Prediction of Chemical Reactivity by Artificial Neural Networks

Abstract

Artificial Neural Networks are, presently, methods with an important role in very many fields. This is illustrated, for example, by the extremely rapid growth of the applications in many areas of chemistry, in the last 15 years [1].The prediction of polar breaking of bonds for a wide range of molecules is a chemical problem that can be solved by artificial neural networks [2,3].To understand and predict which bonds in molecules will break, we have implemented a neural network with architecture 7×3×1 trained by the back-propagation algorithm using seven energetic and electronic parameters as input. The choice of bonds for training was done by two ways: random selection and experimental design technique, using a dataset of 10 molecules.The test of the neural network, trained with two different sets of bonds, shows that a dataset of about 50 bonds is sufficient to the neural network learn the relation between physicochemical parameters and reactivity.

Keywords: Physical Chemistry > Computational Chemistry, physchem/0310002