Efficient Fault Rupture Simulation with a Dual-Stage Fourier Neural Operator and Physics-Based Sampling

Accurately simulating fault rupture dynamics is critical for aftershock prediction but remains computationally prohibitive due to the multiscale nature of earthquake processes. While Fourier Neural Operators (FNOs) offer a promising framework for seismic simulation, their direct application to rupture dynamics is hindered by spectral bias from global processing and resolution loss from uniform downsampling. To overcome these limitations, this paper introduces a novel dual-stage FNO architecture explicitly designed for multiscale rupture simulation. The architecture decouples the problem into a first stage for efficient low-frequency wave propagation in the non-fault zone and a second stage for resolving meter-scale nonlinear rupture dynamics within the fault zone. Then, we propose a physics-based sampling strategy that maintains high resolution in the critical fault zone while coarsening the non-fault zone based on wave-propagation criteria, coupled with an interpolation scheme that enforces conservation of mass, momentum, and energy. Evaluated on the SCEC TPV101 benchmark, our integrated framework achieves a 92.4% reduction in model parameters and a 2.34× speedup in training time compared to a baseline FNO approach, while also reducing the NRMSE in fault zones by 80.1%. Furthermore, the model demonstrates robust generalization to unseen geological parameters.