Compositional machine learning frameworks for predicting mercury solubility in natural gas: Bridging predictive...

Mercury contamination in natural gas streams poses severe risks to both industrial infrastructure and the environment. Its presence, even in trace amounts, can lead to catalyst poisoning, equipment degradation, and the emission of toxic pollutants. As a result, accurate and cost-effective prediction of mercury solubility in natural gas components is essential for ensuring process safety, environmental compliance, and operational efficiency. This study presents a comprehensive machine learning framework based on compositional inputs to estimate mercury solubility across a wide range of pressures, temperatures, and multi-component mixtures. Among the developed models, including Cascade Forward Neural Network (CFNN), Adaptive Boosting Decision Tree (AdaBoost-DT, and Categorical Boosting (CatBoost), the CFNN model exhibited promising accuracy, achieving an R² of 0.9999 and an RMSE of 3.7063 ppb. The model's superiority stems from the integration of an expanded dataset incorporating new experimental measurements for intermediate hydrocarbons (C3-C5), as well as its capacity to generalize across both simple and multicomponent systems. Comparative analysis with conventional thermodynamic models further validated the CFNN's robustness and reliability. Additionally, explainable AI techniques, such as SHAP, verified the physical relevance of key features and preserved the underlying trends of mercury solubility behavior. The resulting predictive tool is not only scientifically sound but also industrially actionable. Its deployment in gas processing operations enables better decision-making and ease solvent selection, early risk mitigation, and optimization of mercury removal systems. Environmentally, the tool supports proactive strategies for limiting mercury emissions, aligning with stricter regulatory standards and sustainability goals. The convergence of high-performance modeling and practical implementation makes this approach a valuable step forward in cleaner, safer, and more efficient natural gas processing.