Identification of Crystalline Materials with Ultra-Low Thermal Conductivity Based on Machine Learning Study
Xinming Wang, Shuming Zeng, Zhuchi Wang, Jun Ni
Abstract
We have developed machine learning (ML) models to predict the thermal conductivity (κ). By analyzing the impact of different choice of descriptors and various algorithms, we find that the XGBoost algorithm based on the descriptors of crystal structural and compositional information can accurately predict the thermal conductivity with an average mean absolute error of 2.13 W m–1 K–1. To understand and interpret the model, we have analyzed the feature importance of the ML model. Among the top five important features, we find that the thermal conductivity is clearly partitioned when projected to the avg(ΔHatomic)–Vpc and Vpa–Vpc dual-descriptor spaces, which exhibit simple heuristic rules useful for searching and designing materials with favorable thermal conductivity. We employ the learned models on the materials in the entire Inorganic Crystallographic Structure Database (ICSD) and find that the heavy elements like Cs, Au, Hg, Tl, and Pb are helpful to reduce κ. Using the prediction results on the ICSD, we have screened the materials with low κ like BiTe2Tl and Cl2CsI and validated them using first-principles calculations, which could indicate prospective thermoelectric materials.