Bayesian Physics‐Informed Neural Networks for Probabilistic Fatigue Life Estimation of Corroded Steel Reinforcement

The mechanical properties of reinforcing steel are fundamental to the durability of concrete structures. In service environments, corrosion of reinforcing bars is pervasive, and corrosion–fatigue failure constitutes a major driver of structural degradation, particularly in bridge infrastructure. Conventional data‐driven approaches, such as probabilistic neural networks ( PNN ), often assume a normal distribution to construct point predictions and confidence intervals. This assumption can obscure the intrinsic distributional characteristics of corrosion–fatigue life and offers limited physical interpretability, thereby constraining reliability in engineering decision‐making. To overcome these limitations, this study proposes a fatigue‐life prediction framework for corroded reinforcing bars based on Bayesian Physically Informed Neural Networks ( BPINN ). The proposed approach employs variational Bayesian inference to update priors into posterior distributions directly from fatigue data, enabling rigorous uncertainty quantification without imposing a predefined distributional form. Moreover, physics‐based constraints are embedded into the network through the training objective, guiding learning toward physically admissible solutions and improving both predictive accuracy and interpretability. Experimental results show that BPINN consistently outperforms conventional neural networks in prediction performance. Unlike PNN , BPINN captures the empirical distributional features of fatigue life and reveals substantial deviations from normality. Overall, integrating physical laws with Bayesian learning enhances robustness and interpretability, underscoring the importance of physics‐grounded uncertainty quantification and distribution‐aware modeling for assessing corrosion–fatigue performance of reinforced concrete structures.