A machine learning-driven approach to predict mechanical degradation associated with matrix cracks in fiber-reinforced composite laminates

Matrix cracking is an early, critical damage mechanism in fiber-reinforced plastic laminates, significantly affecting the mechanical properties such as stiffness and residual strain. This study aims to develop machine learning models capable of predicting stiffness reduction and residual strain in carbon fiber-reinforced plastic and glass fiber-reinforced plastic laminates containing off-axis plies. This work focuses on modeling under conditions with limited experimental data and a high number of input features. Two tree-based machine learning models, namely Random Forest and Light Gradient Boosting Machine (LightGBM), were trained using experimental datasets derived from laminates with various stacking configurations and material systems. Input features included material properties, laminate configurations, applied loading conditions, and crack density. Bayesian optimization was used for hyperparameter tuning, and Recursive Feature Elimination with Cross-Validation (RFECV) was applied to reduce the number of input features while preserving model performance. Both models achieved high predictive accuracy in cross-validation and performance evaluations, with mean cross-validation R² values of up to 0.83 for stiffness reduction and 0.88 for residual strain. LightGBM performed well with the full feature set, whereas Random Forest benefited significantly from feature selection, leading to improved generalization. Validation using previously unseen experimental data confirmed that both models accurately predicted the mechanical behavior of virgin and cracked laminates. Overall, the results suggest that the combination of Random Forest and RFECV is especially effective when working with small datasets and high-dimensional inputs.