Autonomous Defect Analysis of 3D Atom Probe Microscopy Data using Machine Learning

Tetragonal yttria stabilized zirconia (t-YSZ) is a ceramic material that exhibits excellent mechanical strength, fracture toughness, thermal shock resistance, and high ionic conductivity. Its unique combination of properties has led to its widespread use in oxygen sensors, dental restorations, solid oxide fuel cells, and other applications that require exceptional performance and reliability. To increase the tetragonal phase stability of the material, a critical vacancy concentration is desired. Annihilated vacancies, or the presence of excessive vacancies in oxide materials, can result in material degradation and reduced performance. Vacancy serves as a diffusion pathway for atomic species resulting in unwanted atomic arrangements and phase transformation. To mitigate these issues, vacancy mapping and detection are significant. While conventional microscopy tools have limitations in detecting vacancies, advanced characterization techniques such as Atom Probe Tomography (APT) can provide valuable information for predicting local structure including vacancies. APT is a powerful 3D nano-analytical tool that provides atomic positions, structural information, and chemical composition with exceptional spatial resolution. It works by field-evaporating atoms from the surface of a sample and detecting them individually, allowing for precise atom-by-atom reconstruction of the material. While APT can provide valuable insights into the atomic-scale structure, its capability for directly detecting vacancies is limited. This motivates us to complement APT data analysis with an advanced machine learning model to detect vacancy concentration and reconstruction.<br/>To detect vacancies in the microscopic 3D positional data, we build, train, and deploy a deep learning (DL) model, using architectures common in computer vision and image processing tasks. The DL model detects structural features in 3D voxelated images. To train this model on large data, synthetic dataset with known vacancy positions and associated structural and chemical information is simulated using an empirical ball-and-spring model. The learnt DL model is applied on the real APT data for automated vacancy mapping and detection in t-YSZ. By predicting the local structure, including vacancies, the model can provide insights into the defects in the material. This information can help optimize the phase structure and stability of t-YSZ and guide the design of improved oxygen sensors and solid oxide fuel cells.<br/>This approach of combining deep learning with APT data analysis holds great potential for advancing our understanding of vacancy detection and its impact on the mechanical stability of oxide materials like t-YSZ.