Prediction of mixture properties based on QSPR models : the flash point of organic mixtures as a test case
Résumé
Quantitative structure property relationships (QSPR) are increasingly used for the prediction of physico-chemical properties of pure compounds. Such models rely on mathematical relationships linking a target macroscopic property to a series of molecular descriptors of various types: constitutional, topological, geometric and quantum chemical. INERIS developed such models for various hazardous physico-chemical properties like explosibility and flammability. p to now only few works were devoted to the prediction of the properties of mixtures based on QSPR models since they are mainly limited to pure compounds. But, the substances used in industrial processes are mainly mixtures, sometimes complex. So, the development of QSPR models for mixture is of great interest at R&D level for the selection of substances in a new industrial process or to estimate the hazards of mixtures in process safety issues. In particular, the prediction of the flash point of organic mixtures represents a great challenge, since it characterizes flammability hazards of liquids and is a key safety issue in the risk assessment of industrial processes. Mixing rules are commonly used but they need knowledge of the flash point of each pure component. Moreover, most of the existing QSPR models for the prediction of flash points of pure compounds require the knowledge of the boiling point. In this study, new mixture descriptors were proposed by combination of the molecular descriptors calculated for each component of the mixture, taking into account the linear or non-linear dependences of the flash point with the concentration of each compound. Based on these mixture descriptors, new multilinear QSPR models were developed for a large dataset of 435 flash points of organic binary mixtures. Good prediction capabilities were obtained with a prediction error of only 10.3K evaluated on an external validation set.