Prediction of long-period ground motion responses for high-rise buildings using physics-assisted fully convolutional neural network

Far-field long-period ground motions (LPGMs) have a significant long-duration characteristic and can cause severe damages to high-rise buildings in the LPGM potential area. Rapid and accurate prediction of LPGM-induced responses is crucial for earthquake emergency management and loss assessment in large cities. However, dynamic-based methods for calculating structural responses involve intricate physical modeling process, resulting in low computational efficiency and failing to meet the prediction timeliness requirement. Furthermore, current data-driven methods cannot satisfy the prediction accuracy requirement due to the harsh formation condition causing scarcity and particularity of LPGM records. To this end, this paper develops a physics-assisted fully convolutional neural network (PhyFCN) for predicting LPGM-induced response of high-rise buildings. Central to this method lies in encoding the complex seismic motion equation into FCN for formulating an innovative physical loss function. The utilization of this function not only significantly enhances the prediction accuracy and robustness, but also remarkably reduces the dependence on the large number of training samples, i.e., a small number of training samples is sufficient for acquiring the desired parameters. This method integrates strengths of both dynamic-based methods and data-driven methods, enabling it to simultaneously fulfill requirements of prediction timeliness and accuracy. Numerical examples based on two different buildings are employed to verify the effectiveness and superiority of PhyFCN. Results suggest that the incorporation of the physics-assisted mechanism into FCN, even when only one training sample is available, can sharply increase the R value by 28.4 %, while those of MSE and MAE are decreased by 60.2 % and 37.6 %, respectively.