SCEC Award Number 25298 View PDF
Proposal Category Individual Research Project (Single Investigator / Institution)
Proposal Title Towards Machine Learning Accelerating Modeling of Sequences of Earthquakes and Aseismic Slp
Investigator(s)
Name Organization
Ahmed Elbanna University of Illinois at Urbana-Champaign
SCEC Milestones C1-2, C2-2, C3-1 SCEC Groups FARM, RC, ASI
Report Due Date 03/15/2026 Date Report Submitted 05/07/2026
Project Abstract
 Fault zones exhibit multiscale geometric and rheological variations that influence fault dynamics across both short (i.e. sub-seconds)- and long-term (i.e. centuries to millinia) timescales. To capture these processes, earthquake models must resolve fine-scale fault features, on the sub-meter scale, and simulate inertial dynamics of rupture propagation alongside long-term aseismic deformation, potentially on continuously evolving kilometer-scale fault surfaces. Addressing the multiscale nature of the problem, however, demands extensive computational resources and innovative simulation techniques, as traditional methods struggle to handle the intense data requirements and fine resolutions needed for full earthquake cycle simulations. More efficient approaches are needed to address this conundrum of scales.
Recent advancements in scientific machine learning (SciML) offer a promising approach to addressing these challenges. This proposal builds on these advances to (1) Develop an FNO framework to accelerate simulations of dynamic rupture propagation, incorporating heterogeneous shear stress and friction parameters. (2) Develop a hybrid FNO framework to efficiently simulate the long-term earthquake cycle, capturing both seismic and aseismic phases.
Intellectual Merit Earthquake cycle simulations must resolve processes spanning vastly different timescales—from decades-long aseismic creep to seconds-long dynamic rupture—making them prohibitively expensive for large-scale hazard studies. This work develops a multi-phase Fourier Neural Operator (FNO) framework with four specialized surrogates (aseismic, nucleation, seismic, postseismic) coupled sequentially through shared physical state variables. Phase-specific training strategies yield L2 errors below 5% and a ~12,000× speedup over a high-fidelity finite element–spectral boundary integral solver. The framework remains stable across multiple cycles and extends to nonperiodic events via hybrid physics–ML workflows, establishing principled methods for neural operators in multiscale geophysical systems.
Broader Impacts Reducing earthquake cycle runtimes from hours to seconds enables ensemble-scale hazard analysis, sensitivity studies, and scenario generation currently out of reach for physics-based solvers. The phase-decomposed neural operator methodology is transferable to other multiscale geophysical systems including subduction zones, volcanic unrest, and glacial dynamics. The hybrid physics–ML framework contributes broadly to scientific machine learning and provides a training ground for researchers at the intersection of computational geomechanics and deep learning—disciplines central to next-generation natural hazard assessment.
Project Participants USC, UIUC, Caltech
PI: Elbanna, Graduate Student: Napat Tainpakdipat (Defended in January 2026), Post doctoral researcher at Caltech (Dr. Mohamed Abdelmeguid, incoming assistant professor at the University of Houston.
Exemplary Figure Figure 1
Linked Publications

Add missing publication or edit citation shown. Enter the SCEC project ID to link publication.