Transfer learning-enhanced surrogate model for the prediction of structural full-profile seismic time-history response

Structural dynamic analysis is an important tool for simulating seismic damage of buildings. However, the time-history analysis methods are computationally inefficient and current data-driven methods commonly lack the generalization capability to structures. In this study, a transfer learning-enhanced surrogate model is developed for the rapid prediction of structural full-profile (story-level) seismic time-history response. First, a nearly lossless data processing method is used to unify the seismic response data dimensions of various structures, making the proposed method applicable to structures with any number of stories. Then, the structural feature extracted by the graph neural network and the ground motion feature extracted by the multi-scale convolutional neural network are fed into the attention-enhanced long short-term memory layers together for feature fusion, which then outputs the structural full-profile response. The validation results on numerical cases show that the proposed method can accurately predict the seismic time-history response of new structures with a number of stories not exceeding the training set range. After fine-tuning the surrogate model with a few new data, it is applicable to the seismic response prediction of structures with a number of stories exceeding the training set range.