Dresden 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 17: Data-driven Materials Science: Big Data and Workflows II

MM 17.1: Vortrag

Dienstag, 10. März 2026, 14:00–14:15, SCH/A251

Modelling Diffusion Kinetics in Refractory High Entropy Alloys Using Graph Neural Network Database Models — •Klemens Lechner¹, Jiyao Zhang¹, Peter Wagatha², Wolfram Knabl², Helmut Clemens¹, and David Holec¹ — ¹Department of Materials Science, Montanuniversitaet Leoben — ²Plansee SE

Refractory high-entropy alloys (RHEAs) offer exceptional mechanical and thermal properties, such as high-temperature strength, and may exhibit high-temperature oxidation and corrosion resistance. However, their stability at high temperatures has yet to be confirmed. Nonetheless, even thermodynamically unstable solid solutions can have useful applications if the decomposition is slow. This is inherently connected with the (self-)diffusion kinetics. In this study, we present a workflow for the systematic investigation of diffusion kinetics in RHEAs. The necessary diffusion barriers are predicted using a graph neural network (GNN). We train the GNN using an active learning cycle involving molecular statics simulations with a universal machine-learning interatomic potential (uMLIP). The training data of migration barriers are calculated using the Nudged Elastic Band method. By varying the amount of training data, the GNN can be trained to an accuracy that, in theory, can fully mimic that of the uMLIP but with a more efficient computation. This is crucial for larger-scale modeling applications, e.g., the kinetics of decomposition, ordering or clustering of specific elements. We demonstrate the usage and performance of the GNN to quantify self-diffusion in Mo-Nb-Ta-W alloys using the Kinetic Monte Carlo method.

Keywords: Active Learning; High Entropy Alloy; Chemically Complex Alloy; Machine Learning; Diffusion