Defect-limited thermal transport in AlN using pretrained machine-learning interatomic potentials

Abstract

Aluminum nitride (AlN) is an important thermal management material whose high lattice thermal conductivity is strongly suppressed by oxygen impurities. We investigate phonon scattering by oxygen-related defects using pretrained universal machine-learning interatomic potentials (MLIPs), molecular dynamics (MD), and phonon Boltzmann transport calculations. Several pretrained MLIPs are benchmarked against density functional theory for phonon dispersions and pristine thermal conductivity. To balance accuracy and computational speed, we use a fine-tuned version of the compact SevenNet-Nano model for MD simulations. Monte Carlo annealing supports the formation of bound $V_{\mathrm{Al}}(\mathrm{O_N})_3$ complexes, whose scattering differs from that of their isolated constituents. Defect scattering rates extracted from excess spectral energy density (SED) linewidths agree reasonably with harmonic $T$-matrix predictions at low oxygen contents, supporting the independent-scatterer approximation. Incorporating these rates into an iterative Boltzmann transport equation with Bose--Einstein statistics yields thermal conductivities comparable to experimental values and captures their observed decrease with oxygen content. At high concentrations, however, the scattering rate deviates from linear scaling with oxygen content, suggesting limitations of independent-defect and pristine-phonon descriptions. These results demonstrate an efficient approach using pretrained MLIPs to quantify defect-limited thermal transport and provide insights into impurity effects beyond the dilute-defect approximation.