The representative structure of graphene oxide nanoflakes from machine learning

Abstract

In this paper we revisit the structure of graphene oxide, and determine the pure and truly representative structures for graphene nanoflakes using machine learning. Using 20 396 random configurations relaxed at the electronic structure level, we observe the presence of hydroxyl, ether, double bonds, aliphatic (cyclohexane) disruption, defects and significant out-of-plane distortions that go beyond the Lerf–Klinowski model. Based on an diverse list of 224 chemical, structural and topological features we identify 25 archetypal ‘pure’ graphene oxide structures which capture all of the complexity and diversity of the entire data set; and three prototypes that are the truly representative averages in 224-dimensional space. Together these 28 structures, which are shown to be largely robust against changes in thermochemical conditions modeled using ab initio thermodynamics, can be downloaded and used collectively as a small data set for with a fraction of the computational cost in future work, or independently as an exemplar of graphene oxide with the required oxidation.