Molecular dynamics (MD) simulation is a powerful computational method to observe molecular behavior. Although the detection of molecular behavior that characterizes systems is an important task in the study of MD, it is typically difficult and depends on human expert knowledge. Therefore, we propose a novel analysis scheme for MD data using deep neural networks. A key aspect of our scheme is the estimation of statistical distances between different ensembles that are probability distributions over the possible states of systems. This allows us to build low-dimensional embeddings of ensembles to visualize differences between systems in a compact metric space. Furthermore, the molecular behavior that contributes to the differences between systems can also be detected using the trained function of deep neural networks. The applicability of our scheme is demonstrated using three types of MD data. Our scheme could be a powerful tool to clarify the underlying physics in the molecular systems.
ASJC Scopus subject areas
- Materials Science(all)