Structural prediction of carbon cluster isomers with machine-learning potential

Abstract

Structural prediction of low-energy isomers of carbon twelve-atom clusters is carried out using the recently developed machine-learning potential GAP-20. The GAP-20 agrees with density-functional theory calculations regarding geometric structures and average C-C bond lengths for most isomers. However, the GAP-20 substantially lowers the energies of cage-like structures, resulting in a wrong ground state. A comparison of the cohesive energies with the density-functional theory points out that the GAP-20 only gives good results for monocyclic rings. Two multicyclic rings appear as new low-energy isomers, which have yet to be discovered in previous research.