Machine learning-assisted prediction of polymer flooding performance in heterogeneous carbonate reservoirs of the Pre-Caspian Basin: a gradient boosting and physics-informed approach

Abstract

Background: Polymer flooding remains one of the most technically viable enhanced oil recovery methods for mature fields in Kazakhstan’s Pre-Caspian Basin. Predicting its sweep efficiency in heterogeneous carbonate reservoirs is challenging due to complex pore-throat geometry, high-salinity formation waters, and pronounced permeability contrasts. Conventional reservoir simulation is computationally prohibitive for real-time operational decisions and probabilistic uncertainty analysis. Aim: This study develops and validates a hybrid surrogate framework integrating gradient boosting regression with physics-informed constraints derived from polymer transport theory to predict incremental oil recovery factor and injectivity ratio for polymer flooding in carbonate formations of the Pre-Caspian Basin. Materials and Methods: A dataset of 412 polymer flooding operations was compiled from published literature, Society of Petroleum Engineers (SPE) technical reports, and internal field data from three Kazakhstani assets: Tengizchevroil, Karachaganak Petroleum Operating, and the Uzen field operated by Ozenmunaigas. Eighteen reservoir and fluid parameters were used as input features. Gradient boosting regression was trained using a stratified 80/20 split with five-fold cross-validation. Physics-informed penalty terms derived from Darcy-scale polymer transport equations were embedded in the loss function to ensure physical consistency. Model performance was benchmarked against artificial neural networks, support vector regression, and full compositional simulation. Results: The physics-informed gradient boosting model achieved a coefficient of determination R² = 0.924 and root mean square error of 2.31% on the held-out test set for recovery factor prediction, outperforming artificial neural networks (R² = 0.891) and support vector regression (R² = 0.857). The physics penalty reduced physically inconsistent predictions by 78%. SHapley Additive exPlanations (SHAP) analysis identified the permeability variation coefficient, polymer concentration, and formation water salinity as dominant predictors. Surrogate estimates agreed with full compositional simulation within ±1.8% absolute for the Uzen carbonate pilot. Conclusions: The proposed surrogate model provides rapid and reliable predictions of polymer flooding performance suitable for field-scale integration, substantially reducing computational burden for enhanced oil recovery screening and optimization in heterogeneous carbonate reservoirs of Kazakhstan.