Characterizing the Stimulated Reservoir Volume Using Manifold Learning on 3d Motions of Microseismic Sources

Abstract

Abstract Uniform manifold approximation and projection (UMAP) finds a low-dimensional representation of a complex dataset while preserving local structure and global distances present in high-dimensional data. Using UMAP, a high-dimensional microseismic data is transformed into a lower-dimensional, graph-based representation. In this paper, UMAP is applied on accelerometer signals recorded during an intermediate field-scale fracture propagation experiment at the Sanford Underground Research Facility in South Dakota. Our analysis shows that the short-time Fourier Transform of signals recorded by a single channel of the 3D accelerometer is the best feature extraction technique for the desired representation learning. For the first time, we have successfully identified the distinct fracture planes corresponding to each micro-earthquake location using accelerometer data from an intermediate-scale hydraulic stimulation experiment. Due to the scatter of microseismic events and associated uncertainty of microseismic event locations, the assignment of reliable fracture-plane label to a microseismic event is possible only for a small percentage of events that have sufficiently high signal to noise ratio, and/or are favorably located for unambiguous assignment. In this paper, we also present a semi-supervised learning method based on UMAP for label propagation to assign fracture-plane labels to many unlabeled microseismic events, utilizing a limited amount of labeled data, which have pre-assigned fracture-plane labels. The proposed semi-supervised label propagation on a low-dimensional manifold achieves exceptional precision and recall (>92%) while employing just a small percentage (~20%) of microseismic events with fracture-plane labels. With just a small set of labeled microseismic events, it is possible to assign fracture-plane labels to four times as many microseismic events.