A robust machine learning framework for predicting contact angle in nano-assisted chemical EOR

Wettability alteration through nano-assisted chemical Enhanced Oil Recovery (nano-cEOR) critically governs capillary pressure and multiphase flow in porous media. Contact angle (CA) quantifies wettability, yet experimental measurement remains costly, inconsistent, and operationally impractical, creating nanofluid screening bottlenecks. Existing machine learning (ML) models lack nanoparticle (NP)-specific descriptors essential for nano-cEOR physics. This study introduces a comprehensive ML framework predicting CA from 418 experimental data points, integrating NP-specific descriptors (type, concentration, size, functionalization), fluid properties, rock mineralogy, and reservoir conditions. Six algorithms were benchmarked across independent training, testing, and validation datasets with hyperparameter optimization and tenfold nested cross-validation. Multi-method sensitivity analysis integrating Sobol indices, SHAP values, and partial dependence plots identified operational thresholds and lithology-specific strategies. XGBoost Regressor (XGBR) achieved superior validation performance: R² = 0.953, RMSE = 9.24°, MAE = 5.87°, and near-zero prediction bias (- 0.04°). Sensitivity analysis identified a minimum permeability of 0.1 mD for nano-cEOR viability, an optimal salinity window of 30,000-80,000 ppm, and synergistic NP-to-chemical ratios of 1:1 to 1.5:1. Rock-type analysis revealed lithology-dependent formulation inversion, with optimal NPs shifting between carbonates (ZrO₂, TiO₂) and sandstones (Fe₃O₄, CuO). These findings advance nanofluid screening capability, support reservoir-specific formulation design, and provide decision-support guidelines for field-scale nano-cEOR deployment.