Machine Learning Approach for Modeling Surface Tension Between H2-CH4 Gas Mixture and Brine: Implication for Underground Hydrogen Storage in Partially Depleted Gas Reservoirs

Hydrogen (H2) has emerged as a pivotal component in the transition toward sustainable energy systems. However, large-scale deployment is constrained by the lack of sufficient storage infrastructure. Underground hydrogen storage (UHS) in depleted gas reservoirs is considered a promising solution due to its cost-effectiveness and scalability. Accurate estimation of interfacial properties, particularly surface tension (ST), is essential for assessing storage capacity and predicting fluid behavior in porous media. This study explores the use of machine learning (ML) models to predict ST between H2/CH4 gas mixtures and brine under reservoir-relevant conditions. A dataset of experimentally measured ST values was used, incorporating input variables such as salinity, pressure, temperature, and density difference between gas and brine phases. Three artificial neural network (ANN) architectures—feedforward neural network (FFNN), legacy feedforward neural network (LFNN), and pattern recognition neural network (PRNN)—were first evaluated through an initial screening phase. Based on performance metrics, FFNN was selected as the most accurate model and was subsequently used for detailed sensitivity analysis. Two optimization algorithms, Levenberg–Marquardt (LM) and Bayesian Regularization (BR), were applied in the extended analysis. Additionally, the effect of the number of neurons and training epochs on model performance was assessed. Model accuracy was evaluated using R2, mean squared error (MSE), and average absolute percentage error (AAPE), with training dynamics further analyzed using loss function plots. The LM-optimized models consistently outperformed BR configurations, achieving the highest R2 (0.986) during training and maintaining strong generalization in testing (R2 = 0.982). BR models demonstrated robust performance, with all testing R2 values exceeding 0.965. Among the tested cases, LM-20-100 (i.e., FFNN with LM optimization, 20 neurons, and 100 training epochs) offered the best balance of accuracy and computational efficiency. Loss function analysis revealed faster and smoother convergence in LM models, while BR models exhibited more gradual, multi-phase training behavior. This study highlights the capability of ML, specifically FBNNs with LM and BR optimization, to accurately predict ST in complex gas–brine systems relevant to UHS. The results contribute to improving ST estimation under subsurface conditions, enhancing the screening and design of UHS operations in depleted gas reservoirs.