Native point defects are fundamental building blocks of defect engineering in thermoelectric materials. However, it remains unclear which changes induced by point defects reduce lattice thermal conductivity ${\kappa }_{\text{l}}$ at the atomic level. This study develops a neural-network potential (NNP) trained on density functional theory data for lead telluride (PbTe) and applies it to equilibrium and perturbed molecular dynamics (PMD) simulations. Isolated Pb and Te vacancies, antisites, interstitials and Schottky pairs are examined using sufficiently large supercells. The NNP is shown to reproduce harmonic and anharmonic vibrational properties and to accurately predict potential energy and atomic forces at finite temperatures. NNP-based PMD simulations reveal that local thermal transport near defects varies significantly depending on the defect type, exhibiting the lowest ${\kappa }_{\text{l}}$ for interstitial Pb. Atom-resolved ${\kappa }_{\text{l}}$ of point defects and their neighboring atoms correlate with the integrated absolute difference in vibrational density of states between those atoms and bulk atoms. This trend appears even within the harmonic approximation, which neglects anharmonic contributions. This suggests that the difference between the harmonic potential energies of pristine and defect-containing PbTe plays a central role in reducing ${\kappa }_{\text{l}}$, by modifying the force field around point defects through the second-order force constant. In contrast, atomic displacements around point defects show a weak correlation with atom-resolved ${\kappa }_{\text{l}}$ and thus may not be a suitable metric for predicting ${\kappa }_{\text{l}}$ of defect-containing PbTe, although relevant quantities are assumed to affect ${\kappa }_{\text{l}}$ in conventional analytical models. The present NNP enables us to analyze physical quantities accessible via molecular simulations with high accuracy and to understand the underlying mechanisms of thermal transport reduction by point defects, without relying on analytical models.

Close

PrevNext

View figure in article

• Lattice dynamics
• Point defects
• Thermal conductivity
• Density functional theory
• Machine learning
• Molecular dynamics
• Neural network simulations