Predictive modeling of Oil–Water interfacial tension using biosurfactant parameters and machine learning approaches

Abstract

Oil–water interfacial tension (IFT) governs many industrial phenomena in petroleum recovery and environmental remediation, yet its experimental determination under complex biosurfactant–crude oil–water conditions is laborious and resource-intensive. This study aimed to develop a robust machine learning framework capable of accurately predicting IFT from physicochemical descriptors of biosurfactant and crude oil systems, thereby reducing empirical dependency through data-driven modeling. A dataset containing 1480 laboratory measurements was compiled from peer-reviewed sources and characterized by eight explanatory variables, Head charge, Molecular Weight, hydroxyl and carboxyl group counts (OH, COOH), biosurfactant concentration, oil API gravity, oil–air interfacial tension, and acid number. Models including Decision Tree, Random Forest, AdaBoost, Ensemble Learning, Support Vector Regression (SVR), Convolutional Neural Network (CNN), and Multilayer Perceptron (MLP-ANN) were optimized and evaluated using five-fold cross-validation with R², MSE, and AARE % metrics. Results evidenced superior performance for ensemble-based algorithms, particularly Random Forest (R² = 0.957, MSE = 0.46) and AdaBoost (R² = 0.978, MSE = 0.24), exhibiting high stability and minimal prediction error. SHAP (SHapley Additive exPlanations) analysis identified COOH and oil compositional variables (API and acid number) as the most influential predictors, aligning with theoretical expectations regarding polarity and molecular orientation at the oil–water interface. Overall findings demonstrate that properly tuned ensemble learning provides a physically interpretable, highly accurate surrogate to laboratory IFT measurement, revealing clear structural–functional dependencies across biosurfactant systems and supporting its broader integration into predictive material design and green petroleum applications.