Using the recurrent neural networks to predict the results of classical molecular dynamics modeling of crystalline solids

Abstract

In this article, on an example of modeling the equilibrium state of crystalline copper, the possibility of predicting the atomic configuration at the next time step using data on atomic configurations at several previous time steps by a bidirectional recurrent neural network trained on classical molecular dynamics data is shown. A network with a minimum number of hidden layers (the number of layers is three) is used, which corresponds to the required number of nodes for calculating velocity and acceleration when integrating particle motion equations using the finite difference method. The figure of merit of the simulation is based on a comparison of the dynamic structural factor of the system, calculated from the initial molecular dynamic trajectories, and the one calculated from the molecular trajectories predicted by the recurrent neural network. The graphs of the radial distribution function, as well as the histograms of the distribution of the first-order differences of coordinates, constructed from the initial data on the arrangement of atoms and from the data predicted by the neural network, show a good match. Both the graphs of the radial distribution function and the histograms of the distribution of the first-order coordinate differences do not undergo significant changes during the simulation of the equilibrium state (180 ps), which indicates an adequate representation of statistical patterns in molecular trajectories predicted by a recurrent neural network.