4:15 PM - 4:30 PM
[SCG49-10] Time History-Based Probabilistic Seismic Hazard Analysis based on a Ground Motion Generative Model
Keywords:Probabilistic seismic hazard analysis, Ground motion model, Deep generative model
This study proposes a novel framework for probabilistic seismic hazard analysis (PSHA) based on the probability distribution of ground motion time-history data generated by a ground motion generative model.
In general, the exceedance probability of a ground motion intensity measure is evaluated in PSHA using predictions from ground motion models (GMMs). Such hazard analysis results are widely utilized for various purposes. However, in performance-based earthquake engineering (PBEE), exemplified by the PEER-PBEE framework, seismic performance evaluations based on dynamic response analyses of buildings are increasingly common, often requiring ground motion time-history data as the input seismic hazard.
Against this background, we have developed GMMs capable of directly modeling the probability distribution of ground motion time-history data using generative adversarial networks (GANs), a type of deep generative model. We refer to such a GMM as a ground motion generative model (GMGM). Our proposed GMGM evaluates the distribution of ground motion time-history data by incorporating its source, propagation path, and site characteristics, making it promising for PSHA applications.
We construct a GMGM using a method called StyleGAN2 and a strong-motion observed record database of crustal earthquakes in Japan. The constructed GMGM considers moment magnitude as the source characteristic, rupture distance as the propagation path effect, and the average shear-wave velocity in the top 30 m (Vs30) as the site characteristic. Subsequently, we propose a novel formulation of PSHA, where the PSHA integral is computed through Monte Carlo sampling using ground motion time-history data generated by the GMGM. Finally, a numerical experiment is conducted for a specific site and its surrounding active faults to demonstrate the proposed method. The PSHA results are presented and compared with those obtained using existing empirical GMMs, focusing on peak ground velocity, to discuss the validity of the proposed method.
In general, the exceedance probability of a ground motion intensity measure is evaluated in PSHA using predictions from ground motion models (GMMs). Such hazard analysis results are widely utilized for various purposes. However, in performance-based earthquake engineering (PBEE), exemplified by the PEER-PBEE framework, seismic performance evaluations based on dynamic response analyses of buildings are increasingly common, often requiring ground motion time-history data as the input seismic hazard.
Against this background, we have developed GMMs capable of directly modeling the probability distribution of ground motion time-history data using generative adversarial networks (GANs), a type of deep generative model. We refer to such a GMM as a ground motion generative model (GMGM). Our proposed GMGM evaluates the distribution of ground motion time-history data by incorporating its source, propagation path, and site characteristics, making it promising for PSHA applications.
We construct a GMGM using a method called StyleGAN2 and a strong-motion observed record database of crustal earthquakes in Japan. The constructed GMGM considers moment magnitude as the source characteristic, rupture distance as the propagation path effect, and the average shear-wave velocity in the top 30 m (Vs30) as the site characteristic. Subsequently, we propose a novel formulation of PSHA, where the PSHA integral is computed through Monte Carlo sampling using ground motion time-history data generated by the GMGM. Finally, a numerical experiment is conducted for a specific site and its surrounding active faults to demonstrate the proposed method. The PSHA results are presented and compared with those obtained using existing empirical GMMs, focusing on peak ground velocity, to discuss the validity of the proposed method.