Revisiting the structural dynamics of gold clusters by machine learning force field

Abstract

The structural evolution of gold clusters has been investigated by numerous density functional theory (DFT) studies. However, due to the slow computational efficiency of DFT, these studies tend to be scattered and lack systematicness. We have developed a robust machine learning force field (MLFF) of gold. The accuracy and robustness of the MLFF were validated by comparing DFT results. By integrating the highly efficient MLFF, which is about 1000,000 times faster than DFT calculations, with the CALYPSO global search method, we systematically explored Au_n clusters spanning a wide range of sizes (n = 2–55) and uncovered several key issues: (i) revealing the critical transition points from planar to 3D structures (n = 14) and from cage-like to core-shell structures (n > 26); (ii) discovering new stable cluster structures; (iii) conducting an in-depth analysis of the core-shell model. This study shows that MLFF can be used to study complex structural systems like clusters and address systematic issues related to larger clusters. It also indicates the potential of MLFF in tackling more complex problems, including mixed and ligand-protected clusters.