Bokai Song

Reliable durability assessment of prestressed high-strength concrete (PHC) pipe piles in marine environments requires accurate characterization of chloride transport parameters. However, these parameters are inherently layer-dependent because of centrifugal forming, and they are difficult to identify from limited field or experimental observations. This study proposes an inverse-to-forward spatiotemporal interface-constrained physics-informed neural network (IF-STIC-PINN) for identifying layer-dependent chloride diffusion coefficients from sparse short-term observations and subsequently predicting long-term chloride ingress and corrosion initiation. Chloride transport in PHC pipe piles is formulated as a radial convection-diffusion process in a centrifugally stratified annular domain. The governing equations, initial and boundary conditions, and interface continuity conditions are embedded into the neural-network training through physics-informed residuals and automatic differentiation. To represent material heterogeneity, separate subnetworks are constructed for the inner and outer layers of the PHC pipe pile, while interface constraints enforce the continuity of chloride concentration and total flux across the centrifugal-layer boundary. Temporal segmentation is further introduced to enhance the stability of long-term prediction. The effects of monitoring duration, sensor placement, spatial and temporal decomposition, and observation noise are systematically investigated. The results show that monitoring near the inner and outer exposure boundaries for approximately two to three years enables stable identification of layer-dependent diffusion coefficients. The proposed IF-STIC-PINN reconstructs the 100-year chloride concentration field with an relative error of approximately 0.72% and predicts a corrosion initiation time of 26.79 years, compared with 26.72 years obtained from COMSOL simulation. Validation against four published chloride-ingress cases yields concentration errors of 1.60%–2.36% and corrosion-initiation-time errors below 1.05%. These results demonstrate that IF-STIC-PINN provides a physics-consistent, data-efficient, and robust computational framework for translating sparse short-term chloride observations into long-term corrosion-initiation assessment of marine PHC pipe piles.