11:00 AM - 1:00 PM
[SCG44-P18] Investigating clustering and migration patterns of deep LFE activity in Nankai subduction zone using deep-learning and recurrent neural-networks.
Keywords:LFE clustering , migration pattern, DBSCAN algorithm
Deep-learning algorithms have the potential to model and predict complex systems which might be difficult to model using large temporal and spatial data. In Nankai subduction zone, episodic tremor and slip events, besides showing a simple parabolic along-strike migration (Ide, 2010, Ando et al., 2012), also include up-dip and down-dip very fast propagating clusters. RTR events with rapid propagation in the opposite direction of the migrating front have also been observed, but are less understood (Atienza et al., 2018). For a low-frequency (LFE) catalog with more than 500.000 events (Kato and Nakamura, 2020), we applied several unsupervised machine-learning algorithms which can identify clusters in a temporal dataset (OPTICS, HDBSCAN and DBSCAN). After evaluating each model’s performance in regards to identifying and classifing LFE clusters, we choose DBSCAN algorithm (Ester et al., 1998) which was able to identify clusters with variable size and duration.
We confirmed our clustering results (for very large clusters) are in agreement with reported short-term SSE (Hirose et al., 2020; Nishimura, 2020). To capture the short-range variation in clustering determined by RTR, we modeled two very large ETS events by using local polynomial interpolation maps, which can capture the short-range variation. The two consecutive ETS occurred in the same region in Shikoku (Figure a), but have contrasting migration patterns (Figure b-c), depending on the starting location. The first episode started in the central area and migrated along dip in the first stage, continuing with a bilateral along strike-migration, while the second ETS episodes was activated at the Bungo Channel and propagated unilateral along strike direction. We further explore the migration patterns of ETS clusters, as well as intermediate and smaller clusters which occur more frequantly, but are not associated with observed short-term SSE. Our end goal is to train a neural network which can identify and calculate physical parameters for each migration pattern, in order to evaluate physical parameters (diffusion coefficient, average velocity, reoccurrence patterns), which can be ultimately used for modelling all types of tremor migration across the Nankai subduction zone, offering the possibility to further explore the generating mechanism (stress diffusion and/or fluid diffusion) and identify “regions of interest” for real-time monitoring and forecasting.
We confirmed our clustering results (for very large clusters) are in agreement with reported short-term SSE (Hirose et al., 2020; Nishimura, 2020). To capture the short-range variation in clustering determined by RTR, we modeled two very large ETS events by using local polynomial interpolation maps, which can capture the short-range variation. The two consecutive ETS occurred in the same region in Shikoku (Figure a), but have contrasting migration patterns (Figure b-c), depending on the starting location. The first episode started in the central area and migrated along dip in the first stage, continuing with a bilateral along strike-migration, while the second ETS episodes was activated at the Bungo Channel and propagated unilateral along strike direction. We further explore the migration patterns of ETS clusters, as well as intermediate and smaller clusters which occur more frequantly, but are not associated with observed short-term SSE. Our end goal is to train a neural network which can identify and calculate physical parameters for each migration pattern, in order to evaluate physical parameters (diffusion coefficient, average velocity, reoccurrence patterns), which can be ultimately used for modelling all types of tremor migration across the Nankai subduction zone, offering the possibility to further explore the generating mechanism (stress diffusion and/or fluid diffusion) and identify “regions of interest” for real-time monitoring and forecasting.