Liquid flammability ratings predicted by machine learning considering aerosolization.

Area of Science:

• Chemical safety engineering
• Computational chemistry
• Risk assessment

Background:

• Traditional liquid flammability classification relies on flash points (NFPA 704, GHS, OSHA).
• Flash points are inadequate for aerosolized liquids, posing significant fire risks.
• Numerous incidents highlight ignition of liquids below their flash point in aerosol form.

Purpose of the Study:

• To develop novel machine learning (ML) methods for assessing liquid flammability, incorporating aerosolization potential.
• To propose a safety index combining flammability hazards and aerosolization probability.
• To cluster liquids based on flammability and aerosolization using ML algorithms.

Main Methods:

• Utilized two ML approaches on 823 compounds from the Design Institute for Physical Properties 801 database.
• Method 1: Separate rating of flammability and aerosolization, combined via a safety index.
• Method 2: Principal Component Analysis (PCA) for dimensionality reduction and clustering (K-means and Hierarchical Clustering).

Main Results:

• The first ML method effectively combines flammability and aerosolization data into a safety index.
• PCA method automatically weights properties and facilitates clustering.
• Hierarchical Clustering (HC) provided more reasonable aerosolization probability ratings, while K-means Clustering (KC) excelled in flammability clustering.

Conclusions:

• ML offers a robust approach to evaluating liquid flammability beyond traditional flash point metrics.
• Considering aerosolization is critical for accurate risk assessment of liquids.
• The study demonstrates the utility of HC and KC algorithms for distinct aspects of liquid flammability assessment.