Revealing phonon signature of dislocations in silicon carbide using machine-learning interatomic potential

Abstract

Defects are the main performance killer in silicon carbide (SiC) power devices. Among various defect types, dislocations are particularly important, as they affect device reliability. However, first-principles modeling of dislocations is computationally challenging due to their complex, extended structure and topological nature. To overcome this difficulty, we develop a neuroevolution potential (NEP) to enable accurate and large-scale lattice dynamics simulations for defect-containing SiC. To circumvent the difficulty of direct dislocation calculation, the NEP is trained on a first-principles dataset generated by iteratively incorporating various point defects, line defects, and surface structures that are computationally tractable. The resulting NEP reproduces phonon spectra in crystalline and dislocation-containing SiC, indicating its transferability. With this potential, we analyze the phonon characteristics around dislocations in 4H-SiC. Our results reveal localized vibrational modes around dislocation cores, and phonon frequency shifts away from the cores due to dislocation-induced strain fields. This work may facilitate the identification of dislocation phonon signatures and delivers a machine-learning potential that overcomes the computational limitations for large-scale SiC defect simulations.