Delving into carbon dioxide frosting temperature in natural gas mixtures using hyperparameter optimized gradient boosting decision trees

Predicting carbon dioxide frosting temperature within natural gas mixture is critical to prevent costly and hazardous blockages in cryogenic processing facilities. The research develops an accurate and thermodynamically interpretable machine learning model to precisely forecast these phase transition points. A Gradient Boosting Decision Tree (GBDT) model was trained on a comprehensive dataset of 430 experimental points. Advanced hyperparameter tuning strategies were employed, among which Bayesian Probability Improvement (BPI) was identified as the most effective due to its superior balance between parameter exploration space and exploitation of region, preventing overfitting and enhancing model generalization. The resulting GBDT-BPI model demonstrated exceptional predictive accuracy (R 2 test 0.9903 and MSE test 1.75%). To ensure the model’s decisions were scientifically sound, Shapley Additive Explanations (SHAP) were used for interpretation. The SHAP analysis confirmed that the model’s predictions are directly rooted in physical thermodynamics; it correctly identified that increased carbon dioxide concentration and system pressure raise the frosting temperature, which is consistent with the principles of partial pressure and fugacity, while increased methane concentration lowers it, reflecting its role as a thermodynamic diluent. This work provides a robust, validated, and interpretable computational tool for improving the safety and efficiency of natural gas processing operations.