A REVIEW OF RESEARCH ON ANALYSIS, TESTING, AND MACHINE LEARNING OF INTERATOMIC POTENTIALS USING DELOCALIZED NONLINEAR VIBRATIONAL MODES

10.25712/ASTU.1811-1416.2025.04.002

Abstract

Molecular dynamics (MD) is a powerful tool for materials research, allowing one to work with millions or more atoms. However, in molecular dynamics, the quality of the interatomic potential is crucial. To analyze and test a wide range of potentials, we propose using delocalized nonlinear vibrational modes (DNVMs). DNVMs are exact solutions to the equation of atomic motion obtained based on the symmetry of the structure and, unlike analysis based solely on phonon modes, allow vibrations over a wide range of amplitudes and include both the linear (phonon) and nonlinear parts of the vibration. This approach allows testing potentials from both linear and nonlinear physics perspectives. Therefore, this paper presents a review of two studies investigating DNVMs in BCC tungsten. One of them compares existing potentials with respect to ab initio data, while the other already presents a machine-trained potential using DNVM and shows the difference in DNVM reproducibility over a wide range of amplitudes.