The fitting of potential energy and transition moment functions using neural networks: transition probabilities in OH (A2Rþ ! X2P)

Abstract

We have studied the performance of the back-propagation neural network with different architectures and activation functions to fit potential energy curves and dipolar transition moment functions of the OH molecule from the ab initio data points of Bauschlicher and Langhoff [J. Chem. Phys. 87 (1987) 4665]. The neural network fittings are tested in different moments of the training process by computing the vibrational levels, the transition probabilities between A2Rþ and X2P electronic states, and the radiative lifetimes. The results from the neural network fittings are then compared with experimental values, previous results calculated by Bauschlicher and Langhoff and the ones obtained by using of extended Rydberg function fitting.