Convolutional Recurrent Autoencoder for Molecular-Continuum Coupling

Abstract

Molecular-continuum coupled flow simulations are used in many applications to build a bridge across spatial or temporal scales. Hence, they allow to investigate effects beyond flow scenarios modeled by any single-scale method alone, such as a discrete particle system or a partial differential equation solver. On the particle side of the coupling, often molecular dynamics (MD) is used to obtain trajectories based on pairwise molecule interaction potentials. However, since MD is computationally expensive and macroscopic flow quantities sampled from MD systems often highly fluctuate due to thermal noise, the applicability of molecular-continuum methods is limited. If machine learning (ML) methods can learn and predict MD based flow data, then this can be used as a noise filter or even to replace MD computations, both of which can generate tremendous speed-up of molecular-continuum simulations, enabling emerging applications on the horizon.

In this paper, we develop an advanced hybrid ML model for MD data in the context of coupled molecular-continuum flow simulations: A convolutional autoencoder deals with the spatial extent of the flow data, while a recurrent neural network is used to capture its temporal correlation. We use the open source coupling tool MaMiCo to generate MD datasets for ML training and implement the hybrid model as a PyTorch-based filtering module for MaMiCo. It is trained with real MD data from different flow scenarios including a Couette flow validation setup and a three-dimensional vortex street. Our results show that the hybrid model is able to learn and predict smooth flow quantities, even for very noisy MD input data. We furthermore demonstrate that also the more complex vortex street flow data can accurately be reproduced by the ML module.