Data-Driven Prediction of CO۲ MMP for Reservoir Screening & CO۲-EOR under Pure and Mixed-Gas Conditions

Accurate prediction of minimum miscibility pressure (MMP) is crucial for efficient CO۲-enhanced oil recovery (CO۲-EOR) design. Traditional methods, including laboratory slim-tube experiments and empirical correlations, present limitations in cost and applicability. This study introduces a physics-consistent, data-driven workflow utilizing a curated experimental dataset to train and validate a Random Forest regressor. Rigorous cross-validation and a hold-out test set were employed to ensure model robustness. The resulting model achieved a test root mean squared error (RMSE) of ۴.۴۹ MPa and a coefficient of determination (R۲) of ۰.۷۳۸, significantly outperforming linear and baseline models. Feature importance analysis, utilizing permutation importance and partial dependence plots, confirmed physically plausible trends governing MMP behavior. Notably, on a pure-CO۲ benchmark subset (n=۳۱), the model demonstrated reduced error compared to the widely used Yelling-Metcalfe correlation (RMSE ۴.۵۰ vs. ۵.۷۱ MPa). This workflow provides a rapid and reliable screening tool for preliminary feasibility assessments and pressure targeting in CO۲-EOR projects.

کلیدواژه ها:

CO۲-EOR ، MMP ، machine learning ، gas composition ، reservoir engineering