Development of neural network interatomic potentials for accelerated prediction of stable compounds

Construction of machine learning interatomic potentials suitable for guiding unconstrained ab initio

structure searches presents a considerable challenge, as they are required to provide an accurate

description of yet unknown configurations probed in global optimizations. We have developed an

open-source wrapper based on our MAISE package that streamlines all stages of the neural network

(NN) model parameterization. An evolutionary sampling scheme for generating reference structures

improves the NNs’ mapping of regions visited in unconstrained searches, while a stratified training

approach enables the creation of standardized NN models for multiple elements. Among first

applications of such NN potentials is the prediction of new Mg-Ca alloys thermodynamically stable

under ambient or high pressures.