Machine learning assisted phonon dynamics beyond harmonic limit

Exploring phonon dynamics beyond the harmonic limit is crucial for understanding a wide range of physical phenomena that cannot be captured by the harmonic approximation. We present a machine learning (ML) force-field model to perform anharmonic phonon dynamics with the efficiency of ML and quantum accuracy of first principles. We use the nonperturbative frozen phonon formulation method to capture the interatomic forces at finite temperatures using density functional theory (DFT), where configurations of nuclei in a supercell were sampled employing a stochastic Monte Carlo algorithm. A trained deep neural network accurately predicts the Hellman–Feynman (FH) forces of atoms in the supercell of sampled phonon configurations, therefore, bypassing the time-consuming ab initio calculations. A symmetry-invariant descriptor is developed using group-theoretical methods for the supercell, incorporating point-group symmetry of the atomic system into the ML model. Our ML model is benchmarked by accurately producing phonon dispersion where acoustic sum rule (ASR) was preserved by the ML itself. Our work demonstrates a promising ML method for molecular dynamics (MD) simulations in orders of magnitude larger crystals and achieving quantum accuracy beyond that of traditional MD force-field models.