Harnessing Ensemble Learning Techniques for Accurate Interfacial Tension Estimation in Aqueous CO2 Systems

Abstract The interfacial tension (IFT) of aqueous CO2 systems plays a critical role in determining CO2 capillary entry pressure, maximum CO2 storage height, and subsurface storage capacity. Conventional methods for measuring IFT are time-consuming and resource-intensive. This study therefore explores the application of ensemble learning techniques: Gradient Boosting (GradBoost) and Light Gradient-boosting Machine (LightGBM), to predict the IFT of aqueous CO2 systems. A comprehensive dataset of 1570 IFT data points, encompassing six features: pressure (0.1–69.51 MPa), temperature (5.2–196.25℃), monovalent and divalent cation molality (0–5 mol/kg), and methane and nitrogen mole fractions (0–80 mol.%), was compiled from the literature. The data was preprocessed and divided into 70%, 15%, and 15% subsets for model training, testing, and validation. Model performance was optimized through regularization and hyperparameter tuning. Statistical metrics and visualizations were employed for quantitative and qualitative evaluation of the models. The Leverage approach was used to identify potential outliers and ensure model reliability. Sensitivity analysis and feature importance were assessed using permutation importance and the Akaike Information Criterion (AIC). GradBoost and LightGBM exhibited remarkable performance, achieving a coefficient of determination (R2) exceeding 0.98, root mean square error (RMSE) below 2.00 mN/m, mean absolute error (MAE) lower than 1.2 mN/m, and average absolute percentage relative error (AAPRE) less than 1.5% for all data groups. GradBoost surpassed LightGBM in terms of accuracy (higher R2 of 0.99), precision (lower MAE of 0.87 mN/m), consistency (lower RMSE of 1.23 mN/m), and complexity (lower AIC of 53). Furthermore, GradBoost outperformed a committee machine intelligent system, a group method of data handling model, and other robust ensemble models such as random forest and adaptive boosting. Permutation importance and AIC revealed that pressure and monovalent cation molality were the least and most influential features on IFT prediction. Additionally, excluding any of the six input features significantly reduced model performance, with AIC increasing by 6 and 9 folds without temperature and pressure data, respectively. The Leverage approach confirmed the statistical validity and reliability of the GradBoost model, identifying only 3% of the total data points as potential outliers. This study demonstrates the effectiveness of ensemble learning techniques in capturing the complex relationships between variables that govern the IFT of aqueous CO2 systems. The constructed ML models offer a rapid and reliable alternative to conventional methods for IFT evaluation, speeding up workflows, and reducing experimental uncertainties. These advancements hold promise for optimizing CO2 storage strategies and enhancing subsurface CO2 storage capacity.