Construction of a neural network potential for SiC and its application in uniaxial tension simulations

Currently, neural networks have become an attractive tool for achieving highly quantitative accuracy without time-consuming fitting. In this study, we constructed a neural network potential (NNP) for silicon carbide (SiC) by employing neural networks to study the elastic response. The results showed that, compared to existing empirical potential, the NNP for SiC demonstrated greater accuracy in calculating various material properties. Additionally, this potential successfully reproduced the tension-strain behavior of monocrystalline SiC. In describing polycrystalline SiC, there is a limit to the transferability of the NNP, particularly in terms of its dynamical response. This defect could be improved by introducing additional corresponding training samples. This study not only provides a reliable SiC potential, but also contributes to assessing the applicability and limitation of NNPs. In addition, we place emphasis on the significance of the composition of the training set and the optimization of training weights in NNP training.