A New Accurate, Efficient and Transferable Deep Learning Interatomic Potential for Carbon with Advanced Applications

Abstract

Ab initio molecular dynamics (AIMD), based on density functional theory (DFT), provides advantages in terms of accuracy, despite the considerable computational time required. In contrast, traditional empirical potentials are faster but suffer from lower calculation accuracy. However, recent advancements for deep neural networks machine learning potential have enabled a promising fusion of computational efficiency and precision. In this work, the obtained potential function of carbon not only demonstrates strong scalability but also effectively enables the study of the formation mechanisms of amorphous diamond and polycrystalline diamond using C60 crystals and graphene as precursors under high-pressure high-temperature conditions (HPHT). Furthermore, the structure search software (AIRSS) was used to generate numerous initial structures, which were optimized using the machine learning potential, this process led to find new structural clusters of carbon. Interestingly, the predictive capabilities of the machine learning potential for symmetric and asymmetric carbon clusters aligned well with the Gaussian approximation potential (GAP), yet the former demonstrated higher computational efficiency, making it more suitable for carbon material research. The results of this work signify significant progress in the field of carbon machine learning potentials, opening up new possibilities for exploring and understanding carbon materials with improved computational efficacy.

Keywords: machine learning potential, Density functional theory, C60, diamond and amorphous diamond, structure search