For the efficient design of biofuels, it is critical to identify biofuel candidates that show desirable fuel properties before expensive synthesis and experimental property measurements by in silico predictions. To this end, the machine learning (ML) models were developed to rapidly predict two fuel properties: the heat of vaporization (HoV) and cetane number (CN). HoV is one of the key properties determining particulate matter emission and antiknock characteristics. CN quantifies the autoignition reactivity of diesel fuels, and it is one of the criteria used to screen bio-derived diesel candidates. According to the ASTM D975 standard, the acceptable diesel fuel should exceed the minimum CN of 40, while EN590 requires a minimum of 51.<br/>In this work, the ML models were trained against experimental databases of HoV and CN, and their accuracy was assessed using the held-out test data. Comparison of the accuracy with the existing model was also performed to corroborate that the developed models show higher reliability for a broader scope of biofuel candidate molecules. The ML models were utilized to predict the value of CN and HoV and gain insights that lead to rational design principles of biofuels. The additional analysis was performed to elucidate the effect of specific chemical structures and functional groups on HoV and CN. Finally, the ML models were applied to propose desirable biofuel compounds whose HoV and CN were not yet measured.

Development of machine learning models to predict biofuel properties

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.