Extracting Information from Geophysical and Geochemical Signals: Applying Machine Learning Through Data Science Challenges II

The earthquake rupture process reflects complex interactions of stress, fracture and frictional properties. Our fully unsupervised machine learning methods, based on methods for image, text and audio analysis, reveal subtle spectral-temporal variations in similar seismic signals. Specifically, N seismograms comprising an earthquake catalog recorded at a particular station ($N>O(3)$) are converted to spectrograms, then (1) decomposed by non-negative matrix factorization (NMF) into a single dictionary and N activation matrixes, (2) reduced by hidden Markov models (HMM) to N fingerprints and (3) clustered by K-means. These three steps produce clusters of subtly similar signals. Results from 46,000 earthquakes of $0.3<M<1.5$ over 3 years from the entire Geysers Geothermal Field (CA) show clusters that have no reservoir-scale spatial patterns, but clear temporal patterns: events with similar spectral properties repeat on seasonal cycles within each cluster and track closely the water injection rates into the Geysers reservoir (Holtzman et al., Sci. Adv., 2018). The challenge with this powerful new ML tool is to understand the physical causes of the subtle variations in spectral properties. These patterns indicate that changes in the fluid/vapor ratio in fractures (thermo-mechanical state) affect the reservoir acoustic properties and faulting processes (friction and thermal-tectonic-poroelastic stresses). Previous studies focused on a small area in the NW Geysers have identified systematic changes in moment tensor orientations and non-double couple components (e.g. Mart\'inez-Garz\'on et al., GRL, 2017) and stress drop (e.g. Kwiatek et al., JGR, 2015) associated with changes in water injection rate. Here, we apply our ML methods to a catalog of $>$6,000 events ($M>0.2$) over 5 years from the same small area, which behaves as a relatively isolated hydraulic compartment. Initial results show complex temporal clustering with clear correlations to changes in injection rates and moderate correlation to ratios of normal and strike-slip faulting. Our ML approach to seismicity characterization opens many possibilities for real-time identification of changes in microseismicity and many questions on interpreting physical causes of spectral variations.