日本地球惑星科学連合2021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-SS 地震学

[S-SS05] 地殻変動

2021年6月3日(木) 13:45 〜 15:15 Ch.22 (Zoom会場22)

コンビーナ:加納 将行(東北大学理学研究科)、落 唯史(国立研究開発法人産業技術総合研究所 地質調査総合センター 活断層・火山研究部門)、富田 史章(国立研究開発法人 海洋研究開発機構)、座長:田中 愛幸(東京大学理学系研究科)、矢野 恵佑(統計数理研究所)

14:45 〜 15:00

[SSS05-05] 単一矩形断層推定に関するハミルトニアンモンテカルロ法の適用可能性

*山田 太介1、太田 雄策1 (1.東北大学大学院理学研究科附属 地震・噴火予知研究観測センター)

The rapid estimation of the coseismic fault model and its uncertainty are extremely important to predict the possible hazard such as tsunami. Based on such motivation, Ohno et al. (in revision) developed the estimation method of single rectangular fault model deduced from Markov Chain Monte Carlo (MCMC) method in real-time. They adopted conventional Metropolis-Hasting (M-H) method to sample the probability density function. The M-H method, however, require long Markov chain because of the ideal acceptance ration is around 20-30%. Therefore, the large dimensional problem requires the very long mixing time.

To overcome such problem, we investigate the new estimation approach for single rectangular fault model based on the Hamiltonian Monte Carlo (HMC) method. HMC is a MCMC method that uses the derivatives of the density function being sampled to generate efficient transitions spanning the posterior. It uses an approximate Hamiltonian dynamics simulation based on numerical integration which is then corrected by performing a Metropolis acceptance step. To calculate the numerical integration, we adopted the leapfrog integrator, which is a numerical integration algorithm. During the leapfrog integrator, we need to assume the discretization time (e) and number of steps (L). To estimate the appropriate number of leapfrog (L), we adopted the no-U-turn sampler (NUTS). We also assumed the discretization time (e) as 10-4 which was determined by try-and-error. The model parameter vector that contains fault parameters of a single rectangular fault model (Okada 1992). To optimize amount of each step size for unknown parameters, we installed logit transformation for all unknown parameters except for latitude and longitude.

We applied the developed approach to numerical experience which was under the optimum site coverage. Obtained results clearly shows the rapid convergence (~1000 steps) which clearly shorter step number compared with M-H method. We will show the more detail of our developed method and results for more realistic site coverage.