Automated Generation of Structure Datasets for Machine Learning Potentials and Alloys

Abstract We propose a strategy for generating unbiased and systematically extendable training data for machine learning interatomic potentials (MLIP) for multicomponent alloys, called Automated Small SYmmetric Structure Training or ASSYST. Based on exploring the full space of random crystal structures with space groups, it facilitates the construction of training sets for MLIPs in an automatic way without prior knowledge of the material in question. The advantages of this approach are that only cells consisting of few atoms (~10) are needed for the DFT training set, and the size and completeness of the data set can be systematically controlled with very few parameters. We validate that potentials fitted this way can accurately describe a wide range of binary and ternary phases, random alloys, as well as point and extended defects, that have not been part of the training set. Finally, we estimate the binary phase diagrams with good experimental agreement. We demonstrate that the overall excellent performance is not a coincidence, but a consequence of the extensive sampling in phase space of ASSYST. Overall, this means that ASSYST will enable the largely autonomous generation of high-quality DFT reference data and MLIPs.