High throughput generation of high-zT thermoelectrics with precise stoichiometric controls

The pursuit of designing high-performance thermoelectric (TE) materials is often hindered by the complex coupling between structural composition and transport properties. In this study, we present TE-Diffusion, an inverse design framework combining a generative diffusion model with a novel three-channel doping representation designed to capture subtle stoichiometric variations intrinsic to complex doping. Comprehensive evaluations demonstrate that TE-Diffusion significantly outperforms traditional generative adversarial networks (GANs). Specifically, the generated materials achieve a thermodynamic stability rate of 95% and a zT alignment accuracy of 82%, in addition to high doping effectiveness (71%) and electrical neutrality (52%). The model effectively extracts and learns fine-grained empirical correlations between doping stoichiometry and zT performance of high-zT systems from the experimental dataset, without explicit programming of doping rules. Leveraging this capability, we identified 100 previously unreported potential high performance candidates. Notably, the screened material Si_0.03Ge_0.89Sb_0.08Te was validated via first-principles calculations to possess a maximum zT of 2.517 at 800 K. Calculations reveal that this high performance stems from the optimization of carrier concentration and the favorable matching of transport properties exhibiting bipolar behavior. TE-Diffusion effectively reduces the search space and establishes a robust composition-structure-performance mapping, providing a new paradigm for the targeted design of complex doped materials.