DMLoc: Automatic Microseismic Locating Workflow Based on Deep Learning and Waveform Migration

Abstract

Abstract During hydraulic fracturing, real-time acquisition of the spatiotemporal distribution of microseismic in the reservoir is essential in evaluating the risk of induced seismicity and optimizing injection parameters. By integrating deep learning with migration-based location methods, we develop an automatic microseismic locating workflow (named DMLoc). DMLoc applies deep learning to automate phase picking and leverage the phase arrival probability function generated by a convolutional network as the input for waveform migration. The proposed workflow is first applied to the continuous data of the Dawson-Septimus area. Compared with a reference catalog generated by the SeisComP3 software, our method automatically locates 57 additional seismic events (accounting for 43% of the events in the obtained catalog). We further evaluate the performance of DMLoc by applying it to a 35-day continuous microseismic dataset from the Tony Creek Dual Microseismic Experiment. The spatiotemporal distribution of our detected events is consistent with results reported in prior catalogs, demonstrating the effectiveness of our method. In contrast to using raw microseismic records for stacking, DMLoc addresses the issue of inaccurate locating caused by low signal-to-noise ratios and polarity changes. The use of GPUs has substantially accelerated the calculations and enabled DMLoc to output locating results in minutes. This fast and efficient metric could be easily extended to any microseismic monitoring scenario that requires (near) real-time locations and assists in site-based risk mitigation and industrial operation optimization.