ISSN print edition: 0366-6352 ISSN electronic edition: 1336-9075 Registr. No.: MK SR 9/7 Published monthly

Application of machine learning methods to the prediction of physicochemical properties of potential bio-based oxygenated fuel additives

Andrei Dubiniak, Sergey Egazar’yants, and Anton Maximov

Department of Petroleum Chemistry and Organic Catalysis, Faculty of Chemistry, Lomonosov Moscow State University, Moscow, Russia

E-mail: atesller@mail.ru

Received: 13 October 2025  Accepted: 13 April 2026

Abstract:

This study presents a machine learning-driven quantitative structure–property relationship framework to predict key fuel properties of potential bio-based fuel oxygenates that can be synthesized from furfural and its derivatives. Molecular structures from a curated database were encoded as SMILES strings and converted into 2D Mordred descriptors. Four ML algorithms – Kernel Ridge Regression, XGBoost, Artificial Neural Network and LightGBM – were evaluated, with XGBoost demonstrating superior performance for cetane number, density, heat of combustion and melting point prediction. LightGBM excelled at viscosity estimation, while ANN, despite careful optimization, showed slightly lower accuracy for melting point prediction, highlighting the importance of algorithm selection for each target property. The optimized models screened 250 potential biomass-derived oxygenates, identifying 10 top candidates according to diesel fuel criteria. Ether derivatives dominated the selection due to their favorable cetane numbers and synthetic accessibility. Presented results highlight 5-methylfurfural as a particularly versatile and promising compound, with five potential fuel additives derivable through its valorization.

Keywords: Extreme gradient boost; Neural network; Fuel additives properties; Fuel molecular structure; Quantitative structure property relationship

Full paper is available at www.springerlink.com.

DOI: 10.1007/s11696-026-04914-2

Chemical Papers 80 (8) 8789–8800 (2026)