Machine learning regression based group contribution method for cetane and octane numbers prediction of pure fuel compounds and mixtures. (15th November 2020)

• Record Type: Journal Article
• Authors: Li, Runzhao; Herreros, Jose Martin; Tsolakis, Athanasios; Yang, Wenzhao
• Is Part Of: Fuel. Volume 280(2020)
• Publication Date: 2020-11-15
• Subjects: Fuel molecular structure -- Group contribution method -- Machine learning regression -- CN/RON/MON prediction -- Pure fuel compounds & mixtures; Fuel -- Periodicals; Coal -- Periodicals; Coal; Fuel; Periodicals; 662.6
• DOI: 10.1016/j.fuel.2020.118589
• Languages: English
• ISSNs: 0016-2361

Abstract

Graphical abstract: Schematic diagram of fuel ignition quality prediction by machine learning regression based group contribution method. Highlights: CN/RON/MON prediction of pure compounds and mixtures by machine learning based group contribution method. Group contribution method extracts the structural features and transforms into molecular structure matrix. Machine learning regression model correlates the molecular structure matrix and ignition quality matrix. A comprehensive fuel ignition quality database is developed for regression model training and validation. Abstract: Current methods to predict fuel ignition quality usually focus on either cetane numbers or research/motor octane numbers (CN, RON, MON) and most of them apply to pure compounds. A machine learning regression based group contribution method (GCM) is proposed to simultaneously predict CN, RON and MON of pure fuel compounds and mixtures. The GCM extracts the structural features of fuel molecules to build a molecular structure matrix. Then a mathematical model developed by machine learning correlates the molecular structure matrix with ignition quality (CN, RON, MON) matrix. A comprehensive fuel ignition quality database is built for model training which contains 603, 374, 371 compounds for CN, RON and MON respectively. High predictive precision is obtained for CN, RON, MON (R 2 equal to 0.9911, 0.9874, 0.9731) being superior to those obtained by neural network. The method is successfully applied to a wide range of compounds including alkanes, alkenes, alkynes, cycloalkanes, cycloalkenes, aromatics, alcohols, aldehydes/ketones, ethers, esters, acids, furans and fuel mixtures. Three key factors contribute to the high predictive capacity: (i) GCM considers the structural features, functional group interaction and fuel reactivity of fuel molecules; (ii) the built-in machine learning algorithm automatically optimizes the model function and parameters and (iii) the fuel ignition quality database provides adequate model training data for different fuel types. This method provides an effective tool to obtain CN, RON and MON of pure compounds and mixtures and a fundamental understanding of the impact of fuel molecular structures on the ignition quality.