Radial basis function networks for obtaining long range dispersion coefficients from second virial data

A new approach, consisting of using radial basis function networks to obtain the long-range part of diatomic potential energy functions from simulated second virial coefficients, is presented. From these simulated data the artificial neural network was able not only to learn but also to predict properties for systems that were not considered during the training process. Fifteen different diatomic systems were chosen and a leave-three-out approach was applied. A cross-validation procedure was used for analysing the network generalization properties and the relative average error achieved for the three systems was about 5%, providing accurate data for the long-range dispersion coefficients.