Poster Presentation

Earthquake modeling captures the multiscale nature of fault processes, spanning spatial and temporal scales from slow aseismic slip to rapid dynamic rupture. Classical physics-based modeling, while accurate, is computationally expensive. To address this challenge, we present a computationally efficient and quantitatively accurate surrogate modeling approach. Specifically, we develop a Fourier Neural Operator–based framework to approximate the nonlinear equations governing dynamic rupture and earthquake cycle simulations. The surrogate model is trained on synthetic data generated from multiple physics-based simulations and is then applied to previously unseen scenarios. In dynamic rupture modeling, our approach preserves the accuracy of traditional multiscale methods while achieving a speedup of up to 400,000 compared to state-of-the-art conventional solvers. We demonstrate its generalization capability under unseen conditions. Additionally, we apply the FNO-based framework to model aseismic slip within earthquake cycles. Using autoregressive prediction techniques, we obtain a computational speedup of up to 30,000. This development advances the state of the art in computational earthquake dynamics, enabling large-scale statistical analyses, systematic parameter exploration, and probabilistic hazard assessments that were previously infeasible.