Improved explicit data-driven frameworks for accurate prediction of carbon dioxide diffusivity in high viscosity hydrocarbon systems

Precise estimation of CO 2 diffusivity in hydrocarbon mixtures is fundamental for designing effective injection schemes, controlling displacement efficiency, and optimizing the overall performance of CO 2 -enhanced oil recovery processes. This study introduces a novel approach, integrating genetic programming with explainable AI, to accurately predict CO 2 diffusivity in viscous hydrocarbon systems. Utilizing an experimental dataset of 260 samples from the literature, explicit data-driven predictive correlations were developed using the well-established genetic programming (GP) technique. The results confirmed the strong performance of the developed GP-based frameworks, with a coefficient of determination of 0.995 and a root mean square error of 0.17 × 10-10 m/s 2 . Compared to some existing correlations, this approach offers improved accuracy and reduced computational demands. Data quality was confirmed via leverage analysis, and model reliability was established. Shapley plot-based sensitivity analysis revealed pressure as the primary influence on carbon dioxide diffusivity, followed by temperature and carbon dioxide mass fraction, with fluid density having minimal impact. The model's explicit formulation facilitates real-world deployment. Furthermore, Shapley additive explanations enhance interpretability and validate feature importance, making the model user-friendly for carbon dioxide injection applications. Physical validation via trend analysis confirms the paradigm's ability to maintain physical relationships related to independent variable variations. The implemented model can contribute to the improved management and decision-making processes for CO 2 -based enhanced oil recovery (EOR) projects, which have the potential to enhance oil extraction and facilitate carbon dioxide capture and storage. • The CO 2 diffusivity in highly viscous systems was modeled using explainable artificial intelligence. • The GP scheme exhibited the best consistency with an RMSE of 0.17 × 10 -10 m/s 2 . • The physical validity and interpretability of the GP was testified using trend analysis, and SHAP techniques. • The GP model displayed excellent reliability for various scenarios related to CO 2 -based enhanced oil recovery.