Flash point prediction of organic compounds using a group contribution and support vector machine

The flash point is one of the most important properties of flammable liquids. This study proposes a support vector regression (SVR) model to predict the flash points of 792 organic compounds from the DIPPR 801 database. The input variables of the model consist of 65 different functional groups, logarithm of molecular weight and their boiling points in this study. Cross-validation and particle swarm optimization were adopted to find three optimal parameters for the SVR model. Since the prediction largely relies on the selection of training data, 100 training data sets were randomly generated and tested. Moreover, all of the organic compounds used in this model were divided into three major classes, which are non-ring, aliphatic ring, and aromatic ring, and a prediction model was built accordingly for each class. The prediction results from the three-class model were much improved than those obtained from the previous works, with the average absolute error being 5.11–7.15 K for the whole data set. The errors in calculation were comparable with the ones from experimental measurements. Therefore, the proposed model can be implemented to determine the initial flash point for any new organic compounds.