Optimizing Geologic Carbon Storage with Machine Learning: Impurity Effect on Trapping Mechanisms

Abstract Geologic Carbon Storage (GCS) presents a viable approach to mitigating greenhouse gas emissions. While utilizing impure CO2, the typical output from industrial emission sources, offers significant cost advantages compared to the expense of acquiring pure CO2. Nevertheless, the presence of even small impurity concentrations can alter CO2 trapping dynamics and potentially impact storage effectiveness. This study explores the application of machine learning (ML) models to optimize GCS operations, specifically focusing on predicting trapping mechanisms and the influence of impurity concentrations on CO2 behavior within geological saline aquifers. The research employs numerical reservoir simulation models and advanced machine learning predictive tools to identify key factors influencing CO2 storage performance. Our reservoir simulation models were used to generate training data for ML models based on Latin Hypercube Sampling method, with including in total 723 reservoir simulation cases. Further, four ML models for different trapping mechanisms were trained based on these simulation data. These models were first verified with traditional physics-based simulators, and achieved high predictive accuracy with R2 score at 0.99, but only cost at a fraction of the CPU time of the physics-based simulation models. The numerical experiments based on the ML models show that the co-injection of impure gas components such as N2, H2S, CH4, and CO2 can affect the CO2 trapping efficiency as well as the plume migration. Therefore, the ML models provide accurate and efficient predictions of CO2 trapping, which can provide valuable insights for optimizing injection strategies, thus enhancing decision-making for stakeholders.