Phonon-Mediated Thermal Transport in Nanocrystalline Silicon Using Machine-Learning Interatomic Potentials

Abstract

Understanding phonon-mediated heat transport in structurally complex materials remains a central challenge for next-generation electronic and nanomechanical devices, where grain boundaries and interfacial disorder strongly limit thermal dissipation. Although classical interatomic potentials enable large-scale simulations, their limited transferability can lead to inaccuracies in vibrational properties and interfacial phonon scattering. In this work, we develop a machine learning-based framework for modeling thermal transport in bulk and nanocrystalline silicon by combining Gaussian approximation potential and multi-atomic cluster expansion models with lattice-dynamical calculations and molecular dynamics. Harmonic and anharmonic force constants derived from machine-learning interatomic potentials (MLIPs) are used within a unified Phonopy/Phono3py workflow to compute phonon dispersions, lifetimes, and lattice thermal conductivity, providing an internally consistent description of vibrational properties. In nanocrystalline silicon, non-equilibrium molecular dynamics simulations directly quantify the thermal boundary resistance associated with grain boundaries and reveal its sensitivity to interfacial roughness and the underlying interatomic description. Compared with the Stillinger-Weber and Tersoff potentials, the MLIPs provide a quantitatively accurate and internally consistent description of bulk and interfacial phonon transport, enabling better predictive modeling of nanoscale thermal transport in low-dimensional materials.