Pseudo-Dynamic Source Characterization for Geometrically Rough Faults Using Machine Learning

ABSTRACT Accurately predicting intensity and variability of ground motions for future large earthquakes is essential for seismic hazard analysis. Dynamic rupture simulations provide insights into the physics of the time-dependent slip evolution on the fault plane, governed by stress and friction laws, but their computational demands limit their widespread application. In contrast, kinematic simulations offer computational efficiency using predefined spatiotemporal slip patterns that, however, lack physical consistency of dynamic models. Pseudo-dynamic (PD) approaches integrate the physical principles of dynamic rupture into a kinematic framework for efficient ground-motion computation. To capture the complex nonlinear relationships between earthquake source parameters, machine learning (ML) approaches offer a suitable alternative to first-order statistics. In this study, we develop an ML framework using Fourier neural operators to capture the intricate interdependencies between earthquake source parameters extracted from dynamic rupture scenarios. The rupture model generator begins with a stochastic slip distribution and a slip-constrained hypocenter location. Assuming a dynamically consistent Yoffe source time function (STF), the ML model then estimates the kinematic source parameters rupture velocity, peak slip velocity, and rise time. We adjust the fine-scale STF shape to account for small-scale rupture variations for improving high-frequency radiation. In addition, we introduce a method to model fault roughness correlated with stress drop for stochastic slip scenarios. Our PD rupture generator reproduces the mean and standard deviation of ground-motion models for different intensity measures for Mw 6.5 strike-slip scenarios, showing similar ground-motion variability as dynamic models. We prove the reliability of our approach by modeling strong-motion recordings of the 2000 Mw 6.6 Tottori earthquake. We conclude that our method is successful and computationally efficient to model complex interactions between earthquake source parameters for Mw 6.5–7.0 strike-slip events up to a frequency of 5.75 Hz. Our approach improves the accuracy of ground-motion simulations and thus may help advance seismic hazard assessment.