Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
MM: Fachverband Metall- und Materialphysik
MM 20: Topical Session: Thermophysical Properties of Bulk Metallic Glasses and Bulk Metallic Glass-forming Liquids
MM 20.5: Topical Talk
Mittwoch, 19. März 2025, 17:15–17:45, H10
Diffusion and nucleation in Al-Ni melts using machine-learned MD simulations — Johannes Sandberg1,2,3, Leon F. Granz2,3, and •Thomas Voigtmann2,3 — 1Universitë Grenoble-Alpes, Grenoble, France — 2Heinrich-Heine-Universität, Düsseldorf, Germany — 3Deutsches Zentrum für Luft- und Raumfahrt, Köln, Germany
The microstructure that forms during solidification of metallic melts greatly influences the material properties. It depends crucially on the microscopic transport properties, and the initial phase selection in the cricial nucleus. Simulation of these phenomena faces two contradictory demands: while the relevant length and time scales match well that of classical molecular dynamics simulations, the sensitive dependence on details of the interatomic interactions is only captured in much smaller-scale quantum-mechanical simulations. In recent years, machine-learned interaction potentials have helped to reconcilce these requirements, allowing MD simulations to be performed with almost DFT-like accuracy.
I will present results that we have obtained using high-dimensional neural network potentials (HDNNP) to the case of Al-Ni melts and nucleation processes therein. Crucially, we assess the performance of the HDNNP by comparing to structural and dynamical experimental data of the liquids at different compositions. This reveals also how the level of DFT closure chosen in the quantum-mechanical simulations used to train the network influences the prediction of thermophysical quantities.
Keywords: machine learning; molecular dynamics simulations; Al-Ni binary melts