Fig. 1
Representing site-specific interstice distribution in the short and medium range. We define the distance, area, and volume interstice (in ratio) as the fraction of neighboring distances, triangulated surface area, and tetrahedra that are not covered by atom spheres. We characterize the distance, area, and volume interstices across all nearest-neighbor bonds and convex hull simplices and take statistics (mean, min, max, and std) of them to describe the anisotropy of interstice distribution in the short-range order (SRO) around each atom. The SRO feature vectors of all neighbors can be reduced to a single medium-range order (MRO) feature vector by calculating statistics across neighbors (mean, min, max, and std). The SRO and MRO features are then concatenated (⊕) as a representation of the interstice distributions around each atom. The features are then served as input to a ML algorithm (gradient boosting decision tree in this work) to train and predict the heterogeneous plastic response of atoms in MGs.
