AI-based prediction of seismic time-history responses of RC frame structures considering varied structural parameters

In this paper, an end-to-end framework for Intelligent Seismic Response Prediction, ISRPnet, is introduced. ISRPnet comprises a structural parameter module for discretizing reinforced concrete frame structures into a series of static features and an encoder-decoder architecture for encoding seismic loads and autoregressively predicting seismic responses. The model is trained on a data set of 16,544 cases generated through validated fibre-based finite element models. ISRPnet achieves promising performance on both frequent and rare earthquakes. ISRPnet rapidly and highly precisely predicts temporal responses for frequent earthquakes. The peak displacement predictions remain accurate for rare earthquakes. The superiority of the physical loss and the advantages of gated recurrent unit over long short-term memory are analysed in comparative experiments. Verification with unseen seismic waves beyond the training data shows the robust generalization and extrapolation capabilities of the framework. The proposed model accomplishes efficient surrogate computation of the full-process seismic response for a class of RC frame structures.