Simultaneous structure exploration and machine-learned potential fitting

Defining interatomic potentials using ideas from machine learning that treat the problem as a high-dimensional fit of the reference (usually density functional theory) potential energy surface is an exciting new approach for developing accurate potentials. However, because of their variational freedom such potentials require large fitting datasets, with large amounts of manual selection and tuning of configurations by the researcher. We present an iterative method, where a preliminary potential is used to carry out a number of random-structure searches, and selected configurations from the searches are used to fit the next iteration's potential. We test the method on a number of elements with different bonding types, including an insulator, a semiconductor, and a metal. We show how the process converges in a few iterations, and how the resulting potentials reproduce the reference DFT values on a number of bulk and defect properties.