Explainable machine learning frameworks for predicting CO2 deposition conditions in cryogenic natural gas processing systems

Accurate prediction of the deposition temperature of carbon dioxide (DTCD) in natural gas (NG) mixtures is essential for cryogenic gas processing. In this study, advanced machine learning models, including support vector machines (SVM), artificial neural networks (ANN), radial basis function neural networks (RBFNN), and gene expression programming (GEP), were developed to estimate DTCD using only pressure and gas composition as input features. A dataset of 430 experimental measurements was compiled from seven literature sources, encompassing a broad range of pressures and binary/ternary NG mixtures. The ANN model achieved the best performance, with a testing mean absolute percentage errors (MAPE) of 0.60 %, validated through five-fold cross-validation and violin-and-box plots. A transparent GEP-based correlation was also derived, offering a physically interpretable expression consistent with thermodynamic behavior. Trend analysis demonstrated the models’ abilities to capture expected physical patterns in DTCD behavior, and Williams’ plot analysis confirmed the applicability domain and the absence of influential outliers. Shapley additive explanations (SHAP)-based sensitivity analysis confirmed that CO 2 concentration and system pressure are the most influential variables. The proposed models generalize effectively, require minimal input features, and provide a reliable and explainable alternative to conventional equation of state (EoS) methods for cryogenic CO 2 capture system design. • Analysis of 430 experimental DTCD data for natural gas streams. • Design of advanced modeling tools based on ANFIS, GPM and DT methods. • Development of a mathematical correlation for DTCD based on GEP approach. • Simulation of the DTCD variations with respect to operational parameters. • Determination of fundamental factors influencing the DTCD in natural gas streams.