Deep-Embedded-Clustering of Microseismicity Identifies Multiple Failure Mechanisms at The Geysers Geothermal Field

Enhanced Geothermal Systems (EGS) regularly produce microseismic activity as fluid injection perturbs the thermo-hydromechanical state of the reservoir. Understanding the physical origin of this induced seismicity remains challenging due to the large injection and extraction volumes, the mechanical complexity of the reservoir, and the heterogeneity of seismic waveforms. Here, we apply a Deep Embedded Clustering (DEC) framework to perform unsupervised classification of microseismic waveforms recorded at the Geysers geothermal field in northern California during the 2011–2012 stimulation of the Prati-32 EGS well. Using both time-domain waveforms and frequency-domain spectrograms and cepstrograms, the DEC model clusters the convolutional autoencoder latent embeddings to identify three distinct failure classes. To assess which features enable the DEC to separate the events into distinct classes, engineered features are calculated and analyzed using an Extreme Gradient Boosting (XGB) classifier to rank the attributes that define each cluster. The spatiotemporal evolution of the event classes, combined with the feature importance, provides insight into the failure mechanism of each class. The identified classes correspond to: (1) fluid-driven events that migrate spatially following hydraulic diffusion, (2) tectonic fault-slip activity characterized by larger magnitudes and a spatial alignment with pre-existing mapped faults, and (3) a group of events with increased high-frequency to low-frequency ratios suggesting the formation of new fractures. These important insights about the failure mechanisms can be applied for safe operations of EGS and seismicity monitoring during active field production.