← Back to Sessions

Deep Learning Phase Pickers: How Well Can They Detect Induced Seismicity?

Description: 

Deep learning phase picking models can efficiently process large volumes of data typically produced during microseismic monitoring. These models have detected earthquakes not cataloged by existing traditional methods (e.g. STA/LTA, autocorrelation). Additionally detected earthquakes can affect the statistical properties of a catalog (e.g. b-value) which is important for seismic hazard assessments. Deep learning enhanced catalogs may help us study different driving mechanisms of hydraulic fracturing induced seismicity (HFIS) by observing the spatiotemporal evolution of HFIS in greater detail. We compared four existing models (GPD, U-GPD, PhaseNet and EQTransformer) pre-trained on large volumes of regional earthquakes recorded on surface station datasets (100 Hz) and investigated how well they identify seismic phases in high-frequency (2000 Hz) borehole array data. The PNR-1z dataset comprises continuously recorded injection operations at a hydraulic fracturing site in Preston New Road, UK, where operators cataloged >38,000 events using the Coalescence Microseismic Mapping (CMM) method. We generated earthquake catalogs for the PNR-1z dataset to compare (benchmark) against this initial catalog.

Results show that some models, particularly PhaseNet, detects seismic phases robustly within our data: they recover up to 95% of the initial catalog and detected >15,800 additional events (36% increase). PhaseNet’s robust application on our dataset could be due to its exposure to different instrument data during training, as well as its comparatively small model size which likely reduces overfitting to its initial training set. The GPD, U-GPD and EQT models require fine-tuning or re-training to detect more microseismic events. We conclude that PhaseNet can be applied off-the-shelf to detect HFIS in high frequency borehole data. The newly detected events could reveal new insights into the mechanisms controlling the spatiotemporal evolution of seismicity during fluid injections.

Session: Understanding and Managing Induced Seismicity

Type: Oral

Date: 4/19/2023

Presentation Time: 02:45 PM (local time)

Presenting Author: Cindy Lim

Student Presenter: Yes

Invited Presentation: 

Authors