Multi-scale Unsupervised Machine Listening to Geothermal Earthquakes in the Geysers, CA

Abstract

Understanding seismicity patterns in geothermal reservoirs can be an important source of information on the thermodynamic state and geomechanical changes in the reservoir. Here, we apply unsupervised machine learning (feature extraction and clustering) to earthquakes in The Geysers geothermal reservoir in Northern California. The feature extraction approach is comprised of sequential non-negative matrix factorization (NMF) and hidden Markov modeling (HMM), which together reveal subtle spectral differences in among earthquakes. Fingerprints generated from the HMM results are then clustered using k-means. We search empirically for correlations of the clusters with physical aspects of the earthquakes, such as spatial locations, moment tensor components, or event times. In our first study, we found that strong clustering emerged from analysis of 46,000 earthquakes over the entire Geysers region in three years. These clusters formed no clear spatial patterns but very strong temporal patterns, essentially defining seasonal earthquakes types that recur annually, closely correlating with average fluid injection rates. In order to better understand these patterns, we analyze three more data sets: (1) Zooming out in time, 12 years of seismicity over the entire Geysers area; (2) Zooming in in space, to a small area in the NW Geysers, over a 7-year period, and (3) an adjacent smaller area in the NW where experiments in enhanced geothermal systems (EGS) were performed. These analyses are currently works-in-progress, but initial results from (2) reveal no apparent correlations with moment tensor properties (all types of moment tensors occur in each cluster). In all 3, we find strong temporal correlations, but with higher resolution in the correlations between injection rate and clustering. This result strengthens our inference that changes in the thermal-mechanical state (stress changes associated with water and steam content and distribution on fault interfaces) causes subtle variations in seismic source spectra. Sonic representations (sonification) of the seismic data aid in the interpretation of the machine listening results.