Atomic energy in grain boundaries studied by machine learning

Abstract

Grain boundaries (GBs) have been studied for decades, but it remains a challenging task to describe characteristic GB properties by simple structural descriptors, especially at the local atomic level. In this paper, we use the atomic descriptor based on the smooth overlap of atomic positions (SOAP) to study the atomic energy at GBs by using machine learning and propose a route to simplify it. It is found that, compared with conventional local atomic descriptors such as the Voronoi index, excess volume, centrosymmetry, or local entropy, the SOAP vector shows excellent predictive performance for the atomic energy among the 172 Al and 388 Ni coincidence site lattice (CSL) GBs as well as general GBs in the nanocrystalline model. Additionally, we successfully used the datasets of GBs and amorphous models to predict the atomic energy of one another, which proves the similarity between the local atomic environments (LAEs) in GBs and the amorphous state. Furthermore, the distribution of local distortion factors based on the SOAP vector shows the transition in atomic pack ordering from special CSL GBs to general GBs and the amorphous structures. The simplified descriptor we propose can reduce the original SOAP vector from >1000 features to only a few yet still shows the superior predictive performance of the atomic energy at GBs in all cases than the conventional descriptors combined. It is expected that the simplification process can be adapted to study more complex GB behaviors. A simple and efficient descriptor of the LAEs should allow us to have a clearer picture of the structure-property correlation in GBs, which is essential for GB engineering.