Robust machine learning models for calculating the carbon dioxide desublimation point within natural gas mixtures at low temperature conditions

Abstract

Desublimation at low temperatures offers an efficient method for removing CO₂ from gas streams. Accurate prediction of the carbon dioxide desublimation temperature (CDDT) is essential for applying this method in natural gas processing. This investigation aimed to develop predictive tools utilizing machine learning approaches to estimate CDDT within natural gas mixtures. To reach this target, a large data set comprising 430 measurements obtained from published sources, was prepared. These data points cover the CDDT in binary and ternary natural gas mixtures under different pressures and gas fractions. In addition to black-box tools such as Decision Tree (DT), Gaussian Process Method (GPM) and Adaptive Neuro-Fuzzy Inference System (ANFIS) methods, a mathematical equation was developed via Genetic Programming (GP) technique for CDDT calculation. The performances of the designed models were rigorously evaluated through various visual inspections and statistical indices. While all models demonstrated excellent predictive accuracy, the GPM model provided superior results among black-box tools, exhibiting a mean absolute percentage error (MAPE) of 0.99 %. Furthermore, the GP equation achieved an overall MAPE of 0.65 % for the CDDT data. The intelligent models also performed well in predicting the data pertinent to both binary and ternary mixtures. A series of simulations based on the models’ outcomes were carried out to depict the CDDT variations in response to operational parameters, and the findings showed full consistency with previous experimental results. Ultimately, a sensitivity analysis was conducted to pinpoint the dominant factors affecting the CO₂ desublimation. Overall, the outcomes of this research enhance the understanding of CDDT behavior and provide valuable information for optimizing low-temperature CO₂ capture processes used in natural gas purification.