Predicting viscosity of CO2–CH4 binary mixtures using robust white-box machine learning frameworks: implication for carbon capture, utilization, and storage

Carbon capture, utilization, and storage (CCUS) technologies, particularly those involving pure and impure carbon dioxide (CO2) injection for enhanced oil recovery (EOR), are vital for mitigating greenhouse gas emissions while optimizing energy production. The viscosity of carbon dioxide-methane (CO2–CH4) binary systems plays a critical role in determining flow behavior, injectivity, and storage efficiency in subsurface formations. However, direct experimental measurements of viscosity are often costly, time-consuming, and constrained by operational limitations. Furthermore, existing predictive correlations frequently exhibit limited accuracy across wide ranges of pressure, temperature, and composition, hindering their application in practical CCUS and EOR scenarios. This study introduces a white-box machine learning framework based on multi-gene genetic programming (MGGP) to predict the viscosity of CO2–CH4 mixtures with enhanced precision. A comprehensive dataset comprising 742 experimental measurements was utilized to construct explicit mathematical correlations as functions of pressure, temperature, and CO2 mole fraction. Extensive statistical analyses and graphical validations confirmed the high fidelity of the developed models. The MGGP-based schemes achieved a low total RMSE of 2.6343 and an excellent R2 of 0.9942, outperforming four previously established models. Trend analyses and Shapley additive explanations (SHAP) further reinforced the model’s reliability, highlighting the dominant influence of pressure, followed by CO2 mole fraction and temperature, on viscosity behavior. The proposed explicit and user-friendly correlations, combining accuracy with interpretability, provide valuable tools for industrial applications, particularly in the simulation, design, and optimization of CCUS and CO2-EOR projects under a wide range of operating conditions.