Advancing thermoelectric materials discovery through semi-supervised learning and high-throughput calculations

Abstract

The data-driven machine learning technique is widely used to assist in accelerating the design of thermoelectric materials. In this study, we proposed a positive and unlabeled learning (PU learning) method, a semi-supervised learning, to train a classifier to distinguish the positive samples from the unlabeled samples, in which the positive class was labeled by matching the formulas in our dataset with the published article titles. The probabilities that the unlabeled materials belong to the positive class were predicted by PU learning, and 40 candidate thermoelectric materials were determined. The transport properties were calculated by high-throughput first-principles calculations, among which 8 p-type and 12 n-type materials have the maximum theoretical zT values greater than 1. Specifically, a series of AX2 binary compounds, (Cd/Zn)(GaTe2)2 ternary compounds, and Cs(Dy/Ho/Tb)2Ag3Te5 quaternary compounds deserve further investigations in the future.