Neural network prediction of glass-transition temperatures from monomer structure

Abstract

Our goal is to establish the applicability of artificial neural networks to the prediction of physical and mechanical polymer properties from their monomer structures alone. We demonstrate their ability to predict, quickly and accurately, the glass-transition temperatures (T_g) of linear homopolymers. A variety of multilayer, feed-forward artificial neural networks were trained to predict the T_g values of the polymers from their monomer structures using the error back-propagation method. The networks were constructed using the PlaNet and Aspirin/Migraines software. The networks were trained using 360 example monomer structures and tested on an independent set of 89 different monomer structures. In these initial trials, the best networks were able to predict the T_g values for a testing set of polymers with a wide range of structures with an rms error of ca. 35 K based on monomer structure information alone, with no additional information included about the polymers.