Rigorous hybrid machine learning approaches for interfacial tension modeling in brine-hydrogen/cushion gas systems: Implication for hydrogen geo-storage in the presence of cushion gas

In light of the environmental consequences linked to the burning of fossil fuels, there is a mounting push to devise cleaner means of energy generation. The production of hydrogen is gaining momentum as a favored avenue for enabling the shift toward sustainable energy. Effective hydrogen storage is crucial for sustainable energy solutions, and the inclusion of cushion gas in geological storage could aid in maintaining formation pressure during hydrogen reproduction while expanding pore volumes for gas by preventing water presence in pores. Accurately controlling hydrogen behavior and interactions in porous media is crucial for successful storage, with interfacial tension between gas and water playing a significant role in the trapping and mobilizing the phases. Thus, developing a predictive tool to simulate interfacial tension measurements is essential. This study employs various intelligent modeling techniques, including Adaptive Neuro-Fuzzy Inference System, Multilayer Perceptron optimized with Bayesian Regularization, Levenberg-Marquardt, and Scaled Conjugate Gradient algorithm, Grey Wolf Optimizer-based Least Squares Boosting (GWO-LSBOOST), GWO-based Radial Basis Function, GWO-based Least Squares Support Vector Machine, and Extreme Learning Machine (ELM), to develop interfacial tension models for brine-hydrogen/cushion gas systems. Applying these models developed based on diverse theories, structures, and performance characteristics helps identify the optimal predictor for the target parameter. The evaluation of the models was carried out using 2868 experimental data points. The achieved results demonstrated the satisfactory performance of different modeling techniques (with R2 falling within the range of 0.8979 to 0.9960), however, GWO-LSBOOST outperformed others (improving R2 from 0.8979 to 0.9960 and AARE% from 5.5714 % to 0.8060 %, compared to ELM). Leverage outlier identification and trend estimation analysis confirmed the statistical reliability of the dataset and the GWO-LSBOOST model. The formulated strategies offer a promising framework applicable for incorporating the gas-water interfacial tension in simulating various stages of gas injection/ reproduction in underground hydrogen storage.