Mechanical degradation mechanisms of prestressing steel strands subject to stress–corrosion coupling: Probabilistic physics‐constrained neural network modeling

Abstract

Quantifying the mechanical degradation of prestressing steel strands subject to stress–corrosion coupling is essential for safety assessment and service‐life prediction of in‐service prestressed structures. This study conducts long‐term corrosion tests to characterize stress–strain responses and fracture behaviors under different initial stress levels and mass loss ratios. Degradation models for ultimate strength and ultimate strain are established by accounting for stress‐coupling effects. A random pitting finite element model is then developed and validated to generate a sample database covering multiple combinations of corrosion levels and stress conditions. By embedding the experimentally derived degradation models as physical priors, a probabilistic physics‐constrained neural network (PPCNN) is proposed for joint prediction of ultimate strength and ultimate strain, providing both point estimates and prediction intervals. Results show that stress–corrosion coupling accelerates the degradation of strength and ductility and increases the probability of brittle fracture, with a rapid ductile‐to‐brittle transition at a mass loss of approximately 16%–18%. On the independent test set, the PPCNN achieves an R ² /RMSE of 0.9389/29.0372 for ultimate strength and 0.9780/8.3 × 10 ⁻⁴ for ultimate strain, outperforming a purely data‐driven model (0.8629/43.4977 and 0.9259/1.5 × 10 ⁻³ ). Experimental validation further indicates that the PPCNN prediction intervals achieve high coverage, supporting reliable probabilistic interval prediction subject to stress–corrosion coupling.