Unsupervised Learning-Based Analysis of Hydraulic Fracturing-Induced Seismicity

Abstract

Passive seismicity is crucial for optimizing hydraulic fracture treatments and reducing the risk of hydraulic fracturing-induced seismicity. Unsupervised machine learning is suitable for understanding fracturing-induced seismicity because it can process large amounts of data without prior knowledge or labeling and identify patterns and relationships in the data that isn’t apparent to the human eye. Clustering and dimensionality reduction can be used to group similar earthquakes and reveal patterns in the distribution and evolution of seismicity over time. This can provide valuable insights into the behavior of the fractures and help identify the conditions that may lead to induced seismicity. In case of meso-scale (~ 10 m) microseismicity, the analysis is based on Experiment 1 of the EGS Collab Project. A non-linear dimension reduction is applied on the high dimensional features extracted from seismic signals generating embeddings in low dimensions. The different groups obtained by clustering the low dimension embeddings reveal that different clusters are related to distinct fracture networks. A separate study on the same experiment focused on the simultaneous, wide-band hydrophone signals. Low frequency signals (2-80 Hz) were found to be strong related to the fluid injection. Workflows were developed to reliably locate the sources of the low frequency signals and their spatiotemporal distribution was correlated with the distribution of natural fractures. Additionally, workflows based on semi-supervised, graph-based label propagation are developed to reliably extend fracture labels to noisy and unlabeled microseismic data. Lastly, a field (~km) scale microseismic dataset containing high quality moment tensor information was used to explore relationship between the three-dimensional motion recorded at geophones and the deformation occurring at hydraulic fracture planes.