Inversion for the spatial distribution of frictional parameters on the long-term SSE fault with Physics-Informed Neural Networks

Rikuto Fukushima, Masayuki Kano, Kazuro Hirahara, Makiko Ohtani, Kyungjae Im, & Jean-Philippe Avouac

Published September 8, 2024, SCEC Contribution #13859, 2024 SCEC Annual Meeting Poster #139 (PDF)

Various types of fault slip, including earthquakes and slow slip events (SSEs), have been observed in many subduction zones. The diversity of these slip behaviors on the plate interface suggests spatially heterogeneous frictional properties The frictional properties on a long-term SSE fault were estimated using data assimilation, a technique to integrate a physics-based model and observation, from GNSS observations assuming a velocity weakening patch with a prescribed shape (e.g. Hirahara and Nishikiori, 2019; Kano et al., 2024). However, no previous study assessed the spatial variations of frictional properties because of the difficulty of conventional data assimilation to optimize the model parameters in the large dimension model. Instead of the conventional data assimilation method, Fukushima et al. (2023) proposed a new approach using Physics-Informed Neural Networks (PINNs) to constrain the frictional parameters with a single-degree-of-freedom spring slider model. PINN is a deep learning technique that can be used to solve physics-based differential equations and to determine the model parameters from observations. In this study, we extend this method to 3D and estimate the spatial distribution of frictional parameters.

We execute a series of numerical twin experiments assuming the long-term SSEs observed in the Bungo channel, southwest Japan, and demonstrate the performance of the PINN-based approach. For the true SSE model, we set the SSE region as a uniform circular velocity-weakening patch embedded in the velocity-strengthening region. We start with an idealized case where it is assumed that fault slip can be directly observed. In this case, the spatial distribution of frictional parameters is estimated well. We next move to a more realistic case where the synthetic surface displacement velocity data are observed by ideally distributed GNSS stations. The geometry of the velocity weakening region, where the slip instability develops, is well estimated. We find that observation directly above the SSE region is crucial to constrain the frictional parameters. These results suggest that the PINN-based method is a promising approach for estimating the spatial distribution of friction parameters and assimilating GNSS observations in dynamic models of SSE.

Citation
Fukushima, R., Kano, M., Hirahara, K., Ohtani, M., Im, K., & Avouac, J. (2024, 09). Inversion for the spatial distribution of frictional parameters on the long-term SSE fault with Physics-Informed Neural Networks. Poster Presentation at 2024 SCEC Annual Meeting.

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