(511e) A Novel Quantitative Chemical Reactivity Rating System Design By Machine Learning Methodology

The hazard associated with chemical reactivity has posed significant threats to chemical manufacturers and workers with the potential to trigger runaway incidents resulting in tremendous fatalities, injuries, and property loss. According to the incident statistics by the Chemical Safety Board (CSB), 60% of the reactive chemicals that caused severe incidents was not rated as reactive or had “0” accounts for chemical reactivity according to the current NFPA rating criterion. Only 38 out of 137 chemicals listed by the Process Safety Management (PSM) Standard (29 CFR 1910.119) were considered highly reactive based on NFPA instability ratings of “3” or “4.” Moreover, the NFPA rating system regarding chemical reactivity is empirical and qualitative. It is of great importance to developing a new chemical reactivity rating system to facilitate the quantitative screening and classification of highly hazardous chemicals. This study aims to establish a novel framework of reactivity rating criterion by utilizing both intrinsic chemical/physical properties and thermal stability indicators of reactive chemicals with no need of calculating sophisticated QSPR descriptors as the standard practice. First, in order to deal with the highly correlated natures among various intrinsic physical/chemical properties and thermal stability indicators, we utilize specific latent variable based model reduction methodology to get new independent features of reactive chemicals. Then, multiple machine learning algorisms are adopted to generate models of correlating intrinsic physical/chemical properties and thermal stability indicators with the coherent ratings for the reactive chemicals among NFPA and PSM/CSB investigation report. These models are then used to predict the missing reactivity ratings for the mislabelled reactive chemicals in the PSM/CSB investigation report under NFPA rating scheme. Model accuracy and prediction performance are compared among different models by the cross-validation method.