Advances in Machine Learning for Seismic Event Detection

Advances in computational power and storage are facilitating a new era of modelling. As the amount of information captured within datasets has evolved, tools have emerged to better exploit the statistical properties of such datasets. Machine Learning (ML) methods are one such family of techniques, leading the revolution of 'data-driven' solutions which achieve state-of-the-art performance in solving tasks across both business and the sciences. ML is concerned with the automated discovery of the governing relationships within data distributions. The algorithms can be thought of as a suite of generalized algorithms for extracting information. Seismology is a field naturally suited to the application of ML, containing high-quality catalogs of seismic recordings - collected over decades - which are crucial inputs into many seismological studies.

With seismology only starting to widely integrate the latest state-of-the-art ML research over the last few years, the lack of uptake means that huge performance increases may be possible for traditional tasks. The detection of arriving seismicity is one such area. Having more complete seismic catalogs means imaging smaller magnitude events, 'closing the gap' between seismicity observed in nature and what can be simulated in laboratory environments. ... mehrBetter-resolution seismic catalogs are, therefore, crucial for enhancing any physical understanding of seismogenic rupture processes.

This thesis investigates the extent to which ML can improve the task of detecting seismic events. We first explore and propose new methods for seismic phase arrival identification, applying a Convolutional Neural Network (CNN) trained for supervised classification of labelled seismic arrivals. We also propose a novel algorithm for the subsequent stage of the event detection pipeline, the task of seismic phase association. Our proposed association algorithm is designed to operate efficiently in a scalable and robust manner. We adapt a parametric model fitting framework to extract a physical model of the seismic wavefield moveout to associate picks to events. We then turn to the question of how to best evaluate the state-of-the-art ML in seismology. Here, we can also leverage practices from ML-focused fields such as computer vision, and natural language processing. Access to both open-access benchmark datasets and models is crucial for accelerating the research process.
This thesis introduces software specifically designed for this task - SeisBench. It aims to significantly reduce the amount of work required to conduct ML research in seismology, accelerating development and iteration. We finally explore how the performance of the latest ML models varies when moving from well-curated benchmark training datasets to practical pipelines in less well-explored seismic environments.

We hope our exploratory work and subsequent results, including productionized software, will facilitate further in-depth research in applying ML to one of the most fundamental tasks in seismology.