Interpretable machine learning for predicting the remaining life of matrix microcracks in CFRP

Carbon fiber-reinforced polymer (CFRP) composites offer excellent axial mechanical properties but are prone to premature failure under transverse loading due to matrix microcracking. Accurate prediction of the remaining life of these microcracks is essential for structural safety and reliability. Traditional finite element methods (FEM), while effective for initial life prediction based on fracture mechanics, become computationally expensive when simulating crack propagation in large-scale models with multiple parameters. Meanwhile, common machine learning (ML) methods often lack physical interpretability, limiting their practical application. To overcome these issues, this study proposes an integrated framework combining finite element modeling of a representative volume element (RVE) with machine learning to predict the remaining life of transverse microcracks and identify dominant influencing factors. Using ABAQUS and FRANC3D, high-fidelity fracture mechanics simulations generated a dataset linking microcrack size, orientation, stress ratio, and local stress to stress intensity factor ranges and resulting remaining life. Several ML models were trained and evaluated, with the eXtreme Gradient Boosting (XGBoost) algorithm achieving superior performance (R 2 > 0.90 on the test set). Explainable AI (XAI) techniques, specifically SHAP (SHapley Additive exPlanations), were applied to interpret the model. SHAP analysis identified stress ratio (58.1% contribution) and maximum von Mises stress (29.8% contribution) as the dominant factors affecting remaining life, while initial crack geometry had a secondary influence. These findings offer important physical insights into microcrack-driven failure and establish an efficient, interpretable methodology for predicting the remaining life of composite materials.