Black-box and white-box machine learning tools to estimate the frost formation condition during cryogenic CO2 capture from natural gas blends

Cryogenic carbon capture (CCC), a promising technology for the sequestration of CO 2 from natural gas streams, relies on the CO 2 frost formation at cryogenic temperatures. Precise estimation of the CO 2 frost formation temperature (FFT) is vital for optimizing CCC processes and maximizing efficiency. This study deals with the development of machine learning models for predicting the FFT of CO 2 in natural gas blends. A comprehensive dataset, including 430 experimental samples, was assembled from the literature. The foregoing dataset included the FFT data for binary and ternary natural gas blends across a wide range of pressures and component fractions. Three distinct black-box algorithms, including Regression Tree (RT), Radial Basis Function Neural Network (RBF-NN) and Support Vector Machine (SVM) were employed to model FFT. The performance of each model was rigorously explored through diverse statistical and graphical methods. While all developed models demonstrated high accuracy, the RBF-NN model was the superior predictive tool, achieving a mean absolute percentage error (MAPE) of 0.82 % and a standard deviation (SD) of 1.19 % during the validation phase. Also, an explicit correlation for FFT was proposed using the white-box machine learning technique of Gene Expression Programming (GEP), which achieved the MAPE of 0.59 % for all data. The models were also capable of predicting the FFT of CO 2 in both binary and ternary blends, and favorably captured the complicated physical variations of FFT under diverse operating conditions. To gain deeper insights into the most fundamental factors in controlling the FFT of CO 2 , a sensitivity analysis was conducted. The findings of the current study, in turn, contribute to the understanding of FFT of CO 2 and the optimal design of the CCC processes in natural gas industries. • 430 experimental data for FFT of CO 2 in natural gas blends were analyzed. • Three Black-box machine learning tools were employed for modeling. • A simple FFT correlation was also proposed based on the intelligent method of GEP. • The models properly described the physical variations of FFT. • The most affecting factors on FFT were determined through a sensitivity analysis.