Application of Machine Learning for Prediction of Crude Oil Properties

Accurate determination of crude oil properties, including API gravity, kinematic viscosity, sulfur content, and SARA fractions (Saturates, Aromatics, Resins, and Asphaltenes), is essential for optimal reservoir management, pipeline transportation, and refining processes. Traditional laboratory measurements based on ASTM standards are expensive, time-consuming, and require complex experimental setups. This study develops a robust, high-performance machine learning (ML) framework to predict key physical and chemical properties of crude oil utilizing easily measurable physical parameters: density (ρ۱۵), refractive index (nD), and distillation temperatures (T۱۰%, T۵۰%, T۹۰%). Five state-of-the-art ML algorithms — Support Vector Regression (SVR), Random Forest (RF), Multi-Layer Perceptron (MLP-ANN), Gradient Boosting Regressor (GBR), and Extreme Gradient Boosting (XGBoost) — were trained, optimized, and validated using a comprehensive dataset of ۱,۲۰۰ diverse global crude oil samples. The results demonstrate that the XGBoost model outperforms all other models, yielding exceptional prediction accuracy for API gravity (R۲ = ۰.۹۹۱, RMSE = ۰.۷۳), kinematic viscosity at ۴۰ °C (R۲ = ۰.۹۸۲, RMSE = ۱۱.۴۰ cSt), sulfur content (R۲ = ۰.۹۵۸), and asphaltene content (R۲ = ۰.۹۴۲). Feature importance analysis utilizing SHAP (SHapley Additive exPlanations) values revealed that density and refractive index are the most critical predictors. The proposed ML models can serve as highly accurate virtual sensors, offering real-time characterization in upstream and downstream petroleum industries

کلیدواژه ها:

Machine Learning ، Crude Oil Properties ، XGBoost ، Viscosity ، API Gravity ، Predictive Modeling