Training of Machine Learning Potentials for the Modeling of Nucleation in Graphite

The parameterization of machine learning potentials (MLP) for precise characterization of the interaction between carbon atoms in graphite and diamond phases is described. The training set consisted of various allotropic forms of carbon and their compounds. The MLPs are trained using forces, energies, and stress tensors obtained from ab initio simulations. It is shown that the MLPs can accurately reproduce elastic properties and structural parameters of carbon phases. However, the MLPs also predict some unphysical behavior due to the training set limitations and the lack of long-range interactions in the MLPs. In spite of these limitations, MLPs are a promising tool for the accurate characterization of diamond nucleation in graphite.