Physics-Based Earthquake Forecasting with Machine-Learned Reduced-Order Models and Ensemble Kalman Filtering

Hojjat Kaveh, Jean-Philippe Avouac, & Andrew Stuart

Submitted September 7, 2025, SCEC Contribution #14632, 2025 SCEC Annual Meeting Poster #TBD

Accurate forecasting of earthquakes in fault systems remains challenging due to the high dimensionality and strong nonlinearity of governing equations. We present a physics-based data assimilation framework that integrates the Ensemble Kalman Filter (EnKF) with a machine-learned reduced-order model (ROM) of the earthquake cycle. The ROM is constructed by training on high-fidelity numerical simulations of elastodynamic rupture governed by rate-and-state friction, using machine learning to identify the dominant spatiotemporal modes and their nonlinear evolution. The EnKF operates on the reduced state space, assimilating noisy and sparse surface displacement data to update the fault state variables in real time. We evaluate the method using synthetic earthquake sequences that exhibit a broad range of recurrence intervals and rupture sizes, including extreme events. Results show that the machine-learned ROM–EnKF system can reconstruct unobserved fault variables and improve forecasts of upcoming events, with skill depending on observation density, noise level, and assimilation frequency. This approach significantly reduces computational cost while preserving the essential dynamics of the full model, offering a promising avenue for time-dependent earthquake hazard assessment in complex fault systems.

Citation
Kaveh, H., Avouac, J., & Stuart, A. (2025, 09). Physics-Based Earthquake Forecasting with Machine-Learned Reduced-Order Models and Ensemble Kalman Filtering. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Fault and Rupture Mechanics (FARM)