The conformational stabilities and distributions of different metastable states do important roles in many physical or chemical processes especially when they are related to thermodynamic properties. However, driving configurational transition using traditional computational methods is nearly impossible as they are separated by high energy barriers. Moreover, this kind of conformational information is described by the underlying free energy surfaces (FESs), whose dimensionality is typically too high to understand, so it becomes another hindrance to theoretical study. Therefore, it needs to use some enhanced sampling techniques such as metadynamics to calculate conformational FES. To drive some proper low-dimensional collective variables which are required for these techniques, we introduced an approach based on using an autoencoder neural network which is trained iteratively on a dataset without supervised labels. Both an explicit mapping of high dimensional data and an out of sample projection for new snapshot of the molecular simulation become feasible, thus the obtaining reliable FES could be relatively easy and fast. For a demonstration, the conformational FES of cyclooctane was calculated which serves as a benchmark for conformational rare event problems. The free-energy difference among the canonical conformations were computed while FES preserve the essential symmetries.

Conformational free-energy surface of cyclooctane in the collective variables space from autoencoder neural network

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.