← Back to Sessions

Magma Movement Revealed by Unsupervised Spectral Feature Characterization of Seismicity at Axial Seamount

Description: 

Axial Seamount is an active submarine volcano on the Juan de Fuca Ridge which last erupted in April 2015. The eruption was captured by the OOI cabled, 7-station OBS array that started operation 5 months before the eruption. The array records signals from a variety of sources, providing crucial constraints on the dynamics of Axial’s magma system and its complex ring faults. We used supervised machine learning and double-differences to compute a 7-year long, high-precision earthquake catalog of over 160,000 events. Here, we explore this catalog with an unsupervised spectral feature extraction method (SpecUFEx) to characterize time-dependent variations in the event spectrograms. We condensed the spectrograms with nonnegative matrix factorization and hidden Markov model to reduce dimensionality and built fingerprints that best represent the events’ time-variant spectral features. These fingerprints were then clustered by a k-means algorithm to form groupings of events based on their spectral characteristics. Among the groupings we find clusters corresponding to signals of earthquakes, impulsive events generated by hot lava reaching the sea floor, and fin whale calls. Another cluster includes characteristic waveforms with a rich low frequency train following the first arrival. These low-frequency events (LFEs) first appeared in the northern part of the caldera a few hours before the 2015 eruption at depths near the top of the magma chamber (1.5-2 km), in an area of large seafloor deformation. The LFEs then migrated south along the eastern ring fault and intensified and moved upward near where fresh lava first reached the sea floor in the central part of the caldera, consistent with the onset of small deformation recorded at the tilt sensors. We interpret the LFEs as localized brittle failure in response to magma movement prior to the eruption. We show that by combining supervised and unsupervised machine learning we can efficiently discriminate between various types of seismic sources at high spatial and temporal resolution, which in our particular case may help forecast the timing of Axial’s next eruption.

Session: Opportunities and Challenges for Machine Learning Applications in Seismology

Type: Oral

Date: 4/19/2023

Presentation Time: 02:30 PM (local time)

Presenting Author: Kaiwen Wang

Student Presenter: No

Invited Presentation: 

Authors