Application of Random Forest Algorithm for Viscosity Prediction in CO2-dissolved Crude Oil Systems

With the increasing application of CO2-EOR technology, accurate prediction of CO2-dissolved crude oil viscosity has become essential for flow assurance and gathering transportation. In this study, the coupled effects of temperature (20–40 ℃), pressure (0.5–3.5 MPa), and water cut (0–60%) on the viscosity behavior of CO2-saturated crude oil were systematically investigated through 64 laboratory experiments. Increasing temperature significantly reduced viscosity, whereas the pressure effect was comparatively weaker. As water cut increased, viscosity first increased and subsequently decreased sharply around the emulsion transition region.Among the evaluated empirical correlations, the Elsharkawy model showed comparatively better agreement with the experimental data and was therefore selected as the physical basis of the hybrid framework. To improve prediction accuracy under coupled operating conditions, a hybrid Elsharkawy–RF model was developed using Random Forest (RF)-based residual correction. For the independent testing dataset, the proposed model achieved an MAE of 1.3215 mPa·s, an RMSE of 1.8528 mPa·s, and an R² value of 0.9868. Feature importance analysis indicated that water cut was the dominant factor affecting viscosity variation, followed by temperature and pressure. The model provided accurate viscosity predictions within the tested temperature, pressure, and water-cut ranges. However, since the model was developed using a single crude-oil sample and a relatively limited dataset, further validation under broader reservoir and fluid conditions is still required.