Quantification of the Effect of CO2 Storage on CO2-Brine Relative Permeability in Sandstone Reservoirs: An Experimental and Machine Learning Study

Abstract Injection of carbon dioxide (CO2) into subsurface reservoirs is a pivotal component of carbon capture, utilization, and storage (CCUS) technologies, aimed at mitigating the adverse impact of anthropogenic greenhouse gas emissions on climate change. A critical aspect influencing the efficiency of CCUS is the interaction between CO2 and brine within sandstone reservoirs, particularly how CO2 storage affects the wettability of the rock and, consequently, dynamics of fluid flow characterized by the relative permeability and capillary pressure. This study aims to quantitatively assess the relative permeability between CO2 and brine in Berea sandstone, a common analog for reservoir rocks, and the implications of it in CO2 storage. Prior to performing the quantitative assessment, a detailed analysis of static reactivity of the rock-fluid system has been completed. This study integrates experimental and machine learning (ML) approaches to assess the effects of CO2 storage on CO2-brine relative permeability in Berea sandstone. We conducted a 30-day CO2 storage experiment under reservoir conditions (1500 psi and 150°F), analyzing rock and fluid properties before and after storage. These experimental data provided valuable insights into the physical and geochemical changes occurring due to CO2 exposure. Before applying ML approaches, data for this study were collected from a combination of available experimental measurements and literature sources, which were curated to include key input parameters known to influence relative permeability behavior. The dataset consisted of the following features: porosity (%), permeability (mD), interfacial tension (IFT, mN/m), salinity (mg/L), pressure (MPa), threshold capillary pressure (kPa), temperature (°C), and gas saturation (Sg). These features were used as input variables, while the target variables were the gas relative permeability (krg) and water relative permeability (krw). To predict relative permeability curves, an XGBoost machine learning model was trained on the dataset. The input data were processed into feature matrices along with corresponding target variables for both gas and water relative permeability. The model demonstrated high accuracy, with prediction errors for both water and CO2 relative permeability being less than 10% absolute relative error (ARE) when compared to experimental values. After developing the model, we applied the experimental data collected before and after CO2 storage, allowing us to quantify the effect of CO2 storage on changes in relative permeability. This machine learning approach offers a fast and cost-effective alternative to the time-consuming and expensive core flooding experiments. The insights from this study are expected to improve CCUS predictive models and optimize CO2 injection strategies and storage capacity in sandstone reservoirs, thereby contributing to climate change mitigation efforts.