HEIMDALL: a grapH-based sEIsMic Detector And Locator for microseismicity

Abstract

Machine Learning (ML) applications in geoscience are growing exponentially, particularly in the field of seismology. ML has significantly impacted traditional seismological observatory tasks, such as phase picking and association, earthquake detection and location, and magnitude estimation. However, despite promising results, ML-based classical workflows still face challenges in analyzing microseismic data Leveraging recent advances in Deep Learning (DL) methods, we present HEIMDALL: a grapH-based sEIsMic Detector And Locator for microseismicity. This tool utilizes an attention-based, spatiotemporal graph-neural network for seismic event detection and employs a waveform-stacking approach for event location, using output probability functions over a dense regular grid. We applied HEIMDALL to a one-month dataset (December 2018) from the publicly available Hengill Geothermal Field in Iceland, collected during the COSEIMIQ project (active from December 2018 to August 2021). This dataset is ideal for testing seismic event detection and location algorithms due to its high seismicity rate (over 12,000 events in about two years) and the presence of burst sequences with very short interevent times (e.g., less than 5 seconds). We assessed the methodology's performance by comparing our catalog with those obtained by two recent DL methods and one manually compiled by ISOR for the same period. The DL algorithms we considered are: (i) MALMI, a waveform-based location algorithm, and (ii) the recent GENIE graph-neural-network algorithm. For GENIE, we conducted a full repicking of continuous waveforms using the PhaseNet picking algorithm and subsequent retraining of its model to adapt it to the new seismic network. Finally, we highlight the pros and cons of each method and discuss potential improvements for microseismic event detection and location, with a particular focus on induced seismicity monitoring operations at EGS sites.