace, consist of relatively simple structural components featuring an open-cell
medical, and transportation (Sur et al., 2021). With the rapid advance­ topology, which facilitates their fabrication (Ma et al., 2022). Mean­
ment of additive manufacturing which is suitable for creating intricate while, various truss lattice structures can be generated by varying the
geometries, the design possibilities of lattice-based metamaterials have diameter and arrangement of the struts, providing a large design space
been significantly expanded (Kumar et al., 2020). Lattice materials are and versatility for different applications (Gurtner and Durand, 2014;
generally divided into truss (Tancogne-Dejean and Mohr, 2018), plate Tancogne-Dejean and Mohr, 2018; Tancogne-Dejean et al., 2016;
(Liu, 2021), and shell (Xiao et al., 2022) lattice structures according to Challapalli and Li, 2021). Additionally, some truss lattice structures with
their cell characteristics. In general, the stiffest instances of truss-lattices a low relative density exhibit excellent performance in yield strength
only reach less than one third of the theoretical maximum stiffness for (Andersen et al., 2021), offering insights for the design of lightweight
isotropic porous materials (Gurtner and Durand, 2014). According to load-bearing structures.

* Corresponding authors.
E-mail addresses: [email protected] (L. Xiao), [email protected] (W. Song).

https://doi.org/10.1016/j.ijsolstr.2024.112893
Received 7 November 2023; Received in revised form 14 May 2024; Accepted 21 May 2024
Available online 31 May 2024
0020-7683/© 2024 Elsevier Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

reverse generation of lattice structures from mechanical performance

predictions. Garland et al. (Garland et al., 2021) introduced a method
utilizing a 12 × 12 matrix to represent different materials with distinct
pixel values. Afterwards, they predicted the mechanical properties of
two-dimensional lattice structures by CNNs, presenting a new approach
for neural networks to extract spatial features. Yu et al. (Yu et al., 2022)
constructed a CBLSTM network to predict the stress–strain curves of
porous silicone rubber with different geometric structures under
compression, demonstrating the clear advantages of LSTM in predicting
serialized data such as stress–strain curves. Eren et al. (Eren et al., 2024)
obtained a dataset through parametric design and simulated annealing
techniques, and implemented a complex structure-performance rela­
tionship fitting using a deep learning-based 3DGAN model. This enabled
the generation of three-dimensional truss lattice structures of aluminum
alloy with customized mechanical properties. Wilt et al. (Wilt et al.,
2020) used two-dimensional images to represent the distribution of
lattice unit cells with different Poisson’s ratios and employed CNNs to
predict the deformation patterns of the overall structure. The research
results showed that the error between the values predicted by the CNN
and the actual expected values was within 5 %. This study provides a
reference for understanding and predicting strain contours of lattice
Fig. 1. Stress–strain curve of the Tough-2000 tensile test specimen. structures during deformation.
Despite the widespread application of machine learning in the field
Currently, the primary approaches to investigate the mechanical of mechanics, we still face several challenges. The simulation of lattice
properties of lattice structures involve experiments and numerical sim­ structures to obtain datasets requires substantial computational and
ulations. However, the high economic and time costs associated with time costs, which induces that the majority of researchers limit their
experiments (Li et al., 2020; Wang et al., 2023; Xiao et al., 2020), as well studies to the elastic segment (Lee et al., 2022; Challapalli et al., 2021),
as the elevated computational costs and time requirements of numerical as the computational cost of elastic deformation is lower. Furthermore,
simulations (Xiao et al., 2022; Bai et al., 2020), have led to a relatively when using neural networks to extract spatial features of lattice struc­
slow progress in the research on lattice structures. With the advance­ tures, they are typically limited to one-dimensional (Lee et al., 2022;
ments in computer hardware, machine learning has evolved into a data- Challapalli et al., 2021) or two-dimensional (Garland et al., 2021; Yu
driven tool capable of deciphering any nonlinear relationships within et al., 2022) structure descriptors. These descriptors overly rely on
the data. This offers a new avenue for exploring the intricate relation­ manual operations, requiring experienced researchers to manually
ships between lattice structures and their mechanical responses determine many hyperparameters, which increases uncertainty and
(Capuano and Rimoli, 2019; Sepasdar et al., 2022.A). An Artificial computational complexity. Moreover, there is still a lack of research
Neural Network (ANN) was utilized by Lee et al. (Lee et al., 2022) to accurately predicting the complete stress–strain curves of lattice struc­
study the dependency of relative Young’s modulus on the relative den­ tures under large deformation based on machine learning. The complete
sity the body-centered cubic (BCC) lattice structure. The elastic pa­ stress–strain curve can be used to derive important mechanical property
rameters of BCC lattice structures with different strut cross-sections were descriptors such as modulus, strength, toughness, energy absorption,
predicted. Hu et al. (Hu et al., 2022) employed machine learning algo­ and can reflect the overall force evolution process of the lattice structure
rithms to investigate the dominant defects and design features affecting during deformation. Consequently, predicting the stress–strain curve of
the mechanical properties of truss lattice structures. The results indicate lattice structures is of great significance.
that the surface roughness of strut (peak-to-valley height) exceeding 10 In this study, over 20,000 truss lattice structures were randomly
% of the beam diameter significantly reduces the specific modulus and generated using Python scripts initially, and the corresponding beam
strength of the lattice structures, particularly for the bending-dominated element models were established in ABAQUS for numerical simulations
structures. Glaesener et al. (Glaesener et al., 2023) proposed a machine of quasi-static compression tests without considering strain rate and
learning framework that enables the prediction of geometric defects thermal effects. Subsequently, a training dataset and a test dataset
during the manufacturing process solely based on the mechanical required for machine learning were constructed based on the simulation
response of truss lattice structures. This framework provides a basis for results. The three-dimensional truss lattice structures were then encoded
designing and manufacturing truss lattice structures with superior me­ into graphs, which are data structures composed of vertices and edges,
chanical responses. Yu et al. (Yu et al., 2022) introduced a method for capable of fully representing the spatial features of lattice structures. A
representing heterogeneous truss lattice structures using a 4 × 4 matrix. Structure to Sequence Neural Network (abbreviated as Structure2SeqNN
Afterwards, the spatial features of lattice structures were extracted and in the following sections for convenience), based on Graph Convolu­
their mechanical responses were predicted by utilizing ANN. Wu et al. tional Networks (GCN) and Long Short-Term Memory networks (LSTM),
(Wu et al., 2023) adopted the structural characteristics of truss lattice was constructed to predict the macroscopic stress–strain curves of lattice
cell structures as inputs of neural networks to predict the energy ab­ structures under quasi-static compression. Afterwards, a group of lattice
sorption performance of lattice structures. structures were randomly chosen from the test dataset for fabrication
With the rapid development of machine learning algorithms, several using a resin photopolymerization 3D printer, and quasi-static
new neural network frameworks have been proposed, including Con­ compression tests were conducted to validate the simulation results
volutional Neural Network (CNN) (LeCun et al., 1998), Generative and neural network predictions. Finally, the advantages and disadvan­
Adversarial Network (GAN) (Goodfellow et al., 2020), Long Short-Term tages of our proposed Structure2SeqNN and traditional neural networks
Memory networks (LSTM) (Shi et al., 2015), Graph Neural Network were further compared, and the potential application of the proposed
(GNN) (Kipf and Welling, 2016), and others. Challapalli et al. (Challa­ networks was also discussed.
palli et al., 2021) represented truss lattice structures using one-
dimensional vectors and proposed a reverse generation framework for
truss lattice structures based on GAN. This framework enables the

2
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 2. The process of lattice structure modeling.

being lightweight. If the number of struts in the unit cell is too large, the
lattice structure will lose its lightweight characteristics. Therefore, this
paper controls the number of struts in the unit cell to 9 or less, ensuring
that the relative volume fraction of the lattice structure is within 0.15.
Although a unit cell with periodic boundary conditions can predict
the mechanical properties of lattice structures with infinite cells, it can
only be used to obtain the correct deformation mode of lattice structures
with homogeneous deformation. For lattice structures with localized
deformation, a multi-cell assembly model should be adopted instead. In
our manuscript, the simulations were not only used to obtain the
stress–strain curves of the lattice structures, but also aimed to compare
the overall deformation mode of the lattice structure with experimental
results to make the simulated data more credible. Accordingly, a 4 × 4 ×
4 cell layout is adopted here considering the computational load caused
by excessive cell quantities, which is sufficient for investigating the large
Fig. 3. Simulation model of the lattice structure. deformation mechanisms and mechanical properties of lattice struc­
tures. A successful 3D printing begins with a meticulously designed
2. Method model. To satisfy the preparation requirements of the 3D printer, the
length of the lattice cell and the diameter of the strut were specified as
2.1. Numerical simulation 15 mm and 0.8 mm, respectively. SolidWorks was adopted to model the
entire lattice structure and export the design in.stl format, which was
2.1.1. Base material recognizable by PreForm software—a slicing tool for 3D printing. The
The mechanical performance of lattice structures is largely contin­ lattice structure was also produced using the Form3 desktop stereo­
gent on the characteristics of their base materials. In this research, lithography printer and the Tough2000 base material. The employed
Tough-2000 resin was selected as the matrix material for the lattice laser power and spot diameter during fabrication were 250mW and
configurations. Compared to the other resin types, Tough-2000 0.085 mm, while the printer operated at a temperature of 35℃. The XY
demonstrated superior fracture resilience under various conditions resolution of Form3, which describes the printing accuracy on the XY
including compression, tension, bending, and impact. To gain an in- plane (horizontal plane), is 0.025 mm. To maximize the precision of 3D
depth understanding of the mechanical properties of this material, a printed lattice structures, the print layer thickness was set at 0.05 mm.
Form 3 desktop stereolithography printer was adopted to fabricate dog- Upon completion of printing, the lattice structures were immersed in an
bone specimens following the ASTM-D638-14 standard. Subsequently, a alcohol solution to clean off the adhered resin. Finally, the cleaned lat­
universal testing machine was employed to perform quasi-static axial tice structures were cured for 20 min, to enhance the tensile and bending
tension tests at a nominal strain rate of 0.001/s. The stress–strain curve strengths of the material while maintaining its ductility.
presented in Fig. 1 clearly illustrates the high fracture elongation and
high strength of the material. 2.1.3. Modeling and simulation
In this study, the ABAQUS software was employed for finite element
2.1.2. Design and fabrication of lattice-based mechanical metamaterials calculations to generate the dataset required for machine learning. The
The mechanical response of truss lattice structures primarily depends beam element model (Liu et al., 2017) is capable of accurately deter­
on the topology of the unit cells. For cells composed of struts with mining the mechanical properties of lattice structures with uniform strut
different quantities and positions, the corresponding mechanical prop­ diameters, necessitating only a small computational cost. Consequently,
erties of the lattice structures can vary significantly. In the current study, the beam element was used to simulate the lattice structures subjected to
a cubic unit composed of 8 nodes as shown in Fig. 2(a), is regarded as a quasi-static compression. It is worth noting that the beam element model
1/8 unit cell of the truss lattice structure. Each node in this unit is can be rapidly modeled and simulated using Python scripts.
numbered, and a strut in the lattice structure cell can be formed by A finite element model was established as illustrated in Fig. 3. The
connecting nodes with different numbers. For example, connecting node loading system consists of a loading platform, a lattice structure spec­
1 and node 7 can form an inclined strut. In the 1/8 unit cell, there are 28 imen, and a supporting platform. Both the upper loading platform and
different ways to randomly connect the eight vertices pairwise, resulting the lower supporting platform were modeled as discrete three-
in 28 different struts. By controlling the number of struts in the 1/8 dimensional rigid bodies. Reference points were positioned at the cen­
element of the unit cell and the connecting nodes of each strut, then ter of the loading and supporting platforms for setting boundary con­
mirroring the 1/8 unit cell along the three coordinate planes, all possible ditions and reading mechanical responses. In the quasi-static
unit cells of various lattice structures can be obtained. Fig. 2(b) shows a compression process, the lower support platform remained stationary,
randomly generated 1/8 elementary cell composed of 7 struts. The unit while the upper compression platform compressed the lattice structure
cell, after being mirrored along the three coordinate planes, is shown in specimen along the Z-axis at a constant speed of 1 m/s, using the
Fig. 2(c). One of the important reasons for the wide application of lattice ABAQUS/Explicit solver. The ABAQUS/Explicit solver was used to
structures is their ability to maintain high mechanical properties while simulate the quasi-static compression of the lattice structure, which can

3
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 4. Verification of quasi-static simulation state for (a). Total internal energy and total kinetic energy, (b) Stress response under different loading speeds.

Fig. 5. The graph representation process of lattice structure.

significantly improve calculation speed compared to the ABAQUS/ energy and total internal energy of the model under compression at a
Standard solver. According to the research by Santosa et al. (Santosa speed of 1 m/s. Fig. 4 demonstrates that the stress–strain curve under a
et al., 2000), when the total kinetic energy during compression simu­ compression rate of 1 m/s closely approximates the simulation result
lation is much smaller than the total internal energy, and the force­ under a strain rate of 0.001/s. Additionally, during the compression
–displacement response is similar at different velocities, explicit process at 1 m/s, the total kinetic energy is significantly smaller than the
algorithms can be applied to simulate quasi-static experimental pro­ total internal energy, indicating that simulating quasi-static compres­
cesses. So, this paper selected Lattice-C mentioned in section 3.2 and sion at 1 m/s in this study is feasible.
calculated its stress–strain curves under compression conditions for Timoshenko beam elements which consider the shear deformation
strain rates of 0.001/s (loading speed of 0.06 mm/s) and 16.6/s (loading were used in the simulations. They are applicable to both deep beams
speed of 1 m/s). Additionally, we calculated the changes in total kinetic where shear deformation plays a crucial role, and slender beams where

4
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 6. The specific structure of Structure2SeqNN.

shear deformation is less significant (Gümrük and Mines, 2013). The one-dimensional vectors or two-dimensional matrices for machine
lattice struts were assumed with a circular cross section, while all B31 learning studies, which needs to recognize and extract the structure
beam elements were linear and with a diameter of 0.8 mm. In this study, information manually. Recently, some researchers have employed
Python scripts were used to calculate the normal vectors of each element graphs to represent truss lattice structures, which can capture their
to determine the direction of the beam element cross-section. To study three-dimensional spatial characteristics (Ross and Hambleton, 2021;
the large deformation behavior of the lattice structure, a complete Zheng et al., 2023). However, most of them focus on the prediction of
elastic–plastic response of the base material was required. A multilinear simple mechanical variables such as Young’s modulus and Poisson’s
plastic hardening model which took the strain rate and temperature ratio. The research on predicting the complete stress–strain curves of
effects out of consideration was employed to describe the matrix ma­ lattice-based metamaterials based on graph neural networks is still
terial in this study. As depicted in Fig. 1, the true stress–strain curve of insufficient. To circumvent this problem, a novel Structure to Sequence
the matrix material was obtained through quasi-static tensile tests. Neural Network (Structure2SeqNN) is introduced and designed to
Subsequently, the elastic modulus of the base material could be deter­ directly extract features from three-dimensional lattice spatial structures
mined from the initial linear part of the true stress–strain curve which and predict the corresponding mechanical responses.
was 1051 MPa. The Possion’s ratio of 0.3 was assumed. The plastic strain
versus flow stress data was also extracted from the true stress–strain 2.3.1. Graph representation of structures
curve and was directly input into the ABAQUS software as material Due to the powerful expressiveness and interpretability of graph
parameters. This method has also been validated in our previous work structures, graph analysis using deep learning methods has emerged as a
(Yu et al., 2022). Moreover, the density of the base material was 1110 focal point in the field of machine learning. As a high degree of
kg/m3. In this paper, the general contact is employed to consider the abstraction of associated data, graphs are adept at representing the in­
contact between the specimen and two plates as well as the self-contact terrelationships among numerous entities, making them ideally suited
between the struts in the specimen. The contact property with “hard” in for capturing the spatial characteristics inherent to lattice structures. In
the normal direction and penalty friction in the tangential direction is essence, a graph consists of a vertex feature matrix V encapsulating all
adopted. vertex characteristics and an edge adjacency matrix A detailing all
connections between vertices. Therefore, a graph can be abstracted as a
function of the vertex feature matrix V and the adjacency matrix A,
2.2. Quasi-Static compression experiment
concisely denoted as:
To verify the precision of the numerical simulation results for the G = (V, A) (1)
lattice structure, quasi-static compression tests were conducted using an
electronic universal testing machine (SUNS, UTM5004) with a range of Taking the node connection method shown in Fig. 5(a) as an example,
100KN. All 3D printed lattice structures were compressed until reaching the first number next to each vertex in Fig. 5(a) is the vertex number,
densification. The strain rate for the quasi-static compression experi­ and the following three numbers are the relative coordinates of the
ments was set as 0.001/s, and the lattice structure was 60 mm high. vertex. Here, the relative coordinates were introduced to represent the
Based on calculations, the loading speed of the testing machine was relationships and distinguish the features among the 8 nodes in the cube
determined to be 3.6 mm/min. Vaseline was applied between the lattice (Fig. 5(a)). By fixing the coordinates of one node and the edge length of
structure samples and the indenter to minimize friction. A digital camera the cube, the coordinates of the remaining nodes can be determined.
(MER-503-36U3M), positioned directly in front of the samples, was Assuming the coordinates of the first node are (0,0,0), and the edge
employed to monitor the experimental process. The camera was set to length of the cube formed by these 8 nodes is 1, the relative coordinates
capture at a rate of 1 frame per second (fps) with a resolution of of the other nodes are (1,0,0), (1,1,0), (0,1,0), (1,0,1), (0,0,1), (0,1,1),
2048*2048 pixels. Load force and displacement data were recorded and (1,1,1). The number of other vertices connected to each vertex is the
using a sensor located at the top of the indenter. Consequently, the degree of the vertex. In Fig. 5(a), vertices 2, 4, and 5 are connected to
nominal stress–strain curves throughout the compression process of the vertex 1, so the degree of vertex 1 is 3. The relative coordinates and
lattice structure were calculated. degree of each vertex are selected as the features for each vertex. As
shown in Fig. 5(b), the first three columns in the vertex feature matrix
represent the relative coordinates of the vertices, and the fourth column
2.3. Machine learning
represents the degree of each vertex. In the 1/8 unit cell, there are 8
vertices in total, so the shape of the vertex feature matrix for the 1/8 unit
Conventional approaches generally represent lattice structures as

5
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 7. Graph Convolutional Network.

Structure2SeqNN is illustrated in the following figure. The imple­

mentation of Structure2SeqNN is reliant on the open-source neural
network libraries, PyTorch and Py-G.
The encoder primarily consists of Graph Convolutional Network
(GCN) (Kipf and Welling, 2016) and a Fully Connected (FC) Layer,
tasked with extracting the spatial features of the lattice structure. The
multi-layer GCN, as shown in Fig. 6, can update the vertex feature matrix
of lattice-structured nodes based on the adjacency matrix of the lattice
structure, thereby achieving the extraction of spatial features of the
lattice structure. The specific formulation of GCN is as follows:
( )
X(l+1) = σ Lsys X(l) W(l) (3)

In the aforementioned formulation, W(l) denotes the learnable weight

matrix in each layer of the GCN. The function σ() represents a nonlinear
activation function. Herein, the chosen activation function is LEAK­
RULE. The LEAKRULE activation function used in this paper is shown
Fig. 8. The specific structure of Long Short-Term Memory. below, where α is set to 0.01.
{
x, if x ≥ 0
cell is 8 × 4. The adjacency matrix A of the graph, as shown in Fig. 5(c), LEAKRULE = (4)
αx, others
represents the connections between each vertex. The adjacency matrix A
for the 1/8 unit cell is an 8 × 8 square matrix, with the specific formula According to the equation, each GCN layer takes two inputs: a
as follows: symmetrically normalized Laplacian matrix, denoted as Lsys , and a ver­
{ tex feature matrix, denoted as X(l) . Here, X(l) signifies the output from the
1, vi isadjacenttovj
Ａ[i][j] = (2) lth layer. When l = 1, X(0) stands for the vertex feature matrix V,
0, others
composed of relative coordinates of the eight vertices in the lattice
In the above equation, vi and vj respectively represent node i and structure. The symmetrically normalized Laplacian matrix encapsulates
node j. Taking vertex 1 in Fig. 5(a) as an example, vertex 1 is connected the vertex connectivity within the lattice structure, and its calculation is
to vertices 2, 4, and 5. Therefore, in the adjacency matrix A, the second, formulated as follows:
fourth, and fifth elements in the first row are 1, and the remaining ele­ 1 1 1
ments are 0. Similarly, the second, fourth, and fifth elements in the first Lsys = D ̂− 2D
̂− 2A ̂− 2
(5)
column are 1, and the remaining elements are 0. To incorporate more
information into the graph, this paper uses the graph encoded from the A
̂ = A+I (6)
mirrored complete unit cell as the input to the neural network. The
mirrored lattice structure is shown in Fig. 5(d). The complete unit cell D
̂ = D+I (7)
contains 27 vertices, so the shape of the vertex feature matrix V changes
to 27 × 4, and the shape of the adjacency matrix A changes to 27 × 27, as In the above equations, A is the adjacent matrix of the graph, and I is
shown in Fig. 5(e) and Fig. 4(f), respectively. the identity matrix. Following the feature extraction from the graph by
GCN, the output vertex feature matrix is reshaped into a one-
2.3.2. Structure2SeqNN dimensional vector in the order of vertices. Unlike pooling operations,
In this study, a novel neural network architecture denoted as this procedure ensures that no graph features are lost. The flattened
Structure2SeqNN is designed to convert a lattice structure graph into a feature vector is then fed into fully connected layer, also known as
fixed-length vector. Subsequently, this vector is decoded into a serial­ Artificial Neural Network (ANN), for further feature extraction and
ized mechanical response. A schematic representation of the integration.

6
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Table 1 3. Results
Evaluation of the prediction accuracy of different models for stress–strain curve
of lattice structure. 3.1. Training results of Structure2SeqNN
Model MAE R2 MSE
The stress–strain curve is a pivotal characterization of the mechan­
ANN-ANN 0.096272 0.916474 0.046130
ANN-LSTM 0.071131 0.960827 0.020608 ical attributes associated with lattice structures, reflecting significant
GCN-ANN 0.088707 0.932538 0.038333 mechanical parameters encompassing the elastic modulus, yield stress,
Structure2SeqNN 0.058459 0.975748 0.015199 and plateau stress. Consequently, this study employs the stress–strain
curve as a serialized mechanical response of lattice structures, while
deploying the Structure2SeqNN model for the approximation of the
As shown in Fig. 7, the encoder can extract the node feature matrix
intricate non-linear interrelation between lattice structures and their
after GCN processing. The node feature matrix is reshaped into a one-
corresponding stress–strain curve.
dimensional vector and remains a one-dimensional vector after being
In this segment, we harness a compilation of 19,688 data sets,
processed by the FC layer. The decoder element in our proposed model
serving as our training dataset. Remarkably, this dataset only constitutes
employs the Long Short-Term Memory (LSTM) units (Shi et al., 2015), a
a minuscule 0.17 % of the total sample space, thereby it merits classi­
distinctive variant from traditional neural networks which typically
fication as a small-sample training process. Additionally, a collection of
operate on parallel outputs and overlook the temporal relationships
1,035 data sets is employed as the testing dataset, designed to scrutinize
between data points. Conversely, LSTM units process data sequentially,
the training effectiveness of the deployed neural network. The Quantile
conscientiously factoring in the preceding temporal state when
loss function (Koenker and Hallock, 2001) is chosen as the loss function
analyzing the current instance. The mechanism through which the LSTM
of the model, with its detailed mathematical representation illustrated
units retain information across computational operations is attributed to
as follows:
cell states. These cell states, analogous to a conduit, persist throughout
∑ ∑
the entire computational process. Information stored in these cells un­ QuantileLoss = (γ − 1)|yi − ŷi | + (γ)|yi − ŷi | (13)
dergoes modifications at each LSTM neuron, mediated by a gating i=yi <yˆi i=yi ≥yˆi
mechanism. As outlined in the provided Fig. 8, an LSTM unit in­
corporates three gates, namely a forget gate, an input gate, and an here, γ denotes the desired quantile, which resides within the 0 to 1
output gate. The forget gate determines the extent of information to be interval. yi embodies the actual values, and ŷi encapsulates the predicted
discarded from the cell state, while the input gate selects novel infor­ counterparts. The Quantile loss function sanctions over-predictions or
mation to be included. These two gates collaborate to establish a under-predictions variably, hinging on the selected γ. Unlike the Mean
balanced integration of prior memories and new inputs. Subsequently, Square Error (MSE) and Mean Absolute Error (MAE) loss functions
the output gate determines the mechanism of information extraction which evaluate points, the Quantile loss function delineates a prediction
from the updated cell state. One key advantage of LSTM over the interval by defining multiple quantiles, thereby mitigating the impact of
traditional Recurrent Neural Networks (RNNs), which are engineered to outliers and fortifying the robustness of model predictions (Koenker and
handle sequence data, is its use of memory cells and gated information Hallock, 2001; Jadon et al., 2022). Within the scope of this study, we
flow, substantially mitigating the issue of gradient vanishing. The sub­ individually calculate the loss function values when γ equals 0.25, 0.5,
sequent operations conducted by the LSTM are detailed below: and 0.75. Subsequently, the arithmetic mean of the loss values across
these unique quantiles is employed to assess the training performance of
it = σ (Wxi xt + Whi ht− 1 + Wci ◦ ct− 1 + bi ) (8)
the model. In an attempt to evaluate the performance of Structur­
( ) e2SeqNN in structural feature extraction and sequential output, a set of
ft = σ Wxf xt + Whf ht− 1 + Wcf ◦ ct− 1 + bf (9)
comparative experiments was conducted. Initially, the GCN in the
encoder segment of Structure2SeqNN was substituted with an ANN of
ct = ft ◦ ct− 1 + it ◦ tanh(Wxc xt + Whc ht− 1 + bc ) (10)
equivalent layers. Afterwards, the LSTM in the decoder segment of
ot = σ(Wxo xt + Who ht− 1 + Wco ◦ ct + bo ) (11) Structure2SeqNN was similarly replaced with an ANN of matching
layers. Subsequently, these two altered networks were juxtaposed with
ht = ot ◦ tanh(ct ) (12) the original Structure2SeqNN for comparison. Moreover, an ANN
congruent in scale to Structure2SeqNN was chosen to scrutinize the
where ft , it , ct , ot , ht , and ◦ respectively symbolize the forget gate capability of Structure2SeqNN in nonlinear fitting. The specific param­
threshold, the input gate threshold, the updated cell state, the output eters of different neural networks can be found in Appendix A. Addi­
gate threshold, the output information and the Hadamard product. tionally, MAE, MSE and Coefficient of Determination R2 were adopted as
Since LSTM deals with sequential data, in order to convert the performance metrics for the neural network predictions. The mathe­
structural features extracted by the encoder into sequential data, we matical formulations of these metrics are as follows:
divided the interval from 0 to 1 into 101 time steps. These time steps are N
then multiplied by the structural feature vector, transforming the 1 ∑
MAE = |yi − ŷi | (14)
structural features into sequential features. These sequential features are N i=1
then input into the LSTM layer in the decoder for processing, finally
N
outputting a matrix with a sequence length of 101 and a feature length of 1 ∑
MSE = (yi − ŷi )2 (15)
1. In this paper, we discretize the stress–strain curve by taking 101 stress N i=1
values at a constant strain increment during the compression process of
the lattice structure. The most important aspect of discretizing the ∑N
(yi − y)2
stress–strain curve is to accurately and completely display the curve R2 = 1 − ∑Ni=1 2
(16)
i=1 ( yi − y)
̂
change process of the lattice structure during compression. As long as
this can be achieved, the discretization method of the curve is feasible, where yi represents the actual values, ŷi stands for the predicted values,
which will not influence the training process. and y denotes the mean values. The specific calculation results are
shown in Table 1. It can be seen that the MAE and MSE of Strutur­
e2SeqNN are the lowest among the four networks, while its R2 is the
highest.

7
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 9. Training processes of different neural networks:(A) ANN-ANN and ANN-LSTM;(B) GCN-ANN and Structure2SeqNN.

It can be captured from Fig. 9 that at the beginning of the training, lattice structures evolve through three discernible phases: an initial
the test machine loss function values of the four neural networks all linear portion, a plateau region, and finally a densification segment.
decrease with the decrease of the training dataset loss function value. During the linear phase, the nominal stress of Lattice-A advances line­
When the loss value decreases to around 0.142, overfitting phenomena arly in conjunction with the intensifying strain, with the elastic modulus
begin to appear in ANN-ANN, whereas GCN-ANN starts to overfit as the procured from both the experimental and simulation data exhibiting
loss value decreases to 0.128, indicating that GCN demonstrates rela­ close congruence. However, the experimental peak stress at the initial
tively superior feature extraction compared to ANN. Subsequently, stage is somewhat elevated compared to the simulated outcome. This
ANN-LSTM and Structure2SeqNN also exhibit overfitting when the loss discrepancy may be attributed to the residual adhesions from the resin
function value drops to 0.0982 and 0.0885, respectively. This suggests solution at the lattice structure vertices that were not thoroughly
that Structure2SeqNN possesses stronger spatial feature extraction and removed during the fabrication process. Prolonged cleaning durations,
sequence feature output capabilities, thereby offering a larger learning alternatively, could induce defects in the remaining lattice structure.
space and superior learning performance. The numerical simulation predicts an ideal outcome where such defects
To further investigate the proficiency of Structure2SeqNN in fitting are not included. Nevertheless, the trends of the experimental and
the nonlinear relationships between spatial attributes of lattice struc­ simulated curves in the post-yield region are consistent, exhibiting a
tures and their sequential mechanical responses, six distinct lattice slight decline followed by a rapid rebounding in stress. During the
structures were randomly chosen from the test dataset, as illustrated in plateau region, the simulated prediction aligns well with the experi­
Fig. 10. The comparison between the nominal stress–strain curves ob­ mental result, with minor oscillations likely ascribed to the mechanical
tained from numerical simulations and Structure2SeqNN predictions are properties of the base material and the deformation characteristics of the
presented in Fig. 10(a-f). Additionally, regression plots comparing the structure. Due to the beam elements neglecting the diameter of the
Structure2SeqNN predictions to the simulation results are illustrated in struts, the contact between the struts is too weak, ultimately causing the
Fig. 10(g-l). Ideally, the correlation between the actual values and the experimental curve reaches the densification stress earlier than the
values predicted by the neural network is linear. Consequently, in the simulation result. The nominal stress–strain curves of Lattice-B and
regression plots, the actual values are displayed as y = x, depicting an Lattice-C are portrayed in Fig. 11(B) and Fig. 11(C), respectively. It is
optimal situation where the predicted values are equal to the actual apparent that the simulated predictions coincide with the experimental
ones. As captured from the plots, the predictions from Structure2SeqNN results during the linear region, with a minimal discrepancy in initial
are uniformly and tightly dispersed around the line y = x, indicating that peak stresses. After the initial peak stress, a considerable decline is
the Structure2SeqNN predictions approach the optimal results. Mean­ evident in both the experimental and simulated results for Lattice-B and
while, the Structure2SeqNN proposed in this paper requires approxi­ Lattice-C, with a solid degree of fit. Subsequently, the experimental
mately 0.02 ms to predict a lattice structure when using an RTX3090, result of Lattice-B remains relatively stable, except for substantial fluc­
while ABAQUS requires about 10 min using a CPU with 30 cores for each tuations prior to densification. The simulated prediction exhibits larger
case. All the calculations are performed on the same desktop worksta­ oscillations but still agrees well with the experimental curve. Both the
tion. The prediction of stress–strain curves for lattice structures based on experimental and numerical curves of Lattice-C are relatively smooth
Structure2SeqNN significantly accelerates the research progress in lat­ and closely aligned during the plateau region.
tice structures. Subsequently, based on Structure2SeqNN, the stress–strain curves of
the three structures mentioned in the experiments were predicted. As
3.2. Experiment verification of FEM simulations can be seen in Fig. 11, the predicted results fit well with the experi­
mental and simulation results. In summary, the Structure2SeqNN pro­
In the domain of mechanics, the adoption of finite element simula­ posed in this paper can accurately extract the spatial features of lattice
tion results to construct neural network datasets is a widely employed structures and predict their serialized mechanical responses. Structur­
strategy (Yu et al., 2022). As a result, the precision of these simulation e2SeqNN has excellent nonlinear relationship fitting capabilities, which
outcomes significantly affects the performance of the ensuing neural can greatly accelerate the research progress of lattice structures.
network. Within this portion of our study, three structural designs The deformation evolution of Lattice-A is portrayed in Fig. 12(A). As
(Lattice-A, Lattice-B, and Lattice-C) were randomly extracted from the observed in the figure, the deformation pattern predicted by the simu­
test dataset. Afterwards, quasi-static compression experiments and nu­ lation is in substantial agreement with the experimental results. Initially,
merical simulations were conducted on these structures to validate the the deformation of Lattice-A is uniform with a progressively increasing
accuracy of the simulation results and the Structure2SeqNN predictions. load. However, as the load escalates, the first and fourth layers of
The nominal stress–strain trajectory under quasi-static load for Lattice-A begin to deform more significantly and collapse earlier than
Lattice-A is depicted in Fig. 11(A). Generally speaking, these curves of the middle two layers. Fig. 12(B) illustrates the deformation evolution

8
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 10. Comparison of Structure2SeqNN predicted results, STNN predicted results and simulation results:(a-f) stress–strain response for 6 different lattice structures;
(g-l) regression plots of Structure2SeqNN predicted results for 6 different lattice structures.

9
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 11. Comparison of experimental results, simulation results and Structure2SeqNN predicted results:(A) stress–strain response for Lattice-A; (B) stress–strain
response for Lattice-B; (C) stress–strain response for Lattice-C.

Fig. 12. Comparison of deformation mode between experiments and simulations: (A) Lattice-A; (B) Lattice-B; (C) Lattice-C.

10
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 13. The specific process of extracting feature at the graph level: (a) Max pooling; (b) Add pooling; (c) Mean pooling; (d) Vertex feature reshape.

Fig. 14. The error comparison of different image-level feature extraction methods: (a) MAE; (b) R2 ; (c) MSE.

process of Lattice-B. In the early stages of loading, both the simulation deformation across the structure. With further increase in load, the
and experimental results show a strikingly similar bending of the ver­ deformation in both the simulation and experimental results starts to
tical supports on the front face of Lattice-B, indicating a uniform concentrate on the third layer of Lattice-B. The deformation evolution

11
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. 15. The specific structure of Spatial-Temporal NN.

Table 2
Evaluation of the prediction accuracy of Structure2SeqNN and STNN for Fig. B1. Comparison of stress–strain curves between tensile experiment and
stress–strain curve of lattice structure. tensile simulation.
Model MAE R2 MSE

Structure2Seq NN 0.053459 0.975748 0.015199

Spatio-Temporal NN 0.109418 0.908276 0.058795

Fig. B2. The geometric parameters of the tensile test specimen.

Fig. 16. Various mechanical indicators in the stress–strain curve.

process of Lattice-C Fig. 12(C) is similar to the previous scenarios. The

deformation is uniform in the initial loading stage, with a good align­
ment between the experimental and simulation results. As the global
strain increases, the deformation in the simulation and in the experiment
begins to concentrate between the second and third layers of the
structure. However, compared to the experimental results, the defor­
mation in the simulation is more concentrated, and the struts in the
simulation bend more significantly. The reason for this phenomenon
may be the inaccurate consideration of mutual contact between the
struts in the simulation of the beam elements. Nonetheless, an exami­ Fig. C1. Comparison of stress–strain curves in experimental, solid element
nation of the images reveals a consistent mode of strut deformation simulation, and beam element simulation.
across all models, primarily characterized by a twisting at the vertices
which leads to an associated bending deformation pattern of the sur­ different structures, it is evident that the curves obtained by these three
rounding struts. methods are in excellent agreement. Furthermore, the deformation
Comparing the stress–strain curves obtained through simulation, patterns of the lattice structures obtained through simulation closely
Structure2SeqNN predictions, and experimental data for the three

12
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Fig. C2. Comparison of deformation modes in experimental, solid element simulation, and beam element simulation.

align with the experimental results. This implies the reliability of the maximum value within each column of the vertex feature matrix,
simulated data and demonstrates that Structure2SeqNN can accurately thereby capturing only the primary attributes of the graph and omitting
model the nonlinear relationship between lattice outcomes and their most vertex-specific features. The sum pooling and average pooling
associated stress–strain curves. methods incorporate every value within the vertex feature matrix into
their computations, but also neglect the distinctive features attributed to
4. Discussion each individual vertex. These three methodologies could potentially
overlook significant features within the entire graph structure, resulting
4.1. Influence of Graph-Level feature extraction method in an incomplete extraction of graph-level attributes. For small graphs
consisting of merely 27 vertices, this causes the loss of many features.
The Structure2SeqNN defined in this study typifies a graph-level Consequently, the neural networks employing max pooling and average
neural network, necessitating the extraction of features across the pooling methods yield larger prediction errors. Contrastingly, the vertex
entire graph. Conventionally utilized extraction methods encompass feature matrix reshaping technique proposed in this study ensures the
maximum pooling, summation pooling, and average pooling, which comprehensive retention of specific attributes corresponding to each
respectively derive the maximum, sum, and average values for each vertex. Furthermore, during the reshaping process, the vertex features
column in the vertex feature matrix following GCN processing (Varbella are arranged following the order of their respective vertex identifiers,
et al., 2023; Wang et al., 2020; Yang et al., 2023). The specific compu­ thereby preserving the intrinsic interrelationships among the vertices.
tational processes are elucidated in Fig. 13(a)~(c). In addition to these Consequently, the Structure2SeqNN, when employed with the vertex
methods, the present study introduced another graph-level feature matrix reshaping method, exhibits superior predictive efficacy.
extraction method specifically tailored for graphs with a fixed number of Recently, some researchers have proposed novel pooling methods,
vertices. This method involves transposing the vertex feature matrix into such as global attention pooling (Jiyu et al., 2023; Bastek and Koch­
a one-dimensional vector, following the sequential order of the vertices, mann, 2023). To further validate the pooling capability of the proposed
as presented in Fig. 13 (d). in this study, we compare it with global attention pooling. As shown in
In order to thoroughly investigate the influence of graph-level Fig. 14, it can be concluded that the error of the neural network using
feature extraction methods on neural network performance, four global attention pooling has been reduced compared to traditional
distinct neural networks that employ various graph-level feature ag­ pooling methods. However, the reshaping of vertex feature matrix
gregation techniques were designed and trained. Subsequently, metic­ proposed in this paper is still slightly better. The reason may be that the
ulous computation and analysis of the errors associated with these four number of vertices of the object studied in this paper remains un­
networks were undertaken. As displayed in Fig. 14, it is conspicuous that changed, and a fully connected layer were added after the reshaping
the neural networks deploying maximum pooling, sum pooling, and function, which is equivalent to increasing the learnability of this
average pooling techniques exhibit markedly higher MAE and MSE than method. Moreover, the number of vertices in the graphs presented in this
that adopts the vertex feature matrix reshaping method. Meanwhile, the study is relatively small. The number of learnable parameters in global
R2 of the vertex feature reshaping method is also the highest. Particu­ attention pooling is related to the number of vertices, and a smaller
larly striking is the performance of networks leveraging the add pooling number of parameters may lead to incomplete feature extraction from
technique, wherein their MAE and MSE reach approximately double and the graph. Consequently, the application of global attention pooling may
decuple the values associated with the vertex feature matrix reshaping not yield optimal results in the context of this study.
method, respectively. These findings demonstrate that the choice of Generally, the number of vertices in the graphs processed by GNN is
feature extraction method significantly influences the accuracy of the variable. However, the range of this variation should be close to the
predicted results. Compared with the maximum pooling, sum pooling, number of vertices in the graphs within the training dataset. This en­
and average pooling techniques, the vertex feature matrix reshaping sures the accuracy of GNN predictions. It is suggested that the maximum
method offers more precise extraction of spatial features from the lattice number of vertices in a graph be determined based on the training
structures. Specifically, the max pooling method focuses solely on the dataset prior to applying the proposed vertex matrix reshaping method.

13
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

If the number of vertices in a graph input into the neural network is less time consumption for computation is increased, possibly leading to
than this maximum number, zeros can be used to pad the reshaped one- overfitting. Additionally, the graphs used in this study consist of only 27
dimensional vector, ensuring that graphs with different numbers of vertices with relatively simple structures. The combination of relatively
vertices yield one-dimensional vectors of the same length after pooling. simple graphs with an overly complex network structure might result in
This extends the versatility of the vertex feature reshaping method. excessively smooth outputs, failing to capture the specific features of the
graphs. In contrast, the LSTM layer in Structure2Seq NN solely focuses
4.2. Structure2Seq NN or Spatio-Temporal NN? on serializing spatial features of the lattice structure, avoiding any
additional temporal computational overhead. Consequently, the Struc­
The distinct structural features of the lattice structure can be referred ture2Seq NN successfully extracts more specific spatial features
to as spatial-domain features, whereas the corresponding stress states compared to the STNN. Even for relatively simple graphs, such as truss
that change with strain during the quasi-static compression process can lattice structures, Structure2Seq NN achieves remarkably good results.
be interpreted as temporal-domain features. However, the Structur­
e2SeqNN model proposed in this study, consisting of GCN and LSTM, 4.3. The further applications of Structure2SeqNN
should not be categorically identified as a Spatio-Temporal Neural
Network (STNN) (Sahili and Awad, 2023; He et al., 2020). The Struc­ As provided Fig. 16, it is evident that the stress–strain curve of the
ture2SeqNN model solely incorporates the spatial-domain features of lattice structure contains crucial mechanical parameters such as the
the lattice structure as its inputs. In particular, the input to the LSTM elastic modulus, yield stress, plateau stress, and energy absorption ca­
layer is the integrated spatial-domain features obtained from GCN and pacity. Therefore, based on accurate predictions from Structure2SeqNN
Full Connected Layers to be precise, the function of the LSTM layer on the stress–strain curve of the lattice structure, we can effectively
within the Structure2SeqNN model is to serialize the spatial features of extract multiple mechanical parameters of the structure. Moreover, by
the lattice structure, rendering the resultant curve smoother and more employing methods such as genetic algorithms, we can perform targeted
accurate. Conversely, a genuine STNN ought to accept inputs from both optimization for multiple objectives specific to the lattice structure (Hu
spatial-domain features and temporal sequence stimuli. Moreover, the et al., 2023; Pour et al., 2023; Wang et al., 2022).
input of spatial-domain features should dynamically alter in tandem Structure2SeqNN is a highly versatile neural network that serves as a
with the change in time steps. powerful tool in the field of structural analysis and design. For lattice
A novel STNN by integrating GCN into LSTM based on the work of structures, the sequence features encompass not only stress–strain
Seo (Seo et al., 2018) was constructed to evaluate the performance of curves but also other crucial characteristics such as velocity–time curves
Structure2SeqNN. The specific computational process of STNN is illus­ in impact experiments (Shakirzyanova et al., 2021) and absorption
trated in Fig. 15. In this network, GCN is employed to extract spatial coefficient-frequency curves in acoustic experiments (Kumar et al.,
features from the lattice structure, while LSTM processes the temporal 2023; Li et al., 2023). The application of Structure2SeqNN in these
stimuli in the form of time series. The decoding is performed through a diverse domains accelerates the pace of research significantly. Further­
fully connected layer. The proposed STNN consists of a total of 101 time more, the utility of graph representation extends beyond lattice struc­
steps, with each step taking an input graph. It is noteworthy that the tures alone. For instance, Yamaguch et al. (Yamaguchi et al., 2022)
connectivity pattern of vertices in the graph remains constant successfully applied the graph representation method to Tachi-Miura
throughout the time steps, ensuring that the adjacency matrix remains Polyhedron structures, while Meyer et al. (Meyer et al., 2022) imple­
unchanged when fed into STNN. However, the vertex feature matrix mented it for thin-walled lattice structures. These structures generate
keeps changing with each time step. The specific approach is to multiply various sequence features reflecting their mechanical characteristics
the vertex feature matrix at each step with the corresponding temporal when subjected to external loading, all of which can be rapidly predicted
stimulus. This temporal stimulus ranges from 0 to 1, comprising 101 using Structure2SeqNN. In summary, the flexibility and potent
values, effectively incorporating time-domain features into the vertex nonlinear fitting capabilities of Structure2SeqNN position it as a
feature matrix. By synergizing GCN and LSTM within the proposed formidable tool in the field of structural analysis and design. Its ability to
STNN, we achieved joint modeling and analysis of spatial features handle diverse sequence features and adapt to different types of lattice
extracted from the lattice structure and temporal dynamics derived from structures holds immense potential for breakthroughs and advance­
the time series stimuli. The specific formula for LSTM nested in GCN is ments across multiple research domains and engineering practices.
provided as follows:
5. Conclusion
it = σ (Wxi GCN(Xt , At ) + Whi ht− 1 + Wci ◦ ct− 1 + bi ) (17)
( ) In this study, an Structure2SeqNN that accurately predicts stress–­
ft = σ Wxf GCN(Xt , At ) + Whf ht− 1 + Wcf ◦ ct− 1 + bf (18)
strain curves of lattice structures is developed. Initially, over 20,000
ct = ft ◦ ct− 1 + it ◦ tanh(Wxc GCN(Xt , At ) + Whc ht− 1 + bc ) (19) data sets were randomly selected from the sample space to establish the
training and testing datasets. The Structure2SeqNN was then trained
ot = σ(Wxo GCN(Xt , At ) + Who ht− 1 + Wco ◦ ct + bo ) (20) and tested, and its exceptional performance was explored. Subsequently,
the accuracy of the numerical simulation predictions and the predictions
ht = ot ◦ tanh(ct ) (21) of Structure2SeqNN was experimentally validated. These findings offer
promising prospects for the analysis and design of lattice-based meta­
here, GCN(Xt , At ) represents graph convolution operations on the input materials. The following are some key conclusions drawn from this
graph data, while Xt and At are the vertex feature matrix and adjacency research:
matrix at time step t, respectively. The remaining parameters are the
same as described in Section 2. (1) The numerical simulation based on beam elements, without
The same dataset was employed to perform training and testing on considering the strain rate and thermal effects, can accurately
the two networks. The results have been incorporated in Fig. 10 and predict the quasi-static compression results of lattice structures
Table 2, which indicates that the Structure2SeqNN continues to exhibit with the relative density below 15 %. This suggests that the
superior predictive performance. The possible reason could be the dataset derived from the numerical simulation is precise.
integration of GCN into LSTM, which necessitates four graph convolu­ (2) The calculation efficiency is significantly promoted by the pro­
tion operations per layer and 101 computations per graph in the tem­ posed Structure2SeqNN compared with traditional numerical
poral dimension. Accordingly, the complexity of the model as well as the simulations. The prediction for a lattice structure can be finished

14
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

within 0.02 ms by the proposed network when using an Validation, Investigation, Funding acquisition, Conceptualization.
RTX3090, while classical simulation through ABAQUS requires Gaoquan Shi: Validation, Software, Methodology, Investigation,
about 10 min to compute a lattice structure by using the same Formal analysis. Weidong Song: Writing – review & editing, Supervi­
workstation. sion, Project administration.
(3) Using traditional ANN to extract spatial features of lattice struc­
tures requires manual digitization of the spatial characteristics. In
contrast, representing lattice structures as graphs allows for a Declaration of competing interest
more faithful reconstruction of their spatial features, minimizing
human intervention and enhancing the accuracy of spatial The authors declare that they have no known competing financial
feature extraction by the GCN within Structure2SeqNN. interests or personal relationships that could have appeared to influence
(4) The method of vertex feature matrix reshaping effectively ex­ the work reported in this paper.
tracts spatial features of lattice structures compared to conven­
tional pooling methods. It guarantees that the spatial features of Data availability
lattice structures are not lost and enables Structure2SeqNN to
achieve higher prediction accuracy and stronger generalization Data will be made available on request.
capabilities.
(5) The stress–strain curve of lattice structure represents a typical Acknowledgements
time series feature. However, Structure2SeqNN only extracts the
spatial characteristics of the lattice structure without extracting The authors are grateful for the financial support from the National
temporal stimuli. The role of the LSTM layer in Structure2SeqNN Natural Science Foundation of China (12372349, 12002049, 11972092,
is to serialize the spatial features, leading to more accurate and 12172056) and the Beijing Institute of Technology Research Fund Pro­
smooth curves. gram for Young Scholars (XSQD-202102005). Meanwhile, the opening
project of Sichuan Provincial Key Lab of Shock and Vibration of Engi­
CRediT authorship contribution statement neering Materials and Structures (Southwest University of Science and
Technology, 22kfgk02), and the grant from State Key Laboratory of
Lijun Xiao: Writing – review & editing, Writing – original draft, Computational Physics (6142A05230102) are also acknowledged.

A. The specific parameters of different neural networks

Due to the different data structures handled by the various networks compared in this paper, it is difficult to ensure that the number of neurons in
the neural networks is exactly the same. Therefore, when controlling the number of layers in different networks, this paper changes the number of
neurons in each layer of the neural network by powers of 2. For most computer hardware, the efficiency of memory allocation and data processing is
the highest when the size of the data is a power of 2. Consequently, if the number of neurons is a power of 2, matrix operations and data transmission
may be more efficient. In Structure2SeqNN, the number of node features output by each layer of GCN are: 16, 32, 64, 128, 256, and the length of the
vector after reshaping the node feature matrix is 6912. The output feature length of the subsequent fully connected layer is 4112, and the sequence
length of the subsequent two LSTM layers is 101, with the number of output units being 1024 and 1, respectively. In GCN-ANN, the number of node
features output by each layer of GCN are: 16, 32, 64, 128, 256, and the length of the vector after reshaping the node feature matrix is also 6912. The
output feature lengths of the subsequent three fully connected layers are 4112, 1024, and 101, respectively. The input to ANN-ANN and ANN-LSTM is
a vector of length 28, because there are 18 types of struts formed by pairwise connections of nodes in the 1/8 unit cell composed of 8 nodes. Each
element in the vector represents a strut, with the element being 0 indicating that the strut does not exist, while 1 indicating that the strut exists. To
ensure the same number of network layers, ANN-LSTM consists of 6 fully connected layers and 2 LSTM layers, with the number of units in the first 6
fully connected layers being 32, 64, 128, 256, 512, 1024, and the sequence length of the 2 LSTM layers being 101, with the number of output units
being 512 and 1, respectively. ANN-ANN consists of 8 layers of fully connected layers, with the outputs of each layer being 32, 64, 128, 256, 512, 1024,
512, and 101, respectively. The learnable parameters for Structure2SeqNN are 49515592, for GCN-ANN are 32785077, for ANN-LSTM are 3850216,
and for ANN-ANN are 1277893. The input for GCN is a node feature matrix and an adjacency matrix, while ANN takes a one-dimensional vector, and
LSTM processes serialized data with a sequence length of 101. The data format required by different types of neural networks varies, which means the
amount of data input into different neural networks differs. And the number of learnable parameters needed by neural networks to process these data
also varies. Consequently, this study did not fully control for each model to have the same number of learnable parameters. Instead, it ensured that the
number of layers in the neural networks was identical and varied the number of neurons per layer in accordance with the size of the input data and a
consistent rule.

B. A comparison between numerical simulation and experimental results for the tensile test

This paper provided a comparison between numerical simulation and experimental results for the tensile test at a strain rate of 0.001/s. As shown
in the figure below, it can be seen that the experimental and simulated results are close, indicating that the simulations of the lattice structures in this
paper are correct. During the simulation, the elastic modulus E is 1051 MPa, with a Poisson’s ratio of 0.3. The stress–strain relationship in the plastic
region is described by the Swift law:
( )n
σ = A εp + ε0

The Swift parameters A = 0.056MPa, ε0 = 3.309 and n = 5.651 are used. (See Figs. B1 and B2)

15
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

C. A comparison between beam element simulation and solid element simulation for the lattice structure

To validate the accuracy of the beam element simulation, this paper compares the computational results between the beam element model and the
solid element model. In the figure below, it can be observed that the deformation patterns and stress–strain curves during compression of the grid
structure using beam elements closely match the simulation results of the model using solid elements.(See Figs. C1 and C2).

References acoustic meta-absorbers. Appl. Acoust. 212, 109603 https://doi.org/10.1016/j.

apacoust.2023.109603.
LeCun, Y., Bottou, L., Bengio, Y., et al., 1998. Gradient-based learning applied to
Andersen, M.N., Wang, F., Sigmund, O., 2021. On the competition for ultimately stiff and
document recognition. Proc. IEEE 86 (11), 2278–2324. https://doi.org/10.1109/
strong architected materials. Mater. Design 198, 109356. https://doi.org/10.1016/j.
5.726791.
matdes.2020.109356.
Lee, S., Zhang, Z., Gu, G.X., 2022. Generative machine learning algorithm for lattice
Bai, L., Gong, C., Chen, X., et al., 2020. Mechanical properties and energy absorption
structures with superior mechanical properties. Mater. Horiz. 9 (3), 952–960.
capabilities of functionally graded lattice structures: Experiments and simulations.
https://doi.org/10.1039/D1MH01792F (Communication) Mater. Horiz., 2022, 9,
Int. J. Mech. Sci. 182, 105735 https://doi.org/10.1016/j.ijmecsci.2020.105735.
952-960.
Bastek, J., Kochmann, D.M., 2023. Inverse-design of nonlinear mechanical metamaterials
Li, C., Lei, H., Zhang, Z., et al., 2020. Architecture design of periodic truss-lattice cells for
via video denoising diffusion models. ArXiv 2305 (19836). https://doi.org/
additive manufacturing. Addit. Manuf. 34, 101172 https://doi.org/10.1016/j.
10.48550/arXiv.2305.19836.
addma.2020.101172.
Berger, J.B., Wadley, H.N.G., McMeeking, R.M., 2017. Mechanical metamaterials at the
Li, C., Lu, Y., Lan, C., et al., 2023. Noise Reduction in Helicopter Cabins Using
theoretical limit of isotropic elastic stiffness. Nature 543 (7646), 533–537. https://
Microperforated Panel Composite Sound Absorption Structures. Appl. Sci. 13, 8153.
doi.org/10.1038/nature21075.
https://doi.org/10.20944/preprints202306.1509.v1.
Bonatti, C., Mohr, D., 2019. Mechanical performance of additively-manufactured
Li, Z., Wang, X., Li, X., Wang, Z., Zhai, W., 2022. New Class of Multifunctional
anisotropic and isotropic smooth shell-lattice materials: Simulations & experiments.
Bioinspired Microlattice with Excellent Sound Absorption, Damage Tolerance, and
J. Mech. Phys. Solids 122, 1–26. https://doi.org/10.1016/j.jmps.2018.08.022.
High Specific Strength. ACS Appl. Mater. Interfaces. https://doi.org/10.1021/
Capuano, G., Rimoli, J.J., 2019. Smart finite elements: A novel machine learning
acsami.2c19456.
application. Comput. Methods Appl. Mech. Eng. 345, 363–381. https://doi.org/
Ling, C., Cernicchi, A., Gilchrist, M.D., et al., 2019. Mechanical behaviour of additively-
10.1016/j.cma.2018.10.046.
manufactured polymeric octet-truss lattice structures under quasi-static and dynamic
Challapalli, A., Li, G., 2021. Machine learning assisted design of new lattice core for
compressive loading. Mater. Des. 162, 106–118. https://doi.org/10.1016/j.
sandwich structures with superior load carrying capacity. Sci. Rep. 11 (1), 18552.
matdes.2018.11.035.
https://doi.org/10.1038/s41598-021-98015-7.
Liu, Y., 2021. Mechanical properties of a new type of plate–lattice structures. Int. J.
Challapalli, A., Patel, D., Li, G., 2021. Inverse machine learning framework for
Mech. Sci. 192, 106141 https://doi.org/10.1016/j.ijmecsci.2020.106141.
optimizing lightweight metamaterials. Mater. Des. 208, 109937 https://doi.org/
Liu, L., Kamm, P., García-Moreno, F., et al., 2017. Elastic and failure response of
10.1016/j.matdes.2021.109937.
imperfect three-dimensional metallic lattices: the role of geometric defects induced
Chougrani, L., Pernot, J.P., Véron, P., et al., 2017. Lattice structure lightweight
by Selective Laser Melting. J. Mech. Phys. Solids 107, 160–184. https://doi.org/
triangulation for additive manufacturing. Comput. Aided Des. 9, 95–104. https://
10.1016/j.jmps.2017.07.003.
doi.org/10.1016/j.cad.2017.05.016.
Ma, Q., Yan, Z., Zhang, L., et al., 2022. The family of elastically isotropic stretching-
Eren, O., Yüksel, N., Börklü, H.R., et al., 2024. Deep learning-enabled design for tailored
dominated cubic truss lattices. Int. J. Solids Struct. 239, 111451 https://doi.org/
mechanical properties of SLM-manufactured metallic lattice structures. Eng. Appl.
10.1016/j.ijsolstr.2022.111451.
Artif. Intel. 130, 107685 https://doi.org/10.1016/j.engappai.2023.107685.
Meyer, P.P., Bonatti, C., Tancogne-Dejean, T., et al., 2022. Graph-based metamaterials:
Garland, A.P., White, B.C., Jensen, S.C., et al., 2021. Pragmatic generative optimization
Deep learning of structure-property relations. Mater. Des. 223, 111175 https://doi.
of novel structural lattice metamaterials with machine learning. Mater. Des. 203,
org/10.1016/j.matdes.2022.111175.
109632 https://doi.org/10.1016/j.matdes.2021.109632.
Portela, C.M., Greer, J.R., Kochmann, D.M., 2018. Impact of node geometry on the
Glaesener, R.N., Kumar, S., Lestringant, C., et al., 2023. Predicting the influence of
effective stiffness of non-slender three-dimensional truss lattice architectures.
geometric imperfections on the mechanical response of 2D and 3D periodic trusses.
Extreme Mech. Lett. 22, 138–148. https://doi.org/10.1016/j.eml.2018.06.004.
Acta Mater. 254, 118918 https://doi.org/10.1016/j.actamat.2023.118918.
Pour, A.F., Faradonbeh, R.S., Gholampour, A., et al., 2023. Predicting ultimate condition
Goodfellow, I., Pouget-Abadie, J., Mirza, M., et al., 2020. Generative adversarial
and transition point on axial stress–strain curve of FRP-confined concrete using a
networks. Commun. ACM 63 (11), 139–144. https://doi.org/10.1145/3422622.
meta-heuristic algorithm. Compos. Struct. 304, 116387 https://doi.org/10.1016/j.
Gümrük, R., Mines, R.A.W., 2013. Compressive behaviour of stainless steel micro-lattice
compstruct.2022.116387.
structures. Int. J. Mech. Sci. 68, 125–139. https://doi.org/10.1016/j.
Ross E, Hambleton D, 2021. Using graph neural networks to approximate mechanical
ijmecsci.2013.01.006.
response on 3d lattice structures. Proceedings of AAG2020-Advances in
Gurtner, G., Durand, M., 2014. Stiffest elastic networks. Proc. Math. Phys. Eng. Sci 470
Architectural Geometry, 24: 466-485.
(2164), 20130611. https://doi.org/10.1098/rspa.2013.0611.
Sahili Z A, Awad M, 2023. Spatio-Temporal Graph Neural Networks: A Survey. arXiv
Hashin, Z., Shtrikman, S., 1963. A variational approach to the theory of the elastic
preprint arXiv:2301.10569. https://doi.org/10.48550/arXiv.2301.10569.
behaviour of multiphase materials. J. Mech. Phys. Solids 11 (2), 127–140. https://
Santosa, S.P., Wierzbicki, T., Hanssen, A.G., et al., 2000. Experimental and numerical
doi.org/10.1016/0022-5096(63)90060-7.
studies of foam-filled sections. Int. J. Impact Eng 24 (5), 509–534. https://doi.org/
He, Z., Chow, C.Y., Zhang, J.D., 2020. STNN: A spatio-temporal neural network for traffic
10.1016/S0734-743X(99)00036-6.
predictions. IEEE Trans. Intell. Transp. Syst. 22 (12), 7642–7651. https://doi.org/
Seo Y, Defferrard M, Vandergheynst P, et al., 2018. Structured sequence modeling with
10.1109/TITS.2020.3006227.
graph convolutional recurrent networks. Neural Information Processing: 25th
Hu, E., Seetoh, I.P., Lai, C.Q., 2022. Machine learning assisted investigation of defect
International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16,
influence on the mechanical properties of additively manufactured architected
2018, Proceedings, Part I 25. Springer International Publishing. 362-373. https://
materials. Int. J. Mech. Sci. 221, 107190 https://doi.org/10.1016/j.
doi.org/10.1007/978-3-030-04167-0_33.
ijmecsci.2022.107190.
Sepasdar, R., Karpatne, A., Shakiba, M., 2022.A. data-driven approach to full-field
Hu, B., Wang, Z., Du, C., et al., 2023. Multi-objective Bayesian optimization accelerated
nonlinear stress distribution and failure pattern prediction in composites using deep
design of TPMS structures. Int. J. Mech. Sci. 244, 108085 https://doi.org/10.1016/j.
learning. Comput. Methods Appl. Mech. Eng. 397, 115126 https://doi.org/10.1016/
ijmecsci.2022.108085.
j.cma.2022.115126.
Jadon A, Patil A, Jadon S. A, 2022. Comprehensive Survey of Regression Based Loss
Shakirzyanova, V.V., Selivanov, V.V., Silnikov, M.V., et al., 2021. Behavior of multilayer
Functions for Time Series Forecasting. arXiv preprint arXiv:2211.02989. https://doi.
transparent spacecraft elements during high-speed collision with compact impactors.
org/10.1109/TNNLS.2023.3281643.
Acta Astronaut. 180, 119–124. https://doi.org/10.1016/j.actaastro.2020.12.003.
Jiyu, H., Lin, G., Zihan, C., Liukai, C., Haoying, C., Zhiying, C., 2023. A Global Attention
Shi, X., Chen, Z., Wang, H., et al., 2015. Convolutional LSTM network: A machine
Pooling-Based Graph Learning Scheme for Generator-Level Transient Stability
learning approach for precipitation nowcasting. Adv Neural Inf Process Syst. 28
Assessment, in: 2023 IEEE Power & Energy Society General Meeting (PESGM). pp.
https://doi.org/10.48550/arXiv.1506.04214.
1–5. https://doi.org/10.1109/PESGM52003.2023.10253244.
Sur, A., Darvekar, S., Shah, M., 2021. Recent advancements of micro-lattice structures:
Kipf T N, Welling M, 2016. Semi-supervised classification with graph convolutional
application, manufacturing methods, mechanical properties, topologies and
networks. arXiv preprint arXiv:1609.02907. https://doi.org/10.48550/
challenges. Arab. J. Sci. Eng. 46, 11587–11600. https://doi.org/10.1007/s13369-
arXiv.1609.02907.
021-05992-y.
Koenker, R., Hallock, K.F., 2001. Quantile regression. J. Econ. Perspect. 15 (4), 143–156.
Tancogne-Dejean, T., Diamantopoulou, M., Gorji, M.B., et al., 2018. 3D plate-lattices: an
https://doi.org/10.1257/jep.15.4.143.
emerging class of low-density metamaterial exhibiting optimal isotropic stiffness.
Kumar, A., Collini, L., Daurel, A., et al., 2020. Design and additive manufacturing of
Adv. Mater. 30 (45), 1803334. https://doi.org/10.1002/adma.201803334.
closed cells from supportless lattice structure. Addit. Manuf. 33, 101168 https://doi.
Tancogne-Dejean, T., Mohr, D., 2018. Stiffness and specific energy absorption of
org/10.1016/j.addma.2020.101168.
additively-manufactured metallic BCC metamaterials composed of tapered beams.
Kumar, S., Jin, H., Lim, K.M., et al., 2023. Comparative analysis of machine learning
Int. J. Mech. Sci. 141, 101–116. https://doi.org/10.1016/j.ijmecsci.2018.03.027.
algorithms on prediction of the sound absorption coefficient for reconfigurable

16
L. Xiao et al. International Journal of Solids and Structures 300 (2024) 112893

Tancogne-Dejean, T., Mohr, D., 2018. Elastically-isotropic truss lattice materials of Xiao, L., Song, W., Xu, X., 2020. Experimental study on the collapse behavior of graded
reduced plastic anisotropy. Int. J. Solids Struct. 138, 24–39. https://doi.org/ Ti-6Al-4V micro-lattice structures printed by selective laser melting under high
10.1016/j.ijsolstr.2017.12.025. speed impact. Thin Wall Struct. 155, 106970 https://doi.org/10.1016/j.
Tancogne-Dejean, T., Spierings, A.B., Mohr, D., 2016. Additively-manufactured metallic tws.2020.106970.
micro-lattice materials for high specific energy absorption under static and dynamic Xiao, L., Feng, G., Li, S., et al., 2022. Mechanical characterization of additively-
loading. Acta Mater. 116, 14–28. https://doi.org/10.1016/j.actamat.2016.05.054. manufactured metallic lattice structures with hollow struts under static and dynamic
Varbella, A., Gjorgiev, B., Sansavini, G., 2023. Geometric deep learning for online loadings. Int. J. Impact Eng 169, 104333. https://doi.org/10.1016/j.
prediction of cascading failures in power grids. Reliab. Eng. Syst. Saf. 237, 109341 ijimpeng.2022.104333.
https://doi.org/10.1016/j.ress.2023.109341. Xiao, L., Xu, X., Feng, G., et al., 2022. Compressive performance and energy absorption of
Wang, Z., Cao, X., Yang, H., et al., 2023. Additively-manufactured 3D truss-lattice additively manufactured metallic hybrid lattice structures. Int. J. Mech. Sci. 219,
materials for enhanced mechanical performance and tunable anisotropy: Simulations 107093 https://doi.org/10.1016/j.ijmecsci.2022.107093.
& experiments. Thin Wall Struct. 183, 110439 https://doi.org/10.1016/j. Yamaguchi, K., Yasuda, H., Tsujikawa, K., et al., 2022. Graph-theoretic estimation of
tws.2022.110439. reconfigurability in origami-based metamaterials. Mater. Des. 213, 110343 https://
Wang Z, Wei W, Cong G, et al., 2020. Global context enhanced graph neural networks for doi.org/10.1016/j.matdes.2021.110343.
session-based recommendation. Proceedings of the 43rd international ACM SIGIR Yang, Y., Zou, D., He, X., 2023. Graph Neural Network-Based Node Deployment for
conference on research and development in information retrieval. 169-178. https:// Throughput Enhancement. IEEE Trans Neural Netw Learn Syst. https://doi.org/
doi.org/10.1145/3397271.3401142. 10.1109/TNNLS.2023.3281643.
Wang, X., Li, X., Li, Z., Wang, Z., Zhai, W., 2024. Superior Strength, Toughness, and Yin, S., Guo, W., Wang, H., et al., 2021. Strong and tough bioinspired additive-
Damage-Tolerance Observed in Microlattices of Aperiodic Unit Cells. Small manufactured dual-phase mechanical metamaterial composites. J. Mech. Phys.
2307369, 1–13. https://doi.org/10.1002/smll.202307369. Solids 149, 104341. https://doi.org/10.1016/j.jmps.2021.104341.
Wang, Y., Zeng, Q., Wang, J., et al., 2022. Inverse design of shell-based mechanical Yin, H., Zhang, W., Zhu, L., et al., 2023. Review on lattice structures for energy
metamaterial with customized loading curves based on machine learning and absorption properties. Compos. Struct. 304, 116397 https://doi.org/10.1016/j.
genetic algorithm. Comput. Methods Appl. Mech. Eng. 401, 115571 https://doi.org/ compstruct.2022.116397.
10.1016/j.cma.2022.115571. Yu, S., Chai, H., Xiong, Y., et al., 2022. Studying Complex Evolution of Hyperelastic
Wilt, J.K., Yang, C., Gu, G.X., 2020. Accelerating auxetic metamaterial design with deep Materials under External Field Stimuli using Artificial Neural Networks with
learning. Adv. Eng. Mater. 22 (5), 1901266. https://doi.org/10.1002/ Spatiotemporal Features in a Small-Scale Dataset. Adv. Mater. 34 (26), 2200908.
adem.201901266. https://doi.org/10.1002/adma.202200908.
Wu, Y., Mao, Z., Feng, Y., 2023. Energy absorption prediction for lattice structure based Yu, G., Xiao, L., Song, W., 2022. Deep learning-based heterogeneous strategy for
on D2 shape distribution and machine learning. Compos. Struct. 319, 117136 customizing responses of lattice structures. Int. J. Mech. Sci. 229, 107531 https://
https://doi.org/10.1016/j.compstruct.2023.117136. doi.org/10.1016/j.ijmecsci.2022.107531.
Xiao, L., Shi, G., Feng, G., Li, S., Liu, S., Song, W., 2024. Large deformation response of a Zheng, L., Karapiperis, K., Kumar, S., et al., 2023. Unifying the design space and
novel triply periodic minimal surface skeletal-based lattice metamaterial with high optimizing linear and nonlinear truss metamaterials by generative modeling. Nat.
stiffness and energy absorption. Int. J. Solids Struct. 296, 112830. https://doi.org/ Commun. 14, 7563. https://doi.org/10.1038/s41467-023-42068-x.
10.1016/j.ijsolstr.2024.112830. Zok, F.W., Latture, R.M., Begley, M.R., 2016. Periodic truss structures. J. Mech. Phys.
Solids 96, 184–203. https://doi.org/10.1016/j.jmps.2016.07.007.

17