Recent development of large-scale atomic simulation techniques based on machine learning has brought a great promise in chemistry. These simulations are featured by both high speed and high accuracy. This review outlined recent development on three key aspects of atomic simulation based on machine learning potential, i.e., machine learning models and structure descriptors, generation of global potential energy surface training sets, and automatic training of potential functions based on active learning. It is indicated that the designed structure descriptor and feedforward neural network model are suitable for generating a highly complex global potential energy surface. In addition, the applications of LASP software in material and reaction simulations were also selected to illustrate how ML-based atomic simulation could assist the discovery of novel materials and reactions.