Extracting knowledge from molecular mechanics simulations of grain boundaries using machine learning

Abstract

In this paper, we demonstrate that the “process-structure-property” (PSP) paradigm of materials science can be extended to atomistic grain boundary (GB) simulations through the development of a novel framework that addresses the objective identification of the atoms in the grain boundary regions using the centro-symmetry parameter and local regression, and the quantification of the resulting structure by a pair correlation function (PCF) derived from kernel density estimation (KDE). For asymmetric tilt GBs (ATGBs) in aluminum, models were successfully established connecting the GB macro degrees of freedom (treated as process parameters) and energy (treated as property) to a low-rank GB atomic structure approximation derived from principal component analysis (PCA) of the full ensemble of PCFs aggregated for this study. More specifically, it has been shown that the models produced in this study resulted in average prediction errors less than 13 mJ/m2, which is less than the error associated with the underlying simulations when compared with experiments. This demonstration raises the potential for the development and application of PSP linkages from atomistic simulation datasets, and offers a powerful route for extracting high value actionable and transferrable knowledge from such computations.