Anharmonic lattice dynamics in large thermodynamic ensembles with machine-learning force fields: the breakdown of the phonon quasiparticle picture in CsPbBr3.

The harmonic approximation is a very powerful method for describing phonon dispersion relations. However, when the temperature is raised and the potential energy landscape exhibits more anharmonicity, the approximation fails to capture all crystal lattice dynamics properly. Here we study, for the first time, the phonon dispersion of a complex "Dynamic Solid" with machine-learning force fields, by simulating the dynamic structure factor (DSF) $S(\mathbf{q},\omega)$ and the projected velocity autocorrelation function (PVACF) trough large-scale molecular dynamics. These force fields have near first-principles accuracy and the linear scaling computational cost of classical potentials. To asses the strengths and weaknesses of the three methods we start with an analysis based on the classical Morse potential. Hereafter, the methods are applied to the inorganic perovskite: CsPbBr$_{3}$. This perovskite serves as an archetypal example of a wider class of novel perovskite solar-cell materials. Imaginary modes in the harmonic picture of the CsPbBr$_{3}$ structure are absent in the calculated DSF and PVACF, indicating a dynamic stabilization of the crystal. The anharmonic nature of the potential and the presence of rattling Cs$^{+}$ cations, result in the breakdown of the phonon quasi-particle picture. A more realistic picture for CsPbBr$_{3}$ emerges, presenting the phonons by a continuum of accessible frequencies.