PhD Final Defense – Napat Tainpakdipat
- Event Type
- Seminar/Symposium
- Sponsor
- Department of Civil and Environmental Engineering
- Location
- Newmark 2218
- Date
- Jan 7, 2026 9:00 am
- Views
- 1
Scientific Machine Learning for Earthquake Modeling of Fault Dynamics and Earthquake Cycles
Advisor: Professor Ahmed Elbanna
Abstract
Earthquakes are one of the most severe natural hazards, yet limitations in available data prevent the use of purely data-driven approaches. Seismic observations are typically sparse, noisy, and limited in spatial coverage and duration. As a result, physics-based models have emerged to complement existing data and provide insights into earthquake source processes. However, these models face significant challenges, including limited ability to systematically integrate sparse and noisy observational data, poorly constrained fault properties, the inherent ill-posedness of inverse problems, and the multiscale nature of earthquake processes in space and time, which together lead to high computational costs.
Here, we develop a scientific machine learning framework to model earthquake processes by combining physical laws with data-driven learning. The framework addresses forward and inverse problems by leveraging physics-informed learning for data integration and inverse modeling, and operator-learning surrogates to enable computationally efficient multiscale simulations.
To study the interaction between data availability and physical constraints, we first investigate Physics-Informed Neural Networks (PINNs) using a spring–block model governed by a nonlinear rate-and-state friction law. We investigate the ability of PINNs to solve forward problems with and without observational data and show that conventional PINNs struggle in data-free settings due to numerical stiffness and optimization instability. To overcome this limitation, we introduce an adaptive time-stepping strategy that improves training stability and enables PINNs to recover the solutions even in the absence of measurement data. Later, we apply the PINNs framework to inverse modeling. We demonstrate that this approach can successfully infer intrinsic system properties, including material properties, time-dependent frictional evolution, and stability parameters that characterize fault behavior, despite sparse and noisy observations. These results highlight the potential of physics-informed learning to mitigate ill-posedness in inverse earthquake problems.
To address the computational demands of high-fidelity simulations in continuum models, we further develop surrogate models based on Fourier Neural Operators (FNOs). We first introduce an FNO-based framework for accelerating dynamic rupture simulations trained on datasets that include fractal initial shear stress fields, varying frictional parameters, initial slip rates, and diverse nucleation scenarios. This surrogate achieves speed-ups of up to five orders of magnitude compared to conventional numerical solvers and shows promising generalizability to previously unseen conditions. We then extend this operator-learning approach to earthquake cycles by incorporating adaptive time-stepping and recursive prediction to simulate long-term aseismic deformation in earthquake cycles. Finally, we introduce an end-to-end FNO framework that simulates aseismic, nucleation, seismic, and postseismic phases, achieving an overall speed-up of approximately four orders of magnitude and enabling scalable simulations of long-term fault dynamics.