Monitoring Subsurface Fracture Flow Using Unsupervised Deep Learning of Borehole Microseismic Waveform Data

Abstract

Fracture flow is the fluid movement in a fracture or a fracture zone. Since fracture flow can induce long-duration (LD) microseismic events, classifying different types of microseismicity is crucial for reliable monitoring of subsurface fracture flow. We analyze hydraulic fracturing-induced microseismic data recorded by borehole geophones and find four types of microseismic events: two types of short-duration events and two types of LD events. Among the two types of LD events, one contains frequency-drop LD (FDLD) characteristics, and the other exhibits low-frequency LD (LFLD) characteristics. We employ an unsupervised machine-learning algorithm based on the U-Net convolutional network to classify microseismic events. Our study shows that LFLD events occur only during the proppant injection period of hydraulic fracturing and that the spatiotemporal distributions of the LFLD events gradually grow from the fracture stimulation wells outward with time. Also, the cumulative seismic moment of the LFLD events is proportional to the cumulative amount of injected proppant. These results can be used to optimize hydraulic fracturing parameters in unconventional reservoirs.