Predicting long-term degradation of prestressed bridge girders using a hybrid finite element and deep learning approach

Prestressed concrete girders are widely used in bridge and railway infrastructure. However, their long-term performance is highly vulnerable to environmental effects, particularly corrosion, which significantly compromises structural serviceability and durability. Accurately quantifying corrosion-induced deterioration over extended periods remains challenging due to the substantial time and computational demands of conventional finite element analyses. This study introduces an integrated numerical and data-driven framework to efficiently evaluate the long-term corrosion effects on prestressed concrete structures. A prestressed Advanced Railroad Trivet (ART) girder was modelled in ABAQUS, incorporating time-dependent corrosion degradation and 20 years of recorded historical temperature data. Corrosion effects were simulated through progressive reductions in the elastic modulus of prestressing tendons based on empirical relationships, while temperature variations were applied using historical meteorological data to simulate long-term deflection responses. The long-term response predicted in the present study is limited to the components associated with corrosion-induced tendon degradation and temperature variation; concrete creep, shrinkage, and time-dependent prestress losses are not included. To overcome the high computational cost associated with extended Finite Element Analysis (FEA), a Long Short-Term Memory (LSTM) neural network was trained using the finite element-generated dataset. Data normalization, a sliding window approach, and bayesian optimization were implemented to enhance model accuracy. The optimized LSTM model achieved a Root Mean Square Error (RMSE) of 1.42 and a coefficient of determination (R 2 ) of 0.97, showing strong agreement with the original data.