The unintended formation of solid carbon dioxide during the cryogenic processing of natural gas introduces severe operational hazards, pipeline blockages, and financial losses. To address this critical challenge, this study aims to develop a highly accurate, explainable data-driven framework capable of forecasting frosting temperatures based on operating pressures and mixture compositions. By utilizing a comprehensive dataset of 430 experimental data points comprising molar concentrations of methane, carbon dioxide, nitrogen, and ethane, ten distinct machine learning algorithms were trained and rigorously evaluated. The methodology specifically contrasted traditional linear techniques with advanced non-linear ensemble algorithms to identify the most robust predictive tool. Results clearly demonstrate the quantitative superiority of tree-based ensemble methods. Notably, the CatBoost algorithm emerged as the optimal model, achieving exceptional predictive accuracy with a coefficient of determination (R2) of 0.9918 and a Mean Relative Deviation (MRD) of only 0.55% on the unseen test set. To ensure physical reliability, SHapley Additive exPlanations (SHAP) were integrated, revealing that pressure and carbon dioxide concentration act as the primary positive drivers for frost formation, whereas methane concentration serves as the most significant mitigating factor. Ultimately, this research provides a novel, transparent, and highly deployable prognostic tool for chemical engineers. By successfully bridging the gap between high-accuracy machine learning and thermodynamic interpretability, this work establishes a trustworthy academic foundation for future predictive modeling and empowers industrial operators to proactively optimize natural gas purification, reduce energy expenditures, and ensure process safety.
Elham Kariri (Wed,) studied this question.