Higher-order phonon scattering and lattice thermal conductivity prediction via machine learning

Abstract

Higher-order phonon scattering processes, such as three-phonon (3ph) and four-phonon (4ph) interactions, critically influence lattice thermal conductivity (κ_L). However, calculating κ_L with both 3ph and 4ph scattering is challenging due to the intrinsic complexity of high-order interactions and the extensive scattering phase space. In this work, we develop an interpretable machine learning framework that provides fast and accurate predictions of κ_L by accounting for both 3ph and 4ph scattering. We use SHAP analysis to show that the maximum acoustic phonon frequency (f_a,max) and the Grüneisen parameter are the most important features. The f_a,max acts as a key factor that regulates thermal transport by controlling the available scattering channels. Our results show that large phonon bandgaps and flat optical branch dispersions enhance 4ph combination and redistribution processes, respectively. These factors lead to dominant 4ph scattering in certain materials. Moreover, we propose a dual-label regression model that predicts both the κ_L at 300 K and the power-law exponent. This model simulates the variation of κ_L within a specific temperature range instead of predicting the value at a single temperature point. This work not only provides an efficient tool for high-throughput material screening and thermal management design but also provides new insights into higher-order phonon physics.