Abstract

Deep learning (DL)-based seismic picking models can improve the detection of induced seismicity compared to more conventional methods and reveal finer-scale patterns of induced seismicity. We present a high-resolution “deep” seismic catalogue for the Preston New Road-1z (PNR-1z) shale gas site in the UK, developed using DL phase picking on continuous downhole data. Using PhaseNet, we detected over 49,472 events, introducing up to 11,020 new events compared to the previous beamforming-based catalogue (38,452 events). The deep catalogue increases event cluster densities by an average of 42% and lowers the magnitude of completeness (Mc) from -0.2 to -0.5, adding more than 1,700 events to Gutenberg-Richter b-value analyses. The deepcatalogue also reveals new spatio-temporal patterns, indicating fault reactivation north of the PNR wells, contrasting with previous event locations and interpretations. These results demonstrate that DL-enhanced catalogues can refine subsurface interpretations, identify reactivated structures more eUiciently, and contribute to improved hazard assessment in geo-energy systems. However, further improvements are needed, as current models struggle to detect very small events (Mw < -0.5) in downhole data. The development of bespoke AI tools for downhole monitoring also remains challenging due to the limited availability of publicly accessible training datasets. To address this gap, we develop the AI-ready downhole Microseismic Benchmark (AMBER) database. AMBER contains picked event waveforms and seismic noise from 10 downhole microseismic datasets. Over 4,000 events and >150,000 individual P- and S-wave picks have all been manually verified. The database also includes over 10,000 examples of manually verified noise waveforms. AMBER provides a benchmark resource for the development and testing of robust AI tools for downhole microseismic processing.