Deep potential for a face-centered cubic Cu system at finite temperatures

The state-of-the-art method generating potential functions used in molecular dynamics is based on machine learning with neural networks, which is critical for molecular dynamics simulation. This method provides an efficient way for fitting multi-variable nonlinear functions, attracting extensive attention in recent years. Generally, the quality of potentials fitted by neural networks is heavily affected by training datasets and the training process and could be ensured by comprehensively verificating the model accuracy. In this study, we obtained the neural network potential of face-centered cubic (FCC) Cu with the most accurate and adequate training datasets from first-principle calculations and the training process performed by Deep Potential Molecular Dynamics (DeePMD). This potential could not only succeed in reproductions of the variety of properties of Cu at 0 K, but also have a good performance at finite temperatures, such as predicting elastic constants and the melting point. Moreover, our potential has a better generalization capacity to predict the grain boundary energy without including extra datasets about grain boundary structures. These results support the applicability of the method under more practical conditions.