Collaborative Research: CAIG: Unsupervised Deep Clustering of Seismic Time Series Data at Volcanoes

Many natural systems, including those in the solid earth, are typically described by time series data. Seismic data are known to be summations of many concurrent sources with complex frequency patterns, rendering attempts at automated identification of behaviors difficult. In many cases, rapid identification of transient signals is critical to early warning systems and escalations in hazards, such as with volcanic systems.

The advent of machine learning methods has opened the door to unsupervised learning efforts, where a neural network can be tailored specifically to distinguish between different physical behaviors. This project designs an algorithm to learn the various forms of seismic behaviors that exist in a volcanic environment, such as icequakes, lava lake eruptions, ash vent signals, iceberg tremor, teleseismic activity and other phenomena.

The project will focus on Erebus volcano, Antarctica, where there are enough multi-scale arrays and where long-term well-understood data exist. The approach will map any time series segment to a position in a different space where similar seismic behavior types cluster closely and are separated from other families of events. Once validated at Erebus, the method will be tested at other glaciated volcanoes featuring long term seismic arrays, such as the Cascades in the Pacific Northwest and the Aleutian volcanoes in Alaska.

This project also aims to accelerate exposure of machine learning methods to underserved student populations through the hosting of a cross-disciplinary workshop.

Unsupervised time series clustering is a long-standing objective that spans multiple fields including seismology. Whereas classic approaches to clustering involve seismic segments that are tested against human-assigned features such as distributional skew, standard deviation, and others, recent machine learning methods have allowed for the flexible expansion of these efforts through unsupervised non-orthogonal basis pursuit methods.

Here, a multimodal variational autoencoder approach is proposed to process array-based continuous seismic data at Erebus volcano with the objective of blindly resolving families of different events contained within. These include icequakes, lava lake eruptions, ash vent signals, iceberg tremor, teleseismic activity, and different manifestations of ambient noise.

Once trained, transfer learning will be tested at other glaciated volcanoes in the Aleutian Islands and the Cascades, with the objective of generating a community Python package that will allow researchers to process any data set for preliminary behavioral analysis. A workshop will be hosted at University of Texas El Paso in year 3 to promote educational exposure and internship opportunities in machine learning and data science and will be aimed at underserved graduate and postdoc populations seeking to expand their skillsets.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.