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

Area of Science:

• Chemical Engineering
• Environmental Science
• Data Science

Background:

• Mercury contamination in natural gas presents significant risks to industrial equipment and the environment.
• Accurate prediction of mercury solubility is vital for process safety, environmental compliance, and operational efficiency in gas processing.

Purpose of the Study:

• To develop a comprehensive machine learning (ML) framework for estimating mercury solubility in natural gas components.
• To provide an accurate and cost-effective predictive tool for mercury levels across diverse conditions.

Main Methods:

• A machine learning framework utilizing compositional inputs was developed.
• Models evaluated included Cascade Forward Neural Network (CFNN), Adaptive Boosting Decision Tree (AdaBoost-DT), and Categorical Boosting (CatBoost).
• Explainable AI techniques (SHAP) were employed to verify feature relevance and model behavior.

Main Results:

• The CFNN model demonstrated exceptional accuracy with R² of 0.9999 and RMSE of 3.7063 ppb.
• The model's performance was enhanced by an expanded dataset including intermediate hydrocarbons (C3-C5).
• CFNN showed superior generalization across simple and multicomponent systems compared to conventional thermodynamic models.

Conclusions:

• The developed CFNN model offers a robust and reliable tool for predicting mercury solubility in natural gas.
• This ML framework supports improved decision-making in gas processing, including solvent selection and mercury removal optimization.
• The tool aids in mitigating environmental risks associated with mercury emissions, aligning with regulatory standards and sustainability goals.