Abstract
Raman spectroscopy is an effective tool to analyze the structures of various materials as it provides chemical and compositional information. However, the computation demands for Raman spectra are typically significant because quantum perturbation calculations need to be performed beyond ground state calculations. This work introduces a novel route based on deep neural networks (DNNs) and density-functional perturbation theory to access anharmonic Raman spectra for extended systems. Both the dielectric susceptibility and the potential energy surface are trained using DNNs. The ab initio anharmonic vibrational Raman spectra can be reproduced well with machine learning and DNNs. Silicon and paracetamol crystals are used as showcases to demonstrate the computational efficiency.
Highlights
The principle of Raman spectroscopy is based on the phenomena of inelastic light scattering from various materials
Both scitation.org/journal/adv with the trained deep neural networks (DNNs) polarizability model, and the Deep potential molecular dynamics (DPMD) Raman spectrum is evaluated with the DNN trajectories
We find that the prediction error of the neural networks (NN) model for the polarizabilities is within 5% and the learning performances are similar among the six components, which is comparable with the symmetry-adapted Gaussian process regression (GPR) method used in Ref. 21
Summary
The principle of Raman spectroscopy is based on the phenomena of inelastic light scattering from various materials. Raman spectra can be simulated from the timecorrelation function of polarizabilities, which are calculated at each atomic configuration as obtained from the ab initio molecular dynamics (AIMD) trajectories. We recently developed and implemented a real-space formalism for DFPT15 in the all-electron, fullpotential, numerical atomic orbitals based on the Fritz Haber Institute ab initio molecular simulations (FHI-aims) package, which takes advantage of the inherent locality of the basis set to achieve a numerically favorable scaling. Such a real-space DFPT has been applied in lattice dynamics calculations and when computing polarizabilities, dielectric constants, harmonics, and anharmonic Raman spectra, which has demonstrated good computational accuracy, computational efficiency, and parallel scalability.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have