[S-CG40] Science of slow-to-fast earthquakes

[SCG40-02] Unsupervised Event Characterization Reveals Complexity in Earthquake Spectral Patterns and Driving Processes

★Invited Papers

*Kaiwen Wang¹, Felix Waldhauser¹, Maya Tolstoy², William Wilcock², Theresa Sawi¹, David Schaff¹, Yen Joe Tan³ (1.Lamont-Doherty Earth Observatory, Columbia University, 2.University of Washington, 3.The Chinese University of Hong Kong)

Keywords:Earthquakes, Unsupervised Machine Learning

Earthquake signals have been observed at a wide range of frequency bands. They span over a rather continuous spectrum from fast to slow earthquakes, including tectonic/volcanic earthquakes, low-frequency/long-period earthquakes (LFEs/LPs), very low-frequency earthquakes (VLFEs), tremors, slow slip events (SSEs), and hybrid frequency earthquakes. Especially in volcanic environments, various seismic source processes generate signals with a broad range of spectral characteristics. Understanding their behaviors and driving mechanisms may help improve our knowledge of complex tectonic and volcanic processes. In this study, we employ data-driven approaches to mine the seismic data at Axial Seamount, an active submarine volcano on the Juan de Fuca Ridge. We use continuous seismic recordings of the Ocean Observatories Initiative (OOI) cabled Ocean Bottom Seismometers (OBS) array to uncover hidden patterns that are indicative of different source processes. The OBS array started operation ~4 months before the last eruption in April 2015 and captured the intense seismic activity leading up to and during the eruption. We built a comprehensive earthquake catalog using supervised machine learning (ML) and double-difference (DD) methods. We then applied SpecUFEx, an unsupervised machine learning method, to extract time-variant spectral features and search for hidden patterns in the waveforms. We find two distinct earthquake clusters in the spectral feature space that we associate with different source characteristics. Based on the spatiotemporal distribution of the events, their correlation with tidal forces, volcanic activities, and deformation observations, we infer that the events in the first cluster are regular (shear failure) earthquakes that occur on the caldera ring faults and are triggered by stress changes imposed by tidal forces. We call the events in the second cluster mixed-frequency earthquakes (MFEs), and we associate them with brittle crack opening followed by influx of magma or volatiles. Because they appear about 15 hours before the eruption, we believe they represent the beginning of the eruption preparation process. These results offer new insights into the dynamics of Axial’s magma system and its complex faults system. We integrated this semi-supervised event classification module into real-time ML-based catalog production. This enhances the routine seismic monitoring system by facilitating the discrimination and tracking of seismic events across different frequency bands as they occur. It also showcases the capability of data-driven methods in the discovery of signals with unknown signatures. This approach could be potentially generalized to other active volcanic or tectonic regions to advance our understanding of slow to fast earthquakes, especially as the size of seismic datasets continues to grow.