Accelerating thermal conductivity prediction through machine-learning: Two orders of magnitude reduction in phonon-phonon scattering rates calculation

Abstract

The calculation of phonon-phonon scattering rates is a computational bottleneck in first-principles based phonon thermal conductivity prediction of materials. Here, we report a machine learning approach for phonon scattering rates prediction which is capable of predicting thermal conductivity by employing only 5 % of the total computational cost. We test this approach on more than 200 diverse materials and found that when this approach is combined with that of Guo et al. [npj Computational Materials 9, 95 (2023)] for phonon-phonon linewidth prediction, the cumulative speed-up is more than two orders of magnitude while the accuracy of thermal conductivity prediction is preserved to within 10 %. This drastic speedup is translated into computational time reduction for phonon scattering rates calculation from more than 60,000 cpu-hours to less than 500 cpu-hours for considered 230 materials.