Interpretable knowledge-guided machine learning for prediction wettability hydrogen-rocks/minerals-brine system in underground hydrogen storage project

Hung Vo Thanh,
Interpretable knowledge-guided machine learning for prediction wettability hydrogen-rocks/minerals-brine system in underground hydrogen storage project,
International Journal of Hydrogen Energy,
Volume 177,
2025,
151462,
ISSN 0360-3199,
https://doi.org/10.1016/j.ijhydene.2025.151462.
(https://www.sciencedirect.com/science/article/pii/S0360319925044647)
Abstract: Maximizing Underground Hydrogen Storage (UHS) depends on an awareness of the wettability behavior of rock/mineral-brine-hydrogen systems. This work presents an interpretable, knowledge-guided machine learning (ML) framework to accurately and physically consistently forecast hydrogen wettability. 623 experimental contact angle observations spanning several pressure, temperature, salinity, and substrate variables were used in a complete dataset. Two advanced ML models including Catboost and Histogram-based Gradient Boosting (HistGB) were trained and polished using Bayesian optimization in order to guarantee dependability. The proposed methodology combines physics-based constraints including capillary pressure, wettability index, and interfacial tension into the learning process to match predictions with basic surface science concepts unlike typical data-driven approaches. With a R2 of 0.997 and much reduced RMSE over conventional boosting techniques, results show that the knowledge-guided CatBoost model (KCatboost) beats regular ML models. Under SHAP-based feature importance analysis, substrate type is found to be the most important determinant of wettability; pressure and salinity follow in importance. Furthermore established to measure the fit of various geological formations for hydrogen storage was an Optimality Index (OI), which highlights mica and quartz-rich reservoirs as the most suitable. The suggested structure improves forecast accuracy and offers understandable understanding of subsurface wettability behavior, therefore enabling improved decision-making in hydrogen storage processes. These results help to build more strong, physics-based ML methods for uses in energy storage. Importantly, this is among the first studies to integrate capillary pressure, interfacial tension, and work of adhesion physics directly into boosting algorithms for UHS applications, establishing a new benchmark for robust and physically consistent machine learning approaches in subsurface energy storage.
Keywords: UHS; Catboost; HistGB; Physics informed machine learning; Optimal index; Wettability