Functionalized graphene is a versatile medium that has been proposed for the efficient separation of toxic cations from aqueous solutions.1 Recent studies suggest that precisely doped polyoxometalate or metal clusters can be synthesized on graphene to tune their electronic properties and the interfacial solvent structure.2 How these modifications change the separation properties and how one can model and simulate these phenomena constitutes an interesting and challenging subject for theoretical simulations. A meaningful model for such systems often requires large (~nanometer scale) unit cells that can be accessed by classical molecular dynamics (CMD) simulations. However, these often suffer from the uncertain accuracy of force fields. Ab initio simulations are more reliable but are considerably expensive. To combine the advantages of both methods, we apply state-of-the-art machine learning algorithms to accelerate the development of reliable force fields. First, based on ab initio simulation data, we fitted a set of accurate neural network potentials (NNPs)3 for systems of functionalized graphene, water, and metal species. Then, using these NNPs as energetic rank, the artificial bee colony algorithm,4 a popular global optimization method, combined with reinforcement learning and clustering methods, was applied to search for the energetically most favorable distributions of oxide groups on graphene as well as the most stable structures of metallic clusters landed on functionalized graphene. The NNPs were also applied for molecular dynamics simulations to study how different types of affect surface properties towards separations of cations. Our study presents a set of robust methods towards the applications of machine learning tools to chemical problems with a focus on interfacial and separation properties facilitated by novel graphene-supported metal clusters.

Application of machine learning in separation processes using graphene-supported metallic clusters

2019 Fall National Meeting

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