Regensburg 2022 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
O: Fachverband Oberflächenphysik
O 83: Frontiers of Electronic Structure Theory: Focus on Artificial Intelligence Applied to Real Materials 4
O 83.9: Talk
Friday, September 9, 2022, 12:30–12:45, S054
MD-based Raman Spectra using Machine Learning — •Manuel Grumet1, Karin S. Thalmann1, Tomáš Bučko2,3, and David A. Egger1 — 1Department of Physics, Technical University of Munich, Garching, Germany — 2Comenius University in Bratislava, Slovakia — 3Slovak Academy of Sciences, Slovakia
Theoretical calculations of Raman spectra based on molecular dynamics (MD) trajectories allow to directly incorporate both anharmonic and temperature-dependent effects and thus yield more realistic spectra compared to a phonon-based approach [1]. The spectra can be calculated from the Fourier-transformed velocity correlation function of the polarizability tensor α. However, this requires evaluating α for a large number of MD configurations along each trajectory, which has high computational cost if done by ab-initio methods.
We therefore use kernel-based machine learning (ML) methods with density-based descriptors [2, 3] to predict α based on atomic positions. Ab-initio calculations are then only needed for obtaining a training data set, reducing the computational cost significantly. We use a number of test systems, including both solids and small molecules, to test and optimize several different variants of this approach and compare the achieved prediction performances. We also test transferability of the trained models to trajectories at different temperatures.
[1] M. Thomas et al., Phys. Chem. Chem. Phys. 15, 6608 (2013)
[2] A. P. Bartók et al., Phys. Rev. B 87, 184115 (2013)
[3] A. Grisafi et al., Phys. Rev. Lett. 120, 036002 (2018)