Alaukik Saxena

N. Kusampudi, S. Katnagallu, J. Neugebauer, B. Gault, C. Freysoldt

Max-Planck-Institut für Eisenforschung GmbH, D-40237 Düsseldorf, Germany

Machine learning based analysis of atom probe tomography data

Typically, deciphering the relationship between processing, structure, and properties of a specific material through supervised machine learning (ML) necessitates the accumulation of substantial experimental or simulation training datasets. Subsequently, descriptors or features embodying the pertinent information within the data are extracted. The endeavor of deriving descriptors from the training data spans a spectrum of complexity; for instance, it's more straightforward with 2D microstructure images where relevant features can be labeled by a human, yet escalates in complexity when describing 3D atomic environments in atomic-scale simulations. We have architected automated workflows to unveil and isolate descriptors from geometrically intricate 3D atom probe tomography (APT) microstructures leveraging ML. APT is a unique technique that provides 3D elemental distribution with near-atomic resolution for a given material. The developed methods are adaptable across a wide array of 3D APT datasets, thus empowering the prospective construction of high-throughput ML models aimed at unraveling the intricate interplay between material structure and properties. The developed workflows are showcased on APT datasets containing plate-like and linear microstructural features. Further, a SOAP (Smooth Overlap of Atomic Positions)--based ML model is trained to perform the structural analysis on APT data.