Machine Learning Aids Rapid Assessment of Aftershocks: Application to the 2022–2023 Peace River Earthquake Sequence, Alberta, Canada

Abstract

The adoption of machine learning (ML) models has ignited a paradigm shift in seismic analysis, fostering enhanced efficiency in capturing patterns of seismic activity with reduced need for time-consuming user interaction. Here, we investigate automated event detection and extraction of seismic phases using two widely used ML models: EQTransformer and PhaseNet. We applied both the models to four weeks of continuous recordings of aftershocks using a temporary array following the 30 November 2022, ML 5.6 earthquake near Peace River, Alberta, Canada. Both the tools identified >1000 events over the recording period. The aftershocks are located in close proximity to the ML 5.6 mainshock as well as to wastewater disposal operations that were ongoing at the time. Both the methods reveal an aftershock distribution that was not identified by the regional network; however, we find that events detected by PhaseNet have smaller event location errors and better depict subtle fault structures at depth, despite identifying ∼200 events less than EQTransformer. Our results highlight the advantages of using ML models for rapid detection and assessment of seismicity following felt events, which is important for rapidly assessing seismic hazard potential and risk.