Machine Learning–Based Automatic Microseismic Event Detection During the 17 August 2015 Mw 4.6 Induced Earthquake Sequence in Northern Montney, British Columbia, Canada

ABSTRACT Microseismicity monitoring is critical for better understanding fracture growth and reservoir behavior in hydraulic fracturing (HF) operations, as well as fault activation processes of induced earthquakes and associated seismic hazards resulting from anthropogenic processes. Detecting seismic events and picking the first arrivals of P- and S-wave phases are essential tasks in microseismic monitoring and safe resource management. In this study, we conduct a controlled domain-transfer benchmark of state-of-the-art deep learning (DL) models for microseismic event detection on a geophone dataset recorded to monitor HF operations between 10 August and 10 September 2015 in the northern Montney play of British Columbia, Canada. During this period, an induced earthquake of Mw 4.6 occurred. After evaluating multiple pretrained pickers, we select PhaseNet and optimize its performance under significant domain shift by increasing the input data sampling rate from 100 to 500 Hz, significantly enhancing its ability to accurately identify low-signal-to-noise ratio (SNR) microseismic phases. By applying a workflow of automatic phase association and event location algorithms, we build a high-confidence, precision-oriented event catalog that successfully identifies ∼55% of a reference industrial catalog generated via migration-based stacking. The DL catalog shows highly consistent spatiotemporal patterns and moment magnitude (Mw 4.6) detection capabilities. We demonstrate that the detection gap primarily reflects fundamental differences between single-station classifiers and network-level brightness methods, with missed events typically being either true low-SNR arrivals or visually noise-like segments in the reference catalog rather than being obscured by temporal signal overlap. These findings demonstrate that the open-source package-based workflow for automatic event detection and localization may work sufficiently well compared with proprietary commercial software in microseismic monitoring, offering a viable open-source alternative for HF operators and regulatory agencies.