Combined High-Throughput and Machine Learning Approach for Prediction of Lattice Thermal Conductivity

Search for materials via explicit evaluation of thermal conductivity (κl) either experimentally or computationally is very challenging. We carried out high-throughput screening on a dataset containing total of 2691 binary, ternary, and quaternary compounds. The κl values of 120 dynamically stable and nonmetallic compounds are calculated. Among these, 11 ultrahigh and 15 ultralow κl materials are identified. For the machine learning prediction models, the descriptor set is usually tuned via conventional algorithms such as least absolute selection and shrinkage operator (LASSO). However, we generated an extensive property map from high-throughput calculations to design a minimal set of descriptors directly related to the physics of κl. These simple descriptors are 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 room temperature log-scaled κl with a very small root mean square error. The superior performance of the ML model can ensure a reliable and accelerated search for a multitude of low and high κl materials.

Reference: Chem. Mater. 31, 14, 5145-5151, 2019