Nonlinear molecular based modeling of the flash point for application in inherently safer design

New chemical process design strategies utilizing computer-aided molecular design (CAMD) can provide significant improvements in process safety by designing chemicals with required target properties and the substitution of safer chemicals. An important aspect of this methodology concerns the prediction of properties given the molecular structure. This study utilizes one such emerging method for prediction of a hazardous property, flash point (FP), which is in the center of attention in safety studies. Using such a reliable data set comprising 1651 organic and inorganic chemicals, from 79 diverse material classes, and robust dynamic binary particle swarm optimization for the feature selection step resulted in the most efficient molecular features of the FP investigations. Apart from the simple yet precise five-parameter multivariate model, the FP nonlinear behavior was thoroughly investigated by a novel hybrid of particle swarm optimization and support vector regression. Besides, 195 missing experimental FPs of the DIPPR data set are predicted via the presented procedure.

Simple yet accurate estimation of Flash point for 1651 organic and simple inorganics (RMSEall = 28.04 K).Figure optionsDownload as PowerPoint slideHighlights
► Using the most recent released DIPPR data set, the flash point (FP) of 1651 chemicals comprising 79 diverse chemical material classes were theoretically studied and modeled.
► The multivariate regression resulted to simple yet accurate QSPR model prediction of the FPs with an acceptable accuracy and practical applicability in inherently safer design (ISD).
► For investigation of the nonlinear behavior of the FPs with the nonlinear obtained molecular features a hybrid support vector regression with particle swarm optimization is introduced and the results compared to the provided multivariate model and previous QSPR studies.
► Additionally, 195 DIPPR chemical compounds with unknown experimental FPs were estimated by the multivariate and PSO-SVR models.