Machine learning discovery of medium-entropy thermoelectric materials with ultralow lattice thermal conductivity

Abstract

Entropy engineering has proven to be an effective approach for lowering lattice thermal conductivity and enhancing thermoelectric performance. Medium-entropy materials, in particular, offer a desirable balance between electronic transport optimization and phonon scattering, attributed to their intrinsic structural disorder. However, their complex compositions and vast configurational space pose significant challenges to conventional experimental exploration, leading to high costs and limited efficiency in material discovery. To overcome these limitations, we developed a data-driven discovery framework to efficiently explore the vast composition space and predict high-performance thermoelectric materials within the medium-entropy regime. The model was trained using experimentally reported thermoelectric data and applied to materials sourced from the Materials Project. Among the screened candidates, In₂Bi₄Pb₄Se₁₃ was identified as a promising p-type thermoelectric compound. Subsequent DFT calculations predicted a peak zT of 1.37 at 800 K, highlighting the promising thermoelectric performance about In₂Bi₄Pb₄Se₁₃. This high zT arises from the complex crystal structure of In₂Bi₄Pb₄Se₁₃, which induces substantial lattice distortion, strong phonon anharmonicity, and enhanced phonon scattering. Notably, the material exhibits an ultralow lattice thermal conductivity of 0.22 W·m^-1·K^-1 at 800 K. This study provides a robust strategy for exploring complex alloy systems.