Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

Poster #215, Seismology

Microseismic monitoring using deep learning

Cindy Lim, Sacha Lapins, Maximilian J. Werner, & Margarita Segou
Poster Image: 

Poster Presentation

2021 SCEC Annual Meeting, Poster #215, SCEC Contribution #11440 VIEW PDF
Induced seismicity can pose great risks to subsurface fluid injection projects (e.g., hydraulic fracturing for shale gas, enhanced geothermal systems and wastewater injection) and more importantly, to local infrastructure and society. These risks underline the importance of monitoring, modelling and understanding induced seismicity. Operators now frequently record vast seismic datasets to detect and monitor hydraulic fracturing induced seismicity (HFIS) with downhole arrays. Deep learning models can offer rapid event detection in large datasets, which can be useful for real-time risk mitigation strategies or subsequent analysis. Such models have already demonstrated success in detecting regi...onal earthquakes. Here, we examine to what extent several prominent pre-trained neural networks can detect HFIS, including the GPD model, EQTransformer, and a more recent U-GPD model. We use data from Preston New Road, a UK shale gas site, where over 38,000 events (-2.8 ≤ Mw ≤ 1.2) were catalogued in 2018 by the coalescence microseismic mapping (CMM) method. Results show that the models managed to identify up to 76.6% (29,224 of 38,452 events) of the pre-existing CMM catalogue and started to struggle with event detection at Mw < -0.5. Although the models miss smaller events, they are able to detect new events previously uncatalogued by the CMM method (e.g., 980 new events within an hour). The models run very efficiently- at least four times faster than existing methods. However, the models do not generally have good picking accuracy and consistent phase detection of microseismic phases (with low signal-to-noise ratio) as they were trained to pick larger, regional phases. This study indicates that pre-trained deep learning models struggle with accurate phase detection of low magnitude, low SNR events; however, they still offer much potential for microseismic monitoring with their efficiency in analysing large datasets.