[2K1-OS-6] OS-6

[2K1-OS-6-03] Application of Waveform Generative AI Techniques to Earthquake Ground Motion

〇Takahiro Matsui¹, Atsuko Oana¹, Masahito Tsukano, Misaki Yokoji¹, Toru Ishii¹, Kei Furukawa¹, Shunichi Taniguchi² (1. Shimizu Corporation, 2. Lightblue, Co., Ltd.)

Keywords:Earthquake ground motion, Waveform, Machine learning, Earthquake-resistant design

Dynamic response analysis for the earthquake-resistant design requires design earthquake ground motions that take into account specific source, path, and site characteristics corresponding to anticipated earthquakes and construction sites. Among the conventional methods for evaluating design ground motions, simplified methods cannot adequately capture the characteristics of natural ground motions because they rely on a limited number of representative parameters for description. On the other hand, detailed methods require high level of expertise and complex calculations. To address these challenges, this study examines three types of machine learning models that use observed waveforms as training data: (1) a model based on WaveGAN, (2) a model that incorporates the running spectra of earthquake ground motion into the cost function during the training of WaveGAN, and (3) a model based on DiffWave which applies a diffusion model. The seismic waveform characteristics of these models are investigated.