Accelerating the prediction of large carbon clusters via structure search: Evaluation of machine-learning and classical potentials

From as small as single carbon dimers up to giant fullerenes or amorphous nanometer-sized particles, the large family of carbon nanoclusters holds a complex structural variability that increases with cluster size. Capturing this variability and predicting stable allotropes remains a challenging modelling task, crucial to advance technological applications of these materials. While small cluster sizes are traditionally investigated with first-principles methods, a comprehensive study spanning larger sizes calls for a computationally efficient alternative. Here, we combine the stochastic ab initio random structure search algorithm (AIRSS) with geometry optimisations based on interatomic potentials to systematically predict the structure of carbon clusters spanning a wide range of sizes. We first test the transferability and predictive capability of seven widely used carbon potentials, including classical and machine-learning potentials. Results are compared against an analogous cluster dataset generated via AIRSS combined with density functional theory optimizations. The best performing potential, GAP-20, is then employed to predict larger clusters in the nanometer scale, overcoming the computational limits of first-principles approaches. Our complete cluster dataset describes the evolution of topological properties with cluster size, capturing the complex variability of the carbon cluster family. As such, the dataset includes ordered and disordered structures, reproducing well-known clusters, like fullerenes, and predicting novel isomers.