深層学習による構造物地震応答予測における工学的な説明性とその妥当性の検証

2026-06-03·
才田 大聖
,
西尾 真由子
· 0 min read
Abstract
In structural seismic response prediction, the “black-box” nature of deep learning models poses a significant barrier to their practical application. To gain the trust of engineers, it is essential that the rationale behind AI predictions aligns with physical phenomena and is interpretable from an engineering standpoint. This study focuses on “frequency”—a fundamental parameter in structural dynamics—and applies ExSRNet (Explainable Seismic Response Networks), a deep learning model capable of quantifying and visualizing the contribution of individual frequency bands of input ground motions. The validity of the model is evaluated through the seismic response prediction of a seismic isolation reinforced concrete bridge pier. Results show that the frequency contributions estimated by ExSRNet align well with the modal frequencies of the structure during linear responses, and exhibit a broader frequency distribution during nonlinear responses, accurately reflecting the variations in structural characteristics such as stiffness degradation. Furthermore, the model successfully predicts complex nonlinear time-history responses.
Date
2026-06-03 10:15
Event
Location

Nagoya in Japan

events