Coupling the High-Throughput Property Map to Machine Learning for Predicting Lattice Thermal Conductivity

Low thermal conductivity materials are crucial for applications such as thermoelectric conversion of waste heat to useful energy and thermal barrier coatings. On the other hand, high thermal conductivity materials are necessary for cooling electronic devices. However, search for such materials via explicit evaluation of thermal conductivity either experimentally or computationally is very challenging. Here, we carried out high-throughput ab initio calculations, on a dataset containing 195 binary, ternary, and quaternary compounds. The lattice thermal conductivity κ_l values of 120 dynamically stable and nonmetallic compounds are calculated, which span over 3 orders of magnitude. Among these, 11 ultrahigh and 15 ultralow κ_l materials are identified. An analysis of generated property map of this dataset reveals a strong dependence of κ_l on simple descriptors, namely, maximum phonon frequency, integrated Grüneisen parameter up to 3 THz, average atomic mass, and volume of the unit cell. Using these descriptors, a Gaussian process regression-based machine learning (ML) model is developed. The model predicts log-scaled κ_l with a very small root mean square error of ∼0.21. Comparatively, the Slack model, which uses more involved parameters, severely overestimates κ_l. The superior performance of our ML model can ensure a reliable and accelerated search for multitude of low and high thermal conductivity materials.