Lattice dynamics simulation using machine learning interatomic potentials

V. V. Ladygin, P. Yu Korotaev, A. V. Yanilkin, A. V. Shapeev

Research output: Contribution to journalArticlepeer-review

28 Citations (Scopus)

Abstract

In this work, a machine-learning approach was applied to obtaining an interatomic potential for lattice dynamics properties calculation with accuracy close to the one of density functional theory (DFT). The computational efficiency of the potential allows one to access large time and length scales through molecular dynamics simulations. The use of active learning and an automatic training procedure greatly reduces the number of quantum-mechanical calculations for the training set. In order to estimate the accuracy of the obtained potentials, four materials Al, Mo, Ti and U with different phonon and thermodynamic properties were investigated. Phonon properties were calculated using the temperature dependent effective potential method. The potentials reproduce not only harmonic behavior but also anharmonic effects, as shown by the calculation of the third-order force constants. We found that machine-learning potentials reproduce quantum-mechanical data with high accuracy. Furthermore, the vibrational density of states was obtained via velocity autocorrelation function integration, which would be infeasible in direct quantum-mechanical simulations.

Original languageEnglish
Article number109333
JournalComputational Materials Science
Volume172
DOIs
Publication statusPublished - 1 Feb 2020

Keywords

  • Anharmonicity
  • Lattice dynamics
  • Machine learning
  • Moment tensor potentials

Fingerprint

Dive into the research topics of 'Lattice dynamics simulation using machine learning interatomic potentials'. Together they form a unique fingerprint.

Cite this