Dissimilar thermal transport properties in κ-Ga2O3 and β-Ga2O3 revealed by homogeneous nonequilibrium molecular dynamics simulations using machine-learned potentials

The lattice thermal conductivity (LTC) of Ga2O3 is an important property due to the challenge in the thermal management of high-power devices. In this work, we develop machine-learned neuroevolution potentials (NEPs) for single-crystalline β−Ga2O3 and κ−Ga2O3 and demonstrate their accuracy in modeling thermal transport properties. Combining NEP-driven homogeneous non-equilibrium molecular dynamics simulations with tensor analysis, we determine the spatial distributions of LTCs for two Ga2O3 crystals, showing dissimilar thermal behaviors. Specifically, β−Ga2O3 shows isotropic thermal transport properties, with the LTCs along [100], [010], and [001] directions being predicted to be 10.3±0.2, 19.9±0.2, and 12.6±0.2 W/(m K), respectively, consistent with previous experimental measurements. For κ−Ga2O3, our predictions suggest nearly isotropic thermal transport properties, with the LTCs along [100], [010], and [001] being estimated to be 4.5±0.1, 3.9±0.1, and 4.0±0.1 W/(m K). The reduced LTC of κ−Ga2O3 vs β−Ga2O3 stems from its restricted low-frequency phonons up to 5 THz. Furthermore, we find that the β phase exhibits a typical temperature dependence slightly stronger than ∼T−1, whereas the κ phase shows a weaker temperature dependence, ranging from ∼T−0.5 to ∼T−0.7.