Learning grain boundary segregation behavior through fingerprinting complex atomic environments

Abstract

Although continuum-scale segregation is a well-documented behavior in multi-species materials, detailed site-specific behavior remains largely unexplored. This is partially due to the complexity of analyzing materials at the requisite time and length scales for describing segregation with full atomic accuracy. Here, we better evaluate the segregation behavior of disordered grain boundary (GB) atomic environments through leveraging a set of Strain Functional Descriptors (SFDs) to generate an atomic descriptor (i.e., fingerprint). Using this atomic fingerprint, we resolve key relationships between atomic structure and segregation energy. Machine learning (ML) techniques are utilized in concert with this SFD fingerprint to elucidate complex relationships relating segregation potential to changes in specific features of the local Gaussian density captured by the SFDs. Finally, we identify relationships that indicate both individual and joint structure-property correlations. Linking atomic segregation energy to key structural features demonstrates the value of higher-order descriptors for uncovering complex structure-property relationships at an atomic scale.

Describing site-specific segregation in multi-species materials is a computationally complex task that typically requires model simplification, at the expense of atomic accuracy, or limitation to small samples. Here, the relationships between local atomic environments at grain boundaries and their segregation energies are investigated by developing suitable machine learning atomic descriptors.