Impacts of Deep Learning to Detect Induced Seismicity: a Case Study from Preston New Road, UK

Machine learning (ML) phase pickers, like PhaseNet, have emerged as valuable tools for automated phase picking. However, their potential benefits have been largely unexplored in induced microseismicity (Mw < 0). We address this critical research gap by investigating the implications of ML-enhanced phase detection for induced microseismicity at unconventional exploration sites.

We applied PhaseNet to a high-frequency (2000 Hz) borehole seismic dataset from a shale gas exploration site, Preston New Road (PNR-1z), UK. Despite being trained on 100 Hz data, PhaseNet detects over 52,000 earthquakes (with 15,800 being newly identified) in a 3-month period. In comparison, the well operator’s prior method, Coalescence Microseismic Mapping, detected over 38,400 events within the -2.8 < Mw < 1.1 magnitude range.

We address challenges in magnitude estimation due to waveform clipping by calibrating a coda duration-magnitude scale for larger, clipped event waveforms and the new events. We demonstrate that off-the-shelf PhaseNet detects numerous new small events (Mw < −0.5). This results in denser sampling of magnitude distributions and a lower magnitude of completeness, which affects Gutenberg-Richter b-value estimation. These findings have implications for improving seismic hazard assessments and enable a higher resolution analysis of the spatio-temporal evolution of induced microseismicity.