Physics and Explainable Descriptor Informed Machine Learning Probing Lower Bound of Lattice Thermal Conductivity

The advent of artificial intelligence (AI) and machine learning (ML) has revolutionized many aspects of modern science and technology and has sparked significant interest in the material science community in recent years. Despite some early deployment of AI/ML in thermal science area, the power of AI has not been maximized. Existing ML methods for predicting phonon properties of crystals are limited to either small amount of training data or a material-to-material basis, primarily due to the exponential scaling of model parameters with the number of atomic species or elements. This renders high-throughput infeasible when facing large-scale new materials. Here, a quantitative analysis of 4,700 lattice thermal conductivity (LTC) by density functional theory (DFT) spanning all seven crystallographic groups correlated with 12 customized structural and phonon descriptors provides a deep insight towards the two-channel thermal transport. We discover an important exchange between diffuson and propagon transport leading to an ultralow LTC less than 1 W/mK. The trend of two-channel phonon transport properties is reproduced on a large dataset of 31,058 structures by traditional ML models. With trained ML models, ~20 materials were filtered out and further confirmed by DFT to possess ultralow LTC near the lower bound of 0.1 W/mK. Such an understanding of two-channel thermal transport will be useful to tailor materials from the phonon perspective, promoting the discovery of near-future candidates for pushing the LTC to even lower limit.