• Developing a robust data-driven models to predict the viscosity of nano-polymer composites used in enhanced oil recovery (EOR) • Multiple machine learning models were utilized. • Shear rate emerged as the primary element affecting viscosity of composites. • AdaBoost and Random Forest offered high accuracy and scalability for predicting viscosity of such polymeric nanocomposites in EOR systems. The viscosity of nano–polymer composites used in enhanced oil recovery (EOR) is governed by complex and highly nonlinear interactions that limit the applicability of conventional predictive correlations. To address this challenge, this study develops robust data-driven models based on a comprehensive dataset of 1447 experimental points. The dataset spans temperatures from 25 to 90 °C, polymer concentrations of 0.1–15 wt%, nanoparticle concentrations of 0.1–1 wt%, and shear rates between 9.75 and 1021.38 s −1 . Input parameters include temperature, polymer and nanoparticle concentrations, shear rate, and the types of polymers (guar gum, gum arabic, xanthan gum) and nanoparticles (alumina, copper, silica). In addition to viscosity enhancement, nanoparticle incorporation is motivated by their role as sacrificial agents that can reduce polymer adsorption onto rock surfaces and modify wettability in porous media, thereby influencing the effective rheological behavior of polymer solutions in EOR processes. The viscosity values ranged from 1.30 to 2160 cP. The dataset explicitly includes polymer systems both in the absence and presence of nanoparticles, enabling the models to capture viscosity variations before and after nanoparticle incorporation. The results demonstrate that nanoparticle addition generally enhances viscosity through particle–polymer interactions, although the magnitude of this enhancement is strongly dependent on shear rate, polymer concentration, and nanoparticle type. Multiple machine learning models were utilized, including Decision Tree, Random Forest, Adaptive Boosting (AdaBoost), K-Nearest Neighbors (KNN), Convolutional Neural Networks (CNN), and Ensemble Learning. SHAP analysis recognized shear rate as primary element affecting viscosity, while nanoparticle type had the least impact. A correlation-based sensitivity analysis was used to verify well-established rheological trends, including the direct influence of polymer and nanoparticle concentrations and the inverse influence of shear rate and temperature on viscosity. The Leverage method was used to assess data quality, identifying 18 outliers, though all points were retained for modeling. Models were trained and validated using 5-fold cross-validation. AdaBoost (50 estimators) and Random Forest (depth 9) outperformed others, with R 2 values of 0.999254 and 0.998399, and low MSE and AARE% values. CNN showed the weakest performance. Ensemble Learning further improved accuracy (R 2 = 0.999191, MSE = 74.49). Cross-plots and error distribution analyses confirmed the reliability and robustness of the top models. These findings demonstrate that ensemble learning methods especially AdaBoost and Random Forest offer high accuracy and scalability for predicting viscosity in EOR systems. The developed models provide efficient screening and decision-support capabilities for viscosity estimation and nano–polymer composite optimization under laboratory-scale conditions relevant to enhanced oil recovery applications.
Madani et al. (Sun,) studied this question.