Prediction of flexural capacity of corroded reinforced concrete beams by physics-informed neural network

A reasonable prediction of the flexural capacity of corroded reinforced concrete (RC) beams is essential for bridge reliability assessment. This study develops a physics-informed neural network (PINN) model to predict the flexural capacity of corroded RC beams. The performance of this model is evaluated systematically and compared with three deep learning methods, namely deep neural network (DNN), back propagation neural network (BPNN) and radial basis function neural network (RBF), and two classical machine learning algorithms, namely decision tree regression (DTR) and support vector regression (SVR). The effect of a standard loss term weight and three additional loss term weights based on physical knowledge is investigated, and the optimal hyperparameters for the self-adaptive loss balance method are determined to obtain accurate weights for these loss terms. Validation on an independent experimental dataset demonstrated the model's superior generalization capability. The comparison results show that the PINN model achieves a coefficient of determination (R²) of 0.9955, with root mean square error (RMSE) reductions of 7.87%, 20.11%, 10.10%, 78.17%, and 43.91% compared to DNN, BPNN, RBF, DTR, and SVR, respectively. Error analysis indicates that physics-constrained loss and engineering-constrained loss dominate the PINN's errors. On the independent experimental dataset, the PINN model maintains reliable accuracy, with a mean absolute percentage error (MAPE) of 3.4%. Additionally, SHapley Additive exPlanation (SHAP) analysis identified the effective height of section ( h 0 ) and cross-sectional area of tensile reinforcement after corrosion ( A st ) as the most important parameters. This study provides a novel hybrid driven framework for evaluating flexural capacity of corroded RC beams.