Machine learning force field based phonon dispersion prediction

Abstract

First-principles calculations on phonon dynamics using density functional theory (DFT) have proven powerful in estimating the phonon dispersion of crystalline structures. However, it remains a challenging task for defective structures due to the computational cost. The main computational bottleneck of the phonon calculation is obtaining the interatomic force constants in many supercells with different configurations of displacements. Here, we employed a machine learning-based force fields (MLFFs) to accelerate DFT calculations of interatomic force constants of Si-doped HfO₂. We find that the specific phonon band originated from ferroelectric phase disappears, and imaginary modes are enhanced upon the introduction of a 10 % concentration of Si dopants, which is in good agreement with experimental results. This work demonstrates that MLFFs can be a promising application for predicting the phonon dispersion of both crystalline and defective structures.