ML-assisted acceleration of Phonon Scattering Simulation: Methods and Applications

Phonon scattering is a fundamental mechanism that is important in thermal transport and deeply related to thermal conductivity and radiative property, but experimental measurements or first principles calculations including three-phonon and four-phonon scattering are expensive or even unaffordable, preventing the large-scale high-throughput screening of materials. The acceleration of phonon scattering simulations while keeping a high accuracy has been a long-standing open question.
To overcome the computational barriers, we first present a machine learning approach that can predict phonon scattering rates and thermal conductivity with experimental and first principles accuracy. The success of our approach is enabled by mitigating computational challenges associated with the high skewness of phonon scattering rates and their complex contributions to the total thermal resistance. Transfer learning between different orders of phonon scattering can further improve the model performance. Our surrogates offer up to two orders of magnitude acceleration compared to first principles calculations and would enable large-scale thermal transport informatics.
Inspired by the first machine learning model, we further developed a statistical sampling method to accelerate the phonon scattering. This is done by estimating scattering rates from a small sample of scattering processes using maximum likelihood estimation. The calculation of scattering rates and associated thermal conductivity and radiative properties are dramatically accelerated by three to four orders of magnitude. We also derive the confidence interval of our estimation, which is useful for choosing a proper sample size. The accuracy and efficiency of our approach make it ideal for the high-throughput screening of materials for thermal and optical applications. The work is incorporated as a new feature within the FourPhonon open-source package.
Besides acceleration, our model also enables us to use an unprecedented q-mesh (discretized grid in the reciprocal space) to study the lattice dynamics of materials, which was not possible before due to the high computational cost. By using a dense q-mesh for calculating four-phonon scattering of silicon (Si), we achieve a converged thermal conductivity value that agrees much better with experiments. Besides, we did a sophisticated theoretical study of the lattice thermal conductivity and phonon linewidth in bulk hexagonal boron nitride (h-BN). Our simulation revealed the significant or even leading contributions of four-phonon scattering and phonon renormalization. The simulation results showed excellent agreement with the experimental results.