Accelerated Atomic Data Production in Ab Initio Molecular Dynamics with Recurrent Neural Network for Materials Research

Abstract

Ab initio molecular dynamics (AIMD) is a versatile and reliable computational approach to atomic-scale materials science. However, due to the expensive computational cost on the first-principles calculation at each time step, the temporal and spatial scales are significantly limited, hindering its broader applications. Therefore, to accelerate the simulation clock of AIMD, atomic data production in AIMD using a recurrent neural network (RNN) is studied in this research. We demonstrate the feasibility of incorporating RNN-predicted time steps in AIMD, while maintaining its accuracy. The RNN models, which are trained using AIMD simulation results, directly predict atomic velocities and positions of Si atoms, reducing errors by decoupling the position and velocity update procedures from the Newtonian mechanics. Not only the predicted atomic data but also material properties calculated using the predicted data, such as the radial distribution function, temperature, velocity autocorrelation function, phonon density of states, and heat capacity, exhibit excellent agreements with the ground-truth AIMD calculations. Since the RNN prediction is much faster than the first-principles calculation of AIMD, this approach is expected to effectively accelerate AIMD, contributing to computational materials research.