Page 19 - IJAMD-2-3

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

2.6. Summary of predictive models for composite models learn from existing unlabeled data to generate new
materials data. In other words, generative models can create entirely
The applications of predictive ML models in the field of new images, texts, designs, and structures, making them
composite materials are summarized in Table 1, along particularly valuable for various applications, including
data augmentation and innovative design exploration.
111,112
with the advantages and limitations of each type of ML One of the most well-known examples of generative
method. The ML approach should be tailored to the target models is ChatGPT, which demonstrates the potential
application, as no method is superior to another. The ML of this technology in natural language generation. In the
framework should be constructed based on the number of field of composite materials, obtaining experimental and
variable parameters, the required accuracy of predictions, simulation data often demands considerable time and
the type, quantity, and quality of the available data, as cost. Furthermore, the vast range of material combinations
well as the knowledge of the governing equations. It is and configurations leads to diverse shapes, properties,
important to understand the characteristics of the problem and characteristics, making the potential applications of
being modeled to select the most suitable ML method for generative models even more promising.
efficient training and accurate predictions.
3.1. Variational autoencoders (VAEs)
3. Generative models for composite
materials VAEs are the most common types of generative models
designed to generate data in a latent space. 113,114 As depicted
Predictive models are designed to predict the output in Figure 8A, VAEs primarily consist of an encoder and
results based on the given input data, whereas generative a decoder, where the encoder maps input data to a latent

Table 1. Applications, advantages, and limitations of predictive models for composite materials
ML method Applications Advantages Limitations
Generic DNNs - Prediction of nonlinear responses, mechanical - Efficient alternative to traditional - Dependence on large quantities of
properties, structural behaviors of composites, 29-31,33,34 numerical simulation high-quality data
and pressure vessel behavior 32
CNNs -Analysis of microstructure 50,110 - Accurate predictions on - Difficulties in collecting sufficient
- Prediction of stacking angle, 45,51 3D permeability, and image-based data high-quality images.
46
microstructural property 60 -High pre-processing load
- Detection of defect from IR stress measurements,
47
damage from lamb wave signals 59
RNNs - Prediction of nonlinear stress-strain response of - Ability to model sequential, -Need for large training datasets
composites (elasto-plastic, path-dependent), time-dependent data, such as stress - Limited capability in spatial pattern
70
long-term water absorption and swelling, dynamic history or environmental exposure extraction compared to CNNs
65
loading behavior in CMCs and concrete granite -Support for variable-length data - Susceptibility to vanishing/
composites, 66,71 shape transformation in 4D printed - Use of advanced variants (LSTM, exploding gradients in generic
composites, and fatigue behavior of composites 80 GRU, and Bi-RNN) to enhance RNNs
73
- Combination with CNN for spatial-temporal prediction stability, memory
modeling 68,83-85 retention, and directionality
PINNs -Thermochemical curing process 91 - No requirement for labelled training - Accuracy dependence on the
- Predict composition-property relationships of basalt data compatibility of governing equations
fibers, acoustic source localization in anisotropic - Physically-consistent predictions with the modeled process being.
92
composites, and property degradation in carbon beyond the training domain - Limitations due to idealized
90
fiber/epoxy composites 95 assumptions in governing equations
vs. noise and variability in
real-world data
XAI - Analysis of feature importance for composite pressure - Improved transparency in the - Trade-off between explainability and
vessel performance 32 decision-making process and predictive accuracy compared to
- Prediction of the mechanical strength of the prediction reliability complex black-box models
textile-reinforced beam and defects in composite - Visualization of influential factors
107
materials 106
-Design of FRP laminate 104
-Detection of visual delamination 105
Abbreviations: Bi-RNN: Bidirectional recurrent neural network; CMCs: Ceramic matrix composites; CNNs: Convolutional neural network;
DNNs: Deep neural networks; FRP: Fiber-reinforced polymer; GRU: Gated recurrent unit; IR: Infrared; LSTM: Long short-term memory;
ML: Machine learning; PINNs: Physics-informed neural networks; RNNs: Recurrent neural network; XAI: Explainable artificial intelligence.

Volume 2 Issue 3 (2025) 13 doi: 10.36922/IJAMD025210016