Fault dynamics parameter identification using Physics-Informed Neural Networks

Napat Tainpakdipat, & Ahmed E. Elbanna

Published September 8, 2024, SCEC Contribution #13884, 2024 SCEC Annual Meeting Poster #209

Characterization of seismic events requires fusion of data and models related to the earthquake fault dynamics and seismic wave propagation. While this task is crucial for risk assessment and forecasting, it remains challenging due to the complex nature of fault motion and the difficulty in directly measuring several critical parameters, such as fault friction and stress. Our knowledge about the subsurface and the underlying source dynamics relies on information carried by the seismic waves and collected at, typically, sparse locations on the Earth’s surface. The resulting inverse problem is ill-posed and under-constrained and inversion for fault stress and friction has been largely neglected except in a few dynamic inversion studies due to the significant computational cost. A paradigm shift is needed to overcome the computational challenge and enable routine inversion for fault parameters that are critical for source physics.
Here we improve parameter identification in fault dynamics by employing Physics-informed Neural Networks (PINNs), which integrate physical constraints with neural networks. As a proof-of-concept, we apply PINNs to the fully dynamic Burridge-Knopoff spring-block model with nonlinear rate-and-state friction law. We formulate two types of inverse problems. In the first problem, we estimate the constant stiffness term, akin the elastic properties, through a time-independent inverse PINN and track the evolution of friction over time using a time-dependent inverse PINN. In the second problem, we introduce a mixed scheme of time-dependent and time-independent inverse PINNs to solve for the frictional stability parameter, a-b, that determines the fault's velocity-weakening and velocity-strengthening regimes. This parameter is crucial for assessing fault stability and cannot be directly measured in the field. Our results indicate that PINNs effectively identify fault parameters based on accessible real-world data, such as particle velocities. We also demonstrate the robustness of PINN in the presence of noisy data. These findings suggest that PINNs offer a promising and innovative approach for parametric identification in geophysical simulations. The codes used in this study are available on Quakeworx, a new NSF-funded science gateway for earthquake simulations and data.

Key Words
Fault dynamics, Rate-and-state friction, Physics-informed neural networks (PINNs), Inverse problem

Citation
Tainpakdipat, N., & Elbanna, A. E. (2024, 09). Fault dynamics parameter identification using Physics-Informed Neural Networks . Poster Presentation at 2024 SCEC Annual Meeting.


Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)