Physics informed machine learning for prediction hydrogen solubility in aqueous solution to implication for underground hydrogen storage formations

Subsurface formations for hydrogen storage represent a viable possibility for facilitating the transition to a sustainable hydrogen-centric energy economy. Precisely forecasting hydrogen solubility under diverse conditions is essential for the development of efficient and secure storage systems. This research presents an innovative Physics-Informed Machine Learning (PIML) framework employing the Physics-Informed Extreme Gradient Boosting (PI-XGBoost) model to forecast hydrogen solubility in aqueous solutions. The model attains great predicted accuracy by incorporating essential physical factors such as Henry's Law, the Setchenow effect, and gas compressibility, while remaining consistent with established theoretical limits. The PI-XGBoost model exhibited enhanced performance, attaining a coefficient of determination (R2 > 0.98) and markedly reduced mean squared error (MSE) relative to conventional machine learning methods. The model's reliability was validated against physical laws, with predicted-to-expected ratios nearly matching theoretical expectations. Scenario evaluations indicated that the best hydrogen solubility occurs under high-pressure and low-salinity settings, but salinity-induced decreases in solubility pose issues in saline aquifers. Shapley Additive Explanations (SHAP) analysis enhanced interpretability by identifying critical factors that influence predictions. The results underscore the model's capacity to inform reservoir selection, enhance operating efficiency, and mitigate risks in hydrogen storage systems. This study highlights the significance of physics-informed machine learning in enhancing energy storage technologies, connecting theoretical knowledge with practical implementations, and facilitating global decarbonization initiatives.