Page 15 - IJAMD-2-3

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

coefficients, and strain inputs, and showed improved extract spatial information from RVE images, which was
prediction accuracy compared to backpropagation then converted into sequences and fed into the LSTM
neural network and random forest models. RNNs have to predict the stress at each time step. CNN-RNN-based
also been employed to predict the shape transformation models have also been applied to predict the evolution
of 4D-printed active composite structures. Sun et al. of surface roughness in carbon/carbon composites under
73
designed an active composite beam using materials with thermochemical ablation and to assess damage in
83
different coefficients of thermal expansion and treated composites subjected to low-velocity impacts. According
85
the material distribution as sequential data. By applying to a study by Truong et al., combining CNN with GRU
84
an RNN model, they successfully predicted the resulting yielded the highest prediction accuracy compared to using
shape under a stimulus of a 100°C temperature increase, CNN, RNN, GRU, or LSTM individually.
achieving over 2000 times faster performance compared Moreover, traditional RNNs are less robust to new
to conventional finite element simulations. In addition, conditions due to the absence of physically derived patterns,
to enhance the performance of RNNs, bidirectional
RNNs – which process input sequences in both forward which prompted a series of studies to explore physically
and backward directions – have been employed in studies RNN (PRNN) models that incorporate physical constraints
69
69,86
to identify the location and size of defects in laminated directly into the RNN architecture. Maia et al. aimed
composites and to efficiently predict the time-dependent to predict the path- and rate-dependent stress responses
77
damage of composite hydrogen tanks. 78 of composites by embedding the constitutive models into
the material layer of the PRNN, as shown in Figure 5F.
However, RNNs require a large amount of data for This approach enabled them to incorporate the non-linear
training and are difficult to capture spatial patterns elasto-viscoplastic, path- and rate-dependent behavior of
compared to CNNs. To address these limitations, recent the matrix, as well as the nonlinear elastic and anisotropic
studies have explored combining RNNs with techniques characteristics of the fibers, into the RNN architecture.
such as transfer learning 67,79-82 and CNNs. 68,83-85 Cheung As a result, the model accurately predicted stresses with a
et al. combined transfer learning with a GRU-based DL mean absolute error of 5–7 MPa and a relative error within
67
model to reduce the computational cost of predicting the 5–10% under various loading conditions, and was also
nonlinear elasto-plastic behavior of short fiber reinforced used to successfully extrapolate to loading conditions not
composites, as shown in Figure 5D. In this approach, a included in the training data.
mean-field-based RNN model was used with low-fidelity
samples, while high-fidelity full-field simulation data were 2.4. Physics-informed neural networks (PINNs)
employed for fine-tuning. As a result, they successfully DNNs accurately represent the relationship between the
developed a framework with excellent computational given inputs and outputs based on large-scale labeled
efficiency and generalization performance. Jian et al. also datasets. However, due to their purely data-driven learning
80
employed a combination of bidirectional LSTM and transfer approach, the relationship itself remains a black box,
learning to predict the behavior of CFRP composites in the which may limit their ability to reflect physical phenomena
very high cycle fatigue regime. In their approach, transfer reliably. In contrast, PINNs integrate physical laws into the
learning was applied to predict the fatigue behavior of a neural network learning process, as shown in Figure 6A. 87,88
new batch by leveraging existing experimental data. An PINNs do not require any labeled (input-to-output) data
attention mechanism was incorporated to focus on critical for training, and the governing partial differential equation
segments of the fatigue curve, such as regions with abrupt (PDE), as well as initial and boundary conditions that
changes, thereby improving prediction accuracy and describe the process being modeled, are used to train the
convergence speed. However, attention mechanisms, while neural network.
advantageous for highly complex, imbalanced, or long-
dependency data, may lead to overfitting and increased In the case of composite materials, solving the various
computational time when applied to low-complexity or PDEs that describe the elastic behavior, heat transfer,
relatively uniform data. Therefore, careful consideration and damage or fracture models through FEA is time-
78
of the characteristics of the data is essential when applying consuming, and ML-based predictions often fail to
attention mechanisms. As shown in Figure 5E, El Said reflect real physical phenomena. In contrast, predictions
68
employed a deep recurrent CNN that combines CNN generated by PINNs can be consistent with the specified
and LSTM to predict the nonlinear damage behavior physical laws, even outside the training domain, making
of laminated composite structures. In particular, since PINNs a powerful tool for efficiently predicting complex
features such as layup configuration, wrinkles, and ply phenomena in composites research, such as plastic
gaps exhibit spatial characteristics, a CNN was used to deformation, flow through porous media, and heat

Volume 2 Issue 3 (2025) 9 doi: 10.36922/IJAMD025210016