Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 36: Data Driven Materials Science: Big Data and Work Flows – Microstructure-Property-Relationships (joint session MM/CPP)
MM 36.4: Vortrag
Donnerstag, 30. März 2023, 11:00–11:15, SCH A 251
Identifying ordered domains in atom probe tomography using machine learning — •Alaukik Saxena, Navyanth Kusampudi, Shyam Katnagallu, Baptiste Gault, Dierk Raabe, and Christoph Freysoldt — Max-Planck-Institut für Eisenforschung GmbH, Düsseldorf 40237, Germany
Atom probe tomography (APT) is a unique technique that provides 3D elemental distribution with near-atomic resolution for a given material. The spatial resolution of APT is ~1-3 Å in depth and ~3-5 Å in the lateral direction, respectively. Due to the limited spatial resolution, most of the APT data analysis focuses on composition to extract various microstructural features. Here, we aim at identifying additional on-lattice short-range order within an Al-Mg-Li alloy even though the underlying FCC lattice itself is not resolved. We propose a machine learning (ML) methodology to distinguish disordered solid solutions from ordered L12 domains. To encapsulate the local chemistry and noisy structure in APT independent of orientation, we use Smooth Overlap of Atomic Positions (SOAP). To find suitable hyperparameters of the high-dimensional SOAP features, we visualize the data distribution within the latent space of an auto-encoder neural network trained on experimental data with a preliminary classification. After the optimization, synthetic data corresponding to FCC and L12 structures is created with APT level spatial noise and then used for training from scratch a dense neural network for order/disorder classification. The trained model is then able to distinguish between ordered and disordered structures in experimental data.