Dresden 2026 – wissenschaftliches Programm

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

MM: Fachverband Metall- und Materialphysik

MM 13: Data-driven Materials Science: Big Data and Workflows I

MM 13.2: Vortrag

Dienstag, 10. März 2026, 10:30–10:45, SCH/A251

Structural relaxations for nonstoichiometric alloys without forces — •Luca Numrich and Hongbin Zhang — Insitute of Materials Science, Technische Univerität Darmstadt, Darmstadt, Germany

Advances in machine learning and first-principles electronic-structure methods are accelerating materials design, focusing on stoichiometric compounds. Coherent potential approximation (CPA) is a standard approach for alloys with chemical disorder but there is no compatible implementation for forces, hindering the investigation of structural relaxations for nonstoichiometric alloys. For example, in many Ni-based Heusler alloys, a high-temperature disordered B2 phase transforms into the fully ordered L2₁ structure upon cooling via a second-order B2-L2₁ order-disorder transition.

In this work, we integrate Bayesian optimization (BO) with CPA as implemented in the exact muffin-tin orbitals (EMTO) code to identify low-energy crystal structures for nonstoichiometric alloys, as demonstrated for Ni-Mn-Ga-X Heusler alloys. Symmetry analysis is used to reduce the number of independent Wyckoff coordinates, which serve as parameters for BO, while the energy per atom computed via EMTO-CPA defines the objective function to be minimized. Using a Gaussian process surrogate model with a parallel upper confidence bound acquisition function, the framework autonomously proposes candidate structures with progressively lower energies. Thus, instead of a relaxation via interatomic forces, an acquisition function is guiding the relaxation.

Keywords: Bayesian optimization; Alloy; First-principles; Heusler; Machine learning