Near-real-time microseismic monitoring with machine-learning and waveform back-projection at the Utah FORGE geothermal site

Deep geothermal energy exploitation necessitates establishing effective fluid circulation paths for heat transfer and managing induced earthquake risk. By detecting and characterizing induced microseismic events, we can provide insights into the fracture network growth and the induced earthquake risk during hydraulic stimulation and geothermal production in enhanced geothermal systems (EGS). During hydraulic stimulation, monitoring has to be performed in near-real-time to provide timely information for assessing potential earthquake risk and for adjusting the stimulation plan. In addition, high-precision microseismic event location is vital for evaluating the connectivity of the stimulated reservoir and designing the trajectory of the production wells. However, achieving real-time monitoring and high-resolution location in a single monitoring workflow is challenging due to the low signal-to-noise ratio and short inter-event time of microseismic events.

To address these challenges in microseismic monitoring, we build a near-real-time monitoring workflow that integrates machine-learning (ML) techniques for efficient event detection and waveform back-projection methods for high-precision event location. The proposed workflow is designed to utilize various pre-trained ML models to deal with the scarcity issue of training datasets in new EGS sites. We apply the proposed workflow to the microseismic dataset collected at the Utah FORGE geothermal site in a playback mode. Because most pre-trained ML models are trained on local earthquake datasets having larger event magnitudes and lower data sampling rates, we implement and evaluate various strategies, such as re-scaling, re-sampling, and filtering, to enhance the performance of pre-trained models on the microseismic dataset. We compare the obtained ML catalog with a reference catalog built from a conventional workflow consisting of automatic phase picking and manual refinement. Due to the application of ML and waveform back-projection techniques, our workflow can nicely separate microseismic events with very short inter-event times (in terms of a second) and cope with events with significant magnitude/amplitude differences, leading to more reliable event detections. Detailed comparisons show that the accuracy of ML phase identification is comparable to and sometimes even superior to manual picking (with a difference in milliseconds), which contributes to precise event locations.