Silicon potentials investigated using density functional theory fitted neural networks

We present a method for fitting neural networks to geometric and energetic data sets. We then apply this method by fitting a neural network to a set of data generated using the local density approximation for systems composed entirely of silicon. In order to generate atomic potential energy data, we use the Bader analysis scheme to partition the total system energy among the constituent atoms. We then demonstrate the transferability of the neural network potential by fitting to various bulk, surface, and cluster systems.