Fast, easy-to-use, machine learning-developed models of prediction of flash point, heat of combustion, and lower and upper flammability limits for inherently safer design

Abstract

This study proposes easy-to-apply machine learning-developed models, which predict four flammability properties of pure organic compounds: the flash point, heat of combustion, lower flammability limit (LFL), and upper flammability limit (UFL). These flammability properties pose a strong impact on the inherently safer design of industrial processes. Similar to quantitative structure-property relationship (QSPR) or group contribution models, machine learning algorithms are utilized in this study to establish predictive models. Compared to previous models, this study uses readily available variables (i.e., the numbers of atomic elements, molecular weights, and normal boiling points) as default variables without the analysis of detailed molecular structures or the in-depth knowledge of chemistry. This study consists of two steps: Step (1) building multiple linear regression (MLR) models by incorporating default input variables and Step (2) building MLR models by incorporating interaction and transformed variables to improve the predictions from the models in Step 1. In Step 1, an optimal subset of predictors is identified by constructing an MLR model via the sequential floating backward selection (SFBS) algorithm. As a result of Step 1, the two constructed models of the flash point and heat of combustion are found to be adequate, while the predictability of LFL and UFL are insufficient. In Step 2, MLR models incorporating nonlinearity and interaction terms are constructed via the sequential floating forward selection (SFFS) algorithm by selecting the optimal subset of default variables. The results show that all the constructed models in Step 2 are adequate as predictive models; the mean absolute errors (MAEs) of the flash point, heat of combustion, LFL, and UFL are 7.31 (5.67 via the SFBS) [K], 60.6 (61.87 via the SFBS) [kJ/mol], 0.21 (0.19 via the SFBS) [vol.%], and 2.44 (2.33 via the SFBS) [vol.%], respectively. Compared to previous models, the approved models in this study provide highly competitive performance with enhanced simplicity and interpretability.