GSA Connects 2022 meeting in Denver, Colorado

Paper No. 204-10
Presentation Time: 2:00 PM-6:00 PM

DEEP LEARNING BASED GENERIC WELLBORE MODEL FOR WELL LEAKAGE ASSESSMENT IN GEOLOGIC CARBON SEQUESTRATION


BAEK, Seunghwan, Environmental Subsurface Science Group, EED, Pacific Northwest National Laboratory, 902 Battelle Blvd, Richland, WA 99354

Geology carbon sequestration (GCS) is a promising technology to alleviate atmospheric carbon dioxide emissions and increasing climate concern by storing captured CO2 in subsurface reservoirs. Considering the scale of the carbon storage volume needed to be impactful, oil and gas brownfields are an attractive option for immediate storage. The sites, however, have many historical wellbores which were drilled and used for oil and gas production, and they can act as a potential leakage pathway for injected CO2 and formation brine. Therefore, the risk management for GCS operations requires assessment of the potential for leakage through all wells within the Area of Review, which is the CO2 footprint or where reservoir pressures are high enough to lift brine into an overlying underground source of drinking water aquifer through an open conduit. Here, we develop a new physics-centric deep learning wellbore model to predict leakage of the injected CO2 and native brine through leaky wellbores. Multi-physics numerical simulation was used to generate synthetic data sets, and physics-informed features were introduced for model development. Neural network structures and hyperparameters were optimized with an automated searching algorithm. The systematic development leads to high predictive accuracy across a wide of ranges. The robust model provides an efficient way to manage risk management strategies and assess the viability of CO2 sequestration at specific sites for GCS operations.