Broadband Ground-Motion Simulations with Machine-Learning-Based High-Frequency Waves from Fourier Neural Operators

ABSTRACT Seismic hazards analysis relies on accurate estimation of expected ground motions for potential future earthquakes. However, obtaining realistic and robust ground-motion estimates for specific combinations of earthquake magnitudes, source-to-site distances, and site conditions is still challenging due to the limited empirical data. Seismic hazard analysis also benefits from the simulation of ground-motion time histories, whereby physics-based simulations provide reliable time histories but are restricted to a lower frequency for computational reasons and missing information on small-scale earthquake-source and Earth-structure properties that govern high-frequency (HF) seismic waves. In this study, we use densely recorded acceleration broadband (BB) waveforms to develop a machine-learning (ML) model for estimating HF ground-motion time histories from their low-frequency (LF) counterparts based on Fourier Neural Operators (FNOs) and Generative Adversarial Networks (GANs). Our approach involves two separate FNO models to estimate the time and frequency properties of ground motions. In the time domain, we establish a relationship between normalized low-pass filtered and BB waveforms, whereas in the frequency domain, the HF spectrum is trained based on the LF spectrum. These are then combined to generate BB ground motions. We also consider seismological and site-specific factors during the training process to enhance the accuracy of the predictions. We train and validate our models using ground-motion data recorded over a 20 yr period at 18 stations in the Ibaraki province, Japan, considering earthquakes in the magnitude range M 4–7. Based on goodness-of-fit measures, we demonstrate that our simulated time series closely matches recorded observations. To address the ground-motion variability, we employ a conditioned GAN approach. Finally, we compare our results with several alternative approaches for ground-motion simulation (stochastic, hybrid, and ML-based) to highlight the advantages and improvements of our method.