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

MM: Fachverband Metall- und Materialphysik

MM 34: Data Driven Materials Science: Interatomic Potentials / Reduced Dimensions

MM 34.1: Vortrag

Donnerstag, 8. September 2022, 15:45–16:00, H45

Constructing Training Sets for Transferable Moment Tensor Potentials: Application to Defects in Bulk Mg — •Marvin Poul, Liam Huber, Erik Bitzek, and Joerg Neugebauer — Max-Planck-Institut fuer Eisenforschung

Machine learned interatomic potentials promise to bring quantum mechanical accuracy to system sizes that are inaccessible with traditional QM approaches. Here, we present a set of unary Mg Moment Tensor Potentials[1] with different speeds and accuracies in the range of 100–5 meV/atom. We focus on understanding the role of the training data in the fitting process. We discuss several ways in which the structural complexity of the training structures and a physical understanding of them helps to design an efficient training set construction. The resulting potentials are verified on out-of-fold structures, like vacancies, surfaces, and high-symmetry grain boundaries. This work is implemented as a pyiron[2] workflow and we identify challenges and opportunities of a fully automated setup to fit machine-learned potentials.

[1]: https://doi.org/10.1088/2632-2153/abc9fe

[2]: https://doi.org/10.1016/j.commatsci.2018.07.043