Quantitative structure–property relationship studies for estimating boiling points of alcohols using calculated molecular descriptors with radial basis function neural networks

Abstract

Radial basis function (RBF) neural networks were used to link molecular structures and boiling points. The data sets were composed of 106 alcohol compounds, with experimental boiling points values ranging from 64.60 to 245.00 °C, and the number of carbon atoms ranging from 1 to 11. Each compound was characterized by a set of 70 molecular structure descriptors calculated by semi-empirical quantum chemical AM1 and topological method. In the process of model optimization, first, 10 significant molecular descriptors were obtained from the pool of descriptors by objective methods (identity test and pairwise correlation test) and followed by subjective feature reduction method (descriptors can be entered or removed from the model depending on the probability of the F-value). After that, the RBF neural networks were trained to model the structure and property by the orthogonal least squares (OLS) training algorithm. The total database was randomly divided into a training set (75), a validation set (15) and a testing set (15). Simulated with the final optimum RBF neural networks [10- 20 [16] -1], the absolute average residues for the training, the validation, and the testing set were 1.87, 1.74 and 2.86 °C, and mean squared error (MSE) were 5.14, 3.80 and 14.16, and the predictive correlation coefficients R=0.998 (training), 0.998 (validation) and 0.991 (testing). Also, the results based on the best multilinear regression model were presented to evaluate the possible improvement in RBF neural networks. Results show that RBF neural networks can give more satisfactory prediction ability than linear model.