Site-Specific Ground-Motion Waveform Generation Using a Conditional Generative Adversarial Network and Generalized Inversion Technique

Abstract

Accurate ground-motion simulations are essential for seismic hazard assessments and engineering practices. Herein, we propose a novel method combining conditional generative adversarial networks (cGANs) and the generalized inversion technique (GIT) to generate site-specific and variability-controlled strong-motion seismograms. The cGANs calculate synthetic seismogram without amplitude scales. The GIT is to separate the source, path, and site characteristics from the Fourier amplitude spectrum (FAS) of the observed seismograms. This method is applied to plate boundary earthquakes off the Pacific coast of Tohoku, Japan. It successfully generates a set of strong-motion seismograms at a given magnitude, distance, and observation station. The output waveforms reproduce the P and S waves as well as coda waves. We validate the method through a quantitative comparison with observed seismograms in terms of both time-domain duration and frequency-domain amplitude characteristics, using metrics of peak ground acceleration (PGA), peak ground velocity, FASs, response spectra, and waveform duration time. The validation results show that the variation in the PGA of the observed seismograms and the synthetic seismograms has a standard deviation of 0.643, and the duration of the seismograms has a standard deviation of 0.346, comparable to the standard deviations seen in the previous studies. Our approach offers high accuracy in stochastic finite-source modeling for a period of 1 s or shorter. The two features of the method, site-specificity and variability control, can contribute to further improvements in seismic hazard assessment by incorporating empirical information based on observed seismograms.