Neural network potentials for disordered carbon and silicon systems.

Neural network potentials have been used as an alternative method for calculating energies and forces in systems of atoms. In the present work, we improved the classical Behler and Parrinello neural network potential; and, applied our methodologies to systems with different combinations of carbon and silicon, up to several hundreds of particles, arranged in ordered and disordered structure. We will demonstrate how our neural network potential can handle many distinct configurations of disordered materials. Our improvements primarily include - 1) changing the atomic-like potentials for structure-like potentials; and 2) adding four- and five-body interactions to increase the information stored in the features that feed the neural network. With our neural network potential, we can calculate energies and forces for systems of hundreds of atoms, with accuracies close to density functional theory codes, but within the time-frame of a force-field calculation. We use Google’s framework for artificial intelligence, called Tensor Flow - for the implementation of the neural network potential - this enables us to speedup the calculations while keeping the error low.