Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
MM: Fachverband Metall- und Materialphysik
MM 4: Data Driven Material Science: Big Data and Workflows I
MM 4.9: Talk
Monday, March 18, 2024, 12:30–12:45, C 243
Automatic Generation of Atomic Structure Datasets for Machine Learning Potentials: Alloys and Applicatoin to Mg/Al/Ca — •Marvin Poul1, Liam Huber2, and Joerg Neugebauer1 — 1Max-Planck- Institut für Eisenforschung, Düsseldorf, Germany — 2Grey Haven Solutions, Victoria, Canada
We extend a recently proposed strategy for automatically generating training data for machine learning interatomic potentials (MLIP) to alloys.[1]
It is based on small periodic structures (around ten atoms) of various concentrations that are sampled from all crystallographic space groups. These structures are minimized and then again randomly perturbed in positions and cell shape around the obtained local minima. This procedure akin to ab initio random structure search yields samples around the relevant parts of phase space without prior knowledge automatically. Only the cell stoichiometry and the magnitude of the random perturbations remain hyperparameters in this approach.
We explore the natural question of how well potentials can extrapolate in the combinatorically large concentration space and test that they accurately describe structures near the convex hull as well as larger super cells of random alloys. Finally we verify the potentials on binary phase diagrams (and defect phase diagrams) in the ternary Mg/Al/Ca system.
This opens the way for automatic parametrization of MLIPs, promising to bring ab initio accuracy to a large number of systems at scale.
[1]: https://doi.org/10.1103/PhysRevB.107.104103
Keywords: Machine Learning Potentials; Potentials; Workflows; MTP; Phase Diagrams