Achieving long-term monitoring CO2 and geothermal fluids using deep learning

Challenges of using seismic data for long-term monitoring fluid plume and its characteristics (permeability) in the subsurface exist while simultaneously minimizing the potential for leaks and earthquakes (seismicity) for both the successful sequestration of carbon dioxide and the ubiquitous and effective recovery of geothermal fluids. Deep learning using neural networks (NN) have shown promise in solving highly nonlinear seismic inversion problems. However, the feasibility to field data is primarily hindered by the scarcity of field seismic data for NN training. We present a physics-informed deep learning solution to a limited amount of data collected in field settings. The key procedure of the proposed NN is to incorporate physically simulated synthetic datasets for training. To do so, we use site-specific geological information, fluid flow physics, rock physics, and seismic modeling to generate realistic synthetic datasets that closely match some of the field data. Our results from two case studies (FrioII CO2 injection and newberry geothermal site) suggest training NN on physics-informed synthetic datasets and applying the learned weights to field data is a viable approach to estimate field CO2 saturation and MEQs in EGS. This method effectively addresses the scarcity of field training data, indicating the feasibility of long-term CCS and EGS seismic monitoring.