日本地球惑星科学連合2022年大会

講演情報

[J] 口頭発表

セッション記号 S (固体地球科学) » S-CG 固体地球科学複合領域・一般

[S-CG51] 機械学習による固体地球科学の牽引

2022年5月22日(日) 10:45 〜 12:15 102 (幕張メッセ国際会議場)

コンビーナ:久保 久彦(国立研究開発法人防災科学技術研究所)、コンビーナ:小寺 祐貴(気象庁気象研究所)、直井 誠(京都大学)、コンビーナ:矢野 恵佑(統計数理研究所)、座長:久保 久彦(国立研究開発法人防災科学技術研究所)、岡崎 智久(理化学研究所革新知能統合研究センター)、直井 誠(京都大学)

11:15 〜 11:30

[SCG51-08] Physics-Informed Neural NetworkによるDislocation Modelの解法

*岡崎 智久1平原 和朗1,2、上田 修功1 (1.理化学研究所革新知能統合研究センター、2.香川大学)

Recently, Raissi et al. (2019) introduced a physics-informed neural network (PINN) to solve partial differential equations (PDEs). By incorporating a target PDE and its boundary and initial conditions into a loss function using automatic differentiation, PINNs search for a latent solution without training data. Moreover, PINNs can be applied to both forward and inverse problems with almost identical network architectures, which is appealing to geophysical applications.

This study applies PINNs to dislocation models to obtain static crustal deformation caused by fault ruptures. A characteristic of dislocation models is that a displacement field is discontinuous across a dislocation surface and cannot be directly modeled by neural networks. We therefore set an appropriate coordinate system to separate coordinate values of the two sides of a displacement discontinuity.

In experiments, PINNs are applied to anti-plane problems (i.e. infinitely long strike-slip faults). We first compare PINN's solutions with analytical solutions (Segall, 2010) for simple problems on a vertical fault in the uniform half-space. We then solve complex problems such as curved faults, topography, and heterogeneous media for which analytical approaches are difficult. PINNs have an advantage that continuous shapes and variations can be modeled without discretization, in contrast to traditional methods such as the finite difference method and the finite element method. PINNs have a potential to solve a wide variety of modeling applications in crustal deformation.