A transfer learning PINN incorporating multi-physics subnetwork for corrosion fatigue life prediction of high-performance structural steels and butt welds using multi-source data

Corrosion fatigue is a critical degradation mechanism that significantly limits the service life of high-performance structural steels and butt welds used in marine and coastal infrastructure. Traditional fatigue life assessment relies on extensive laboratory testing, which is costly, time-consuming, and insufficient to capture the complex interactions between cyclic loading and aggressive marine environments. To overcome these limitations, this study develops a physics–data-driven framework with multi-physics subnetwork embedding for predicting the corrosion fatigue life of multi-source high-performance structural steels and their butt welds. A physics-informed neural network (PINN) is constructed by integrating data screening, feature reduction, and normalization with physical constraints derived from fatigue evolution. A multi-physics-subnetwork-embedded PINN (MPS-PINN) is further introduced to adaptively calibrate continuum damage mechanics parameters from multi-source corrosion–fatigue data. Compared with conventional artificial neural networks, the MPS-PINN improves prediction accuracy by an average of 17.79% and reduces physical-constraint violations from 34% to below 1%. For butt welds, the transfer-learning-enhanced MPS-PINN (TL-MPS-PINN) achieves an average accuracy gain of 17.12%. SHAP indicates that corrosion damage and applied stress dominate steel fatigue performance, whereas groove angle plays a critical role in weld behavior. The proposed frameworks offer accurate and physically consistent tools for marine structural design and life-cycle management.