Received: 20 March 2024 Revised: 8 September 2024 Accepted: 9 September 2024

DOI: 10.1002/pc.29082

REVIEW ARTICLE

A review on recent applications of machine learning

in mechanical properties of composites

Yi Liang | Xinyue Wei | Yongyue Peng | Xiaohan Wang | Xiaoting Niu

College of Materials Science and

Engineering, Northeast Forestry Abstract
University, Harbin, China Composites are undergoing extensive research and utilization due to their
excellent mechanical properties, driven by human needs. Traditionally, the
Correspondence
Xiaoting Niu, College of Materials Science research methods in materials science predominantly rely on empirical theory
and Engineering, Northeast Forestry or experimental trial and error approaches. However, the increased complexity
University, Harbin 150040, China.
of composite materials results in a greater intricacy in their mechanical behav-
Email: xiaotingniu@nefu.edu.cn
ior. Consequently, the utilization of traditional research methods may not
Funding information achieve sufficient efficiency. Materials science is rapidly transitioning into a
Science and Technology Innovation
Platform Construction Project in
data-driven era, with machine learning (ML) emerging as a potent tool to expe-
Heilongjiang Province, Grant/Award dite materials development and enhance properties prediction. Significant
Number: DL41320601 advancements have been achieved in the application of ML to the study of
composite mechanics. In this review article, we elucidate various ML methods
employed in the construction of constitutive models for isotropic and aniso-
tropic composites, and delve into the research on construction ML models that
leverage input data derived from composite processes, structures, and environ-
mental conditions to predict material mechanical properties. Additionally, we
summarize recent noteworthy ML applications in composite design and opti-
mization. Finally, possible prospective viewpoints are proposed for future
development, with the aim of providing essential scientific guidance for
advancing material science and technology through ML.
Highlights
• Machine learning can address complexity in constitutive model of the aniso-
tropic composites.
• Machine learning predicts mechanical properties of composites well by pro-
cess and structure.
• Machine learning enhances efficiency in inverse design to optimize
composites.
• Limitations, challenges, development trends of ML in composites.

KEYWORDS
composites, constitutive model, inverse design, machine learning, mechanical properties

Polymer Composites. 2025;46:1939–1960. wileyonlinelibrary.com/journal/pc © 2024 Society of Plastics Engineers. 1939

 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1940 LIANG ET AL.

1 | INTRODUCTION

Composites consist of two or more substances with distinct

physical and chemical properties. When combined, they
exhibit unique characteristics different from those of indi-
vidual components. Typically, each type of composite
material comprises a matrix (continuous phase) and a
reinforcement material (discontinuous phase). Common
examples include wood-based composites, metal-based
composites, ceramic-based composites, and organic com-
posites, among others. With the rapid advancement of mod-
ern manufacturing, composites have found extensive
applications across various industries. Their ability to over-
come limitations associated with traditional single mate- F I G U R E 1 Changes in the number of publications that
rials, particularly in terms of physical and mechanical include “ML” and both “ML” and “material” keywords in abstracts
performance, has made them integral in fields, such as elec- published from January 2000 to December 2022.3
tronics, aerospace, medicine, automotive, and mechanical
processing.1 In the realm of materials science and engineer-
ing, there is an ongoing exploration of composites with discovery of materials with exceptional performance and
enhanced mechanical properties to meet the increasingly applicability.3 ML aims to simulate human learning
intricate and stringent demands of applications. The inves- behavior using computers, that is, to establish logical
tigation of material mechanical property is a focal point in relationships through accumulated experience to make
efficiently utilizing and fabricating these composites. In the judgments and decisions under various conditions. ML
process of mechanical research, researchers often conduct entails computers leveraging algorithms to learn from
experiments and theoretical analyses by proposing mathe- experience, with the objective of enhancing performance
matical models. These models typically involve regular geo- in specific tasks. The experience is derived from collected
metric shapes, boundary conditions, and material response data, which is processed into a computer-readable format
functions expressed through explicit expressions. They termed “features”. Tasks are categorized into regression
serve to abstract real and intricate mechanical problems and classification based on whether the predicted values
into mathematical models, and then the analytical solu- are continuous or discrete, respectively. Performance is
tions of the problems are sought. However, as the complex- measured by metrics assessing the error between pre-
ity of composites increases, their mechanical behavior dicted values and actual data. Constructing a fundamen-
becomes more intricate, posing challenges in solving prob- tal ML model involves at least data collection, feature
lems with more complex geometric models and boundary engineering, algorithm selection, and model evaluation.
conditions. Furthermore, the traditional “trial and error” Machine learning primarily include supervised learning,
method of testing composite material performance or unsupervised learning and reinforcement learning.5 Among
designing new materials relies heavily on experiential guid- them, supervised learning is predominantly applied for
ance and experimentation, consuming significant time and materials research. A crucial step in applying ML is to estab-
resources. This approach may also result in uncertainties lish a relevant model that accurately characterizes the rela-
and randomness. Machine learning (ML) emerges as a tionship between features and material properties. During
potent tool for unveiling hidden patterns in data, with flexi- this process, mapping models for objects and targets are con-
ble functionalities and the capacity to understand various structed based on known material data. Various ML
problems. It plays a pivotal role in modern materials sci- methods can be employed to establish such mappings.
ence research by efficiently processing multidimensional There are a lot of choices one can make, Kevin et al.4 give a
data and providing insights into the intricate mechanical non-exhaustive overview of the supervised learning algo-
behavior of advanced composites.2 rithm landscape, shown in Figure 2. Support vector machine
In the past two decades, the application of ML in (SVM) holds unique advantages in addressing challenges,
materials science has demonstrated significant vitality, as such as small-volume data, nonlinearity, and high-
depicted in the Figure 1 illustrating the trend of the dimensional pattern recognition. Decision tree is a predictive
increasing number of relevant publications each year model that recursively splits datasets, represented by a tree
(data sourced from search results on Web of Science). structure to depict decision pathways. Each node represents
The data-rich field of materials has seen the remarkable a feature or attribute, and each branch corresponds to a
effectiveness of ML as a powerful tool for accelerating the decision or outcome. The organized segmentation of data
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1941

FIGURE 2 Overview of the ML algorithm landscape by Kevin et al.4

and final decision derivation are achieved by evaluating spe- the depth of the ANN is increased, the network gains the
cific feature values at each node. It can deal with data with capacity to transform data in increasingly intricate ways,
missing attributes and has good interpretability, but is easy thereby augmenting the complexity of the learned rela-
to overfit. Random forest (RF) integrates its predictive tionship between inputs and outputs by effectively intro-
results by constructing an ensemble of decision trees and ducing additional variables. Simultaneously, it found
combining them through voting or averaging. K-Nearest applications in materials research. Common deep learn-
Neighbors (KNN) is an algorithm that requires a lot of com- ing methods include multilayer perceptron, convolu-
putation and high sample balance. Artificial neural network tional neural network (CNN), recurrent neural networks
(ANN) abstracts neurons as nodes, employing function (RNN), and their derivatives.8
approximation and estimation methods to transform input Machine learning can enhance the theoretical advance-
samples into desired output data through numerous inter- ment of materials science by leveraging its application in
connected nodes. Artificial neural network exhibits strong analyzing the mechanical properties. By extensively explor-
self-learning and fault-tolerant capabilities, albeit requiring ing and scrutinizing vast amounts of material data, intricate
large datasets and having relatively weaker interpretability. correlations between material mechanics and structure that
Ensemble model is a machine learning approach that com- elude traditional theories can be unveiled, thereby enriching
bines multiple individual models to improve overall predic- and refining the theoretical framework of materials science.
tive performance through methods, such as averaging, Machine learning enables the development of more precise
voting, or stacking. Commonly used ML methods in mate- constitutive models for composite materials, revealing pat-
rials science include SVM, decision tree, RF, ANN and so terns in their mechanical behavior under various condi-
on, each having its advantages and limitations.6 tions. Additionally, ML demonstrates remarkable capability
At the beginning of the 21st century, deep learning in handling high-dimensional and multi-variable material
propelled ML back into the forefront of research, emerg- data. By comprehensively analyzing the composition, struc-
ing as one of the most widely applied techniques in ture, and performance of composite, their inherent charac-
ML. It represents an extension of the traditional neural teristics and interaction mechanisms can be revealed,
network architectures, incorporating more hidden layers providing a solid theoretical foundation for material design
and a more diverse range of activation functions, which and optimization. In terms of practical guidance, ML can
map the input to the output response of each neuron.7 As expedite the research process of composite materials while
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1942 LIANG ET AL.

broadening their application domains. It enables rapid per- composites.9,10 Traditional constitutive models are typi-
formance prediction and facilitates the screening of poten- cally represented by explicit mathematical functions, often
tial candidate materials from a wide range of material requiring modeling based on inherent physical mecha-
combinations, thereby minimizing the need for unnecessary nisms of materials combined with certain assumptions.
repetition. The integration of ML enables holistic design The model parameters are then calibrated through experi-
based on the mechanical properties, leading to the creation mental testing, making the establishment of the model
of high-performance materials suitable for fields, such as straightforward when mechanical characteristics are
aerospace engineering, electronic information technology, clearly understood. However, when dealing with complex
and construction. composites, studying the mechanical behavior of intricate
In this era of science propelled by data, ML has demon- nonlinear systems poses a significant challenge in scien-
strated the capability to address numerous complexities tific research. This implies that traditional approaches face
within the field of composites mechanics. In recent years, challenges in dealing with the emerging complexity, since
researchers have been actively exploring the use of ML to both empirical and theoretical constitutive models require
establish constitutive models for composites. Furthermore, the incorporation of an increasing number of parameters,
ML-based methods have gained widespread attention in thereby restricting their practical utility.
prediction of composite mechanical properties, with many Machine learning possesses remarkable capabilities
scholars constructing highly accurate and robustly general- in describing complex mapping relationships and simu-
izable predictive models. Moreover, ML-guided design and lating intricate physical constitutions. Given a sufficient
development of composites have enabled the creation of amount of training data, ML models can learn complex
new materials with target properties. This research direc- nonlinear mappings from high-dimensional feature vec-
tion has achieved substantial progress, and this review aims tors to the desired outputs.11 The construction of consti-
to summarize the advancements that ML has made in the tutive models in ML primarily relies on experimental
field of composites mechanics. This includes the construc- data and enables the establishment of highly nonlinear
tion of material constitutive models, prediction of material relationships between composite variables. In recent
mechanical properties, and the material design and optimi- years, ML has become a focal point in the research efforts
zation. The constitutive model construction can provide a on constitutive models for various materials.12–14 There-
theoretical basis, enabling the prediction of mechanical fore, many studies have utilized ML methods to construct
properties. This, in turn, facilitates guidance for material constitutive models for composites, leveraging the flexi-
design and optimization. The systematic research frame- bility and applicability of ML techniques.
work presented aims to deepen the understanding of mate-
rial mechanics and achieve targeted material design. These
studies indicate that ML holds significant potential in han- 2.1 | Isotropic composite
dling mechanical data for composites and elucidating com-
plex or ambiguous mechanical mechanisms. Isotropic composites exhibit uniformity, with the internal
structure and components evenly distributed in different
directions, without specific orientations or arrangements.
2 | CONSTRUCTION OF This uniformity results in similar performance characteris-
CONSTITUTIVE MODELS tics in all directions. The constitutive model for these
materials can be described by a nonlinear curve, and ML
Constitutive models are frameworks used to elucidating demonstrates a powerful capability to approximate complex
the material mechanical behavior, thereby facilitating to nonlinear relationships in the study of material constitutive
understand their physical and mechanical responses relations.15 Several studies employ ML methods to construct
under diverse conditions. Faced with complex materials, models that approximate the complex nonlinear functions
continuous improvement of constitutive models enhances or experimental curves of composites, thereby describing the
our comprehension of the intricate interplay between constitutive relationship of these materials. These researches
material structure and property. This in-depth exploration predominantly center on utilizing the influencing factors of
of constitutive models is crucial for advancements of composite material mechanical behavior to develop ML
materials science. The study of constitutive models for models for predicting stress–strain curve. It aims to explore
composites is significant in material research as it not only the relationship between these factors and composite
facilitates the explanation and prediction of material prop- mechanical behavior, employing implicit representations
erties but also expedites the discovery and application of instead of explicit functional constitutive models.
materials. Currently, this field has seen the development The thermal deformation behavior and constitutive
of numerous theoretical and empirical models for relations of metal matrix composites have been investigated
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1943

In addition to process conditions, scholars also utilize

structural parameters of composites as input parameters in
ML models for predicting stress–strain curve. The volume
fraction of composites is a commonly used structural param-
eter, which quantifies the relative proportion of each com-
ponent's volume to overall volume of the entire material. It
directly influences the mechanical property and behavior of
the composites. The distribution of carbon fibers in carbon
fiber nonwoven cloth composite (CFNWC) differs from that
in traditional woven carbon fiber composite. The carbon
fibers are randomly distributed in a disordered manner
within a two-dimensional plane in the material. As a result,
F I G U R E 3 Schematic illustration of the BPANN CFNWC exhibits similar performance in all directions, mak-
architecture.16 BPANN comprises an input layer, one or more ing it an isotropic composite. The volume fraction of carbon
hidden layers, and an output layer. The input layer consists of three fiber, as a crucial microstructural parameter, exerts a signifi-
nodes representing temperature, strain and strain rate, while the cant influence on the material properties. He et al.20 initially
output layer consists of one node representing stress. The selection applied principal component analysis (PCA) to reduce the
of the number of hidden layers and nodes in the hidden layer is dimensionality of stress–strain curves for CFNWC. Subse-
determined through experimentation. The weights and thresholds quently, an ANN model was constructed using carbon fiber
are adjusted by forward propagation of experimental data and back volume fraction as input parameters to accurately predict
propagation of errors to minimize target errors.
the material mechanical behavior. In addition, the study
conducted a comparison between BPANN and Bayesian
using ML methods. Notably, some scholars have estab- neural network (BNN), revealing that BNNs exhibit signifi-
lished constitutive models for these composites based on cantly superior performance in stress–strain curve predic-
backpropagation artificial neural networks (BPANN). tion. BPANN requires a large training set to achieve high
During modeling of these studies, the input parameters accuracy, otherwise it may lead to overfitting of predictions.
include temperature, strain rate, and strain, while stress is In contrast, BNN weights are probability distributions that
regarded as the output variable. The BPANN structure in describe weight uncertainty, enabling the model to provide
these studies is shown in Figure 3. Zhong et al.17 designed both predicted values and prediction uncertainties. More-
a BPANN with a single hidden layer to establish the consti- over, BNN fits posterior distributions rather than annotated
tutive equations for (2vol%Ti5Si3 + 5vol%TiBw)/TC11 com- values, making it highly robust. The CFNWC dataset in the
posites. Y.S. Wang et al.18 employed a BPANN with a aforementioned study is small in size, so the BNN perfor-
double hidden layer to establish the constitutive model for mance is better.
Ti-based metallic glassy matrix composite. Yu and collabo- The ML has been employed to study the constitutive
rators19 attempted to utilize BPANN with double and triple relationship of isotropic composite. Several scholars have
hidden layers to construct the constitutive models for developed constitutive models utilizing ANN. Different
Ta-rich particle reinforced Zr-based bulk metallic compos- scholars adopt ANNs with different hidden layers. Gener-
ites. The results indicate that the ANN with a double hid- ally, multi-layer ANNs exhibit stronger generalization
den layer exhibited stronger generalization capabilities abilities. However, there are still challenges that need to
while ensuring accuracy. This suggests that shallow ANN be addressed, such as the time-consuming determination
may be sufficient to address the issues related to isotropic of optimal structure and the lack of physical interpretation
metal matrix composites, and more complex network struc- in knowledge representation. In addition to process condi-
tures could potentially lead to the problem of overfitting. tions, scholars have also incorporated volume fraction, a
There are some unresolved issues in composite material structural parameter of composites, as an input for ML
constitutive models based on ANN. The optimal structure models to predict stress–strain curves. In the study of com-
of ANN models in these studies often requires time- posites, it is necessary to take multiple factors into account
consuming trial-and-error methods, such as determining to select the appropriate ML model and parameters.
the number of neurons, selecting transfer functions, and
choosing optimization algorithms. Additionally, knowledge
is represented through weight matrices and biases in these 2.2 | Anisotropic composite
models, lacking specific physical meanings and clear
expressions, making this method suitable for numerical Anisotropic composites exhibit significant differences in
analysis but challenging for theoretical analysis. various directions, often arising from the non-uniformity
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1944 LIANG ET AL.

of microstructures or orientations. The structure and their strengths and limitations. Given the diversity in
mechanical performance of anisotropic materials are experimental objects, datasets, and results among studies,
influenced by processing conditions. ML research is determining the optimal ML model for investigating the
underway to predict the stress–strain curve of anisotropic composite mechanical properties poses a formidable chal-
materials based on these processing conditions. The per- lenge. The variables influencing the mechanical behavior
formance of thermoplastic composites is profoundly of composites are numerous and exhibit highly nonlinear
affected by their processing conditions. In order to under- characteristics. Consequently, the establishment of a uni-
stand their behavior under different processing condi- versally applicable ML-based constitutive model poses a
tions, Mohammed et al.21 employed ANN to predict the significant challenge.
stress–strain relationship of unidirectional glass fiber- However, owing to strong generalization capabilities
polypropylene composites processed at various cooling and advantages in multi-output predictions, ANN have
rates. Hamed and Reza22 utilized ML methods to forecast become the mainstream method for establishing constitu-
the complete stress–strain behavior of polymer-based tive models of composites and has become the most com-
composites by constructing an ANN based on the compo- monly used ML approach in this field. Additionally, in
sition, environment, and processing conditions. Achiev- comparison to isotropic materials, ANN frameworks
ing uniform effects in anisotropic materials through applied to anisotropic materials tend to exhibit higher
processing conditions is challenging. Adjusting the complexity and involve a greater number of selected input
structure can modify anisotropic composites to meet parameters. Shack et al.26 employed ANN to construct a
mechanical property requirements in different directions. stress–strain curve prediction model for steel-confined
Therefore, more attention was given to the structure. ultra-high-performance concrete (UHPC), employing a
Additionally, due to the need to consider the mechanical network composed of 3 hidden layers and 21 hidden neu-
behavior in different directions, the feature engineering rons. When selecting input parameters, they considered
process in constructing ML models for anisotropic com- parameters influencing UHPC failure modes, axial bearing
posites is typically more complex. Constitutive models for capacity, and stress–strain behavior. Additionally, in order
anisotropic materials require thorough consideration of to maintain consistency with existing design theories, they
multiple factors and often necessitate a larger dataset for determined 15 input parameters based on model parame-
fitting compared to their isotropic counterparts. ters and design specifications. Other researchers utilized
At present, constructing constitutive models through 23 material features related to steel fiber parameters, PVA
ML involves a diverse range of input parameters and fiber parameters, and material composition as inputs to
frameworks. However, there is a lack of standardized build an ANN with 2 hidden layers for predicting the ten-
methodology to guide the selection of these parameters sile stress–strain curve27 and compressive stress–strain
and frameworks. Furthermore, there is no universally curve28 of hybrid fiber-reinforced concrete.
applicable and application-independent ML algorithm Given the more complex mechanical behavior of aniso-
recognized as the most effective. As a result, various clas- tropic materials, more sophisticated ML models are typically
sical ML methods are applied to construct constitutive employed. In addition to conventional ML methods, some
models for different composites. Hasanzadeh et al.23 used researchers have applied deep learning, a rising approach in
ML methods to simulate the compression stress–strain recent years, to the study of the constitutive relationship of
curve of basalt fiber-reinforced high-performance con- composites. Deep neural networks represent a powerful ML
crete. They employed three methods, including polyno- strategy, and input features are gradually combined through
mial regression (PR), support vector regression (SVR), multiple connected hidden layers to generate an output.29
and linear regression (LR). Experimental results indi- Syntactic foams exhibit a complex internal microstructure
cated that the PR exhibited higher accuracy and reliabil- with discontinuities. In order to forecast the stress–strain
ity in their study. Gautham and Saptarshi24 applied three behavior of composite foams, Haotian and Pavana30 intro-
ML methods (K-means, gaussian mixture model and duced a parameterization-based neural network, and used
SVM) to the study of the constitutive relationship of this trained neural network model to link the microstructure
cement-based composites. The results showed that K- of composites to nonlinear responses. Transitioning to a dif-
means performed best in the available dataset, while the ferent aspect of exploring structure and mechanical
SVM algorithm might be more suitable for larger and responses by deep learning, the stress–strain curve of com-
more complex datasets. Gayatri et al.25 used ANN posites was predicted by combining PCA and CNN. Charles
and gaussian process regression models to predict the et al.31 used a dataset comprising 100,000 different micro-
woven composite stress–strain curve. Both algorithms structures of composites and their corresponding stress–
demonstrated strong correlations between experimental strain curve to train and evaluate the performance of CNN.
data and predicted data. Different ML algorithms have A flow diagram is shown in Figure 4. The process begins by
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1945

FIGURE 4 Workflow of composites constitutive relationship constructed by Charles et al.31

preprocessing the target parameters (stress–strain curve), Carbon fiber fabric reinforced woven composites
using PCA to reduce their dimensionality. These curves present a higher level of complexity compared to tradi-
are transformed into PCA coordinate representations, fit- tional unidirectional composites. In response to this com-
ting a PCA model to the training set, and applying the plexity, a combined multiscale method and ML approach
trained model to compress both the training and testing was proposed. Dug-Joong et al.35 implemented this
sets. Next, composite designs are used as input for the approach, using 900 RVEs to train a deep neural net-
CNN, making predictions in PCA coordinate space. The work, aiming to predict stress–strain curve of carbon
loss function is weighted based on the explained variance fiber fabric-reinforced woven composites. They per-
ratio of each component, with the input of CNN being a formed a micro-mechanical multiscale analysis to con-
binary matrix representing the composite design (0 for soft duct progressive damage analysis. The analysis used
block, 1 for hard block). Finally, the CNN predicts unseen stress amplification factors to transfer stress between
composite designs, and PCA predictions are transformed microscale and mesoscale simulations, applying respec-
back into conventional stress–strain curves. To analyze the tive failure criteria for fibers and matrix at the microscale.
results, the predicted material descriptors are compared This resulted in a stress–strain curve database for differ-
with descriptors derived from both predicted and actual ent yarn geometries and composition characteristics
stress–strain curve. In contrast to Charles et al., Gana- under tensile, compressive and shear loading.
pathi32 incorporated 2-point quantification and dimension- The stress–strain curve prediction of fiber reinforced
ality reduction (PCA) of microstructure images in building plastic composites was addressed through a parallel
CNN for predicting composite stress–strain curve effort, wherein Bassam proposed a multiscale data-driven
(as shown in the Figure 5). Similarly, Do-Won et al.33 pro- modeling framework. Notably, this framework seam-
posed a method for studying the transverse mechanical lessly integrated a deep recurrent convolutional neural
behavior of unidirectional composites using a CNN. Subse- network.
quently, they applied the same method, combined with In a parallel effort, a multiscale data-driven modeling
PCA, to construct a CNN for predicting stress–strain curve framework was proposed by Bassam36 to predict stress–
of unidirectional composites.34 They first generated repre- strain curve of fiber reinforced plastic. The noteworthy
sentative volume elements (RVEs) with arbitrarily distrib- aspect of this framework lies in its integration of a deep
uted circular fibers, converted them into corresponding recurrent convolutional neural network. The stress–
binary microstructure images, and used these images as strain dataset exhibits sequential features, thus prompt-
inputs for model training. The stress–strain curve for each ing the investigation of RNN for analyzing the mechani-
RVE were obtained from finite element (FE) simulations, cal behavior of composites. Since traditional RNN limited
considering the interfacial debonding phenomenon. history storage, long short-term memory (LSTM) neural
Additionally, to efficiently learn stress–strain curve in a network was developed as a solution. Qiang et al.37
lower-dimensional space, the study employed PCA. applied LSTM to study the stress–strain relationship of
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1946 LIANG ET AL.

FIGURE 5 Workflow for predicting stress–strain curve of unidirectional composites by Ganapathi.32

unidirectional metal matrix composites. Jaewook et al.38

used feature engineering methods such as PCA and
recursive feature elimination with cross validation to find
F I G U R E 6 The mechanics system of ML as a subsystem. I s
the minimum and optimal feature sets required to predict
and Os represent the input and output, respectively. The ML input
the tensile behavior of polymer matrix composites
(ik ) is derived from the system input ðI s Þ, while the ML output (ol )
(PMCs). They trained LSTM and feedforward neural net-
is utilized to determine the system output (Os ). When solving the
work models using the optimal feature set and tensile test constitutive relation problem, ik represents strain and ol represents
data from 1270 PMCs, successfully achieving the predic- stress.
tion of tensile stress–strain curve. Yunce et al.39 built a
RNN (gated recurrent unit) to accurately predict the non-
linear constitutive relationship of mesh/membrane com- biases that may impede ML models in effectively captur-
posites. The model was constructed with input data as ing the true constitutive features of composites. While
composite descriptors and deformation paths, and output some constitutive model data is directly based on experi-
data as stress and fracture strength. mental data, these relationships tend to be relatively sim-
The success of ML relies significantly on the ability ple, such as one-dimensional stress–strain relationships.
to collect and interpret large datasets. Training ML To establish intricate constitutive models using ML with
models often requires a substantial amount of data.29 limited data availability, some researchers integrated ML
M. Mozaffar et al.40 even utilized data from up to 15,000 with other models to form a coupled mechanical system.
loading paths to train an RNN model for describing a The framework of mechanics system of ML as a subsystem
two-dimensional elastic–plastic constitutive model. In sit- is shown in the Figure 6. Fei et al.45 maximized the multi-
uations where it is impractical to efficiently conduct suffi- functional FE analysis capabilities of Abaqus and deep neu-
cient experiments to obtain data, some researchers ral network's powerful learning abilities. They integrated
employ transfer learning methods,41,42 while others build deep neural network with Abaqus to form a coupled
datasets by collecting data from other journal articles and mechanical system used to learn the progressive damage
technical reports.26 Another common approach involves constitutive relationship of fiber-reinforced composites. They
generating data through numerical simulations, such as also introduced a FE coupled positive definite deep neural
FE analysis43 and molecular dynamics.44 However, the network mechanical system to learn the nonlinear constitu-
utilization of numerical simulations to generate data tive relationships. This approach was employed to investi-
necessitates certain assumptions, which could introduce gate the damage constitutive relationship in unidirectional
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1947

fiber-reinforced composite laminate plates.46 Xin Liu et al.47 some researchers are leveraging the accuracy and flexibil-
also employed an ANN model as a component or subsystem ity of ML techniques to efficiently predict the material
within a mechanical system. The inputs and outputs of the mechanical properties. Data is gathered through mechan-
system were obtained from experiments, where the input ical tests, FE simulations, and the collection of data from
for the ANN came from the system's input, and the output other literature sources during the ML model building
of the ANN contributed to the system's output. Instead of process. Researchers primarily use data related to the
constructing highly nonlinear relationships using neural processing conditions and structural parameters of com-
network, the goal of this study was to develop an ANN- posites as inputs for the model, with various mechanical
enhanced system to train the model using measurable data properties serving as the final outputs. With the self-
within the system, thereby expanding the training dataset learning and self-processing capabilities of ML methods,
for general mechanical problems. Ultimately, they used this models are capable of acquiring the mapping relation-
developed system to learn the constitutive relationship of ships between inputs and outputs within dataset. This
fiber reinforced composites and the failure initiation criteria. approach provides a more flexible and efficient tool for
The constitutive relationship of anisotropic materials material design and optimization.
has been effectively established using ML. Due to the
intricate mechanical behavior exhibited by anisotropic
materials, more sophisticated ML models, such as CNN 3.1 | Prediction of mechanical properties
and RNN in deep learning, are frequently employed. based on process
Furthermore, studies combining multiscale methods with
ML have been conducted. In addition to encountering The formation of composites involves the combination of
similar challenges faced in isotropic material research, a two or more distinct materials, typically encompassing
scarcity of data volume is also observed. The efficacy of fibers, matrices, and other supplementary components.
ML heavily relies on extensive datasets. Hence, various Processing parameters directly impact the properties and
approaches are adopted for data acquisition. When con- interactions of individual components within composites,
fronted with limited data availability, some researchers thus influencing their mechanical performance. For
resort to transfer learning or numerical simulation tech- instance, processing parameters can affect the final per-
niques or integrate ML with other models to construct formance of the polymer composites, such as tempera-
coupled mechanical systems. For anisotropic composites, ture, screw speed, and filler weight.48 The type of fibers
advanced and complex ML methods and strategies are and mixed design parameters significantly influence the
required. Emphasizing the significance of data acquisi- mechanical properties of cement based composites,49 and
tion and processing along with model integration and material composition also plays a decisive role in the
coupling provides a promising direction toward addres- mechanical performance of composites.50 Therefore,
sing the issue of limited data. many researchers use the processing parameters of mate-
rial preparation as input parameters to construct ML
models for predicting the mechanical properties. Simulta-
3 | PREDICTION OF neously, they utilize the established predictive models to
MECHANICAL PROPERTIES correlate the complex relationship between processing
and performance. This aids in optimizing processes and
Material descriptors can be obtained from material con- preparing composites with enhanced properties.
stitutive models, such as modulus, strength, and tough- As a powerful machine learning tool, ANN have been
ness. However, constructing accurate constitutive models successfully employed in this field. In the comparative study
is quite complex, and even with advanced models, it is of using ANN and statistical method (multiple linear regres-
challenging to simulate all scenarios. Therefore, in some sion) to predict particleboard mechanical properties by
cases, directly predicting and assessing the mechanical pressing conditions, experimental results showed that ANN
properties of materials based on application requirements outperformed multiple linear regression.51 Wood-based
has become increasingly important. Currently, obtaining composites have been continuously evolving and finding
the mechanical properties typically involves experimental applications in various engineering fields. These composites
testing methods, such as tension, shear, impact, and are formed by integrating wood veneers, wood fibers, or par-
fatigue tests on composite samples. Alternatively, numer- ticles through various adhesives, resulting in composites,
ical simulation using FE analysis is employed to predict such as fiberboard, particleboard, plywood, and wood-
composite mechanical properties. However, these polymer composites. In recent years, numerous studies have
methods still face challenges such as being destructive, been conducted for predicting the mechanical properties of
lacking precision, and being time-consuming. As a result, diverse wood-based composites. With processing conditions
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1948 LIANG ET AL.

(time, pressure and temperature) as inputs, ANN could be ensemble algorithm was also employed. XGBoost is an
employed to predict the screw withdrawal strength of ensemble algorithm based on boosting classification and
particleboard,52 the Young's modulus and toughness regression tree. Wei et al.62 employed four ML methods,
of mycelium-based wood composites,53 the modulus of elas- including SVR, RF, BPANN, and gradient boosting ensem-
ticity (MOE) and modulus of rupture (MOR) of fiberboard.54 ble learning. They used component content, physical
Other than that, additive content was used as input to build properties of graphite-based nanomaterials, and curing
an ANN to predict the tensile strength of the foamed wood conditions as inputs to construct ML models for the pre-
fiber reinforced thermoplastic starch composite.55 Exploring diction of the uniaxial compressive strength in graphite-
laminated veneer lumber, M. Nazerian et al.56 developed an based nanomaterial-reinforced cementitious composites.
ANN for predicting the MOE and MOR, with material pro- The findings demonstrated that XGBoost exhibited supe-
cessing conditions, including press time, nanocellulose con- rior predictive performance, but this model introduced
tent, urea-to-formaldehyde (UF) molar ratio, and weight additional complexity and increases computational time.
ratio of starch glue to the UF resin, as inputs. However, these comparative studies employing multi-
Extended to other composites, ANN also plays a pow- ple ML methods did not thoroughly investigate the
erful role. The bending strength and breaking load of underlying factors contributing to variations in the per-
fiber cement composites were predicted by flocculation formance of such methods. It is noteworthy that the
process parameters.57 Evaluating zea fiber–polyester choice of a particular machine learning method varies
composites, Balaji et al.58 constructed an ANN using pro- across studies, and the accuracy, efficiency, and effective-
cessing parameters to evaluate the tensile, bending, and ness of ML methods also exhibit variations among differ-
impact properties. The processing parameters included ent studies. Therefore, selecting an optimal machine
different ratios of molding pressure, fiber length, and learning method remains a persistent endeavor.
fiber content by weight percentage. Udaya et al.59 To improve prediction efficiency, several recently
selected reinforcement material content, sintering time, developed algorithms have been used to optimize these
and ball milling time as input parameters to train ANN ML hyperparameters. Non-dominant sorting genetic
models for predicting the density and hardness of powder Algorithm-II (NSGA2) is a computationally efficient and
metallurgy-produced multi walled carbon nanotubes and highly convergent method for multi-objective optimiza-
fly ash/Al composites. tion. Beilong et al.63 proposed a multi-objective optimiza-
SVR demonstrates remarkable nonlinear mapping and tion SVR using NSGA2. They used particle bonding
generalization ability in predicting composite mechanical process parameters as inputs to predict the MOE, MOR
properties. Mohammad and Vahab60 used material mixing and internal bonding strength of particleboard. The
proportions as inputs to build ANN and SVR models to results demonstrated that the prediction error of
predict the concrete compressive strength. The results NSGA2-SVR was significantly lower compared to SVR.
demonstrated that SVR exhibited superior predictive capa- ML models of recycled aggregate concrete were devel-
bilities compared to ANN. Unfortunately, the specific rea- oped by Yiming and Cise64 to predict the compressive
sons why SVR was superior to ANN were not explained in strength. Model inputs included the mixing proportions
detail in this study. The utilization of SVR in composites, and properties of recycled aggregates. They employed
nevertheless, poses certain limitations. The process of ker- two standard algorithms, namely SVR and ANN, as well
nel function selection in SVR, relying on trial and error, as two optimized hybrid models, namely Particle Swarm
poses a drawback. Moreover, the SVR performance is diffi- Optimization-based SVR (PSO-SVR) and Gray Wolf
cult to explain due to the complexity in mapping a non- Optimizer-based SVR (GWO-SVR). The results indicated
linear input space to a high-dimensional feature space. that GWO-SVR achieved the best prediction accuracy
The multitude of node sizes and internal connections with an R2 of 0.9056.
renders ANN and SVR models as opaque “black boxes”. With the advancement of ML technologies, many
Therefore, it is difficult to generate a transparent mathe- researchers have applied updated ML methods to accu-
matical formula from these models to depict the func- rately forecast the composite mechanical properties.
tional correlation between the inputs and outputs. To Soroush et al.65 created an artificial language to represent
overcome this common problem, RF can be chosen. the physicochemical information of specific mixtures and
M. Schubert et al.61 used multiple processing parameters their components. They devised a character-level N-grams
like drying temperature and pressing temperature as approach for feature extraction, as well as introduced a
inputs to construct ML models predicting the compres- method for automating the configuration of deep learning
sive strength of wood fiberboard. They employed RF, models. This methodology was used to forecast the com-
ANN and SVM, with RF showing the best generalization pressive strength of complex concrete mixtures. Mostafa
performance. In addition to aforementioned techniques, et al.66 investigated the utility of deep learning in predicting
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1949

the mechanical properties of basalt fiber and nanoclay rein- and density as inputs to train an ANN. The trained ANN
forced polymer concrete composite. Instead of constructing was then employed to predict the MOR and MOE of
a dataset based on available experimental data in the litera- structural plywood. In order to predict the apparent com-
ture, all training and validation data were prepared based pression modulus of cork-rubber composites, Helena
on experimental investigations conducted within this et al.69 train an ANN model that employed the shape fac-
study. The weight percentages of components were utilized tor and hardness Shore A of a cork-rubber block as input
as input variables to build a deep learning model, aiming parameters.
to predict the material properties (flexural, compressive, In the investigation of the correlation between the
and splitting tensile strengths). Bouchaib et al.67 used structural characteristics and mechanical properties of
nanoclay and compatibilizer content, as well as processing fiber-reinforced composites, researchers often combine
parameters (screw speed and feed rate), as inputs to numerical analysis of RVEs with ML techniques.
construct a backpropagation deep neural network for pre- Researchers generally first establish the RVEs of the com-
dicting the tensile modulus and tensile strength in clay- posites, and then derive the structure parameters from
reinforced polyethylene nanocomposites. them. These parameters are subsequently used as inputs to
Various ML methods have advantages and disadvan- construct ML models for predict the mechanical properties.
tages in predicting composite mechanical properties. For example, Xiaoxuan et al.70 investigated the mechanical
Among them, ANN have been successfully applied in this behavior of unidirectional fiber-reinforced polymer com-
domain. However, similar to SVR, they are often consid- posite laminate under transverse loading by deep neural
ered as “black box” models. SVR has significant non- network. They initially collected a dataset generated from
linear mapping and generalization capabilities, but it has the discrete element method simulations of 2000 RVEs.
limitations in kernel function selection. On the other The deep neural network model was then constructed with
hand, RF can overcome the “black box” issue and has fiber radius and fiber volume fraction as inputs for predict-
demonstrated good generalization performance in certain ing the transverse Young's modulus and tensile strength.
studies. Ensemble algorithms exhibit superior predictive ANN is also a widespread employed ML tool in the
performance but introduce complexity and computa- existing study on predicting mechanical properties based on
tional time requirements. To enhance prediction effi- structure. ANN could be applied to predict the elastic modu-
ciency, novel algorithms optimize hyperparameters lus of continuous SiC fiber-reinforced SiC ceramic matrix
within ML frameworks. With advancements in machine (SiC/SiC) mini-composites using structural parameters (fiber
learning technology, newer approaches like deep learn- volume ratio and porosity),71 and the elastic properties of
ing have been implemented. However, the choice of ML short fiber-reinforced materials using fiber volume fraction
method varies across different studies regarding accuracy, and fiber geometric properties.72 The application of optimi-
efficiency, and effectiveness levels. Therefore, selecting zation algorithms in studies has been employed to enhance
the most suitable ML method requires continuous efforts. the prediction efficiency of ANN. A genetic algorithm-
optimized ANN was constructed to explore the nonlinear
relationship between the microstructure of unidirectional
3.2 | Prediction of mechanical properties carbon fiber-reinforced polymer composites and their trans-
based on structure verse mechanical properties, including the transverse elastic
modulus, tensile strength, and compressive strength.73
Many composites have diverse compositions and complex In addition to ANN, various other ML techniques have
structures. The relationship between structure and per- been effectively employed in this domain. Pathan et al.74
formance of these materials has long been a focal and employed a gradient-boosted tree regression model for
challenging issue in research. Structure plays a pivotal predicting the macroscopic stiffness and yield strength
role in deciding the composite mechanical properties. based on microscopic structural images of unidirectional
Using ML methods to identify implicit quantitative rela- fiber-reinforced polymer composites. Wenjing Wang
tionships between structure and mechanical performance et al.75 predicted the tensile performance of plant fiber-
enables the effective prediction of mechanical properties reinforced composites using ML methods based on influ-
based on structural information. Understanding the ential factors (fiber volume fraction, contact interface, and
structure-performance relationship is crucial for material resin type). The employed ML methods included categori-
design and applications. The composite mechanical prop- cal feature-supported gradient boosting, RF, extreme gra-
erties can be predicted using ML models that take struc- dient boosting, decision trees, and gradient boosting
tural parameters as inputs. Some researchers have decision trees. The results indicated that the fiber volume
applied this approach to wood-based composites. Fran- fraction was a pivotal parameter in determining the com-
cisco et al.68 used material thickness, moisture content, posite tensile strength, with gradient boosting decision
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1950 LIANG ET AL.

tree showing superior performance (R2 value of 0.786, and investigated the relationship between climate factors and
the lowest MAE and RMSE values of 5.904 and the internal bonding strength of particleboard using either
7.456 MPa, respectively). multiple regression analysis and ANN. Similarly, Fini
Convolutional neural network is primarily designed et al.81 utilized ANN to investigate the elastic, superelastic,
for image data types as input for model training. Some and viscoelastic properties of wood-plastic composites at
studies have constructed CNN models to accurately pre- different temperatures. O. Abuomar et al.82 employed
dict the composite mechanical properties, directly using SVM to study the diverse mechanical responses exhibited
structural images as input data without the need to extract by nanocomposites of vapor-grown carbon nanofiber/
and process structural parameters. Govindaraj et al.76 used vinyl ester under different formulations, processing, and
the binary distribution map of materials (0 representing environmental conditions. Hooman Dadras et al.83 devel-
stiffness, 1 representing flexibility) as input to construct a oped an ANN to investigate the impact of immersion time
CNN. This model accurately predicted the stiffness, bend- in acidic environments on the mechanical properties
ing, and tensile mechanical properties of natural fiber of glass fiber-reinforced polymer composites. These
composites. The CNN based on transfer learning were endeavors collectively underscore the necessity of acquir-
used in Ref. [77] for predicting the mechanical properties ing a comprehensive understanding of how environmental
of two-phase material, thus establishing the structure– factors influence the performance of composites.
property relationship of composites. For composites fea- Non-destructive testing (NDT) of material mechanical
turing arbitrary shapes and inclusions, Sang Ye et al.78 properties is a method for evaluating these properties with-
introduced a ML-based approach to effectively predict the out causing permanent damage or destruction to the mate-
material elastic properties. They generated multiple data- rial. This technique aims to identify defects, cracks,
sets using the FE method and developed a CNN to directly deformations, fatigue, corrosion, or other issues by detect-
predict the effective Young's modulus and Poisson's ratio ing and analyzing the material's internal and surface fea-
from microstructure images. Lei et al.79 proposed a multi- tures, allowing for the assessment of mechanical properties
modal fusion prediction model that integrates information and structural integrity.84 Researchers have increasingly
from composite microstructure and composition. They combined NDT with ML, using data obtained through
first used a CNN to extract features related to the material NDT as input to build models for more accurate and reli-
microstructure, while a multilayer perceptron was utilized able predictions of composite material mechanical proper-
for extracting composition features. Subsequently, they ties. Guanglin et al.85 integrated NDT methods based on
designed an adaptive feature vector fusion module to bending and longitudinal vibration frequencies with ANN
adjust the impact of different modalities on mechanical to predict the MOE of bamboo-wood composites. Way
properties, achieving high-level feature fusion. The neural et al.86 applied undamaged stress wave and ANN for struc-
network was then utilized to predict the mechanical prop- tural damage detection in medium-density fiberboard. Sla-
erties of composites. To summarize, the relationship womir et al.87 used the non-destructive Schmidt hammer
between composite structure and properties is intricate yet method combined with ANN for predicting the subsurface
pivotal. ML techniques, such as ANN and CNN, play a sig- tensile strength of cementitious composites.
nificant role in discerning the quantitative correlation Microwave detection is an effective method for NDT of
between structure and mechanical properties for precise composites, as microwave signals can penetrate and inter-
prediction. The efficiency of ANN prediction can be fur- act with the internal structure of these materials. The
ther enhanced through optimization algorithms. Addition- microwave signal of the material is acquired through
ally, CNN demonstrates effective application in directly microwave reflection. To effectively identify defect sources
predicting the mechanical properties of composite mate- and optimize the processing of microwave signals for dis-
rials based on structural images. tinguishing between defect and non-defect signals, ML
was employed. The contribution of k-means clustering to
detect defect down to 1 mm, which provides a consider-
3.3 | Prediction of mechanical properties able advantage in glass fiber-reinforced polymer compos-
based on others ites.88 Furthermore, terahertz wave detection, another
microwave testing method, has been combined with deep
In addition to process and structure, different environ- learning in existing studies for defect detection in compos-
mental conditions significantly impact material perfor- ites. The terahertz detection process requires specialized
mance. Therefore, it is essential to consider the usage expertise for accurate defect analysis, rendering it both
environment of composites to ensure their reliable and time-consuming and costly. Moreover, as the demand for
safe performance in practical applications. A noteworthy defect detection grows in reliability and accuracy, data
example comes from the work of Hideaki and Ken,80 who analysis complexity is significantly rising. The obstacle,
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1951

nonetheless, can be effectively tackled by harnessing the processes to engineer composites that fulfill performance
power of deep learning. To quantitatively distinguish over- requirements. Traditional approaches to composite
lapping signals from micro-defects in glass fiber reinforced material design often rely on empirical methods and trial-
polymer composites through terahertz signal processing, and-error, necessitating extensive experimentation and spe-
researchers proposed CNN,89 a full-scale promoted CNN90 cialized knowledge, resulting in time-consuming and costly
and deformable attention CNN.91 processes.99 With the advancement of modern computer
Thermography testing is a nondestructive testing technology and simulation tools, ML has emerged as a
technique to evaluate the temperature distribution on the highly promising method for inverse design.100 ML has
surface or within an object.92 Nondestructive testing demonstrated significant successes in various fields, such
takes input signals (temperature distribution in the ther- as photonic device design,101 architectural structure,102 and
mal imager) to output information about defects. This is material103 design, as well as the design and optimization
referred to as an inverse problem, which is challenging to of thermoelectric generators.104,105 ML methods enable
solve. The ML was harnessed to confront this challenge. engineers and researchers to rapidly identify material con-
Vahid and Mohammad93 combined infrared thermogra- figurations that meet specific requirements, accelerating
phy with the k-nearest neighbors algorithm for the detec- the processes of material design and optimization.106,107
tion of voids and defects in fiber reinforced composites. Furthermore, integrating ML approaches with computa-
Ali et al.94 proposed a thermal imaging detection method tional and simulation-based design methods enables more
for fatigue damage in glass fiber-epoxy composites based efficient design of material structure and process. In the
on the kernel ridge regression model. Muzammil et al.95 field of composites, the rapid development of ML technol-
utilized SVM and Barker coded thermography to detect ogy offers new opportunities for enhanced design and
defect of carbon fiber reinforced polymer composites. optimization.
Although the prediction goal is that the image form can The ML-based design and optimization of materials
be done using CNN, the scholars chose conventional ML are primarily achieved through three main approaches.
methods because of the simplicity and faster implementa- The first approach entails designing materials based on
tion along with hyperparameter optimization. target property.108 ML model is constructed by using tar-
The utilization of ML proves to be an influential tool get property as input and design parameters as output.
in data mining, especially when dealing with noisy and This model predicts the optimal design parameters to
incomplete datasets that exhibit high nonlinearity. The achieve the desired property, guiding the preparation and
complexity of modeling and finding solutions escalates sig- optimization of materials.
nificantly when guided waves traverse through an aniso- The second approach is based on known material prop-
tropic medium, and becomes even more intricate when erties (shown in Figure 7) and encompasses three main
attempting to simulate their interaction with defects.96 steps: descriptor generation and dimensionality reduction,
Therefore, A. Mardanshahi et al.97 combined guided wave model construction and validation, as well as material pre-
propagation and ML to detect matrix cracking in compos- diction and experimental validation.109 The process begins
ites. They extracted Lamb wave signal data from glass/ by collecting descriptors or features of materials, requiring
epoxy cross-ply laminated composites with different a deep understanding of domain-specific knowledge to
matrix cracking densities and used this data to build SVM, ensure the accuracy and relevance of descriptors. Next, a
linear vector quantization neural network, and multilayer mapping model is established between descriptors and tar-
perceptron neural network models. Laser ultrasonic tech- get performance, constructing ML models with descriptors
nology enables precise and non-contact NDT of compos- as input and target property as output. The model construc-
ites. However, the limited signal-to-noise ratio of laser tion process relies on known material data and utilizes vari-
ultrasonic signals poses a primary challenge. ANN can ous ML methods to establish this mapping. Sufficient data
uncover the underlying information in the signal. A com- is crucial to ensure the reliability of the predictive property.
bination of ANN and laser ultrasonic technology detected Finally, inverse design is conducted to find new materials
cracks in carbon fiber reinforced plastic, offering fast rec- with the desired property. In this phase, ML models predict
ognition speed and high accuracy (99.6%).98 potential candidate materials that meet specific design cri-
teria. These candidates are then experimentally validated in
the laboratory to confirm actual properties. The focus of
4 | DESIGN AN D OPT I MI ZATI O N this method is on selecting ML algorithms and composite
OF COMPOSITES descriptors to map the input space to the output space. The
trained ML model serves as a filter to explore the design
Inverse design is a method that starts from the desired space of materials, identifying potential candidate materials
properties and optimizes factors, such as structure and that meet design criteria.
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1952 LIANG ET AL.

FIGURE 7 Workflow for material discovery and design of based on known material properties.109

The third approach involves directly utilizing inverse the relationship between material mechanical properties
ML methods to drive material inverse design. This method and processes to achieve process design for composites.
mainly employs ML for exploring and optimizing structure Hot pressing is crucial in plywood production as it
design, using inverse ML techniques to explore material directly impacts the mechanical property. In the plywood
structures that meet specific property targets. Researchers industry, ML can be employed as a cost-effective
achieve composite material structure design by constructing approach to determine the optimal pressing parameters,
neural network models, with prediction models serving as reducing the need for excessive experiments, labor, time,
feedback mechanisms based on desired property. Various energy consumption, and high expenses. Hasan et al.115
approaches, such as leverages generative inverse and surro- inputted pressing parameters into an ANN to construct a
gate forward neural network model,110 generative adversar- predictive model for the bonding strength, bending
ial networks (GAN),111 graph neural networks,112 and strength, and elastic modulus of plywood. The subse-
more, can be utilized. Taking the research conducted by quent utilization of the trained ANN facilitated the opti-
Chun-Teh and Grace as an illustrative example of imple- mization of process parameters, obtaining the optimal
menting this approach in composite inverse design, they pressing temperature and time for wood veneer/
proposed the use of generative inverse design network polystyrene composite plywood. In addition to the hot
(GIDN).113 GIDN framework comprises two deep neural pressing parameters, ANN can also be used to determine
networks: the “predictor” and the “designer” (as illustrated the optimal veneer drying temperature of plywood, based
in Figure 8). In the predictor, weights and biases are learn- on the desired bonding shear strength.116
ing variables optimized to minimize the difference between The development of numerical optimization algo-
ML predictions and actual outcomes. In the designer, the rithms aims to enhance the efficiency of machine learning
values of weights and biases are adopted from the predictor models in solving optimization problems and accelerate
and treated as constants. Initial designs generated from the retrieval of optimal process parameter schemes. Yang
samples drawn from a gaussian distribution are fed into the et al.117 employed ML to study the relationship between
designer as input, and optimized designs as output. mechanical properties (limit strength, flexural strength,
Through a feedback loop, the optimized designs are vali- and impact) and injection molding process during the co-
dated using physics-based models. Subsequently, these blending process of short glass fiber and polytetrafluor-
designs can be incorporated into the previous training data oethylene reinforced polycarbonate composites. Initially,
for the next round of training and design iterations. an ANN was constructed using process parameters to pre-
dict the mechanical properties. The subsequent applica-
tion of optimization algorithms, including simulated
4.1 | Process design annealing and genetic algorithms, enabled individual
searches to identify the optimal mixture ratio settings.
In material development, ML directly associates material Another intriguing approach to process design is to
performance with the manufacturing process, which can build a predictive model using the desired process param-
avoid errors accumulated due to the complexity of multi- eters as outputs. Cotton fiber is commonly used as a filler
ple mechanisms.114 Some studies utilize ML to establish material for enhancing the polypropylene composite
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1953

F I G U R E 8 The framework
of GIDN by Chun-Teh and
Grace.113

mechanical properties. Kazi et al.108 constructed an ANN composites that meet mechanical property requirements
for predicting the target filler content of cotton fiber/ provides significantly enhanced efficiency. Checkerboard
polypropylene composites based on desired properties composite is a type of composite material with two phases
(tensile strength and elastic modulus). The thermoform- (one phase being soft and ductile, while the other phase is
ing process involves preheating laminated sheets to the stiff and brittle). Diab et al.119 established a CNN to pre-
desired forming temperature, pressing and cooling them dict the mechanical properties (stiffness, strength, and
between molds, resulting in producing the final molded toughness) of two-dimensional checkerboard composites.
part. To achieve defect-free parts of fiber reinforced com- Subsequently, they utilized a genetic algorithm to find the
posites in the thermoforming process, engineers need to optimal microstructure design that achieves maximum
conduct numerous physical experiments and simulations mechanical property. In the inverse design process of
to determine appropriate process parameter combina- establishing design space and searching candidate com-
tions. These methods are costly due to manual labor posites with target property, in addition to the previously
hours, equipment and raw material usage, and mold mentioned genetic algorithm, GAN methods can also be
manufacturing costs. Hence, ML method was applied to used. The composite structure design integrates the GAN
the thermoforming of woven composites to expedite pro- and CNN, as illustrated in Figure 9. Kazuya et al.120 pro-
cess optimization. In order to predict process parameters posed a ML framework to predict the mechanical proper-
that yield good molding product quality, Long et al.118 ties based on the phase separation structure of polymer
obtained full-field contour data from FE simulation alloys, followed by conducting inverse design to generate
results, and developed an ANN. Subsequently, the pro- material structures based on these properties. The authors
cess parameters were optimized by minimizing the maxi- employed a CNN for forward analysis by using the
mum slip path length and maximizing the region with a Young's modulus as the output. Inverse design was
predetermined shear angle range in the lamination. The achieved by combining GAN and CNN in a random
findings showed that ANN could effectively optimize pro- search. In this framework, the CNN and GAN were
cess parameters to achieve enhanced product quality. ML trained independently with distinct roles. This method is
has also been utilized in the design of fiber reinforced more cost-effective and efficient since GANs with high
polymer composites. computational costs do not require retraining. Leveraging
CNN, GAN and genetic algorithms, Chao et al.121 pro-
posed an adaptive generative design method to design
4.2 | Structure design composite. The method effectively explores the design
space and generates a GAN for creating candidate designs
Compared to the trial-and-error method of synthesizing by adopting adaptive learning and optimization strategies.
materials in the laboratory, ML methods employed to The reduction in the size of design variables significantly
explore predefined design spaces and find candidate expedites the optimization process.
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1954 LIANG ET AL.

F I G U R E 9 Composite structure design combining GAN and CNN. (A) Flow diagram. The blue area signifies forward analysis, while
the yellow area signifies inverse design. (B)120 and (C)121 Architectures for ML. (A) The generator of GAN. (B) The discriminator of GAN.
(C) The architecture of CNN for mechanical properties prediction.

Unlike the application in process design, research in inverse (generation) problems in composite material
structure design often employs the direct construction of design. The prediction model was initially constructed,
inverse ML models. The use of predictive ML models is and then utilized as a feedback mechanism in the neu-
better suited for composite design problems characterized ral network generation model. The proposed prediction-
by a smaller design space. When confronted with more generation model can generate composite material
complex material design and optimization problems, microstructures that meet various requirements based
there may exist a vast number of candidate materials. In on the desired elastic properties. Interpenetrating phase
such extensive design space, despite forward modeling composite (IPC) is advanced multiphase composites.
tools are very powerful, the complexity of composite Darshil et al.124 employed a deep neural network to gen-
design poses challenges in the application of predictive erate IPC with specified properties by integrating mathe-
models. Therefore, many researchers prefer to directly matically known triply periodic minimal surfaces
adopt inverse ML models to accomplish structure design (TPMS), enabling inverse material design. Their first
while satisfying mechanical property requirements of step involved generating IPC structures based on TPMS,
composite design and optimization without exhaustive combining three base TPMS surfaces in varying propor-
exploration the entire design space. tions to develop new TPMS-based IPC structures. The
Due to the tremendous design space on numerous second step was to establish a homogenized FE analysis
voxels, the rational designing the property distribution solver to assess the effective elastic properties of the
of active composites poses a significant challenge. A combined TPMS IPCs. Finally, a deep neural network
new approach combining RNN and evolutionary algo- architecture was trained using data generated from the
rithms was proposed to guide the design of active com- first two steps to solve the inverse problem—predicting
posites, effectively finding suitable designs to achieve the geometric parameters of TPMS IPC structures for
multiple target shape changes of varying complexity.122 given desired material properties, including volume frac-
Ashank et al.123 developed a computational framework tion, shear modulus, bulk modulus, Young's modulus,
based on CNN to address the forward (prediction) and Poisson's ratio, and so forth.
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1955

5 | CONCLUSION AND P ROSPECT involve leveraging more advanced ML techniques to

address these challenges.
Machine learning has made significant strides in the Current research often treats the construction of con-
study of composite mechanics. This paper provides an stitutive models, prediction of mechanical properties, and
overview of the ML applications in three aspects of com- the design and optimization of materials in composites as
posites: the constitutive model construction, the mechan- independent problems rather than interconnected com-
ical property prediction, and the composite design and ponents of a comprehensive mechanical system. This
optimization. In the application of constructing constitu- approach may lead to redundant efforts, increased com-
tive models for isotropic and anisotropic composites, ML plexity, and costs. Through mutual feedback and optimi-
offers flexibility and applicability in dealing with non- zation, the research on composites becomes more
linear systems, overcoming the issues of excessive param- systematic and effective. Moreover, the intrinsic mecha-
eters and limited practicality associated with traditional nisms of the composite mechanical behavior using ML
methods. In the realm of predicting mechanical proper- have not received sufficient attention and in-depth explo-
ties, various ML models are employed, leveraging multi- ration at the current stage, leaving significant research
source data related to the processes, structures, and gaps in this field. Exploring advanced research methods
environmental conditions of composites. Additionally, to unveil the essence of material mechanics behavior is
the integration of NDT techniques aids in damage identi- necessary. Currently, most ML-based models are unable
fication and the prediction of mechanical properties. ML to comprehensively simulate the composite mechanical
is pivotal in the design and optimization of composites. It behavior. Future research should focus on constructing
achieves this by either designing based on known proper- models that describe the more complete mechanical
ties or target performance, or directly employing inverse behavior, integrating material viscoelasticity, damage,
ML methods. This facilitates the efficient development of and fracture properties, as well as considering the influ-
composites by tailoring material processes and structures ence of environmental factors like temperature, humid-
to meet specific requirements. ity, and light. Such research endeavors will provide
The use of ML in the study of the mechanics of com- deeper insights for optimizing the performance of com-
posites is in the ascendant, but it faces several challenges posites and material design.
that need to be addressed. Current research still grapples Machine learning models typically require a large
with issues such as the quantity, quality, and diversity of amount of data for accurate training. However, acquiring
data, as well as challenges related to model generalization experimental data for composites is often expensive and
and interpretability. To better address these issues in the time-consuming, leading to a contradiction between the
research methods for studying the composite mechanical need for data quantity and the available data. Data quality
properties based on ML, further in-depth research efforts is also a concern, as low-quality and noisy data can intro-
are needed. duce errors in ML models. The field of composites encom-
While ML has demonstrated significant advantages in passes various material types, structures, and application
constructing constitutive models of composites, the ideal scenarios, necessitating diverse data to train ML models to
approach would be to achieve the reconstruction of consti- cover the mechanical behavior of materials in different situ-
tutive models with well-defined physical meanings based ations. Techniques like data augmentation, meta-learning,
on the complex mechanical behaviors of composites. This and transfer learning can be employed to minimize data
remains a challenging and yet-to-be-completed task. More- requirements while ensuring diversity. Improving data
over, ML methods are limited in some highly complex com- quality involves addressing issues like reducing data noise,
posites, such as wood-based composites, due to factors like minimizing experimental errors, and developing more pre-
data availability, complexity, heterogeneity, specific proper- cise measurement techniques. Employing suitable methods
ties and applications. Materials like metal-based and for assessing data quality helps filter out high-quality data,
cement-based composites have accumulated rich experi- thereby enhancing ML model accuracy. Additionally, we
mental data, whereas data for wood-based composites are propose to establish an open and collaborative platform,
relatively scarce. Wood-based composites, with their intri- facilitating data sharing to foster the development of larger-
cate structure resulting from the combination of natural scale datasets. Furthermore, constructing knowledge graphs
wood and other components, pose challenges in simplifying to store and retrieve vast amounts of material information
their constitutive relationships into forms suitable for ML represents an effective strategy to enhance ML applications.
algorithms. However, as technology advances and the In addition to challenges in building datasets, there
demand for these materials increases, there is potential for are pressing issues regarding model generalization and
future research on constitutive models for biomass compos- interpretability. The generalization ability of ML models
ites, including wood-based composites. This research may to new datasets remains a significant challenge, as
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1956 LIANG ET AL.

models that perform well on training data may underper- learning. Chem Rev. 2020;120(16):8066-8129. doi:10.1021/acs.
form on new data, particularly in the modeling of complex chemrev.0c00004
material mechanics. Machine learning models often pro- 5. Suwardi A, Wang F, Xue K, et al. Machine learning-driven
biomaterials evolution. Adv Mater. 2022;34:1. doi:10.1002/
vide results in a black-box fashion, making it challenging to
adma.202102703
understand the inner workings of the model. This lack of 6. Juan Y, Dai Y, Yang Y, Zhang J. Accelerating materials dis-
interpretability can lead to a lack of trust in the model, covery using machine learning. J Mater Sci Technol. 2021;79:
especially in scientific and engineering applications where 178-190. doi:10.1016/j.jmst.2020.12.010
interpretability is crucial. By integrating ML models with 7. Karatzas P, Melagraki G, Ellis LA, et al. Development of deep
known principles of mechanics and leveraging prior knowl- learning models for predicting the effects of exposure to engi-
edge, the required amount of data can be reduced. This neered nanomaterials on Daphnia magna. Small. 2020;16(36):
e2001080. doi:10.1002/smll.202001080
fusion approach enhances the accuracy of ML models in
8. Cassola S, Duhovic M, Schmidt T, May D. Machine learning
describing complex materials' mechanical behavior while
for polymer composites process simulation – a review. Com-
maintaining interpretability. Exploring ways to incorporate pos B Eng. 2022;246:110208. doi:10.1016/j.compositesb.2022.
more prior knowledge into ML models is a promising 110208
research direction, aiming to develop interpretable ML 9. Chen Y, Zhao Z, Li D, Guo Z, Dong L. Constitutive model-
models that improve understanding of predictions and ing for linear viscoelastic fiber-reinforced composites. Com-
enhance trustworthiness in practical applications. The pos Struct. 2021;263:113679. doi:10.1016/j.compstruct.2021.
establishment of interpretable models in the material field 113679
10. Rique O, Liu X, Yu W, Pipes RB. Constitutive modeling for
may begin with model explanation methods, such as What-
time- and temperature-dependent behavior of composites.
If Explanations. Besides, some ML models might be benefit Compos B Eng. 2020;184:107726. doi:10.1016/j.compositesb.
from being explicitly designed to extract factual knowledge 2019.107726
or causal interpretations. 11. Salmenjoki H, Alava MJ, Laurson L. Machine learning plastic
The rapid development of ML algorithms has not deformation of crystals. Nat Commun. 2018;9(1):5307. doi:10.
been consistently reflected in the study of composite 1038/s41467-018-07737-2
mechanical properties. Therefore, there is an opportunity 12. Gao W. A comprehensive review on identification of the geo-
to bridge this gap by promptly introducing advanced and material constitutive model using the computational intelli-
gence method. Adv Eng Inform. 2018;38:420-440. doi:10.1016/
efficient algorithms from the broader field of ML to over-
j.aei.2018.08.021
come the challenges faced by research on the mechanical 13. Zhang P, Yin ZY, Jin YF. State-of-the-art review of machine
properties of composites in this new era. learning applications in constitutive modeling of soils. Arch
Comput Methods Eng. 2021;28(5):3661-3686. doi:10.1007/
A C K N O WL E D G M E N T S s11831-020-09524-z
This work was funded by the Science and Technology 14. Maia MA, Rocha IBCM, Kerfriden P, van der Meer FP.
Innovation Platform Construction Project in Heilongjiang Physically recurrent neural networks for path-dependent het-
erogeneous materials: embedding constitutive models in a
Province (DL41320601).
data-driven surrogate. Comput Methods Appl Mech Eng. 2023;
407:115934. doi:10.1016/j.cma.2023.115934
DATA AVAILABILITY STATEMENT 15. Ali U, Muhammad W, Brahme A, Skiba O, Inal K. Applica-
The data that support the findings of this study are avail- tion of artificial neural networks in micromechanics for poly-
able from the corresponding author upon reasonable crystalline metals. Int J Plast. 2019;120:205-219. doi:10.1016/j.
request. ijplas.2019.05.001
16. Shi S-x, Liu X-s, Zhang X-y, Zhou K-c. Comparison of flow
ORCID behaviors of near beta Ti-55511 alloy during hot compression
based on SCA and BPANN models. Trans Nonferrous Met
Xiaoting Niu https://orcid.org/0009-0002-7981-6776
Soc China. 2021;31(6):1665-1679. doi:10.1016/S1003-6326(21)
65606-6
R EF E RE N C E S 17. Liang Z, Yu F, Yinyang W, Yongdong X. Constitutive relation-
1. Hasan M, Zhao J, Jiang Z. Micromanufacturing of composite ship of (Ti5Si3+TiBw)/TC11 composites based on BP neural
materials: a review. Int J Extrem Manuf. 2019;1(1):012004. doi: network. Mater Today Commun. 2022;32:103973. doi:10.1016/
10.1088/2631-7990/ab0f74 j.mtcomm.2022.103973
2. Hart GLW, Mueller T, Toher C, Curtarolo S. Machine learn- 18. Wang YS, Linghu RK, Zhang W, Shao YC, Lan AD, Xu J.
ing for alloys. Nat Rev Mater. 2021;6(8):730-755. doi:10.1038/ Study on deformation behavior in supercooled liquid region
s41578-021-00340-w of a Ti-based metallic glassy matrix composite by artificial
3. Li C, Zheng K. Methods, progresses, and opportunities of mate- neural network. J Alloys Compd. 2020;844:155761. doi:10.
rials informatics. InfoMat. 2023;5(8):e12425. doi:10.1002/inf2. 1016/j.jallcom.2020.155761
12425 19. Yu G, Bao X, Xu X, et al. Constitutive modeling of Ta-rich
4. Jablonka KM, Ongari D, Moosavi SM, Smit B. Big-data sci- particle reinforced Zr-based bulk metallic composites in the
ence in porous materials: materials genomics and machine supercooled liquid region by using evolutionary artificial
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1957

neural network. J Alloys Compd. 2023;938:168488. doi:10. neural networks. Compos B Eng. 2021;225:109314. doi:10.
1016/j.jallcom.2022.168488 1016/j.compositesb.2021.109314
20. He ZC, Huo SL, Li E, Cheng HT, Zhang LM. Data-driven 34. Kim DW, Go MS, Lim JH, Lee S. Data-driven stress and strain
approach to characterize and optimize properties of carbon curves of the unidirectional composites by deep neural net-
fiber non-woven composite materials. Compos Struct. 2022; works with principal component analysis and selective-data
297:115961. doi:10.1016/j.compstruct.2022.115961 augmentation. Compos Struct. 2023;313:116902. doi:10.1016/j.
21. Kabbani MS, El Kadi HA. Predicting the effect of cooling compstruct.2023.116902
rate on the mechanical properties of glass fiber–polypropylene 35. Kim DJ, Kim GW, Baek J-h, Nam B, Kim HS. Prediction of
composites using artificial neural networks. J Thermoplast Com- stress-strain behavior of carbon fabric woven composites by
pos Mater. 2019;32(9):1268-1281. doi:10.1177/0892705718792351 deep neural network. Compos Struct. 2023;318:117073. doi:10.
22. Esmaeili H, Rizvi R. An accelerated strategy to characterize 1016/j.compstruct.2023.117073
mechanical properties of polymer composites using the 36. El Said B. Predicting the non-linear response of composite
ensemble learning approach. Comput Mater Sci. 2023;229: materials using deep recurrent convolutional neural networks.
112432. doi:10.1016/j.commatsci.2023.112432 Int J Solids Struct. 2023;276:112334. doi:10.1016/j.ijsolstr.2023.
23. Hasanzadeh A, Vatin NI, Hematibahar M, Kharun M, 112334
Shooshpasha I. Prediction of the mechanical properties of 37. Chen Q, Jia R, Pang S. Deep long short-term memory neural
basalt fiber reinforced high-performance concrete using network for accelerated elastoplastic analysis of heteroge-
machine learning techniques. Materials. 2022;15(20):7165. neous materials: an integrated data-driven surrogate
doi:10.3390/ma15207165 approach. Compos Struct. 2021;264:113688. doi:10.1016/j.
24. Gautham S, Sasmal S. Constitutive relations of nanoscale compstruct.2021.113688
hydration products present in engineered cementitious com- 38. Lee J, Lee N, Son J, Shin D. An LSTM model with optimal fea-
posites from machine learning assisted experimental nanoin- ture selection for predictions of tensile behavior and tensile
dentation. J Build Eng. 2023;75:106912. doi:10.1016/j.jobe. failure of polymer matrix composites. Korean J Chem Eng.
2023.106912 2023;40(9):2091-2101. doi:10.1007/s11814-023-1502-3
25. Marrivada GV, Chaganti PK, Sujith R. Experimental investiga- 39. Zhang Y, Tao Q, Liu Y, Wang C. Mesh/membrane composite
tion and machine learning prediction of mechanical properties with superior mechanical performance: a deep learning-based
of graphene nanoplatelets based triaxial braided composites. design. Compos Sci Technol. 2022;230:109735. doi:10.1016/j.
Mater Today Commun. 2023;34:105305. doi:10.1016/j.mtcomm. compscitech.2022.109735
2022.105305 40. Mozaffar M, Bostanabad R, Chen W, Ehmann K, Cao J,
26. Hiew SY, Bin TK, Raman SN, Kong D, Hafezolghorani M. Pre- Bessa MA. Deep learning predicts path-dependent plasticity.
diction of ultimate conditions and stress–strain behaviour of Proc Natl Acad Sci. 2019;116(52):26414-26420. doi:10.1073/
steel-confined ultra-high-performance concrete using sequen- pnas.1911815116
tial deep feed-forward neural network modelling strategy. Eng 41. Zhang Z, Liu Q, Wu D. Predicting stress–strain curves using
Struct. 2023;277:115447. doi:10.1016/j.engstruct.2022.115447 transfer learning: knowledge transfer across polymer composites.
27. Liu F, Ding W, Qiao Y, Wang L. An artificial neural network Mater Des. 2022;218:110700. doi:10.1016/j.matdes.2022.110700
model on tensile behavior of hybrid steel-PVA fiber reinforced 42. Park D, Park M, Ryu S. Expanding design spaces in digital
concrete containing fly ash and slag power. Front Struct Civ composite materials: a multi-input deep learning approach
Eng. 2020;14(6):1299-1315. doi:10.1007/s11709-020-0712-6 enhanced by transfer learning and multi-kernel network. Adv
28. Liu F-y, Ding W-q, Qiao Y-f, Wang L-b, Chen Q-y. Compres- Theory Simul. 2023;6(11):2300465. doi:10.1002/adts.202300465
sive behavior of hybrid steel-polyvinyl alcohol fiber-reinforced 43. Zhang C, Li Y, Jiang B, Wang R, Liu Y, Jia L. Mechanical
concrete containing fly ash and slag powder: experiments and properties prediction of composite laminate with FEA and
an artificial neural network model. J Zhejiang Univ Sci A. machine learning coupled method. Compos Struct. 2022;299:
2021;22(9):721-735. doi:10.1631/jzus.A2000379 116086. doi:10.1016/j.compstruct.2022.116086
29. Peng GCY, Alber M, Buganza Tepole A, et al. Multiscale model- 44. Jin W, Pei J, Xie P, Chen J, Zhao H. Machine learning-based
ing meets machine learning: what can we learn? Arch Comput prediction of mechanical properties and performance of
Methods Eng. 2021;28(3):1017-1037. doi:10.1007/s11831-020- nickel–graphene nanocomposites using molecular dynamics
09405-5 simulation data. ACS Appl Nano Mater. 2023;6(13):12190-
30. Feng H, Prabhakar P. Parameterization-based neural net- 12199. doi:10.1021/acsanm.3c01919
work: predicting non-linear stress–strain response of compos- 45. Tao F, Liu X, Du H, Yu W. Learning composite constitutive
ites. Eng Comput. 2023;40:1621-1635. doi:10.1007/s00366-023- laws via coupling Abaqus and deep neural network. Compos
01849-0 Struct. 2021;272:114137. doi:10.1016/j.compstruct.2021.114137
31. Yang C, Kim Y, Ryu S, Gu GX. Prediction of composite micro- 46. Tao F, Liu X, Du H, Yu W. Finite element coupled positive
structure stress-strain curves using convolutional neural networks. definite deep neural networks mechanics system for constitu-
Mater Des. 2020;189:108509. doi:10.1016/j.matdes.2020.108509 tive modeling of composites. Comput Methods Appl Mech Eng.
32. Ammasai Sengodan G. Prediction of two-phase composite 2022;391:114548. doi:10.1016/j.cma.2021.114548
microstructure properties through deep learning of reduced 47. Liu X, Tao F, Yu W. A neural network enhanced system for
dimensional structure-response data. Compos B Eng. 2021; learning nonlinear constitutive law and failure initiation crite-
225:109282. doi:10.1016/j.compositesb.2021.109282 rion of composites using indirectly measurable data. Compos
33. Kim DW, Lim JH, Lee S. Prediction and validation of the Struct. 2020;252:112658. doi:10.1016/j.compstruct.2020.112658
transverse mechanical behavior of unidirectional composites 48. Khanam PN, AlMaadeed M, AlMaadeed S, et al. Optimization
considering interfacial debonding through convolutional and prediction of mechanical and thermal properties of
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1958 LIANG ET AL.

graphene/LLDPE nanocomposites by using artificial neural 62. Dong W, Huang Y, Lehane B, Ma G. Multi-objective design
networks. Int J Polym Sci. 2016;2016:1-15. doi:10.1155/2016/ optimization for graphite-based nanomaterials reinforced
5340252 cementitious composites: a data-driven method with machine
49. Hossain KMA, Anwar MS, Samani SG. Regression and learning and NSGA-II. Construct Build Mater. 2022;331:
artificial neural network models for strength properties of 127198. doi:10.1016/j.conbuildmat.2022.127198
engineered cementitious composites. Neural Comput Appl. 63. Zhang B, Hua J, Cai L, Gao Y, Li Y. Multi-objective optimi-
2018;29(9):631-645. doi:10.1007/s00521-016-2602-3 zation of particle gluing operating parameters in particle-
50. Reddy BRS, Premasudha M, Panigrahi BB, Cho K, board production based on improved machine learning
Reddy NGS. Modeling CONSTITUENT–PROPERTY relationship of algorithms. J Wood Sci. 2022;68(1):61. doi:10.1186/s10086-
polyvinylchloride composites by neural networks. Polym Com- 022-02068-9
pos. 2020;41(8):3208-3217. doi:10.1002/pc.25612 64. Peng Y, Unluer C. Modeling the mechanical properties of
51. Tiryaki S, Aras U, Kalaycıoglu H, Erişir E, Aydın A. Predictive recycled aggregate concrete using hybrid machine learning
models for modulus of rupture and modulus of elasticity of algorithms. Resour Conserv Recycl. 2023;190:106812. doi:10.
particleboard manufactured in different pressing conditions. 1016/j.resconrec.2022.106812
High Temp Mater Process. 2017;36(6):623-634. doi:10.1515/ 65. Mahjoubi S, Barhemat R, Meng W, Bao Y. Deep learning from
htmp-2015-0203 physicochemical information of concrete with an artificial
52. Kurt R. Control of system parameters by estimating screw language for property prediction and reaction discovery.
withdrawal strength values of particleboards using artificial Resour Conserv Recycl. 2023;190:106870. doi:10.1016/j.
neural network-based statistical control charts. J Wood Sci. resconrec.2023.106870
2022;68(1):64. doi:10.1186/s10086-022-02065-y 66. Hassani Niaki M, Pashaian M, Ghorbanzadeh Ahangari M.
53. Yang L, Qin Z. Mycelium-based wood composites for light Experimental assessment of the efficiency of deep learning
weight and high strength by experiment and machine learn- method in predicting the mechanical properties of polymer
ing. Cell Reports Phys Sci. 2023;4(6):101424. doi:10.1016/j.xcrp. concretes and composites. J Build Eng. 2023;76:107107. doi:10.
2023.101424 1016/j.jobe.2023.107107
54. Nwobi-Okoye CC, Anyichie MK, Atuanya CU. RSM and ANN 67. Zazoum B, Triki E, Bachri A. Modeling of mechanical proper-
modeling for production of Newbouldia Laevies fibre and ties of clay-reinforced polymer nanocomposites using deep
recycled high density polyethylene composite: multi objective neural network. Materials. 2020;13(19):4266. doi:10.3390/
optimization using genetic algorithm. Fibers Polym. 2020; ma13194266
21(4):898-909. doi:10.1007/s12221-020-9597-1 68. Fernandez FG, de Palacios P, Esteban LG, Garcia-Iruela A,
55. Zeng GS, Hu C, Zou S, Zhang L, Sun G. BP neural network Rodrigo BG, Menasalvas E. Prediction of MOR and MOE of
model for predicting the mechanical performance of a foamed structural plywood board using an artificial neural network
wood-fiber reinforced thermoplastic starch composite. Polym and comparison with a multivariate regression model. Com-
Compos. 2019;40(10):3923-3928. doi:10.1002/pc.25252 pos B Eng. 2012;43(8):3528-3533. doi:10.1016/j.compositesb.
56. Nazerian M, Razavi SA, Partovinia A, Vatankhah E, 2011.11.054
Razmpour Z. Prediction of the bending strength of a lami- 69. Lopes H, Silva SP, Machado J. Application of artificial neural
nated veneer lumber (LVL) using an artificial neural network. networks to predict mechanical behaviour of cork-rubber
Mech Compos Mater. 2020;56(5):649-664. doi:10.1007/s11029- composites. Neural Comput Appl. 2021;33(20):14069-14078.
020-09911-4 doi:10.1007/s00521-021-06048-w
57. Negro C, Alonso A, Blanco A, Tijero J. Breaking load and 70. Ding X, Hou X, Xia M, Ismail Y, Ye J. Predictions of macro-
bending strength prediction in manufacture of fibre cement scopic mechanical properties and microscopic cracks of unidi-
composites using artificial neural networks and a flocculation rectional fibre-reinforced polymer composites using deep
sensor. Compos Part A Appl Sci Manuf. 2005;36(12):1617- neural network (DNN). Compos Struct. 2022;302:116248. doi:
1626. doi:10.1016/j.compositesa.2005.04.008 10.1016/j.compstruct.2022.116248
58. Balaji N, Jayabal S. Artificial neural network modeling of 71. Tang L, Mu F, Chuwei Z. Prediction of tensile elastic modulus
mechanical behaviors of zea fiber–polyester composites. Proc of SiC/SiC mini-composites with the artificial neural network.
Inst Mech Eng Part E J Process Mech Eng. 2016;230(1):45-55. Acta Mech. 2023;234(10):4733-4748. doi:10.1007/s00707-023-
doi:10.1177/0954408914539939 03640-0
59. Devadiga U, Poojary RKR, Fernandes P. Artificial neural net- 72. Mentges N, Dashtbozorg B, Mirkhalaf SM. A
work technique to predict the properties of multiwall carbon micromechanics-based artificial neural networks model for
nanotube-fly ash reinforced aluminium composite. J Mater elastic properties of short fiber composites. Compos B Eng.
Res Technol. 2019;8(5):3970-3977. doi:10.1016/j.jmrt.2019. 2021;213:108736. doi:10.1016/j.compositesb.2021.108736
07.005 73. Li M, Zhang H, Li S, Zhu W, Ke Y. Machine learning and
60. Rahmati M, Toufigh V. Evaluation of geopolymer concrete at materials informatics approaches for predicting transverse
high temperatures: an experimental study using machine mechanical properties of unidirectional CFRP composites
learning. J Clean Prod. 2022;372:133608. doi:10.1016/j.jclepro. with microvoids. Mater Des. 2022;224:111340. doi:10.1016/j.
2022.133608 matdes.2022.111340
61. Schubert M, Lukovic M, Christen H. Prediction of mechanical 74. Pathan MV, Ponnusami SA, Pathan J, et al. Predictions of the
properties of wood fiber insulation boards as a function of mechanical properties of unidirectional fibre composites by
machine and process parameters by random forest. Wood Sci supervised machine learning. Sci Rep. 2019;9(1):13964. doi:10.
Technol. 2020;54(3):703-713. doi:10.1007/s00226-020-01184-3 1038/s41598-019-50144-w
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIANG ET AL. 1959

75. Wang W, Wu Y, Liu W, Fu T, Qiu R, Wu S. Tensile based on K-means clustering algorithm. IEEE Access. 2021;9:
performance mechanism for bamboo fiber-reinforced, palm 4820-4828. doi:10.1109/ACCESS.2020.3048147
oil-based resin bio-composites using finite element simulation 89. Kim HS, Park DW, Kim SI, Oh GH, Kim HS. Non-destructive
and machine learning. Polymers. 2023;15(12):2633. doi:10. detection of thin micro-defects in glass reinforced polymer
3390/polym15122633 composites using a terahertz electro–magnetic wave based on
76. Ramkumar G, Sahoo S, Anitha G, et al. An unconventional a convolution neural network. Compos B Eng. 2023;257:
approach for analyzing the mechanical properties of natural 110694. doi:10.1016/j.compositesb.2023.110694
fiber composite using convolutional neural network. M R, ed. 90. Xu Y, Zhou H, Cui Y, et al. Full scale promoted convolution
Adv Mater Sci Eng. 2021;2021:1-15. doi:10.1155/2021/5450935 neural network for intelligent terahertz 3D characterization of
77. Xu Y, Weng H, Ju X, et al. A method for predicting mechani- GFRP delamination. Compos B Eng. 2022;242:110022. doi:10.
cal properties of composite microstructure with reduced data- 1016/j.compositesb.2022.110022
set based on transfer learning. Compos Struct. 2021;275: 91. Wang X, Xu Y, Cui Y, et al. DA-CNN-based similar terahertz
114444. doi:10.1016/j.compstruct.2021.114444 signal identification for intelligent characterization of internal
78. Ye S, Li B, Li Q, Zhao HP, Feng XQ. Deep neural network debonding defects of composites under high-resolution mode.
method for predicting the mechanical properties of compos- Compos Struct. 2023;322:117412. doi:10.1016/j.compstruct.
ites. Appl Phys Lett. 2019;115(16):161901. doi:10.1063/1. 2023.117412
5124529 92. Just-Agosto F, Serrano D, Shafiq B, Cecchini A. Neural net-
79. Song L, Wang D, Liu X, Yin A, Long Z. Prediction of mechan- work based nondestructive evaluation of sandwich compos-
ical properties of composite materials using multimodal ites. Compos B Eng. 2008;39(1):217-225. doi:10.1016/j.composi
fusion learning. Sensors Actuators A Phys. 2023;358:114433. tesb.2007.02.023
doi:10.1016/j.sna.2023.114433 93. Daghigh V, Naraghi M. Machine learning-based defect char-
80. Korai H, Watanabe K. Predicting the strength reduction acterization in anisotropic materials with IR-thermography
of particleboard subjected to various climatic conditions synthetic data. Compos Sci Technol. 2023;233:109882. doi:10.
in Japan using artificial neural networks. Eur J Wood 1016/j.compscitech.2022.109882
Wood Prod. 2017;75(3):385-396. doi:10.1007/s00107-016- 94. Sarhadi A, Albuquerque RQ, Demleitner M, Ruckdäschel H,
1056-8 Eder MA. Machine learning based thermal imaging damage
81. Fini SHA, Farzaneh M, Erchiqui F. Study of the elastic behav- detection in glass-epoxy composite materials. Compos Struct.
iour of wood–plastic composites at cold temperatures using 2022;295:115786. doi:10.1016/j.compstruct.2022.115786
artificial neural networks. Wood Sci Technol. 2015;49(4):695- 95. Parvez MM, Mishra SK, Nandini K, et al. Machine learning
705. doi:10.1007/s00226-015-0717-9 based models for defect detection in composites inspected by
82. Abuomar O, Nouranian S, King R, Ricks TM, Lacy TE. Com- barker coded thermography: a qualitative analysis. Adv Eng
prehensive mechanical property classification of vapor-grown Softw. 2023;178:103425. doi:10.1016/j.advengsoft.2023.103425
carbon nanofiber/vinyl ester nanocomposites using support 96. Haywood-Alexander M, Dervilis N, Worden K, Cross EJ,
vector machines. Comput Mater Sci. 2015;99:316-325. doi:10. Mills RS, Rogers TJ. Structured machine learning tools for
1016/j.commatsci.2014.12.029 modelling characteristics of guided waves. Mech Syst Signal
83. Dadras H, Teimouri A, Barbaz-Isfahani R, Saber-Samandari S. Process. 2021;156:107628. doi:10.1016/j.ymssp.2021.107628
Indentation, finite element modeling and artificial neural net- 97. Mardanshahi A, Nasir V, Kazemirad S, Shokrieh MM. Detec-
work studies on mechanical behavior of GFRP composites in tion and classification of matrix cracking in laminated com-
an acidic environment. J Mater Res Technol. 2023;24:5042- posites using guided wave propagation and artificial neural
5058. doi:10.1016/j.jmrt.2023.04.146 networks. Compos Struct. 2020;246:112403. doi:10.1016/j.com
84. Najeeb MI, Syamsir A, Amir SMM, Khan T, Sebaey TA. pstruct.2020.112403
Failure analysis of plant fibre-reinforced composite in civil 98. Fu LL, Yang JS, Li S, Luo H, Wu JH. Artificial neural
building materials using non-destructive testing methods: cur- network-based damage detection of composite material using
rent and future trend. J Nat Fibers. 2023;20(2):2246654. doi:10. laser ultrasonic technology. Measurement. 2023;220:113435.
1080/15440478.2023.2246654 doi:10.1016/j.measurement.2023.113435
85. You G, Wang B, Li J, Chen A, Sun J. The prediction of MOE 99. Tagade PM, Adiga SP, Pandian S, Park MS, Hariharan KS,
of bamboo-wood composites by ANN models based on the Kolake SM. Attribute driven inverse materials design using
non-destructive vibration testing. J Build Eng. 2022;59:105078. deep learning Bayesian framework. Npj Comput Mater. 2019;
doi:10.1016/j.jobe.2022.105078 5(1):127. doi:10.1038/s41524-019-0263-3
86. Long W, Rice RW. Detection of structural damage in 100. Unni R, Yao K, Zheng Y. Deep convolutional mixture density
medium density fiberboard panels using neural network network for inverse Design of Layered Photonic Structures.
method. J Compos Mater. 2008;42(11):1133-1145. doi:10.1177/ ACS Photonics. 2020;7(10):2703-2712. doi:10.1021/acsphotoni
0021998308090455 cs.0c00630
87. Czarnecki S, Moj M. Comparative analyses of selected neural 101. Yeung C, Tsai R, Pham B, et al. Global inverse design across
networks for prediction of sustainable cementitious composite multiple photonic structure classes using generative deep learn-
subsurface tensile strength. Appl Sci. 2023;13(8):4817. doi:10. ing. Adv Opt Mater. 2021;9(20):2100548. doi:10.1002/adom.
3390/app13084817 202100548
88. Shrifan NHMM, Jawad GN, Isa NAM, Akbar MF. Microwave 102. Li M, Liu Y, Wong BCL, Gan VJL, Cheng JCP. Automated
nondestructive testing for defect detection in composites structural design optimization of steel reinforcement using
 15480569, 2025, 3, Downloaded from https://4spepublications.onlinelibrary.wiley.com/doi/10.1002/pc.29082 by Central Library, Wiley Online Library on [17/08/2025]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
1960 LIANG ET AL.

graph neural network and exploratory genetic algorithms. network. Eur J Wood Wood Prod. 2022;80(4):907-922. doi:10.
Autom Constr. 2023;146:104677. doi:10.1016/j.autcon.2022. 1007/s00107-022-01818-2
104677 116. Ozsahin S, Aydin I. Prediction of the optimum veneer drying
103. Peng B, Wei Y, Qin Y, et al. Machine learning-enabled con- temperature for good bonding in plywood manufacturing by
strained multi-objective design of architected materials. Nat means of artificial neural network. Wood Sci Technol. 2014;
Commun. 2023;14(1):6630. doi:10.1038/s41467-023-42415-y 48(1):59-70. doi:10.1007/s00226-013-0583-2
104. Zhu Y, Newbrook DW, Dai P, de Groot CHK, Huang R. 117. Yang YK, Yang RT, Tzeng CJ. Optimization of mechanical
Artificial neural network enabled accurate geometrical characteristics of short glass fiber and polytetrafluoroethylene
design and optimisation of thermoelectric generator. Appl reinforced polycarbonate composites using the neural net-
Energy. 2022;305:117800. doi:10.1016/j.apenergy.2021. work approach. Expert Syst Appl. 2012;39(3):3783-3792. doi:
117800 10.1016/j.eswa.2011.09.079
105. Maduabuchi C. Thermo-mechanical optimization of thermoelec- 118. Tan LB, Nhat NDP. Prediction and optimization of process
tric generators using deep learning artificial intelligence algo- parameters for composite thermoforming using a machine
rithms fed with verified finite element simulation data. Appl learning approach. Polymers. 2022;14(14):2838. doi:10.3390/
Energy. 2022;315:118943. doi:10.1016/j.apenergy.2022.118943 polym14142838
106. Raccuglia P, Elbert KC, Adler PDF, et al. Machine-learning- 119. Abueidda DW, Almasri M, Ammourah R, Ravaioli U,
assisted materials discovery using failed experiments. Nature. Jasiuk IM, Sobh NA. Prediction and optimization of mechani-
2016;533(7601):73-76. doi:10.1038/nature17439 cal properties of composites using convolutional neural net-
107. Zhou T, Gani R, Sundmacher K. Hybrid data-driven and works. Compos Struct. 2019;227:111264. doi:10.1016/j.compst
mechanistic modeling approaches for multiscale material ruct.2019.111264
and process design. Engineering. 2021;7(9):1231-1238. doi:10. 120. Hiraide K, Hirayama K, Endo K, Muramatsu M. Application
1016/j.eng.2020.12.022 of deep learning to inverse design of phase separation struc-
108. Kazi MK, Eljack F, Mahdi E. Optimal filler content for cotton ture in polymer alloy. Comput Mater Sci. 2021;190:110278.
fiber/PP composite based on mechanical properties using arti- doi:10.1016/j.commatsci.2021.110278
ficial neural network. Compos Struct. 2020;251:112654. doi:10. 121. Qian C, Tan RK, Ye W. Design of architectured composite
1016/j.compstruct.2020.112654 materials with an efficient, adaptive artificial neural network-
109. Zhou T, Song Z, Sundmacher K. Big data creates new oppor- based generative design method. Acta Mater. 2022;225:117548.
tunities for materials research: a review on methods and doi:10.1016/j.actamat.2021.117548
applications of machine learning for materials design. Engi- 122. Sun X, Yue L, Yu L, et al. Machine learning-evolutionary
neering. 2019;5(6):1017-1026. doi:10.1016/j.eng.2019.02.011 algorithm enabled design for 4D-printed active composite
110. Ha CS, Yao D, Xu Z, et al. Rapid inverse design of metamater- structures. Adv Funct Mater. 2022;32(10):2109805. doi:10.
ials based on prescribed mechanical behavior through 1002/adfm.202109805
machine learning. Nat Commun. 2023;14(1):5765. doi:10. 123. Ashank CS, Garg P, Kumar A, Agnihotri PK, Agrawal M.
1038/s41467-023-40854-1 Deep neural networks based predictive-generative framework
111. Sanchez-Lengeling B, Aspuru-Guzik A. Inverse molecular with data augmentation for designing composite materials.
design using machine learning: generative models for matter Model Simul Mater Sci Eng. 2022;30(7):075003. doi:10.1088/
engineering. Science. 2018;361(6400):360-365. doi:10.1126/ 1361-651X/ac88e8
science.aat2663 124. Patel D, Yang R, Wang J, Rai R, Dargush G. Deep learning-
112. Wang Q, Zhang L. Inverse design of glass structure with deep based inverse design framework for property targeted novel
graph neural networks. Nat Commun. 2021;12(1):5359. doi:10. architectured interpenetrating phase composites. Compos
1038/s41467-021-25490-x Struct. 2023;312:116783. doi:10.1016/j.compstruct.2023.116783
113. Chen C, Gu GX. Generative deep neural networks for inverse
materials design using backpropagation and active learning.
Adv Sci. 2020;7(5):1902607. doi:10.1002/advs.201902607
114. Shen C, Wang C, Wei X, Li Y, van der Zwaag S, Xu W. Physi- How to cite this article: Liang Y, Wei X, Peng Y,
cal metallurgy-guided machine learning and artificial intelli-
Wang X, Niu X. A review on recent applications of
gent design of ultrahigh-strength stainless steel. Acta Mater.
machine learning in mechanical properties of
2019;179:201-214. doi:10.1016/j.actamat.2019.08.033
115. Ozturk H, Demir A, Demirkir C. Optimization of pressing composites. Polym Compos. 2025;46(3):1939‐1960.
parameters for the best mechanical properties of wood doi:10.1002/pc.29082
veneer/polystyrene composite plywood using artificial neural