Interfacial tension modeling in CO2-brine systems using machine learning for enhanced CO2 storage

This study introduces an advanced machine learning (ML) framework to predict interfacial tension (IFT) in CO 2 -brine systems, a key factor in optimizing carbon capture and storage (CCS) processes. Accurate IFT predictions are critical for enhancing CO 2 trapping mechanisms, including structural, residual, solubility, and mineral trapping. Using a dataset of 1255 experimental IFT measurements, comprehensive preprocessing steps—such as outlier detection and feature standardization—were applied to improve data quality. Six ML models, including gradient boosting, extra trees, categorical boosting (CatBoost), random forest, extreme gradient boosting (XGB), and light gradient boosting machine (LGBM), were developed and rigorously evaluated. Among these, CatBoost demonstrated superior performance with an R 2 of 0.986 and a mean absolute percentage error (MAPE) of 2.349%. A stacking ensemble methodology was employed to enhance predictive accuracy further, integrating base models using Lasso regression. This approach achieved performance metrics: R 2 = 0.988, RMSE = 1.158 mN/m, and MAPE = 2.202%. Feature importance analysis using SHapley Additive exPlanations (SHAP) identified density difference (DD), pressure ( P ), and temperature ( T ) as the most influential features governing IFT. An analytical expression derived from the stacking model provides interpretable insights into the nonlinear relationships between features and IFT. This robust framework minimizes reliance on time-consuming experimental measurements and accelerates CCS project workflows by delivering reliable IFT predictions under diverse conditions. These findings underscore the framework’s potential to advance CCS optimization and contribute to climate change mitigation efforts.