Machine learning-based interfacial tension prediction for nanoparticles-assisted enhanced oil recovery

Abstract

Accurate interfacial tension (IFT) prediction is crucial for optimizing enhanced oil recovery (EOR), as traditional measurement methods are time-consuming, costly, and often impractical under reservoir conditions. This study presents a novel machine learning (ML) framework for robust IFT prediction. A comprehensive dataset of 474 data points was curated from published literature, uniquely encompassing a wide variability in reservoir conditions and complex chemical formulations, including 10 distinct nanoparticle (NP) types, 4 surfactant classes, and various polymers. Multiple ML models were developed, including Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGBoost), Random Forest (RF), and novel hybrid ensembles (ANN-RF and ANN-XGBoost). The results demonstrated that hybrid and deep learning approaches significantly outperformed simpler models. The hybrid ANN-RF model achieved superior validation accuracy (R 2 = 0.973, RMSE = 1.730), while the standalone ANN (Test R 2 = 0.942) and XGBoost (Test R 2 = 0.940) also showed strong generalization. A comprehensive sensitivity analysis identified chemical additive concentration as the dominant predictive factor, followed by NP concentration. Crucially, it revealed that salinity’s significant impact is driven primarily by high-order interaction effects. This work establishes a powerful, validated data-driven tool, highlighting the potential of ML to reduce experimentation and accelerate the design of optimized chemical EOR formulations. The novelty lies in combining hybrid models with comprehensive sensitivity analysis to deconstruct the complex, interaction-driven physics of NP-assisted EOR systems.