Phonon spectrum analysis using Gaussian process regression

Abstract

This work employs Machine Learning (ML) tools to generate phonon spectra from ab initio Molecular Dynamics (MD) simulations. For this end, an ML algorithm is trained using two-body and Smooth Overlap of Atomic positions (SOAP) descriptors on the MD trajectory of 54 carbon atoms in a solid cubic diamond structure calculated employing Density Functional Theory (DFT) in a canonical ensemble, with energies and forces as targets.The accuracy of the ML force field is studied using the following two approaches. Firstly, MD simulations using ML methods starting from the same geometry as the original DFT trajectory are computed and compared to the original runs, together with simulations where the ML algorithm was trained with smaller DFT datasets to inspect robustness against data reduction.Secondly, the trained ML algorithm is also used to calculate forces for new systems, specifically finite displacements from equilibrium to compute the second derivatives of the energy surface from which phonon spectra can be generated. 250 datapoints were sufficient to obtain accurate phonon dispersions.