SKM 2023 – wissenschaftliches Programm

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

O: Fachverband Oberflächenphysik

O 67: Poster: Electronic Structure of Surfaces

O 67.7: Poster

Mittwoch, 29. März 2023, 18:00–20:00, P2/EG

Comparison of machine learning strategies in the high-throughput exploration of ABO₂ delafossites — •Armin Sahinovic, Benjamin Geisler, and Rossitza Pentcheva — Fakultät für Physik, Universität Duisburg-Essen

The advent of machine learning introduced new techniques to considerably expedite materials discovery. This raises a fundamental question about how they balance interpretability versus accuracy. We address this aspect by comparing ensemble-based active learning (AL) of neural networks (NN) [1] and the sure independence screening and sparsifying operator (SISSO) [2] for the prediction of formation energies and lattice parameters in ABO₂ delafossite oxides. To this end, we generate a consistent dataset from first principles. Element embedding is found to be superior to scalar input strategies, e.g., atomic properties. In conjunction with AL, the NNs reach DFT accuracy, allowing for a significant acceleration of high-throughput materials screening. In contrast, the precision of the physically interpretable SISSO descriptors is limited by the high data complexity. We combine ABO₂ infinite-layer, ABO₃ perovskite [1] and the delafossite data to extend the unsupervised AL into the structural space, thereby enhancing the sample efficiency in the spirit of transfer learning. Finally, we compile a phase diagram that compares the relative stability of the three distinct oxide materials classes.

[1] A. Sahinovic and B. Geisler, Phys. Rev. Research 3, L042022 (2021); J. Phys.: Condens. Matter 34, 214003 (2022)

[2] R. Ouyang et al., Phys. Rev. Materials 2, 083802 (2018)