Deep-learning coarse-grained particle method for crystalline metallic materials

A novel coarse-grained method for crystalline metallic materials is developed based on deep learning. By proposing a methodology to decompose the potential energy of crystal atom clusters into interior and exterior parts, we establish the coarse-grained particle model. Two independent neural networks are innovatively employed to respectively predict the interior energy and interaction between particles, as both two parts are complex and difficult to be expressed explicitly. Typical examples and interpretability analysis verify that the networks can accurately learn atomic information. The proposed method is implemented in both a single crystal copper model and a high-entropy alloy FeNiCrCoCu model. Accuracy comparable with molecular dynamics simulations is obtained. Efficiency analysis demonstrates the great potential of our method for multiscale simulation.