Molecular Dynamics Simulations of Lattice Thermal Conductivity with Machine-Learning Anharmonic Interaction

The lattice thermal conductivity of crystals with strong anharmonic interaction is of particular interest, which cannot be adequately evaluated by linearized Boltzmann transport equation. In this work, we adopt compressive sensing, a machine learning technique to obtain high-order force constants from a small amount of training data. By considering strong heredity effects, we show that the dominant anharmonic interactions are short ranged. This largely shrinks the number of force constants needed in the expansion. To calculate the lattice thermal conductivity, molecular dynamics simulations that include anharmonic interactions up to the sixth order are performed. Test results for Si and NaCl will be presented. The anharmonic interaction in NaCl is found to be more significant than that in Si, which is consistent with the fact that NaCl exhibits a smaller thermal conductivity at above 100 K. Simulation results for highly anharmonic thermoelectric materials SnSe and GeSe that exhibit very low thermal conductivity will also be discussed.

*In collaboration with Jing Wang, Jie-Cheng Chen, and Martin Callsen