Development of deep potential models for silicon carbide and investigation of damage evolution mechanisms

Traditional empirical interatomic potentials exhibit limited accuracy in describing the high-temperature structural response of silicon carbide (SiC), while first-principles simulations remain computationally expensive. In this work, a deep neural network interatomic potential for SiC was constructed based on the density functional theory (DFT) datasets. An attention-enhanced descriptor was introduced to improve the representation of bond-length and bond-angle correlations. A concurrent learning workflow based on the deep potential generator (DP-GEN) was employed to iteratively refine sampling of defect configurations and improve force prediction accuracy. The final deep potential model demonstrates close agreement with DFT, achieving root-mean-square errors of 0.42meV/atom for energy and 49meV/? for force. Molecular dynamics (MD) simulations using the developed potential significantly outperform DFT in efficiency, reaching 34 steps per second at 1300K. Structural evaluation shows that the model reproduces radial distribution characteristics and suppresses high-frequency atomic fluctuations at elevated temperatures. Nano-indentation simulations further reveal temperature-accelerated bond breaking, local structural disorder and expansion of subsurface damage regions. The proposed potential enables reliable atomistic modelling of SiC under severe thermomechanical conditions, providing support for understanding high-temperature damage evolution and extending machinelearning- based simulation of complex processing environments.