Page 17 - IJAMD-2-3

International Journal of AI for
Materials and Design AI applications in composite materials

demonstrated high precision and reliability, with errors fabric, including localized variations. Alternatively, PINNs
in the order of 10 . Compared to conventional numerical can be combined with numerical solvers to enhance
−5
methods, such as the finite element method and finite training when simulating multi-scale problems, such as
difference approaches, the PINN framework reduced the the dual-scale permeability of fibrous reinforcements.
100
computation time by over an order of magnitude. Furthermore, applying active learning to efficiently select
important training data is expected to provide a more
To better represent the response of real fibrous 101
reinforcements, Lee et al. incorporated experimental data effective learning.
89
of the permeability and compressibility relations into the 2.5. Explainable AI (XAI)
PINN training process to explain the remaining physics
not expressed in the governing equations, as presented XAI is an approach that aims to provide clear and
in Figure 6B. Real-time simulations of transverse interpretable explanations for the results generated by DL
102,103
permeability measurements were also conducted to extract models. Typically, AI functions as a black box, making
the permeability relation using experimental data of the predictions without offering sufficient insight into the
flow rate and injection pressure. reasoning behind them. However, as illustrated in Figure 7A,
the reliability of AI-driven results can be improved by
These examples demonstrate how PINNs are utilizing XAI, which provides a transparent explanation of
transforming composite materials research using more the decision-making process. This transparency not only
reliable, faster, and data-efficient methods for prediction, enhances the reliability of AI systems but also enables users
design, and analysis. While PINNs offer significant to better understand and interpret the factors influencing
potential in the field of composite materials, conventional the model’s predictions, making AI more accessible in
PINN models may struggle to fully capture the complex various applications. When using AI for the analysis and
multi-physics and multi-scale issues inherent in composite design of composite materials, simply obtaining the output
materials. Obtaining a good balance between physics and values of the results is insufficient. In particular, composite
data can also be challenging, potentially resulting in slower materials exhibit nonlinear properties and are sensitive to
learning rates or difficulties in convergence. In addition, the numerous factors, requiring careful analysis. While XAI is
training process is sensitive to network hyperparameters not a type of DL model, such as DNNs or CNNs, it has
and can fall into local minima, and research is ongoing to been widely used with ML models to predict composite
improve the training performance of PINNs. 96 defects, damage, and mechanical properties, explaining
To overcome these limitations, it is necessary to the decision-making process and improving the reliability
enhance model performance by utilizing hybrid models of predictive models.
that combine data-driven neural networks with physics- Several studies have highlighted the potential of XAI in
based learning, enabling the simultaneous integration of this field. 32,104-109 Hong et al. developed a predictive model
32
physical laws and data-based training. 97,98 for the behavior of a composite pressure vessel and utilized
XAI to analyze the impact of stacking angle and layer
A notable advancement in this field is the development
of a transfer learning-based multi-fidelity PINN framework thickness on its performance using permutation feature
106
(mfPINN) to mitigate the effects of sparse and noisy importance for each design parameter. Daghigh et al.
sensor data, as well as the limited accuracy of idealized developed an XAI-based ML model for predicting defect
90
governing equations. A low-fidelity PINN model is characteristics (depth, size, and thickness) in composite
first trained based on the governing physics. Then, the materials using infrared thermography (IRT) data and
evaluated the feature importance of each parameter.
information learned is transferred to a data-driven DNN, Yossef et al. predicted the flexural strength of composite
104
which is trained on a smaller set of high-fidelity data, as laminates using a decision tree-based gradient boosting
shown in Figure 6C. The mfPINN framework is used for algorithm as the ML model. They combined the ML model
acoustic source localization in anisotropic composites and with XAI to analyze how the design parameters influence
is shown to accurately predict the location of defects or laminate strength and to provide a visual interpretation
signal sources within composite materials. By integrating of the relationships between these factors (Figure 7B).
high-fidelity and low-fidelity data, the mfPINN enhances Azad and Kim developed an XAI-based model with
105
localization accuracy while preventing data overfitting vision transformers to detect damage in CFRP composites,
and reducing the need for expensive high-resolution data, as shown in Figure 7C. Their approach provided visual
which is often difficult to obtain in practice.
insights into areas of delamination, enhancing the damage
In addition, Hanna et al. used PINNs with CNNs to detection process. Similarly, Song et al. applied XAI to
107
99
predict the 2D permeability tensor field of a glass fiber textile-reinforced mortar beams to predict the bending
Volume 2 Issue 3 (2025) 11 doi: 10.36922/IJAMD025210016