Gaussian Process Approach to Constructing Transferable Force Fields for Thiolate-Protected Gold Nanoclusters.

Gold nanoparticles can exhibit unique physical and chemical properties, such as plasmon resonances or photoluminescence. These nanoparticles have many atoms, which leads to high computational costs for density functional theory (DFT) calculations. In this work, we used the FLARE++ (fast learning of atomistic rare events) code and incorporated an active learning algorithm to construct force fields for gold thiolate-protected nanoclusters. We started training the force field using Au(SCH) as the initial structure and then applied the trained force field to perform molecular dynamics (MD) simulations. We then validated the machine learning force field using different types of gold nanoclusters as testing models. The test results were integrated into the existing database and retrained again. The final force fields show success in predicting energies for nanoclusters not only in the training database but also outside the database. These tests revealed that the force field has achieved quantum mechanical level accuracy in some key performance metrics.