Lattice Thermal Conductivity from First Principles and Active Learning with Gaussian Process Regression

Abstract

The lattice thermal conductivity ($\kappa_{\ell}$) is a key materials property in power electronics, thermal barriers, and thermoelectric devices. Identifying a wide pool of compounds with low $\kappa_{\ell}$ is particularly important in the development of materials with high thermoelectric efficiency. The present study contributed to this with a reliable machine learning (ML) model based on a training set consisting of 268 cubic compounds. For those, $\kappa_{\ell}$ was calculated from first principles using the temperature-dependent effective potential (TDEP) method based on forces and phonons calculated by density functional theory (DFT). 238 of these were preselected and used to train an initial ML model employing Gaussian process regression (GPR). The model was then improved with active learning (AL) by selecting the 30 compounds with the highest GPR uncertainty as new members of an expanded training set. This was used to predict $\kappa_{\ell}$ of the 1574 cubic compounds in the \textsc{Materials Project} (MP) database with a validation R2-score of 0.81 and Spearman correlation of 0.93. Out of these, 27 compounds were predicted to have very low values of $\kappa_{\ell}$ ($\leq 1.3$ at 300~K), which was verified by DFT calculations. Some of these have not previously been reported in the literature, suggesting further investigations of their electronic thermoelectric properties.