14:45 〜 15:00
[SCG45-29] 物理深層学習を用いた2010年豊後水道スロースリップイベントの時間発展モデル構築：GEONET実データ解析
キーワード：物理深層学習、スロースリップ、地震サイクルシミュレーション、速度状態依存摩擦則、GNSS
Slow slip events (SSEs) have been observed in many subduction zones and are understood to result from frictional unstable slip on the plate interface. Previous observations show that aseismic and seismic slips occur at different regions, suggesting that frictional properties are heterogeneous. We are however lacking methods to constrain spatial distribution of frictional properties.
One of the possible approaches to address this problem is data assimilation, a method that incorporates the observational data into a physics-based model. Kano et al. (2024) applied the data assimilation method to the 2010 slow slip event in the Bungo region, southwest Japan, and estimated the frictional parameters of a rate and state friction (Dieterich, 1979) from GNSS observation. However, due to the difficulty of optimizing a large number of unknown parameters, their model imposed a strong assumption that frictional properties are uniform on a prescribed fault patch. As a result, the simplified model exhibits significant residuals from GNSS observation in some stations, failing to capture some important characteristics of the slip patterns.
With the recent development of machine learning, a new method known as Physics-Informed Neural Networks (PINNs) has been proposed to estimate the model parameter from observation (Raissi et al., 2019). PINN is a deep learning-based method to solve the PDEs governing the physics-based model while simultaneously inferring the controlling parameters from the data). Fukushima et al. (2024, ESS Open Archive) developed a PINN-based method to estimate fault slip and frictional parameters from geodetic observations. They conducted numerical experiments using the synthetic slow slip event data and successfully recovered the fault slip and the spatial distribution of frictional parameters.
In this study, we applied the PINN-based method to the 2010 slow slip event in the Bungo region. Using the same actual GNSS data as Kano et al., (2024), we estimate the fault slip and the spatial distribution of frictional parameters with rate and state friction. The estimated fault slip evolution shows the localized nucleation of SSE that was not captured in the previous study. Our obtained model, which includes the spatial distribution of frictional parameters, shows better consistency with the geodetic observation data.
One of the possible approaches to address this problem is data assimilation, a method that incorporates the observational data into a physics-based model. Kano et al. (2024) applied the data assimilation method to the 2010 slow slip event in the Bungo region, southwest Japan, and estimated the frictional parameters of a rate and state friction (Dieterich, 1979) from GNSS observation. However, due to the difficulty of optimizing a large number of unknown parameters, their model imposed a strong assumption that frictional properties are uniform on a prescribed fault patch. As a result, the simplified model exhibits significant residuals from GNSS observation in some stations, failing to capture some important characteristics of the slip patterns.
With the recent development of machine learning, a new method known as Physics-Informed Neural Networks (PINNs) has been proposed to estimate the model parameter from observation (Raissi et al., 2019). PINN is a deep learning-based method to solve the PDEs governing the physics-based model while simultaneously inferring the controlling parameters from the data). Fukushima et al. (2024, ESS Open Archive) developed a PINN-based method to estimate fault slip and frictional parameters from geodetic observations. They conducted numerical experiments using the synthetic slow slip event data and successfully recovered the fault slip and the spatial distribution of frictional parameters.
In this study, we applied the PINN-based method to the 2010 slow slip event in the Bungo region. Using the same actual GNSS data as Kano et al., (2024), we estimate the fault slip and the spatial distribution of frictional parameters with rate and state friction. The estimated fault slip evolution shows the localized nucleation of SSE that was not captured in the previous study. Our obtained model, which includes the spatial distribution of frictional parameters, shows better consistency with the geodetic observation data.