Crystalline structure and grain boundary identification in nanocrystalline aluminum using K-means clustering

Abstract

K-means clustering was carried out to identify the atomic structure of nanocrystalline aluminum. For this purpose, per-atom physical quantities were calculated bymeans ofmolecular dynamics simulations, such as the potential energy, stress components, and atomic volume. Statistical analysis revealed that potential energy, atomic volume and von Mises stress were relevant parameters to distinguish between fcc atoms and grain boundary atoms. These three parameters were employed with the K-means algorithm to establish two clusters, one corresponding to fcc atoms and another to GB atoms. When comparing the K-means classification performance with that of CNA, an F-1 score of 0.969 and a Matthews correlation coefficient of 0.859 were achieved. This approach differs from other traditional methods in that the quantities employed here do not require input settings such as the number of nearest neighbor nor a cut-off value. Therefore, K-means clustering could be eventually used to inspect the atomic structure in more complex systems.

K-means clustering was carried out to identify the atomic structure of nanocrystalline aluminum. For this purpose, per-atom physical quantities were calculated bymeans ofmolecular dynamics simulations, such as the potential energy, stress components, and atomic volume. Statistical analysis revealed that potential energy, atomic volume and von Mises stress were relevant parameters to distinguish between fcc atoms and grain boundary atoms. These three parameters were employed with the K-means algorithm to establish two clusters, one corresponding to fcc atoms and another to GB atoms. When comparing the K-means classification performance with that of CNA, an F-1 score of 0.969 and a Matthews correlation coefficient of 0.859 were achieved. This approach differs from other traditional methods in that the quantities employed here do not require input settings such as the number of nearest neighbor nor a cut-off value. Therefore, K-means clustering could be eventually used to inspect the atomic structure in more complex systems.