Advanced generalized machine learning models for predicting hydrogen-brine interfacial tension in underground hydrogen storage systems.

The global transition to clean energy has highlighted hydrogen (H) as a sustainable fuel, with underground hydrogen storage (UHS) in geological formations emerging as a key solution. Accurately predicting fluid interactions, particularly interfacial tension (IFT), is critical for ensuring reservoir integrity and storage security in UHS. IFT is key in fluid behavior, influencing structural and residual trapping capacities. However, measuring IFT for H-brine systems is challenging due to H's volatility and the complexity of reservoir conditions. This study applies machine learning (ML) techniques to predict IFT between H and brine across various salt types, concentrations, and gas compositions. A dataset was used with variables such as temperature, pressure, brine salinity, and gas composition (H, CH, CO). Several ML models, including Random Forests (RF), Gradient Boosting Regressor (GBR), Extreme Gradient Boosting Regressor (XGBoost), Artificial Neural Networks (ANN), Decision Trees (DT), and Linear Regression (LR), were trained and evaluated. RF, GBR, and XGBoost achieved R values over 0.99 in training, 0.97 in testing, and all exceeded 0.975 in validation. These top models achieved RMSE values below 1.3 mN/m and MAPE values under 1.5%, confirming their high predictive accuracy. Residual frequency analysis and APRE results further confirmed these ensemble models' low bias and high reliability, with error distributions centered near zero. DT performed slightly lower, with R values of 0.93, while LR struggled to model the non-linear behavior of IFT. A novel salt equivalency metric was introduced, transforming multiple salt variables into a single parameter and improving model generalization while maintaining high prediction accuracy (R = 0.98). Sensitivity analysis and SHAP (Shapley Additive Explanations) analysis revealed temperature as the dominant factor influencing IFT, followed by CO concentration and pressure, while divalent salts (CaCl, MgCl) exhibited a stronger impact than monovalent salts (NaCl, KCl). This study optimizes hydrogen storage by offering a generalized, high-accuracy ML model that captures nonlinear fluid interactions in H-brine systems. Integrating real-world experimental data with ML-driven insights enhances reservoir simulation accuracy, improves hydrogen injection strategies, and supports the global transition toward sustainable energy storage solutions.