SKM 2021 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 4: Poster Session II: Nonlinear Dynamics, Simulations and Machine Learning
DY 4.8: Poster
Dienstag, 28. September 2021, 17:30–19:30, P
Calculating Raman Spectra using Kernel-Based Machine Learning — •Manuel Grumet1, Karin S. Thalmann1, Tomáš Bučko2,3, and David A. Egger1 — 1Technical University of Munich, Germany — 2Comenius University in Bratislava, Slovakia — 3Slovak Academy of Sciences, Slovakia
First-principles theoretical predictions of Raman spectra are possible using either a phonon-based approach or molecular dynamics (MD) simulations. In both cases, the polarizability tensor of the system, α, is the central quantity. Specifically, the Raman spectrum is obtained from Fourier-transformed velocity autocorrelation functions (VACs) of tensor invariants of α in the MD method [1]. This requires a large number of evaluations of α and thus leads to high computational cost.
We use kernel-based machine learning (ML) to reduce the number of polarizability calculations needed. In this approach, a subset of all configurations serves as a training data set, and polarizabilities for all other configurations are predicted using ML methods. In particular, we obtain the polarizabilites using kernel ridge regression with descriptors based on the atomic neighbourhood density around each atom [2,3].
We apply these methods to a number of test systems, consisting of small molecules and simple solids. We compare different descriptors with regard to the size of the training data set required to obtain accurate predictions for polarizabilties and Raman spectra.
[1] M. Thomas et al., Phys. Chem. Chem. Phys. 15, 6608 (2013)
[2] N. Raimbault et al., New J. Phys. 21, 105001 (2019)
[3] A. P. Bartók et al., Phys. Rev. B 87, 184115 (2013)