Prediction of Gasoline Blend Ignition Characteristics Using Machine Learning Models

Research Octane Number (RON), among other autoignition related properties, is a primary indicator of the grade of spark-ignition (SI) fuels. However, in many cases, the blending of various gasoline components affects the RON of the final fuel product in a nonlinear way. Currently, the lack of precise predictive models for RON challenges the accurate blending and production of commercial SI fuels. This study compares popular Machine Learning (ML) algorithms and evaluates their potential to develop state-of-the-art models able to predict key SI fuel properties. Typical gasoline composition was simplified and represented by a palette of seven characteristic molecules, including five hydrocarbons and two oxygenated species. Ordinary Least Squares (OLS), Nearest Neighbors (NN), Support Vector Machines (SVM), Decision Trees (DT), and Random Forest (RF) algorithms were trained, cross-validated, and tested using a database containing 243 gasoline-like fuel blends with known RON. Best results were obtained with nonlinear SVM algorithms able to reproduce synergistic and antagonistic molecular interactions. The Mean Absolute Error (MAE) on the test set was equal to 0.9, and the estimator maintained its accuracy when alterations were performed on the training data set. Linear methods performed better using molar compositions while predictions on a volumetric basis required nonlinear algorithms for satisfactory accuracy. Developed models allow one to quantify the nonlinear blending behavior of different hydrocarbons and oxygenates accounting for those effects during fuel blending and production. Moreover, these models contribute to a deeper understanding of the phenomena that will facilitate the introduction of alternative gasoline recipes and components.