(661g) Predicting the Shapes of Small Gold Nanoparticles: Force Fields Vs. Density-Functional Theory

Gold nanocrystals are widely used as antibacterial agents, catalysts, in medical applications, and for chemical sensing, where the shape plays a crucial role in performance. While solution-phase synthesis provides a pathway to create shape-controlled nanocrystals, predicting and understanding the equilibrium shapes remains a major theoretical challenge. Quantum chemical methods such as density-functional theory (DFT) offer high accuracy but are limited due to the high computational cost. In contrast, empirical force fields, such as the Embedded Atom Method (EAM), are more efficient, but often yield contradictory shape predictions. Recently, the development of machine learning force fields (MLFFs) offers a promising compromise, presenting DFT-like accuracy with reduced cost, yet their reliability remains dependent on the training dataset and domain.

In this study, we investigated the shapes and melting of small gold nanoparticles with 55 and 147 atoms - both known as magic numbers for icosahedral symmetry, using parallel tempering molecular dynamics (PTMD) simulations based on four different force fields. We used DFT to reoptimize the force-field results, as well as to assess the energies of shapes published in three other studies based on force fields. To classify the diverse gold nanocrystal structures in our studies, we employed unsupervised machine learning techniques: K-means clustering and the Gaussian mixture model, based on atomic environment signatures derived from common neighbor analysis (CNA). Our findings reveal substantial discrepancies in predicting minimum-energy structures and thermodynamic shape distributions across force fields, emphasizing the need for improved force field development. This study advances our understanding of shape evolution in nanoscale gold and the limitations of current theoretical models in capturing realistic nanocrystal behavior.