Descriptor-driven UMAP (D-UMAP) for visualizing and analyzing structural evolution in SiC nano-grinding

Abstract

This study introduces a descriptor-driven uniform manifold approximation and projection (D-UMAP) framework to visualize and analyze the complex structural evolution in SiC nano-grinding. Various atomistic descriptors, including global (Coulomb, Ewald, Sine, MBTR) and local (LMBTR, ACSF, SOAP) types, were systematically evaluated for their efficacy in characterizing SiC polytypes and capturing dynamic structural transformations during simulated 4H-SiC melting. The D-UMAP framework, which integrates optimized descriptors with the UMAP algorithm, successfully mapped high-dimensional descriptor spaces to low-dimensional manifolds, preserving structural relationships and enabling intuitive visualization of continuous structural evolution. Results from nano-grinding simulations on SiC with dopants, micropipes, and inclusions demonstrate D-UMAP’s ability to reveal subtle evolutionary pathways and distinct material states, offering critical insights for process optimization and material property control in precision manufacturing.