High Temperature Thermal Conductivity by Machine Learning driven Atomistic Simulations

Understanding thermal transport is important for many high temperature applications. However, the conventional ab initio lattice dynamics method ignored the temperature dependence of force constants, thus cannot be applied to energetically unstable crystals including body centered cubic phase of group IV metals, CmCm phase of SnSe and β-Cu2Se, which all display soft phonon modes at static limit while they are all stable phases at high temperatures. Conventionally, this instability problem is approached by iteratively solving the high temperature harmonic force constants, but predicting thermal conductivity requires third- and even forth-order anharmonic force constants. In addition, the accuracy of the force constants are significantly affected by the artificial truncation of the Taylor expansion of the potential energy surface. Using machine learning, we developed interatomic potentials for these energetically-unstable high temperature phases to study the thermal conductivity. Using machine learning-driven molecular dynamics, the soft phonon modes are observed to renormalize to a positive frequency and the thermal conductivity is predicted using Green-Kubo method including all orders of phonon anharmonicity.