Modeling Gas–Brine Surface Tension Using Data-Driven Techniques for Underground Hydrogen Storage: A Focus on Depleted Gas Reservoirs

Hydrogen (H₂) is gaining momentum as a clean energy carrier, yet large-scale storage remains a major challenge. Storing hydrogen in depleted gas reservoirs is a promising option, leveraging existing infrastructure and proven containment reliability. A key factor influencing storage efficiency is the surface tension (ST) at the gas-brine interface, which controls capillary trapping and fluid flow. However, direct ST measurements are costly and time-consuming, motivating the use of predictive tools. This study applies machine learning (ML) algorithms to model ST between hydrogen-methane mixtures and brine under reservoir conditions. A data set of 1050 experimental measurements was used to train and test several ML algorithms. The models include an Adaptive Neuro-Fuzzy Inference System (ANFIS) and artificial neural networks (ANNs) optimized via Bayesian regularization. Model performance was evaluated using the coefficient of determination (<i>R</i> ²), mean squared error (MSE), and mean absolute percentage error (MAPE). Model-agnostic methods such as partial-dependence plots and permutation feature importance were used for interpretation. All models achieved strong predictive performance (<i>R</i> ² > 0.96), significantly better than the ST estimated by empirical correlations. The cascade-forward backpropagation neural network (CFBN) showed the highest training accuracy (<i>R</i> ² = 0.9919, MSE = 0.001, MAPE = 0.334), while the feedforward neural network (FNN) performed best in testing (<i>R</i> ² = 0.9849, MSE = 0.0019, MAPE = 0.286). ANFIS yielded the lowest MAPE in both training (0.0186) and testing (0.0200). Density difference was identified as the most influential feature. These results confirm that ML models can efficiently and accurately predict ST in underground hydrogen storage systems, reducing experimental demands and accelerating site evaluation.