Abstract
Predicting hydrogen solubility in aqueous environments requires sophisticated modeling to decipher complex parameter interactions. By employing diverse machine learning algorithms, researchers developed advanced predictive frameworks capable of identifying intricate relationships among critical input parameters and hydrogen solubility. The methodology involved multiple algorithmic techniques, including supervised learning algorithms like artificial neural networks, regression approaches, support vector regression, ensemble methods, and gradient boosting algorithms, each designed to capture nuanced interactions between pressure, salinity, temperature, and hydrogen solubility. Data integrity was ensured using the Monte Carlo outlier detection algorithm to rigorously screen experimental datasets, guaranteeing reliable model training. Performance evaluation utilized comprehensive metrics and visualization techniques to assess predictive accuracy. Empirical results identified CatBoost, Random Forest, Gradient Boosting, and Support Vector Regression as top performers, with CatBoost achieving the greatest forecasting precision, evidenced by an R-squared value of 0.9756 for testing, together with mean squared error of 0.0012 (testing), and mean relative deviation percentages of 57.35% (testing). Sensitivity analyses revealed key insights into hydrogen solubility dynamics, showing positive correlations with temperature and pressure, and a nverse relationship with salinity. SHAP analysis further identified pressure and salinity as the most influential factors. These findings validate the efficacy of sophisticated machine learning techniques and enhance the scientific understanding of hydrogen solubility mechanisms in aqueous systems.