14:30 〜 14:45
[SCG55-09] Physics Informed Deep Learning for Inverse Modeling of Coseismic Crustal Deformation
Physics-informed neural networks (PINNs) have been proposed to solve both the forward and inverse problems by incorporating partial differential equations (PDEs) into the loss function of neural networks (NNs) (Raissi et al. 2019). Okazaki et al. (2022) applied PINNs to forward modeling of coseismic crustal deformation in anti-plane (i.e. strike-slip faults) dislocations. In the forward modeling, the loss function consists of the governing PDE, the displacement discontinuity on the fault, the traction continuity on the faults, and the traction-free condition on the ground.
In this study, we apply PINNs to the inverse modeling of slip distributions from coseismic crustal deformation. In typical inverse problems (e.g. seismic tomography), observational data (e.g. travel time) and the estimated model (e.g. velocity structure) represent different physical variables; therefore, two NNs are constructed to represent the data and model variables. In contrast, in the slip inversion, data (surface displacement) and the model (displacement discontinuity on the fault) represent the same physical variable (i.e. displacement); therefore, only one NN is required. The inverse modeling can be simply formulated by removing the displacement discontinuity on the fault and adding the data residual on the ground. The slip distribution can be estimated as the difference in displacement between the two sides of the fault.
To investigate the characteristics of the PINN-based slip inversion, we present a preliminary study on a deterministic (i.e. non-Bayesian) inversion analysis in anti-plane and in-plane (i.e. normal/reverse faults) dislocations. We carry out synthetic tests to systematically evaluate the performance on different problem settings. In particular, the weight of the data residual term in the loss function is essential for the stable estimation and should be determined in trial and error or adaptively based on the magnitude of observational noises.
In this study, we apply PINNs to the inverse modeling of slip distributions from coseismic crustal deformation. In typical inverse problems (e.g. seismic tomography), observational data (e.g. travel time) and the estimated model (e.g. velocity structure) represent different physical variables; therefore, two NNs are constructed to represent the data and model variables. In contrast, in the slip inversion, data (surface displacement) and the model (displacement discontinuity on the fault) represent the same physical variable (i.e. displacement); therefore, only one NN is required. The inverse modeling can be simply formulated by removing the displacement discontinuity on the fault and adding the data residual on the ground. The slip distribution can be estimated as the difference in displacement between the two sides of the fault.
To investigate the characteristics of the PINN-based slip inversion, we present a preliminary study on a deterministic (i.e. non-Bayesian) inversion analysis in anti-plane and in-plane (i.e. normal/reverse faults) dislocations. We carry out synthetic tests to systematically evaluate the performance on different problem settings. In particular, the weight of the data residual term in the loss function is essential for the stable estimation and should be determined in trial and error or adaptively based on the magnitude of observational noises.