Page 13 - IJAMD-2-3

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

Similarly, as with DNNs, many studies have also utilized or spatial-order relationships. Unlike feedforward neural
transfer learning techniques to enhance the training networks or CNNs, RNNs receive data one step at a time in
efficiency of CNNs. Kojima et al. used transfer learning sequential order, as shown in Figure 5A, and have feedback
47
with CNNs to predict internal defects in carbon fiber- connections in the hidden layers to recognize the sequential
reinforced polymer (CFRP) laminates by combining data structure of the data. This architecture is advantageous in
generated through FEA with infrared stress distributions problems where the order of inputs is important, and it
obtained from experiments. Figure 4D presents the also offers the flexibility to handle variable-length data.
ground truth of defects for each layer of the composite However, RNNs have a drawback known as the long-term
laminate, together with the predicted results obtained dependency problem, where the influence of distant inputs
through transfer learning. By fine-tuning these models diminishes over time despite their relevance. In addition, as
with a limited set of experimental data, transfer learning weights are multiplied across time steps, the model suffers
improves their generalization capabilities across various from vanishing or exploding gradient issues. To address
composite structures. The results show that as the amount these problems, advanced RNN-based models, such as
of experimental stress measurement data increases, transfer long short-term memory (LSTM) and gated recurrent unit
learning-based defect prediction outperforms traditional (GRU), are commonly used (Figure 5A). LSTM utilizes
FEA-based approaches in terms of efficiency. Moreover, Liu three gates to selectively store, transmit, and discard
et al. aimed to diagnose damage in composite materials information, and uses a memory cell to retain long-term
59
using transfer learning with CNNs by leveraging Lamb wave data. GRU, a simplified version of LSTM, applies an update
signals generated from numerical simulations and applying gate to determine how much of the previous information
them to experimental data. By integrating transfer learning to keep, and a reset gate to decide how much of the past
with domain adversarial learning, the model achieved information to forget, thereby overcoming the limitations
high generalization performance across complex domain of conventional RNNs.
discrepancies, leading to highly accurate delamination Since composites exhibit highly nonlinear and time-
prediction. In addition to cases of domain adaptation, dependent behavior, RNNs have been actively studied
where only the domains differ while the tasks remain the and applied in various research areas. 65,66,70-76 Chen et al.
70
same as in the aforementioned studies, transfer learning can predicted the nonlinear behavior of composites, such as
also be applied when there are differences in the source and the time-dependent nonlinear elastic-plastic response that
target tasks. Xu et al. extracted vector representations of varies with the strain/loading path, using an LSTM network
60
the geometric morphology and distribution characteristics capable of processing the entire time sequence. As a result,
of two-phase composites from images to establish similarity the trained LSTM model accurately predicted stress with
between the source and target domains. Their method a maximum error of <4%, and the model successfully
enabled the use of small datasets through transfer learning, reproduced the stress-strain hysteresis curves even under
yielding high prediction accuracy for mechanical properties, nonlinear cyclic loading conditions. Yousefi et al. also
65
such as stiffness, strength, and toughness. Techniques employed an LSTM DL model, as shown in Figure 5B,
such as data augmentation and transfer learning not only to predict the long-term water absorption and thickness
enhance the efficiency of predictive models but also enable swelling behavior of a composite composed of hydrophilic
broad applicability across diverse tasks and DL models. cellulose fibers and hydrophobic polypropylene. By
In addition, developing hybrid models by integrating training the model with the initial 200 h of data obtained
CNNs with other ML/DL methods holds promise for from immersion experiments, they demonstrated that
enhancing their ability to capture and analyze complex the LSTM can effectively predict the long-term physical
composite material structures. 61-64 Among various behavior of the composite for the subsequent 1300 h.
approaches, attention mechanism-based CNNs have Furthermore, RNN-based models have also been utilized
demonstrated strong capabilities in capturing complex to predict the behavior of composites under dynamic
failure behaviors. Liu et al. incorporated an attention- loading conditions. Borkowski et al. employed a GRU to
64
71
based loss function into a CNN framework, enabling the predict the response of ceramic matrix composites (CMCs)
model to focus more on high-damage regions in thick under non-monotonic loading and microstructural
composites with wrinkles, thereby achieving effective variations, and demonstrated that stable predictions
prediction of failure modes. could be achieved without overfitting across 60 test sets.
66
Gao et al. applied an LSTM to predict the axial stress of
2.3. Recurrent neural networks (RNNs) concrete-granite composites under impact conditions, as
RNNs are a type of DL model designed to process sequential shown in Figure 5C. The trained model was able to predict
data, and they excel in handling data that contain temporal stress responses to new impact velocities, joint roughness

Volume 2 Issue 3 (2025) 7 doi: 10.36922/IJAMD025210016