Active learning-driven global search for neutral gold clusters via neural network potential

Abstract

The structural prediction of metal nanoclusters is hindered by the extremely complex potential energy surface and the prohibitive cost of first-principles calculations. Here, we develop an efficient structure-prediction framework that tightly integrates machine-learning interatomic potentials with global optimization. Neural network atomic potentials are iteratively trained to achieve density-functional-theory accuracy and coupled with a genetic algorithm to enable reliable exploration of complex energy landscapes. As a stringent benchmark, the framework is applied to neutral Au_n clusters (n = 30–45), where it robustly identifies low-energy structures at an affordable computational cost and reveals a non-monotonic structural evolution from hollow cage-like motifs to multi-core-cage building blocks over a critical size range. Notably, this transition exhibits pronounced differences from that of the corresponding anionic clusters, highlighting the potential of the proposed active-learning workflow as an extensible strategy for investigating metal clusters with complex electronic structures.