Thermodynamic Variable-Augmented Ensemble Machine Learning for Feature Selection and Interfacial Tension Prediction in Underground Storage Systems

Reliable prediction of interfacial tension (IFT) is critical for optimizing engineering processes like hydrogen separation, high-density energy storage and fluid displacement in underground oil reservoir systems. The goals of this study are: (a) to assess the efficacy of optimized ensemble machine learning (EML) models and (b) parametric sensitivity analysis to interpret feature importance while predicting interfacial tension (IFT) across an extensive range of thermodynamic conditions. The dataset comprises 498 experimentally measured IFT records collected under diverse conditions of temperature (T), pressure (P), and gas compositions of hydrogen (H2), carbon dioxide (CO2), and methane (CH4) in complex reservoir brine systems. The EML algorithms of extreme gradient boosting (XGB), random forest (RF), decision tree (DT), light gradient boosting (LGB) and adaptive boosting (ADB) are applied to construct the data-driven computational models to obtain IFT. The model’s effectiveness and robustness are evaluated using statistical performance indices. SHapley Additive exPlanations (SHAP) analysis is adopted to interpret and rank feature importance, thereby ensuring the model’s transparency and interpretability. The XGB model showed excellent performance, with high determination of coefficient (R2, 83%) and least root mean square error (RMSE, 4.81 mN/m) compared to other models of RF (81, 5.06), DT (79.94, 5.23), LGB (77.52, 5.54) and ADB (69.32, 6.46) accordingly. Based on SHAP scores, it is revealed that T is the most significant parameter to control IFT, and we ranked them higher to lower weighty order, like T, H2, P, CO2 and CH4. This study demonstrates the efficacy of EML and physics-guided computational models, coupled with feature selection techniques, in capturing IFT phenomena relevant to a subsurface hydrogen–brine geological system. It also establishes an interpretable artificial intelligent framework to uphold the project design, risk evaluation, and optimization of underground hydrogen storage systems in saline aquifers and other geologically complex formations for energy transition.