Neural modeling of torsional potential hypersurfaces in non-rigid molecules

This paper presents a study on the ability of neural networks to model torsional potential hypersurfaces in non-rigid molecules. Using the potential function for the methyl torsion of acetaldehyde, we find that standard models do not represent accurately periodic function in its full range of definition. However, the functions are correctly described in any non-periodic zone. This behaviour arises from the periodic nature of the actual function and from the local character of the training methods. A new periodic activation function is defined, which enhances greatly the results for the periodic and non-periodic cases. The new activation function permits description of any periodic function of any number of arguments using a neural network with only one hidden layer. (C) 1998 Elsevier Science Ltd. All rights reserved.