Applications of neural networks to fitting interatomic potential functions

It is shown that neural networks can be used to fit a two-element many-body potential function. The system chosen is the C-H combination for which a m any-body potential formulation due to Brenner exists. Comparison between th is potential and the neural network indicates good agreement with both stru cture and energetics of the small C-H clusters and bulk carbon. However, be cause of the networks complicated structure, molecular dynamics simulations run at about a factor of 60-80% slower than with the Brenner many-body for malism.