Subsurface carbon capture and storage has the potential to substantially reduce net carbon dioxide (CO2) emissions and provide potentially important benefits for the environment, US energy security and economic growth. Assuring safe operation of carbon storage facilities requires effective methods for monitoring complex CO2 reservoirs over decades to detect fluid leakage. While state?of?the?art electromagnetic sensing technology is now well suited for cost effective and accurate long?term monitoring, it is currently limited by computationally intensive inversion methods that require weeks or more to provide subsurface images of the CO2 distribution. Facility operators require near real?time, low cost, and actionable information about the condition of the reservoir. This project aims to address this bottleneck by developing a system for the rapid imaging of CO2 storage reservoirs over time (4D imaging) by applying deep learning advanced computational methods to high quality electromagnetic sensor data recorded with a recently proven technology. Deep (machine) learning approaches have the promise of reducing computing time from weeks to minutes compared to the physics?based inversion modeling that is currently in use. In Phase I, we will demonstrate a deep learning model suitable for replacing time?consuming inversions for CO2 storage monitoring. We will train deep learning models with synthetic data (and known targets) with real world environmental noise added from our extensive field survey database. Deep learning models will be optimized and evaluated using statistical methods to provide not only the prediction of CO2 leakage events, but also its associated uncertainty. Importantly, we will use field survey data acquired during a previous DOE?funded project on an active CO2 injection site to enable us to compare our best deep learning model to conventional inversions. The principal benefit of this project, if successful, is a substantial improvement in CO2 reservoir monitoring capabilities at reduced operational costs, which is anticipated to bring about more widespread adoption of this technology. This will be achieved by combining fast deep learning computational methods optimized for time?lapse monitoring with cost?effective electric field sensors. Additional commercial applications could include enhanced oil recovery (EOR) for CO2 and other fluids and ultimately even subsurface imaging in general.
