以下の特別セミナーを開催いたします。皆様是非ご参加ください。

日時：2026年6月15日（月）午後3時から午後4時半まで
場所：京都大学宇治キャンパス　地震災害研究センター棟200号室（セミナー室）

講演者: Dr. Nicolas CHARTIER (理化学研究所　革新知能統合研究センター　防災科学チーム　基礎科学特別研究員）

タイトル：Towards accelerated forward modelling for Bayesian inference of fault friction and seismicity

Abstract:
Geodetic and seismic observations of subduction zones, such as displacements from GNSS, carry information about frictional parameters and seismicity that current inversion frameworks cannot fully exploit. Quasi-dynamic Sequences of Earthquakes and Aseismic Slip (SEAS) solvers and Epidemic-Type Aftershock Sequence (ETAS) phenomenological models are essential tools for earthquake and seismic hazard forecasting. Geophysicists have dedicated significant effort to go beyong assumptions such as that of a linear, homogeneous elastic half-space, or that of a planar fault geometry. Bayesian inference of physical parameters remains limited by the computational cost of forward models, especially for SEAS if frictional parameters are spatially variable.

I joined the Disaster Resilience Science Team at RIKEN in May 2026 as a special postdoctoral researcher. This talk outlines my background and introduces my research proposal, with the ambition of fostering discussion and collaborative projects.

In the first part, I will give an overview of my past works on simulations and statistical inference in Cosmology. These works include the “CARPool” method that combines simulations with cheap surrogates to reduce variance on statistics of observables, and the constraining of cosmological parameters from galaxy surveys using Simulation-Based Inference (SBI). In addition, I will introduce the public ltu-ili inference code, still being improved in the Learning the Universe Collaboration, as well as an example of ML surrogate for a bottleneck of cosmological simulations.Then, I will explain SBI in more details, a framework that trains Neural Density Estimators on simulation suites to approximate posterior distributions of parameters, even when the likelihood is intractable. In addition, I will emphasize the need for posterior validation metrics, which allow to assess fundamental issues such as the effects of model misspecification on real-world applications.

In the second part, I introduce the geophysics context. I will summarize recent SEAS and crustal deformation results from DRST members and collaborators using Physics-Informed Neural Networks (PINNs), before presenting my first geophysics project: constraining the distribution of frictional parameters on the Bungo Channel non-planar fault geometry, using the BEM solver PyQuake3D. Finally, we will review the state of the art in SEAS and ETAS inversion, including recent contributions from DPRI members, and identify problems that motivate investigating SBI and neural surrogate forward models. Dealing with spatial distributions of frictional parameters in a Bayesian sense will be openly discussed. Another particular challenge in SEAS is that fault geometry is prescribed rather than inferred. Whether learned surrogates for Green's functions conditioned on geometry could constitute a useful contribution, although beyond the scope of the current proposal, is an open question.