AbstractLayered IV-V-VI semiconductors have immense potential for thermoelectric (TE) applications due to their intrinsically ultralow lattice thermal conductivity. However, it is extremely difficult to assess their TE performance via experimental trial-and-error methods. Here, we present a machine-learning-based approach to accelerate the discovery of promising thermoelectric candidates in this chalcogenide family. Based on a dataset generated from high-throughput ab initio calculations, we develop two highly accurate-and-efficient neural network models to predict the maximum ZT (ZTmax) and corresponding doping type, respectively. The top candidate, n-type Pb2Sb2S5, is successfully identified, with the ZTmax over 1.0 at 650 K, owing to its ultralow thermal conductivity and decent power factor. Besides, we find that n-type Te-based compounds exhibit a combination of high Seebeck coefficient and electrical conductivity, thereby leading to better TE performance under electron doping than hole doping. Whereas p-type TE performance of Se-based semiconductors is superior to n-type, resulting from large Seebeck coefficient induced by high density-of-states near valence band edges.
High-throughput ab initio calculations on dielectric constant and band gap of non-oxide dielectrics
