Machine Learning-Based Assessment of Impurity Effects on Thermophysical Properties in CO2-Rich Mixtures for CCUS Applications

Summary This study presents an integrated statistical-machine learning (ML) framework to quantify and interpret the impact of impurities on the thermophysical properties of CO2-rich mixtures relevant to carbon, capture, utilization, and storage (CCUS) applications. The framework combines classical statistical metrics, such as correlations, ANOVA F-statistics, and Kolmogorov-Smirnov tests, with ML interpretability tools, such as feature importance, permutation analysis, partial dependence, and SHAP values. This dual approach enables a comprehensive examination of both linear and nonlinear dependencies between impurity composition, operating conditions, and transport properties. The analysis revealed that most impurities decrease CO2 mixture density and viscosity, while increasing the saturation pressure (Psat). ML interpretation further demonstrated that impurity interactions are nonlinear and can be modulated by phase state and thermodynamic conditions. The results underscore that ML methodologies can reproduce established physical trends and uncover hidden dependencies traditional analyses can overlook. Overall, the developed framework enhances the understanding of impurity-driven thermodynamic behavior and provides a foundation to ensure risk-free CO2 transport in CCUS systems.