Accelerating the search for carbon cluster isomers via machine-learning potential

Abstract Cage-like isomers of C22 clusters are explored with machine-learning potential GAP-20 and density-functional theory (DFT) calculations. Using the GAP-20 to approximate the energy landscape helps reduce the search time and provides superior starting structures for subsequent DFT geometry optimization. However, the Jahn-Teller distortion is not predicted by the GAP-20. The relative GAP-20 energies are overestimated, and the vibrational modes/frequencies are poorly characterized. The thermodynamic and chemical stabilities of the cages are discussed, and the simulated infrared spectra are provided.