Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
O: Fachverband Oberflächenphysik
O 43: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 1
O 43.8: Vortrag
Mittwoch, 7. September 2022, 12:15–12:30, S054
Machine learning for estimation of spin models in undoped cuprates — •Denys Y. Kononenko1, Ulrich K. Rößler1, Jeroen van den Brink1,2, and Oleg Janson1 — 1Institute for Theoretical Solid State Physics, IFW Dresden, Dresden, Germany — 2Institute for Theoretical Physics, TU Dresden, Dresden, Germany
Undoped cuprates tailor a fascinating variety of low-dimensional and frustrated spin models, which can be indirectly characterized by the transfer integrals. The estimation of transfer integrals is related to a relatively complicated computational procedure which includes besides DFT calculation also a Wannierization. We propose a data-driven approach to replace this computationally demanding procedure.
We employ the Gaussian Process Regression model, trained on the results of high-throughput DFT calculations to estimate transfer integrals in undoped cuprates. The model learns from data the dependency between the local crystal environment of copper atoms pair and the corresponding value of transfer integral. The site position function of the local crystal environment is represented as a finite-dimensional vector composed of decomposition coefficients in the truncated basis of Zernike 3D functions [1]. The vector descriptor incorporates the spatial configuration and chemical composition of the local crystal environment. The proposed approach can be utilized for a rapid assessment of the spin models of new cuprates using structural information as the only input.
[1] M. Novotni and R. Klein, Computer Aided Design 36, 1047 (2004)