A residual stiffness prediction approach for carbon fiber reinforced composite materials based on interpretable machine learning algorithms

The conventional standard regression models often struggle to accurately predict the residual stiffness of Carbon Fiber Reinforced Polymers (CFRP), while data-driven approaches typically lack interpretability. To address these challenges, we introduce a novel method that integrates Back Propagation Neural Network (BPNN) with Symbolic Regression (SR). A stiffness degradation dataset is constructed using static and fatigue test data from T800. Key features, including stress level, normalized life, and strength, are selected through methods such as Pearson Correlation Coefficient (PCC), Max-Relevance and Min-Redundancy (mRMR), and SHAP analysis. SR is employed to uncover clear physical principles, while BPNN effectively captures complex relationships among multiple parameters. The results indicate that SR significantly outperforms traditional models in predicting the combined effects of stress level and normalized life. Additionally, BPNN demonstrates greater accuracy in predicting the interactions among stress level, normalized life and strength, maintaining low prediction errors across varying conditions. This integrated framework successfully merges physical interpretability with the capacity to model intricate relationships, offering a valuable tool for precise and transparent fatigue damage assessment in composite materials.