CO₂ Solubility Prediction in Saline Brines for Geological Sequestration

Quantifying CO₂ solubility in brine is a critical aspect of evaluating the efficiency and security of geological carbon storage (GCS), particularly in low-salinity subsurface formations. This study presents a novel and comprehensive machine learning (ML)-based framework for predicting CO₂ solubility in NaCl-bearing aqueous systems under a wide range of thermodynamic conditions representative of low saline aquifers. A suite of supervised learning algorithms such as ExtraTrees, CatBoost, Random Forest, Decision Tree, AdaBoost, Ridge, and second-degree polynomial Ridge regression were developed and rigorously assessed. Among these, the ExtraTrees model demonstrated superior predictive performance, achieving an R² of 0.9954, RMSE of 0.0279, and MAE of 0.0217. The polynomial Ridge model also yielded competitive results among conventional regressors. Model interpretability was enhanced using SHAP, which identified pressure as the most influential factor affecting CO₂ solubility, followed by temperature, while salinity had a limited impact within the studied range. These findings align with established thermodynamic principles, including Henry’s Law and the inverse relationship between temperature and gas solubility. To address dataset limitations, a data augmentation approach based on Gaussian copula was proposed. The ExtraTrees model retained robust performance on the augmented dataset with an R² of 0.87, confirming its generalizability. This work introduces an accurate and interpretable ML methodology for estimating CO₂ solubility in low-salinity brines. The results underscore the critical role of pressure in enhancing CO₂ dissolution and provide actionable insights for optimizing GCS operations. The proposed framework highlights the promise of data-driven models in supporting sustainable and safe subsurface carbon sequestration strategies.