Development of Neural-Network Interatomic Potentials for Structural Materials

Abstract

The validity of the molecular dynamics (MD) simulation is highly dependent on the accuracy or reproducibility of interatomic potentials used in the MD simulation. The neural-network (NN) interatomic potential is one of promising interatomic potentials based on machine-learning method. However, there are some parameters that should be determined heuristically before making the NN potential, such as the shape and number of basis functions. We have developed a new approach to select only relevant basis functions from a lot of candidates systematically and less heuristically without loosing the accuracy of the potential. The present NN potential for Si system shows very good agreements with the results obtained using ab-initio calculations.