Unsupervised Learning for Defect Identification in Yttria-Stabilized Zirconia from Atom Probe Tomography Data

The properties of ceramic materials are strongly influenced by nanoscale defects such as oxygen vacancies and grain boundaries which drive phase transformations and affect functionality. Yttria-Stabilized Zirconia (YSZ), a widely used ceramic, is particularly sensitive to such defects at the atomic scale. In this work, to investigate these nanoscale features in YSZ we employ Atom Probe Tomography (APT) focusing on the Local Atomistic Environment (LAE) of yttrium atoms. We introduce a computational framework that constructs atomic neighborhoods using Voronoi tessellations, represents them as graphs, and quantifies their structure through customizable signature functions. These signature functions, when integrated with unsupervised learning techniques, enable the extraction of meaningful patterns and defect signatures from APT datasets. By representing LAEs as graph-based structures, our approach captures intricate local details of composition, coordination, and connectivity. This representation is particularly effective for indirectly identifying defects such as oxygen vacancies. Our analysis reveals significant differences in yttrium environments of YSZ sample at grain boundaries compared to those within grains, with grain boundary regions showing fewer and less connected oxygen ions. Additionally, we distinguish between two distinct types of grain boundaries of YSZ sample based on their local atomic environments. Finally, we demonstrate that a graph based order parameter derived from the LAE enables the detection of subtle variations in oxygen ion coordination and connectivity features indicative of potential vacancies that are otherwise difficult to resolve. While applied here to YSZ, the framework is generalizable to other material systems. Its flexibility in designing material specific signature functions makes it a powerful and versatile tool for atomic scale characterization of complex materials.