A Framework of Data-driven Seismic Performance Assessment of Structures using Earthquake Ground Motion Time Histories Generated by Machine Learning Techniques

Abstract

Observed ground motion time history records have been utilized in various ways to develop earthquake engineering. Especially in performance-based seismic design, dynamic response analysis is effective when strong motion time histories are used as input ground motions. Selection of appropriate ground motions has become a critical issue. Probabilistic methods such as PSHA (Probabilistic Seismic Hazard Analysis) are generally introduced to explicitly consider the uncertainties in seismic hazard. When evaluating the performance of structures, it is important to consider the key characteristics of seismic hazard, including appropriate treatment of uncertainties. However, conventional approach characterizes ground motion using a scalar intensity measure, such as spectral acceleration at fundamental period, which oversimplifies ground motion characteristics, including temporal characteristics in time history. Although dynamic response analysis is effective for evaluating the building performance considering such temporal characteristics, the probabilistic evaluation of uncertainties in ground motion time histories is still challenging. We have developed a novel method to predict ground motion time histories by utilizing GAN (Generative Adversarial Networks), a deep generative model proposed in the field of deep learning. GAN is a well-known model in image generation that can generate realistic data following the distribution of the learned data set. In this study, we present a novel framework for probabilistically evaluating the seismic performance of structures using time history data generated following the distribution of strong-motion observation records. We also discuss how earthquake ground motion time histories generated by data-driven technologies based on machine learning such as GAN can be utilized to develop a framework of performance-based seismic design.