Abstract

Prescreening candidate structures with reliable classical potentials is an effective way to accelerate ab initio ground state searches. Given the growing popularity of machine learning force fields, surprisingly little work has been dedicated to quantifying their advantages over traditional potentials in global structure optimizations. In this study, we have developed a neural network (NN) model and systematically benchmarked it against a commonly used Gupta potential and an embedded atom model in the search for stable Au-N clusters (30 <= N <= 80). An efficient simultaneous optimization of clusters in the full size range was achieved with our recently introduced multitribe evolutionary algorithm. Density functional theory (DFT) evaluations of candidate configurations identified with the three classical models revealed that the NN structures were lower in energy by at least 10 meV/atom for 30 of the 51 sizes. We also demonstrated that DFT evaluation of all NN-relaxed structures during evolutionary searches resulted in finding even more stable configurations, which highlights the need for further improvement of the NN accuracy to avoid excessive DFT calculations. Overall, the global searches produced putative ground states with matching or lower DFT energies compared to all previously reported Au clusters with 30-80 atoms.