Comprehensive Assessment of Supervised Machine Learning Models for Prediction of Oil Recovery Factor and NPV in Surfactant-Polymer Flooding: Bayesian Optimization and Stacking Ensembles

Surfactant-polymer (sp) flooding is recognized as an effective chemical enhanced oil recovery (eor) method, where accurate prediction of oil recovery factor (rf) and net present value (npv) is vital for field development planning and economic analysis. this study systematically evaluates a range of supervised machine learning algorithms—including catboost, artificial neural networks (ann), xgboost, lightgbm, and gradient boosting regressor (gbr)— for forecasting rf and npv based on experimental sp flooding data. baseline model results were established using default hyperparameters, followed by comprehensive two-stage hyperparameter tuning using grid search and bayesian optimization with optuna, along with five-fold cross-validation to ensure robustness. catboost and ann consistently achieved the highest predictive accuracy. in addition, ensemble stacking was then performed by combining top-performing models, further enhancing prediction reliability and generalization. additional post-processing using quantile adjustment (linear residual correction) addressed residual bias and improved calibration between predicted and observed values. furthermore, model performance was benchmarked using standard statistical metrics and comparative graphical analysis. also, the results demonstrate that integrating well-established supervised learning methods with systematic optimization, stacking, and output calibration offers a robust and practical framework for accurate prediction of sp flooding outcomes. moreover, this approach provides valuable support for data-driven decision- making in eor project design and evaluation. furthermore, the proposed framework achieved strong predictive accuracy in the all-stacking ensemble with cross-validation, yielding an r² of 0.978 and aapre of 2.71 for recovery factor, and an r² of 0.944 and aapre of 6.18 for net present value. ultimately, then applying quantile adjustment to the all-stacking ensemble, the performance remained competitive, with an r² of 0.964 and aapre of 3.61 for recovery factor, and an r² of 0.924 and aapre of 7.94 for net present value, further demonstrating the robustness of the approach.