Efficient Training of Machine Learning Potentials by a Randomized Atomic-System Generator

Machine learning
potentials provide an efficient and comprehensive
tool to simulate large-scale systems inaccessible by conventional
first-principles methods still in a similar level of accuracy. One
critical issue in constructing machine learning potentials is to build
training data sets cost-effectively that can represent the potential
energy surface in a wide range of configurations. We develop a scheme
named randomized atomic-system generator (RAG) to produce the training
sets that widely cover the potential energy surface by combining the
random sampling and structural optimization. We apply the scheme to
construct the machine learning potentials for simulation of chalcogen-based
phase change materials. Constructed machine learning potentials successfully
simulate the dynamics of melting and crystallization processes of
binary GeTe at a level comparable to first-principles simulations.
The visual analysis shows that the RAG-generated training set represents
the crystallization process including the amorphous phases. From the
velocity autocorrelation function obtained from the molecular dynamics
simulations, we calculate the phonon density of states to analyze
the vibrational properties during crystallization.