Phonon Thermal Transport in Bi2Te3 from a Deep-Neural-Network Interatomic Potential

${\mathrm{Bi}}_{2}{\mathrm{Te}}_{3}$ is a widely used thermoelectric material with strong anharmonicity. Determination of its thermal conductivity requires consideration of the high-order phonon scattering process, which makes it extremely time consuming and challenging to accurately calculate its thermal conductivity by obtaining high-order force constants based on density-functional theory. In this work, a deep-neural-network potential is developed to reproduce phonon dispersion and predict the lattice thermal conductivity of ${\mathrm{Bi}}_{2}{\mathrm{Te}}_{3}$. The equilibrium molecular dynamics simulations combined with this potential are performed to calculate the lattice thermal conductivity and the results nicely match the experimental values. Meanwhile, we find the generalized gradient approximation with the DFT-D3 functional can accurately reproduce the experimental lattice constants of ${\mathrm{Bi}}_{2}{\mathrm{Te}}_{3}$ and provide a description of the phonon dispersion in ${\mathrm{Bi}}_{2}{\mathrm{Te}}_{3}$ as well as the local density approximation. Furthermore, we explore the influence of the native point defects on thermal conductivity, and find that $\mathrm{Te}$ vacancies have the most significant effect on the reduction of thermal conductivity, owing to the appreciable inhibition of phonon propagation speed by $\mathrm{Te}$ vacancies, and the additional scattering among original low-frequency optical phonons and the fresh low-frequency optical phonons moving downward from high frequency region, which provides some theoretical guidance for reducing thermal conductivity in experimental research.