Learning physical states of bulk crystalline materials from atomic trajectories in molecular dynamics simulation

Melting of crystalline material is a common physical phenomenon, yet it remains elusive owing to the diversity in physical pictures. In this work, we proposed a deep learning architecture to learn the physical states (solid- or liquid-phase) from the atomic trajectories of the bulk crystalline materials with four typical lattice types. The method has ultra-high accuracy (higher than 95%) for the classification of solid-liquid atoms during the phase transition process and is almost insensitive to temperature. The atomic physical states are identified from atomic behaviors without considering any characteristic threshold parameter, which yet is necessary for the classical methods. The phase transition of bulk crystalline materials can be correctly predicted by learning from the atomic behaviors of different materials, which confirms the close correlation between atomic behaviors and atomic physical states. These evidences forecast that there should be a more general undiscovered physical quantity implicated in the atomic behaviors and elucidate the nature of bulk crystalline melting.