Development of an interatomic potential for Fe-He by neural network

Abstract

Ferritic steel is a widely used structural material for both nuclear fusion and advanced fission reactors, yet the presence of helium degrades the mechanical properties of materials. Molecular dynamics (MD) simulation is a commonly used approach to study the mechanism of helium effect in iron from the atomic scale. The accuracy and reliability of MD simulations depends critically on the empirical interatomic potentials. In this work, we develop a high-dimensional neural network potential (HDNNP) to the Fe-He system that incorporates an extensive dataset of 5205 atomic configurations from ab-initio electronic structure calculations. The developed potential not only well reproduces the bulk properties of bcc iron, but also predicts the binding properties and relative stabilities of helium clusters accurately in the iron matrix. Overall, the average error of the binding energies predicted by HDNNP is 0.04 eV per atom as compared to 0.18 eV per atom for previously reported empirical potentials (EPs). The present Fe-He HDNNP is used to perform MD simulations of helium migration in the iron matrix and it is found that migration energies of both interstitial and substitutional helium are highly consistent with ab-initio results.