Surrogate modeling for the long-term behavior of PC bridges via FEM analyses and long short-term neural networks

Accurate prediction of long-term behavior of prestressed concrete (PC) bridges is computationally expensive and complex. One main reason is that accurate finite element (FE) modeling of coupled "creep-shrinkage-crackingrelaxation" behavior is still far from accuracy and sufficiency. To this end, this study proposes a practical surrogate modeling procedure to predict long-term behavior of large-scale PC bridges, by combining the long shortterm memory (LSTM) neural networks with principal component analysis (PCA). The FE modeling and its implementation are firstly presented, and it shows that the B4 model, with additional consideration of concrete cracking and nonlinear concrete creep, could effectively capture the long-term deflection of the real bridge. Five critical variables are randomly generated via Latin Hypercube Sampling (LHS), including creep coefficient, shrinkage coefficient, concrete strength, prestressing level, and ambient humidity. Subsequently, mid-span deflection, prestressing loss and cracking patterns are extracted from FE results to form the database, with the aid of Python script. Before training, high-dimensional data is reduced through principal component analysis (PCA) technique. Long short-term neural networks are subsequently introduced, with customized loss functions, to achieve accurate and robust predictions. Results show that the proposed PCA-LSTM surrogate models are efficient and accurate in predicting long-term behavior of PC bridges.