5:15 PM - 6:45 PM
[STT38-P07] Acquisition of stochastic differential equation representations to characterize low-frequency tremors in seismic waveform data using deep learning
Keywords:low frequency tremor, slow earthquake, stochastic differential equation, SDE, deep learning, signature kernel
Slow earthquakes are slip phenomena occurring at significantly slower velocities compared to conventional earthquakes, potentially linked to plate boundary seismic activities, and have been studied from various perspectives.
Slow earthquakes can be categorized into short-duration low frequency earthquakes (LFEs) ranging from 2-8Hz, very low frequency earthquakes (VLFs) lasting about 20 seconds, and slow slip events (SSEs) extending from several days to months. This study focuses on VLFs.
Slow earthquakes, when observed over the long term, can be considered to be modeled through stochastic differential equations (SDEs) due to their behavior resembling Brownian motion, thereby capturing their unique structural characteristics. Pioneering work by Ide[1] elucidated that the distinct properties of slow earthquakes can be extensively explained by a model governed by SDEs, where changes in the size of the circular fault generate seismic waves.
Moreover, recent advancements have explored acquiring representations of SDEs from time-series observation data using deep learning. Issa et al. [2] proposed a method to learn SDEs formulated as neural networks from time-series data, measuring similarities in the data with the Signature Kernel metric.
This research aims to learn the form of SDEs from actual slow earthquake data at multiple observation points using the method by Issa et al., proposing and discussing a method for predicting seismic wave velocities at other locations when seismic wave velocities related to slow earthquakes are observed at a certain location.
[1]Ide, S. (2008), A Brownian walk model for slow earthquakes, Geophys. Res. Lett., 35, L17301, doi:10.1029/2008GL034821.
[2]Issa, Z., Horvath, B., Lemercier, M., & Salvi, C. (2023). Non-adversarial training of Neural SDEs with signature kernel scores. arXiv preprint arXiv:2305.16274.
Slow earthquakes can be categorized into short-duration low frequency earthquakes (LFEs) ranging from 2-8Hz, very low frequency earthquakes (VLFs) lasting about 20 seconds, and slow slip events (SSEs) extending from several days to months. This study focuses on VLFs.
Slow earthquakes, when observed over the long term, can be considered to be modeled through stochastic differential equations (SDEs) due to their behavior resembling Brownian motion, thereby capturing their unique structural characteristics. Pioneering work by Ide[1] elucidated that the distinct properties of slow earthquakes can be extensively explained by a model governed by SDEs, where changes in the size of the circular fault generate seismic waves.
Moreover, recent advancements have explored acquiring representations of SDEs from time-series observation data using deep learning. Issa et al. [2] proposed a method to learn SDEs formulated as neural networks from time-series data, measuring similarities in the data with the Signature Kernel metric.
This research aims to learn the form of SDEs from actual slow earthquake data at multiple observation points using the method by Issa et al., proposing and discussing a method for predicting seismic wave velocities at other locations when seismic wave velocities related to slow earthquakes are observed at a certain location.
[1]Ide, S. (2008), A Brownian walk model for slow earthquakes, Geophys. Res. Lett., 35, L17301, doi:10.1029/2008GL034821.
[2]Issa, Z., Horvath, B., Lemercier, M., & Salvi, C. (2023). Non-adversarial training of Neural SDEs with signature kernel scores. arXiv preprint arXiv:2305.16274.