Explainable advanced modelling of interfacial tension in H₂ - CO₂ - CH₄ - brine systems for sustainable subsurface...

Hydrogen is increasingly recognized as a pillar of future low-carbon energy systems, offering flexibility across multiple sectors and contributing to deep decarbonization goals. To enable large-scale deployment, Underground Hydrogen Storage (UHS) in geological formations such as saline aquifers is gaining considerable attention as a strategic option for ensuring secure and scalable energy reserves. The safety, efficiency, and operational success of UHS heavily depend on accurately determining interfacial tension (IFT) between hydrogen/cushion gas and the in-situ brine, as this parameter governs several critical phenomena related to storage performance. This study aims to develop a reliable and explainable machine learning framework to predict IFT in H₂/cushion gas-brine systems across diverse thermodynamic and compositional conditions. Two input schemes were explored: the first included nine variables covering detailed salt composition, pressure, temperature, and the presence of cushion gas (represented by average critical temperature, T_cm); the second used a simplified set of four inputs, namely equivalent salinity, pressure, temperature, and T_cm. A curated database of approximately 500 experimentally measured IFT data points was employed to train and validate three models: radial basis function neural network (RBFNN), generalized regression neural network (GRNN), and categorical boosting (CatBoost). Among these, the CatBoost-based model under the simplified scheme achieved the highest predictive accuracy (R² = 0.9979, RMSE = 0.5136 mN/m), with excellent physical consistency and generalization. Trend analysis confirmed the model's ability to replicate expected IFT behavior under varying conditions. Comparative benchmarking showed that the CatBoost model outperformed several state-of-the-art models from the literature. A leverage-based applicability domain analysis revealed that 97 % of the data points fall within acceptable influence and residual bounds, confirming robustness. Beyond accuracy, the model is highly interpretable globally and locally using SHAP and LIME techniques, making it suitable for real-time deployment in UHS operations to support safe injection design and long-term storage reliability.