Molecular dynamics (MD) is a simulation method wherein the Newton’s equation of motion is solved for all the atomic degrees of freedom. It has enabled prediction of mechanical and dynamical properties of vast range of molecular assemblies and can treat large-scale systems by parallelization. However, as the time scale that MD simulations are limited by the clock cycle of processors, time scales are typically limited up to the order of 10^10 time steps, which corresponds to microseconds for all-atom potentials. Recently, graph neural network has been found to be capable of predicting the long-time evolution of the dynamics of glassy liquid that relaxes very slowly, solely from the initial particle positions (V. Bapst et al., Nat Phys. 16, 448 (2020) ). Motivated by their work, I will present several showcase calculation, and discuss possibilities of further development for extending the power of machine learning.