Machine learning for effective on-site Coulomb interaction parameters in transition metal oxides in DFT+U method Takahiro Matsumoto, Kenji Nawa, Koji Nakamura
Graduate school of Engineering, Mie University
Abstract
Theoretical understanding of strongly-correlated magnetic materials is of importance for realizing novel magneto-optical device. To treat large complex systems, however, calculations may be practically limited due to huge computational resources demanded. In order to overcome this problem, we have combined neural network technique to density functional theory (DFT+\(U\) method), and applied to NiO as test calculations and extended to yttrium iron garnet (YIG) that contains 160 atoms in a unit. In prediction, such as for dielectric spectra, two parameters may be optimized- effective on-site Coulomb interaction parameter, \(U_{\rm eff}\), and the relaxation time, \(1/\tau \). We thus constructed a neural network to learn a relation between the parameters and the calculated dielectric spectra, and then determined the optimal values to reproduce experimental spectra. For the NiO, we obtained \(U_{\rm eff}\) of 6.1 eV and \(1/\tau \) of 0.38 eV, which are reasonably close to literature values. In the case of YIG, the results interestingly predict to two different \(U_{\rm eff}\) values, \(U_{\rm eff}^{\rm O}\) and \(U_{\rm eff}^{\rm T}\), for two inequivalent octahedral and tetrahedral Fe cation sites. The calculated spectra fairly agree with experiments. The detailed results and discussion including the methodology will be presented.
