Predicting CO2 solubility in brines using density as a proxy for detailed compositional data of brines: A data-driven modeling approach

Traditional thermodynamic or data-driven CO₂ solubility models require detailed compositional data of brine, which involves time-consuming laboratory procedures. This study proposes a novel machine learning framework that uses brine density at atmospheric conditions as a proxy for ionic composition, enabling CO₂ solubility prediction with only three inputs: temperature, pressure, and density. Using an accumulated experimental dataset comprising 9050 observations in water and NaCl brines across diverse temperature (273.15 to 478.15 K), pressure (1 to 1510 bar), and salinity (0 to 6 mol/kg) conditions, multiple regression and neural network models were developed and compared. A Back Propagation Neural Network model coupled with Adam optimizer demonstrated the best performance, achieving an R² score of 0.999 and a mean squared error of 0.00004 on the validation set. Trend analysis confirmed that the model captures the complex relationship between CO₂ solubility and temperature, pressure, and salinity, in accordance with thermodynamic principles. The model also demonstrated strong generalizability across other chloride-based brines and two produced water samples.This is the first study to generalize CO₂ solubility prediction using brine density as a full proxy for composition across wide operational conditions. The results suggest that atmospheric brine density is a reliable surrogate for brine composition, which enables assessment of CO₂ solubility trapping potential in field settings where full compositional data are unavailable. Model's practical utility was demonstrated by estimating CO₂ solubility trapping for 445 storage sites in central Gulf of Mexico, where over 111 megatons of storage capacity via solubility trapping were projected.