Abstract Scope |
A new ML pipeline is developed to study the structures and properties of amorphous solids. We employ SOAP descriptor to encode the local environments, which are then fed into extreme gradient boosting tree algorithm to train, learn, and eventually predict the global configurational energy of metallic glasses. We identify 40 important unique local environments (ULEs) that are most responsible for the energy of a given glass sample. A designed decoding stage is then employed to decompose a sample’s 3N degrees of freedom configuration into a 40-dimension probability vector via frequency mapping of those ULEs. The obtained probability spectrum barcode is thus regarded as a signature representation of an interested sample. We further demonstrate that, through the analysis of occupational factions and fluctuations of the barcode, one can simultaneously estimate a sample’s global energy and level of structural heterogeneity. The physical interpretations of these ULEs and their implications are also discussed. |