Phonon local non-equilibrium at Al/Si interface from machine learning molecular dynamics

Abstract

All electronics are equipped with metal/semiconductor junctions, resulting in resistance to thermal transport. The nanoscale phononic complexities, such as phonon local non-equilibrium and inelastic scattering, add to the computational or experimental characterization difficulty. Here, we use a neural network potential (NNP) trained by ab initio data, demonstrating near-first-principles precision more accurate than classical potentials used in molecular dynamics (MD) simulations to predict thermal transport at the Al/Si interface. The interfacial thermal conductance of 380 ± 33 MW / m 2 K from our NNP-MD simulations is in good agreement with the previous experimental consensus while considering the crucial physics of interfacial bonding nature, phonon local non-equilibrium, and inelastic scattering. Furthermore, we extract phonon mode insights from the NNP-MD simulations to reveal the decrease in local non-equilibrium of the longitudinal acoustic modes at the Al/Si interface. Our work demonstrates the utility of a machine learning MD to predict and extract accurate insights about interfacial thermal transport.