Liquid flammability ratings predicted by machine learning considering aerosolization

• Published in: Journal of hazardous materials 2020-03, Vol.386, p.121640-121640, Article 121640
• Main Authors: Yuan, Shuai, Zhang, Zhuoran, Sun, Yue, Kwon, Joseph Sang-II, Mashuga, Chad V.
• Format: Article
• Language: English
• Subjects: Aerosol flammability; Flammability rating; Liquid flammability; Machine learning; Risk assessment
• ISSN: 0304-3894; 1873-3336
• DOI: 10.1016/j.jhazmat.2019.121640

[Display omitted] •Current liquid flammability standards are flash point based.•Liquids can ignite below the flash point when an aerosol.•Viscosity & surface tension predict aerosolization which affect liquid flammability.•Unsupervised learning applied to database of 823 substances resulting in 6 groups.•Aerosol Flammability Safety Index has advantage compared with PCA approach. Liquid flammability is classified based on flash point as in NFPA 704, GHS, and OSHA. However, flash points become insignificant when the liquid is in aerosol form, which is evident from numerous incidents revealing that a liquid can be ignited below its flash point when an aerosol. In this study, two machine learning (ML) methods are utilized to propose liquid flammability ratings while considering aerosolization. 823 compounds from the Design Institute for Physical Properties 801 database are used in this study. The first method rates the liquid flammable hazards and probability of aerosolization separately and then uses the proposed safety index to combine the contribution of flammable hazards and aerosolization. The second method uses Principal Component Analysis (PCA) to create two principal components, then clusters the liquids based on these two principal components. The PCA method advange is the weight of each property is automatically considered. A traditional risk assessment utilizes a risk matrix, this study uses two ML clustering algorithms are applied, K-means Clustering (KC) and Hierarchical Clustering (HC). Based on expert judgment, the HC algorithm gives a more reasonable rating of the probability of aerosolization, while the KC algorithm has a more reasonable rating on liquid flammability clustering.