Deep learning based microearthquake location prediction at Newberry EGS using physics-informed synthetic dataset

This study introduces an efficient machine learning workflow for quickly predicting microearthquake (MEQ) locations within Enhanced Geothermal Systems (EGS), focusing on the Newberry EGS site. We leverage on a high-resolution field 3D P-velocity model to simulate physics-informed synthetic MEQ events and corresponding acoustic waveforms. We introduce a deep learning-based method namely, probabilistic multilayer perceptron (PMLP), to predict MEQ locations based on time lags from the simulated waveforms. Unlike most conventional deep learning methods, the PMLP enables uncertainty quantification. In synthetic tests, the model demonstrates low prediction errors (~ 40 m distance error). Our approach involves an initial time investment in creating synthetic waveforms, however, the fast prediction speeds (few seconds) present a compelling advantage for potential real-time EGS monitoring. Ongoing work includes applications to 2012 and 2014 field waveforms.