Molecular simulation has already become a powerful tool for investigating life principles at the molecular level. The past 50 years witnessed that molecular simulation has enabled the quantitative characterization of both kinetic and thermodynamic properties of complicated molecular processes, such as protein folding and conformational changes. In recent years, the application of machine learning algorithms represented by deep learning has further advanced the development of molecular simulation. This work provides a review on machine learning methods in molecular simulation, focusing on the important progress made by machine learning algorithms in improving the accuracy of molecular force fields, the efficiency of molecular simulation conformation sampling, and also the processing of high-dimensional simulation data. On this basis, this review gives an outlook for future research based on machine learning techniques to further overcome the accuracy and effciency bottleneck of molecular simulation, expand the scope of molecular simulation, and realize the integration of computational simulation and experimental results.