Machine learning potential-assisted design for thermoelectric performance in anharmonic CsPbBr3 and CsPbI3

Abstract

The structural stability and thermoelectric properties of halide perovskites were investigated using a combination of machine learning approaches and density functional theory (DFT). A neuroevolution potential (NEP) was trained via a feedforward neural network and subsequently employed in molecular dynamics (MD) simulations to probe the phase stability of CsPbBr₃ and CsPbI₃. The simulations reveal a temperature-driven phase transition from the harmonic γ-phase to the anharmonic β- and α-phases. Notably, the high-temperature α-phase combines low thermal conductivity with structural stability, suggesting its promise as a thermoelectric material. The self-consistent phonon (SCPH) method addresses the limitations of the harmonic approximation by renormalizing phonon frequencies, thereby eliminating the unphysical imaginary modes. Accurate electron–phonon interactions were critical for carrier transport analysis. Carrier mobility, calculated via the Boltzmann transport equation (BTE) with Wannier-interpolated electron–phonon coupling matrices, was found to be overestimated compared to experiments. This discrepancy is attributed to additional carrier scattering from dopants in n- or p-type semiconductors under operating conditions. At 600 K, the maximum predicted ZT for α-CsPbI₃ is 1.59 under ideal conditions and 0.65 under realistic conditions, corresponding to thermoelectric conversion efficiencies of 10.0% and 5.3% in power generation devices.