Testing a multi-station deep learning seismic phase picker trained on surface array data in a high frequency downhole setting

Summary

Induced seismicity can damage infrastructure, cause injury and generate negative public perceptions for geo-energy operations. To monitor, forecast and understand induced earthquake physics, phase picking models must accurately detect and classify microearthquake signals. Deep learning (DL) algorithms can detect and classify seismic signals to the accuracy of a human analyst in real-time. Recent advances have introduced multi-station DL algorithms which leverage two-dimensional (2-D) spatial information to improve the detection of low signal-to-noise ratio arrivals. However, most such models are trained on surface networks, and applying them to borehole arrays which differ in sampling rate, station geometry and waveform characteristics, remains an open challenge. Here, we develop input adaptation strategies that allow a surface-trained multi-station DL picker to be applied effectively to borehole data, without retraining. Our central finding is that rotating the input traces so that the observed P-and S-wave energies align with the input channels the model has learned during training substantially improves detection. We demonstrate this by applying CubeNet, a multi-station DL model, to detect induced seismicity at the Preston New Road 2 (PNR2) hydraulic fracturing site, UK using a borehole array and an existing catalogue. Using station geometry information and rotating the borehole traces to mimic surface traces increases successful P and S picks by 9.5 % points and 13 % points respectively for −1.5 ≤ Mw ≤ 1.2. We also develop a workflow to explore how the increases in the number of picks affect the quality and number of phase associations and located earthquakes. By rotating the borehole traces, we show a 65 % increase in the number of earthquakes located in a one hour period (203 to 335 events). Moreover, CubeNet can operate 1.86 times faster than real-time data recording and our relatively computationally efficient workflow can reproduce previous location estimates. We recall 100 % of earthquakes above Mw -0.2, but only 18 % of the existing catalogue because just ~16 % of CubeNet’s original training dataset consisted of events with Mw < -0.2. We conclude that these input adaptation strategies can be implemented to other surface-trained multi-station DL pickers applied to borehole arrays. Our proposed methods offer a practical route to deploying existing models on new borehole geometries without retraining. While CubeNet proves useful for near real-time monitoring of larger events, detecting smaller (Mw < −0.2) events will require transfer learning on low SNR events, or indeed the development of bespoke models for downhole datasets.