A Novel Hybrid Machine Learning Model to Predict Viscosity of CO2 Mixtures with Impurities

Summary This work presents a two-stage hybrid physics machine learning (ML) framework for predicting the dynamic viscosity (µ) of CO2-rich mixtures containing impurities. A total of 1689 experimental datapoints with CO2 molar fractions above 0.5 are processed using six established µ models (Aasberg-Petersen (AP), Pedersen (Ped), Pedersen with CO2 as a reference fluid (CO2 Ped), Lohrenz bray Clark (LBC), Lennard Jones (LJ), and Residual Entropy Scaling (RES)), which serve as the physical baselines. Although these models generally provide reliable estimates, deviations can reach up to 40%. In the first stage, a Random Forest (RF) classifier selects the most accurate model for each datapoint by mapping the system's thermodynamic conditions to the model with the lowest deviation from experimental behavior. While certain models, such as CO2 Ped, consistently achieve high accuracy, the classifier enables other models to excel within their respective domains of validity, resulting in improved overall performance. In the second stage, an RF regression based corrective layer refines residual errors using absolute or relative error targets, further enhancing predictive accuracy. The hybrid framework is evaluated on both seen and unseen impurity compositions, demonstrating that classification alone achieves high accuracy, while the corrective regressor ensures robustness for models that may still exhibit limited accuracy. Notably, the Ped model benefited from this corrective layer, with substantial improvements in predictive performance for C2+ unseen impurities. This integrated methodology preserves the physical interpretability of µ models while significantly improving their predictive reliability for CO2-rich multicomponent systems.