A COMBINED APPLICATION OF 2 DIFFERENT NEURAL-NETWORK TYPES FOR THE PREDICTION OF CHEMICAL-REACTIVITY

A multilayer neural network trained by the back-propagation algorithm is developed that is able to predict which single bonds in aliphatic molecules will break preferentially. Each potential bond breaking is described by seven empirical physicochemical parameters that allow the treatment of large datasets of organic molecules. This makes the approach outlined in this paper attractive for an automatic learning from reaction databases. It is demonstrated that a Kohonen network can be used as a basis for the selection of a training set for a supervised learning method. For training a multilayer neural network this selection gives results that arc superior to a random selection and also to an experimental design technique. A detailed analysis of the Kohonen mapping shows that the chemical similarity of bond breakings is perceived by the topology-conserving Kohonen mapping of a multidimensional space.