Accurate determination of the viscosity of carbon dioxide (CO 2 ) mixed with nitrogen (N 2 ) is vital for enhanced oil recovery (EOR) and carbon capture, utilisation and storage (CCUS/CCS). The determination of this important thermophysical property is usually through costly and time‐consuming experiments, which is not ideal for field recovery planning and rapid decision‐making. On the other hand, the conventional modelling relies largely on equations of state (EoS) and empirical correlations, which can be inaccurate for CO 2 –N 2 viscosity, particularly near supercritical conditions due to simplifying assumptions and limited transferability. Consequently, machine‐learning (ML) methods have gained popularity for fast and accurate prediction. Hence, in this study, ~3036 literature data points spanning pressures of 0.00127–160.99 MPa and temperature of 66.55–575.15 K were collected, cleaned and pre‐processed. Then, using pre‐processed data, several ML models, including gradient boosting (GB), extreme gradient boosting (XGBoost), LightGBM, CatBoost, random forest, three multilayer perceptron artificial neural networks (MLP‐ANNs), a stacking ensemble and a group method of data handling (GMDH) were developed. The developed models were benchmarked to predict CO 2 –N 2 viscosity as a function of temperature, pressure and the mole fractions of CO 2 –N 2 in the mixture. The analysis of the results indicate that the GB achieved the best performance with a correlation coefficient ( R 2 ) of 0.9933 ± 0.0011, root mean square error (RMSE) of 4.83 ± 0.39 μPa·s and mean absolute error (MAE) of 2.34 ± 0.10 μPa·s (mean ± 95% CI) for the test dataset, outperforming all other ML models and the utilised literature correlations. In addition, based on GMDH, two practical explicit equations within temperature ranges of T 300 K that predict the experimental viscosity with high accuracy were proposed. The sensitivity analysis also shows that the pressure has the highest positive impact, while temperature exhibited a comparably strong negative effect on viscosity.
Alwerfali et al. (Thu,) studied this question.