A Comprehensive AI-Model for H2-Brine Interfacial Tension Prediction: Implications for Underground Hydrogen Storage

Abstract The global focus on clean energy has made hydrogen a crucial part of the future energy framework. Among various large-scale hydrogen technologies, underground hydrogen storage (UHS) in geological formations, particularly saline aquifers, presents a strategic solution. However, the effective storage of hydrogen underground is influenced by fluid-fluid interactions, notably interfacial tension (IFT) between hydrogen and brine. Traditionally, IFT has been measured using experimental methods, such as the pendant drop technique, which are often limited by high costs and technical challenges. This study aims to develop accurate, data-driven machine learning (ML) models to predict H2–brine IFT under diverse operational conditions, incorporating factors such as temperature, pressure, salinity, and gas composition. To support this, a dataset of IFT measurements was compiled from various literature, covering a wide range of thermodynamic and chemical environments. The dataset included different gas compositions like H2, CO2, CH4, and multiple salt types including NaCl, CaCl2, and KCl. Data preprocessing included outlier removal, unit consistency checks, and Min–Max scaling. ML models using Random Forest (RF) and Artificial Neural Networks (ANNs) were developed due to their strengths in interpretability and pattern recognition. The dataset was split into training, testing, and validation sets for robust evaluation. A strategy was also implemented to generalize input parameters, simplifying the model while maintaining its predictive power. The results showed high prediction accuracy, with the RF model achieving a Mean Absolute Percentage Error (MAPE) of 1.88% on the training data and 3.42% on the testing data. The ANN model achieved MAPE values of 2.14% for training and 3.78% for testing. Temperature (R = –0.60) and CO2 concentration (R = –0.47) had the strongest correlations with IFT, while divalent salts like CaCl2 significantly influenced IFT (R ≈ 0.32–0.33). Generalized input transformations, such as NaCl-equivalent salinity and bulk gas properties, provided stronger correlations with IFT. These findings highlight the innovation of using physically meaningful input generalizations in ML-based IFT modeling, which promotes safer and more efficient underground hydrogen storage systems and supports the transition to a sustainable hydrogen economy.