3D Cloud Component Analysis of Atomic Structures

Abstract

We present a method for decomposing atomic structures into physically meaningful components by converting discrete atomic coordinates into continuous three-dimensional density fields. Each element is represented by a Gaussian-smeared density map with values ranging from 0 to 1, computed efficiently through a bin-then-blur approach with periodic boundary conditions. The sum of all element densities, truncated at unity, defines the material region; its complement defines the vacuum. Every voxel is first assigned a chemical formula from the set of elements present above a threshold; the resulting formula map is then cleaned so that only regions with a genuine bulk interior-measured by the Euclidean distance to their own boundary-survive as components. Thin surface terminations (e.g., a Ga monolayer on GaAs) and one-to-two-voxel boundary layers between two crystals are absorbed by the neighboring stable region, so the decomposition contains exactly the bulk-like chemical-formula components and the vacuum, with no surface or interface components. For each component we determine the crystal phase with a neural prototype classifier: one phase if the trusted interior atoms vote unanimously, two phases (e.g., crystalline and amorphous Si) if they split into two confident groups, in which case the component is divided into two. Interfaces and surfaces are then derived as boundaries-material-material and material-vacuum-and every atom is labeled bulk, surface, interface, or vertex, with vertex reserved for geometric corners of a component. Inside crystalline components, inner defects are detected from coordination-number and local-density deviations. We demonstrate the approach on Si/GaAs and Si/SiO2 heterojunctions, crystalline/amorphous silicon junctions, vacancy-containing crystals, and bulk crystals.