Lighting Up a 1 km Fault near a Hydraulic Fracturing Well Using a Machine Learning-Based Picker

Abstract The development of portable nodal array in the recent years greatly improved the seismic monitoring ability across multiple scales. The dense arrays also directly benefit microseismic monitoring by providing relatively low-cost surface recordings. However, the rapid growth of seismic data is accompanied by the increased demand for efficient seismic phase picking. On the other hand, machine learning-based phase picking techniques achieved high stability and accuracy, showing promising potential to replace human labors and traditional automatic pickers. In this study, we applied a state-of-the-art package on newly collected nodal array data around a hydraulic fracturing well in southwestern China. The array consists of up to 85 nodes with an average station spacing of less than a kilometer. Within the hydraulic fracturing stimulation periods, we detected ∼3000 seismic events with magnitude down to ∼−2. After waveform cross-correlation-based relocation, the 1979 relocated events clearly light up a 1 km long fault structure and several fractures. Furthermore, the frequency–magnitude distribution of the catalog exhibits weak bilinear features with relatively low b-value (0.88) and a moderate coefficient of variation (Cv ∼2). The nature and origin of the observed earthquake cluster are then discussed and defined based on the industrial information, high-resolution earthquake catalog, and basic statistics. Finally, we summarized our experience and provided recommendations for applying similar approaches to other local scale, surface microseismic monitoring scenarios.