Generation of a Novel Equation for Molecular Boiling Point Estimation Using a Neural Network

Abstract

We have developed a novel property estimation equation with a group contribution scheme for molecular properties (boiling points), in the standard condition using a three layers perceptron type neural network and are equipped MolWorks™ with it. 142 groups are newly defined as a set to reproduce the differences of isomers and to realize more accurate predictions than are available with usual methods. 765 data of molecular boiling points are selected for training of the neural network. 953 data were applied to evaluate the efficiency of the equation. The correlation of observed and predicted molecular boiling points by this work is better than the values obtained by Joback's equation. The equation is applicable to estimate a wide thermal range, including high and low temperature regions. Furthermore, the equation well reproduces the differences of boiling points for not only ortho-, meta-, and para-isomers but also for cis-and trans-isomers.