Time-series prediction of velocity trajectories using generative AI for estimating molecular properties

As the high computational cost of molecular dynamics (MD) simulations is a significant issue, it is important to investigate the applicability and performance of Generative Adversarial Networks (GANs), known as one of the machine learning approaches. In the present study, molecular information for liquid argon (Ar) systems at various temperatures and densities in equilibrium states was employed as an example of prediction via GANs. Time series data of molecular velocities calculated using MD were extracted as training samples for the GANs, and the predictive capability of the GANs was evaluated by examining the reproducibility of sample distributions and self-diffusion coefficient from different calculation methods. It was demonstrated that the samples of molecular velocity generated by the GANs replicated the MD simulation distributions. Mean squared displacements (MSD) and velocity autocorrelation functions (VACF) calculated from the generated velocity trajectories via GANs showed qualitative reproducibility of those calculated by the MD simulations. Regarding the self-diffusion coefficient, we found that the GANs reproducibility of the MD results tends to decline under conditions of low density and high temperature. Furthermore, it was found that the self-diffusion coefficients predicted using GANs did not significantly depend on the calculation methods used for the self-diffusion coefficient.