Thermal conductivity of selenium crystals based on machine learning potentials.

Abstract

Selenium, as an important semiconductor material, exhibits significant potential for understanding lattice dynamics and thermoelectric applications through its thermal transport properties. Conventional empirical potentials are often unable to accurately describe the phonon transport properties of selenium crystals, which limits in-depth understanding of their thermal conduction mechanisms. To address this issue, this study developed a high-precision machine learning potential (MLP), with training datasets generated via ab initio molecular dynamics simulations. Validation demonstrated that the phonon dispersion relations calculated by the MLP showed excellent agreement with density functional theory results. Using this potential, we systematically investigated the thermal transport properties of trigonal (t-Se) and monoclinic selenium (m-Se). The results demonstrate that t-Se exhibits higher thermal conductivity. Phonon density of states analysis reveals that this originates from its chain-like structure (where intrachain atoms are connected by strong covalent bonds while interchain interactions occur through weaker van der Waals forces), which enables stronger thermal transport compared to the ring-like structure of m-Se. The electronic structure calculations further reveal that the bandgap of t-Se is significantly smaller than that of m-Se (by approximately 0.7 eV). Therefore, although t-Se exhibits a relatively large lattice thermal conductivity, its higher electrical conductivity (σ) (six orders of magnitude difference) and Seebeck coefficient compensate for this disadvantage, enabling t-Se to achieve a high ZT (σ/κ ratio) and superior thermoelectric potential.