Abbas Albalola, Yanbu Aramco Sinopec Refining Company (YASREF), Saudi Arabia

This study develops and compares four machine learning regression models to predict Research Octane Number (RON) from specific gravity (S.G) and aromatic hydrocarbon content (v%) using a consolidated dataset of 108 unique Cooperative Fuel Research (CFR) engine measurements. The dataset spans RON 102.3-107.5, S.G 0.8377-0.8611, and Aromatics 81.48 89.96 V%, partitioned into training (75 samples, 70%), test (22 samples, 20%), and validation (11 samples, 10%) subsets. Four regression approaches were evaluated: (i) Linear regression using S.G and Aromatics as predictors; (ii) Polynomial regression incorporating derived quadratic and interaction terms (S.G², Aromatics², S.GxAromatics); (iii) Ridge regression with L2 regularization (alpha=0.001, selected by 5-fold cross-validation); and (iv) Random Forest regression (300 trees, maximum depth 4). Model performance was assessed using the coefficient of determination (R²), root-mean-square error (RMSE), and standard deviation of prediction bias, evaluated separately on training, test, and validation subsets to distinguish between fitting accuracy and true generalization. Linear and Polynomial models achieved the highest test-set R² (0.902 and 0.909, respectively), indicating superior generalization to unseen data. Pearson correlations between RON and predictors ranged from r=0.926 (S.G) to r=0.948 (S.GxAromatics), indicating strong positive relationships. Ridge regression (Test R²=0.901) provided comparable performance with enhanced coefficient stability, while Random Forest exhibited overfitting characteristics (Train R²=0.960, Test R²=0.889). Results demonstrate that Linear regression offers the optimal balance of interpretability, simplicity, and generalization performance for production deployment, with Polynomial regression as a viable alternative if marginal accuracy gains justify increased model complexity. Recommendations for real-time deployment include daily bias monitoring, extrapolation flagging, and periodic model recalibration.

Research Octane Number, Soft Sensor, Machine Learning, Linear Regression, Polynomial Regression, Ridge Regression, Random Forest, CFR Engine, Specific Gravity, Aromatic Content.