Explainable Advanced Machine Learning Paradigms for Predicting Wettability of Hydrogen Mixtures and the Implications for Underground Hydrogen Storage

Abstract

Abstract Hydrogen has emerged as a versatile, carbon-zero energy carrier and industrial feedstock, with projections indicating it will contribute approximately 12% of the global energy mix by 2050. Large-scale underground storage and a comprehensive understanding of hydrogen behavior, such as wettability in porous media, are essential to meet such demand, ensure sustainability, and minimize hydrogen loss and contamination. However, current experimental techniques for determining wettability are costly, time-consuming, and labor-intensive. Nevertheless, the extensive data support the deployment of machine learning for time-efficient and cost-effective wettability estimation. This study integrates Bayesian Optimization with advanced machine learning paradigms: Extremely Randomized Trees (ET), Light Gradient Boosting Machine, and Multilayer Perceptron to predict hydrogen wettability, considering 1650 data points extracted from published peer-reviewed articles in reputable journals. The predictors include pressure (3.45–20.68 MPa), temperature (303–343K), brine salinity (0.34–3.42 mol/kg), density difference between brine and gas mixture (835–1087 kg/m3), and gas mole fractions (5–70 mol%). The dataset was partitioned into 70% for training and 30% for testing. Model performance was evaluated using statistical indices. Thereafter, the outstanding paradigm was juxtaposed against existing models and deployed to compute the capillary entry pressure, column height, and storage capacity of a typical repository. The Leverage method was then applied to validate data integrity and model reliability, while explainability was ascertained via permutation importance and Shapley Additive Explanations (SHAP). The Bayesian-optimized paradigms demonstrated considerable predictive capabilities, registering root mean square error (RMSE) and mean absolute percentage error (MAPE) below 2.6° and 5.7%, respectively. ET recorded the highest overall coefficient of determination (R2) of 0.881 and the least RMSE of 1.315° and MAPE of 2.442% and ranked ahead of existing paradigms (prediction errors between 3 and 8°). 99.58% of the dataset was categorized as valid data, while only 0.42% were vertical suspect and good high-leverage instances, signifying data authenticity and model reliability. From permutation importance and SHAP, nitrogen mole fraction and density difference showed the most significant and the least influence on wettability, respectively. Furthermore, the analysis revealed that increasing nitrogen fraction, pressure, and salinity favor hydrogen wettability and withdrawal. Rising temperature and density difference minimize the effect, while the methane fraction showed a non-monotonic effect. Most importantly, ET estimates for capillary entry pressure (11.0–11.7 MPa), column height (1135–1196 m), and storage capacity (298–1795 kg H2/m2) of a repository in the Saudi Arabian Cenozoic Harrat Uwayridh field were closer to the field results, highlighting generalizability and practical relevance. The developed paradigms enable accurate and efficient prediction of wettability, complementing experimental and simulation efforts. Their application can also enhance hydrogen loss mitigation and geo-storage optimization, with potential for adaptation in related fields to improve predictive performance and support data-driven decisions.