Presentation Information
[P22-11]Mapping deep fault geometries by clustering of deep learning enhanced hypocenter distributions and their normal vectors in Hokkaido
*Admore MPUANG1, Takahiro SHIINA1, Yasunori SAWAKI1, Ayaka TAGAMI1, Takahiko UCHIDE1 (1. Geological Survey of Japan, AIST)
The oblique subduction of the Pacific plate beneath Hokkaido drives the collision between the SW-bound Kuril forearc sliver and the NE Japan arc along central Hokkaido (Kimura., 1994, 1996). This collision results in a complex thickened crust (e.g., Gou et al., 2019; Iwasaki et al., 2004) and several deep crustal earthquakes. Most of the region’s seismicity occurs at blind faults of unknown geometries. Moreover, inferred geometries of some of the deep seismic activity, such as the causative fault of the 2017 Mw 6.6 event estimated by Kobayashi et al., 2019, show no relation to known surface traces. Knowledge of fault geometries is important to understand rupture processes and regional tectonics. In this study, we determine deep fault geometries in the Hokkaido region, based on relocated hypocenter distributions and their normal vectors. Sawaki et al. (2025) developed a new method to reconstruct intricate fault models based on hierarchical clustering of hypocenter distributions and normal vectors. We first relocated hypocenters from the JMA catalog reported since 2000, with a minimum magnitude of 0, using tomoDD (Zhang & Thurber, 2003, 2006) and a three-dimensional velocity model (Nishida et al., 2008). We use only inland events occurring within the crust. Our results in the whole Hokkaido region show intricate fault geometries that do not necessarily reflect known surface traces or surface geology. In the Eastern Iburi region, where a Mw 6.6 earthquake occurred in 2018, we find steeply dipping multiple fault segments related to the Mw 6.6 event as well as prior events. The fault segments show a general N-S trend similar to the regional gravity anomaly, highlighting the relation to deep regional structures and tectonics. To improve resolution of fault geometries, we apply a retrained PhaseNet model (Zhu and Beroza, 2019; Naoi et al, 2024) to pick seismic P and S phases from continuous seismic waveforms recorded since 2004 by inland seismic stations. To create our deep learning enhanced seismic catalog, we used the GaMMA phase associator (Zhu et al., 2022). The preliminary catalog suggests previously hidden seismic activity occurring mostly in central Hokkaido, highlighting hidden faults. The detection of additional small events by deep-learning is expected to improve fault plane determination accuracy, and possibly delineate the shallow extensions of deep faults, if present. From our results, we will provide a region-wide fault geometry map with important insights for seismic hazard assessments in the Hokkaido region.
This study is supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217. We used data from the JMA Unified Earthquake Catalog, and waveform data from Hi-net, JMA and Hokkaido University seismic networks.
This study is supported by MEXT Project for Seismology toward Research Innovation with Data of Earthquake (STAR-E) Grant Number JPJ010217. We used data from the JMA Unified Earthquake Catalog, and waveform data from Hi-net, JMA and Hokkaido University seismic networks.