Machine-learning potential for phonon transport in AlN with defects in multiple charge states

Abstract

Understanding phonon transport properties in defect-laden AlN is important for their device applications. Here, we construct a machine-learning potential to describe phonon transport with $ab$ $initio$ accuracy in pristine and defect-laden AlN, following the template of Behler-Parrinello-type neural network potentials (NNPs) but extending them to consider multiple charge states of defects. The high accuracy of our NNP in predicting second- and third-order interatomic force constants is demonstrated through calculations of phonon bands, three-phonon anharmonic, phonon-isotope and phonon-defect scattering rates, and thermal conductivity. In particular, our NNP accurately describes the difference in phonon-related properties among various native defects and among different charge states of the defects. They reveal that the phonon-defect scattering rates induced by V$_{N}^{3+}$ are the largest, followed by V$_{Al}^{3-}$, and that V$_{N}^{1+}$ is the least effective scatterer. This is further confirmed by the magnitude of the respective depressions of the thermal conductivity of AlN. Our findings reveal the significance of the contribution from structural distortions induced by defects to the elastic scattering rates. The present work shows the usefulness of our NNP scheme to cost-efficiently study phonon transport in partially disordered crystalline phases containing charged defects.