05. Micro- and Nano-Scale Transport, MEMS

Molecular dynamics (MD) simulations are frequently used for the investigation of thermal properties and phenomena at the nanometer scales. However, the recent research target has become more complex and larger in scale, which has led to a significant increase in computational cost of MD simulations. To solve the aforementioned problem, machine learning-based approaches have been studied from multiple perspectives for the prediction of thermal properties in various conditions and systems. Generative AI, in particular Generative Adversarial Networks (GAN), is attracting attention due to its performance. Nevertheless, the related works are still insufficient to show its generic applicability, and further investigation is required regarding the scope of application. In the present study, MD simulations and prediction using GAN are implemented for the thermal equilibrium Ar system in order to investigate how to apply the GAN to MD simulations. The time-series trajectory of molecular velocities is extracted from the MD simulation, which were used as the training samples in the calculation of GAN. The time-series transition of temperature and the mean squared displacement (MSD) were calculated through molecular trajectories predicted by the GAN. Furthermore, self-diffusion coefficient was calculated from MSD predicted by the GAN in comparison with that calculated from MD samples. We discussed the effects of various hyperparameters on the prediction accuracy of the self-diffusion coefficients under various conditions.