Fine tuning machine learning potentials for searching the structures of charge-neutral Cu, Ag, and Au clusters

The structure of a cluster plays a decisive role in determining its physical and chemical properties. However, as cluster size increases, the number of possible isomers grows exponentially, and first-principles calculations become computationally demanding, posing significant challenges for structure prediction. To address this challenge, we examined an efficient method for searching low-energy structures of atomic clusters integrating machine learning interatomic potentials (MLIPs) and a Comprehensive Genetic Algorithm (CGA). We constructed training datasets for Cu, Ag, and Au clusters and trained the Orb-v2 potential using two strategies: fine-tuning a pre-trained model and training from scratch. Our benchmarking results demonstrate that the fine-tuned model achieves significantly lower energy prediction mean absolute errors (5 ∼ 7 meV/atom) compared to the pre-trained model (36 ∼ 368 meV/atom). Compared to training from scratch, fine tuning a pre-trained model requires 50% less epochs to train and demonstrates better accuracy. Integrating the fine-tuned MLIP with CGA enables efficient and effective searching for low-energy structures of atomic clusters, successfully reproducing known low-energy configurations of small clusters with less than 55 atoms.