Regensburg 2022 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

MM: Fachverband Metall- und Materialphysik

MM 20: Computational Materials Modelling: HEA, Alloys & Nanostructures

MM 20.7: Talk

Wednesday, September 7, 2022, 12:00–12:15, H44

Alchemical machine learning for high entropy alloys — •Nataliya Lopanitsyna, Guillaume Fraux, and Michele Ceriotti — École Polytechnique Fédérale de Lausanne, Switzerland

High entropy alloys (HEAs) are a class of metallic materials composed of five or more principal elements. Interest in HEAs has grown over the last decades due to their exceptional structural and mechanical properties. HEAs are particularly challenging for atomistic modeling. Machine-learning (ML) models have emerged as a promising alternative to inaccurate empirical forcefields and very demanding first-principles simulations, with the ability to deliver the accuracy of first principle methods with lower computational resources. However, the complexity of ML models grows exponentially with the number of different elements due to the unfavourable scaling of their associated feature space sizes, limiting the chemical diversity of the systems tackled thus far. To address the problems arising from the high feature space dimensionality, first, we propose a chemical embedding compression scheme to reduce the dimensionality of the feature space required for multi-component systems, based on the framework of Willatt et al [ Phys. Chem. Chem. Phys., 2018 ], and implemented in PyTorch. Second, we generate a dataset of several thousands configurations, assembled from 25 d-block elements, which aims to represent cross-elemental interactions, evaluating their energies and forces at the DFT level. We demonstrate the effectiveness of the alchemical ML model in learning the energetics of this extremely diverse dataset, and provide showcase calculations of the properties of some realistic HEA compositions.