Learning from laboratory earthquakes: A transfer learning approach to active source seismic monitoring

Concerns about seismic hazards associated with fluid injection necessitate close monitoring of geothermal reservoirs to anticipate and minimize the threat of induced earthquakes. Here, we apply machine learning to a series of datasets from laboratory-scale friction experiments coupled with time-lapse active ultrasonic monitoring to predict laboratory earthquakes. We demonstrate how accurately and how far in advance ultrasonic (velocity and amplitude) features could foretell the shear stress state of laboratory faults. The outstanding question is whether the knowledge learned from laboratory experiments could be transferred to field applications. In this study, we address a number of challenges associated with transitioning from laboratory to the field. We use transfer learning and stacked generalization to investigate how the knowledge gained from one experiment could be utilized to make accurate predictions on a small-sized dataset from a different experiment. Our results show that with much less training, both methods outperform a standalone model that has been trained with no prior knowledge. Moreover, stacked generalization provides superior performance to that of transfer learning. The proposed methods introduce a prospective approach for induced seismicity prediction using time-lapse active source seismic data with potential applications in monitoring of geothermal reservoirs, carbon storage sites, and unconventional energy reservoirs.