Abstract
In this study, an explainable seismic response surrogate model, which can be deemed valid from an engineering perspective, was developed. The model employs SincNet convolution to segment features by frequency band and an attention mechanism to quantitatively determine the contribution of features from each frequency band to the model’s predictive outcomes. For verification, a surrogate model was constructed for the seismic response analysis of isolated RC bridge piers, and its accuracy, along with the contribution of frequency bands as determined by attention, was evaluated. The results demonstrated that the surrogate model could predict seismic responses with sufficient accuracy. Furthermore, the contribution of frequency bands as identified through attention was found to be reasonable in relation to the natural frequencies of the RC bridge piers.