Robust computational algorithm for temperature-dependant phonon frequencies and lifetimes calculation beyond the perturbation approximation

Recent advancements of machine-learning interatomic potentials provide new opportunities to accurately simulate lattice vibrations of complex materials using a computer cluster of moderate size. We have implemented a robust algorithm to compute temperature-dependent phonon frequencies and lifetimes of crystals based on the molecular dynamics (MD) simulations of atomic velocity-velocity correlation functions. This method calculates effects of all orders of lattice anharmonicity beyond the perturbation approximation. A robust reciprocal q-space symmetrization algorithm significantly reduces numerical fluctuations found in the MD simulations with thousand-atom material systems over hundreds of picoseconds. We have adopted this algorithm to predict phonon frequencies and lifetimes of crystalline silicon from 300K to 1500K. Our MD simulation results are compared to results from perturbation calculations based on the bare-phonon spectra and third order lattice anharmonicity, as well as the latest neutron-scattering data. This comparison shows that the onset of the break-down of single relaxation time approximation starts around 600K in many optical phonon modes. Current developments are in the works to accurately predict the lattice thermal conductivity from these temperature-dependent phonon properties. This approach can be readily extended to structurally more complex materials, such as nano-materials or materials interfaces.