Interpretable multi-branch physics-informed neural networks for structural health assessment of reinforced concrete beams

Abstract

The relationship between observable damage and residual shear performance remains unclear. This study proposes an interpretable multi-branch Physics-Informed Neural Network (PINN) to predict the residual shear capacity of Reinforced Concrete (RC) beams using 583 experimental samples. The model uses a physics-guided multi-branch architecture to separate concrete, reinforcement, and crack-induced effects, improving both accuracy and interpretability. Monotonic and range constraints were applied to ensure physically consistent degradation and valid predictions. The proposed model achieved superior performance, with an R² of 0.934 and an RMSE of 6.44%, outperforming conventional machine learning models. Ablation studies confirmed the effectiveness of both the multi-branch architecture and physics-informed constraints. SHAP analysis further revealed that crack width is the dominant factor governing residual shear capacity and captured the transition from concrete-dominated to reinforcement-dominated shear resistance with increasing cracking.