第１部門④

[21006-11-02] Deep kernel learning surrogate model with extracting seismic feature for seismic risk analysis (Proceedings of Symposium on Applied Mechanics)

*Taisei Saida¹, Mayuko Nishio¹ (1. University of Tsukuba)

Keywords:surrogate model, Gaussian process regression, explainability, seismic risk analysis, seismic response analysis

This study developed a deep kernel learning surrogate model to reduce computational costs of a seismic risk analysis. This model uses convolutional neural networks (CNN) to extract seismic load features. Furthermore, the explainability of the surrogate model are obtained by estimating the contribution of each part of the seismic load by Gradient-weighted Class Activation Mapping (Grad-CAM) and by estimating the contribution of each structural parameter by ARD. In the validation, a surrogate model was constructed for the seismic response analysis of isolated RC piers. It was shown that the computational cost can be reduced and that the model has explainability.