Real-time hydraulic fracturing monitoring using deep learning clustering of microseismic data

How to effectively monitor fracture growth using induced microseismicity during fracture stimulation is an unsolved problem. We observe two distinct types of long-duration (LD) microseismic events recorded using borehole geophones at an unconventional oil field: one exhibits frequency-drop long-duration (FDLD) characteristics, and the other exhibits low-frequency long-duration (LFLD) characteristics. We discover that only the LFLD events, but not the FDLD events, are associated with the fracture growth. To cluster LFLD, FDLD, and non-LD signals from terabytes of continuously-recorded microseismic data, we design a 7-layer U-Net convolutional network to cluster the microseismic events in real time. We find that LFLD events occur only during the proppant injection period, and their spatial distributions expand away from the fracture stimulation wells with time. We hypothesize that the LFLD events are induced by fluid movement in the proppant-opened fractures. We find that the cumulative seismic moment of LFLD events is proportional to the cumulative amount of injected proppant. Detecting LFLD events using deep learning clustering provides a new approach to real-time monitoring of fluid flow and hydraulic fracture growth during fracture stimulation.