Toward generalized models for estimating molecular weights and acentric factors of pure chemical compounds

In this work, four prompt and robust techniques have been used to introduce new generalized models for estimation of the physical properties of pure substances, including molecular weight and acentric factor. These methods were developed based on radial basis function (RBF) neural networks, group method of data handling (GMDH), multilayer perceptron (MLP), and least square support vector machine (LSSVM) techniques. Models were introduced based on a set of experimental data including 563 pure compounds that were collected from available literature. Input parameters for estimation of molecular weight were considered as specific gravity and normal boiling point. Critical temperature, critical pressure and normal boiling point were selected as inputs for estimation of the acentric factor. Statistical and graphical error analyses normal boiling point revealed that all of the developed models are accurate. The designed RBF models give the most accurate results with an AAPRE of 5.98% and 1.92% for molecular weight and acentric factor, respectively. The developed GMDH models are in the form of simple correlations, which can be used easily in hand calculation problems without any need to computers. Comparison of the developed models with the available methods showed that all of the developed models are more accurate than the existing methods. Using the relevancy factor, the impact of each input parameter on the output results was determined. Additionally, to find out the applicability region of the developed models, and to demonstrate the reliability of the models, the Leverage method has been used. There are few data out of the applicability domain of the proposed models. All the statistical and graphical resolutions, demonstrate the reliability of the developed models in estimating the molecular weight and acentric factor.