Advancing Microseismic Event Detection with Deep Learning Phase Pickers

: Towards High-Resolution Catalogue Development and Real-Time Monitoring of Induced Seismicity

Abstract

Human-induced seismicity is an expected outcome of fluid injection during geo- energy operations, as microseismic events often signal successful fracture stimulation. Larger or felt events, by contrast, can occur as an unintended consequence and pose seismic hazards. Monitoring induced seismicity is therefore essential for understanding the subsurface response and managing seismic hazards during industrial activities such as hydraulic fracturing (HF). However, traditional seismic processing workflows often struggle to detect and accurately characterise the large number of microseismic events that occur during stimulation, especially in real-time. This thesis presents a complete microseismic event catalogue workflow, spanning detection, location and magnitude estimation, with a focus on apply- ing deep learning (DL) for enhanced event detection. We evaluate the workflow using data from the Preston New Road-1z (PNR-1z) HF shale gas well in the United Kingdom (UK), which offers a uniquely rich dataset through continuous high-frequency downhole monitoring from an adjacent well, along with detailed injection records. The first part of the study assesses four pre-trained DL phase pickers and demonstrates that PhaseNet was the most consistent and highest- performing model overall, achieving the highest recall (∼95%) from the benchmark catalogue (which had over 38,000) and identifying over 49,000 events. PhaseNet and U-GPD are more computationally efficient than the previous beamforming- based method, achieving real-time or faster processing speeds on standard laptop hardware. Both models also identify additional events, corresponding to increases of 41% and 22%, respectively, over the benchmark catalogue. To address the known issue of false positives in DL detections, we developed a novel validation method, the Linear Event Moveout Filter (LinMEF), which exploits P-wave moveout patterns along the linear borehole array. LinMEF successfully removed ∼94% of false detections in a test sample, substantially improving catalogue reliability. Sub- sequently, with the PhaseNet detections, we used a broader spatial search volume beyond that of the original beamforming workflow to determine event locations using absolute travel-time methods (NonLinLoc). This approach revealed new spatial patterns of seismicity, including the activation of seismogenic structures north of the injection wells that were not identified in the previous catalogue. The expanded spatial coverage and increased event cluster densities enable improved structural interpretations and provide a higher-resolution view of fault activation during stimulation. Notably, the largest events were spatially aligned with seismic discontinuities identified in this study using existing 3D seismic reflection data, supporting interpretations of slip on pre-existing fault structures on the north side of the two injection wells. The final stage of the workflow focused on robust moment magnitude (Mw) estimation using borehole waveform data and subsequent analyses of the frequency-magnitude distributions (FMDs). In contrast with previously underestimated contractor Mw, our downhole Mw estimates were consistent with those derived from surface broadband stations, confirming the viability of using downhole data for reliable magnitude estimation. Although we inferred a log-linear relationship between Mw and coda duration (tcoda), coda duration magnitudes (Md) proved less consistent and their tcoda measurements more sensitive to noise and overlapping events than Mw. Importantly, the DL-enhanced detection made the catalogue more complete and extended it down to lower magnitudes, bet- ter capturing the population of smaller events that are critical to characterising FMDs. This enhanced representation enabled robust and statistically significant changes in the Gutenberg-Richter b-value. We interpreted the temporal changes in b-value as a shift from early-stage tensile fracturing (high b-values) to fault reactivation, involving proportionally larger events (lower b-values). In summary, we show that the detection rate and computational efficiency offered by DL phase pickers can meaningly improve seismic processing workflows and indicates clear potential for real-time implementation during microseismic monitoring. While the catalogue size increased by approximately 30%, a more modest gain compared to the order-of-magnitude improvements seen in some regional-scale studies, this likely reflects the strength of the benchmark catalogue, which was produced using a computationally intensive and robust beamforming-based workflow. Further improvements may require bespoke DL models tailored to specific datasets. However, the development of such models remains challenging due to limited access to large, high-quality, labelled training datasets, which are often proprietary. This thesis demonstrates that, even using existing pre-trained models, DL offers a scalable and efficient alternative to current workflows and can enable detailed structural analyses for improved hazard assessment for regulatory applications in geo-energy exploration.

Date of Award	30 Sept 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Max Werner (Supervisor), Margarita Segou (Supervisor) & James P Verdon (Supervisor)